Accurate and real-time prediction of laser cladding morphology is critical for quality control in directed energy deposition processes. This study proposes a physics-informed neural network (PINN) for predicting the cross-sectional profiles of laser cladding by three input parameters: laser power, scanning speed, and powder flowrate. The advantage of our method is that PINN embeds a learnable empirical model. The coefficients of the model are collaboratively optimized with the network weights via physics-consistency loss and data loss. Experimental results demonstrate that the proposed PINN outperforms purely data-driven and fixed-parameter empirical models, especially under conditions of limited training data. The trained model is successfully deployed in a digital twin-based laser cladding system. Thus, without the need for online geometric sensing, the system can achieve real-time morphology prediction and three-dimensional geometric structure reconstruction solely through process parameters. The proposed method provides an efficient and robust predictive model for digital twin-enabled laser cladding applications.
Quality control in polyester fiber polymerization is a multi-objective process-parameter optimization problem complicated by nonlinear parameter-quality coupling and the limited reuse of abnormal-condition knowledge. Existing surrogate-assisted optimization methods typically search within fixed engineering boundaries and lack a mechanism for converting unstructured process knowledge into computable constraints. To address this limitation, this study develops PLMC as a LightRAG empowered multi-objective optimization framework that integrates knowledge-to-constraint transformation, surrogate-based objective evaluation, and knowledge-constrained dual-population co-evolution. Using a polyester polymerization process knowledge graph, PLMC retrieves mechanistic causes and control measures. A domain-adapted large language model then uses this information to generate condition-specific prior knowledge constraints. These constraints are embedded into PLMC-DP, a knowledge-constrained dual-population co-evolution module. The global population explores the engineering feasible space, whereas the knowledge-guided population searches within the prior-constraint region. Candidate solutions are evaluated using a surrogate error compensation-enhanced Kolmogorov–Arnold Network (SEC-KAN). SEC-KAN predicts the effects of key process parameters on intrinsic viscosity, diethylene glycol content, and terminal carboxyl group concentration while correcting local surrogate bias using historical residuals. Under the tested settings, experiments on industrial production-line data show that the complete PLMC framework achieves better Pareto-set quality and convergence stability than the baseline algorithms. The resulting hypervolume (HV) and inverted generational distance (IGD) are 0.8387 and 0.0473, respectively. Stable process-control performance across fully drawn yarn (FDY), partially oriented yarn (POY), and draw textured yarn (DTY) product scenarios indicates the applicability of PLMC to diverse polyester products.
The uniform grinding of epoxy resin coatings on aircraft skins is a key challenge in aviation maintenance; traditional manual grinding is inefficient and produces inconsistent results. This study addresses the problem of uneven material removal caused by sandpaper wear during the grinding of epoxy resin coatings by proposing an optimization method that combines the Preston equation with the XGBoost machine learning algorithm. By establishing a theoretical model relating the Preston coefficient kp to process parameters (feed rate, rotational speed, and grinding pressure) and using it to guide feature engineering, and by combining it with 16 sets of orthogonal experimental data, an XGBoost model was trained to predict the decay trend of kp along the grinding path. Furthermore, a binary search method was employed to optimize the robot’s feed rate, thereby achieving dynamic compensation during the coating grinding process. Experimental results demonstrate that the optimized feed rate strategy significantly improves the uniformity of coating grinding, with the maximum error in coating grinding thickness reduced by 51, 38.47, 62.5, and 46.67
Aluminum extrusion is a critical plastic forming technology for manufacturing high-performance aluminum alloy profiles, with products widely applied in industries such as aerospace, transportation, and construction. In the aluminum alloy extrusion process, the precise setting of process parameters has a decisive influence on product quality and production efficiency. However, this process is characterized by high dimensionality, non-linearity, and time-varying properties, which traditional methods struggle to address. To address this problem, this paper proposes a prediction framework integrating a stability-aware multi-view feature selection method (SAM-FS) and a DDPG-driven adaptive stacking model (DA-Stacking). SAM-FS constructs a comprehensive evaluation method based on the entropy weight method (EWM) by introducing a stability bias term. By integrating results from statistical tests, information theory, and embedded modeling, it effectively screens key features that possess both discriminative power and robustness. DA-Stacking, inspired by the mixture of experts architecture, integrates multiple heterogeneous base learners as specialized experts and utilizes a deep deterministic policy gradient (DDPG) agent combined with an attention mechanism to achieve adaptive fusion weight allocation for heterogeneous base learners according to real-time operating conditions. Evaluated on a real-world production dataset from an aluminum profile manufacturing enterprise in China, the proposed framework achieves enhanced predictive performance, demonstrating competitive advantages over baseline methods, such as TabPFN and AutoGluon, by reducing the MAE by 7.2 R^2 value.
