Chang Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Chang Liu has a total of 69 co-authors. Here are the top co-authors by number of shared publications:
- Feifei He
Key Laboratory of Advanced Equipment Intelligent Manufacturing Technology of Yunnan Province, Kunming University of Science & Technology, Kunming, 650500, People’s Republic of China
10 shared publications - Tao Liu
Kunming University of Science and Technology
10 shared publications - Sen Wang
Kunming University of Science and Technology
6 shared publications Xing Wu
Key Laboratory of Advanced Equipment Intelligent Manufacturing Technology of Yunnan Province, Kunming University of Science & Technology, Kunming, 650500, People’s Republic of China
5 shared publicationsTao Liu
4 shared publications- Xiaoqin Liu
Kunming University of Science and Technology
4 shared publications
Latest Publications
- Open Access
- 30 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Physics-Guided Multi-Source Information Fusion Method Based on Multi-Graph Correlation Learning
Multi-source fault diagnosis of rotating machinery requires the fusion of complementary sensor information and physical relationships. Existing methods often rely on feature concatenation, single-relation graphs, or fixed physical topologies, limiting state-adaptive yet physically plausible relation modelling. We propose the Physics-Guided Multi-Graph Correlation Learning Network (PhyMGC-Net), which learns relations in deep-feature, equipment-physical, and hierarchical representation spaces. The method constructs a spectral-topological feature relation graph, a composite physical graph, and a hierarchical relation graph. The composite physical graph uses device-structure constraints, mechanism-informed rules, and observable signal responses to estimate sample-dependent edge weights without requiring complete dynamic parameters. The three graphs are independently encoded and fused through inter-graph relation mapping and sample-dependent gating. On the BJTU-RAO cross-speed task, PhyMGC-Net achieves 95.56% accuracy and 95.52% Macro-F1, exceeding the best baseline by 7.52 and 11.57 percentage points. On the KAIST dataset, it achieves 93.56% accuracy and 93.54% Macro-F1. Ablation results show higher mean performance and lower variability than incomplete graph configurations. PhyMGC-Net provides a state-adaptive, physics-guided approach to structured fusion of heterogeneous sensor information for rotating machinery fault diagnosis.
- 30 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Physics-Guided Multi-Source Information Fusion Method Based on Multi-Graph Correlation Learning
Multi-source fault diagnosis of rotating machinery requires the fusion of complementary sensor information and physical relationships. Existing methods often rely on feature concatenation, single-relation graphs, or fixed physical topologies, limiting state-adaptive yet physically plausible relation modelling. We propose the Physics-Guided Multi-Graph Correlation Learning Network (PhyMGC-Net), which learns relations in deep-feature, equipment-physical, and hierarchical representation spaces. The method constructs a spectral-topological feature relation graph, a composite physical graph, and a hierarchical relation graph. The composite physical graph uses device-structure constraints, mechanism-informed rules, and observable signal responses to estimate sample-dependent edge weights without requiring complete dynamic parameters. The three graphs are independently encoded and fused through inter-graph relation mapping and sample-dependent gating. On the BJTU-RAO cross-speed task, PhyMGC-Net achieves 95.56% accuracy and 95.52% Macro-F1, exceeding the best baseline by 7.52 and 11.57 percentage points. On the KAIST dataset, it achieves 93.56% accuracy and 93.54% Macro-F1. Ablation results show higher mean performance and lower variability than incomplete graph configurations. PhyMGC-Net provides a state-adaptive, physics-guided approach to structured fusion of heterogeneous sensor information for rotating machinery fault diagnosis.
- Open Access
- 16 May 2026
- Published by Elsevier in SSRN Electronic Journal
Physics-Constrained Cross-Modal Alignment and Knowledge Enhancement: An Interpretable Framework for Non-Contact Bearing Fault Diagnosis
To address the limitations of single-modal signals in feature representation for the non-contact fault detection of machine tool bearings, this paper proposes an interpretable cross-modal knowledge enhancement method.First, a Residual Self-Attention Dimensionality Reduction Module (RSADM) is developed to process high-dimensional sparse frequency-domain features via residual connections, ensuring the high-fidelity preservation of the physical traceability of fault features. This is integrated with a multi-head self-attention mechanism to adaptively weight fault-sensitive feature dimensions.Subsequently, an acoustic-vibration cross-modal shared alignment space (VA-FAS) is established to map features extracted from a dual-branch network into a unified representation space. Leveraging Wasserstein distance as an explicit physical constraint to minimize the distribution discrepancy between intra-class acoustic and vibration features, the cross-modal alignment is endowed with a definitive physical connotation of minimum transport cost for fault information.Finally, a cross-modal fusion loss function (W-CMFF Loss) is designed by integrating data-driven and physical constraint losses. By dynamically adjusting balancing weights during training, the cross-modal alignment and fault classification losses are co-optimized, ensuring precise classification while strengthening intra-class aggregation and inter-class separation.Validation results on real-world acoustic-vibration datasets from two experimental testbeds demonstrate that the proposed method effectively enhances the cross-modal alignment and inter-class separability of features. The diagnostic accuracy of knowledge-enhanced non-contact acoustic signals significantly surpasses that of unenhanced signals, while exhibiting superior anti-noise robustness and feature interpretability.
