AI/ML Based Interoperable Anomaly Detection in Advanced Manufacturing Using MTConnect-Derived Multi-Machine Data

AI/ML Based Interoperable Anomaly Detection in Advanced Manufacturing Using MTConnect-Derived Multi-Machine Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article AI/ML Based Interoperable Anomaly Detection in Advanced Manufacturing Using MTConnect-Derived Multi-Machine Data Harshkumar Kiritbhai Parmar, Shivakumar Raman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9358105/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 5 You are reading this latest preprint version Abstract Early detection of abnormal process behavior is critical in smart manufacturing, as it contributes to reduced downtime, improved process reliability, and prevention of quality-related losses. However, many anomaly-detection studies in manufacturing still rely on isolated sensor datasets or single-machine setups, limiting their extensibility across different machines and production environments. This leaves a gap in leveraging standardized MTConnect-derived data for anomaly detection in a more interoperable multi-machine setting. To address this gap, this study develops an AI/ML-based anomaly-detection framework using MTConnect-derived data from multiple milling machine groups. Anomaly labels are assigned at the experiment level using documented run remarks, and statistical features are extracted from machine, spindle, load, vibration, energy, and runtime variables. The study first examines a single-machine baseline and then expands to a combined dataset built from two feature-compatible machine groups, imi_vm20i and imi_vmx30ui. Logistic Regression, Random Forest, and XGBoost are used for anomaly classification. In 5-fold stratified cross-validation on the combined dataset, Random Forest achieved the highest mean accuracy (0.9455 ± 0.0727), Logistic Regression achieved the highest recall (0.9000 ± 0.2000) and F1-score (0.8000 ± 0.2667), and XGBoost achieved the highest ROC-AUC (0.9222 ± 0.1556). Feature-importance analysis showed that vibration-related variables were the most influential in the full model, while an ablation study showed that Power/Energy/Runtime features performed best as a standalone group. A separate robustness experiment on the structurally different tmf_vf10 machine group showed a clear drop in performance, highlighting the importance of feature compatibility for interoperable anomaly detection. Overall, the results show that MTConnect-derived data can support anomaly detection across compatible machine groups, while also revealing the limits of transferability when machine-data structure changes. MTConnect anomaly detection smart manufacturing machine learning milling interoperable manufacturing data predictive monitoring Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Modern manufacturing systems increasingly depend on the ability to detect abnormal process behavior early. In milling operations, issues such as excessive vibration, clamp failure, tool breakage, or unstable cutting conditions can quickly affect process stability, product quality, and machine health. If these problems are not identified in time, they can lead to scrap, downtime, and unnecessary cost. Consequently, anomaly detection has become an important part of smart manufacturing and real-time shop-floor monitoring [ 1 ]. A large portion of the existing work in this area relies on isolated sensor datasets, machine-specific signals, or single-machine data pipelines. While those studies are useful, they often remain tied to a particular setup and are difficult to extend across different machines or production environments. In practice, manufacturing systems need approaches that are not only accurate, but also easier to scale and reuse across equipment. This is where standardized machine data becomes especially important. MTConnect is valuable in this context because it provides a common structure for machine-generated data. Rather than relying on proprietary tags, controller-specific naming conventions, or ad hoc data pipelines, MTConnect provides a standardized representation of machine, spindle, load, vibration, and runtime information [ 2 ]. Prior studies have shown that MTConnect can support interoperable monitoring, asset knowledge automation, and broader digital-manufacturing integration across heterogeneous systems [ 3 ]-[ 5 ]. Recent work has also examined the role of MTConnect in enabling communication within and across manufacturing enterprises [ 22 ]. Although anomaly detection can still be performed without MTConnect, such approaches usually require more custom preprocessing, manual data mapping, and machine-specific integration. Accordingly, MTConnect is not a prerequisite for anomaly detection, but it substantially improves the practicality of building frameworks that are reusable, interpretable, and scalable across machines. Even with the growing use of MTConnect for machine connectivity and interoperability, its role in anomaly detection across multiple machine groups remains limited. Much of the published work either focuses on MTConnect-enabled monitoring and data integration or develops anomaly-detection methods in single-machine or application-specific settings [ 1 ], [ 3 ]-[ 5 ]. Limited work exists that integrates these two directions within a unified framework for anomaly detection using MTConnect-derived data across multiple machine groups, while also examining how performance changes when machine-data structures are not fully compatible. Problem Detecting abnormal machining behavior early is important for reducing downtime, quality loss, and process instability. However, doing this consistently across different machines remains difficult because most existing approaches are developed and validated on single-machine or application-specific datasets, making their scalability and reusability unclear. Gap Although MTConnect has been widely adopted for machine connectivity and data standardization, its potential for enabling anomaly detection across multiple machine groups has not been systematically examined. In particular, no prior study has evaluated how MTConnect-derived anomaly-detection performance changes when feature compatibility varies across machines, a scenario that is common in real manufacturing environments. Solution To address this gap, this paper develops an AI/ML-based anomaly-detection framework using MTConnect-derived multi-machine data. Unlike prior work that focuses either on MTConnect-based monitoring or on single-machine anomaly detection, this framework integrates experiment-level labeling, statistical feature engineering, model comparison, interpretability analysis, and robustness evaluation under both feature-compatible and feature-heterogeneous machine-data conditions. The explicit evaluation of transferability limits across structurally different machine groups distinguishes this work from existing studies. The main contributions of this work are summarized as follows: An interoperable anomaly-detection framework is developed using MTConnect-derived manufacturing data for milling-process monitoring. The framework is evaluated across two compatible machine groups using a common set of standardized features, enabling multi-machine anomaly detection rather than single-machine analysis alone. Interpretability is improved through both feature-importance analysis and feature-group ablation, showing how vibration, power/energy/runtime, axis-load, and process/spindle variables contribute differently to anomaly detection. A separate robustness experiment on a structurally different machine group highlights the transferability limits of anomaly detection when machine-data feature compatibility is reduced. The study provides empirical evidence that standardized machine data supports but does not automatically guarantee cross-machine anomaly-detection transferability, offering practical guidance for deploying interoperable monitoring systems in heterogeneous manufacturing environments. The remainder of this study is organized as follows. Section 2 reviews related work on MTConnect-based data integration, manufacturing anomaly detection, and heterogeneous industrial-data learning. Section 3 describes the dataset, anomaly-labeling process, and feature engineering. Section 4 presents the modeling approach and evaluation strategy. Section 5 reports and discusses the results, including cross-validation, interpretability, ablation, and robustness analysis. Section 6 concludes this study and outlines future directions. 2. Related Work Prior work related to this study can be grouped into three main areas: MTConnect-based machine-data integration, machine-learning methods for anomaly detection in manufacturing, and learning under heterogeneous industrial data conditions. These areas come together in this paper because the objective is not only to detect anomalies, but to do so using standardized machine data in a way that can extend beyond a single machine. Early MTConnect-related work mainly focused on machine connectivity, monitoring, and interoperability Liao [ 6 ]. showed that MTConnect data can support a self-aware machine framework for condition monitoring and failure-related reasoning. Venkatesh [ 7 ]. demonstrated how MTConnect can be used to automate asset knowledge and machine-status interpretation. In a broader digital-manufacturing context, Helu et al [ 8 ]. emphasized the value of standards-based architectures for integrating heterogeneous manufacturing systems and supporting the digital thread. More recently, metrology-integrated digital twin frameworks have also been explored for advanced manufacturing applications [ 25 ]. Beyond review-level findings, several recent studies have addressed anomaly detection and condition monitoring in machining through different technical approaches. Liu [ 14 ] proposed a digital-twin-based framework for real-time tool condition monitoring, using simulation-driven anomaly detection tied to a specific machine setup. Xie [ 15 ] developed an unsupervised method combining multi-scale prototype augmentation with multi-sensor data for manufacturing process monitoring, though their approach was designed for within-process anomaly detection rather than cross-machine evaluation. Orabi [ 16 ] introduced an adaptive adversarial transformer-based model for smart manufacturing anomaly detection, demonstrating strong performance but within a single application context. In the area of tool-condition monitoring specifically, Schueller and Saldaña [ 17 ] examined the generalizability of ensemble machine learning models across cutting conditions, providing relevant insight into transferability challenges, though without using standardized data protocols such as MTConnect. Mohanraj [ 18 ] combined wavelet features with machine learning for end-milling tool monitoring, focusing on signal-processing techniques rather than cross-machine interoperability. While these studies advance the state of the art in their respective directions, none of them combines standardized MTConnect-derived data with multi-machine anomaly detection and explicit evaluation of transferability under heterogeneous feature conditions. Recent work has also emphasized the importance of explainability in manufacturing ML systems. Johannssen. [ 19 ] discussed explainable AI as a requirement for trustworthy intelligent process monitoring, while Gross [ 20 ] demonstrated explainable machine learning for milling quality prediction. These studies reinforce the value of interpretability in manufacturing contexts, which the present work addresses through feature-importance analysis and feature-group ablation. Transferability is another open issue. Even when anomaly detection performs well on one dataset, that does not mean the same model will transfer effectively to another machine or operating environment. Michau and Fink [ 13 ] discussed unsupervised transfer learning for anomaly detection and highlighted the broader challenge of operating-condition transfer when data structure changes. Transfer learning has also been applied to defect analysis in additive manufacturing contexts [ 23 ], reinforcing the broader relevance of cross-domain transferability in manufacturing quality applications. Table 1 Comparison of related studies and positioning of the present work. Study Main Focus Data Basis Single / Multi Machine MTConnect-Centered Interpretability Heterogeneous Validation Liao.[ 6 ] Self-aware machine monitoring MTConnect data Single-machine oriented Yes Limited No Venkatesh et al. [ 7 ] Asset knowledge automation MTConnect data Not focused on anomaly detection Yes Limited No Helu et al. [ 8 ] Digital thread / integration Standards-based manufacturing data System-level Partly No No Review / anomaly detection studies [ 1 ], [ 9 ], [ 10 ] Manufacturing anomaly detection and ML Sensor / manufacturing datasets Mostly single or application-specific No / limited Varies Rare Heterogeneous industrial-data studies [ 11 ]–[ 13 ] Heterogeneity, imbalance, transferability Industrial / anomaly datasets Multi-source or cross-condition No Limited Yes This work AI/ML-based interoperable anomaly detection MTConnect-derived multi-machine data Multi-machine Yes Yes Yes Taken together, the reviewed studies show strong progress in MTConnect-based interoperability, machine monitoring, and manufacturing anomaly detection. However, a clear gap remains in combining these directions into a single framework for AI/ML-based anomaly detection using MTConnect-derived data across multiple machine groups, while also examining what happens when feature compatibility changes across machines. This paper addresses that gap by developing a multi-machine anomaly-detection framework using MTConnect-derived milling data, evaluating it on two compatible machine groups, and then testing robustness on a structurally different machine group [ 1 ], [ 6 ]–[ 13 ]. 3. Dataset and Feature Engineering 3.1 Dataset Overview The dataset used in this study is derived from a publicly available multi-sensor metal-milling anomaly dataset [ 21 ] organized into three machine groups: imi_vm20i, imi_vmx30ui, and tmf_vf10. Each machine group contains multiple experiment folders, where each folder represents one machining run and includes an MTConnect-derived _mtc.csv file together with other supporting files such as labels, audio, video, or NC-code data. Since the focus of this work is interoperable anomaly detection using standardized machine information, only the MTConnect-derived data files were used, and other available modalities (audio, video, NC-code) were not included in the analysis. A key characteristic of the dataset is that the three machine groups are not equally compatible from a feature perspective. The imi_vm20i and imi_vmx30ui groups share a relatively rich and similar MTConnect-derived feature structure, which makes them suitable for combined multi-machine analysis. In contrast, tmf_vf10 contains a smaller and structurally different feature set. This difference is significant because the study is concerned not only with anomaly detection accuracy, but also with how feature compatibility affects transferability across machine groups. Based on this structure, the dataset was used in three stages: Single-machine baseline : imi_vm20i Main compatible multi-machine dataset : imi_vm20i + imi_vmx30ui Heterogeneous-machine robustness dataset : tmf_vf10 Table 2 Dataset composition used in this study. Machine Group Normal Runs Anomalous Runs Total Runs Role in Study imi_vm20i 30 7 37 Single-machine baseline + combined main dataset imi_vmx30ui 17 2 19 Combined main dataset tmf_vf10 8 5 13 Heterogeneous-machine robustness experiment 3.2 Anomaly Label Definition The anomaly labels used in this paper were assigned at the experiment level. Although some folders included label.csv files, those labels did not consistently represent normal-versus-anomalous status across all machine groups. Instead, the final anomaly labels were derived from the documented experiment remarks in the dataset summary, which provided a more consistent and practically meaningful basis for classification. Runs with no abnormal remark were labeled as normal (0). Runs with remarks indicating process abnormalities, such as excessive vibration, clamp failure, tool breakage, residual chip adhesion, chatter, blown insert, or early stopping, were labeled as anomalous (1). This approach reflects how abnormal runs are often identified in real manufacturing environments, where process logs, operator notes, and documented run outcomes may be more reliable than perfectly segmented point-level fault annotations. The use of experiment-level labels also matches the modeling strategy adopted in this paper. Since each machining run is converted into one experiment-level feature vector, the target variable naturally represents whether the run as a whole should be considered normal or anomalous. 