This research concerns the characterization and detection of process instabilities of the wire arc directed energy deposition (WA-DED) additive manufacturing process, popularly known as wire arc additive manufacturing (WAAM), through a multi-sensor approach resorting to an adaptive signal segmentation and a physics-informed feature extraction methodology. The objective is to detect two main types of instabilities in WAAM, namely, wire stubbing and droplet overgrowth through monitoring and machine learning of multi-sensor data streams. To realize this objective, current, voltage, acoustic signatures, and high-speed melt pool imaging data were acquired during processing. Six physically intuitive features were extracted from each of the three fundamental phases of the electric arc, i.e., arcing, short circuit and arc ignition. These features were employed as inputs to a computationally tractable, support vector machine model trained to differentiate between stable and unstable processing states. The approach predicted the onset of process instability in thin walls with statistical accuracy exceeding 95
Additive Manufacturing is an upcoming technology to produce metal structures in industry as complex near net-shaped structures can be built. One commonly used method is Laser Powder Bed Fusion (PBF-LB), that uses lasers to melt metal powder layer by layer. Alongside advantages that come with this technology, the variety of adjustable process parameters is challenging for optimizing the process due to complex correlations of those. Additionally, the resulting mechanical properties can be challenging to optimize empirically based on the vast multi-dimensional parameter space. To solve this problem, several Machine Learning (ML) approaches have been used in the scope of different applications. With the aim of predicting tensile properties of PBF-LB parts, printed with AlSi10Mg0.5, three ML models (Linear Regression, XGBoost, and Multi-Layer Perceptrons (MLP)) were trained as black box models on a newly created dataset. To include as much process information as possible in the training data, an additional mathematically based dimensionless metrics model was used. The model gained insights into the thermodynamic conditions of certain parameter sets, which enabled a more even distribution of different processing conditions in the dataset. This equation-based white box model, combined with the subsequent ML-based black-box models formed the grey-box modeling approach. Due to the small size of the dataset of only 50 data vectors, strong overfitting was observed. This could be minimized for the case of the MLP by means of a customized architecture and strong regularization. It was shown, that the dataset enabled sufficient predictions by the MLP model of yield and tensile strength with R^2_test values of up to 88.3 % and 86.2 % , respectively. Elongation at break turned out to be more challenging to predict with an R^2_test of up to 75.6 % . This highlighted the benefit of combining the phenomenological melt mode model for data creation and a ML model for sufficient prediction of mechanical properties in a resource efficient manner.
Fine-grained insulator anomaly detection remains challenging because abnormal regions are often small, weakly contrasted, morphologically diverse, and embedded in repetitive structures or cluttered outdoor backgrounds. Conventional RGB-based detectors may fail to preserve subtle structural discontinuities, while direct feature fusion can introduce redundant or conflicting information. To address these issues, this study proposes DBRM-Net, a dual-branch residual framework that jointly processes RGB images and an RGB-derived Surface Structural Modality (SSM). Rather than introducing a new sensing modality or new low-level operators, SSM organizes complementary luminance, gradient, directional, Laplacian, roughness, and local-residual descriptors into a strictly aligned six-channel structural representation. This task-driven representation provides supplementary evidence for weak boundaries, local roughness variations, and subtle surface disturbances without requiring additional sensing hardware or cross-sensor registration. The Prior-Aware Mixture Fusion (PAMF) module models cross-branch agreement, discrepancy, and residual priors to improve forward feature integration, whereas the Adaptive Contextual Topology Injection Fusion (ACTIF) module selectively reuses anomaly-relevant structural evidence while preserving RGB semantic continuity. Experiments on two insulator anomaly datasets show that DBRM-Net achieves AP_50 and AP_50:95 values of 88.7 and 65.4, respectively, on the primary IFD dataset. Compared with the strongest SSM-based baseline, the improvement is modest but consistent, while DBRM-Net uses fewer parameters, lower computational cost, and higher detector-side inference speed. Repeated-run, ablation, robustness, and visualization analyses further demonstrate stable optimization and complementary contributions from the structural representation and staged interaction modules. Overall, DBRM-Net provides a favorable accuracy-efficiency trade-off for fine-grained insulator anomaly detection under complex imaging conditions.
