Towards Smart Aluminum Smelting: An AI-Driven Approach for Real Time Thermal Anomaly Detection Using Distributed Optical Fiber Sensors

Towards Smart Aluminum Smelting: An AI-Driven Approach for Real Time Thermal Anomaly Detection Using Distributed Optical Fiber Sensors | 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 Method Article Towards Smart Aluminum Smelting: An AI-Driven Approach for Real Time Thermal Anomaly Detection Using Distributed Optical Fiber Sensors Songsong Wang, Yueqiang Zhu, Zhengguo Xu, Tiejun Wang, Wei Zheng, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7993883/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Abnormal temperature rise in the cathode steel bars of electrolytic aluminum cells, a core smelting equipment, is a major cause of furnace leakage accidents. However, traditional thermocouple and infrared temperature measurement technologies are susceptible to electromagnetic interference under extreme operating conditions such as strong magnetic fields, high temperatures, and severe corrosion, making continuous and accurate monitoring difficult. To address this technical bottleneck, this study developed an intelligent monitoring system that integrates distributed fiber optic sensing and stacking ensemble learning. By deploying electromagnetic interference-resistant fiber optic sensors on the surface of the cathode steel bars, continuous temperature data acquisition was achieved. A multidimensional feature set was constructed, including time series lag, rolling statistics, and periodic features, and the prediction performance of models such as CatBoost, LightGBM, and random forest was systematically compared. Ultimately, a stacking ensemble strategy was employed to combine the strengths of each base model to achieve accurate cathode steel bar temperature prediction (RMSE 0.99 on the test set). This system can proactively identify abnormal temperature rises, reducing warning time by over 85% compared to manual inspections, providing a reliable technical path for intelligent safety monitoring in electrolytic aluminum production. Materials Informatics material characterization advanced materials composites Materials Science Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION The aluminum electrolysis industry serves as a critical pillar in the supply chain of fundamental raw materials [ 1 – 5 ] , where operational safety, stability, and efficiency are of paramount importance. Within this industry, the electrolytic cell operates under extreme conditions—including high temperatures (∼950°C), intense magnetic fields, and highly corrosive environments—posing significant challenges to structural integrity and continuous operation. Under such conditions, abnormal temperature rises in cathode steel bars act as a primary precursor to lining failure, molten metal penetration, and catastrophic leakage accidents. Thus, the development of continuous, accurate, and real-time thermal monitoring techniques is essential to ensure safe operation and enable intelligent predictive maintenance in aluminum production facilities. Conventional temperature monitoring methods, such as thermocouples and infrared pyrometry, exhibit considerable limitations under these extreme industrial settings. Thermocouples, being electrically based, are highly susceptible to strong electromagnetic interference, leading to signal distortion and unreliable data acquisition. Moreover, their point-sensing nature restricts spatial resolution and large-scale deployment. Although non-contact infrared thermometry avoids direct exposure, it suffers from intermittent sampling, an inability to provide real-time continuous monitoring, and significant operational safety risks for personnel. Recent attempts using wireless sensing still rely on electronic principles and have not fundamentally overcome electromagnetic vulnerability. These inherent drawbacks severely limit the ability to achieve early warning of thermal anomalies, underscoring the urgent need for robust monitoring solutions compatible with modern smart smelting processes. In response to these challenges, distributed optical fiber sensing (DOFS) [ 6 – 11 ] has emerged as a promising alternative. Leveraging optical fibers as both sensor and signal transmission medium, this technology exhibits intrinsic immunity to electromagnetic interference, high temperature tolerance, and corrosion resistance. Notably, Raman scattering-based distributed temperature sensing (DTS) systems enable continuous temperature profiling along the entire length of the fiber with meter-scale spatial resolution. Such capability provides an ideal solution for detecting localized hot spots on cathode steel bars. Successful applications in pipeline integrity management and power cable monitoring have established a solid foundation for its implementation in harsh industrial environments such as aluminum electrolysis cells. Despite the advantages of DOFS in data acquisition, effectively translating high-volume spatiotemporal temperature data into reliable early-warning insights remains a nontrivial task. While conventional machine learning models (e.g., Random Forest [ 14 ] , Gradient Boosting [ 12 ] ) have been widely used for temperature prediction, they often encounter performance ceilings in complex, multi-variable industrial settings. Ensemble learning strategies, particularly Stacking integration, have demonstrated superior generalization ability by combining the strengths of multiple base learners (e.g., CatBoost, LightGBM [ 13 ] ) and leveraging a meta-learner to optimize final predictions. Integrating such advanced machine learning architectures with high-fidelity fiber sensing data presents a compelling pathway toward building a truly data-driven early-warning system. This study aims to develop an intelligent monitoring and early-warning framework for cathode steel bar temperature in aluminum electrolytic cells by integrating distributed fiber optic sensing with a Stacking-based ensemble learning architecture. The main contributions include: (1) deploying an anti-interference distributed optical fiber temperature sensing system to achieve real-time, full-coverage temperature field acquisition; (2) constructing a multi-dimensional feature set incorporating temporal, statistical, and periodic characteristics from the denoised temperature signals; (3) systematically evaluating and comparing the performance of individual machine learning models such as CatBoost, LightGBM, and Random Forest, and further proposing a Stacking ensemble framework to synergistically combine their predictive strengths; and (4) validating the system's capability for early detection of abnormal temperature rises and significant reduction of warning time through field experiments, thereby providing a reliable and intelligent solution for safety monitoring in aluminum electrolysis production. MATERIALS AND METHODS A distributed fiber optic temperature measurement system was deployed on the cathode steel bars of a 380 kA aluminum electrolytic cell. Temperature prediction was performed using CatBoost, LightGBM, Random Forest, and a Bayesian-optimized multi-layer perceptron (MLP) [ 15 ] hyperparameter-based single model. A stacking ensemble learning strategy was also employed for temperature prediction, with CatBoost, LightGBM, and Random Forest as base learners and linear regression as a meta-learner. Data Collection To obtain continuous temperature data from the cathode steel rods of aluminum electrolytic cells under strong magnetic fields and high temperatures, this study deployed a distributed fiber optic temperature measurement system on the bottom of the all-graphite cathode steel rods in a 380 kA prebaked aluminum electrolytic cell. The system used a distributed fiber optic temperature measurement device (model: DTS-BLY-5S(STD)) with a spatial resolution of 1 m. A polyimide-armored multimode optical fiber (model: MM62.5/125/155HTPI, temperature resistance 380°C) was coiled into a ring structure and placed in a custom packaging device to ensure close contact between the fiber