3.3 Feature Compatibility Across Machine Groups One of the central design decisions in this work was how to handle feature compatibility across machine groups. For the main multi-machine analysis, only the two compatible machine groups, imi_vm20i and imi_vmx30ui, were combined. These two groups shared a common set of MTConnect-derived variables related to process behavior, spindle motion, axis loads, vibration statistics, energy, and runtime. This shared feature space made it possible to construct a unified dataset without forcing aggressive feature reduction. The tmf_vf10 machine group, however, did not provide the same feature richness. Its MTConnect-derived variables were more limited and followed a different structure, including variables such as cycle time, motion time, spindle load, spindle speed, and axis-related power or motion signals. As a result, tmf_vf10 was not merged directly into the main multi-machine dataset. Instead, it was used separately to test how the proposed framework behaves when applied to a machine group with reduced and structurally different feature availability. This separation is important for the overall contribution of this study. It allows the study to evaluate anomaly detection under both feature-compatible and feature-heterogeneous conditions, which helps clarify the practical limits of interoperable anomaly detection. 3.4 Experiment-Level Statistical Features Each _mtc.csv file was treated as one complete machining experiment. Rather than using raw time-series rows directly for classification, experiment-level statistical features were extracted from each selected MTConnect-derived variable. For each retained variable, four summary statistics were computed: mean, standard deviation, minimum, and maximum. This produced one feature vector per experiment. For example, a variable such as Spindle_Speed contributed the following derived features: Spindle_Speed_mean, Spindle_Speed_std, Spindle_Speed_min, and Spindle_Speed_max. The same procedure was applied to vibration-related, energy-related, load-related, and runtime-related variables. This representation was chosen for practical and methodological reasons. First, it provides a compact way to summarize full-run behavior without depending on identical run lengths across experiments. Second, it improves interpretability by linking anomaly detection to familiar summary characteristics such as variability, extremes, and average operating levels. Third, it makes it easier to compare compatible machine groups without requiring detailed temporal alignment or sequence modeling. For the main combined dataset, only common MTConnect-derived variables shared by imi_vm20i and imi_vmx30ui were retained. For tmf_vf10, the same statistical summarization procedure was used, but only on the smaller machine-specific feature set available in that group. 3.5 Final Datasets Used for Modeling Based on the feature-engineering process, three experiment-level datasets were prepared for modeling: Baseline dataset: 37 runs from imi_vm20i Main combined dataset: 56 runs from imi_vm20i and imi_vmx30ui Robustness dataset: 13 runs from tmf_vf10 The baseline dataset was used to verify that the anomaly-labeling and feature-engineering strategy produced a meaningful anomaly-detection problem. The combined dataset was used as the main dataset for model comparison, cross-validation, feature-importance analysis, and ablation study. The tmf_vf10 dataset was used separately for a robustness experiment to evaluate performance under heterogeneous machine-data conditions. 3.6 Data Cleaning and Preprocessing Before modeling, the experiment-level feature tables were checked for missing values and nonnumeric entries. Variables that became completely empty after numeric conversion were removed, particularly in the tmf_vf10 robustness dataset, where several machine-specific variables did not provide usable numeric observations after aggregation. For the remaining features, missing values were handled using median imputation during model training. At the end of this stage, the study produced clean experiment-level datasets with anomaly labels derived from documented run remarks and features extracted from standardized MTConnect-derived machine data. 4. Methodology 4.1 Modeling Objective The objective of this study was to determine whether experiment-level statistical features extracted from MTConnect-derived manufacturing data can distinguish normal and anomalous milling runs across multiple machine groups. More specifically, the methodology was designed to answer three related questions: Can anomaly detection be performed reliably on a single MTConnect-based machine group? Can the same framework be extended across multiple compatible machine groups using a common feature space? What happens when the framework is applied to a structurally different machine group with reduced feature compatibility? These questions directly shaped the experimental design and the staged evaluation strategy adopted in this study. 4.2 Classification Models Three supervised machine-learning models were selected for anomaly classification: Logistic Regression Random Forest XGBoost These models were selected because they represent different learning behaviors-linear, nonlinear ensemble, and gradient boosting, while remaining well suited to structured tabular manufacturing data. Although anomaly detection is often approached using unsupervised methods such as Isolation Forest or One-Class SVM, a supervised framing was adopted in this study for two reasons. First, experiment-level anomaly labels were available from documented run remarks, making supervised classification both feasible and appropriate. Second, supervised models allow direct evaluation using standard classification metrics such as precision, recall, F1-score, and ROC-AUC, which makes model comparison and interpretability more straightforward. The use of unsupervised methods remains a relevant direction for future work, particularly in settings where labeled anomaly data are not available. Logistic Regression was used as a simple baseline model to test whether anomaly separation could be captured through relatively direct relationships in the engineered feature space. Random Forest was chosen because it is well suited to tabular industrial data, can capture nonlinear interactions, and also, supports interpretability through feature importance. XGBoost was included as a boosting-based model that often performs well on structured datasets and provides a useful comparison to both linear and ensemble-tree methods. All three models were trained using their default hyperparameter configurations as implemented in scikit-learn (Logistic Regression and Random Forest) and the XGBoost Python package. Specifically, Random Forest used 100 estimators with no maximum depth constraint, and XGBoost used a learning rate of 0.3 with a maximum depth of 6 and 100 boosting rounds. No systematic hyperparameter optimization (e.g., grid search or Bayesian tuning) was performed. This decision was deliberate: the primary objective of this study was to evaluate the informativeness of MTConnect-derived features and the effect of machine compatibility on anomaly detection, rather than to maximize model-specific performance through extensive tuning. The potential effect of hyperparameter optimization is acknowledged as a limitation and a direction for future investigation. Together, these models allow the study to compare linear, nonlinear ensemble, and boosting-based approaches within one anomaly-detection framework. 4.3 Preprocessing During Model Training The final experiment-level feature tables were used as model inputs. Missing values in the retained features were handled through median imputation. Logistic Regression was trained on imputed and standardized features because the model is sensitive to feature scale. Random Forest and XGBoost were trained on the imputed feature values directly without standardization. This preprocessing strategy was deliberately kept minimal to ensure that the results reflect differences in feature representation, machine compatibility, and model behavior rather than the effects of extensive model-specific tuning. 4.4 Experimental Design 4.4.1 Experiment 1: Single-Machine Baseline The first experiment used only the imi_vm20i machine group. This baseline was included to validate the overall pipeline, including experiment-level anomaly labeling, MTConnect-derived feature extraction, and model training. A held-out train-test split was employed to verify that the extracted feature space provided sufficient separability between normal and anomalous runs. This experiment was not intended to serve as the main evidence of this study. Instead, it provided a controlled starting point before moving to the stronger multi-machine setting. 4.4.2 Experiment 2: Compatible Multi-Machine Analysis The main experiment of this study used the combined dataset built from imi_vm20i and imi_vmx30ui. These two machine groups were selected because they share a sufficiently similar set of MTConnect-derived variables. Only the variables common to both groups were retained so that the resulting anomaly-detection problem reflected a genuinely shared feature space. A held-out split was first used to obtain an initial benchmark. However, because the combined dataset still contained a limited number of anomaly cases, this split was treated as an exploratory benchmark rather than the main result. The primary evaluation for the combined dataset therefore relied on cross-validation. 4.4.3 Experiment 3: Five-Fold Stratified Cross-Validation For the combined multi-machine dataset, 5-fold stratified cross-validation was used as the main evaluation method. Stratification helped preserve the approximate anomaly-to-normal ratio across folds, which was important because the dataset was imbalanced. This evaluation was used to compare Logistic Regression, Random Forest, and XGBoost in terms of mean performance and variability across folds. Compared with a single train-test split, this provided a more reliable estimate of model behavior and formed the basis of the main model-comparison results reported in this study. 4.4.4 Experiment 4: Feature Importance and Ablation To better understand how the models made decisions, Random Forest feature importance was extracted from the combined dataset. This analysis was used to identify which MTConnect-derived variables contributed most strongly to anomaly detection. A feature-group ablation study was also performed using Random Forest and 5-fold stratified cross-validation. The features were divided into four families: Process/Spindle Axis Load Vibration Power/Energy/Runtime Each group was evaluated independently, and an additional All Features configuration was included. This experiment was designed to determine whether certain groups of MTConnect-derived variables were sufficient on their own and whether combining all available features necessarily improved anomaly-detection performance. 4.4.5 Experiment 5: Heterogeneous-Machine Robustness The final experiment used the tmf_vf10 machine group as a robustness test. This group had a reduced and structurally different feature set compared with the two main machine groups. Because only 13 runs were available, a conventional train-test split, or 5-fold cross-validation would have been unstable and highly sensitive to the specific partition. For that reason, Leave-One-Out Cross-Validation (LOOCV) was used. This experiment was not intended as a direct competitor to the main compatible-machine analysis. Instead, it was designed to assess how the proposed framework behaves when the MTConnect-derived feature space changes substantially. In that sense, it serves as a transferability test under heterogeneous machine-data conditions. 4.5 Evaluation Metrics Model performance was evaluated using the following metrics: Accuracy Precision Recall F1-score ROC-AUC Accuracy measures the overall proportion of correctly classified runs. Precision reflects how many predicted anomalies were truly anomalous. Recall indicates how many actual anomalies were correctly detected. F1-score provides a balance between precision and recall and is especially useful when the classes are imbalanced. ROC-AUC measures the ability of a model to rank anomalous runs ahead of normal runs across decision thresholds. For the main combined dataset, all five metrics were reported in the cross-validation analysis. For the tmf_vf10 LOOCV robustness experiment, ROC-AUC was not emphasized because each fold contained only one test sample, making ROC-AUC undefined or unstable in many folds. As a result, the robustness discussion focused mainly on accuracy, precision, recall, and F1-score. 4.6 Interpretation Strategy The methodology in this paper was designed not only to compare predictive performance, but also to explain it. Model-comparison results were used to evaluate classification strength, feature-importance analysis was used to identify influential MTConnect-derived variables, and the ablation study was used to determine which feature families contributed most effectively. The robustness experiment then tested how strongly the framework depends on feature compatibility across machine groups. This interpretation strategy is central to the contribution of this work. The study is not simply asking whether anomaly detection works, but also why it works on compatible machine groups and why it weakens under heterogeneous machine-data conditions. 4.7 Summary of Method In summary, the methodology combines experiment-level anomaly labeling, statistical feature engineering, supervised classification, cross-validation, interpretability analysis, and robustness testing. This design allows the study to move beyond a single-machine proof-of-concept and provide a more realistic assessment of how MTConnect-derived manufacturing data can support interoperable anomaly detection across multiple machine groups. 5. Results and Discussion This section evaluates the proposed framework in a staged way. A single-machine baseline is first used to verify that experiment-level MTConnect-derived statistical features can support anomaly classification. The analysis is then extended to a combined multi-machine dataset built from the two compatible machine groups, imi_vm20i and imi_vmx30ui, where the main model-comparison and cross-validation results are reported. To improve interpretability, feature-importance analysis and feature-group ablation are also performed. Finally, robustness is assessed on the structurally different tmf_vf10 machine group to examine the limits of transferability under heterogeneous feature conditions. 5.1 Single-Machine Baseline The first experiment was conducted on the imi_vm20i machine group (37 runs: 30 normal, 7 anomalous). This baseline was used to confirm that the experiment-level labeling strategy and MTConnect-derived statistical feature representation produced a meaningful anomaly-classification problem. Using the extracted experiment-level features, the initial baseline showed strong class separation between normal and anomalous runs. This result provided an important validation of the overall pipeline, including anomaly-label assignment from documented run remarks, feature generation from MTConnect-derived variables, and supervised model training. However, because this experiment was limited to one machine group and a relatively small number of anomaly cases, it was treated as a proof-of-concept rather than the main evidence of this work. Overall, the single-machine baseline confirmed that experiment-level statistical summaries of MTConnect-derived variables can carry useful anomaly-related information. The next step was therefore to examine whether the same framework could be extended to a compatible multi-machine setting. 