Accurate forecasting of intermittent demand remains a persistent challenge for deep learning methods, partly due to the inadequacy of conventional error metrics in capturing zero-demand forecasting performance. This study tackles this gap by introducing Z
One of the most significant challenges to broad adoption of large-format additive manufacturing processes is the occurrence of defects within the manufactured part. Image-based anomaly detection has shown promise in detecting defects such as gaps, mini cracks, delamination, and more, yet improvements are needed to make the approach accessible to inexperienced users. In order to distinguish between normal and faulty data, it is essential to find an adequate model structure through properly tuned parameters. In this article, we focus on the use and tuning of principal component analysis (PCA) models for automatic detection of deviations from typical printer behavior with low-cost thermal cameras. In this case, the number of principal components is a hyper-parameter that needs to be tuned to capture important variation while rejecting noise. We propose a new method to do so that does not depend on strong assumptions of normality but makes use of an efficient row-wise cross-validation scheme. We test the algorithm on multiple data sets and compare to a previously established cross-validation method. We show that dimension reduction methods, like PCA, are useful for anomaly detection in additive manufacturing and that robust model tuning is feasible in absence of distributional assumptions
Few-shot defect detection holds significant importance for adapting to complex industrial environments and enhancing detection accuracy. Addressing the issue where existing few-shot defect detection methods are prone to compromised feature representation under varying defect scales, this study proposes a distance-guided prototype network (DGPN) with explicit meta learning. Building upon the prototype network, our method leverages distance information between normal and anomalous image features to guide feature transformation and fusion processes. An attention gating block (AGB) is introduced to convert distance representation vectors into feature maps while enhancing the capture capability for fine-grained targets. A multiscale feature fusion module (MSFF) is proposed to further strengthen the network’s ability to extract multiscale features and semantic information. Additionally, an upsample fusion module (UF) is designed to fully integrate and exchange multiscale contextual information between shallow and deep layers, decoding semantic information and spatial details to improve the precision of small defect detection. Extensive experiments on Industrial-5i, Visa, and a self-built chip dataset validate the superiority of the proposed method and the practical applicability.
Additive manufacturing (AM) imparts machine-specific fingerprints into the surface texture of printed parts, which can be used to identify the machine or factory of origin. Deep learning methods can detect these fingerprints even when they are not detectable by humans; however, these methods suffer from poor data efficiency and have not been shown to generalize to diverse camera views and other practical imaging conditions. This study develops a novel deep learning network referred to as a Learned Region Fingerprint Model (LRFM) that combines a Differentiable Patch Selection (DPS) module that identifies key textural features, a Fingerprinting (FP) module that extracts latent identifiable features from the image patches, and a consolidation network that aggregates extracted features and makes source predictions. To develop the LRFM, we designed and produced a total of 1,620 AM parts with nine different designs from six contract manufacturers. Data collection is enabled by a unique custom robotic imaging system which photographed each part from 132 unique view angles, creating a dataset of 213,840 images with highly varied appearance due to diverse lighting and shadows. The LRFM predicts the manufacturing source with 98
High-fidelity simulations of laser welding capture complex thermo-fluid phenomena, including phase change, free-surface deformation, and keyhole dynamics, but their computational cost limits large-scale process exploration and real-time use. In this work, we present the Laser Processing Fourier Neural Operator (LP-FNO), a Fourier Neural Operator (FNO)-based surrogate model that learns the parametric solution operator of laser processing from multiphysics simulations generated with FLOW-3D WELD®. Through a reformulation of the transient problem in a reference frame moving with the laser and the application of temporal averaging, the system is recast into a quasi-steady setting suitable for operator learning, even in the stable keyhole welding regime. The proposed LP-FNO maps process parameters to three-dimensional temperature fields and melt-pool boundaries across a broad process window spanning conduction and keyhole regimes using the non-dimensional normalized enthalpy formulation. The model achieves an average temperature relative error of approximately 2.5% and intersection-over-union scores for melt-pool segmentation above 0.9. Compared with a fully connected coordinate network, U-Net, and DeepONet, LP-FNO provides the strongest overall accuracy–efficiency trade-off for the present task, while retaining the resolution-invariant evaluation capability characteristic of FNOs. We demonstrate that an LP-FNO model trained on coarse-resolution data can be evaluated on finer grids, yielding accurate super-resolved predictions in mesh-converged conduction regimes, whereas discrepancies in keyhole regimes reflect unresolved dynamics in the coarse-mesh training data. These results indicate that LP-FNO provides an efficient surrogate modeling framework for laser welding, enabling prediction of full three-dimensional fields and phase interfaces over wide parameter ranges in just tens of milliseconds, up to a hundred thousand times faster than traditional finite-volume multiphysics simulations.