and the cathode steel rod. The contact length between the packaging and the cathode steel rod was greater than 1 meter. All temperature measurement data was transmitted to the industrial computer in the control room via the insulated optical fiber to prevent electromagnetic interference. Table 1 Distributed fiber optic temperature measurement system Parameters Value Remark Temperature measurement principle ROTDR Temperature measurement range 0-450°C Temperature esolution 1°C Spatial resolution 1m Feature Engineering To comprehensively characterize the dynamic characteristics of the cathode steel rod temperature field in the electrolytic cell, this study constructed a feature set from multiple dimensions: time, space, and frequency. Feature engineering primarily involves constructing features based on physical mechanisms and deep time series feature mining. First, based on the electrolytic aluminum process mechanism, the following physically interpretable features were constructed: The electrolytic aluminum production process is influenced by factors such as diurnal load fluctuations in the power grid and periodic pole-changing operations, exhibiting significant periodicity. The timestamps of the data points were converted into cyclical features using sine/cosine transforms to effectively encode temporal information. This method maps linear time information onto a periodic curve, enabling the model to identify periodic temperature variations (such as daytime and nighttime differences) while avoiding the boundary discontinuities inherent in directly encoding discrete time features (such as hours and minutes). To capture the inertial dependencies of temperature changes, features incorporating historical temperature values were constructed. Lag features characterize the correlation between the current temperature and data from preceding time steps, facilitating the model's learning of temperature autocorrelation characteristics, which are crucial for predicting short-term temperature evolution. To extract trend information within localized time windows, sliding window statistics were calculated. The rolling mean smooths random fluctuations and highlights short-term trends; the rolling standard deviation quantifies the intensity of instantaneous temperature fluctuations and serves as an effective metric for identifying anomalous operating conditions e.g., abrupt temperature changes). Following feature construction, samples containing missing values resulting from lag and sliding window operations were removed to ensure data quality. The resultant feature matrix comprises the original temperature signal, physical features, periodic features, lag features, and statistical features, providing an information-rich input for the machine learning model that is both physically interpretable and temporally regular. Table 2 Aluminum electrolytic cell operating conditions Parameters Value Remark Groove type 380 KA Electrolysis temperature 935 ~ 965°C Molecular ratio 2.2 ~ 2.4 Collection duration 2 months Machine Learning Methods This section details the machine learning models employed, key hyperparameter configurations, and the ensemble learning strategy utilized in this study. Four distinct prediction models were selected and optimized. Their key hyperparameter configurations are summarized below: CatBoost, specialized in processing categorical features with inherent robustness to gradient bias, was configured with 1000 iterations, a learning rate of 0.1, and a decision tree depth of 6. LightGBM, an efficient gradient boosting framework implementation utilizing a leaf-wise growth algorithm, employed 31 leaf nodes and a feature sampling fraction of 0.9. Random Forest, representing the Bagging (Bootstrap Aggregating) ensemble algorithm, constructs multiple decision trees and utilizes a voting mechanism for final prediction, demonstrating strong resistance to overfitting. Its hyperparameters included 100 decision trees and a minimum of 5 samples required for internal node splitting. Bayesian optimization was employed to efficiently identify the optimal hyperparameter combination for the Multilayer Perceptron (MLP). The optimized architecture featured hidden layer sizes of (17, 43) and a learning rate of 0.00001. A stacking ensemble strategy was implemented, characterized by the following structural design: optimized CatBoost, LightGBM, and Random Forest models functioned as base learners in the first layer to capitalize on the distinct advantages of each algorithm in capturing data patterns. A Linear Regression model served as the meta-learner in the second layer, assigned to learn the optimal combination of base learner predictions. To produce unbiased meta-features for training the meta-learner and rigorously prevent information leakage, the predictions from the base learners corresponding to the training samples were produced using a 5-fold cross-validation procedure. Table 3. Machine learning model hyperparameter settings Model Key hyperparameters value Optimization methods CatBoost iterations 1000 learning_rate 0.1 depth 6 LightGBM num_leaves 31 Bayesian Optimization bagging_freq 5 feature_fraction 0.9 Random Forest n_estimators 100 min_samples_split 5 MLP random_state 42 Bayesian Optimization learning_rate_init 0.00001 RESULTS AND DISCUSSION Table 4 Performance evaluation indicators of each model on the test set Model MSE RMSE (°C) ​​MAE (°C)​ R² CatBoost 1.4231 1.1930 0.9490 0.9857 LightGBM 2.9052 1.7045 1.3587 0.9706 Random Forest 1.8041 1.3432 1.0819 0.9819 MLP 5.1418 2.2676 1.7866 0.9483 Stacking 0.5768 0.7451 0.5184 0.9944 Overall Model Performance and Comparative Analysis The quantitative evaluation of predictive models on the test set is systematically summarized in Table 4 . The Stacking ensemble model demonstrated superior performance across all metrics, achieving the lowest MSE (0.5768), RMSE (0.7451°C), and MAE (0.5184°C), alongside the highest R² value (0.9944). This signifies an nearly perfect fit to the true temperature distribution of the cathode steel bars. Among the base learners, CatBoost exhibited the strongest individual performance (R² = 0.9857, RMSE = 1.1930°C), followed by Random Forest (R² = 0.9819, RMSE = 1.3432°C), LightGBM, and MLP. The substantial performance gap between the Stacking model and the best single model (CatBoost) underscores the critical advantage of the ensemble strategy in this complex industrial prediction task. Interpretability and Feature Contribution of Base Learners The exceptional capability of the Stacking framework stems from its effective integration of diverse base learners. The high inter-model correlations (0.998–0.999) depicted in Fig. 1 (the scatter matrix of base learner predictions) indicate that CatBoost, LightGBM, and Random Forest capture highly consistent overall temperature trends. This consistency provides a stable foundation for the meta-learner's integration. More importantly, SHAP analysis (Figures S1, S2 and S3) reveals the distinct feature focus of each base learner, explaining their complementary nature. For instance, the CatBoost model (Figure S1) assigns paramount importance to rolling_mean_10and lagged features such as lag_9and lag_7, highlighting its strength in leveraging short-term temporal trends. In contrast, the LightGBM model (Figure S2) also prioritizes rolling_mean_10but shows a different weighting of subsequent lag features (e.g., lag_5, lag_9), indicating a marginally different perception of temporal dependencies. Similarly, the Random Forest model (Figure S3) demonstrates its unique feature interaction patterns. This diversity in feature importance, despite high prediction correlation, is a key source of the Stacking model's robustness, allowing it to mitigate individual model biases. Integration Mechanism and Meta-Learner Analysis The core of the Stacking approach lies in the optimal combination of these base learners by the meta-learner. The strong correlations shown in the meta-learner correlation matrix (Fig. 3 ) confirm the high alignment in prediction directions among base models. The meta-learner's