5.2 Compatible Multi-Machine Held-Out Split The second experiment used the combined dataset built from the two compatible machine groups, imi_vm20i and imi_vmx30ui. This dataset contained 56 runs, including 47 normal and 9 anomalous runs. Since these two machine groups shared a common MTConnect-derived feature space, they were used to form the main dataset for the core anomaly-detection analysis. A held-out train-test split was first used to obtain an initial benchmark for the combined dataset. The results are summarized in Table 3 . Table 3 Held-out split model performance on the combined imi_vm20i and imi_vmx30ui dataset. Model Accuracy Precision Recall F1-score ROC-AUC Logistic Regression 1.0000 1.0000 1.0000 1.0000 1.0000 Random Forest 0.9286 1.0000 0.5000 0.6667 1.0000 XGBoost 0.9286 1.0000 0.5000 0.6667 0.9583 On this split, Logistic Regression achieved the strongest overall performance, while Random Forest and XGBoost remained highly precise but showed lower recall. These results indicate strong class separability in the engineered feature space for this particular partition. However, because a single held-out split can be sensitive to the specific train-test partition, especially in a dataset with limited anomaly cases, it served only as an initial benchmark. The main evaluation of this study is therefore based on 5-fold stratified cross-validation. 5.3 Five-Fold Stratified Cross-Validation on the Compatible Multi-Machine Dataset To obtain a more robust estimate of model performance, 5-fold stratified cross-validation was performed on the combined compatible-machine dataset. This evaluation was used as the main basis for comparing Logistic Regression, Random Forest, and XGBoost Table 4 Five-fold stratified cross-validation results on the combined imi_vm20i and imi_vmx30ui dataset Model Accuracy (mean ± std) Precision (mean ± std) Recall (mean ± std) F1-score (mean ± std) ROC-AUC (mean ± std) Logistic Regression 0.9091 ± 0.1408 0.8400 ± 0.3200 0.9000 ± 0.2000 0.8000 ± 0.2667 0.9044 ± 0.1196 Random Forest 0.9455 ± 0.0727 0.8000 ± 0.4000 0.7000 ± 0.4000 0.7333 ± 0.3887 0.9167 ± 0.1667 XGBoost 0.9091 ± 0.0575 0.7000 ± 0.4000 0.6000 ± 0.3742 0.6000 ± 0.3266 0.9222 ± 0.1556 The cross-validation results provide a more realistic picture than the earlier held-out split. Random Forest achieved the highest mean accuracy ( 0.9455 ± 0.0727 ), Logistic Regression achieved the highest mean recall ( 0.9000 ± 0.2000 ) and mean F1-score ( 0.8000 ± 0.2667 ), and XGBoost achieved the highest mean ROC-AUC ( 0.9222 ± 0.1556 ). These results indicate that the proposed MTConnect-derived feature space is informative for anomaly detection across the two compatible machine groups, but they also show that different models emphasize different strengths. From a practical perspective, the results suggest that Random Forest offers the strongest average classification accuracy, while Logistic Regression provides a stronger balance for recovering anomaly cases. XGBoost, in contrast, ranks anomalous and normal runs effectively overall, but at the default decision threshold it shows lower recall and F1-score. This comparison highlights the multi-objective nature of the anomaly-detection problem: model selection depends on whether the operational priority is overall accuracy, anomaly sensitivity, or ranking quality. 5.4 Feature Importance Analysis To better understand which MTConnect-derived variables contributed most strongly to anomaly detection, feature importance was extracted from the Random Forest model trained on the combined compatible-machine dataset. The ranked features were dominated by vibration-related variables together with runtime-related measures. Among the most influential variables were Program_Runtime_Seconds_std, Program_Runtime_Seconds_mean, smarms_min, ymarms_min, zmarms_mean, zmarms_min, ymapeak_min, and zmapeak_mean, along with selected spindle and load variables such as Spindle_Speed_mean, Feed_Rate_mean, and Y_Axis_Load_max. To improve interpretation, the features were also grouped into four families: Vibration, Power/Energy/Runtime, Axis Load, and Process/Spindle. The grouped importance analysis showed that Vibration had the largest total contribution (0.6442), followed by Power/Energy/Runtime (0.2060), Axis Load (0.0784), and Process/Spindle (0.0713). These results indicate that anomaly detection in the compatible multi-machine dataset is driven primarily by dynamic process-instability signatures, especially vibration-related behavior, with additional support from runtime and energy-related characteristics. This is physically reasonable because the anomaly labels include events such as excessive vibration, clamp failure, tool breakage, chatter, and early stopping, all of which can affect machine dynamics and process duration. This emphasis on interpretability aligns with recent calls for explainable AI in manufacturing process monitoring [ 19 ], [ 20 ]. 5.5 Feature-Group Ablation Study Although the grouped importance analysis suggested that vibration features dominate the full-model importance structure, this does not necessarily mean that the vibration group alone provides the best standalone classification performance. To evaluate the contribution of each feature family more directly, an ablation study was performed using Random Forest and 5-fold stratified cross-validation. The following feature groups were evaluated independently: Vibration, Power/Energy/Runtime, Axis Load, Process/Spindle, and All Features. Table 5 Ablation-study results using Random Forest and 5-fold cross-validation. Feature Group Num. Features Accuracy (mean ± std) Precision (mean ± std) Recall (mean ± std) F1-score (mean ± std) ROC-AUC (mean ± std) Power/Energy/Runtime 12 0.9636 ± 0.0727 1.0000 ± 0.0000 0.8000 ± 0.4000 0.8667 ± 0.2667 0.9333 ± 0.1333 Vibration 64 0.9455 ± 0.0727 0.8000 ± 0.4000 0.7000 ± 0.4000 0.7333 ± 0.3887 0.9222 ± 0.1556 All Features 80 0.9455 ± 0.0727 0.8000 ± 0.4000 0.7000 ± 0.4000 0.7333 ± 0.3887 0.9167 ± 0.1667 Process/Spindle 8 0.9091 ± 0.1408 0.6000 ± 0.4899 0.5000 ± 0.4472 0.5333 ± 0.4522 0.7778 ± 0.2485 Axis Load 16 0.8561 ± 0.1667 0.2000 ± 0.4000 0.3000 ± 0.4000 0.2933 ± 0.3887 0.8911 ± 0.1800 The ablation study revealed that Power/Energy/Runtime was the strongest standalone feature family, achieving the highest mean accuracy (0.9636), mean F1-score (0.8667), and mean ROC-AUC (0.9333). Vibration also performed strongly but did not outperform the Power/Energy/Runtime subset. Interestingly, the All-Features configuration did not improve over the best standalone group, suggesting that a compact and physically meaningful feature subset may be more effective than a larger aggregated feature space for this anomaly-detection task. This finding complements the feature-importance analysis rather than contradicting it. While vibration-related variables dominate the total importance structure of the full model, the ablation study shows that Power/Energy/Runtime forms the most effective standalone feature family. This suggests that vibration contributes many individually important variables within the full model, while Power/Energy/Runtime provides a more compact and directly discriminative feature set when used on its own. 5.6 Heterogeneous-Machine Robustness Analysis To assess robustness under a structurally different machine-data profile, a separate experiment was conducted using the tmf_vf10 machine group. Unlike imi_vm20i and imi_vmx30ui, this machine had a reduced and differently structured MTConnect-derived feature set, with only 13 runs in total ( 8 normal and 5 anomalous ). Because of the small sample size, Leave-One-Out Cross-Validation (LOOCV) was used instead of a standard train-test split or 5-fold cross-validation. Table 6 Heterogeneous-machine robustness results on tmf_vf10 using LOOCV. Model Accuracy_mean Precision_mean Recall_mean F1_mean Logistic Regression 0.4615 0.0769 0.0769 0.0769 Random Forest 0.6154 0.1538 0.1538 0.1538 XGBoost 0.5385 0.1538 0.1538 0.1538 The heterogeneous-machine results were substantially weaker than those obtained on the compatible multi-machine dataset. Among the evaluated models, Random Forest achieved the highest mean accuracy (0.6154), while all three models showed low precision, recall, and F1-score. Since each LOOCV fold contained only one test sample, ROC-AUC was not emphasized for this experiment. These findings are important because they define the boundary of the proposed framework. The results suggest that the approach works well across compatible machine groups with a shared MTConnect-derived feature structure, but degrades when applied to a structurally different machine group with reduced feature richness and different signal semantics. In other words, data standardization supports interoperability at the structural level, but effective anomaly detection additionally requires sufficient feature compatibility across machines. 5.7 Summary of Findings The staged evaluation produced three main findings. First, the single-machine baseline confirmed that experiment-level statistical features extracted from MTConnect-derived data can support meaningful anomaly classification. Second, the compatible multi-machine analysis showed strong cross-validated performance, with Random Forest achieving the highest mean accuracy (0.9455), Logistic Regression the highest recall (0.9000) and F1-score (0.8000), and XGBoost the highest ROC-AUC (0.9222). Feature-importance analysis identified vibration-related variables as the most influential in the full model, while the ablation study showed that Power/Energy/Runtime provided the strongest standalone performance. Third, the tmf_vf10 robustness experiment showed a clear performance drop, indicating that effective anomaly detection depends on feature compatibility across machine groups, not only on data standardization. 5.8 Limitations Several limitations of this study should be acknowledged. First, the dataset used in this work is relatively small. The main combined dataset contains 56 runs with only 9 anomalous cases, and the robustness dataset contains 13 runs. Although stratified cross-validation and LOOCV were used to mitigate the effect of limited samples, the high standard deviations observed in some metrics (e.g., Recall of 0.9000 ± 0.2000 for Logistic Regression) indicate that the performance estimates remain sensitive to fold composition. Future studies with larger and more balanced datasets would strengthen the generalizability of these findings. Oversampling techniques such as those proposed for imbalanced pattern recognition [ 24 ] could also be explored to address class imbalance in future extensions of this framework. Second, only supervised classification methods were used in this study. In many real manufacturing settings, labeled anomaly data may be scarce or unavailable, which would favor unsupervised or semi-supervised approaches such as Isolation Forest or One-Class SVM. The supervised framing adopted here was appropriate given that experiment-level labels were available from documented run remarks, but extending the framework to unsupervised settings would broaden its practical applicability. Third, the models were trained using default or minimally tuned hyperparameters. This choice was intentional to keep the focus on feature representation and machine compatibility rather than model-specific optimization. However, systematic hyperparameter tuning could potentially improve the reported results, particularly for XGBoost, which is known to be sensitive to parameter settings. Fourth, the statistical features used in this study (mean, standard deviation, minimum, maximum) provide a compact experiment-level summary but do not capture temporal patterns within a machining run. Anomalies that develop gradually or occur only during specific phases of a cut may not be fully represented by these summaries. Finally, the cross-machine evaluation was limited to three machine groups from a single public dataset. While this was sufficient to demonstrate the framework and examine feature-compatibility effects, broader validation across additional machine types, facilities, and operating conditions would be needed to confirm the scalability of the approach. 6. Conclusion This study presented an AI/ML-based anomaly-detection framework using MTConnect-derived manufacturing data from multiple milling machine groups. Anomaly labels were assigned at the experiment level using documented run remarks, and statistical features were extracted from process, spindle, load, vibration, energy, and runtime variables. The framework was evaluated through a staged design covering single-machine baseline validation, compatible multi-machine classification, interpretability analysis, and heterogeneous-machine robustness testing. The results demonstrated that MTConnect-derived features can support effective anomaly detection when the machine groups share a compatible feature structure. The compatible multi-machine analysis produced strong cross-validated classification performance across all three models, while the interpretability analysis revealed that both vibration signatures and power/energy/runtime characteristics carry meaningful anomaly-related information. The finding that a compact feature subset (Power/Energy/Runtime) outperformed the full feature set in the ablation study has practical implications, suggesting that smaller, physically meaningful feature groups may be preferable to larger aggregated representations in some manufacturing contexts. Equally important was the robustness experiment, which showed that performance degrades substantially when the machine-data structure changes. This result highlights a distinction that is often overlooked in the literature: MTConnect supports interoperability at the data-structure level, but effective cross-machine anomaly detection still requires sufficient feature compatibility. Standardized data is a necessary foundation, but not a sufficient one. These findings contribute to the growing body of work on smart manufacturing monitoring by providing empirical evidence of both the potential and the practical limits of MTConnect-derived anomaly detection across machines. Future work should focus on extending the framework to larger and more diverse machine populations, exploring transfer-learning or domain-adaptation strategies for heterogeneous machine-data profiles, incorporating temporal and multimodal signal representations, and moving toward online or near-real-time deployment for shop-floor decision support. Abbreviations ATC Apparent Tardiness Cost (dispatching rule) CMM Coordinate Measuring Machine CNC Computer Numerical Control CPFR Collaborative Planning, Forecasting, and Replenishment DT Digital Twin EDD Earliest Due Date (dispatching rule) EV Electric Vehicle IIoT Industrial Internet of Things KPI Key Performance Indicator MILP Mixed-Integer Linear Programming MQTT Message Queuing Telemetry Transport MTConnect An open, royalty-free standard for streaming manufacturing equipment data OEE Overall Equipment Effectiveness OEM Original Equipment Manufacturer OPC UA Open Platform Communications Unified Architecture OTFR On-Time Fulfillment Rate PLC Programmable Logic Controller SKU Stock Keeping Unit Tier-1 First-tier direct supplier to an OEM in a manufacturing supply chain VMI Vendor-Managed Inventory WIP Work-In-Process WT Weighted Tardiness Declarations Funding Not applicable. Competing interests The authors declare no competing interests. Authors' contributions Harshkumar K. Parmar contributed to conceptualization, methodology, software implementation, formal analysis, data curation, visualization, and writing - original draft preparation. Shivakumar Raman contributed to supervision, technical guidance, review, and editing. Both authors read and approved the final manuscript. Declaration of Generative AI in Scientific Writing During the preparation of this manuscript, the authors used ChatGPT to assist with language editing and structural revision. The authors reviewed and edited the output as needed and take full responsibility for the final content of the manuscript. Data Availability Statement The dataset used in this study is publicly available and was originally published by Ströbel et al. [21]. The experimental analysis was conducted using Python with scikit-learn and XGBoost. Feature engineering and statistical summarization were performed using pandas and NumPy. The code and processed feature tables can be made available upon reasonable request to the corresponding author. Declaration of Generative AI in Scientific Writing During the preparation of this manuscript, I used an AI-based language tool (ChatGPT) to assist with language editing and structural revision. References Elía I, Pagola M (2025) Anomaly detection in Smart-manufacturing era: A review. Eng Appl Artif Intell 139:109578Part B MTConnect, Institute (2026) MTConnect Standard Part 1.0 – Fundamentals , Version 2.5.0, Jan Lei P, Zheng L, Li C, Li X (2016) MTConnect Enabled Interoperable Monitoring System for Finish Machining Assembly Interfaces of Large-scale Components. Procedia CIRP 56:378–383 Lei P, Zheng L, Wang L, Wang Y, Li C, Li X (2017) MTConnect compliant monitoring for finishing assembly interfaces of large-scale components: A vertical tail section application. J Manuf Syst Liu C, Xu X, Peng Q, Zhou Z (2018) MTConnect-based Cyber-Physical Machine Tool: A case study. Procedia CIRP 72:492–497 Liao L, Minhas R, Rangarajan A, Kurtoglu T, de Kleer J (2014) A Self-Aware Machine Platform in Manufacturing Shop Floor Utilizing MTConnect Data, Proceedings of the Annual Conference of the PHM Society , vol. 6, no. 1 Venkatesh S, Ly S, Manning M, Michaloski J, Proctor F Automating Asset Knowledge with MTConnect, in Proc. ASME 2016 Int. Manufacturing Science and Engineering Conference (MSEC (2016)), Blacksburg, VA, USA, 2016, Paper No. MSEC2016-8629, doi: ), Blacksburg, VA, USA, 2016, Paper No. MSEC2016-8629. 