Chenyu Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Chenyu Liu has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Konstantinos Gryllias
KU Leuven
5 shared publications - Dandan Peng
Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong, China
3 shared publications Bin Luo
2 shared publicationsHui Cheng
2 shared publications- Zhuyun Chen
Guangdong University of Technology
2 shared publications - Alexandre Mauricio
KU Leuven
1 shared publication
Latest Publications
- Open Access
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Heterogeneous Feature Fusion-Empowered Multimodal Large Language Models for Bearing Fault Diagnosis
The deployment of general Large Language Models (LLMs) in industrial fault diagnosis faces severe challenges due to the substantial adaptation gap between linguistic knowledge and high-dimensional physical signals. These challenges are mainly manifested in representational heterogeneity, reliability risks in the reasoning process, and blurred decision boundaries under long-tail distributions. To address these issues, this study proposes Heterogeneous Feature Fusion-Empowered Multimodal Large Language Model (HFF-MLLM), a framework that empowers LLMs through deep heterogeneous feature fusion and physical-information-based constraints. Specifically, the proposed architecture first establishes a cross-modal alignment mechanism based on dual-path projection, which maps the local dynamics of one-dimensional time series and the global spectral patterns of two-dimensional continuous wavelet transform images into a unified latent space. By leveraging complementary multimodal semantics, this design effectively overcomes the performance bottleneck associated with shallow fusion strategies. Meanwhile, to ensure the credibility of the reasoning process, a hybrid reasoning strategy is introduced. This module uses key signal indicators as explicit physical-feature guidance and imposes conditional constraints on the generation process via Low-Rank Adaptation (LoRA), thereby effectively suppressing the non-factual reasoning bias induced by purely data-driven paradigms. In addition, to address manifold collapse in imbalanced scenarios, a dynamic balance calibration mechanism integrating focal loss and entropy regularization is employed to reshape discriminative decision boundaries and suppress model overconfidence. Extensive empirical evaluations conducted on four datasets (CWRU, JNU, SEU, and their mixture) and across four LLM backbones and two Pre-trained Language Model (PLM) backbones verify the superior generalization capability of the proposed framework, which achieves peak accuracies of 99.32\% and 99.66\% on the JNU and SEU datasets, respectively.
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHeterogeneous Feature Fusion-Empowered Multimodal Large Language Models for Bearing Fault Diagnosis
The deployment of general Large Language Models (LLMs) in industrial fault diagnosis faces severe challenges due to the substantial adaptation gap between linguistic knowledge and high-dimensional physical signals. These challenges are mainly manifested in representational heterogeneity, reliability risks in the reasoning process, and blurred decision boundaries under long-tail distributions. To address these issues, this study proposes Heterogeneous Feature Fusion-Empowered Multimodal Large Language Model (HFF-MLLM), a framework that empowers LLMs through deep heterogeneous feature fusion and physical-information-based constraints. Specifically, the proposed architecture first establishes a cross-modal alignment mechanism based on dual-path projection, which maps the local dynamics of one-dimensional time series and the global spectral patterns of two-dimensional continuous wavelet transform images into a unified latent space. By leveraging complementary multimodal semantics, this design effectively overcomes the performance bottleneck associated with shallow fusion strategies. Meanwhile, to ensure the credibility of the reasoning process, a hybrid reasoning strategy is introduced. This module uses key signal indicators as explicit physical-feature guidance and imposes conditional constraints on the generation process via Low-Rank Adaptation (LoRA), thereby effectively suppressing the non-factual reasoning bias induced by purely data-driven paradigms. In addition, to address manifold collapse in imbalanced scenarios, a dynamic balance calibration mechanism integrating focal loss and entropy regularization is employed to reshape discriminative decision boundaries and suppress model overconfidence. Extensive empirical evaluations conducted on four datasets (CWRU, JNU, SEU, and their mixture) and across four LLM backbones and two Pre-trained Language Model (PLM) backbones verify the superior generalization capability of the proposed framework, which achieves peak accuracies of 99.32\% and 99.66\% on the JNU and SEU datasets, respectively.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-16
Multihead Self-Attention Incremental Network for Rotating Machinery Fault Diagnosis Across Nonstationary Operating Conditions
The advancement of artificial intelligence has enabled significant progress in data-driven fault diagnosis for rotating machinery. However, in real-world industrial environments, such machinery often operates under time-varying and non-stationary conditions, leading to continuous changes in data distributions. These variations pose a major challenge to conventional deep learning models, which are prone to catas-trophic forgetting when sequentially learning new diagnostic tasks. To address this, this paper proposes a novel Multi-head Self-attention Incremental Network (MSIN) for fault diagnosis across non-stationary operating conditions. MSIN employs a stack of multi-head self-attention encoder layers to effectively extract fault-relevant features from frequency-domain signals. To alleviate catastrophic forgetting, the model integrates a selective exemplar replay strategy and introduces a joint cross-head knowledge distillation loss, enabling effective knowledge retention across tasks. The proposed method is evaluated on two benchmark datasets under both stationary and non-stationary conditions. Experimental results demonstrate that MSIN achieves superior diagnostic accuracy and knowledge retention compared to existing Incremental Learning methods, offering a promising solution for continuous fault diagnosis in dynamic industrial settings. The source code of the MSIN model is available at: https://github.com/bankuaimianbao/MSIN-Non-stationary-Code.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-16