feature importance analysis (Fig. 4 ), which displays the linear regression coefficients, provides direct insight into the integration strategy. The significantly higher weight assigned to the LightGBM predictions suggests that the meta-learner identifies its output as the most reliable anchor or baseline prediction among the ensemble. The weights for CatBoost and Random Forest are then adjusted to correct residual errors and capture nuances missed by LightGBM. This weighted combination, visualized conceptually in Fig. 4 , effectively reduces variance and enhances generalization, leading to the observed performance leap. The use of 5-fold cross-validation during the generation of meta-features ensured that this integration was learned without data leakage, guaranteeing the model's validity on unseen data. Engineering Implications for Intelligent Manufacturing The practical value of this Stacking-based monitoring system is profound for intelligent manufacturing in the aluminum industry. The achieved MAE of 0.52°C and RMSE of 0.75°C (Table 4 , Figure S1) represent a critical threshold for reliable early-warning systems. In practical terms, this high precision allows for the detection of subtle abnormal thermal rises long before they escalate into potential hazards like lining failure, thereby providing a crucial window for preventive maintenance. The model's interpretability, afforded by the SHAP analysis, is equally valuable. It transforms the model from a "black box" into a decision-support tool. Plant operators can not only receive alerts but also understand the contributing factors (e.g., a sharp change captured by rolling_std_10or a persistent deviation in rolling_mean_10), facilitating root cause analysis and informed operational decisions. This aligns perfectly with the core objectives of Industry 4.0—moving from reactive to predictive and ultimately prescriptive maintenance. CONCLUSIONS This study developed a distributed fiber-optic temperature measurement system for monitoring cathode steel bars in aluminum electrolytic cells, operating successfully under extreme conditions like intense EMI, high temperatures, and dust. It acquired high-frequency temperature time-series data from actual production. Multiple ML algorithms—CatBoost, LightGBM, RF, and Bayesian-optimized MLP—were applied and evaluated. Integrating fiber-optic sensing with stacking ensemble ML established a real-time monitoring system. Validation shows it proactively identifies abnormal temperature increases, reducing response time versus traditional methods, improving safety and energy control. This supports the industry's intelligent transformation. Future research will focus on transfer learning, multi-parameter fusion, and dynamic online learning. The system delivers a high-precision solution and a generalizable framework for other extreme industrial environments. Declarations Availability of data and materials Not applicable. Financial support and sponsorship This paper was supported by Smart Aluminum Plant Technological Innovation Project of Inner Mongolia Huomei-Hongjun Aluminum Electric Co., Ltd (K20231334) References Fu R (2023) Optimization of Electrolytic Aluminum Production Process by Intelligent Control. In 2023 International Seminar on Computer Science and Engineering Technology (SCSET) ; IEEE, ; pp 695–698. https://doi.org/10.1109/scset58950.2023.00157 . 智慧铝厂 Yao L, Long W, Li T, Tang D, Yi J, An Improved UKFNN (2019) Based on Iterative Renewal for Intelligent Modeling of Aluminum Electrolysis. In 2019 IEEE 4th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC) ; IEEE, ; pp 271–275. https://doi.org/10.1109/iaeac47372.2019.8998081 . 智慧铝厂 Sun Y, Gui W, Chen X, Xie Y, Xie S, Zou ZA (2023) Dynamic Spatial Distributed Information Clustering Method for Aluminum Electrolysis Cell. Eng Appl Artif Intell 126:106793. https://doi.org/10.1016/j.engappai.2023.106793 智慧铝厂 Wang J, Xie Y, Xie S, Chen X (2024) Eng Appl Artif Intell 134:108664. https://doi.org/10.1016/j.engappai.2024.108664 智慧铝厂 Development of Data-Knowledge-Driven Predictive Model and Multi-Objective Optimization for Intelligent Optimal Control of Aluminum Electrolysis Process Hancock J, Khoshgoftaar TM (2020) CatBoost for Big Data: An Interdisciplinary Review. Research Square. Springer Science and Business Media LLC August 19. https://doi.org/10.21203/rs.3.rs-54646/v1 Du W, Hu Z, Li C, Chen Z, Chen W, Qiu S (2023) Distributed Optical Fiber Sensing Technology in Operation Status Monitoring of GIL Equipment. Appl Math Nonlinear Sci 9(1). https://doi.org/10.2478/amns.2023.2.00274 OKAMOTO K (1994) Special Issue on Fiber-Optics. Fiber-Optic Distributed-Temperature Sensing. rle 22 (4), 276–283. https://doi.org/10.2184/lsj.22.276 Liu Y, Lei T, Sun Z, Wang C, Liu T (2012) Application of Distributed Optical Fiber Temperature System in Online Monitoring and Fault Diagnosis of Smart Grid. In 2012 Asia-Pacific Power and Energy Engineering Conference ; IEEE, ; pp 1–4. https://doi.org/10.1109/appeec.2012.6307675 . 分布式光纤 Du W, Hu Z, Li C, Chen Z, Chen W, Qiu S (2023) Distributed Optical Fiber Sensing Technology in Operation Status Monitoring of GIL Equipment. Appl Math Nonlinear Sci 9(1). https://doi.org/10.2478/amns.2023.2.00274 分布式光纤 Ma S, Xu Y, Pang Y, Zhao X, Li Y, Qin Z, Liu Z, Lu P, Bao X (2022) Optical Fiber Sensors for High-Temperature Monitoring: A Review. Sensors 22 (15), 5722. https://doi.org/10.3390/s22155722 . 分布式光纤 Deng Y, Jiang J (2022) Optical Fiber Sensors in Extreme Temperature and Radiation Environments: A Review. IEEE Sens J 22(14):13811–13834. https://doi.org/10.1109/jsen.2022.3181949 分布式光纤 Lian Z, Zhang C, Liu C, Zhao H, Liu F, He A, Bo F, Qi X, Jia X, Liu P, Chen C (2025) Intelligent Energy Optimization for Electrolytic Aluminum Using Industrial Data-Driven and Knowledge-Guided Modeling. Metall Mater Trans B 56(5):5187–5199. https://doi.org/10.1007/s11663-025-03691-9 智慧铝厂 Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu T-Y (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree, LightGBM Rigatti SJ, Random Forest (2017) J Insur Med 47(1):31–39. https://doi.org/10.17849/insm-47-01-31-39.1 Pinkus A (1999) Approximation Theory of the MLP Model in Neural Networks. Acta Numerica 8:143–195. https://doi.org/10.1017/s0962492900002919 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted 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-7993883","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":537790594,"identity":"591b12a1-c5b1-4b8e-a139-a839dabcd73c","order_by":0,"name":"Songsong Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1ElEQVRIiWNgGAWjYBACNvbmw4//VLDV27c3HyBOCx/PsTQDnjN8CQY8xxKI0yIn4WMgwdsil2AgkWNApMMkeAwMJBvM8swZcj7eeMNgJ6fbQEiLdFvBA8MdacWWDWc3W85hSDY2O0BIi8zhDQaJZ44xNhzs3SbNw3AgcRtBLRJAXxxs+8/YcJjnGbFaUgwkG9vYEjcc42EjUgswkI0ZzrAZS/awGVvOMSDCL/LtwKhkqGCT45d//PDGmwo7OYJaUAAwwElRDtFCqo5RMApGwSgYEQAAw0tCUWmCv30AAAAASUVORK5CYII=","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Songsong","middleName":"","lastName":"Wang","suffix":""},{"id":537791003,"identity":"ed92a638-421e-4542-af61-0ed11cda3681","order_by":1,"name":"Yueqiang Zhu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Yueqiang","middleName":"","lastName":"Zhu","suffix":""},{"id":537791004,"identity":"dff5c7dc-f972-440c-8ad1-05dacabfb192","order_by":2,"name":"Zhengguo Xu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Zhengguo","middleName":"","lastName":"Xu","suffix":""},{"id":537791005,"identity":"e3ef6ff9-2798-4094-8975-3588235569cc","order_by":3,"name":"Tiejun Wang","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Tiejun","middleName":"","lastName":"Wang","suffix":""},{"id":537791006,"identity":"cd93c6e0-aa55-4241-8522-19e134e0899a","order_by":4,"name":"Wei Zheng","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zheng","suffix":""},{"id":537791007,"identity":"7536722e-7b07-4f8a-a24a-d33118d86398","order_by":5,"name":"Wei Zhu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zhu","suffix":""},{"id":537791008,"identity":"77cd241a-633e-410d-8cbc-9518e0ab4126","order_by":6,"name":"Liangliang