10.1115/MSEC2016-8629 Helu M, Hedberg T Jr., Barnard Feeney A (2017) Reference architecture to integrate heterogeneous manufacturing systems for the digital thread. CIRP J Manufact Sci Technol 19:191–195 Theissler A, Pérez-Velázquez J, Kettelgerdes M, Elger G (2021) Predictive maintenance enabled by machine learning: Use cases and challenges in the automotive industry. Reliab Eng Syst Saf 215:107864 Kausik AK, Rashid AB, Baki RF, Maktum MMJ (2025) Machine learning algorithms for manufacturing quality assurance: A systematic review of performance metrics and applications. Array 26:100393 Kamm S, Veekati SS, Müller T, Jazdi N, Weyrich M (2023) A survey on machine learning based analysis of heterogeneous data in industrial automation. Comput Ind 149:103930 de Giorgio A, Cola G, Wang L (2023) Systematic review of class imbalance problems in manufacturing. J Manuf Syst 71:620–644. 10.1016/j.jmsy.2023.10.014 Michau G, Fink O (2021) Unsupervised transfer learning for anomaly detection: Application to complementary operating condition transfer, Knowledge-Based Systems , vol. 216, Art. 106816. 10.1016/j.knosys.2021.106816 Liu Z, Lang Z-Q, Gui Y, Zhu Y-P, Laalej H (2024) Digital twin-based anomaly detection for real-time tool condition monitoring in machining. J Manuf Syst 75:163–173. 10.1016/j.jmsy.2024.06.004 Xie Z, Zhang Z, Chen J, Feng Y, Pan X, Zhou Z, He S (2024) Data-driven unsupervised anomaly detection of manufacturing processes with multi-scale prototype augmentation and multi-sensor data. J Manuf Syst 77:26–39. 10.1016/j.jmsy.2024.08.027 Orabi M, Tran KP, Egger P, Thomassey S (2024) Anomaly detection in smart manufacturing: An Adaptive Adversarial Transformer-based model. J Manuf Syst 77:591–611. 10.1016/j.jmsy.2024.09.021 Schueller A, Saldaña C (2022) Generalizability analysis of tool condition monitoring ensemble machine learning models. J Manuf Process 84. 10.1016/j.jmapro.2022.10.064 Mohanraj T, Yerchuru J, Krishnan H, Aravind RSN, Yameni R (2021) Development of tool condition monitoring system in end milling process using wavelet features and Hoelder’s exponent with machine learning algorithms. Measurement 173:108671. 10.1016/j.measurement.2020.108671 Johannssen A, Qiu P, Yeganeh A, Chukhrova N (2025) Explainable AI for trustworthy intelligent process monitoring. Comput Ind Eng 209:111407. 10.1016/j.cie.2025.111407 Gross D, Spieker H, Gotlieb A, Knoblauch R, Elmansori M (2024) Efficient Milling Quality Prediction with Explainable Machine Learning. IFAC-PapersOnLine 58:43–48. 10.1016/j.ifacol.2024.09.085 Ströbel R, Kuck M, Oexle F, Kader H, Puchta A, Noack B, Fleischer J (2025) A multimodal dataset for process monitoring and anomaly detection in industrial CNC milling. Data Brief 63:112207. 10.1016/j.dib.2025.112207 Harshkumar K, Parmar, Raman S (2025) MTConnect for Communication within and across Manufacturing Enterprises, in IISE Annual Conference. Proceedings , pp. 1–11, Institute of Industrial and Systems Engineers (IISE) Ahsan MM, Raman S, Liu Y, Siddique Z (2024) Defect analysis of 3D printed object using transfer learning approaches. Expert Syst Appl 253:124293 Ahsan MM, Raman S, Siddique Z (2023) Bsgan: A novel oversampling technique for imbalanced pattern recognitions, arXiv preprint arXiv:2305.09777 Samadi H, Ahsan MM, Raman S (2025) Metrology and manufacturing-integrated digital twin (MM-DT) for advanced manufacturing: Insights from coordinate measuring machine (CMM) and FARO arm measurements. Next Res 2(2):100299 Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Major Revisions Needed 27 Apr, 2026 Reviewers agreed at journal 14 Apr, 2026 Reviewers invited by journal 14 Apr, 2026 Editor assigned by journal 10 Apr, 2026 First submitted to journal 09 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9358105","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":622967663,"identity":"d8e7c9ea-e8b8-4f8c-b20d-771b1cc52c51","order_by":0,"name":"Harshkumar Kiritbhai Parmar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYBACAyBmZjgA4TB+qLCRA7N4iNXCLHEmzZg0LQy8LYcTGwhpMWc/+/hzwZk78ubTDj/7INlwOL1fIoHxwds23Fose9LNpGfceGY453aa8YzCHem5M2ckMBvOxaPF4EAaGzPPh8OMM6QTjBkkz1jnbriRwCbNi0/L+WfMn4Fa7GdIp39m4G1jTje4kcD+G6+WG2kM0jw3DifOkM4xBmpxTgBqYWPGr+UZmzTPmWfJQC3FoEA2nNnzsFlyzjl8DksDOuzYHVugwzaDolKenz354Ic3Zbi1QMEBZA5jA0H16FpGwSgYBaNgFKACAOWHVd5Qcn0wAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0005-8855-8678","institution":"The University of Oklahoma - Norman Campus: The University of Oklahoma","correspondingAuthor":true,"prefix":"","firstName":"Harshkumar","middleName":"Kiritbhai","lastName":"Parmar","suffix":""},{"id":622967664,"identity":"ff2cbfcc-be86-45af-82ec-02a7e3e5469c","order_by":1,"name":"Shivakumar Raman","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Shivakumar","middleName":"","lastName":"Raman","suffix":""}],"badges":[],"createdAt":"2026-04-08 14:11:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9358105/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9358105/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107473790,"identity":"8ef2a868-cd33-4919-b7c5-c721d22aa06e","added_by":"auto","created_at":"2026-04-21 21:34:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1046310,"visible":true,"origin":"","legend":"\u003cp\u003eOverall framework of the proposed AI/ML-based interoperable anomaly-detection framework using MTConnect-derived multi-machine data\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/b20882bc907967a65c50c2e2.png"},{"id":107871344,"identity":"75b71a68-9e21-47f4-9484-ab13374ebae4","added_by":"auto","created_at":"2026-04-27 07:48:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":217271,"visible":true,"origin":"","legend":"\u003cp\u003eMean 5-fold cross-validation performance across Logistic Regression, Random Forest, and XGBoost on the combined multi-machine dataset\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/4053d864e9949a1cc33fedde.png"},{"id":107704454,"identity":"ba3315f5-e7a4-483d-b2a0-5d3500da61b6","added_by":"auto","created_at":"2026-04-24 08:45:27","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":265866,"visible":true,"origin":"","legend":"\u003cp\u003eTop 15 Random Forest feature importances for anomaly detection on the combined multi-machine dataset\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/821d5e2d0840f415904a1550.png"},{"id":107489850,"identity":"eab4a2db-e49b-4827-bad0-898535c721a3","added_by":"auto","created_at":"2026-04-22 02:49:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":114586,"visible":true,"origin":"","legend":"\u003cp\u003eRelative contribution of grouped MTConnect-derived feature families to Random Forest importance.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/96671bade08d0547586ec776.png"},{"id":107473792,"identity":"6983a806-0d29-4e72-b5ec-b6da5bd8905c","added_by":"auto","created_at":"2026-04-21 21:34:11","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":160654,"visible":true,"origin":"","legend":"\u003cp\u003eAblation study results comparing mean F1-score across feature groups using Random Forest.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/7fd23d7dddd90a5f31982441.png"},{"id":108180954,"identity":"ce0d8993-c829-457b-a61e-2ea78e74fa52","added_by":"auto","created_at":"2026-04-30 08:55:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2208975,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9358105/v1/59dfb0c1-efc9-48e8-86d7-ccf1850979f8.pdf"}],"financialInterests":"","formattedTitle":"AI/ML Based Interoperable Anomaly Detection in Advanced Manufacturing Using MTConnect-Derived Multi-Machine Data","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eModern manufacturing systems increasingly depend on the ability to detect abnormal process behavior early. In milling operations, issues such as excessive vibration, clamp failure, tool breakage, or unstable cutting conditions can quickly affect process stability, product quality, and machine health. If these problems are not identified in time, they can lead to scrap, downtime, and unnecessary cost. Consequently, anomaly detection has become an important part of smart manufacturing and real-time shop-floor monitoring [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eA large portion of the existing work in this area relies on isolated sensor datasets, machine-specific signals, or single-machine data pipelines. While those studies are useful, they often remain tied to a particular setup and are difficult to extend across different machines or production environments. In practice, manufacturing systems need approaches that are not only accurate, but also easier to scale and reuse across equipment. This is where standardized machine data becomes especially important.\u003c/p\u003e \u003cp\u003eMTConnect is valuable in this context because it provides a common structure for machine-generated data. Rather than relying on proprietary tags, controller-specific naming conventions, or ad hoc data pipelines, MTConnect provides a standardized representation of machine, spindle, load, vibration, and runtime information [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Prior studies have shown that MTConnect can support interoperable monitoring, asset knowledge automation, and broader digital-manufacturing integration across heterogeneous systems [\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]-[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Recent work has also examined the role of MTConnect in enabling communication within and across manufacturing enterprises [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Although anomaly detection can still be performed without MTConnect, such approaches usually require more custom preprocessing, manual data mapping, and machine-specific integration. Accordingly, MTConnect is not a prerequisite for anomaly detection, but it substantially improves the practicality of building frameworks that are reusable, interpretable, and scalable across machines.\u003c/p\u003e \u003cp\u003eEven with the growing use of MTConnect for machine connectivity and interoperability, its role in anomaly detection across multiple machine groups remains limited. Much of the published work either focuses on MTConnect-enabled monitoring and data integration or develops anomaly-detection methods in single-machine or application-specific settings [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]-[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Limited work exists that integrates these two directions within a unified framework for anomaly detection using MTConnect-derived data across multiple machine groups, while also examining how performance changes when machine-data structures are not fully compatible.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eProblem\u003c/strong\u003e \u003cp\u003eDetecting abnormal machining behavior early is important for reducing downtime, quality loss, and process instability. However, doing this consistently across different machines remains difficult because most existing approaches are developed and validated on single-machine or application-specific datasets, making their scalability and reusability unclear.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eGap\u003c/strong\u003e \u003cp\u003eAlthough MTConnect has been widely adopted for machine connectivity and data standardization, its potential for enabling anomaly detection across multiple machine groups has not been systematically examined. In particular, no prior study has evaluated how MTConnect-derived anomaly-detection performance changes when feature compatibility varies across machines, a scenario that is common in real manufacturing environments.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eSolution\u003c/strong\u003e \u003cp\u003eTo address this gap, this paper develops an AI/ML-based anomaly-detection framework using MTConnect-derived multi-machine data. Unlike prior work that focuses either on MTConnect-based monitoring or on single-machine anomaly detection, this framework integrates experiment-level labeling, statistical feature engineering, model comparison, interpretability analysis, and robustness evaluation under both feature-compatible and feature-heterogeneous machine-data conditions. The explicit evaluation of transferability limits across structurally different machine groups distinguishes this work from existing studies.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe main contributions of this work are summarized as follows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eAn interoperable anomaly-detection framework is developed using MTConnect-derived manufacturing data for milling-process monitoring.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe framework is evaluated across two compatible machine groups using a common set of standardized features, enabling multi-machine anomaly detection rather than single-machine analysis alone.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eInterpretability is improved through both feature-importance analysis and feature-group ablation, showing how vibration, power/energy/runtime, axis-load, and process/spindle variables contribute differently to anomaly detection.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eA separate robustness experiment on a structurally different machine group highlights the transferability limits of anomaly detection when machine-data feature compatibility is reduced.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe study provides empirical evidence that standardized machine data supports but does not automatically guarantee cross-machine anomaly-detection transferability, offering practical guidance for deploying interoperable monitoring systems in heterogeneous manufacturing environments.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe remainder of this study is organized as follows. Section 2 reviews related work on MTConnect-based data integration, manufacturing anomaly detection, and heterogeneous industrial-data learning. Section 3 describes the dataset, anomaly-labeling process, and feature engineering. Section 4 presents the modeling approach and evaluation strategy. Section 5 reports and discusses the results, including cross-validation, interpretability, ablation, and robustness analysis. Section 6 concludes this study and outlines future directions.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003ePrior work related to this study can be grouped into three main areas: MTConnect-based machine-data integration, machine-learning methods for anomaly detection in manufacturing, and learning under heterogeneous industrial data conditions. These areas come together in this paper because the objective is not only to detect anomalies, but to do so using standardized machine data in a way that can extend beyond a single machine.