Multihead Self-Attention Incremental Network for Rotating Machinery Fault Diagnosis Across Nonstationary Operating Conditions
The advancement of artificial intelligence has enabled significant progress in data-driven fault diagnosis for rotating machinery. However, in real-world industrial environments, such machinery often operates under time-varying and non-stationary conditions, leading to continuous changes in data distributions. These variations pose a major challenge to conventional deep learning models, which are prone to catas-trophic forgetting when sequentially learning new diagnostic tasks. To address this, this paper proposes a novel Multi-head Self-attention Incremental Network (MSIN) for fault diagnosis across non-stationary operating conditions. MSIN employs a stack of multi-head self-attention encoder layers to effectively extract fault-relevant features from frequency-domain signals. To alleviate catastrophic forgetting, the model integrates a selective exemplar replay strategy and introduces a joint cross-head knowledge distillation loss, enabling effective knowledge retention across tasks. The proposed method is evaluated on two benchmark datasets under both stationary and non-stationary conditions. Experimental results demonstrate that MSIN achieves superior diagnostic accuracy and knowledge retention compared to existing Incremental Learning methods, offering a promising solution for continuous fault diagnosis in dynamic industrial settings. The source code of the MSIN model is available at: https://github.com/bankuaimianbao/MSIN-Non-stationary-Code.
- 12 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cybernetics
- Vol. PP (21682267) , 1-14
AeroGPT: Leveraging Large-Scale Audio Model for Aero-Engine Bearing Fault Diagnosis
Aerospace engines, as critical components in the aviation and aerospace industries, require continuous and accurate fault diagnosis to ensure operational safety and prevent catastrophic failures. While deep learning techniques have been extensively studied in this context, they typically output logits or confidence scores, necessitating postprocessing to obtain actionable insights. Furthermore, the potential of large-scale audio models for this task remains largely untapped. To address these limitations, this article proposes AeroGPT, a novel framework that transfers knowledge from the general audio domain to aero-engine bearing fault diagnosis. AeroGPT leverages a large-scale audio model and incorporates vibration signal alignment (VSA) to adapt general audio knowledge to domain-specific vibration patterns, along with generative fault classification (GFC) to directly generate interpretable fault labels. This approach eliminates the need for label postprocessing and supports interactive, interpretable, and actionable fault diagnosis, thereby enhancing industrial applicability. Through comprehensive experimental validation on two aero-engine bearing datasets, AeroGPT achieves 98.94% accuracy on the Dynamic and Identification Research Group (DIRG) dataset and 100% accuracy on Harbin Institute of Technology (HIT) bearing dataset, outperforming representative deep learning approaches. Qualitative analysis and further discussion also demonstrate its potential for interactive diagnosis and real-world deployment, highlighting the promise of large-scale audio models to advance fault diagnosis in aerospace applications.
- 12 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cybernetics
- Vol. PP (21682267) , 1-14
AeroGPT: Leveraging Large-Scale Audio Model for Aero-Engine Bearing Fault Diagnosis
Aerospace engines, as critical components in the aviation and aerospace industries, require continuous and accurate fault diagnosis to ensure operational safety and prevent catastrophic failures. While deep learning techniques have been extensively studied in this context, they typically output logits or confidence scores, necessitating postprocessing to obtain actionable insights. Furthermore, the potential of large-scale audio models for this task remains largely untapped. To address these limitations, this article proposes AeroGPT, a novel framework that transfers knowledge from the general audio domain to aero-engine bearing fault diagnosis. AeroGPT leverages a large-scale audio model and incorporates vibration signal alignment (VSA) to adapt general audio knowledge to domain-specific vibration patterns, along with generative fault classification (GFC) to directly generate interpretable fault labels. This approach eliminates the need for label postprocessing and supports interactive, interpretable, and actionable fault diagnosis, thereby enhancing industrial applicability. Through comprehensive experimental validation on two aero-engine bearing datasets, AeroGPT achieves 98.94% accuracy on the Dynamic and Identification Research Group (DIRG) dataset and 100% accuracy on Harbin Institute of Technology (HIT) bearing dataset, outperforming representative deep learning approaches. Qualitative analysis and further discussion also demonstrate its potential for interactive diagnosis and real-world deployment, highlighting the promise of large-scale audio models to advance fault diagnosis in aerospace applications.
