Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes

Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes | 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 Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes Fengtao Wang, Quanjie Li, Xinxin Zeng, Weiwei Liu, Zhiqi Luo, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7002574/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 Laser Directed Energy Deposition (L-DED) faces key challenges such as unstable printing processes, low forming quality, and poor interlayer consistency. An urgent need is to establish real-time monitoring and process control mechanisms to improve its industrial applicability. However, in practical applications, the construction of monitoring systems is constrained by two major bottlenecks. The scarcity of high-quality defect samples limits the generalization capability of deep learning models. Additionally, due to confidentiality concerns regarding printing paths and product defects, users generally refuse to share raw data, further hindering centralized model training and large-scale deployment. To address these challenges, this study proposes a lightweight neural network integrated into a federated transfer learning framework, termed SK-MNV4. The network incorporates a unified inverted bottleneck module, a lightweight multi-head attention mechanism, and a selective kernel fusion strategy to enable efficient multi-scale feature extraction and robust defect representation, significantly enhancing performance under data privacy and limited-sample conditions. The proposed federated transfer learning scheme treats independently collected melt pool image datasets from three representative alloys (IN738, TB6, Ti-10Mo) in L-DED experiments as heterogeneous clients for federated learning. It ensures local data processing and prevents raw melt pool image exposure, while enabling cross-material transfer learning with IN718 melt pool images as the target domain, thereby improving the model's generalization ability. We used T-SNE to compare and visualize each model. We compare our model with mainstream models (ConvNeXt, DenseNet, EfficientNet, and Swin Transformer) in terms of accuracy and inference efficiency. Comparative experiments show that SK-MNV4 outperforms these models and shows 97.3% accuracy and 0.001 seconds inference efficiency on average, demonstrating superior potential for applications in scenarios with strong real-time requirements, data sensitivity, and sparse defect distributions. Directed Energy Deposition Federated Transfer Learning Melt Pool Monitoring Data Privacy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 1. Introduction Additive manufacturing (AM), also known as 3D printing, is a manufacturing process that creates objects layer by layer by adding material [ 1 , 2 ]. Laser direct energy deposition is one of the most dominant AM forms in current manufacturing practices [ 3 ]. It represents a novel approach to manufacturing high-performance components for aerospace, medical, energy, and automotive components [ 4 ], as well as repairing existing metal parts [ 5 , 6 ]. Due to the inherent characteristics of L-DED, such as high thermal gradients and complex melt pool dynamics, the process often exhibits significant variability, making it difficult to control the quality of the final parts. Despite the significant scientific and engineering efforts in industry and academia devoted to L-DED, enhancing the production efficiency and fabrication quality remains a significant challenge [ 7 ]. The production of high-quality printed parts is a significant concern, as any defects could damage the part's structural integrity [ 8 ]. Therefore, process monitoring is an essential way to improve L-DED component quality and decrease variability in mechanical performance. Recent studies on L-DED process monitoring have primarily focused on data-driven defect detection and quality prediction, leveraging sensing techniques, such as thermal imaging and melt pool visualization, combined with deep learning or unsupervised algorithms. Baumgartl et al. [ 9 ] employed thermographic off-axis imaging and deep learning neural network architectures to detect delamination and splatter defects. Chen et al. [ 10 ] introduced an in-situ acoustic-based defect detection strategy for L-DED, leveraging convolutional neural networks and acoustic signal analysis to achieve high accuracy in identifying cracks and keyhole pores. Ren et al. [ 11 ] developed an unsupervised recognition model using LSTM-Autoencoder and K-means clustering for in-situ quality monitoring of DED-printed parts, demonstrating accurate differentiation between qualified and unqualified depositions. However, all of the aforementioned methods are constrained by data scarcity. Training high-performance deep learning models typically relies on large-scale, diverse, and high-quality datasets. This challenge is particularly pronounced in tasks involving micro-defect identification, where it is especially difficult to acquire multimodal sensing data, such as melt pool images, thermal profiles, and acoustic signals [ 12 , 13 ] that adequately represent typical processing conditions. These limitations collectively lead to insufficient data dimensionality and poor representativeness, making it difficult for the resulting models to achieve the desired levels of detection accuracy and generalization capability [ 14 ]. Transfer learning addresses data scarcity by transferring pre-trained model parameters from a source domain to a similar target task [ 18 ]. For example, Peng et al. [ 16 ] developed a transfer learning-based PINN model to predict 3D temperature fields in laser-based directed energy deposition, achieving an RMSE of 6.365 K and an average error of 1.28%. However, most existing solutions remain limited in addressing the dual challenges of data scarcity and privacy constraints, especially when dealing with heterogeneous materials and rare defect types. These gaps underscore the urgent need for a generalizable and privacy-preserving framework that can maintain high defect detection accuracy even under limited and decentralized datasets. In the process of L-DED, data such as melt pool images and process parameters often involve key information about the manufacturing workflow. Due to the high emphasis on data privacy, manufacturers are generally unwilling to share such sensitive data. Meanwhile, intense market competition among companies further aggravates data closure. This overprotection of data privacy limits cross-organizational data sharing and the development of collaborative monitoring mechanisms, and makes it difficult for traditional transfer learning approaches to acquire broad prior knowledge. To address this challenge, researchers introduce an federated learning (FL) framework that enables privacy-preserving collaborative training by retaining local melt pool image data and exchanging only model parameters. This decentralized approach not only alleviates data-sharing concerns among L-DED manufacturers but also enhances the model’s cross-material generalization capability under heterogeneous and data-scarce conditions [ 17 , 18 ]. FL is a machine learning approach where multiple devices collaboratively train a model without sharing raw data. Instead, only model updates are exchanged, preserving user privacy. It improves data security, reduces communication costs, and enables learning from decentralized data sources, making it ideal for privacy-sensitive applications [ 19 ]. For instance, Mehta et al. [ 20 ] proposed a FL-based semantic segmentation method for defect detection in additive manufacturing, addressing data availability and privacy issues while achieving performance similar to centralized learning. Li et al. [ 21 ] proposed a federated transfer learning fault diagnosis model that preserves data privacy and enables cross-domain diagnosis by aligning feature distributions. These studies demonstrate that FL can effectively balance privacy preservation and model performance in additive manufacturing. Moreover, by enabling cross-domain generalization and outperforming centralized methods, FL proves to be a promising approach for intelligent defect detection and quality control in heterogeneous and data-sensitive environments. Through our literature review, we found that few studies have explored the integration of FL techniques in L-DED applications. To address the aforementioned challenges, this study proposes an SK-MNV4 lightweight network within a FL framework, integrating FL and transfer learning to address the problems of data privacy protection and sample scarcity in the L-DED process. Specifically, this network adopts a unified inverted bottleneck structure (UIB), a Mobile-MQA attention mechanism, and a selective convolution kernel module (SK) to achieve efficient multi-scale feature extraction and channel fusion. The present study performs L-DED experiments with multiple metal alloys. We use three alloys (i.e., IN738, TB6, and Ti-10Mo) as federated clients and execute distributed pre-training to ensure that the molten pool image data is processed locally, ensuring privacy security. The Federated Averaging (FedAvg) algorithm aggregates local model updates to build a global model with cross-material generalization capabilities. Subsequently, the model is migrated to the IN718 material with only a small number of labeled samples, followed by fine-tuning the classification layer. Combined with visualization analysis such as Grad-CAM and t-SNE, we further verify the robustness of the model in discriminating area positioning and category differentiation. This method takes into account high precision, low latency and good scalability, and provides a feasible path for intelligent defect warning in the L-DED process. The key contributions of this study are as follows: (1) A novel FL framework is introduced for defect detection in L-DED, enabling privacy-preserving model training across multiple material domains. (2) The lightweight SK-MNV4 network is proposed and validated, showing strong performance in both accuracy and efficiency. (3) A federated transfer learning strategy is developed to extend model generalization to new materials using minimal target data. The structure of this paper is as follows: Section 2 provides an overview of the fundamental theoretical concepts of FL; Section 3 describes the proposed SK-MNV4 network; Section 4 explains the experimental design approach; Section 5 presents a comparative analysis of the proposed network against other models, followed by conclusions in Section 6 . 2. Related works Due to the complexity and instability of the deposition process, defects such as porosity, cracking, and incomplete fusion frequently occur, necessitating robust real-time monitoring and intelligent defect detection systems. Recent advances in machine learning, especially deep learning, have driven the development of intelligent monitoring solutions. Techniques like FL and transfer learning have been introduced to further enhance the adaptability, scalability, and efficiency of these systems under real-world constraints such as data privacy, small sample size, and platform heterogeneity [ 22 ]. 2.1 Challenges and advances in monitoring L-DED processes L-DED is a complex thermal-mechanical process characterized by rapid heating and cooling cycles, intricate melt pool dynamics, and sensitivity to numerous parameters [ 3 ]. These factors contribute to a high likelihood of defect formation, thereby driving the need for real-time, intelligent monitoring technologies [ 5 ]. In recent years, sensor technologies—ranging from visible-light imaging and infrared thermography to acoustic emission—have been increasingly adopted to capture process signals [ 23 , 24 ]. Combined with deep learning-based approaches, such as convolutional neural networks (CNNs), these systems have shown promise in classifying and localizing defects based on real-time image and signal data [ 9 ]. Nevertheless, several fundamental bottlenecks limit their practical deployment. Foremost among these is the issue of data scarcity: high-quality, annotated datasets are often difficult to obtain due to the costly and time-consuming nature of metallographic inspection and defect labeling. These limitations are further exacerbated when targeting rare or subtle defect types, such as micro-porosity, incomplete fusion, or keyhole-induced vaporization, where ground truth labeling is difficult to standardize and highly sensitive to measurement noise and operator subjectivity. Consequently, existing datasets often suffer from insufficient size and representativeness, ultimately restricting the learning capacity of deep models and impairing their ability to generalize across unseen process conditions [ 11 ]. Additionally, generalization across platforms remains a significant issue. Data heterogeneity induced by differences in equipment, materials, and process settings results in domain shifts that compromise the robustness and transferability of models trained on a single source domain [ 25 ]. Moreover, concerns over data privacy and proprietary information limit data sharing across institutions, leading to fragmented datasets and inconsistent model performance [ 26 ]. As a result, collaborative dataset construction is limited, and models suffer from poor generalizability and overfitting. Finally, existing monitoring algorithms are often computationally intensive, challenging their deployment in real-time or edge-computing environments. Addressing these constraints calls for more flexible, privacy-preserving, and data-efficient modeling strategies [ 18 ]. 2.2 Federated learning for privacy-preserving collaborative modeling FL has emerged as a promising paradigm to address data privacy and collaboration barriers in intelligent manufacturing. It enables decentralized model training across multiple clients or devices by exchanging model parameters instead of raw data, thus preserving confidentiality while building a robust global model [ 17 ]. In the context of additive manufacturing, FL facilitates collaborative learning among different manufacturing units or institutions, enabling them to share model intelligence without compromising proprietary data [ 30 ]. This capability is critical in industries where intellectual property and trade secrets are paramount, making FL a cornerstone for secure, collaborative innovation. The integration of FL with techniques like transfer learning has further amplified its potential. These hybrid approaches enhance defect detection across heterogeneous platforms, enabling models to adapt to diverse machines, materials, and operational conditions. For instance, FL has allowed manufacturers using different machines and materials to collaboratively train models that adapt to various working conditions [ 21 ]. For instance, lightweight neural networks and communication-efficient strategies have been proposed to make FL viable on resource-constrained edge devices, broadening its applicability in real-time industrial settings [ 21 ]. Nonetheless, the practical implementation of FL in industrial settings still encounters notable challenges. One major issue is communication overhead, as federated training requires frequent parameter exchange, which may be infeasible under limited network infrastructure [ 27 ]. Additionally, the non-independent and identically distributed nature of manufacturing data across clients significantly affects convergence speed and model accuracy [ 19 ]. FL systems are also vulnerable to system failures, straggling devices, and even malicious participants, raising concerns about robustness and trustworthiness. Most critically, there exists a fundamental trade-off between model performance and privacy preservation, which remains an open problem for real-world deployment [ 26 ]. Addressing these challenges requires innovative solutions, such as adaptive aggregation algorithms to handle non-IID data and robust encryption protocols to mitigate security risks. Furthermore, developing scalable FL frameworks that balance computational efficiency with privacy will be essential for widespread adoption in dynamic manufacturing ecosystems, ultimately unlocking the full potential of AI-driven process optimization. 