- 16 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPhysics-Constrained Cross-Modal Alignment and Knowledge Enhancement: An Interpretable Framework for Non-Contact Bearing Fault Diagnosis
To address the limitations of single-modal signals in feature representation for the non-contact fault detection of machine tool bearings, this paper proposes an interpretable cross-modal knowledge enhancement method.First, a Residual Self-Attention Dimensionality Reduction Module (RSADM) is developed to process high-dimensional sparse frequency-domain features via residual connections, ensuring the high-fidelity preservation of the physical traceability of fault features. This is integrated with a multi-head self-attention mechanism to adaptively weight fault-sensitive feature dimensions.Subsequently, an acoustic-vibration cross-modal shared alignment space (VA-FAS) is established to map features extracted from a dual-branch network into a unified representation space. Leveraging Wasserstein distance as an explicit physical constraint to minimize the distribution discrepancy between intra-class acoustic and vibration features, the cross-modal alignment is endowed with a definitive physical connotation of minimum transport cost for fault information.Finally, a cross-modal fusion loss function (W-CMFF Loss) is designed by integrating data-driven and physical constraint losses. By dynamically adjusting balancing weights during training, the cross-modal alignment and fault classification losses are co-optimized, ensuring precise classification while strengthening intra-class aggregation and inter-class separation.Validation results on real-world acoustic-vibration datasets from two experimental testbeds demonstrate that the proposed method effectively enhances the cross-modal alignment and inter-class separability of features. The diagnostic accuracy of knowledge-enhanced non-contact acoustic signals significantly surpasses that of unenhanced signals, while exhibiting superior anti-noise robustness and feature interpretability.
- Open Access
- 3 March 2026
- Published by Elsevier in SSRN Electronic Journal
BiMSFNet: A Binarized Multi-Scale Feature Fusion Network for Privacy-Preserving Enhanced Diagnosis on Edge Devices
Rotating machinery is critical equipment in modern industry, and the health status of its bearings directly affects system safety and efficiency. Due to the complexity of the deep learning diagnostic model and the large number of parameters, it is difficult to deploy it in resource-constrained edge embedded devices in real time, and there is also a risk of data privacy leakage. To address these issues, this paper proposes a Binarized Multi-Scale Feature Fusion Network (BiMSFNet) for edge embedded devices. First, in data preprocessing, quantization and bit-plane decomposition convert 2D time–frequency maps into multi-dimensional binary matrices, achieving information transformation and initial privacy protection. Moreover, by leveraging the inherent discretization and non-continuity of the binary neural network itself further resist gradient inversion attacks at the model level. Then, by combining binary neural networks with adaptive weighting, a lightweight model is constructed. Finally, integrating multi-dilation-rate convolutions with adaptive weighting enables multi-scale feature extraction while drastically reducing model size and training cost. Experiments show that BiMSFNet achieves 99.54% average accuracy on a self-built testbed and 99.18% on the SEU bearing dataset with only 0.169M parameters and 17.6MFLOPs, 1–2 orders of magnitude lower than mainstream lightweight models, and effectively resists gradient inversion attacks on the CWRU dataset. Deployed on a Raspberry Pi 5B, inference time is 2.8ms per sample, meeting the real-time, low-power, and privacy requirements of edge scenarios.