Lv","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Liangliang","middleName":"","lastName":"Lv","suffix":""},{"id":537791009,"identity":"0956674d-0627-4e74-b275-447d719adbb3","order_by":7,"name":"Bo Qiu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Bo","middleName":"","lastName":"Qiu","suffix":""},{"id":537791010,"identity":"629886d9-dbfb-45e7-b596-401b666e452b","order_by":8,"name":"Kailiang Pan","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Kailiang","middleName":"","lastName":"Pan","suffix":""}],"badges":[],"createdAt":"2025-10-31 03:13:54","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7993883/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7993883/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95224518,"identity":"4697d7db-fd39-4aa9-b0bd-ed565c445a09","added_by":"auto","created_at":"2025-11-05 16:23:50","extension":"doc","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":965147,"visible":true,"origin":"","legend":"","description":"","filename":"Paperv1.0.doc","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/bc8a72d3d286f513f3e364be.doc"},{"id":95224192,"identity":"d3206d1b-fca6-45ef-ba4b-8ce6ccbd47ff","added_by":"auto","created_at":"2025-11-05 16:23:29","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7993883.json","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/f33084e5b8de378fa8b1b67a.json"},{"id":95101938,"identity":"c52cfcc0-1544-437f-a10b-572205217035","added_by":"auto","created_at":"2025-11-04 10:13:02","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":63740,"visible":true,"origin":"","legend":"","description":"","filename":"rs79938830enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/5c63126076576d16929ee1d8.xml"},{"id":95101944,"identity":"e54abaa9-0890-4a61-8cb9-b02857876e0e","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":373209,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/ba128fa9d4876d6551798de9.png"},{"id":95223999,"identity":"f9134d97-32a6-4eb5-9860-b3bbeadf42a0","added_by":"auto","created_at":"2025-11-05 16:23:11","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":161914,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/e115027b8e755b2ac17c6dbc.png"},{"id":95101936,"identity":"8b8201ed-d49f-44d5-a063-48f31dadad07","added_by":"auto","created_at":"2025-11-04 10:13:02","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49934,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/989cd6109806a26e1035a40d.png"},{"id":95101946,"identity":"426bb258-615b-4377-8dc7-5fe53b70de6d","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":176235,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/6211cfb97184f0ec7ac627df.png"},{"id":95101941,"identity":"f75e624a-55ea-4e28-980e-837feb6318f2","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":311,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/35aff8f9260c895c32fda384.jpeg"},{"id":95101945,"identity":"325e886f-d012-451f-bfee-5eb90f14857a","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":311,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/263b38486a051f671181a22d.jpeg"},{"id":95224068,"identity":"243c0d57-753d-4693-97c4-284d3efba423","added_by":"auto","created_at":"2025-11-05 16:23:15","extension":"wmf","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3652,"visible":true,"origin":"","legend":"","description":"","filename":"image1.wmf","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/30ba01879b8de4f39666e120.wmf"},{"id":95225191,"identity":"77177d44-2740-4237-9932-600c71a674c9","added_by":"auto","created_at":"2025-11-05 16:24:41","extension":"wmf","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":36470,"visible":true,"origin":"","legend":"","description":"","filename":"image2.wmf","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/d965519b508113477ac7190c.wmf"},{"id":95224249,"identity":"d7c469ab-1580-4dfd-aba8-d09c193af3b8","added_by":"auto","created_at":"2025-11-05 16:23:31","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78351,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/22787c7be762bb6cad98b055.png"},{"id":95224700,"identity":"7fc9abe6-bb74-43ac-84ce-16fc99ff2807","added_by":"auto","created_at":"2025-11-05 16:24:11","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44424,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/c3eade407bf8cd95ec78c16e.png"},{"id":95101953,"identity":"32ea0529-8e0f-4b00-ab1c-d5e2b1595b00","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21518,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/b6b8cadb3da92420f8c1f9a0.png"},{"id":95224290,"identity":"9bf66cf8-f896-440b-9219-1b28ee572a76","added_by":"auto","created_at":"2025-11-05 16:23:34","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":48604,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/92ecb1cf213c6ea4ca00b271.png"},{"id":95101948,"identity":"69785ef9-684a-44a4-8688-c497c63459a8","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":346,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/ede5cc01ac42ea3a388327bf.png"},{"id":95101952,"identity":"88bb09a2-a8af-4b20-a77b-067ffd3c9436","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":346,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/09fafc1f6829356cfa45680b.png"},{"id":95101954,"identity":"b9dd23ac-7b29-4ae0-8d39-10e47bed7a22","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":778,"visible":true,"origin":"","legend":"","description":"","filename":"Onlineimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/c3c431096da48b2ee60f98bf.png"},{"id":95101956,"identity":"4861c3a6-5a0b-4646-bceb-e680e422a5c5","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1505,"visible":true,"origin":"","legend":"","description":"","filename":"Onlineimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/432f0fd058244adc73e38252.png"},{"id":95101958,"identity":"5fa35a09-500d-4452-8324-e392fa9542de","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62158,"visible":true,"origin":"","legend":"","description":"","filename":"rs79938830structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/9706eb96dcb02c05e876616e.xml"},{"id":95101955,"identity":"940dec84-c9c6-4cfb-96fe-d326ebb584d1","added_by":"auto","created_at":"2025-11-04 10:13:03","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70641,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/c593ddb5a8befb80b2519fb8.html"},{"id":95224440,"identity":"29c6a35e-994e-452d-97f9-2d3004829cd2","added_by":"auto","created_at":"2025-11-05 16:23:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":315964,"visible":true,"origin":"","legend":"\u003cp\u003eMultiple aluminum electrolytic cell layout: can be divided into 4, 8,16 or 32 line\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/7c1a24f9c846822d5ee82e43.png"},{"id":95101939,"identity":"b08f7573-3b00-43dc-aef9-6cfdf1b6d9c7","added_by":"auto","created_at":"2025-11-04 10:13:02","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":202960,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance evaluation of the Stacking ensemble model for temperature prediction. Quantitative metrics—MSE, RMSE, and R².\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/61b9ad4b6e015b92977b2e5e.png"},{"id":95101934,"identity":"e9c5fa34-e37b-4fd8-a7b5-3bd04f784948","added_by":"auto","created_at":"2025-11-04 10:13:02","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":56911,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation matrix of base learner predictions within the Stacking ensemble framework.​​ The heatmap visualizes the Pearson correlation coefficients between the prediction vectors of the three base learners: Random Forest, LightGBM, and CatBoost.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/eb6f636ddef48e6dbe1d0595.png"},{"id":95225297,"identity":"9e90af82-ae00-438b-a986-b872b02662a3","added_by":"auto","created_at":"2025-11-05 16:24:50","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":118045,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of base learner contributions and prediction consistency in the Stacking ensemble framework.