- 20 January 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (4) , 040201
Special issue: advancements in pattern recognition and machine learning for predictive maintenance in industry applications
- 20 January 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (4) , 040201
Special issue: advancements in pattern recognition and machine learning for predictive maintenance in industry applications
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Unsupervised Anomaly Detection for Aircraft PRSOV With Random Projection-Based Inner Product Prediction
Pressure Regulated Shutoff Valves (PRSOV) are critical components of aircrafts’ environmental control system to regulate the carbon pressure, ensuring it within a safe and comfortable range. However, due to the harsh and dynamic operational environment, PRSOVs are susceptible to various anomalies, which may lead to severe flight incidents such as cabin depressurization. To this end, this paper proposed an unsupervised anomaly detection approach for PRSOV with random projection-based inner product prediction (RPDP-AD). It leverages distance information in a randomly projected space to construct supervisory signals. This enables our model to encode distribution patterns and data structures of normal samples from unlabeled data. Our model is optimized to minimize the prediction errors in the random projection space. We introduced a distribution difference metric to minimize the discrepancy between learnable mappings and random mappings in low-dimensional space. Anomalous samples will have greater distribution differences, which allows for unsupervised anomaly detection. This paper explored the connection between neural networks and random projection theory for industrial anomaly detection. To further support its effectiveness, we also provide a theoretical analysis of inner product preservation property via Johnson-Lindenstrauss theorem. Experiments were conducted on a PRSOV dataset including 7 types of anomalies with different random projection methods. Results demonstrate that our method outperforms other data-driven counterparts. Our source code will be available at https://github.com/Z-yiwei/RPDP-AD.
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Unsupervised Anomaly Detection for Aircraft PRSOV With Random Projection-Based Inner Product Prediction
Pressure Regulated Shutoff Valves (PRSOV) are critical components of aircrafts’ environmental control system to regulate the carbon pressure, ensuring it within a safe and comfortable range. However, due to the harsh and dynamic operational environment, PRSOVs are susceptible to various anomalies, which may lead to severe flight incidents such as cabin depressurization. To this end, this paper proposed an unsupervised anomaly detection approach for PRSOV with random projection-based inner product prediction (RPDP-AD). It leverages distance information in a randomly projected space to construct supervisory signals. This enables our model to encode distribution patterns and data structures of normal samples from unlabeled data. Our model is optimized to minimize the prediction errors in the random projection space. We introduced a distribution difference metric to minimize the discrepancy between learnable mappings and random mappings in low-dimensional space. Anomalous samples will have greater distribution differences, which allows for unsupervised anomaly detection. This paper explored the connection between neural networks and random projection theory for industrial anomaly detection. To further support its effectiveness, we also provide a theoretical analysis of inner product preservation property via Johnson-Lindenstrauss theorem. Experiments were conducted on a PRSOV dataset including 7 types of anomalies with different random projection methods. Results demonstrate that our method outperforms other data-driven counterparts. Our source code will be available at https://github.com/Z-yiwei/RPDP-AD.