- 3 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBiMSFNet: A Binarized Multi-Scale Feature Fusion Network for Privacy-Preserving Enhanced Diagnosis on Edge Devices
Rotating machinery is critical equipment in modern industry, and the health status of its bearings directly affects system safety and efficiency. Due to the complexity of the deep learning diagnostic model and the large number of parameters, it is difficult to deploy it in resource-constrained edge embedded devices in real time, and there is also a risk of data privacy leakage. To address these issues, this paper proposes a Binarized Multi-Scale Feature Fusion Network (BiMSFNet) for edge embedded devices. First, in data preprocessing, quantization and bit-plane decomposition convert 2D time–frequency maps into multi-dimensional binary matrices, achieving information transformation and initial privacy protection. Moreover, by leveraging the inherent discretization and non-continuity of the binary neural network itself further resist gradient inversion attacks at the model level. Then, by combining binary neural networks with adaptive weighting, a lightweight model is constructed. Finally, integrating multi-dilation-rate convolutions with adaptive weighting enables multi-scale feature extraction while drastically reducing model size and training cost. Experiments show that BiMSFNet achieves 99.54% average accuracy on a self-built testbed and 99.18% on the SEU bearing dataset with only 0.169M parameters and 17.6MFLOPs, 1–2 orders of magnitude lower than mainstream lightweight models, and effectively resists gradient inversion attacks on the CWRU dataset. Deployed on a Raspberry Pi 5B, inference time is 2.8ms per sample, meeting the real-time, low-power, and privacy requirements of edge scenarios.
- 1 March 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (6) , 065524
A physics-informed graph convolutional network for bearing fault diagnosis via dynamic constraint-guided feature learning
To overcome the issues of limited diagnostic reliability and poor generalization in the process of bearing fault diagnosis due to the lack of physical mechanism constraints and the opacity of model structure,this paper proposes a bearing fault diagnosis method that integrates Physics-Informed Neural Networks (PINNs) with Graph Convolutional Neural Networks (GCNs). By incorporating bearings dynamic differential equations, the fault mechanism knowledge is embedded into the model, enabling a fusion of data-driven deep learning and physical information.Firstly, by introducing the bearing dynamics equation, the bearing vibration signal is transformed into a feature vector that satisfies the physical constraints, thereby achieving the integration of data and physical mechanism. Then, based on the feature vectors, physical residuals are further reintroduced in the process of constructing the graph edge weights, so that the generated graph structure conforms to the dynamic laws of bearing faults. Finally, a physical residual derived from the two-degree-of-freedom bearing fault dynamics model is introduced into the loss function of GCN, enabling GCN to learn physical features consistent with the fault behavior at the graph node level. Experimental results demonstrate that the proposed method significantly reduces physical loss, achieving reductions of up to 96.02% and 75.55% compared with other models, and enhances physical consistency, achieving diagnostic accuracies of up to 94.59% and 93.78% on rotating machinery fault simulation test bench and the bearing vibration data under time-varying rotational speed conditions from the University of Ottawa, respectively. These results indicate that the proposed method has stability and generalization capability across different operating conditions. Ablation studies further confirm the importance of different physics-informed modules in improving diagnostic accuracy and model generalization performance.
- 1 March 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (6) , 065524
A physics-informed graph convolutional network for bearing fault diagnosis via dynamic constraint-guided feature learning
To overcome the issues of limited diagnostic reliability and poor generalization in the process of bearing fault diagnosis due to the lack of physical mechanism constraints and the opacity of model structure,this paper proposes a bearing fault diagnosis method that integrates Physics-Informed Neural Networks (PINNs) with Graph Convolutional Neural Networks (GCNs). By incorporating bearings dynamic differential equations, the fault mechanism knowledge is embedded into the model, enabling a fusion of data-driven deep learning and physical information.Firstly, by introducing the bearing dynamics equation, the bearing vibration signal is transformed into a feature vector that satisfies the physical constraints, thereby achieving the integration of data and physical mechanism. Then, based on the feature vectors, physical residuals are further reintroduced in the process of constructing the graph edge weights, so that the generated graph structure conforms to the dynamic laws of bearing faults. Finally, a physical residual derived from the two-degree-of-freedom bearing fault dynamics model is introduced into the loss function of GCN, enabling GCN to learn physical features consistent with the fault behavior at the graph node level. Experimental results demonstrate that the proposed method significantly reduces physical loss, achieving reductions of up to 96.02% and 75.55% compared with other models, and enhances physical consistency, achieving diagnostic accuracies of up to 94.59% and 93.78% on rotating machinery fault simulation test bench and the bearing vibration data under time-varying rotational speed conditions from the University of Ottawa, respectively. These results indicate that the proposed method has stability and generalization capability across different operating conditions. Ablation studies further confirm the importance of different physics-informed modules in improving diagnostic accuracy and model generalization performance.