\u003c/p\u003e \u003cp\u003eEarly MTConnect-related work mainly focused on machine connectivity, monitoring, and interoperability Liao [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. showed that MTConnect data can support a self-aware machine framework for condition monitoring and failure-related reasoning. Venkatesh [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. demonstrated how MTConnect can be used to automate asset knowledge and machine-status interpretation. In a broader digital-manufacturing context, Helu et al [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. emphasized the value of standards-based architectures for integrating heterogeneous manufacturing systems and supporting the digital thread. More recently, metrology-integrated digital twin frameworks have also been explored for advanced manufacturing applications [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBeyond review-level findings, several recent studies have addressed anomaly detection and condition monitoring in machining through different technical approaches. Liu [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] proposed a digital-twin-based framework for real-time tool condition monitoring, using simulation-driven anomaly detection tied to a specific machine setup. Xie [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] developed an unsupervised method combining multi-scale prototype augmentation with multi-sensor data for manufacturing process monitoring, though their approach was designed for within-process anomaly detection rather than cross-machine evaluation. Orabi [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] introduced an adaptive adversarial transformer-based model for smart manufacturing anomaly detection, demonstrating strong performance but within a single application context. In the area of tool-condition monitoring specifically, Schueller and Salda\u0026ntilde;a [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] examined the generalizability of ensemble machine learning models across cutting conditions, providing relevant insight into transferability challenges, though without using standardized data protocols such as MTConnect. Mohanraj [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] combined wavelet features with machine learning for end-milling tool monitoring, focusing on signal-processing techniques rather than cross-machine interoperability. While these studies advance the state of the art in their respective directions, none of them combines standardized MTConnect-derived data with multi-machine anomaly detection and explicit evaluation of transferability under heterogeneous feature conditions.\u003c/p\u003e \u003cp\u003eRecent work has also emphasized the importance of explainability in manufacturing ML systems. Johannssen. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] discussed explainable AI as a requirement for trustworthy intelligent process monitoring, while Gross [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] demonstrated explainable machine learning for milling quality prediction. These studies reinforce the value of interpretability in manufacturing contexts, which the present work addresses through feature-importance analysis and feature-group ablation.\u003c/p\u003e \u003cp\u003eTransferability is another open issue. Even when anomaly detection performs well on one dataset, that does not mean the same model will transfer effectively to another machine or operating environment. Michau and Fink [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] discussed unsupervised transfer learning for anomaly detection and highlighted the broader challenge of operating-condition transfer when data structure changes. Transfer learning has also been applied to defect analysis in additive manufacturing contexts [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], reinforcing the broader relevance of cross-domain transferability in manufacturing quality applications.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of related studies and positioning of the present work.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMain Focus\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eData Basis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSingle / Multi Machine\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMTConnect-Centered\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eInterpretability\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eHeterogeneous Validation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiao.[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSelf-aware machine monitoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMTConnect data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSingle-machine oriented\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVenkatesh et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsset knowledge automation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMTConnect data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNot focused on anomaly detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHelu et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDigital thread / integration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStandards-based manufacturing data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSystem-level\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePartly\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReview / anomaly detection studies [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eManufacturing anomaly detection and ML\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensor / manufacturing datasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMostly single or application-specific\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo / limited\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eVaries\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRare\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeterogeneous industrial-data studies [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u0026ndash;[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHeterogeneity, imbalance, transferability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIndustrial / anomaly datasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMulti-source or cross-condition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThis work\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAI/ML-based interoperable anomaly detection\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eMTConnect-derived multi-machine data\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eMulti-machine\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTaken together, the reviewed studies show strong progress in MTConnect-based interoperability, machine monitoring, and manufacturing anomaly detection. However, a clear gap remains in combining these directions into a single framework for AI/ML-based anomaly detection using MTConnect-derived data across multiple machine groups, while also examining what happens when feature compatibility changes across machines. This paper addresses that gap by developing a multi-machine anomaly-detection framework using MTConnect-derived milling data, evaluating it on two compatible machine groups, and then testing robustness on a structurally different machine group [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10 CR11 CR12\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u0026ndash;[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e"},{"header":"3. Dataset and Feature Engineering","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset Overview\u003c/h2\u003e \u003cp\u003eThe dataset used in this study is derived from a publicly available multi-sensor metal-milling anomaly dataset [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] organized into three machine groups: imi_vm20i, imi_vmx30ui, and tmf_vf10. Each machine group contains multiple experiment folders, where each folder represents one machining run and includes an MTConnect-derived _mtc.csv file together with other supporting files such as labels, audio, video, or NC-code data. Since the focus of this work is interoperable anomaly detection using standardized machine information, only the MTConnect-derived data files were used, and other available modalities (audio, video, NC-code) were not included in the analysis.\u003c/p\u003e \u003cp\u003eA key characteristic of the dataset is that the three machine groups are not equally compatible from a feature perspective. The imi_vm20i and imi_vmx30ui groups share a relatively rich and similar MTConnect-derived feature structure, which makes them suitable for combined multi-machine analysis. In contrast, tmf_vf10 contains a smaller and structurally different feature set. This difference is significant because the study is concerned not only with anomaly detection accuracy, but also with how feature compatibility affects transferability across machine groups.\u003c/p\u003e \u003cp\u003eBased on this structure, the dataset was used in three stages:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSingle-machine baseline\u003c/b\u003e: imi_vm20i\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMain compatible multi-machine dataset\u003c/b\u003e: imi_vm20i\u0026thinsp;+\u0026thinsp;imi_vmx30ui\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eHeterogeneous-machine robustness dataset\u003c/b\u003e: tmf_vf10\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDataset composition used in this study.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMachine Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNormal Runs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAnomalous Runs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTotal Runs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRole in Study\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eimi_vm20i\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSingle-machine baseline\u0026thinsp;+\u0026thinsp;combined main dataset\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eimi_vmx30ui\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCombined main dataset\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etmf_vf10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHeterogeneous-machine robustness experiment\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Anomaly Label Definition\u003c/h2\u003e \u003cp\u003eThe anomaly labels used in this paper were assigned at the experiment level. Although some folders included label.csv files, those labels did not consistently represent normal-versus-anomalous status across all machine groups. Instead, the final anomaly labels were derived from the documented experiment remarks in the dataset summary, which provided a more consistent and practically meaningful basis for classification.\u003c/p\u003e \u003cp\u003eRuns with no abnormal remark were labeled as normal (0). Runs with remarks indicating process abnormalities, such as excessive vibration, clamp failure, tool breakage, residual chip adhesion, chatter, blown insert, or early stopping, were labeled as anomalous (1). This approach reflects how abnormal runs are often identified in real manufacturing environments, where process logs, operator notes, and documented run outcomes may be more reliable than perfectly segmented point-level fault annotations.\u003c/p\u003e \u003cp\u003eThe use of experiment-level labels also matches the modeling strategy adopted in this paper. Since each machining run is converted into one experiment-level feature vector, the target variable naturally represents whether the run as a whole should be considered normal or anomalous.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Feature Compatibility Across Machine Groups\u003c/h2\u003e \u003cp\u003eOne of the central design decisions in this work was how to handle feature compatibility across machine groups. For the main multi-machine analysis, only the two compatible machine groups, imi_vm20i and imi_vmx30ui, were combined. These two groups shared a common set of MTConnect-derived variables related to process behavior, spindle motion, axis loads, vibration statistics, energy, and runtime. This shared feature space made it possible to construct a unified dataset without forcing aggressive feature reduction.\u003c/p\u003e \u003cp\u003eThe tmf_vf10 machine group, however, did not provide the same feature richness. Its MTConnect-derived variables were more limited and followed a different structure, including variables such as cycle time, motion time, spindle load, spindle speed, and axis-related power or motion signals. As a result, tmf_vf10 was not merged directly into the main multi-machine dataset. Instead, it was used separately to test how the proposed framework behaves when applied to a machine group with reduced and structurally different feature availability.\u003c/p\u003e \u003cp\u003eThis separation is important for the overall contribution of this study. It allows the study to evaluate anomaly detection under both feature-compatible and feature-heterogeneous conditions, which helps clarify the practical limits of interoperable anomaly detection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Experiment-Level Statistical Features\u003c/h2\u003e \u003cp\u003eEach _mtc.csv file was treated as one complete machining experiment. Rather than using raw time-series rows directly for classification, experiment-level statistical features were extracted from each selected MTConnect-derived variable. For each retained variable, four summary statistics were computed: mean, standard deviation, minimum, and maximum.\u003c/p\u003e \u003cp\u003eThis produced one feature vector per experiment. For example, a variable such as Spindle_Speed contributed the following derived features: Spindle_Speed_mean, Spindle_Speed_std, Spindle_Speed_min, and Spindle_Speed_max. The same procedure was applied to vibration-related, energy-related, load-related, and runtime-related variables.\u003c/p\u003e \u003cp\u003eThis representation was chosen for practical and methodological reasons. First, it provides a compact way to summarize full-run behavior without depending on identical run lengths across experiments. Second, it improves interpretability by linking anomaly detection to familiar summary characteristics such as variability, extremes, and average operating levels. Third, it makes it easier to compare compatible machine groups without requiring detailed temporal alignment or sequence modeling.\u003c/p\u003e \u003cp\u003eFor the main combined dataset, only common MTConnect-derived variables shared by imi_vm20i and imi_vmx30ui were retained. For tmf_vf10, the same statistical summarization procedure was used, but only on the smaller machine-specific feature set available in that group.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Final Datasets Used for Modeling\u003c/h2\u003e \u003cp\u003eBased on the feature-engineering process, three experiment-level datasets were prepared for modeling:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eBaseline dataset: 37 runs from imi_vm20i\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMain combined dataset: 56 runs from imi_vm20i and imi_vmx30ui\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRobustness dataset: 13 runs from tmf_vf10\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe baseline dataset was used to verify that the anomaly-labeling and feature-engineering strategy produced a meaningful anomaly-detection problem. The combined dataset was used as the main dataset for model comparison, cross-validation, feature-importance analysis, and ablation study. The tmf_vf10 dataset was used separately for a robustness experiment to evaluate performance under heterogeneous machine-data conditions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Data Cleaning and Preprocessing\u003c/h2\u003e \u003cp\u003eBefore modeling, the experiment-level feature tables were checked for missing values and nonnumeric entries. Variables that became completely empty after numeric conversion were removed, particularly in the tmf_vf10 robustness dataset, where several machine-specific variables did not provide usable numeric observations after aggregation. For the remaining features, missing values were handled using median imputation during model training.\u003c/p\u003e \u003cp\u003eAt the end of this stage, the study produced clean experiment-level datasets with anomaly labels derived from documented run remarks and features extracted from standardized MTConnect-derived machine data.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Methodology","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Modeling Objective\u003c/h2\u003e \u003cp\u003eThe objective of this study was to determine whether experiment-level statistical features extracted from MTConnect-derived manufacturing data can distinguish normal and anomalous milling runs across multiple machine groups. More specifically, the methodology was designed to answer three related questions:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eCan anomaly detection be performed reliably on a single MTConnect-based machine group?