2.3 Transfer learning strategies for data-efficient defect detection Transfer Learning (TL) offers a compelling solution to the limitations of data scarcity and generalization in deep learning [ 14 ]. By leveraging knowledge from source domains or pre-trained models, TL enables efficient learning in target domains with limited data. In additive manufacturing, TL techniques have been adopted to improve model performance in tasks such as defect classification and process anomaly detection [ 28 ]. The most common strategy involves fine-tuning pre-trained models for manufacturing-specific datasets. These models capture low-level features that are still relevant in industrial imagery and have demonstrated decent performance with small datasets [ 29 ]. In addition to model-based transfer, feature-based domain adaptation techniques such as maximum mean discrepancy and adversarial learning have been used to bridge the distribution gap between source and target domains [ 12 ]. Some studies have also explored the integration of TL with FL, creating collaborative frameworks that can generalize across different factories, machines, or even types of defects. However, the use of TL in additive manufacturing is still hampered by several obstacles. One critical challenge is domain mismatch, as pre-trained models are often based on natural images, while manufacturing data has distinct visual and statistical properties [ 14 ]. This can lead to negative transfer, where the imported knowledge degrades rather than improves performance. Another issue is the semantic gap between tasks, especially when the source task differs significantly from the target task [ 28 ]. Additionally, transferability is hard to quantify or predict, and there is no unified framework to determine when and how transfer learning will be effective [ 30 ]. Furthermore, most TL approaches are static and single-source, while real-world manufacturing involves dynamic, multi-source, and continuously evolving scenarios [ 22 ]. There is an urgent need for multi-source, adaptive, and lifelong transfer learning frameworks that can adapt to new tasks and domains over time. Lastly, maintaining model interpretability, robustness, and safety during the transfer process remains a critical yet unresolved issue, particularly for high-stakes industrial applications [ 14 , 28 ]. 2.4 Motivation and overview of the proposed framework To address these multifaceted challenges holistically, this study proposes a novel federated transfer learning framework, which tightly integrates a lightweight yet powerful backbone network, termed SK-MNV4, with privacy-preserving FL and small-sample transfer learning strategies. The core idea of the proposed framework is to collaboratively leverage decentralized, heterogeneous melt pool datasets without compromising data ownership or security, while simultaneously improving the model's adaptability to unseen material domains. Specifically, the proposed approach operates in two key stages. In the first stage, FL is employed to pre-train a global model across multiple decentralized datasets, each corresponding to a distinct alloy system (e.g., IN738, TB6, Ti-10Mo). Each dataset acts as a federated client, conducting local model training without uploading raw melt pool images to the central server. Only model updates, such as weights or gradients, are securely transmitted for aggregation using FedAvg algorithm. This design ensures that sensitive manufacturing data remains confined within each client, thereby addressing privacy and confidentiality concerns. Furthermore, by learning from diverse material domains in a decentralized manner, the global model gains robust generalization capabilities, mitigating the effects of data heterogeneity across clients. At the heart of the FL framework is the SK-MNV4 network, an enhanced version of MobileNetV4 tailored for defect detection tasks in L-DED environments. SK-MNV4 introduces several architectural innovations, including UIB blocks, lightweight multi-head attention mechanisms, and selective kernel (SK) feature fusion modules. These modifications significantly enhance the model’s ability to capture multi-scale melt pool features while maintaining a compact parameter footprint and fast inference speed, making it highly suitable for real-time deployment on resource-constrained industrial platforms. In the second stage, to further adapt the pre-trained global model to a new target domain characterized by a different material system (e.g., IN718) with small sample availability, a transfer learning procedure is applied. The pre-trained SK-MNV4 model is fine-tuned using a limited number of labeled target domain samples, focusing primarily on updating the classification layers while retaining the generalized feature extraction capabilities learned during federated training. This two-stage approach not only minimizes the dependence on large labeled datasets but also significantly enhances model adaptability and reduces the risk of negative transfer commonly encountered when source and target domains differ. Through the combination of FL and transfer learning, the proposed framework is designed to achieve three pivotal objectives: (1) Enhance defect detection performance under data scarcity conditions by leveraging prior knowledge learned from related material domains; (2) Preserve data privacy and confidentiality during collaborative model development, ensuring compliance with industrial data protection requirements; and (3) Improve model adaptability and robustness across heterogeneous material domains, enabling reliable defect detection in real-world, dynamic manufacturing environments. 3. Experiment 3.1 Materials In this L-DED experiment, three materials—IN738, TB6, and Ti-10Mo—were utilized as the training dataset for FL, while IN718 served as the validation dataset. The experimental metal powder employed had a diameter range of 53–160 µm. As part of the preparation, the metal powder was pre-dried at 120°C for 4 hours in a vacuum drying oven. The nickel-based substrate measured 200 mm × 200 mm × 5 mm. To prevent oxidation of the substrate surface, which could adversely affect the final results, a polishing treatment was conducted on the substrate. 3.2 Experiment device This experiment was conducted on an L-DED experimental equipment. Figure 1 (a) illustrates the basic operational schematic of the L-DED system. The laser emitter GmbH LDF 4000-100VGP can emit laser power up to 4400W, and the laser is released through the laser head (Gaggenau YC-52). To prevent interference from metal powder oxidation on the experimental results, we adopted 99.99% pure argon gas as shielding gas. The water chiller was employed for temperature stabilization of the laser head. Beneath the laser head was a static workbench for fixing the substrate. The laser head outfitted with a CCD camera can capture the melt pool image, as illustrated in Fig. 1 (c) . The CCD camera was installed coaxially on the laser head with a frequency of 50Hz and a pixel resolution of 500*500 for image acquisition. a local computer connected the CCD camera via a serial port to save the captured images locally. 3.3 Experiment design Due to the significant impact of four process variables, namely, laser power, beam diameter, scan speed, and powder supply rate, on the quality of the formed parts [ 31 ], we designed and conducted experiments with four different materials under varying conditions, as detailed in Table 1 . To label the melt pool images of each single-track deposit, we performed post-processing on the deposits. It should be noted that we adopted a line-cutting approach to select a stable portion between 5-10mm (Fig. 2 ), given that the laser emission was initially unstable. Table 1 Experimental parameters under different deposition materials Parameters IN738 TB6 Ti-10Mo IN718 Laser Power(W) 800 ~ 2600 500 ~ 2000 500 ~ 1400 1000 ~ 3000 Beam Diameter(mm) 2, 3 2 2 1, 2, 3 Scan Speed(mm/min) 200 ~ 2000 300 ~ 1200 420 ~ 1140 800 ~ 1900 Powder Supply Rate(g/min) 3.6 ~ 30 5, 7, 9 6, 8, 10 9.3, 18.3, 29.4 This section of deposits was polished using 240, 600, 1200, and 2000 grit sandpapers and a polishing cloth. Subsequently, we captured magnified images of the cross-sections using an optical mirror and measured the contact angle (α), track width (W), melt pool height (H), and depth of penetration(D) of the melt pool images using ImageJ software. Figure 3 (a) displays the cross-section of an experimental single-track deposit, while Fig. 3 (b) illustrates a schematic model showcasing the necessary measurements for the cross-section of the single-track deposit. Previous studies have shown that different shapes of the melt pool imply different quality [ 32 ]. The desired shape for a melt pool, which is free from any defects or porosities, is considered to be a semi-circular shape in its conduction mode[ 33 ]. When the depth of the melt pool exceeds half of its width, it can be considered as being in keyhole mode, which is often caused by excessive energy density [ 33 – 36 ]. It is worth noting that keyhole mode does not necessarily result in the formation of porosity [ 33 , 35 ]. Additionally, it is important to control the contact angle of the melt pool, as excessive contact angles can lead to porosity during interlayer deposition. In situations where the melt pool inadequately penetrates the underlying material, it can be identified as a "lack of fusion" occurrence, frequently leading to incomplete fusion between adjacent layers [ 36 – 38 ]. Therefore, to make the melt pool classification mechanism more reasonable, this paper integrates two indicators, namely dilution rate and contact angle. Dilution rate is defined as D/(H + D) for ease of calculation, and the contact angle is the average of the left and right contact angles. Based on the research mentioned above, the melt pool quality is classified as: (1) Normal Melt (NM): 10%< dilution rate < 30% and θ 30% or θ > 80°; and (3) Lack of Melt (LM): dilution rate < 10%. The grouping of the four different materials for single-track deposits is illustrated in Fig. 4 . After categorizing each material, we randomly selected 2000 melt pool images from each category of IN738, TB6, and Ti-10Mo as the training set. Each material's data serves as a FL client. Additionally, we extracted 500 images from each category of IN718, with 300 images designated for the FL validation set and 200 images allocated for a small sample dataset in federated transfer learning. 4. Proposed method 4.1 Overview of the Federated Transfer Learning Framework The framework illustrated in Fig. 5 shows the L-DED using the SK-MNV4 model. It begins with three distinct datasets—IN738, TB6, and Ti-10Mo—each representing a different alloy, acting as federated clients for local training. These models are trained independently on their respective datasets, ensuring data privacy. The SK-MNV4 model integrates various modules, including the UIB, Mobile MQA, and SK Block, to enhance feature extraction and fusion capabilities. The local model updates from each client are aggregated using the FedAvg algorithm to form a global model. This global model is then fine-tuned with a small labeled dataset (IN718) before being deployed for real-time melt pool defect monitoring, including the detection of defects such as spattering. This approach efficiently addresses the challenges of data privacy and small sample sizes while enabling high-accuracy monitoring in diverse material domains. 4.2 Fundamentals of Federated Learning and FedAvg Algorithm Mathematically, the goal of FL is to solve the global optimization problem by aggregating local model updates from a set of clients. Formally, let the global objective function \(\:f\left(w\right)\) represent the loss function of a shared model, where \(\:w\) are the global model parameters. The objective is to minimize the weighted sum of local loss functions: $${w_{t+1}}=\sum\limits_{{t=1}}^{N} {\frac{{{n_i}}}{{\sum {_{{j=1}}^{N}{n_j}} }}} {w_i}(t)$$ 1 where \(\:{p}_{i}\) is the weight assigned to client \(\:i\) (often based on the number of data samples it holds), and \(\:{L}_{i}\left(w\right)\) represents the local loss function of client \(\:i\) . This approach allows the server to aggregate model updates from clients without ever directly accessing the clients’ data. The process of FL is illustrated in Fig. 6 . The server distributes the global model to individual users, who then utilize their local data to train local models and subsequently upload the model parameters to the server. The server employs an aggregation algorithm to consolidate the local model parameters, thereby updating the global model, which is then retransmitted to the users. This iterative process continues until the maximum number of federated iterations is reached, resulting in the final federated global model. The primary algorithm used in FL is FedAvg, a simple yet effective approach for aggregating local model updates. Each client \(\:i\) computes alocal model update \(\:\varDelta\:{w}_{i}\) using its local data and sends this update to the central server. The server then aggregates the local updates to form a new global model. The aggregation is typically a weighted average: $$\hbox{min} (w)\sum\limits_{{i=1}}^{N} {{p_i}{L_i}(w)}$$ 2 where \(\:{n}_{i}\) is the number of data points on client \(\:i\) , and \(\:{w}_{i}\left(t\right)\) is the model update sent by client \(\:i\) at iteration \(\:t\) , This aggregation step ensures that the global model is updated by considering the local learning from all participating clients 4.3 The proposed SK-MNV4 The proposed SK-MNV4 is an improvement upon MobileNetV4 [ 41 ], with its overall structure illustrated in Fig. 7 (a) . It primarily consists of UIB, Moblie_MQA, and SK blocks [ 39 ]. The UIB, depicted in Fig. 7 (b) , comprises four main components: Extra DW, Inverted Bottleneck (IB), ConvNext, and feed-forward network (FFN). The UIB not only serves as the fundamental structural unit of the SK-MNV4 backbone but also embodies a highly optimized architectural convergence of depthwise separability, residual learning, and adaptive feature transformation. By design, it addresses the limitations inherent in traditional convolutional backbones when applied to real-time industrial tasks, such as melt pool classification in L-DED, where lightweight computation and fine-grained spatial awareness are simultaneously required. The IB, at the heart of UIB, leverages an initial 1×1 pointwise convolution to project low-dimensional input tensors into a higher-dimensional latent space, enabling rich channel-wise representation learning. This operation is followed by a 3×3 depthwise convolution, which performs spatial filtering independently on each channel, drastically reducing the number of required parameters compared to traditional convolutions while retaining essential local texture and shape information—a crucial property for distinguishing fine-scale melt pool geometries. The IB concludes with a second 1×1 pointwise convolution that compresses the expanded feature maps back to a compact representation, often with a residual shortcut if the input and output dimensions match. This residual connection is critical for stabilizing gradient propagation in deep networks and supports efficient training convergence. In contrast to IB, Extra DW convolution introduces a pre-expansion 1×1 convolution prior to the standard IB expansion step. This module can be interpreted as a channel reweighting mechanism that primes the feature space before expansion. By applying early-stage channel mixing, the Extra DW path promotes feature diversity and enables richer interaction between channels prior to spatial filtering. Further extending this foundation, the ConvNext block replaces the conventional depthwise convolution between the expansion and projection layers with a streamlined residual feed-forward structure inspired by next-generation convolutional architectures. This architectural refinement preserves the expressiveness of channel-wise transformations while reducing sensitivity to kernel size selection, thereby enhancing the robustness and computational efficiency of feature learning within the SK-MNV4 backbone. The final subcomponent, the FFN, comprises two stacked 1×1 convolutions with nonlinear activations in between. This sequence emulates the transformation logic found in Transformer-style architectures, where linear projections sandwiching a nonlinear function are used to refine channel-specific semantics. In the context of melt pool monitoring, FFN contributes to the discrimination of subtle thermal and morphological features. In the broader context of the SK-MNV4 architecture, UIB blocks serve as the primary conduit for hierarchical representation learning, enabling the model to transition smoothly from shallow, detail-rich features to abstract, class-specific embeddings. Their synergy with attention-enhancing modules (Mobile-MQA) and scale-aware fusion mechanisms (SK blocks) empowers the network to retain critical melt pool characteristics while discarding redundant or noisy background information. The result is a compact yet powerful encoder capable of supporting real-time, high-accuracy defect classification in federated and data-scarce settings. Mobile-MQA is introduced as an efficient attention mechanism tailored for real-time industrial applications such as melt pool image classification in L-DED. Traditional multi-head attention (MHA) mechanisms, while powerful in modeling long-range dependencies, incur significant computational overhead due to their reliance on high-resolution dot-product operations across all query-key-value interactions. To mitigate this, Mobile-MQA adopts a spatial reduction (SR) strategy for the key and value matrices, effectively reducing their spatial resolution prior to attention computation. This results in a substantial reduction in memory usage and inference latency, making the mechanism well-suited for deployment on edge devices. The formula for Mobile MQA [ 40 ] is presented as follows: $${\text{Mobil}}{{\text{e}}_M}_{{QA(X)}}={\text{Concat}}({\text{attentio}}{{\text{n}}_1},...,{\text{attentio}}{{\text{n}}_n}){W^o}$$ 3 The computation of each attention head, \(attentio{n_j}\) , is further defined as: $${\text{attentio}}{{\text{n}}_j}={\text{Softmax}}(\frac{{(X{W^{Qj}}){{(SR(X){W^K})}^T}}}{{\sqrt {{d_k}} }})(SR(X){W^V})$$ 4 Let \(\:X\) be the input feature. The \(\:n\) attention heads compute different attention weights, \(\:{attention}_{j}\) . These attention outputs are then concatenated and mapped back to the input space dimensions through the weight matrix \(\:{W}^{o}\) .In the case of single-head attention, the operation \(\:SR\left(X\right)\) is applied to reduce computational complexity, where \(\:SR\left(X\right)\) represents the downsampling of keys and values. The query, key, and value are linearly projected using the projection matrices \(\:{W}^{{Q}_{j}}\) , \(\:{W}^{K}\) , and \(\:{W}^{V}\) , respectively. A scaling factor, \(\:\sqrt{{d}_{k}}\) , is applied, where \(\:{d}_{k}\) is the dimension of the key vector, to stabilize the numerical range of the Softmax function. By down sampling the keys and values spatially, the computational cost is reduced, while the query retains high resolution, preserving more detailed information. This design effectively strikes a balance between performance and computational efficiency. The SK block (as shown in Fig. 7 (c) ) is a feature extraction and fusion methodology aimed at enhancing the representational capacity of feature maps. Initially, two parallel convolutional branches with different kernel sizes (3×3 and 5×5) operate on the input feature map to extract features from varying receptive fields. The feature maps generated by these branches are then summed along the channel dimension, facilitating the integration of diverse information. Following this, an average pooling operation is applied to statistically capture spatial information across the channels, yielding a C×1×1 feature map that represents channel significance. This feature map is subsequently mapped to a lower dimension through a fully connected layer and then restored to the original channel count to obtain channel-specific weights. These weights are multiplied element-wise with their corresponding feature maps, leading to a weighted fusion that culminates in an output feature map with enhanced expressiveness. The entire process effectively leverages multi-branch convolutions, attention mechanisms, and weighted feature maps, resulting in substantial improvements in information richness and classification performance of the final output. 