- 29 December 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 0455f2
Fault mechanism transfer of rolling components for imbalanced sample fault diagnosis
This paper addresses the problem of imbalanced samples in the ball screw pair in engineering, as well as challenges such as poor model training performance and difficulties in edge deployment. A fault diagnosis method based on mechanism transfer is proposed. By analyzing the fault mechanism similarities in rolling components, the knowledge distillation framework is utilized to explore feature information transfer for cross-object fault diagnosis knowledge sharing. First, the fault mechanism similarities between rolling bearings and ball screw pairs are studied, and a unified description of fault mechanisms in rolling components is explored. This is achieved by leveraging the comprehensive knowledge of rolling bearings to complete the sample space of ball screw pairs, thereby improving fault diagnosis performance. Then, a dual-layer feature alignment knowledge distillation (DLFA-KD) framework is constructed. ResNet50 is used as the teacher model, and a lightweight student model is designed with multi-kernel depthwise separable convolutions (MK-DSC). Maximum mean discrepancy (MMD) and L2 distance are introduced to guide dual-layer feature alignment, and cross-entropy loss is used to optimize classification performance. Finally, a comprehensive sample set is constructed using multi-condition bearing data, and an imbalanced sample set is created using multi-condition ball screw pair data for mechanism transfer fault diagnosis performance verification. The experimental results show that the proposed method significantly improves model performance under various levels of sample imbalance, with an accuracy of up to 99.675%. Moreover, the student model’s parameter count is only 2% of that of the teacher model, and after edge deployment, the inference speed increases by 98.1 times. This method provides a new approach for edge-based intelligent fault diagnosis in industrial scenarios with imbalanced samples.
- 29 December 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 0455f2
Fault mechanism transfer of rolling components for imbalanced sample fault diagnosis
This paper addresses the problem of imbalanced samples in the ball screw pair in engineering, as well as challenges such as poor model training performance and difficulties in edge deployment. A fault diagnosis method based on mechanism transfer is proposed. By analyzing the fault mechanism similarities in rolling components, the knowledge distillation framework is utilized to explore feature information transfer for cross-object fault diagnosis knowledge sharing. First, the fault mechanism similarities between rolling bearings and ball screw pairs are studied, and a unified description of fault mechanisms in rolling components is explored. This is achieved by leveraging the comprehensive knowledge of rolling bearings to complete the sample space of ball screw pairs, thereby improving fault diagnosis performance. Then, a dual-layer feature alignment knowledge distillation (DLFA-KD) framework is constructed. ResNet50 is used as the teacher model, and a lightweight student model is designed with multi-kernel depthwise separable convolutions (MK-DSC). Maximum mean discrepancy (MMD) and L2 distance are introduced to guide dual-layer feature alignment, and cross-entropy loss is used to optimize classification performance. Finally, a comprehensive sample set is constructed using multi-condition bearing data, and an imbalanced sample set is created using multi-condition ball screw pair data for mechanism transfer fault diagnosis performance verification. The experimental results show that the proposed method significantly improves model performance under various levels of sample imbalance, with an accuracy of up to 99.675%. Moreover, the student model’s parameter count is only 2% of that of the teacher model, and after edge deployment, the inference speed increases by 98.1 times. This method provides a new approach for edge-based intelligent fault diagnosis in industrial scenarios with imbalanced samples.
- 29 December 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 0455e8
Adaptive compressive feature extraction method for bearing signals via dual-branch wavelet convolution joint sparse sensing
To address the issues of difficult sparsity evaluation in traditional compressive sensing and strong randomness in signal compressive sampling, an adaptive deep compressive feature extraction method based on dual-branch wavelet convolution joint sparse sensing is proposed and successfully applied to bearing vibration signals. Firstly, a feature extraction module with a dual-branch heterogeneous wavelet convolution architecture is designed to fully leverage the advantages of different-sized convolution kernels in feature extraction, capturing bearing vibration data features at multiple scales. Secondly, a deep compressive feature reduction module is developed based on the energy-preserving property of compressive sensing. An energy-and-information-entropy co-driven compression strategy is formulated to deeply optimize the reduction process, effectively improving the quality of compressive features. Then, a novel loss function with joint spatial-spectral optimization is constructed to enhance the model’s ability to learn sparse features, and an Adaptive Dynamic Weight Nonlinear Transition (ADWNT) mechanism is proposed to adaptively adjust the loss function weights. Finally, the proposed method is validated using data from a thrust bearing limit state performance test bench and a rotating machinery fault simulation test bench. Experimental results demonstrate that under strong noise conditions, the proposed method not only achieves a Pearson Correlation Coefficient (PCC) exceeding 90% and an approximate 18 dB improvement in the signal-to-noise ratio (SNR) of fault characteristic frequencies—significantly outperforming traditional methods and basic deep learning methods to enable efficient extraction of bearing fault features—but also overcomes the limitations of traditional compressive sensing in feature extraction under strong background noise, while maintaining a favorable balance between model complexity and performance.