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eCan the same framework be extended across multiple compatible machine groups using a common feature space?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWhat happens when the framework is applied to a structurally different machine group with reduced feature compatibility?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThese questions directly shaped the experimental design and the staged evaluation strategy adopted in this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Classification Models\u003c/h2\u003e \u003cp\u003eThree supervised machine-learning models were selected for anomaly classification:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLogistic Regression\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRandom Forest\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eXGBoost\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThese models were selected because they represent different learning behaviors-linear, nonlinear ensemble, and gradient boosting, while remaining well suited to structured tabular manufacturing data. Although anomaly detection is often approached using unsupervised methods such as Isolation Forest or One-Class SVM, a supervised framing was adopted in this study for two reasons. First, experiment-level anomaly labels were available from documented run remarks, making supervised classification both feasible and appropriate. Second, supervised models allow direct evaluation using standard classification metrics such as precision, recall, F1-score, and ROC-AUC, which makes model comparison and interpretability more straightforward. The use of unsupervised methods remains a relevant direction for future work, particularly in settings where labeled anomaly data are not available.\u003c/p\u003e \u003cp\u003eLogistic Regression was used as a simple baseline model to test whether anomaly separation could be captured through relatively direct relationships in the engineered feature space. Random Forest was chosen because it is well suited to tabular industrial data, can capture nonlinear interactions, and also, supports interpretability through feature importance. XGBoost was included as a boosting-based model that often performs well on structured datasets and provides a useful comparison to both linear and ensemble-tree methods.\u003c/p\u003e \u003cp\u003eAll three models were trained using their default hyperparameter configurations as implemented in scikit-learn (Logistic Regression and Random Forest) and the XGBoost Python package. Specifically, Random Forest used 100 estimators with no maximum depth constraint, and XGBoost used a learning rate of 0.3 with a maximum depth of 6 and 100 boosting rounds. No systematic hyperparameter optimization (e.g., grid search or Bayesian tuning) was performed. This decision was deliberate: the primary objective of this study was to evaluate the informativeness of MTConnect-derived features and the effect of machine compatibility on anomaly detection, rather than to maximize model-specific performance through extensive tuning. The potential effect of hyperparameter optimization is acknowledged as a limitation and a direction for future investigation.\u003c/p\u003e \u003cp\u003eTogether, these models allow the study to compare linear, nonlinear ensemble, and boosting-based approaches within one anomaly-detection framework.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Preprocessing During Model Training\u003c/h2\u003e \u003cp\u003eThe final experiment-level feature tables were used as model inputs. Missing values in the retained features were handled through median imputation. Logistic Regression was trained on imputed and standardized features because the model is sensitive to feature scale. Random Forest and XGBoost were trained on the imputed feature values directly without standardization.\u003c/p\u003e \u003cp\u003eThis preprocessing strategy was deliberately kept minimal to ensure that the results reflect differences in feature representation, machine compatibility, and model behavior rather than the effects of extensive model-specific tuning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Experimental Design\u003c/h2\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.4.1 Experiment 1: Single-Machine Baseline\u003c/h2\u003e \u003cp\u003eThe first experiment used only the imi_vm20i machine group. This baseline was included to validate the overall pipeline, including experiment-level anomaly labeling, MTConnect-derived feature extraction, and model training. A held-out train-test split was employed to verify that the extracted feature space provided sufficient separability between normal and anomalous runs.\u003c/p\u003e \u003cp\u003eThis experiment was not intended to serve as the main evidence of this study. Instead, it provided a controlled starting point before moving to the stronger multi-machine setting.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e4.4.2 Experiment 2: Compatible Multi-Machine Analysis\u003c/h2\u003e \u003cp\u003eThe main experiment of this study used the combined dataset built from imi_vm20i and imi_vmx30ui. These two machine groups were selected because they share a sufficiently similar set of MTConnect-derived variables. Only the variables common to both groups were retained so that the resulting anomaly-detection problem reflected a genuinely shared feature space.\u003c/p\u003e \u003cp\u003eA held-out split was first used to obtain an initial benchmark. However, because the combined dataset still contained a limited number of anomaly cases, this split was treated as an exploratory benchmark rather than the main result. The primary evaluation for the combined dataset therefore relied on cross-validation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e4.4.3 Experiment 3: Five-Fold Stratified Cross-Validation\u003c/h2\u003e \u003cp\u003eFor the combined multi-machine dataset, \u003cb\u003e5-fold stratified cross-validation\u003c/b\u003e was used as the main evaluation method. Stratification helped preserve the approximate anomaly-to-normal ratio across folds, which was important because the dataset was imbalanced.\u003c/p\u003e \u003cp\u003eThis evaluation was used to compare Logistic Regression, Random Forest, and XGBoost in terms of mean performance and variability across folds. Compared with a single train-test split, this provided a more reliable estimate of model behavior and formed the basis of the main model-comparison results reported in this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e4.4.4 Experiment 4: Feature Importance and Ablation\u003c/h2\u003e \u003cp\u003eTo better understand how the models made decisions, Random Forest feature importance was extracted from the combined dataset. This analysis was used to identify which MTConnect-derived variables contributed most strongly to anomaly detection.\u003c/p\u003e \u003cp\u003eA feature-group ablation study was also performed using Random Forest and 5-fold stratified cross-validation. The features were divided into four families:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eProcess/Spindle\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAxis Load\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eVibration\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePower/Energy/Runtime\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eEach group was evaluated independently, and an additional \u003cb\u003eAll Features\u003c/b\u003e configuration was included. This experiment was designed to determine whether certain groups of MTConnect-derived variables were sufficient on their own and whether combining all available features necessarily improved anomaly-detection performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.4.5 Experiment 5: Heterogeneous-Machine Robustness\u003c/h2\u003e \u003cp\u003eThe final experiment used the tmf_vf10 machine group as a robustness test. This group had a reduced and structurally different feature set compared with the two main machine groups. Because only 13 runs were available, a conventional train-test split, or 5-fold cross-validation would have been unstable and highly sensitive to the specific partition. For that reason, Leave-One-Out Cross-Validation (LOOCV) was used.\u003c/p\u003e \u003cp\u003eThis experiment was not intended as a direct competitor to the main compatible-machine analysis. Instead, it was designed to assess how the proposed framework behaves when the MTConnect-derived feature space changes substantially. In that sense, it serves as a transferability test under heterogeneous machine-data conditions.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eModel performance was evaluated using the following metrics:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAccuracy\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePrecision\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRecall\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eF1-score\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eROC-AUC\u003c/b\u003e \u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eAccuracy measures the overall proportion of correctly classified runs. Precision reflects how many predicted anomalies were truly anomalous. Recall indicates how many actual anomalies were correctly detected. F1-score provides a balance between precision and recall and is especially useful when the classes are imbalanced. ROC-AUC measures the ability of a model to rank anomalous runs ahead of normal runs across decision thresholds.\u003c/p\u003e \u003cp\u003eFor the main combined dataset, all five metrics were reported in the cross-validation analysis. For the tmf_vf10 LOOCV robustness experiment, ROC-AUC was not emphasized because each fold contained only one test sample, making ROC-AUC undefined or unstable in many folds. As a result, the robustness discussion focused mainly on accuracy, precision, recall, and F1-score.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Interpretation Strategy\u003c/h2\u003e \u003cp\u003eThe methodology in this paper was designed not only to compare predictive performance, but also to explain it. Model-comparison results were used to evaluate classification strength, feature-importance analysis was used to identify influential MTConnect-derived variables, and the ablation study was used to determine which feature families contributed most effectively. The robustness experiment then tested how strongly the framework depends on feature compatibility across machine groups.\u003c/p\u003e \u003cp\u003eThis interpretation strategy is central to the contribution of this work. The study is not simply asking whether anomaly detection works, but also why it works on compatible machine groups and why it weakens under heterogeneous machine-data conditions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.7 Summary of Method\u003c/h2\u003e \u003cp\u003eIn summary, the methodology combines experiment-level anomaly labeling, statistical feature engineering, supervised classification, cross-validation, interpretability analysis, and robustness testing. This design allows the study to move beyond a single-machine proof-of-concept and provide a more realistic assessment of how MTConnect-derived manufacturing data can support interoperable anomaly detection across multiple machine groups.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Results and Discussion","content":"\u003cp\u003eThis section evaluates the proposed framework in a staged way. A single-machine baseline is first used to verify that experiment-level MTConnect-derived statistical features can support anomaly classification. The analysis is then extended to a combined multi-machine dataset built from the two compatible machine groups, imi_vm20i and imi_vmx30ui, where the main model-comparison and cross-validation results are reported. To improve interpretability, feature-importance analysis and feature-group ablation are also performed. Finally, robustness is assessed on the structurally different tmf_vf10 machine group to examine the limits of transferability under heterogeneous feature conditions.\u003c/p\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Single-Machine Baseline\u003c/h2\u003e \u003cp\u003eThe first experiment was conducted on the imi_vm20i machine group (37 runs: 30 normal, 7 anomalous). This baseline was used to confirm that the experiment-level labeling strategy and MTConnect-derived statistical feature representation produced a meaningful anomaly-classification problem.\u003c/p\u003e \u003cp\u003eUsing the extracted experiment-level features, the initial baseline showed strong class separation between normal and anomalous runs. This result provided an important validation of the overall pipeline, including anomaly-label assignment from documented run remarks, feature generation from MTConnect-derived variables, and supervised model training. However, because this experiment was limited to one machine group and a relatively small number of anomaly cases, it was treated as a proof-of-concept rather than the main evidence of this work.\u003c/p\u003e \u003cp\u003eOverall, the single-machine baseline confirmed that experiment-level statistical summaries of MTConnect-derived variables can carry useful anomaly-related information. The next step was therefore to examine whether the same framework could be extended to a compatible multi-machine setting.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Compatible Multi-Machine Held-Out Split\u003c/h2\u003e \u003cp\u003eThe second experiment used the combined dataset built from the two compatible machine groups, imi_vm20i and imi_vmx30ui. This dataset contained 56 runs, including 47 normal and 9 anomalous runs. Since these two machine groups shared a common MTConnect-derived feature space, they were used to form the main dataset for the core anomaly-detection analysis.\u003c/p\u003e \u003cp\u003eA held-out train-test split was first used to obtain an initial benchmark for the combined dataset. The results are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHeld-out split model performance on the combined imi_vm20i and imi_vmx30ui dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9286\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9286\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9583\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eOn this split, Logistic Regression achieved the strongest overall performance, while Random Forest and XGBoost remained highly precise but showed lower recall. These results indicate strong class separability in the engineered feature space for this particular partition. However, because a single held-out split can be sensitive to the specific train-test partition, especially in a dataset with limited anomaly cases, it served only as an initial benchmark. The main evaluation of this study is therefore based on 5-fold stratified cross-validation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Five-Fold Stratified Cross-Validation on the Compatible Multi-Machine Dataset\u003c/h2\u003e \u003cp\u003eTo obtain a more robust estimate of model performance, 5-fold stratified cross-validation was performed on the combined compatible-machine dataset. This evaluation was used as the main basis for comparing Logistic Regression, Random Forest, and XGBoost\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFive-fold stratified cross-validation results on the combined imi_vm20i and imi_vmx30ui dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-score (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eROC-AUC (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.9091\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1408\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.8400\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.9000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.9044\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1196\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.9455\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.7000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.7333\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.9167\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1667\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.9091\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0575\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.7000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.6000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3742\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.6000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3266\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.9222\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1556\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe cross-validation results provide a more realistic picture than the earlier held-out split. \u003cb\u003eRandom Forest\u003c/b\u003e achieved the highest mean accuracy (\u003cb\u003e0.9455\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727\u003c/b\u003e), \u003cb\u003eLogistic Regression\u003c/b\u003e achieved the highest mean recall (\u003cb\u003e0.9000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2000\u003c/b\u003e) and mean F1-score (\u003cb\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2667\u003c/b\u003e), and \u003cb\u003eXGBoost\u003c/b\u003e achieved the highest mean ROC-AUC (\u003cb\u003e0.9222\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1556\u003c/b\u003e). These results indicate that the proposed MTConnect-derived feature space is informative for anomaly detection across the two compatible machine groups, but they also show that different models emphasize different strengths.