5. Results and discussions In this study, we compare the proposed SK-MNV4 model with several state-of-the-art network architectures, including EfficientNet [ 41 ], ConvNeXt [ 42 ], MobileVIT [ 43 ], DenseNet [ 44 ], and Swin Transformer (SwinT) [ 45 ]. We utilize the pretrained models of these three materials for the classification of new materials, thereby validating their respective generalization performances. 5.1 Pre-Training Performance of SK-MNV4 in Federated Learning Accuracy is a crucial performance metric that reflects the proportion of correctly predicted samples relative to the total number of samples. A higher accuracy indicates superior classification performance of the model. As illustrated in Fig. 8 , even without the application of transfer learning, the model exhibits a commendable ability to recognize new materials through the processes of FL. Among the evaluated models, SwinT demonstrated the lowest performance, with an accuracy of only 72.88%. Despite having a substantial number of learnable parameters, it is not well-suited for the classification of melt pool images addressed in this study. Other network models achieved classification accuracies exceeding 78%, signifying robust generalization capabilities. Notably, our proposed SK-MNV4 model achieved an accuracy of 84.77%, thereby validating its strong generalization performance. We employed a confusion matrix to conduct a detailed analysis of the model's predictive outcomes. In this matrix, the horizontal axis represents the true labels, while the vertical axis denotes the predicted labels. We focused on analyzing the top four network models based on accuracy ranking. As illustrated in Fig. 9 (a) , MobileVIT demonstrates commendable performance in classification, particularly achieving an accuracy of 84.33% for class OM; however, its prediction accuracy for class LM is only 71.67%. Figure 9 (b) shows that EfficientNet excels in identifying class OM, attaining an accuracy of 86.67%, although it tends to misclassify LM as NM. In Fig. 9 (c) , ConvNeXt exhibits a high recognition rate for both NM and LM, but its performance in recognizing OM is relatively moderate. Our proposed model, SK-MNV4 (as shown in Fig. 9 (d) ), performs exceptionally well, with accuracy rates exceeding 82% across all labels and a misclassification rate below 10%. Nevertheless, there remains a need for further improvement in accuracy through subsequent transfer learning efforts. 5.2 Transfer Learning Performance Evaluation After constructing the pre-trained network, a gap remained between the validation accuracy and the actual monitoring accuracy. To address this issue, we employed a dataset comprising 600 images representing three categories of IN718. We implemented transfer learning on the fully connected layers of the network model and performed a comparative experimental analysis. 5.2.1 Classification Accuracy and Model Parameters After transfer learning, we compiled statistics on accuracy and the number of parameters, as shown in Table 2 . We found that SwinT achieved an accuracy of only 85.8% post-transfer learning, indicating its unsuitability for the melt pool image recognition task due to relatively low identification accuracy. DenseNet recorded a recognition accuracy of 90.5%, demonstrating a certain level of monitoring capability. Meanwhile, MobileVIT maintained a commendable accuracy of 92.1% while utilizing the fewest parameters. EfficientNet and ConvNeXt achieved higher accuracies of 92.3% and 95.8%, respectively. Notably, our proposed SK-MNV4 obtained the highest accuracy at 97.3%, showcasing its ability for high-precision melt pool identification following small sample transfer learning. Table 2 Highest accuracy and number of parameters in different models Model Highest accuracy Parameter (M) SK-MNV4 0.973 3.6 ConvNeXt 0.958 27.8 EfficientNet 0.923 6.5 MobileVIT 0.921 0.9 DenseNET 0.905 6.9 SwinT 0.848 85.8 5.2.2 Validation times Due to the high requirements for real-time monitoring of the melt pool, the speed at which models process images is crucial. The computational times of various models for processing 900 melt pool images are illustrated in Fig. 10 . It is evident that the SwinT model, with a large number of parameters, exhibits a significantly longer validation time of 4.03 seconds compared to the other models. Notably, although MobileVIT has the fewest parameters at just 0.9M, its validation time of 0.969 seconds is not the shortest. This can be attributed to the high computational complexity of the self-attention mechanism within MobileVIT, which necessitates greater computational resources and time for processing each layer. In contrast, the proposed SK-MNV4 demonstrates the shortest processing time of only 0.918 seconds, indicating its superior capability for real-time identification of melt pool images. 5.2.3 Data visualization T-SNE is employed to embed high-dimensional data into a lower-dimensional space for visualizing classification outcomes. We selected the top four models based on accuracy for analysis. As illustrated in Fig. 11 (a) , MobileVIT can approximately differentiate between three distinct categories; however, misclassification is prevalent at the boundaries between these categories. Figure 11 (b) shows that EfficientNet effectively separates OM from the other categories, although a significant number of LM instances are incorrectly identified as NM. Figure 11 (c) illustrates that ConvNeXt excels at distinguishing LM from other categories, yet there is still some misidentification of OM and NM. Finally, the proposed SK-MNV4 model, as depicted in Fig. 11 (d) , achieves effective separation among all categories with considerable distance between them, indicating superior classification performance and a reduced rate of misclassification compared to the other network models. 5.2.4 Grad-CAM Grad-CAM is a technique used in deep learning for visualizing the areas of an image that are most important for classification. It works by computing the gradient of the score for a target class concerning the feature maps of CNN, and then multiplying each feature map by its corresponding gradient before summing them up. The resulting heatmap highlights the regions of the input image that were most relevant for the CNN's decision. In the context of classifying melt pool images in L-DED to generate thermal maps, Grad-CAM can be utilized to identify the image regions that are most relevant for a deep learning model's classification decision. This can aid researchers and engineers in gaining a better understanding of the decision-making process of the model. As shown in Fig. 12 , Fig. 12 (a) to (c) displays original melt pool images for three different cases: NM, OM, and LM. In Fig. 12 (d) to (f) , corresponding heatmaps generated by Grad-CAM are presented. For the NM case, the melt pool exhibits a well-defined circular boundary, aligning with the expected ideal formation quality. From Fig. 12 (d) , it is evident that the network model bases its decision on the distinct circular boundary of the melt pool. In the OM case, which often occurs under high laser intensity conditions, resulting in irregular melt pool boundaries and the generation of plumes [ 46 ], the network model relies on the presence of plumes near the melt pool for its decision-making, as observed in Fig. 12 (e) .In the LM case, typically associated with lower laser intensity, Fig. 12 (f) reveals that the network model makes decisions based on the darker regions within the melt pool. Additionally, from Fig. 12 (d) , it is observed that despite the presence of minimal spatter interference near the melt pool, our developed network model, SK-MNV4, does not prioritize such interference, highlighting the model's robustness. 6. Conclusions In this study, we propose a novel federated transfer learning framework tailored for defect detection in L-DED, addressing critical challenges such as small sample size, data privacy, and heterogeneous material domains. At the core of the framework lies SK-MNV4, a lightweight yet expressive backbone network designed to balance classification accuracy with computational efficiency. The architecture integrates UIB modules, Mobile-MQA attention mechanisms, and SK blocks to enable efficient multi-scale feature extraction, global context modeling, and adaptive feature fusion. To preserve industrial data privacy while enhancing model generalization, we adopt a FL strategy across decentralized datasets representing different alloy systems. This enables collaborative model pretraining without sharing raw melt pool images. Subsequent transfer learning is performed using limited target domain data to adapt the pre-trained model to new materials. Experimental evaluations across multiple benchmark networks—including EfficientNet, ConvNeXt, DenseNet, Swin Transformer, and MobileViT—demonstrate that SK-MNV4 achieves the highest classification accuracy (97.3%) following small-sample fine-tuning, with the shortest inference time (0.918 s), indicating strong potential for real-time deployment. Additionally, ablation analyses using confusion matrices, t-SNE embeddings, and Grad-CAM visualizations confirm the model’s superior discriminative capability and robust interpretability. These findings validate the efficacy of combining FL and transfer learning with a lightweight, modular architecture to realize accurate and scalable melt pool monitoring under data-constrained and privacy-sensitive conditions. However, there are still several limitations to this work that warrant further attention. The current framework primarily focuses on single-modal image data, which may not fully capture the thermal or structural anomalies that manifest in other sensing domains. The reliance on melt pool images alone could potentially limit the model's responsiveness to subtle defect patterns, especially under varying lighting or occlusion conditions. In addition, while the FL setting successfully addresses data privacy concerns, the framework presumes stable client participation and consistent network conditions during training. In practical industrial environments, system failures, asynchronous updates, or client-side heterogeneity may introduce instability that could impact overall convergence and model performance. Future research will consider the integration of multi-modal sensor inputs, combining acoustic, thermal, and visual signals within a unified architecture to enrich the representation space and improve defect localization under complex process conditions. Expanding the framework toward federated personalization, where models are dynamically adapted to individual client distributions, may also enhance robustness in non-independent and identically distributed scenarios. Furthermore, embedding continual learning capabilities into the SK-MNV4 framework would enable adaptive model updates over time without retraining from scratch, supporting long-term deployment across evolving materials and process settings. These extensions are expected to further reinforce the scalability and practical utility of the proposed system in real-world additive manufacturing applications. Declarations Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author Contribution F.W. implemented the federated learning framework and carried out the main model training and evaluation. Q.L. performed data preprocessing, ablation studies, and comparative experiments. X.Z. and W.L. assisted in data labeling and image acquisition from the L-DED platform. Z.L. contributed to software validation and visualization. J.G. helped with experimental setup and parameter tuning. S.P. supervised the entire project, provided critical revision of the manuscript, and secured funding. All authors reviewed and approved the final manuscript. Acknowledgement We would like to thank the following individuals for their helpful discussions and suggestions during the course of this research:Zhinong Li, Nanchang Hangkong UniversityJiawei Xiang, Wenzhou UniversityGuangyi Ma, Dalian University of TechnologyTheir insights contributed to the refinement of the methodology and validation of the experimental design. Data availability The codes of the algorithms involved this study have been uploaded to GitHub ( https://github.com/shi-tong/SK-MNV4.git ) References Mohanavel V, Ashraff Ali K S, Ranganathan K, Allen Jeffrey J, Ravikumar M M and Rajkumar S (2021) The roles and applications of additive manufacturing in the aerospace and automobile sector Materials Today: Proceedings 47 405–9 Li H, Gao F, Jiao J, Liu Z, Ji D and Lin J (2023) Acoustic Emission-Based Cross-Domain Process Health Monitoring for Additive Manufacturing IEEE Transactions on Instrumentation and Measurement 72 1–8 Wang H, Liu W, Tang Z, Wang Y, Mei X, Saleheen K M, Wang Z and Zhang H (2020) Review on adaptive control of laser-directed energy deposition OE 59 070901 Kumar R, Kumar M and Chohan J S (2021) The role of additive manufacturing for biomedical applications: A critical review Journal of Manufacturing Processes 64 828–50 Liu T-S, Chen P, Qiu F, Yang H-Y, Jin N T Y, Chew Y, Wang D, Li R, Jiang Q-C and Tan C (2024) Review on laser directed energy deposited aluminum alloys Int. J. Extrem. Manuf. 6 022004 Piscopo G and Iuliano L (2022) Current research and industrial application of laser powder directed energy deposition Int J Adv Manuf Technol 119 6893–917 Wang C, Tan X P, Tor S B and Lim C S (2020) Machine learning in additive manufacturing: State-of-the-art and perspectives Additive Manufacturing 36 101538 Dilip J J S, Zhang S, Teng C, Zeng K, Robinson C, Pal D and Stucker B (2017) Influence of processing parameters on the evolution of melt pool, porosity, and microstructures in Ti-6Al-4V alloy parts fabricated by selective laser melting Prog Addit Manuf 2 157–67 Baumgartl H, Tomas J, Buettner R and Merkel M (2020) A deep learning-based model for defect detection in laser-powder bed fusion using in-situ thermographic monitoring Prog Addit Manuf 5 277–85 Chen L, Yao X, Tan C, He W, Su J, Weng F, Chew Y, Ng N P H and Moon S K (2023) In-situ crack and keyhole pore detection in laser directed energy deposition through acoustic signal and deep learning Additive Manufacturing 69 103547 Ren W, Wen G, Zhang Z and Mazumder J (2022) Quality monitoring in additive manufacturing using emission spectroscopy and unsupervised deep learning Materials and Manufacturing Processes 37 1339–46 Shin S-J, Lee J-H, Jadhav S and Kim D B (2024) Material-Adaptive Anomaly Detection Using Property-Concatenated Transfer Learning in Wire Arc Additive Manufacturing Int. J. Precis. Eng. Manuf. 