- 29 December 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 0455e8
Adaptive compressive feature extraction method for bearing signals via dual-branch wavelet convolution joint sparse sensing
To address the issues of difficult sparsity evaluation in traditional compressive sensing and strong randomness in signal compressive sampling, an adaptive deep compressive feature extraction method based on dual-branch wavelet convolution joint sparse sensing is proposed and successfully applied to bearing vibration signals. Firstly, a feature extraction module with a dual-branch heterogeneous wavelet convolution architecture is designed to fully leverage the advantages of different-sized convolution kernels in feature extraction, capturing bearing vibration data features at multiple scales. Secondly, a deep compressive feature reduction module is developed based on the energy-preserving property of compressive sensing. An energy-and-information-entropy co-driven compression strategy is formulated to deeply optimize the reduction process, effectively improving the quality of compressive features. Then, a novel loss function with joint spatial-spectral optimization is constructed to enhance the model’s ability to learn sparse features, and an Adaptive Dynamic Weight Nonlinear Transition (ADWNT) mechanism is proposed to adaptively adjust the loss function weights. Finally, the proposed method is validated using data from a thrust bearing limit state performance test bench and a rotating machinery fault simulation test bench. Experimental results demonstrate that under strong noise conditions, the proposed method not only achieves a Pearson Correlation Coefficient (PCC) exceeding 90% and an approximate 18 dB improvement in the signal-to-noise ratio (SNR) of fault characteristic frequencies—significantly outperforming traditional methods and basic deep learning methods to enable efficient extraction of bearing fault features—but also overcomes the limitations of traditional compressive sensing in feature extraction under strong background noise, while maintaining a favorable balance between model complexity and performance.
- 19 November 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 045585
A knowledge migration method based on multi-axis state coupling and multimodal data fusion for CNC process pattern recognition
To address the challenges of extracting steady-state machining intervals in complex Computer numerical control (CNC) cutting, the high complexity of multimodal data fusion for machine-tool condition characterization, and the difficulty of deploying external sensors on structurally diverse CNC machines, this paper proposes a knowledge-transfer method that mines coupled multi-axis state information, fuses multimodal data to extract CNC machining features, and establishes a knowledge-distillation framework to guide cross-modal feature migration for CNC process-mode recognition. First, a multi-axis synchronous acquisition platform is employed to concurrently capture vibration and current signals under coordinated multi-axis machining, enabling the revelation of inter-axis state synergies across different process modes and the reliable extraction of steady-state machining intervals in complex cutting. Second, a Vibration-Guided Cross-Attention Mechanism (VGCAM) module is introduced together with a Self-Distilled Spatial Attention Mechanism (SDSAM). These components suppress redundancy in multimodal data while injecting high-frequency vibration cues into the current stream, thereby enhancing cross-modal feature representation and improving the efficiency of multimodal fusion. Finally, a multi-stage knowledge-distillation framework is constructed in which a teacher network extracts and fuses multimodal features; the fused representations and process knowledge are then transferred to a single-modal student network. The distilled student model achieves substantially fewer parameters and lower inference latency, and offers greater deployment convenience than external vibration sensors and similar add-on instrumentation. The proposed method is validated on a CNC milling testbed using nylon and aluminum workpieces. Results show that the student network, when supplied only with current root-mean-square (RMS) features, achieves high process-mode recognition accuracy while significantly reducing network parameters and inference latency, demonstrating the feasibility of on-site industrial deployment.