\u003c/p\u003e \u003cp\u003eFrom a practical perspective, the results suggest that Random Forest offers the strongest average classification accuracy, while Logistic Regression provides a stronger balance for recovering anomaly cases. XGBoost, in contrast, ranks anomalous and normal runs effectively overall, but at the default decision threshold it shows lower recall and F1-score. This comparison highlights the multi-objective nature of the anomaly-detection problem: model selection depends on whether the operational priority is overall accuracy, anomaly sensitivity, or ranking quality.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Feature Importance Analysis\u003c/h2\u003e \u003cp\u003eTo better understand which MTConnect-derived variables contributed most strongly to anomaly detection, feature importance was extracted from the Random Forest model trained on the combined compatible-machine dataset. The ranked features were dominated by vibration-related variables together with runtime-related measures.\u003c/p\u003e \u003cp\u003eAmong the most influential variables were Program_Runtime_Seconds_std, Program_Runtime_Seconds_mean, smarms_min, ymarms_min, zmarms_mean, zmarms_min, ymapeak_min, and zmapeak_mean, along with selected spindle and load variables such as Spindle_Speed_mean, Feed_Rate_mean, and Y_Axis_Load_max.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo improve interpretation, the features were also grouped into four families: Vibration, Power/Energy/Runtime, Axis Load, and Process/Spindle. The grouped importance analysis showed that Vibration had the largest total contribution (0.6442), followed by Power/Energy/Runtime (0.2060), Axis Load (0.0784), and Process/Spindle (0.0713).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThese results indicate that anomaly detection in the compatible multi-machine dataset is driven primarily by dynamic process-instability signatures, especially vibration-related behavior, with additional support from runtime and energy-related characteristics. This is physically reasonable because the anomaly labels include events such as excessive vibration, clamp failure, tool breakage, chatter, and early stopping, all of which can affect machine dynamics and process duration. This emphasis on interpretability aligns with recent calls for explainable AI in manufacturing process monitoring [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Feature-Group Ablation Study\u003c/h2\u003e \u003cp\u003eAlthough the grouped importance analysis suggested that vibration features dominate the full-model importance structure, this does not necessarily mean that the vibration group alone provides the best standalone classification performance. To evaluate the contribution of each feature family more directly, an ablation study was performed using Random Forest and 5-fold stratified cross-validation. The following feature groups were evaluated independently: Vibration, Power/Energy/Runtime, Axis Load, Process/Spindle, and All Features.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAblation-study results using Random Forest and 5-fold cross-validation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNum. Features\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-score (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROC-AUC (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePower/Energy/Runtime\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.9636\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e1.0000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.8667\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.9333\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1333\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVibration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.9455\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.7000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.7333\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.9222\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1556\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAll Features\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.9455\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.7000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.7333\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.9167\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1667\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProcess/Spindle\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.9091\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1408\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.6000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4899\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.5000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4472\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.5333\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4522\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.7778\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2485\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxis Load\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.8561\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.2000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.3000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.2933\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.8911\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1800\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe ablation study revealed that Power/Energy/Runtime was the strongest standalone feature family, achieving the highest mean accuracy (0.9636), mean F1-score (0.8667), and mean ROC-AUC (0.9333). Vibration also performed strongly but did not outperform the Power/Energy/Runtime subset. Interestingly, the All-Features configuration did not improve over the best standalone group, suggesting that a compact and physically meaningful feature subset may be more effective than a larger aggregated feature space for this anomaly-detection task.\u003c/p\u003e \u003cp\u003eThis finding complements the feature-importance analysis rather than contradicting it. While vibration-related variables dominate the total importance structure of the full model, the ablation study shows that Power/Energy/Runtime forms the most effective standalone feature family. This suggests that vibration contributes many individually important variables within the full model, while Power/Energy/Runtime provides a more compact and directly discriminative feature set when used on its own.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e5.6 Heterogeneous-Machine Robustness Analysis\u003c/h2\u003e \u003cp\u003eTo assess robustness under a structurally different machine-data profile, a separate experiment was conducted using the tmf_vf10 machine group. Unlike imi_vm20i and imi_vmx30ui, this machine had a reduced and differently structured MTConnect-derived feature set, with only \u003cb\u003e13 runs\u003c/b\u003e in total (\u003cb\u003e8 normal\u003c/b\u003e and \u003cb\u003e5 anomalous\u003c/b\u003e). Because of the small sample size, \u003cb\u003eLeave-One-Out Cross-Validation (LOOCV)\u003c/b\u003e was used instead of a standard train-test split or 5-fold cross-validation.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHeterogeneous-machine robustness results on tmf_vf10 using LOOCV.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy_mean\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision_mean\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall_mean\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1_mean\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.4615\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.0769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0769\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.6154\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5385\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1538\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe heterogeneous-machine results were substantially weaker than those obtained on the compatible multi-machine dataset. Among the evaluated models, Random Forest achieved the highest mean accuracy (0.6154), while all three models showed low precision, recall, and F1-score. Since each LOOCV fold contained only one test sample, ROC-AUC was not emphasized for this experiment.\u003c/p\u003e \u003cp\u003eThese findings are important because they define the boundary of the proposed framework. The results suggest that the approach works well across compatible machine groups with a shared MTConnect-derived feature structure, but degrades when applied to a structurally different machine group with reduced feature richness and different signal semantics. In other words, data standardization supports interoperability at the structural level, but effective anomaly detection additionally requires sufficient feature compatibility across machines.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e5.7 Summary of Findings\u003c/h2\u003e \u003cp\u003eThe staged evaluation produced three main findings. First, the single-machine baseline confirmed that experiment-level statistical features extracted from MTConnect-derived data can support meaningful anomaly classification. Second, the compatible multi-machine analysis showed strong cross-validated performance, with Random Forest achieving the highest mean accuracy (0.9455), Logistic Regression the highest recall (0.9000) and F1-score (0.8000), and XGBoost the highest ROC-AUC (0.9222). Feature-importance analysis identified vibration-related variables as the most influential in the full model, while the ablation study showed that Power/Energy/Runtime provided the strongest standalone performance. Third, the tmf_vf10 robustness experiment showed a clear performance drop, indicating that effective anomaly detection depends on feature compatibility across machine groups, not only on data standardization.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e5.8 Limitations\u003c/h2\u003e \u003cp\u003eSeveral limitations of this study should be acknowledged. First, the dataset used in this work is relatively small. The main combined dataset contains 56 runs with only 9 anomalous cases, and the robustness dataset contains 13 runs. Although stratified cross-validation and LOOCV were used to mitigate the effect of limited samples, the high standard deviations observed in some metrics (e.g., Recall of 0.9000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2000 for Logistic Regression) indicate that the performance estimates remain sensitive to fold composition. Future studies with larger and more balanced datasets would strengthen the generalizability of these findings. Oversampling techniques such as those proposed for imbalanced pattern recognition [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] could also be explored to address class imbalance in future extensions of this framework.\u003c/p\u003e \u003cp\u003eSecond, only supervised classification methods were used in this study. In many real manufacturing settings, labeled anomaly data may be scarce or unavailable, which would favor unsupervised or semi-supervised approaches such as Isolation Forest or One-Class SVM. The supervised framing adopted here was appropriate given that experiment-level labels were available from documented run remarks, but extending the framework to unsupervised settings would broaden its practical applicability.\u003c/p\u003e \u003cp\u003eThird, the models were trained using default or minimally tuned hyperparameters. This choice was intentional to keep the focus on feature representation and machine compatibility rather than model-specific optimization. However, systematic hyperparameter tuning could potentially improve the reported results, particularly for XGBoost, which is known to be sensitive to parameter settings.\u003c/p\u003e \u003cp\u003eFourth, the statistical features used in this study (mean, standard deviation, minimum, maximum) provide a compact experiment-level summary but do not capture temporal patterns within a machining run. Anomalies that develop gradually or occur only during specific phases of a cut may not be fully represented by these summaries.\u003c/p\u003e \u003cp\u003eFinally, the cross-machine evaluation was limited to three machine groups from a single public dataset. While this was sufficient to demonstrate the framework and examine feature-compatibility effects, broader validation across additional machine types, facilities, and operating conditions would be needed to confirm the scalability of the approach.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study presented an AI/ML-based anomaly-detection framework using MTConnect-derived manufacturing data from multiple milling machine groups. Anomaly labels were assigned at the experiment level using documented run remarks, and statistical features were extracted from process, spindle, load, vibration, energy, and runtime variables. The framework was evaluated through a staged design covering single-machine baseline validation, compatible multi-machine classification, interpretability analysis, and heterogeneous-machine robustness testing.\u003c/p\u003e \u003cp\u003eThe results demonstrated that MTConnect-derived features can support effective anomaly detection when the machine groups share a compatible feature structure. The compatible multi-machine analysis produced strong cross-validated classification performance across all three models, while the interpretability analysis revealed that both vibration signatures and power/energy/runtime characteristics carry meaningful anomaly-related information. The finding that a compact feature subset (Power/Energy/Runtime) outperformed the full feature set in the ablation study has practical implications, suggesting that smaller, physically meaningful feature groups may be preferable to larger aggregated representations in some manufacturing contexts.\u003c/p\u003e \u003cp\u003eEqually important was the robustness experiment, which showed that performance degrades substantially when the machine-data structure changes. This result highlights a distinction that is often overlooked in the literature: MTConnect supports interoperability at the data-structure level, but effective cross-machine anomaly detection still requires sufficient feature compatibility. Standardized data is a necessary foundation, but not a sufficient one.\u003c/p\u003e \u003cp\u003eThese findings contribute to the growing body of work on smart manufacturing monitoring by providing empirical evidence of both the potential and the practical limits of MTConnect-derived anomaly detection across machines. Future work should focus on extending the framework to larger and more diverse machine populations, exploring transfer-learning or domain-adaptation strategies for heterogeneous machine-data profiles, incorporating temporal and multimodal signal representations, and moving toward online or near-real-time deployment for shop-floor decision support.