Automated defect inspection in smart manufacturing requires both high accuracy and real-time performance, yet existing approaches often prioritize either rapid localization or precise classification. A hybrid framework combining You Only Look Once (YOLO) with a Vision Transformer (ViT) is proposed for automated defect detection in industrial production environments. YOLOv11 performs rapid defect localization while a ViT classifier provides fine-grained discrimination between defective and non-defective products. A real-world dataset of 8,054 images was collected from a plastic fork manufacturing line under practical production conditions, comprising 6,157 non-faulty and 1897 faulty samples. Several YOLO variants and multiple classification backbones were systematically evaluated to identify the most suitable components for the proposed pipeline. Experimental results showed that pretrained YOLOv11m achieved the best detection performance, reaching a mean average precision (mAP @ 0.5) of 0.8263, while the optimized ViT-Base classifier with a modified classification head achieved an accuracy of 0.9703 and an F1-score of 0.9373. When integrated into the proposed hybrid framework, the system attained an image-level accuracy of 0.9574, precision of 0.9759, recall of 0.9679, and an F1-score of 0.9719 on the test set. The complete framework evaluated the 1,209 unique test images across 2,417 total operational instances in 47.394 s, corresponding to an average inference time of 19.6 ms per instance, demonstrating its suitability for real-time industrial deployment. Combining YOLO-based region localization with transformer-based classification improves defect inspection performance by reducing background noise and enabling more reliable decision-making, providing a practical solution for automated quality control in Industry 4.0 manufacturing systems.
Identifying surface flaws on metals is a crucial component of quality assurance within smart manufacturing. Current detection approaches still face difficulties in simultaneously ensuring high accuracy and fast processing speed under complex industrial conditions. These scenarios involve background interference, varying scales, and low-contrast features. A novel high-precision real-time detection framework, DPSNet, is proposed to address these challenges. Firstly, a polarized enhancement adaptive module is designed. Through a polarized attention mechanism and frequency domain filtering strategies, background noise is effectively suppressed, and blurred defect features are sharpened. To accommodate variations in defect size and geometry, DIKM dynamically combines multiple convolutional branches with different receptive-field configurations. Furthermore, a geometric-aware collaborative loss is constructed. By introducing inner bounding box scaling and shape consistency constraints, the regression precision and convergence stability of the bounding boxes are significantly improved. Evaluations on the Tianchi Aluminum and GC10 industrial datasets show that DPSNet achieves mAP50 scores of 65.4
As application engineers (AE) in production systems usually spend 5 to 10 min on locating relevant documents, 65–70
Accurate characterization of the temporal evolution of machining tool wear is a critical requirement for intelligent machining, predictive maintenance, and surface quality assurance in modern manufacturing processes. Multi-sensor monitoring has been widely adopted in existing studies; however, most approaches rely on discrete-time learning, handcrafted features, or isolated prediction of tool wear and surface roughness, limiting their ability to capture continuous degradation behavior and the intrinsic coupling between tool condition and surface integrity. To overcome these limitations, this study presents a Degradation Dynamics Layer (DDL) based on a Neural Controlled Differential Equation (NCDE) network for multi-sensor operational signal analysis, modeling the temporal evolution of machining tool wear with a shared latent representation. Cutting forces and machining parameters acquired during CNC turning are fused through a continuous-time latent degradation representation that explicitly encodes progressive wear dynamics and surface roughness evolution. The proposed model is implemented in Python using the PyTorch framework and evaluated on a publicly available CNC turning dataset from the Aeronautics Institute of Technology (ITA), involving AISI H13 steel under varying cutting conditions. Experimental results demonstrate that the DDL–NCDE approach achieves an RMSE of 2.89 and an R2 value of 0.971, outperforming state-of-the-art models such as ANN, Attention-GRU, and CNN–BiLSTM demonstrates improved predictive performance compared to internally implemented baseline models. The shared latent degradation state also enhances interpretability, revealing the coupled behavior between tool wear progression and surface quality. These findings confirm that continuous-time, multi-sensor learning through the DDL–NCDE framework significantly improves predictive accuracy and robustness, making it suitable for real-time predictive maintenance and intelligent manufacturing.