25 383–408 Deng T, Li Y, Liu X and Wang L (2023) FL-based collaborative manufacturing for complex parts J Intell Manuf 34 3025–38 Tang Y, Rahmani Dehaghani M and Wang G G (2023) Review of transfer learning in modeling additive manufacturing processes Additive Manufacturing 61 103357 Tang Y, Rahmani Dehaghani M, Sajadi P and Wang G G (2024) Selecting subsets of source data for transfer learning with applications in metal additive manufacturing J Intell Manuf Peng S, Yang S, Gao B, Liu W, Wang F and Tang Z (2025) Prediction of 3D temperature field through single 2D temperature data based on transfer learning-based PINN model in laser-based directed energy deposition Journal of Manufacturing Processes 138 140–56 Qi P, Chiaro D, Guzzo A, Ianni M, Fortino G and Piccialli F (2024) Model aggregation techniques in FL: A comprehensive survey Future Generation Computer Systems 150 272–93 Wu Y, Zhang S, Yu W, Liu Y, Gu Q, Zhou D, Chen H and Cheng W (2023) Personalized FL under Mixture of Distributions Proceedings of the 40th International Conference on Machine Learning International Conference on Machine Learning (PMLR) pp 37860–79 Guan H, Yap P-T, Bozoki A and Liu M (2024) FL for medical image analysis: A survey Pattern Recognition 151 110424 Mehta M and Shao C (2022) FL-based semantic segmentation for pixel-wise defect detection in additive manufacturing Journal of Manufacturing Systems 64 197–210 Li Z, Li Z and Gu F (2024) Intelligent diagnosis method for machine faults based on federated transfer learning Applied Soft Computing 163 111922 Zhu X, Jiang F, Guo C, Xu D, Wang Z and Jiang G (2023) Surface morphology inspection for directed energy deposition using small dataset with transfer learning Journal of Manufacturing Processes 93 101–15 Petrich J, Snow Z, Corbin D and Reutzel E W (2021) Multi-modal sensor fusion with machine learning for data-driven process monitoring for additive manufacturing Additive Manufacturing 48 102364 Wang R, Garcia D, Kamath R R, Dou C, Ma X, Shen B, Choo H, Fezzaa K, Yu H Z and Kong Z (James) (2022) In situ melt pool measurements for laser powder bed fusion using multi sensing and correlation analysis Sci Rep 12 13716 Bisheh M N, Wang X, Chang S I, Lei S and Ma J (2023) Image-based characterization of laser scribing quality using transfer learning J Intell Manuf 34 2307–19 Anon (2025) Personalized federated unsupervised learning for nozzle condition monitoring using vibration sensors in additive manufacturing Robotics and Computer-Integrated Manufacturing 93 102940 Mehta M, Bimrose M V, McGregor D J, King W P and Shao C (2024) FL enables privacy-preserving and data-efficient dimension prediction and part qualification across additive manufacturing factories Journal of Manufacturing Systems 74 752–61 Yang Z, Lu Y, Yeung H and Krishnamurty S (2019) Investigation of Deep Learning for Real-Time Melt Pool Classification in Additive Manufacturing 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) pp 640–7 Jiang H-Z, Li Z-Y, Feng T, Wu P-Y, Chen Q-S, Feng Y-L, Chen L-F, Hou J-Y and Xu H-J (2021) Effect of Process Parameters on Defects, Melt Pool Shape, Microstructure, and Tensile Behavior of 316L Stainless Steel Produced by Selective Laser Melting Acta Metall. Sin. (Engl. Lett.) 34 495–510 Scime L and Beuth J (2019) Melt pool geometry and morphology variability for the Inconel 718 alloy in a laser powder bed fusion additive manufacturing process Additive Manufacturing 29 100830 Ertay D S, Naiel M A, Vlasea M and Fieguth P (2021) Process performance evaluation and classification via in-situ melt pool monitoring in directed energy deposition CIRP Journal of Manufacturing Science and Technology 35 298–314 Li J, Zhou Q, Cao L, Wang Y and Hu J (2022) A convolutional neural network-based multi-sensor fusion approach for in-situ quality monitoring of selective laser melting Journal of Manufacturing Systems 64 429–42 Tang Y, Rahmani Dehaghani M, Sajadi P and Wang G G (2024) Selecting subsets of source data for transfer learning with applications in metal additive manufacturing J Intell Manuf Menon N, Mondal S and Basak A (2022) Transferring Process Mapping Knowledge across SS316L and IN718 in Laser Directed Energy Deposition Using Machine Learning Mochi V H, Núñez H H L, Ribeiro K S B and Venter G S (2023) Real-time prediction of deposited bead width in L-DED using semi-supervised transfer learning Int J Adv Manuf Technol 129 5643–54 Feenstra D R, Molotnikov A and Birbilis N (2021) Utilisation of artificial neural networks to rationalise processing windows in directed energy deposition applications Materials & Design 198 109342 Akbari P, Ogoke F, Kao N-Y, Meidani K, Yeh C-Y, Lee W and Barati Farimani A (2022) MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning Additive Manufacturing 55 102817 Ho S, Zhang W, Young W, Buchholz M, Jufout S A, Dajani K, Bian L and Mozumdar M (2021) DLAM: Deep Learning Based Real-Time Porosity Prediction for Additive Manufacturing Using Thermal Images of the Melt Pool IEEE Access 9 115100–14 Qin D, Leichner C, Delakis M, Fornoni M, Luo S, Yang F, Wang W, Banbury C, Ye C, Akin B, Aggarwal V, Zhu T, Moro D and Howard A (2024) MobileNetV4 -- Universal Models for the Mobile Ecosystem Liu G SK-MobileNet: A Lightweight Adaptive Network Based on Complex Deep Transfer Learning for Plant Disease Recognition Arabian Journal for Science and Engineering Tan M and Le Q V (2020) EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks Liu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T and Xie S (2022) A ConvNet for the 2020s Mehta S and Rastegari M (2022) MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer Huang G, Liu Z, Maaten L van der and Weinberger K Q (2018) Densely Connected Convolutional Networks Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S and Guo B (2021) Swin Transformer: Hierarchical Vision Transformer using Shifted Windows Coeck S, Bisht M, Plas J and Verbist F (2019) Prediction of lack of fusion porosity in selective laser melting based on melt pool monitoring data Additive Manufacturing 25 347–56 Additional Declarations No competing interests reported. Supplementary Files Network.zip 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-7002574","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":485035824,"identity":"2834401f-aee2-46d4-b0aa-358adb55f557","order_by":0,"name":"Fengtao Wang","email":"","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":false,"prefix":"","firstName":"Fengtao","middleName":"","lastName":"Wang","suffix":""},{"id":485035825,"identity":"6f750863-03fa-42e1-b508-6b13f8055a0b","order_by":1,"name":"Quanjie Li","email":"","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":false,"prefix":"","firstName":"Quanjie","middleName":"","lastName":"Li","suffix":""},{"id":485035826,"identity":"d1b9bec6-7d77-4aec-8ed3-7c42243d4cdc","order_by":2,"name":"Xinxin Zeng","email":"","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":false,"prefix":"","firstName":"Xinxin","middleName":"","lastName":"Zeng","suffix":""},{"id":485035827,"identity":"e50d58f5-ddc9-424d-8dae-90d403590cb6","order_by":3,"name":"Weiwei Liu","email":"","orcid":"","institution":"Dalian University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Weiwei","middleName":"","lastName":"Liu","suffix":""},{"id":485035828,"identity":"530b9312-f470-4e75-a7cd-0b7198f2c4c9","order_by":4,"name":"Zhiqi Luo","email":"","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":false,"prefix":"","firstName":"Zhiqi","middleName":"","lastName":"Luo","suffix":""},{"id":485035829,"identity":"c6b454e1-fae7-4b4a-adae-ed3d1b3e8823","order_by":5,"name":"Jianan Guo","email":"","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":false,"prefix":"","firstName":"Jianan","middleName":"","lastName":"Guo","suffix":""},{"id":485035830,"identity":"97f31a40-49a7-4432-84fb-5364668bde0e","order_by":6,"name":"Shitong Peng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYDACCQglB6HYiNLCDKaMIapJ0ZLYQLQW+dn9xx583FGbPn9+jwHDh7LDDPyzG/BrMbhzmN1w5pnjuRuO8Rgwzjh3mEHizgECWiSS2aR5247lbmDjMWDmbTsMFEkg4LAZQC1/246ly7cBtfwlRgvDDaAWxraaBAagw5gZidFicCPZ3LC37YDhhmNpBQd7zqXzSNwg6LDEZw9+ttXJyzcf3vjgR5m1HP8MQg6DxMVhMOsAEPMQVA/VUkeMwlEwCkbBKBipAACuoEBonZUGRQAAAABJRU5ErkJggg==","orcid":"","institution":"Shantou University, Ministry of Education","correspondingAuthor":true,"prefix":"","firstName":"Shitong","middleName":"","lastName":"Peng","suffix":""}],"badges":[],"createdAt":"2025-06-29 12:08:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7002574/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7002574/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":86776870,"identity":"c59bb8d9-2f82-425e-8dd4-8d5d31caba83","added_by":"auto","created_at":"2025-07-15 12:47:30","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":134946,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Schematic diagram of L-DED device operation. (b) KUKA six-axis robot. and (c) Laser head.\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/961f88869dafa2891ec93594.jpg"},{"id":86776871,"identity":"20ce508c-034c-467e-854c-7e7ed4c45695","added_by":"auto","created_at":"2025-07-15 12:47:30","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":47270,"visible":true,"origin":"","legend":"\u003cp\u003eSingle-track deposit scanning process.\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/8c424dad49d21ac1e37b2571.jpg"},{"id":86776872,"identity":"4cfffc0d-3fc4-4200-a97a-f6a219024987","added_by":"auto","created_at":"2025-07-15 12:47:30","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":38175,"visible":true,"origin":"","legend":"\u003cp\u003eThe cross-section of a single-track deposit (a) experimental cross-section (b) schematic diagram\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/8ba7ce9c6c73fafb5780a7e6.jpg"},{"id":86778448,"identity":"e788c71e-9a3f-4392-a621-f0885a379759","added_by":"auto","created_at":"2025-07-15 12:55:30","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":114135,"visible":true,"origin":"","legend":"\u003cp\u003eGrouping of cross-sectional information for different materials.\u003c/p\u003e","description":"","filename":"fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/3d9990cb3ad304a29639f0df.jpg"},{"id":86776878,"identity":"8d358baa-8759-4948-83cc-5d8a97e4b6a6","added_by":"auto","created_at":"2025-07-15 12:47:30","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":183111,"visible":true,"origin":"","legend":"\u003cp\u003eFederated Learning Framework for SK-MNV4 in L-DED Defect Detection\u003c/p\u003e","description":"","filename":"fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/211bf5d41a69cea3c81b2859.png"},{"id":86776875,"identity":"dec94ce7-e2fb-4708-a140-70a7d3fae8a1","added_by":"auto","created_at":"2025-07-15 12:47:30","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":126109,"visible":true,"origin":"","legend":"\u003cp\u003eThe process of FL.\u003c/p\u003e","description":"","filename":"fig6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/b68a3aec913687b3b1bec6d1.jpg"},{"id":86778461,"identity":"89c34c3e-b404-4097-b790-07a6546e5730","added_by":"auto","created_at":"2025-07-15 12:55:31","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":217972,"visible":true,"origin":"","legend":"\u003cp\u003eSK-MNV4. (a) Overall structure. (b) UIB block. (c) SK block.\u003c/p\u003e","description":"","filename":"fig7.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/9d9de2e88ecc65f3c2e43752.png"},{"id":86778449,"identity":"aaabd121-fd42-4165-9dc7-820452be5f95","added_by":"auto","created_at":"2025-07-15 12:55:30","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":77632,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy of different models.\u003c/p\u003e","description":"","filename":"fig8.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/0c5798d2b8d085ce126a2898.png"},{"id":86778735,"identity":"5372f95e-0703-44b7-af62-cd51d0d739e4","added_by":"auto","created_at":"2025-07-15 13:03:30","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":381973,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix of different models.\u003c/p\u003e","description":"","filename":"fig9.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/1550fb5a5beead1be9ed59c3.png"},{"id":86778451,"identity":"e708d687-3233-4329-8f70-04a3ea807fa7","added_by":"auto","created_at":"2025-07-15 12:55:30","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":36804,"visible":true,"origin":"","legend":"\u003cp\u003eComputational time of different models under the validation set.\u003c/p\u003e","description":"","filename":"fig10.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/03153af7cb89c7cfe976304f.png"},{"id":86778737,"identity":"f31ffe90-039e-41b1-b0db-f8dc74c54925","added_by":"auto","created_at":"2025-07-15 13:03:31","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":273982,"visible":true,"origin":"","legend":"\u003cp\u003eT-SNE visualization of feature embeddings from different models\u003c/p\u003e","description":"","filename":"fig11.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/dd7ba9787e60813e603fc091.png"},{"id":86776891,"identity":"031a3e14-9dc4-4f57-b00e-deb8d3d557f1","added_by":"auto","created_at":"2025-07-15 12:47:31","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":271008,"visible":true,"origin":"","legend":"\u003cp\u003eMelt pool images: (a) NM (b) OM (c) LM, related Grad-CAM outputs: (d) NM (e) OM (f) LM.\u003c/p\u003e","description":"","filename":"fig12.png","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/5fc030ad931ed2d7110d776e.png"},{"id":91429530,"identity":"29454da1-6ed3-49bf-83dd-b37b1975f357","added_by":"auto","created_at":"2025-09-16 12:02:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2810217,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/49a44f63-9f0e-44fb-aad2-944b7c2ab3b4.pdf"},{"id":86776912,"identity":"9693c8ec-ee32-4351-b70d-61bb366ce523","added_by":"auto","created_at":"2025-07-15 12:47:32","extension":"zip","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":30257989,"visible":true,"origin":"","legend":"","description":"","filename":"Network.zip","url":"https://assets-eu.researchsquare.com/files/rs-7002574/v1/7a04a493d4699e19717dfb5c.zip"}],"financialInterests":"No competing interests reported.","formattedTitle":"Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAdditive manufacturing (AM), also known as 3D printing, is a manufacturing process that creates objects layer by layer by adding material [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Laser direct energy deposition is one of the most dominant AM forms in current manufacturing practices [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. It represents a novel approach to manufacturing high-performance components for aerospace, medical, energy, and automotive components [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], as well as repairing existing metal parts [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Due to the inherent characteristics of L-DED, such as high thermal gradients and complex melt pool dynamics, the process often exhibits significant variability, making it difficult to control the quality of the final parts. Despite the significant scientific and engineering efforts in industry and academia devoted to L-DED, enhancing the production efficiency and fabrication quality remains a significant challenge [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. The production of high-quality printed parts is a significant concern, as any defects could damage the part's structural integrity [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Therefore, process monitoring is an essential way to improve L-DED component quality and decrease variability in mechanical performance.\u003c/p\u003e\u003cp\u003eRecent studies on L-DED process monitoring have primarily focused on data-driven defect detection and quality prediction, leveraging sensing techniques, such as thermal imaging and melt pool visualization, combined with deep learning or unsupervised algorithms. Baumgartl et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] employed thermographic off-axis imaging and deep learning neural network architectures to detect delamination and splatter defects. Chen et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] introduced an in-situ acoustic-based defect detection strategy for L-DED, leveraging convolutional neural networks and acoustic signal analysis to achieve high accuracy in identifying cracks and keyhole pores. Ren et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] developed an unsupervised recognition model using LSTM-Autoencoder and K-means clustering for in-situ quality monitoring of DED-printed parts, demonstrating accurate differentiation between qualified and unqualified depositions. However, all of the aforementioned methods are constrained by data scarcity. Training high-performance deep learning models typically relies on large-scale, diverse, and high-quality datasets. This challenge is particularly pronounced in tasks involving micro-defect identification, where it is especially difficult to acquire multimodal sensing data, such as melt pool images, thermal profiles, and acoustic signals [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] that adequately represent typical processing conditions. These limitations collectively lead to insufficient data dimensionality and poor representativeness, making it difficult for the resulting models to achieve the desired levels of detection accuracy and generalization capability [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Transfer learning addresses data scarcity by transferring pre-trained model parameters from a source domain to a similar target task [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. For example, Peng et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] developed a transfer learning-based PINN model to predict 3D temperature fields in laser-based directed energy deposition, achieving an RMSE of 6.365 K and an average error of 1.28%. However, most existing solutions remain limited in addressing the dual challenges of data scarcity and privacy constraints, especially when dealing with heterogeneous materials and rare defect types. These gaps underscore the urgent need for a generalizable and privacy-preserving framework that can maintain high defect detection accuracy even under limited and decentralized datasets.