​​ The top three panels show scatter plots comparing the predictions of individual base learners (Random Forest, LightGBM, and CatBoost) against the final Stacking model predictions\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/f01a79f01697e4cdab88f347.png"},{"id":95230474,"identity":"1e23269a-663e-4a0f-8389-194a90509d07","added_by":"auto","created_at":"2025-11-05 16:37:33","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1273362,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7993883/v1/65639700-37f1-4603-9d61-1e527ac0827f.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eTowards Smart Aluminum Smelting: An AI-Driven Approach for Real Time Thermal Anomaly Detection Using Distributed Optical Fiber Sensors\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eThe aluminum electrolysis industry serves as a critical pillar in the supply chain of fundamental raw materials\u003csup\u003e[\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e, where operational safety, stability, and efficiency are of paramount importance. Within this industry, the electrolytic cell operates under extreme conditions\u0026mdash;including high temperatures (\u0026sim;950\u0026deg;C), intense magnetic fields, and highly corrosive environments\u0026mdash;posing significant challenges to structural integrity and continuous operation. Under such conditions, abnormal temperature rises in cathode steel bars act as a primary precursor to lining failure, molten metal penetration, and catastrophic leakage accidents. Thus, the development of continuous, accurate, and real-time thermal monitoring techniques is essential to ensure safe operation and enable intelligent predictive maintenance in aluminum production facilities.\u003c/p\u003e\u003cp\u003eConventional temperature monitoring methods, such as thermocouples and infrared pyrometry, exhibit considerable limitations under these extreme industrial settings. Thermocouples, being electrically based, are highly susceptible to strong electromagnetic interference, leading to signal distortion and unreliable data acquisition. Moreover, their point-sensing nature restricts spatial resolution and large-scale deployment. Although non-contact infrared thermometry avoids direct exposure, it suffers from intermittent sampling, an inability to provide real-time continuous monitoring, and significant operational safety risks for personnel. Recent attempts using wireless sensing still rely on electronic principles and have not fundamentally overcome electromagnetic vulnerability. These inherent drawbacks severely limit the ability to achieve early warning of thermal anomalies, underscoring the urgent need for robust monitoring solutions compatible with modern smart smelting processes.\u003c/p\u003e\u003cp\u003eIn response to these challenges, distributed optical fiber sensing (DOFS)\u003csup\u003e[\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e has emerged as a promising alternative. Leveraging optical fibers as both sensor and signal transmission medium, this technology exhibits intrinsic immunity to electromagnetic interference, high temperature tolerance, and corrosion resistance. Notably, Raman scattering-based distributed temperature sensing (DTS) systems enable continuous temperature profiling along the entire length of the fiber with meter-scale spatial resolution. Such capability provides an ideal solution for detecting localized hot spots on cathode steel bars. Successful applications in pipeline integrity management and power cable monitoring have established a solid foundation for its implementation in harsh industrial environments such as aluminum electrolysis cells.\u003c/p\u003e\u003cp\u003eDespite the advantages of DOFS in data acquisition, effectively translating high-volume spatiotemporal temperature data into reliable early-warning insights remains a nontrivial task. While conventional machine learning models (e.g., Random Forest\u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e, Gradient Boosting\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e) have been widely used for temperature prediction, they often encounter performance ceilings in complex, multi-variable industrial settings. Ensemble learning strategies, particularly Stacking integration, have demonstrated superior generalization ability by combining the strengths of multiple base learners (e.g., CatBoost, LightGBM\u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e) and leveraging a meta-learner to optimize final predictions. Integrating such advanced machine learning architectures with high-fidelity fiber sensing data presents a compelling pathway toward building a truly data-driven early-warning system.\u003c/p\u003e\u003cp\u003eThis study aims to develop an intelligent monitoring and early-warning framework for cathode steel bar temperature in aluminum electrolytic cells by integrating distributed fiber optic sensing with a Stacking-based ensemble learning architecture. The main contributions include: (1) deploying an anti-interference distributed optical fiber temperature sensing system to achieve real-time, full-coverage temperature field acquisition; (2) constructing a multi-dimensional feature set incorporating temporal, statistical, and periodic characteristics from the denoised temperature signals; (3) systematically evaluating and comparing the performance of individual machine learning models such as CatBoost, LightGBM, and Random Forest, and further proposing a Stacking ensemble framework to synergistically combine their predictive strengths; and (4) validating the system's capability for early detection of abnormal temperature rises and significant reduction of warning time through field experiments, thereby providing a reliable and intelligent solution for safety monitoring in aluminum electrolysis production.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003eA distributed fiber optic temperature measurement system was deployed on the cathode steel bars of a 380 kA aluminum electrolytic cell. Temperature prediction was performed using CatBoost, LightGBM, Random Forest, and a Bayesian-optimized multi-layer perceptron (MLP)\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e hyperparameter-based single model. A stacking ensemble learning strategy was also employed for temperature prediction, with CatBoost, LightGBM, and Random Forest as base learners and linear regression as a meta-learner.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eData Collection\u003c/h2\u003e\u003cp\u003eTo obtain continuous temperature data from the cathode steel rods of aluminum electrolytic cells under strong magnetic fields and high temperatures, this study deployed a distributed fiber optic temperature measurement system on the bottom of the all-graphite cathode steel rods in a 380 kA prebaked aluminum electrolytic cell. The system used a distributed fiber optic temperature measurement device (model: DTS-BLY-5S(STD)) with a spatial resolution of 1 m. A polyimide-armored multimode optical fiber (model: MM62.5/125/155HTPI, temperature resistance 380\u0026deg;C) was coiled into a ring structure and placed in a custom packaging device to ensure close contact between the fiber and the cathode steel rod. The contact length between the packaging and the cathode steel rod was greater than 1 meter. All temperature measurement data was transmitted to the industrial computer in the control room via the insulated optical fiber to prevent electromagnetic interference.\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\u003eDistributed fiber optic temperature measurement system\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eParameters\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRemark\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTemperature measurement principle\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eROTDR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTemperature measurement range\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0-450\u0026deg;C\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTemperature esolution\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u0026deg;C\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpatial resolution\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1m\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eFeature Engineering\u003c/h3\u003e\n\u003cp\u003eTo comprehensively characterize the dynamic characteristics of the cathode steel rod temperature field in the electrolytic cell, this study constructed a feature set from multiple dimensions: time, space, and frequency. Feature engineering primarily involves constructing features based on physical mechanisms and deep time series feature mining.