Vision-based depth perception is essential for robotics embodiment, industrial inspection, and measurement. However, strong specular reflections, weak surface textures and severe occlusions of cluttered stacking parts make reliable depth perception particularly challenging. Existing depth perception methods often suffer from poor generalization due to unreliable photometric cues and the scarcity of annotated real-world depth data. To address these challenges, this paper proposes a digital twin-inspired geometry-aware monocular depth estimation framework for cluttered stacking parts. A calibrated digital twin is constructed through material and appearance modeling, pose estimation and scene reconstruction, as well as illumination modeling and relighting, enabling the generation of large-scale multimodal training data with precision geometric ground truth. Based on this data, a normal-guided depth estimation network is designed, where surface normals are explicitly embedded into the depth estimation process as transferable geometric priors through feature-level fusion. In addition, a geometry-aware virtual-to-real transfer strategy is introduced to adapt the pretrained model to real industrial images emphasizing geometric. Experimental results demonstrate that the proposed method achieves accurate and structurally consistent depth estimation for cluttered stacking parts and exhibits strong generalization performance, providing an effective solution for depth perception in complex industrial environments.
Thermal deformation of CNC machine tools is still among the key constraints on machining accuracy, energy efficiency, and sustainability. In this investigation, we present a cyber-physical system that combines real-time sensor-based thermal monitoring and digital twin-based finite element simulations to reduce thermal errors in high-speed machining processes. A low-cost AD590 sensor array was installed at thermally critical locations and experimentally validated against quartz thermometer measurements with a thermal compensation accuracy of up to 10
Vision-based real-time locating system (RTLS) approaches are promising for material tracking in smart manufacturing, in which multi-object tracking (MOT) plays a key role in autonomous logistics and process automation. However, conventional tracking algorithms often suffer from identity switches and trajectory fragmentation due to frequent occlusions, multiple visually indistinguishable objects, and nonlinear motion patterns in manufacturing environments. To overcome these limitations, this study develops a controlled, scaled-down, manufacturing-like conveyor testbed and proposes a prediction-assisted BoT-SORT framework integrated with a Seq2Seq LSTM module. Unlike conventional Kalman filter-based methods that rely on a constant-velocity assumption, the proposed Seq2Seq LSTM learns complex kinematic patterns from historical displacement sequences to predict object positions during occlusion. Experimental results show that the proposed method consistently improves tracking robustness in the controlled testbed. In the short-term occlusion scenario, the proposed framework reduced the mean number of identity switches (IDSWs) from 4.75 to 0.75 compared with the baseline BoT-SORT, corresponding to a reduction of approximately 84.2
Efficient monitoring and regeneration of cutting tools are essential for maintaining product quality and process continuity in modern manufacturing environments. Consistent and robust assessment of tool wear remains a practical challenge, particularly when evaluations rely heavily on expert interpretation of visual data. This study presents a vision-based inspection framework for the evaluation of gear hob cutters that integrates defect detection, wear segmentation, and dimensional wear quantification within a unified deep learning pipeline. The system combines a lightweight Ghost Slim U-Net architecture with a compound loss function and is designed for operation within a controlled industrial inspection setup. When evaluated on an expert-annotated dataset acquired from an industrial inspection station, the framework achieved an average F_1 -score of 0.886± 0.004 for wear segmentation. Concurrently, dimensional wear estimation yielded a mean absolute error of 4.05± 0.23μ m . Additional experiments conducted under controlled perturbations of illumination, focus blur, and surface contamination provided a comparative assessment of robustness under non-ideal imaging conditions. The obtained results suggest that the proposed framework may serve as an operator-assistance tool for semi-automated inspection, supporting more consistent wear assessment while retaining expert oversight in regeneration decisions.