\u003c/p\u003e\u003cp\u003eIn the process of L-DED, data such as melt pool images and process parameters often involve key information about the manufacturing workflow. Due to the high emphasis on data privacy, manufacturers are generally unwilling to share such sensitive data. Meanwhile, intense market competition among companies further aggravates data closure. This overprotection of data privacy limits cross-organizational data sharing and the development of collaborative monitoring mechanisms, and makes it difficult for traditional transfer learning approaches to acquire broad prior knowledge. To address this challenge, researchers introduce an federated learning (FL) framework that enables privacy-preserving collaborative training by retaining local melt pool image data and exchanging only model parameters. This decentralized approach not only alleviates data-sharing concerns among L-DED manufacturers but also enhances the model\u0026rsquo;s cross-material generalization capability under heterogeneous and data-scarce conditions [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. FL is a machine learning approach where multiple devices collaboratively train a model without sharing raw data. Instead, only model updates are exchanged, preserving user privacy. It improves data security, reduces communication costs, and enables learning from decentralized data sources, making it ideal for privacy-sensitive applications [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. For instance, Mehta et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] proposed a FL-based semantic segmentation method for defect detection in additive manufacturing, addressing data availability and privacy issues while achieving performance similar to centralized learning. Li et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] proposed a federated transfer learning fault diagnosis model that preserves data privacy and enables cross-domain diagnosis by aligning feature distributions. These studies demonstrate that FL can effectively balance privacy preservation and model performance in additive manufacturing. Moreover, by enabling cross-domain generalization and outperforming centralized methods, FL proves to be a promising approach for intelligent defect detection and quality control in heterogeneous and data-sensitive environments. Through our literature review, we found that few studies have explored the integration of FL techniques in L-DED applications.\u003c/p\u003e\u003cp\u003eTo address the aforementioned challenges, this study proposes an SK-MNV4 lightweight network within a FL framework, integrating FL and transfer learning to address the problems of data privacy protection and sample scarcity in the L-DED process. Specifically, this network adopts a unified inverted bottleneck structure (UIB), a Mobile-MQA attention mechanism, and a selective convolution kernel module (SK) to achieve efficient multi-scale feature extraction and channel fusion. The present study performs L-DED experiments with multiple metal alloys. We use three alloys (i.e., IN738, TB6, and Ti-10Mo) as federated clients and execute distributed pre-training to ensure that the molten pool image data is processed locally, ensuring privacy security. The Federated Averaging (FedAvg) algorithm aggregates local model updates to build a global model with cross-material generalization capabilities. Subsequently, the model is migrated to the IN718 material with only a small number of labeled samples, followed by fine-tuning the classification layer. Combined with visualization analysis such as Grad-CAM and t-SNE, we further verify the robustness of the model in discriminating area positioning and category differentiation. This method takes into account high precision, low latency and good scalability, and provides a feasible path for intelligent defect warning in the L-DED process. The key contributions of this study are as follows: (1) A novel FL framework is introduced for defect detection in L-DED, enabling privacy-preserving model training across multiple material domains. (2) The lightweight SK-MNV4 network is proposed and validated, showing strong performance in both accuracy and efficiency. (3) A federated transfer learning strategy is developed to extend model generalization to new materials using minimal target data.\u003c/p\u003e\u003cp\u003eThe structure of this paper is as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e provides an overview of the fundamental theoretical concepts of FL; Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the proposed SK-MNV4 network; Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e4\u003c/span\u003e explains the experimental design approach; Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents a comparative analysis of the proposed network against other models, followed by conclusions in Section \u003cspan refid=\"Sec22\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e"},{"header":"2. Related works","content":"\u003cp\u003eDue to the complexity and instability of the deposition process, defects such as porosity, cracking, and incomplete fusion frequently occur, necessitating robust real-time monitoring and intelligent defect detection systems. Recent advances in machine learning, especially deep learning, have driven the development of intelligent monitoring solutions. Techniques like FL and transfer learning have been introduced to further enhance the adaptability, scalability, and efficiency of these systems under real-world constraints such as data privacy, small sample size, and platform heterogeneity [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Challenges and advances in monitoring L-DED processes\u003c/h2\u003e\u003cp\u003eL-DED is a complex thermal-mechanical process characterized by rapid heating and cooling cycles, intricate melt pool dynamics, and sensitivity to numerous parameters [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These factors contribute to a high likelihood of defect formation, thereby driving the need for real-time, intelligent monitoring technologies [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In recent years, sensor technologies\u0026mdash;ranging from visible-light imaging and infrared thermography to acoustic emission\u0026mdash;have been increasingly adopted to capture process signals [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Combined with deep learning-based approaches, such as convolutional neural networks (CNNs), these systems have shown promise in classifying and localizing defects based on real-time image and signal data [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Nevertheless, several fundamental bottlenecks limit their practical deployment.\u003c/p\u003e\u003cp\u003eForemost among these is the issue of data scarcity: high-quality, annotated datasets are often difficult to obtain due to the costly and time-consuming nature of metallographic inspection and defect labeling. These limitations are further exacerbated when targeting rare or subtle defect types, such as micro-porosity, incomplete fusion, or keyhole-induced vaporization, where ground truth labeling is difficult to standardize and highly sensitive to measurement noise and operator subjectivity. Consequently, existing datasets often suffer from insufficient size and representativeness, ultimately restricting the learning capacity of deep models and impairing their ability to generalize across unseen process conditions [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAdditionally, generalization across platforms remains a significant issue. Data heterogeneity induced by differences in equipment, materials, and process settings results in domain shifts that compromise the robustness and transferability of models trained on a single source domain [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Moreover, concerns over data privacy and proprietary information limit data sharing across institutions, leading to fragmented datasets and inconsistent model performance [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. As a result, collaborative dataset construction is limited, and models suffer from poor generalizability and overfitting. Finally, existing monitoring algorithms are often computationally intensive, challenging their deployment in real-time or edge-computing environments. Addressing these constraints calls for more flexible, privacy-preserving, and data-efficient modeling strategies [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Federated learning for privacy-preserving collaborative modeling\u003c/h2\u003e\u003cp\u003eFL has emerged as a promising paradigm to address data privacy and collaboration barriers in intelligent manufacturing. It enables decentralized model training across multiple clients or devices by exchanging model parameters instead of raw data, thus preserving confidentiality while building a robust global model [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. In the context of additive manufacturing, FL facilitates collaborative learning among different manufacturing units or institutions, enabling them to share model intelligence without compromising proprietary data [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. This capability is critical in industries where intellectual property and trade secrets are paramount, making FL a cornerstone for secure, collaborative innovation.\u003c/p\u003e\u003cp\u003eThe integration of FL with techniques like transfer learning has further amplified its potential. These hybrid approaches enhance defect detection across heterogeneous platforms, enabling models to adapt to diverse machines, materials, and operational conditions. For instance, FL has allowed manufacturers using different machines and materials to collaboratively train models that adapt to various working conditions [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. For instance, lightweight neural networks and communication-efficient strategies have been proposed to make FL viable on resource-constrained edge devices, broadening its applicability in real-time industrial settings [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eNonetheless, the practical implementation of FL in industrial settings still encounters notable challenges. One major issue is communication overhead, as federated training requires frequent parameter exchange, which may be infeasible under limited network infrastructure [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Additionally, the non-independent and identically distributed nature of manufacturing data across clients significantly affects convergence speed and model accuracy [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. FL systems are also vulnerable to system failures, straggling devices, and even malicious participants, raising concerns about robustness and trustworthiness. Most critically, there exists a fundamental trade-off between model performance and privacy preservation, which remains an open problem for real-world deployment [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Addressing these challenges requires innovative solutions, such as adaptive aggregation algorithms to handle non-IID data and robust encryption protocols to mitigate security risks. Furthermore, developing scalable FL frameworks that balance computational efficiency with privacy will be essential for widespread adoption in dynamic manufacturing ecosystems, ultimately unlocking the full potential of AI-driven process optimization.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Transfer learning strategies for data-efficient defect detection\u003c/h2\u003e\u003cp\u003eTransfer Learning (TL) offers a compelling solution to the limitations of data scarcity and generalization in deep learning [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. By leveraging knowledge from source domains or pre-trained models, TL enables efficient learning in target domains with limited data. In additive manufacturing, TL techniques have been adopted to improve model performance in tasks such as defect classification and process anomaly detection [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. The most common strategy involves fine-tuning pre-trained models for manufacturing-specific datasets. These models capture low-level features that are still relevant in industrial imagery and have demonstrated decent performance with small datasets [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. In addition to model-based transfer, feature-based domain adaptation techniques such as maximum mean discrepancy and adversarial learning have been used to bridge the distribution gap between source and target domains [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Some studies have also explored the integration of TL with FL, creating collaborative frameworks that can generalize across different factories, machines, or even types of defects.\u003c/p\u003e\u003cp\u003eHowever, the use of TL in additive manufacturing is still hampered by several obstacles. One critical challenge is domain mismatch, as pre-trained models are often based on natural images, while manufacturing data has distinct visual and statistical properties [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This can lead to negative transfer, where the imported knowledge degrades rather than improves performance. Another issue is the semantic gap between tasks, especially when the source task differs significantly from the target task [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Additionally, transferability is hard to quantify or predict, and there is no unified framework to determine when and how transfer learning will be effective [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Furthermore, most TL approaches are static and single-source, while real-world manufacturing involves dynamic, multi-source, and continuously evolving scenarios [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. There is an urgent need for multi-source, adaptive, and lifelong transfer learning frameworks that can adapt to new tasks and domains over time. Lastly, maintaining model interpretability, robustness, and safety during the transfer process remains a critical yet unresolved issue, particularly for high-stakes industrial applications [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Motivation and overview of the proposed framework\u003c/h2\u003e\u003cp\u003eTo address these multifaceted challenges holistically, this study proposes a novel federated transfer learning framework, which tightly integrates a lightweight yet powerful backbone network, termed SK-MNV4, with privacy-preserving FL and small-sample transfer learning strategies. The core idea of the proposed framework is to collaboratively leverage decentralized, heterogeneous melt pool datasets without compromising data ownership or security, while simultaneously improving the model's adaptability to unseen material domains.\u003c/p\u003e\u003cp\u003eSpecifically, the proposed approach operates in two key stages. In the first stage, FL is employed to pre-train a global model across multiple decentralized datasets, each corresponding to a distinct alloy system (e.g., IN738, TB6, Ti-10Mo). Each dataset acts as a federated client, conducting local model training without uploading raw melt pool images to the central server. Only model updates, such as weights or gradients, are securely transmitted for aggregation using FedAvg algorithm. This design ensures that sensitive manufacturing data remains confined within each client, thereby addressing privacy and confidentiality concerns. Furthermore, by learning from diverse material domains in a decentralized manner, the global model gains robust generalization capabilities, mitigating the effects of data heterogeneity across clients.