\u003c/p\u003e\u003cp\u003eFirst, based on the electrolytic aluminum process mechanism, the following physically interpretable features were constructed:\u003c/p\u003e\u003cp\u003eThe electrolytic aluminum production process is influenced by factors such as diurnal load fluctuations in the power grid and periodic pole-changing operations, exhibiting significant periodicity. The timestamps of the data points were converted into cyclical features using sine/cosine transforms to effectively encode temporal information. This method maps linear time information onto a periodic curve, enabling the model to identify periodic temperature variations (such as daytime and nighttime differences) while avoiding the boundary discontinuities inherent in directly encoding discrete time features (such as hours and minutes).\u003c/p\u003e\u003cp\u003eTo capture the inertial dependencies of temperature changes, features incorporating historical temperature values were constructed. Lag features characterize the correlation between the current temperature and data from preceding time steps, facilitating the model's learning of temperature autocorrelation characteristics, which are crucial for predicting short-term temperature evolution.\u003c/p\u003e\u003cp\u003eTo extract trend information within localized time windows, sliding window statistics were calculated. The rolling mean smooths random fluctuations and highlights short-term trends; the rolling standard deviation quantifies the intensity of instantaneous temperature fluctuations and serves as an effective metric for identifying anomalous operating conditions e.g., abrupt temperature changes). Following feature construction, samples containing missing values resulting from lag and sliding window operations were removed to ensure data quality. The resultant feature matrix comprises the original temperature signal, physical features, periodic features, lag features, and statistical features, providing an information-rich input for the machine learning model that is both physically interpretable and temporally regular.\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\u003eAluminum electrolytic cell operating conditions\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eParameters\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRemark\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGroove type\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e380 KA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eElectrolysis temperature\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e935\u0026thinsp;~\u0026thinsp;965\u0026deg;C\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMolecular ratio\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2.2\u0026thinsp;~\u0026thinsp;2.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCollection duration\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2 months\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003eMachine Learning Methods\u003c/h3\u003e\n\u003cp\u003eThis section details the machine learning models employed, key hyperparameter configurations, and the ensemble learning strategy utilized in this study.\u003c/p\u003e\u003cp\u003eFour distinct prediction models were selected and optimized. Their key hyperparameter configurations are summarized below:\u003c/p\u003e\u003cp\u003eCatBoost, specialized in processing categorical features with inherent robustness to gradient bias, was configured with 1000 iterations, a learning rate of 0.1, and a decision tree depth of 6. LightGBM, an efficient gradient boosting framework implementation utilizing a leaf-wise growth algorithm, employed 31 leaf nodes and a feature sampling fraction of 0.9. Random Forest, representing the Bagging (Bootstrap Aggregating) ensemble algorithm, constructs multiple decision trees and utilizes a voting mechanism for final prediction, demonstrating strong resistance to overfitting. Its hyperparameters included 100 decision trees and a minimum of 5 samples required for internal node splitting. Bayesian optimization was employed to efficiently identify the optimal hyperparameter combination for the Multilayer Perceptron (MLP). The optimized architecture featured hidden layer sizes of (17, 43) and a learning rate of 0.00001.\u003c/p\u003e\u003cp\u003eA stacking ensemble strategy was implemented, characterized by the following structural design: optimized CatBoost, LightGBM, and Random Forest models functioned as base learners in the first layer to capitalize on the distinct advantages of each algorithm in capturing data patterns. A Linear Regression model served as the meta-learner in the second layer, assigned to learn the optimal combination of base learner predictions. To produce unbiased meta-features for training the meta-learner and rigorously prevent information leakage, the predictions from the base learners corresponding to the training samples were produced using a 5-fold cross-validation procedure.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eTable 3. Machine learning model hyperparameter settings\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"586\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 20.2297%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey hyperparameters\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e\u003cstrong\u003evalue\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOptimization methods\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003eCatBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003eiterations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e1000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003elearning_rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e0.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003edepth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003eLightGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003enum_leaves\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003eBayesian Optimization\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003ebagging_freq\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003efeature_fraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003en_estimators\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003emin_samples_split\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003erandom_state\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003eBayesian Optimization\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.4645%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 33.2446%;\"\u003e\n \u003cp\u003elearning_rate_init\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17.9142%;\"\u003e\n \u003cp\u003e0.00001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 29.2828%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"RESULTS AND DISCUSSION","content":"\u003cp\u003e\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\u003ePerformance evaluation indicators of each model on the test set\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\u003eMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRMSE (\u0026deg;C)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e​​MAE (\u0026deg;C)​\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eR\u0026sup2;\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCatBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.4231\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.1930\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9490\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9857\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLightGBM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2.9052\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.7045\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.3587\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9706\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\u003e1.8041\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.3432\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.0819\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9819\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMLP\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5.1418\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e2.2676\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.7866\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9483\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStacking\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.5768\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.7451\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.5184\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.9944\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\n\u003ch3\u003eOverall Model Performance and Comparative Analysis\u003c/h3\u003e\n\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe quantitative evaluation of predictive models on the test set is systematically summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The Stacking ensemble model demonstrated superior performance across all metrics, achieving the lowest MSE (0.5768), RMSE (0.7451\u0026deg;C), and MAE (0.5184\u0026deg;C), alongside the highest R\u0026sup2; value (0.9944). This signifies an nearly perfect fit to the true temperature distribution of the cathode steel bars. Among the base learners, CatBoost exhibited the strongest individual performance (R\u0026sup2; = 0.9857, RMSE\u0026thinsp;=\u0026thinsp;1.1930\u0026deg;C), followed by Random Forest (R\u0026sup2; = 0.9819, RMSE\u0026thinsp;=\u0026thinsp;1.3432\u0026deg;C), LightGBM, and MLP. The substantial performance gap between the Stacking model and the best single model (CatBoost) underscores the critical advantage of the ensemble strategy in this complex industrial prediction task.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eInterpretability and Feature Contribution of Base Learners\u003c/h2\u003e\u003cp\u003eThe exceptional capability of the Stacking framework stems from its effective integration of diverse base learners. The high inter-model correlations (0.998\u0026ndash;0.999) depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (the scatter matrix of base learner predictions) indicate that CatBoost, LightGBM, and Random Forest capture highly consistent overall temperature trends. This consistency provides a stable foundation for the meta-learner's integration. More importantly, SHAP analysis \u003cb\u003e(Figures S1, S2 and S3)\u003c/b\u003ereveals the distinct feature focus of each base learner, explaining their complementary nature. For instance, the CatBoost model (Figure S1) assigns paramount importance to rolling_mean_10and lagged features such as lag_9and lag_7, highlighting its strength in leveraging short-term temporal trends. In contrast, the LightGBM model (Figure S2) also prioritizes rolling_mean_10but shows a different weighting of subsequent lag features (e.g., lag_5, lag_9), indicating a marginally different perception of temporal dependencies. Similarly, the Random Forest model (Figure S3) demonstrates its unique feature interaction patterns. This diversity in feature importance, despite high prediction correlation, is a key source of the Stacking model's robustness, allowing it to mitigate individual model biases.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eIntegration Mechanism and Meta-Learner Analysis\u003c/h3\u003e\n\u003cp\u003eThe core of the Stacking approach lies in the optimal combination of these base learners by the meta-learner. The strong correlations shown in the meta-learner correlation matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) confirm the high alignment in prediction directions among base models. The meta-learner's feature importance analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), which displays the linear regression coefficients, provides direct insight into the integration strategy. The significantly higher weight assigned to the LightGBM predictions suggests that the meta-learner identifies its output as the most reliable anchor or baseline prediction among the ensemble. The weights for CatBoost and Random Forest are then adjusted to correct residual errors and capture nuances missed by LightGBM. This weighted combination, visualized conceptually in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, effectively reduces variance and enhances generalization, leading to the observed performance leap. The use of 5-fold cross-validation during the generation of meta-features ensured that this integration was learned without data leakage, guaranteeing the model's validity on unseen data.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eEngineering Implications for Intelligent Manufacturing\u003c/h3\u003e\n\u003cp\u003eThe practical value of this Stacking-based monitoring system is profound for intelligent manufacturing in the aluminum industry. The achieved MAE of 0.52\u0026deg;C and RMSE of 0.75\u0026deg;C (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, Figure S1) represent a critical threshold for reliable early-warning systems. In practical terms, this high precision allows for the detection of subtle abnormal thermal rises long before they escalate into potential hazards like lining failure, thereby providing a crucial window for preventive maintenance. The model's interpretability, afforded by the SHAP analysis, is equally valuable. It transforms the model from a \"black box\" into a decision-support tool. Plant operators can not only receive alerts but also understand the contributing factors (e.g., a sharp change captured by rolling_std_10or a persistent deviation in rolling_mean_10), facilitating root cause analysis and informed operational decisions. This aligns perfectly with the core objectives of Industry 4.0\u0026mdash;moving from reactive to predictive and ultimately prescriptive maintenance.\u003c/p\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eThis study developed a distributed fiber-optic temperature measurement system for monitoring cathode steel bars in aluminum electrolytic cells, operating successfully under extreme conditions like intense EMI, high temperatures, and dust. It acquired high-frequency temperature time-series data from actual production. Multiple ML algorithms\u0026mdash;CatBoost, LightGBM, RF, and Bayesian-optimized MLP\u0026mdash;were applied and evaluated. Integrating fiber-optic sensing with stacking ensemble ML established a real-time monitoring system. Validation shows it proactively identifies abnormal temperature increases, reducing response time versus traditional methods, improving safety and energy control. This supports the industry's intelligent transformation. Future research will focus on transfer learning, multi-parameter fusion, and dynamic online learning. The system delivers a high-precision solution and a generalizable framework for other extreme industrial environments.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFinancial support and sponsorship\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis paper was supported by Smart Aluminum Plant Technological Innovation Project of Inner Mongolia Huomei-Hongjun Aluminum Electric Co., Ltd (K20231334)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eFu R (2023) Optimization of Electrolytic Aluminum Production Process by Intelligent Control. In \u003cem\u003e2023 International Seminar on Computer Science and Engineering Technology (SCSET)\u003c/em\u003e; IEEE, ; pp 695\u0026ndash;698. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/scset58950.2023.00157\u003c/span\u003e\u003cspan address=\"10.1109/scset58950.2023.00157\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 智慧铝\u0026amp;#21378\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYao L, Long W, Li T, Tang D, Yi J, An Improved UKFNN (2019) Based on Iterative Renewal for Intelligent Modeling of Aluminum Electrolysis. In \u003cem\u003e2019 IEEE 4th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC)\u003c/em\u003e; IEEE, ; pp 271\u0026ndash;275. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/iaeac47372.2019.8998081\u003c/span\u003e\u003cspan address=\"10.1109/iaeac47372.2019.8998081\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 智慧铝\u0026amp;#21378\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun Y, Gui W, Chen X, Xie Y, Xie S, Zou ZA (2023) Dynamic Spatial Distributed Information Clustering Method for Aluminum Electrolysis Cell. Eng Appl Artif Intell 126:106793. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.engappai.2023.106793\u003c/span\u003e\u003cspan address=\"10.1016/j.engappai.2023.106793\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e智慧铝厂\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang J, Xie Y, Xie S, Chen X (2024) Eng Appl Artif Intell 134:108664. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.engappai.2024.108664\u003c/span\u003e\u003cspan address=\"10.1016/j.engappai.2024.108664\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e智慧铝厂 Development of Data-Knowledge-Driven Predictive Model and Multi-Objective Optimization for Intelligent Optimal Control of Aluminum Electrolysis Process\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHancock J, Khoshgoftaar TM (2020) CatBoost for Big Data: An Interdisciplinary Review. Research Square. Springer Science and Business Media LLC August 19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21203/rs.3.rs-54646/v1\u003c/span\u003e\u003cspan address=\"10.21203/rs.3.rs-54646/v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDu W, Hu Z, Li C, Chen Z, Chen W, Qiu S (2023) Distributed Optical Fiber Sensing Technology in Operation Status Monitoring of GIL Equipment. Appl Math Nonlinear Sci 9(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2478/amns.2023.2.00274\u003c/span\u003e\u003cspan address=\"10.2478/amns.2023.2.00274\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOKAMOTO K (1994) Special Issue on Fiber-Optics. Fiber-Optic Distributed-Temperature Sensing. \u003cem\u003erle 22\u003c/em\u003e (4), 276\u0026ndash;283. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2184/lsj.22.276\u003c/span\u003e\u003cspan address=\"10.2184/lsj.22.276\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Y, Lei T, Sun Z, Wang C, Liu T (2012) Application of Distributed Optical Fiber Temperature System in Online Monitoring and Fault Diagnosis of Smart Grid. In \u003cem\u003e2012 Asia-Pacific Power and Energy Engineering Conference\u003c/em\u003e; IEEE, ; pp 1\u0026ndash;4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/appeec.2012.6307675\u003c/span\u003e\u003cspan address=\"10.1109/appeec.2012.6307675\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 分布式光\u0026amp;#32420\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDu W, Hu Z, Li C, Chen Z, Chen W, Qiu S (2023) Distributed Optical Fiber Sensing Technology in Operation Status Monitoring of GIL Equipment. Appl Math Nonlinear Sci 9(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2478/amns.2023.2.00274\u003c/span\u003e\u003cspan address=\"10.2478/amns.2023.2.00274\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e分布式光纤\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa S, Xu Y, Pang Y, Zhao X, Li Y, Qin Z, Liu Z, Lu P, Bao X (2022) Optical Fiber Sensors for High-Temperature Monitoring: A Review. \u003cem\u003eSensors 22\u003c/em\u003e (15), 5722. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/s22155722\u003c/span\u003e\u003cspan address=\"10.3390/s22155722\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 分布式光\u0026amp;#32420\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeng Y, Jiang J (2022) Optical Fiber Sensors in Extreme Temperature and Radiation Environments: A Review. IEEE Sens J 22(14):13811\u0026ndash;13834. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/jsen.2022.3181949\u003c/span\u003e\u003cspan address=\"10.1109/jsen.2022.3181949\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e分布式光纤\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLian Z, Zhang C, Liu C, Zhao H, Liu F, He A, Bo F, Qi X, Jia X, Liu P, Chen C (2025) Intelligent Energy Optimization for Electrolytic Aluminum Using Industrial Data-Driven and Knowledge-Guided Modeling. Metall Mater Trans B 56(5):5187\u0026ndash;5199. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11663-025-03691-9\u003c/span\u003e\u003cspan address=\"10.1007/s11663-025-03691-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e智慧铝厂\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKe G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu T-Y (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree, LightGBM\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRigatti SJ, Random Forest (2017) J Insur Med 47(1):31\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.17849/insm-47-01-31-39.1\u003c/span\u003e\u003cspan address=\"10.17849/insm-47-01-31-39.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePinkus A (1999) Approximation Theory of the MLP Model in Neural Networks. Acta Numerica 8:143\u0026ndash;195. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1017/s0962492900002919\u003c/span\u003e\u003cspan address=\"10.1017/s0962492900002919\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Materials Informatics, material characterization, advanced materials, composites, Materials Science","lastPublishedDoi":"10.21203/rs.3.rs-7993883/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7993883/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAbnormal temperature rise in the cathode steel bars of electrolytic aluminum cells, a core smelting equipment, is a major cause of furnace leakage accidents. However, traditional thermocouple and infrared temperature measurement technologies are susceptible to electromagnetic interference under extreme operating conditions such as strong magnetic fields, high temperatures, and severe corrosion, making continuous and accurate monitoring difficult. To address this technical bottleneck, this study developed an intelligent monitoring system that integrates distributed fiber optic sensing and stacking ensemble learning. By deploying electromagnetic interference-resistant fiber optic sensors on the surface of the cathode steel bars, continuous temperature data acquisition was achieved. A multidimensional feature set was constructed, including time series lag, rolling statistics, and periodic features, and the prediction performance of models such as CatBoost, LightGBM, and random forest was systematically compared. Ultimately, a stacking ensemble strategy was employed to combine the strengths of each base model to achieve accurate cathode steel bar temperature prediction (RMSE\u0026thinsp;\u0026lt;\u0026thinsp;2.1\u0026deg;C, R\u0026sup2; \u0026gt;0.99 on the test set). This system can proactively identify abnormal temperature rises, reducing warning time by over 85% compared to manual inspections, providing a reliable technical path for intelligent safety monitoring in electrolytic aluminum production.\u003c/p\u003e","manuscriptTitle":"Towards Smart Aluminum Smelting: An AI-Driven Approach for Real Time Thermal Anomaly Detection Using Distributed Optical Fiber Sensors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-04 10:12:58","doi":"10.21203/rs.3.rs-7993883/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fbcf3289-23ee-4804-92b2-acf544dd1cd3","owner":[],"postedDate":"November 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-11-04T10:12:58+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-04 10:12:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7993883","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7993883","identity":"rs-7993883","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}