- 19 November 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 045585
A knowledge migration method based on multi-axis state coupling and multimodal data fusion for CNC process pattern recognition
To address the challenges of extracting steady-state machining intervals in complex Computer numerical control (CNC) cutting, the high complexity of multimodal data fusion for machine-tool condition characterization, and the difficulty of deploying external sensors on structurally diverse CNC machines, this paper proposes a knowledge-transfer method that mines coupled multi-axis state information, fuses multimodal data to extract CNC machining features, and establishes a knowledge-distillation framework to guide cross-modal feature migration for CNC process-mode recognition. First, a multi-axis synchronous acquisition platform is employed to concurrently capture vibration and current signals under coordinated multi-axis machining, enabling the revelation of inter-axis state synergies across different process modes and the reliable extraction of steady-state machining intervals in complex cutting. Second, a Vibration-Guided Cross-Attention Mechanism (VGCAM) module is introduced together with a Self-Distilled Spatial Attention Mechanism (SDSAM). These components suppress redundancy in multimodal data while injecting high-frequency vibration cues into the current stream, thereby enhancing cross-modal feature representation and improving the efficiency of multimodal fusion. Finally, a multi-stage knowledge-distillation framework is constructed in which a teacher network extracts and fuses multimodal features; the fused representations and process knowledge are then transferred to a single-modal student network. The distilled student model achieves substantially fewer parameters and lower inference latency, and offers greater deployment convenience than external vibration sensors and similar add-on instrumentation. The proposed method is validated on a CNC milling testbed using nylon and aluminum workpieces. Results show that the student network, when supplied only with current root-mean-square (RMS) features, achieves high process-mode recognition accuracy while significantly reducing network parameters and inference latency, demonstrating the feasibility of on-site industrial deployment.
- 12 November 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 045574
A rolling bearing fault diagnosis method based on multi-feature fusion of BiLSTM-DSC- transformer under strong noise background
In the backdrop of the rapid development of intelligent manufacturing technologies, bearings, as core components of mechanical equipment, play a crucial role in monitoring operational status and diagnosing faults to ensure production safety and efficiency. However, bearings often operate under severe environmental noise and frequently changing working conditions, which significantly constrain the noise resistance and generalization performance of intelligent fault diagnosis methods. Therefore, this paper proposes a multi-feature fusion method for rolling bearing fault diagnosis based on BiLSTM-DSC-Transformer (BDT). Firstly, a new nonlinear activation function, R-Swish, is designed and applied to a lightweight deep separable convolution (DSC) feature extraction module in both time and frequency domains, ensuring the model’s feature extraction capability and operational efficiency. Secondly, an adjustable BiLSTM-Transformer model is constructed, fully leveraging the advantages of Transformer in capturing long-range dependencies and the bidirectional data processing mechanism of BiLSTM to reveal the intrinsic relationships among multiple features. Finally, a feature fusion module is established to integrate multi-feature information, combined with a linear classification layer, to achieve high-accuracy fault diagnosis of bearings under complex working conditions. Experimental results from two case studies demonstrate that the proposed method exhibits excellent robustness and generalization ability under strong noise and varying load conditions, providing a new and effective technical solution for the field of bearing fault diagnosis.
- 12 November 2025
- Published by IOP Publishing in Engineering Research Express
- Vol. 7 (4) , 045574
A rolling bearing fault diagnosis method based on multi-feature fusion of BiLSTM-DSC- transformer under strong noise background
In the backdrop of the rapid development of intelligent manufacturing technologies, bearings, as core components of mechanical equipment, play a crucial role in monitoring operational status and diagnosing faults to ensure production safety and efficiency. However, bearings often operate under severe environmental noise and frequently changing working conditions, which significantly constrain the noise resistance and generalization performance of intelligent fault diagnosis methods. Therefore, this paper proposes a multi-feature fusion method for rolling bearing fault diagnosis based on BiLSTM-DSC-Transformer (BDT). Firstly, a new nonlinear activation function, R-Swish, is designed and applied to a lightweight deep separable convolution (DSC) feature extraction module in both time and frequency domains, ensuring the model’s feature extraction capability and operational efficiency. Secondly, an adjustable BiLSTM-Transformer model is constructed, fully leveraging the advantages of Transformer in capturing long-range dependencies and the bidirectional data processing mechanism of BiLSTM to reveal the intrinsic relationships among multiple features. Finally, a feature fusion module is established to integrate multi-feature information, combined with a linear classification layer, to achieve high-accuracy fault diagnosis of bearings under complex working conditions. Experimental results from two case studies demonstrate that the proposed method exhibits excellent robustness and generalization ability under strong noise and varying load conditions, providing a new and effective technical solution for the field of bearing fault diagnosis.