\u003c/p\u003e\u003cp\u003eAt the heart of the FL framework is the SK-MNV4 network, an enhanced version of MobileNetV4 tailored for defect detection tasks in L-DED environments. SK-MNV4 introduces several architectural innovations, including UIB blocks, lightweight multi-head attention mechanisms, and selective kernel (SK) feature fusion modules. These modifications significantly enhance the model\u0026rsquo;s ability to capture multi-scale melt pool features while maintaining a compact parameter footprint and fast inference speed, making it highly suitable for real-time deployment on resource-constrained industrial platforms.\u003c/p\u003e\u003cp\u003eIn the second stage, to further adapt the pre-trained global model to a new target domain characterized by a different material system (e.g., IN718) with small sample availability, a transfer learning procedure is applied. The pre-trained SK-MNV4 model is fine-tuned using a limited number of labeled target domain samples, focusing primarily on updating the classification layers while retaining the generalized feature extraction capabilities learned during federated training. This two-stage approach not only minimizes the dependence on large labeled datasets but also significantly enhances model adaptability and reduces the risk of negative transfer commonly encountered when source and target domains differ.\u003c/p\u003e\u003cp\u003eThrough the combination of FL and transfer learning, the proposed framework is designed to achieve three pivotal objectives: (1) Enhance defect detection performance under data scarcity conditions by leveraging prior knowledge learned from related material domains; (2) Preserve data privacy and confidentiality during collaborative model development, ensuring compliance with industrial data protection requirements; and (3) Improve model adaptability and robustness across heterogeneous material domains, enabling reliable defect detection in real-world, dynamic manufacturing environments.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Experiment","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Materials\u003c/h2\u003e\u003cp\u003eIn this L-DED experiment, three materials\u0026mdash;IN738, TB6, and Ti-10Mo\u0026mdash;were utilized as the training dataset for FL, while IN718 served as the validation dataset. The experimental metal powder employed had a diameter range of 53\u0026ndash;160 \u0026micro;m. As part of the preparation, the metal powder was pre-dried at 120\u0026deg;C for 4 hours in a vacuum drying oven. The nickel-based substrate measured 200 mm \u0026times; 200 mm \u0026times; 5 mm. To prevent oxidation of the substrate surface, which could adversely affect the final results, a polishing treatment was conducted on the substrate.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Experiment device\u003c/h2\u003e\u003cp\u003eThis experiment was conducted on an L-DED experimental equipment. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u003cb\u003e(a)\u003c/b\u003e illustrates the basic operational schematic of the L-DED system. The laser emitter GmbH LDF 4000-100VGP can emit laser power up to 4400W, and the laser is released through the laser head (Gaggenau YC-52). To prevent interference from metal powder oxidation on the experimental results, we adopted 99.99% pure argon gas as shielding gas. The water chiller was employed for temperature stabilization of the laser head. Beneath the laser head was a static workbench for fixing the substrate. The laser head outfitted with a CCD camera can capture the melt pool image, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u003cb\u003e(c)\u003c/b\u003e. The CCD camera was installed coaxially on the laser head with a frequency of 50Hz and a pixel resolution of 500*500 for image acquisition. a local computer connected the CCD camera via a serial port to save the captured images locally.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Experiment design\u003c/h2\u003e\u003cp\u003eDue to the significant impact of four process variables, namely, laser power, beam diameter, scan speed, and powder supply rate, on the quality of the formed parts [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], we designed and conducted experiments with four different materials under varying conditions, as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. To label the melt pool images of each single-track deposit, we performed post-processing on the deposits. It should be noted that we adopted a line-cutting approach to select a stable portion between 5-10mm (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), given that the laser emission was initially unstable.\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\u003eExperimental parameters under different deposition materials\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=\"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\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\u003eIN738\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTB6\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTi-10Mo\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIN718\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLaser Power(W)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e800\u0026thinsp;~\u0026thinsp;2600\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e500\u0026thinsp;~\u0026thinsp;2000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e500\u0026thinsp;~\u0026thinsp;1400\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1000\u0026thinsp;~\u0026thinsp;3000\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBeam Diameter(mm)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2, 3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1, 2, 3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eScan Speed(mm/min)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e200\u0026thinsp;~\u0026thinsp;2000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e300\u0026thinsp;~\u0026thinsp;1200\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e420\u0026thinsp;~\u0026thinsp;1140\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e800\u0026thinsp;~\u0026thinsp;1900\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePowder Supply Rate(g/min)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3.6\u0026thinsp;~\u0026thinsp;30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e5, 7, 9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e6, 8, 10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e9.3, 18.3, 29.4\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\u003e\u003c/p\u003e\u003cp\u003eThis section of deposits was polished using 240, 600, 1200, and 2000 grit sandpapers and a polishing cloth. Subsequently, we captured magnified images of the cross-sections using an optical mirror and measured the contact angle (α), track width (W), melt pool height (H), and depth of penetration(D) of the melt pool images using ImageJ software. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003cb\u003e(a)\u003c/b\u003e displays the cross-section of an experimental single-track deposit, while Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003cb\u003e(b)\u003c/b\u003e illustrates a schematic model showcasing the necessary measurements for the cross-section of the single-track deposit.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003ePrevious studies have shown that different shapes of the melt pool imply different quality [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. The desired shape for a melt pool, which is free from any defects or porosities, is considered to be a semi-circular shape in its conduction mode[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. When the depth of the melt pool exceeds half of its width, it can be considered as being in keyhole mode, which is often caused by excessive energy density [\u003cspan additionalcitationids=\"CR34 CR35\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. It is worth noting that keyhole mode does not necessarily result in the formation of porosity [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Additionally, it is important to control the contact angle of the melt pool, as excessive contact angles can lead to porosity during interlayer deposition. In situations where the melt pool inadequately penetrates the underlying material, it can be identified as a \"lack of fusion\" occurrence, frequently leading to incomplete fusion between adjacent layers [\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Therefore, to make the melt pool classification mechanism more reasonable, this paper integrates two indicators, namely dilution rate and contact angle. Dilution rate is defined as D/(H\u0026thinsp;+\u0026thinsp;D) for ease of calculation, and the contact angle is the average of the left and right contact angles. Based on the research mentioned above, the melt pool quality is classified as: (1) Normal Melt (NM): 10%\u0026lt; dilution rate\u0026thinsp;\u0026lt;\u0026thinsp;30% and θ\u0026thinsp;\u0026lt;\u0026thinsp;80\u0026deg;; (2) Over Melt (OM): dilution rate\u0026thinsp;\u0026gt;\u0026thinsp;30% or θ\u0026thinsp;\u0026gt;\u0026thinsp;80\u0026deg;; and (3) Lack of Melt (LM): dilution rate\u0026thinsp;\u0026lt;\u0026thinsp;10%. The grouping of the four different materials for single-track deposits is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eAfter categorizing each material, we randomly selected 2000 melt pool images from each category of IN738, TB6, and Ti-10Mo as the training set. Each material's data serves as a FL client. Additionally, we extracted 500 images from each category of IN718, with 300 images designated for the FL validation set and 200 images allocated for a small sample dataset in federated transfer learning.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Proposed method","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Overview of the Federated Transfer Learning Framework\u003c/h2\u003e\u003cp\u003eThe framework illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the L-DED using the SK-MNV4 model. It begins with three distinct datasets\u0026mdash;IN738, TB6, and Ti-10Mo\u0026mdash;each representing a different alloy, acting as federated clients for local training. These models are trained independently on their respective datasets, ensuring data privacy. The SK-MNV4 model integrates various modules, including the UIB, Mobile MQA, and SK Block, to enhance feature extraction and fusion capabilities. The local model updates from each client are aggregated using the FedAvg algorithm to form a global model. This global model is then fine-tuned with a small labeled dataset (IN718) before being deployed for real-time melt pool defect monitoring, including the detection of defects such as spattering. This approach efficiently addresses the challenges of data privacy and small sample sizes while enabling high-accuracy monitoring in diverse material domains.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Fundamentals of Federated Learning and FedAvg Algorithm\u003c/h2\u003e\u003cp\u003eMathematically, the goal of FL is to solve the global optimization problem by aggregating local model updates from a set of clients. Formally, let the global objective function \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:f\\left(w\\right)\\)\u003c/span\u003e\u003c/span\u003e represent the loss function of a shared model, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:w\\)\u003c/span\u003e\u003c/span\u003e are the global model parameters. The objective is to minimize the weighted sum of local loss functions:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$${w_{t+1}}=\\sum\\limits_{{t=1}}^{N} {\\frac{{{n_i}}}{{\\sum {_{{j=1}}^{N}{n_j}} }}} {w_i}(t)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the weight assigned to client \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e (often based on the number of data samples it holds), and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{i}\\left(w\\right)\\)\u003c/span\u003e\u003c/span\u003e represents the local loss function of client \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e. This approach allows the server to aggregate model updates from clients without ever directly accessing the clients\u0026rsquo; data.\u003c/p\u003e\u003cp\u003eThe process of FL is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The server distributes the global model to individual users, who then utilize their local data to train local models and subsequently upload the model parameters to the server. The server employs an aggregation algorithm to consolidate the local model parameters, thereby updating the global model, which is then retransmitted to the users. This iterative process continues until the maximum number of federated iterations is reached, resulting in the final federated global model.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe primary algorithm used in FL is FedAvg, a simple yet effective approach for aggregating local model updates. Each client \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e computes alocal model update \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\varDelta\\:{w}_{i}\\)\u003c/span\u003e\u003c/span\u003e using its local data and sends this update to the central server. The server then aggregates the local updates to form a new global model. The aggregation is typically a weighted average:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\hbox{min} (w)\\sum\\limits_{{i=1}}^{N} {{p_i}{L_i}(w)}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{n}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the number of data points on client \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{w}_{i}\\left(t\\right)\\)\u003c/span\u003e\u003c/span\u003e is the model update sent by client \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e at iteration \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:t\\)\u003c/span\u003e\u003c/span\u003e, This aggregation step ensures that the global model is updated by considering the local learning from all participating clients\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.3 The proposed SK-MNV4\u003c/h2\u003e\u003cp\u003eThe proposed SK-MNV4 is an improvement upon MobileNetV4 [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], with its overall structure illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e\u003cb\u003e(a)\u003c/b\u003e. It primarily consists of UIB, Moblie_MQA, and SK blocks [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. The UIB, depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e\u003cb\u003e(b)\u003c/b\u003e, comprises four main components: Extra DW, Inverted Bottleneck (IB), ConvNext, and feed-forward network (FFN). The UIB not only serves as the fundamental structural unit of the SK-MNV4 backbone but also embodies a highly optimized architectural convergence of depthwise separability, residual learning, and adaptive feature transformation. By design, it addresses the limitations inherent in traditional convolutional backbones when applied to real-time industrial tasks, such as melt pool classification in L-DED, where lightweight computation and fine-grained spatial awareness are simultaneously required.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe IB, at the heart of UIB, leverages an initial 1\u0026times;1 pointwise convolution to project low-dimensional input tensors into a higher-dimensional latent space, enabling rich channel-wise representation learning. This operation is followed by a 3\u0026times;3 depthwise convolution, which performs spatial filtering independently on each channel, drastically reducing the number of required parameters compared to traditional convolutions while retaining essential local texture and shape information\u0026mdash;a crucial property for distinguishing fine-scale melt pool geometries. The IB concludes with a second 1\u0026times;1 pointwise convolution that compresses the expanded feature maps back to a compact representation, often with a residual shortcut if the input and output dimensions match. This residual connection is critical for stabilizing gradient propagation in deep networks and supports efficient training convergence.\u003c/p\u003e\u003cp\u003eIn contrast to IB, Extra DW convolution introduces a pre-expansion 1\u0026times;1 convolution prior to the standard IB expansion step. This module can be interpreted as a channel reweighting mechanism that primes the feature space before expansion. By applying early-stage channel mixing, the Extra DW path promotes feature diversity and enables richer interaction between channels prior to spatial filtering. Further extending this foundation, the ConvNext block replaces the conventional depthwise convolution between the expansion and projection layers with a streamlined residual feed-forward structure inspired by next-generation convolutional architectures. This architectural refinement preserves the expressiveness of channel-wise transformations while reducing sensitivity to kernel size selection, thereby enhancing the robustness and computational efficiency of feature learning within the SK-MNV4 backbone. The final subcomponent, the FFN, comprises two stacked 1\u0026times;1 convolutions with nonlinear activations in between. This sequence emulates the transformation logic found in Transformer-style architectures, where linear projections sandwiching a nonlinear function are used to refine channel-specific semantics. In the context of melt pool monitoring, FFN contributes to the discrimination of subtle thermal and morphological features. In the broader context of the SK-MNV4 architecture, UIB blocks serve as the primary conduit for hierarchical representation learning, enabling the model to transition smoothly from shallow, detail-rich features to abstract, class-specific embeddings. Their synergy with attention-enhancing modules (Mobile-MQA) and scale-aware fusion mechanisms (SK blocks) empowers the network to retain critical melt pool characteristics while discarding redundant or noisy background information. The result is a compact yet powerful encoder capable of supporting real-time, high-accuracy defect classification in federated and data-scarce settings.