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eATC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eApparent Tardiness Cost (dispatching rule)\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCMM\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCoordinate Measuring Machine\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCNC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eComputer Numerical Control\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCPFR\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCollaborative Planning, Forecasting, and Replenishment\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eDT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDigital Twin\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eEDD\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEarliest Due Date (dispatching rule)\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eEV\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eElectric Vehicle\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eIIoT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIndustrial Internet of Things\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eKPI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eKey Performance Indicator\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMILP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMixed-Integer Linear Programming\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMQTT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMessage Queuing Telemetry Transport\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMTConnect\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAn open, royalty-free standard for streaming manufacturing equipment data\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOEE\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOverall Equipment Effectiveness\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOEM\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOriginal Equipment Manufacturer\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOPC UA\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOpen Platform Communications Unified Architecture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOTFR\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOn-Time Fulfillment Rate\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003ePLC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eProgrammable Logic Controller\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eSKU\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eStock Keeping Unit\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eTier-1\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFirst-tier direct supplier to an OEM in a manufacturing supply chain\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eVMI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eVendor-Managed Inventory\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eWIP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWork-In-Process\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eWT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWeighted Tardiness\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHarshkumar K. Parmar contributed to conceptualization, methodology, software implementation, formal analysis, data curation, visualization, and writing - original draft preparation. Shivakumar Raman contributed to supervision, technical guidance, review, and editing. Both authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Generative AI in Scientific Writing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this manuscript, the authors used ChatGPT to assist with language editing and structural revision. The authors reviewed and edited the output as needed and take full responsibility for the final content of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study is publicly available and was originally published by Str\u0026ouml;bel et al. [21]. The experimental analysis was conducted using Python with scikit-learn and XGBoost. Feature engineering and statistical summarization were performed using pandas and NumPy. The code and processed feature tables can be made available upon reasonable request to the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Generative AI in Scientific Writing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this manuscript, I used an AI-based language tool (ChatGPT) to assist with language editing and structural revision.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eEl\u0026iacute;a I, Pagola M (2025) Anomaly detection in Smart-manufacturing era: A review. Eng Appl Artif Intell 139:109578Part B\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMTConnect, Institute (2026) \u003cem\u003eMTConnect Standard Part 1.0 \u0026ndash; Fundamentals\u003c/em\u003e, Version 2.5.0, Jan\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLei P, Zheng L, Li C, Li X (2016) MTConnect Enabled Interoperable Monitoring System for Finish Machining Assembly Interfaces of Large-scale Components. Procedia CIRP 56:378\u0026ndash;383\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLei P, Zheng L, Wang L, Wang Y, Li C, Li X (2017) MTConnect compliant monitoring for finishing assembly interfaces of large-scale components: A vertical tail section application. J Manuf Syst\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu C, Xu X, Peng Q, Zhou Z (2018) MTConnect-based Cyber-Physical Machine Tool: A case study. Procedia CIRP 72:492\u0026ndash;497\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiao L, Minhas R, Rangarajan A, Kurtoglu T, de Kleer J (2014) A Self-Aware Machine Platform in Manufacturing Shop Floor Utilizing MTConnect Data, \u003cem\u003eProceedings of the Annual Conference of the PHM Society\u003c/em\u003e, vol. 6, no. 1\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVenkatesh S, Ly S, Manning M, Michaloski J, Proctor F Automating Asset Knowledge with MTConnect, in \u003cem\u003eProc. ASME 2016 Int. Manufacturing Science and Engineering Conference (MSEC\u003c/em\u003e(2016)), Blacksburg, VA, USA, 2016, Paper No. MSEC2016-8629, doi: ), Blacksburg, VA, USA, 2016, Paper No. MSEC2016-8629. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1115/MSEC2016-8629\u003c/span\u003e\u003cspan address=\"10.1115/MSEC2016-8629\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHelu M, Hedberg T Jr., Barnard Feeney A (2017) Reference architecture to integrate heterogeneous manufacturing systems for the digital thread. CIRP J Manufact Sci Technol 19:191\u0026ndash;195\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTheissler A, P\u0026eacute;rez-Vel\u0026aacute;zquez J, Kettelgerdes M, Elger G (2021) Predictive maintenance enabled by machine learning: Use cases and challenges in the automotive industry. Reliab Eng Syst Saf 215:107864\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKausik AK, Rashid AB, Baki RF, Maktum MMJ (2025) Machine learning algorithms for manufacturing quality assurance: A systematic review of performance metrics and applications. Array 26:100393\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKamm S, Veekati SS, M\u0026uuml;ller T, Jazdi N, Weyrich M (2023) A survey on machine learning based analysis of heterogeneous data in industrial automation. Comput Ind 149:103930\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Giorgio A, Cola G, Wang L (2023) Systematic review of class imbalance problems in manufacturing. J Manuf Syst 71:620\u0026ndash;644. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmsy.2023.10.014\u003c/span\u003e\u003cspan address=\"10.1016/j.jmsy.2023.10.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMichau G, Fink O (2021) Unsupervised transfer learning for anomaly detection: Application to complementary operating condition transfer, \u003cem\u003eKnowledge-Based Systems\u003c/em\u003e, vol. 216, Art. 106816. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.knosys.2021.106816\u003c/span\u003e\u003cspan address=\"10.1016/j.knosys.2021.106816\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Lang Z-Q, Gui Y, Zhu Y-P, Laalej H (2024) Digital twin-based anomaly detection for real-time tool condition monitoring in machining. J Manuf Syst 75:163\u0026ndash;173. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmsy.2024.06.004\u003c/span\u003e\u003cspan address=\"10.1016/j.jmsy.2024.06.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie Z, Zhang Z, Chen J, Feng Y, Pan X, Zhou Z, He S (2024) Data-driven unsupervised anomaly detection of manufacturing processes with multi-scale prototype augmentation and multi-sensor data. J Manuf Syst 77:26\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmsy.2024.08.027\u003c/span\u003e\u003cspan address=\"10.1016/j.jmsy.2024.08.027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrabi M, Tran KP, Egger P, Thomassey S (2024) Anomaly detection in smart manufacturing: An Adaptive Adversarial Transformer-based model. J Manuf Syst 77:591\u0026ndash;611. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmsy.2024.09.021\u003c/span\u003e\u003cspan address=\"10.1016/j.jmsy.2024.09.021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchueller A, Salda\u0026ntilde;a C (2022) Generalizability analysis of tool condition monitoring ensemble machine learning models. J Manuf Process 84. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmapro.2022.10.064\u003c/span\u003e\u003cspan address=\"10.1016/j.jmapro.2022.10.064\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohanraj T, Yerchuru J, Krishnan H, Aravind RSN, Yameni R (2021) Development of tool condition monitoring system in end milling process using wavelet features and Hoelder\u0026rsquo;s exponent with machine learning algorithms. Measurement 173:108671. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.measurement.2020.108671\u003c/span\u003e\u003cspan address=\"10.1016/j.measurement.2020.108671\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohannssen A, Qiu P, Yeganeh A, Chukhrova N (2025) Explainable AI for trustworthy intelligent process monitoring. Comput Ind Eng 209:111407. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.cie.2025.111407\u003c/span\u003e\u003cspan address=\"10.1016/j.cie.2025.111407\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGross D, Spieker H, Gotlieb A, Knoblauch R, Elmansori M (2024) Efficient Milling Quality Prediction with Explainable Machine Learning. IFAC-PapersOnLine 58:43\u0026ndash;48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ifacol.2024.09.085\u003c/span\u003e\u003cspan address=\"10.1016/j.ifacol.2024.09.085\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStr\u0026ouml;bel R, Kuck M, Oexle F, Kader H, Puchta A, Noack B, Fleischer J (2025) A multimodal dataset for process monitoring and anomaly detection in industrial CNC milling. Data Brief 63:112207. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.dib.2025.112207\u003c/span\u003e\u003cspan address=\"10.1016/j.dib.2025.112207\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarshkumar K, Parmar, Raman S (2025) MTConnect for Communication within and across Manufacturing Enterprises, in \u003cem\u003eIISE Annual Conference. Proceedings\u003c/em\u003e, pp. 1\u0026ndash;11, Institute of Industrial and Systems Engineers (IISE)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhsan MM, Raman S, Liu Y, Siddique Z (2024) Defect analysis of 3D printed object using transfer learning approaches. Expert Syst Appl 253:124293\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhsan MM, Raman S, Siddique Z (2023) Bsgan: A novel oversampling technique for imbalanced pattern recognitions, \u003cem\u003earXiv preprint arXiv:2305.09777\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSamadi H, Ahsan MM, Raman S (2025) Metrology and manufacturing-integrated digital twin (MM-DT) for advanced manufacturing: Insights from coordinate measuring machine (CMM) and FARO arm measurements. Next Res 2(2):100299\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-advanced-manufacturing-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jamt","sideBox":"Learn more about [The International Journal of Advanced Manufacturing Technology](https://www.springer.com/journal/170)","snPcode":"170","submissionUrl":"https://submission.nature.com/new-submission/170/3","title":"The International Journal of Advanced Manufacturing Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"MTConnect, anomaly detection, smart manufacturing, machine learning, milling, interoperable manufacturing data, predictive monitoring","lastPublishedDoi":"10.21203/rs.3.rs-9358105/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9358105/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEarly detection of abnormal process behavior is critical in smart manufacturing, as it contributes to reduced downtime, improved process reliability, and prevention of quality-related losses. However, many anomaly-detection studies in manufacturing still rely on isolated sensor datasets or single-machine setups, limiting their extensibility across different machines and production environments. This leaves a gap in leveraging standardized MTConnect-derived data for anomaly detection in a more interoperable multi-machine setting. To address this gap, this study develops an AI/ML-based anomaly-detection framework using MTConnect-derived data from multiple milling machine groups. Anomaly labels are assigned at the experiment level using documented run remarks, and statistical features are extracted from machine, spindle, load, vibration, energy, and runtime variables. The study first examines a single-machine baseline and then expands to a combined dataset built from two feature-compatible machine groups, imi_vm20i and imi_vmx30ui. Logistic Regression, Random Forest, and XGBoost are used for anomaly classification. In 5-fold stratified cross-validation on the combined dataset, Random Forest achieved the highest mean accuracy (0.9455\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0727), Logistic Regression achieved the highest recall (0.9000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2000) and F1-score (0.8000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.2667), and XGBoost achieved the highest ROC-AUC (0.9222\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1556). Feature-importance analysis showed that vibration-related variables were the most influential in the full model, while an ablation study showed that Power/Energy/Runtime features performed best as a standalone group. A separate robustness experiment on the structurally different tmf_vf10 machine group showed a clear drop in performance, highlighting the importance of feature compatibility for interoperable anomaly detection. Overall, the results show that MTConnect-derived data can support anomaly detection across compatible machine groups, while also revealing the limits of transferability when machine-data structure changes.\u003c/p\u003e","manuscriptTitle":"AI/ML Based Interoperable Anomaly Detection in Advanced Manufacturing Using MTConnect-Derived Multi-Machine Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-21 21:34:07","doi":"10.21203/rs.3.rs-9358105/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major Revisions Needed","date":"2026-04-27T08:34:42+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2026-04-14T12:30:35+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-14T12:20:37+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-10T09:25:37+00:00","index":"","fulltext":""},{"type":"submitted","content":"The International Journal of Advanced Manufacturing Technology","date":"2026-04-09T15:13:14+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-advanced-manufacturing-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jamt","sideBox":"Learn more about [The International Journal of Advanced Manufacturing Technology](https://www.springer.com/journal/170)","snPcode":"170","submissionUrl":"https://submission.nature.com/new-submission/170/3","title":"The International Journal of Advanced Manufacturing Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"4014e824-3c60-4253-9d9d-cef4aec77000","owner":[],"postedDate":"April 21st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T12:46:06+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-21 21:34:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9358105","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9358105","identity":"rs-9358105","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}