\u003c/p\u003e\u003cp\u003eMobile-MQA is introduced as an efficient attention mechanism tailored for real-time industrial applications such as melt pool image classification in L-DED. Traditional multi-head attention (MHA) mechanisms, while powerful in modeling long-range dependencies, incur significant computational overhead due to their reliance on high-resolution dot-product operations across all query-key-value interactions. To mitigate this, Mobile-MQA adopts a spatial reduction (SR) strategy for the key and value matrices, effectively reducing their spatial resolution prior to attention computation. This results in a substantial reduction in memory usage and inference latency, making the mechanism well-suited for deployment on edge devices. The formula for Mobile MQA [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] is presented as follows:\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$${\\text{Mobil}}{{\\text{e}}_M}_{{QA(X)}}={\\text{Concat}}({\\text{attentio}}{{\\text{n}}_1},...,{\\text{attentio}}{{\\text{n}}_n}){W^o}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe computation of each attention head, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(attentio{n_j}\\)\u003c/span\u003e\u003c/span\u003e, is further defined as:\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$${\\text{attentio}}{{\\text{n}}_j}={\\text{Softmax}}(\\frac{{(X{W^{Qj}}){{(SR(X){W^K})}^T}}}{{\\sqrt {{d_k}} }})(SR(X){W^V})$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eLet \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:X\\)\u003c/span\u003e\u003c/span\u003e be the input feature. The \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:n\\)\u003c/span\u003e\u003c/span\u003e attention heads compute different attention weights, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{attention}_{j}\\)\u003c/span\u003e\u003c/span\u003e. These attention outputs are then concatenated and mapped back to the input space dimensions through the weight matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}^{o}\\)\u003c/span\u003e\u003c/span\u003e.In the case of single-head attention, the operation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:SR\\left(X\\right)\\)\u003c/span\u003e\u003c/span\u003e is applied to reduce computational complexity, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:SR\\left(X\\right)\\)\u003c/span\u003e\u003c/span\u003e represents the downsampling of keys and values. The query, key, and value are linearly projected using the projection matrices \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}^{{Q}_{j}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}^{K}\\)\u003c/span\u003e\u003c/span\u003e, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}^{V}\\)\u003c/span\u003e\u003c/span\u003e, respectively. A scaling factor, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\sqrt{{d}_{k}}\\)\u003c/span\u003e\u003c/span\u003e, is applied, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{d}_{k}\\)\u003c/span\u003e\u003c/span\u003e is the dimension of the key vector, to stabilize the numerical range of the Softmax function. By down sampling the keys and values spatially, the computational cost is reduced, while the query retains high resolution, preserving more detailed information. This design effectively strikes a balance between performance and computational efficiency.\u003c/p\u003e\u003cp\u003eThe SK block (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e\u003cb\u003e(c)\u003c/b\u003e) is a feature extraction and fusion methodology aimed at enhancing the representational capacity of feature maps. Initially, two parallel convolutional branches with different kernel sizes (3\u0026times;3 and 5\u0026times;5) operate on the input feature map to extract features from varying receptive fields. The feature maps generated by these branches are then summed along the channel dimension, facilitating the integration of diverse information. Following this, an average pooling operation is applied to statistically capture spatial information across the channels, yielding a C\u0026times;1\u0026times;1 feature map that represents channel significance. This feature map is subsequently mapped to a lower dimension through a fully connected layer and then restored to the original channel count to obtain channel-specific weights. These weights are multiplied element-wise with their corresponding feature maps, leading to a weighted fusion that culminates in an output feature map with enhanced expressiveness. The entire process effectively leverages multi-branch convolutions, attention mechanisms, and weighted feature maps, resulting in substantial improvements in information richness and classification performance of the final output.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Results and discussions","content":"\u003cp\u003eIn this study, we compare the proposed SK-MNV4 model with several state-of-the-art network architectures, including EfficientNet [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], ConvNeXt [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e], MobileVIT [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], DenseNet [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e], and Swin Transformer (SwinT) [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. We utilize the pretrained models of these three materials for the classification of new materials, thereby validating their respective generalization performances.\u003c/p\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Pre-Training Performance of SK-MNV4 in Federated Learning\u003c/h2\u003e\u003cp\u003eAccuracy is a crucial performance metric that reflects the proportion of correctly predicted samples relative to the total number of samples. A higher accuracy indicates superior classification performance of the model. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, even without the application of transfer learning, the model exhibits a commendable ability to recognize new materials through the processes of FL. Among the evaluated models, SwinT demonstrated the lowest performance, with an accuracy of only 72.88%. Despite having a substantial number of learnable parameters, it is not well-suited for the classification of melt pool images addressed in this study. Other network models achieved classification accuracies exceeding 78%, signifying robust generalization capabilities. Notably, our proposed SK-MNV4 model achieved an accuracy of 84.77%, thereby validating its strong generalization performance.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe employed a confusion matrix to conduct a detailed analysis of the model's predictive outcomes. In this matrix, the horizontal axis represents the true labels, while the vertical axis denotes the predicted labels. We focused on analyzing the top four network models based on accuracy ranking. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e\u003cb\u003e(a)\u003c/b\u003e, MobileVIT demonstrates commendable performance in classification, particularly achieving an accuracy of 84.33% for class OM; however, its prediction accuracy for class LM is only 71.67%. Figure\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e\u003cb\u003e(b)\u003c/b\u003e shows that EfficientNet excels in identifying class OM, attaining an accuracy of 86.67%, although it tends to misclassify LM as NM. In Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e\u003cb\u003e(c)\u003c/b\u003e, ConvNeXt exhibits a high recognition rate for both NM and LM, but its performance in recognizing OM is relatively moderate. Our proposed model, SK-MNV4 (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e\u003cb\u003e(d)\u003c/b\u003e), performs exceptionally well, with accuracy rates exceeding 82% across all labels and a misclassification rate below 10%. Nevertheless, there remains a need for further improvement in accuracy through subsequent transfer learning efforts.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Transfer Learning Performance Evaluation\u003c/h2\u003e\u003cp\u003eAfter constructing the pre-trained network, a gap remained between the validation accuracy and the actual monitoring accuracy. To address this issue, we employed a dataset comprising 600 images representing three categories of IN718. We implemented transfer learning on the fully connected layers of the network model and performed a comparative experimental analysis.\u003c/p\u003e\u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\u003ch2\u003e5.2.1 Classification Accuracy and Model Parameters\u003c/h2\u003e\u003cp\u003eAfter transfer learning, we compiled statistics on accuracy and the number of parameters, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. We found that SwinT achieved an accuracy of only 85.8% post-transfer learning, indicating its unsuitability for the melt pool image recognition task due to relatively low identification accuracy. DenseNet recorded a recognition accuracy of 90.5%, demonstrating a certain level of monitoring capability. Meanwhile, MobileVIT maintained a commendable accuracy of 92.1% while utilizing the fewest parameters. EfficientNet and ConvNeXt achieved higher accuracies of 92.3% and 95.8%, respectively. Notably, our proposed SK-MNV4 obtained the highest accuracy at 97.3%, showcasing its ability for high-precision melt pool identification following small sample transfer learning.\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\u003eHighest accuracy and number of parameters in different models\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=\"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\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\u003eHighest accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eParameter (M)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSK-MNV4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.973\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eConvNeXt\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.958\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e27.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEfficientNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.923\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e6.5\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMobileVIT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.921\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDenseNET\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e6.9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSwinT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.848\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.8\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=\"Sec19\" class=\"Section3\"\u003e\u003ch2\u003e5.2.2 Validation times\u003c/h2\u003e\u003cp\u003eDue to the high requirements for real-time monitoring of the melt pool, the speed at which models process images is crucial. The computational times of various models for processing 900 melt pool images are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. It is evident that the SwinT model, with a large number of parameters, exhibits a significantly longer validation time of 4.03 seconds compared to the other models. Notably, although MobileVIT has the fewest parameters at just 0.9M, its validation time of 0.969 seconds is not the shortest. This can be attributed to the high computational complexity of the self-attention mechanism within MobileVIT, which necessitates greater computational resources and time for processing each layer. In contrast, the proposed SK-MNV4 demonstrates the shortest processing time of only 0.918 seconds, indicating its superior capability for real-time identification of melt pool images.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec20\" class=\"Section3\"\u003e\u003ch2\u003e5.2.3 Data visualization\u003c/h2\u003e\u003cp\u003eT-SNE is employed to embed high-dimensional data into a lower-dimensional space for visualizing classification outcomes. We selected the top four models based on accuracy for analysis. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e\u003cb\u003e(a)\u003c/b\u003e, MobileVIT can approximately differentiate between three distinct categories; however, misclassification is prevalent at the boundaries between these categories. Figure\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e\u003cb\u003e(b)\u003c/b\u003e shows that EfficientNet effectively separates OM from the other categories, although a significant number of LM instances are incorrectly identified as NM. Figure\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e\u003cb\u003e(c)\u003c/b\u003e illustrates that ConvNeXt excels at distinguishing LM from other categories, yet there is still some misidentification of OM and NM. Finally, the proposed SK-MNV4 model, as depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e\u003cb\u003e(d)\u003c/b\u003e, achieves effective separation among all categories with considerable distance between them, indicating superior classification performance and a reduced rate of misclassification compared to the other network models.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section3\"\u003e\u003ch2\u003e5.2.4 Grad-CAM\u003c/h2\u003e\u003cp\u003eGrad-CAM is a technique used in deep learning for visualizing the areas of an image that are most important for classification. It works by computing the gradient of the score for a target class concerning the feature maps of CNN, and then multiplying each feature map by its corresponding gradient before summing them up. The resulting heatmap highlights the regions of the input image that were most relevant for the CNN's decision.\u003c/p\u003e\u003cp\u003eIn the context of classifying melt pool images in L-DED to generate thermal maps, Grad-CAM can be utilized to identify the image regions that are most relevant for a deep learning model's classification decision. This can aid researchers and engineers in gaining a better understanding of the decision-making process of the model. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(a) to (c)\u003c/b\u003e displays original melt pool images for three different cases: NM, OM, and LM. In Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(d) to (f)\u003c/b\u003e, corresponding heatmaps generated by Grad-CAM are presented. For the NM case, the melt pool exhibits a well-defined circular boundary, aligning with the expected ideal formation quality. From Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(d)\u003c/b\u003e, it is evident that the network model bases its decision on the distinct circular boundary of the melt pool. In the OM case, which often occurs under high laser intensity conditions, resulting in irregular melt pool boundaries and the generation of plumes [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e], the network model relies on the presence of plumes near the melt pool for its decision-making, as observed in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(e)\u003c/b\u003e.In the LM case, typically associated with lower laser intensity, Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(f)\u003c/b\u003e reveals that the network model makes decisions based on the darker regions within the melt pool. Additionally, from Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e\u003cb\u003e(d)\u003c/b\u003e, it is observed that despite the presence of minimal spatter interference near the melt pool, our developed network model, SK-MNV4, does not prioritize such interference, highlighting the model's robustness.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"6. Conclusions","content":"\u003cp\u003eIn this study, we propose a novel federated transfer learning framework tailored for defect detection in L-DED, addressing critical challenges such as small sample size, data privacy, and heterogeneous material domains. At the core of the framework lies SK-MNV4, a lightweight yet expressive backbone network designed to balance classification accuracy with computational efficiency. The architecture integrates UIB modules, Mobile-MQA attention mechanisms, and SK blocks to enable efficient multi-scale feature extraction, global context modeling, and adaptive feature fusion.\u003c/p\u003e\u003cp\u003eTo preserve industrial data privacy while enhancing model generalization, we adopt a FL strategy across decentralized datasets representing different alloy systems. This enables collaborative model pretraining without sharing raw melt pool images. Subsequent transfer learning is performed using limited target domain data to adapt the pre-trained model to new materials. Experimental evaluations across multiple benchmark networks\u0026mdash;including EfficientNet, ConvNeXt, DenseNet, Swin Transformer, and MobileViT\u0026mdash;demonstrate that SK-MNV4 achieves the highest classification accuracy (97.3%) following small-sample fine-tuning, with the shortest inference time (0.918 s), indicating strong potential for real-time deployment. Additionally, ablation analyses using confusion matrices, t-SNE embeddings, and Grad-CAM visualizations confirm the model\u0026rsquo;s superior discriminative capability and robust interpretability. These findings validate the efficacy of combining FL and transfer learning with a lightweight, modular architecture to realize accurate and scalable melt pool monitoring under data-constrained and privacy-sensitive conditions.\u003c/p\u003e\u003cp\u003eHowever, there are still several limitations to this work that warrant further attention. The current framework primarily focuses on single-modal image data, which may not fully capture the thermal or structural anomalies that manifest in other sensing domains. The reliance on melt pool images alone could potentially limit the model's responsiveness to subtle defect patterns, especially under varying lighting or occlusion conditions. In addition, while the FL setting successfully addresses data privacy concerns, the framework presumes stable client participation and consistent network conditions during training. In practical industrial environments, system failures, asynchronous updates, or client-side heterogeneity may introduce instability that could impact overall convergence and model performance. Future research will consider the integration of multi-modal sensor inputs, combining acoustic, thermal, and visual signals within a unified architecture to enrich the representation space and improve defect localization under complex process conditions. Expanding the framework toward federated personalization, where models are dynamically adapted to individual client distributions, may also enhance robustness in non-independent and identically distributed scenarios. Furthermore, embedding continual learning capabilities into the SK-MNV4 framework would enable adaptive model updates over time without retraining from scratch, supporting long-term deployment across evolving materials and process settings. These extensions are expected to further reinforce the scalability and practical utility of the proposed system in real-world additive manufacturing applications.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eDeclaration of Competing Interest\u003c/h2\u003e\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eF.W. implemented the federated learning framework and carried out the main model training and evaluation. Q.L. performed data preprocessing, ablation studies, and comparative experiments. X.Z. and W.L. assisted in data labeling and image acquisition from the L-DED platform. Z.L. contributed to software validation and visualization. J.G. helped with experimental setup and parameter tuning. S.P. supervised the entire project, provided critical revision of the manuscript, and secured funding. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eWe would like to thank the following individuals for their helpful discussions and suggestions during the course of this research:Zhinong Li, Nanchang Hangkong UniversityJiawei Xiang, Wenzhou UniversityGuangyi Ma, Dalian University of TechnologyTheir insights contributed to the refinement of the methodology and validation of the experimental design.\u003c/p\u003e\u003ch2\u003eData availability\u003c/h2\u003e\u003cp\u003eThe codes of the algorithms involved this study have been uploaded to GitHub (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/shi-tong/SK-MNV4.git\u003c/span\u003e\u003cspan address=\"https://github.com/shi-tong/SK-MNV4.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMohanavel V, Ashraff Ali K S, Ranganathan K, Allen Jeffrey J, Ravikumar M M and Rajkumar S (2021) The roles and applications of additive manufacturing in the aerospace and automobile sector Materials Today: Proceedings 47 405\u0026ndash;9\u003c/li\u003e\n\u003cli\u003eLi H, Gao F, Jiao J, Liu Z, Ji D and Lin J (2023) Acoustic Emission-Based Cross-Domain Process Health Monitoring for Additive Manufacturing IEEE Transactions on Instrumentation and Measurement 72 1\u0026ndash;8\u003c/li\u003e\n\u003cli\u003eWang H, Liu W, Tang Z, Wang Y, Mei X, Saleheen K M, Wang Z and Zhang H (2020) Review on adaptive control of laser-directed energy deposition OE 59 070901\u003c/li\u003e\n\u003cli\u003eKumar R, Kumar M and Chohan J S (2021) The role of additive manufacturing for biomedical applications: A critical review Journal of Manufacturing Processes 64 828\u0026ndash;50\u003c/li\u003e\n\u003cli\u003eLiu T-S, Chen P, Qiu F, Yang H-Y, Jin N T Y, Chew Y, Wang D, Li R, Jiang Q-C and Tan C (2024) Review on laser directed energy deposited aluminum alloys Int. J. Extrem. Manuf. 6 022004\u003c/li\u003e\n\u003cli\u003ePiscopo G and Iuliano L (2022) Current research and industrial application of laser powder directed energy deposition Int J Adv Manuf Technol 119 6893\u0026ndash;917\u003c/li\u003e\n\u003cli\u003eWang C, Tan X P, Tor S B and Lim C S (2020) Machine learning in additive manufacturing: State-of-the-art and perspectives Additive Manufacturing 36 101538\u003c/li\u003e\n\u003cli\u003eDilip J J S, Zhang S, Teng C, Zeng K, Robinson C, Pal D and Stucker B (2017) Influence of processing parameters on the evolution of melt pool, porosity, and microstructures in Ti-6Al-4V alloy parts fabricated by selective laser melting Prog Addit Manuf 2 157\u0026ndash;67\u003c/li\u003e\n\u003cli\u003eBaumgartl H, Tomas J, Buettner R and Merkel M (2020) A deep learning-based model for defect detection in laser-powder bed fusion using in-situ thermographic monitoring Prog Addit Manuf 5 277\u0026ndash;85\u003c/li\u003e\n\u003cli\u003eChen L, Yao X, Tan C, He W, Su J, Weng F, Chew Y, Ng N P H and Moon S K (2023) In-situ crack and keyhole pore detection in laser directed energy deposition through acoustic signal and deep learning Additive Manufacturing 69 103547\u003c/li\u003e\n\u003cli\u003eRen W, Wen G, Zhang Z and Mazumder J (2022) Quality monitoring in additive manufacturing using emission spectroscopy and unsupervised deep learning Materials and Manufacturing Processes 37 1339\u0026ndash;46\u003c/li\u003e\n\u003cli\u003eShin S-J, Lee J-H, Jadhav S and Kim D B (2024) Material-Adaptive Anomaly Detection Using Property-Concatenated Transfer Learning in Wire Arc Additive Manufacturing Int. J. Precis. Eng. Manuf. 25 383\u0026ndash;408\u003c/li\u003e\n\u003cli\u003eDeng T, Li Y, Liu X and Wang L (2023) FL-based collaborative manufacturing for complex parts J Intell Manuf 34 3025\u0026ndash;38\u003c/li\u003e\n\u003cli\u003eTang Y, Rahmani Dehaghani M and Wang G G (2023) Review of transfer learning in modeling additive manufacturing processes Additive Manufacturing 61 103357\u003c/li\u003e\n\u003cli\u003eTang Y, Rahmani Dehaghani M, Sajadi P and Wang G G (2024) Selecting subsets of source data for transfer learning with applications in metal additive manufacturing J Intell Manuf\u003c/li\u003e\n\u003cli\u003ePeng S, Yang S, Gao B, Liu W, Wang F and Tang Z (2025) Prediction of 3D temperature field through single 2D temperature data based on transfer learning-based PINN model in laser-based directed energy deposition Journal of Manufacturing Processes 138 140\u0026ndash;56\u003c/li\u003e\n\u003cli\u003eQi P, Chiaro D, Guzzo A, Ianni M, Fortino G and Piccialli F (2024) Model aggregation techniques in FL: A comprehensive survey Future Generation Computer Systems 150 272\u0026ndash;93\u003c/li\u003e\n\u003cli\u003eWu Y, Zhang S, Yu W, Liu Y, Gu Q, Zhou D, Chen H and Cheng W (2023) Personalized FL under Mixture of Distributions Proceedings of the 40th International Conference on Machine Learning International Conference on Machine Learning (PMLR) pp 37860\u0026ndash;79\u003c/li\u003e\n\u003cli\u003eGuan H, Yap P-T, Bozoki A and Liu M (2024) FL for medical image analysis: A survey Pattern Recognition 151 110424\u003c/li\u003e\n\u003cli\u003eMehta M and Shao C (2022) FL-based semantic segmentation for pixel-wise defect detection in additive manufacturing Journal of Manufacturing Systems 64 197\u0026ndash;210\u003c/li\u003e\n\u003cli\u003eLi Z, Li Z and Gu F (2024) Intelligent diagnosis method for machine faults based on federated transfer learning Applied Soft Computing 163 111922\u003c/li\u003e\n\u003cli\u003eZhu X, Jiang F, Guo C, Xu D, Wang Z and Jiang G (2023) Surface morphology inspection for directed energy deposition using small dataset with transfer learning Journal of Manufacturing Processes 93 101\u0026ndash;15\u003c/li\u003e\n\u003cli\u003ePetrich J, Snow Z, Corbin D and Reutzel E W (2021) Multi-modal sensor fusion with machine learning for data-driven process monitoring for additive manufacturing Additive Manufacturing 48 102364\u003c/li\u003e\n\u003cli\u003eWang R, Garcia D, Kamath R R, Dou C, Ma X, Shen B, Choo H, Fezzaa K, Yu H Z and Kong Z (James) (2022) In situ melt pool measurements for laser powder bed fusion using multi sensing and correlation analysis Sci Rep 12 13716\u003c/li\u003e\n\u003cli\u003eBisheh M N, Wang X, Chang S I, Lei S and Ma J (2023) Image-based characterization of laser scribing quality using transfer learning J Intell Manuf 34 2307\u0026ndash;19\u003c/li\u003e\n\u003cli\u003eAnon (2025) Personalized federated unsupervised learning for nozzle condition monitoring using vibration sensors in additive manufacturing Robotics and Computer-Integrated Manufacturing 93 102940\u003c/li\u003e\n\u003cli\u003eMehta M, Bimrose M V, McGregor D J, King W P and Shao C (2024) FL enables privacy-preserving and data-efficient dimension prediction and part qualification across additive manufacturing factories Journal of Manufacturing Systems 74 752\u0026ndash;61\u003c/li\u003e\n\u003cli\u003eYang Z, Lu Y, Yeung H and Krishnamurty S (2019) Investigation of Deep Learning for Real-Time Melt Pool Classification in Additive Manufacturing 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) pp 640\u0026ndash;7\u003c/li\u003e\n\u003cli\u003eJiang H-Z, Li Z-Y, Feng T, Wu P-Y, Chen Q-S, Feng Y-L, Chen L-F, Hou J-Y and Xu H-J (2021) Effect of Process Parameters on Defects, Melt Pool Shape, Microstructure, and Tensile Behavior of 316L Stainless Steel Produced by Selective Laser Melting Acta Metall. Sin. (Engl. Lett.) 34 495\u0026ndash;510\u003c/li\u003e\n\u003cli\u003eScime L and Beuth J (2019) Melt pool geometry and morphology variability for the Inconel 718 alloy in a laser powder bed fusion additive manufacturing process Additive Manufacturing 29 100830\u003c/li\u003e\n\u003cli\u003eErtay D S, Naiel M A, Vlasea M and Fieguth P (2021) Process performance evaluation and classification via in-situ melt pool monitoring in directed energy deposition CIRP Journal of Manufacturing Science and Technology 35 298\u0026ndash;314\u003c/li\u003e\n\u003cli\u003eLi J, Zhou Q, Cao L, Wang Y and Hu J (2022) A convolutional neural network-based multi-sensor fusion approach for in-situ quality monitoring of selective laser melting Journal of Manufacturing Systems 64 429\u0026ndash;42\u003c/li\u003e\n\u003cli\u003eTang Y, Rahmani Dehaghani M, Sajadi P and Wang G G (2024) Selecting subsets of source data for transfer learning with applications in metal additive manufacturing J Intell Manuf\u003c/li\u003e\n\u003cli\u003eMenon N, Mondal S and Basak A (2022) Transferring Process Mapping Knowledge across SS316L and IN718 in Laser Directed Energy Deposition Using Machine Learning\u003c/li\u003e\n\u003cli\u003eMochi V H, N\u0026uacute;\u0026ntilde;ez H H L, Ribeiro K S B and Venter G S (2023) Real-time prediction of deposited bead width in L-DED using semi-supervised transfer learning Int J Adv Manuf Technol 129 5643\u0026ndash;54\u003c/li\u003e\n\u003cli\u003eFeenstra D R, Molotnikov A and Birbilis N (2021) Utilisation of artificial neural networks to rationalise processing windows in directed energy deposition applications Materials \u0026amp; Design 198 109342\u003c/li\u003e\n\u003cli\u003eAkbari P, Ogoke F, Kao N-Y, Meidani K, Yeh C-Y, Lee W and Barati Farimani A (2022) MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning Additive Manufacturing 55 102817\u003c/li\u003e\n\u003cli\u003eHo S, Zhang W, Young W, Buchholz M, Jufout S A, Dajani K, Bian L and Mozumdar M (2021) DLAM: Deep Learning Based Real-Time Porosity Prediction for Additive Manufacturing Using Thermal Images of the Melt Pool IEEE Access 9 115100\u0026ndash;14\u003c/li\u003e\n\u003cli\u003eQin D, Leichner C, Delakis M, Fornoni M, Luo S, Yang F, Wang W, Banbury C, Ye C, Akin B, Aggarwal V, Zhu T, Moro D and Howard A (2024) MobileNetV4 -- Universal Models for the Mobile Ecosystem\u003c/li\u003e\n\u003cli\u003eLiu G SK-MobileNet: A Lightweight Adaptive Network Based on Complex Deep Transfer Learning for Plant Disease Recognition Arabian Journal for Science and Engineering\u003c/li\u003e\n\u003cli\u003eTan M and Le Q V (2020) EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks\u003c/li\u003e\n\u003cli\u003eLiu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T and Xie S (2022) A ConvNet for the 2020s\u003c/li\u003e\n\u003cli\u003eMehta S and Rastegari M (2022) MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer\u003c/li\u003e\n\u003cli\u003eHuang G, Liu Z, Maaten L van der and Weinberger K Q (2018) Densely Connected Convolutional Networks\u003c/li\u003e\n\u003cli\u003eLiu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S and Guo B (2021) Swin Transformer: Hierarchical Vision Transformer using Shifted Windows\u003c/li\u003e\n\u003cli\u003eCoeck S, Bisht M, Plas J and Verbist F (2019) Prediction of lack of fusion porosity in selective laser melting based on melt pool monitoring data Additive Manufacturing 25 347\u0026ndash;56\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"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":"Directed Energy Deposition, Federated Transfer Learning, Melt Pool Monitoring, Data Privacy","lastPublishedDoi":"10.21203/rs.3.rs-7002574/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7002574/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLaser Directed Energy Deposition (L-DED) faces key challenges such as unstable printing processes, low forming quality, and poor interlayer consistency. An urgent need is to establish real-time monitoring and process control mechanisms to improve its industrial applicability. However, in practical applications, the construction of monitoring systems is constrained by two major bottlenecks. The scarcity of high-quality defect samples limits the generalization capability of deep learning models. Additionally, due to confidentiality concerns regarding printing paths and product defects, users generally refuse to share raw data, further hindering centralized model training and large-scale deployment. To address these challenges, this study proposes a lightweight neural network integrated into a federated transfer learning framework, termed SK-MNV4. The network incorporates a unified inverted bottleneck module, a lightweight multi-head attention mechanism, and a selective kernel fusion strategy to enable efficient multi-scale feature extraction and robust defect representation, significantly enhancing performance under data privacy and limited-sample conditions. The proposed federated transfer learning scheme treats independently collected melt pool image datasets from three representative alloys (IN738, TB6, Ti-10Mo) in L-DED experiments as heterogeneous clients for federated learning. It ensures local data processing and prevents raw melt pool image exposure, while enabling cross-material transfer learning with IN718 melt pool images as the target domain, thereby improving the model's generalization ability. We used T-SNE to compare and visualize each model. We compare our model with mainstream models (ConvNeXt, DenseNet, EfficientNet, and Swin Transformer) in terms of accuracy and inference efficiency. Comparative experiments show that SK-MNV4 outperforms these models and shows 97.3% accuracy and 0.001 seconds inference efficiency on average, demonstrating superior potential for applications in scenarios with strong real-time requirements, data sensitivity, and sparse defect distributions.\u003c/p\u003e","manuscriptTitle":"Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-15 12:47:26","doi":"10.21203/rs.3.rs-7002574/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":"a45e1c14-da39-46da-9fc6-52b7d3fb5e77","owner":[],"postedDate":"July 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-16T11:54:11+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-15 12:47:26","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7002574","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7002574","identity":"rs-7002574","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}