- 8 February 2024
- Published by ASME International in Journal of Engineering for Gas Turbines and Power
- Vol. 146 (7) , 1-10
Condition Monitoring of Wind Turbine Drivetrain Bearings
As an extremely valuable and reliable energy source, the renewable energy is developing around the world at an unprecedented pace. By the end of 2022, the global installed capacity of offshore wind turbines is expected to reach 46.4 GW, among which 33.9 GW in Europe. Efficiencies in Operations and Maintenance (O&M) offer potential to achieve significant cost savings as it accounts for around 20%–30% of overall offshore wind farm costs. A recent study of approximately 350 offshore wind turbines indicates that gearboxes might have to be replaced as early as 6.5 years. Therefore sensing and condition monitoring (CM) systems are needed in order to obtain reliable information on the state and condition of different critical parts. The development and use of such technologies will allow companies to schedule actions at the right time. The reduced costs of O&M enable the wind energy at a competitive price thus strengthening productivity of the wind energy sector. At the academic level, a plethora of methodologies have been proposed during the last decades focusing toward early and accurate fault detection. Among others, Envelope Analysis is one of the most important methodologies, where the envelope of the vibration signal is estimated, usually after filtering around a selected frequency band excited by impacts due to the faults. Different tools, such as Kurtogram, have been proposed in order to accurately select the optimum filter parameters (center frequency and bandwidth). Cyclostationary analysis and corresponding methodologies, i.e., the cyclic spectral correlation and the cyclic spectral coherence, have been proved as powerful tools for CM. On the other hand the application, test, and evaluation of such tools in general industrial cases is still rather limited. Therefore the main aim of this paper is the application and evaluation of advanced diagnostic techniques and diagnostic indicators, including the Enhanced Envelope Spectrum and the Spectral Flatness on real-world vibration data collected from vibration sensors on gearboxes in multiple wind turbines over an extended period of time of nearly four years. The diagnostic indicators are compared with classical statistic time and frequency indicators, i.e., Kurtosis, Crest Factor, etc. and their effectiveness is evaluated based on the successful detection of two failure events.
- 8 February 2024
- Published by ASME International in Journal of Engineering for Gas Turbines and Power
- Vol. 146 (7) , 1-10
Condition Monitoring of Wind Turbine Drivetrain Bearings
As an extremely valuable and reliable energy source, the renewable energy is developing around the world at an unprecedented pace. By the end of 2022, the global installed capacity of offshore wind turbines is expected to reach 46.4 GW, among which 33.9 GW in Europe. Efficiencies in Operations and Maintenance (O&M) offer potential to achieve significant cost savings as it accounts for around 20%–30% of overall offshore wind farm costs. A recent study of approximately 350 offshore wind turbines indicates that gearboxes might have to be replaced as early as 6.5 years. Therefore sensing and condition monitoring (CM) systems are needed in order to obtain reliable information on the state and condition of different critical parts. The development and use of such technologies will allow companies to schedule actions at the right time. The reduced costs of O&M enable the wind energy at a competitive price thus strengthening productivity of the wind energy sector. At the academic level, a plethora of methodologies have been proposed during the last decades focusing toward early and accurate fault detection. Among others, Envelope Analysis is one of the most important methodologies, where the envelope of the vibration signal is estimated, usually after filtering around a selected frequency band excited by impacts due to the faults. Different tools, such as Kurtogram, have been proposed in order to accurately select the optimum filter parameters (center frequency and bandwidth). Cyclostationary analysis and corresponding methodologies, i.e., the cyclic spectral correlation and the cyclic spectral coherence, have been proved as powerful tools for CM. On the other hand the application, test, and evaluation of such tools in general industrial cases is still rather limited. Therefore the main aim of this paper is the application and evaluation of advanced diagnostic techniques and diagnostic indicators, including the Enhanced Envelope Spectrum and the Spectral Flatness on real-world vibration data collected from vibration sensors on gearboxes in multiple wind turbines over an extended period of time of nearly four years. The diagnostic indicators are compared with classical statistic time and frequency indicators, i.e., Kurtosis, Crest Factor, etc. and their effectiveness is evaluated based on the successful detection of two failure events.
- 8 November 2022
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
- Vol. 238 (2-3) , 295-303
A novel H∞ observer for sensor fault diagnosis of vehicle electronic stability control system with disturbance and input delay
This paper mainly focuses on the sensor fault diagnosis method based on H∞ observer in vehicle electronic stability control (ESC) system with disturbance and input delay. Firstly, the four-wheel steering vehicle model is established. Secondly, the system state and sensor fault are reconstructed into new system variables by a special matrix transformation method, based on which a new H∞ observer is developed to estimate system state and sensor fault signals. Then, based on Lyapunov method and linear matrix inequality (LMI) theory, the conditions for asymptotic stability and H∞ performance of the estimation error system are proved. Finally, Carsim and Matlab/Simulink co-simulation are used to show the effectiveness of the method.
- 8 November 2022
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
- Vol. 238 (2-3) , 295-303
A novel H∞ observer for sensor fault diagnosis of vehicle electronic stability control system with disturbance and input delay
This paper mainly focuses on the sensor fault diagnosis method based on H∞ observer in vehicle electronic stability control (ESC) system with disturbance and input delay. Firstly, the four-wheel steering vehicle model is established. Secondly, the system state and sensor fault are reconstructed into new system variables by a special matrix transformation method, based on which a new H∞ observer is developed to estimate system state and sensor fault signals. Then, based on Lyapunov method and linear matrix inequality (LMI) theory, the conditions for asymptotic stability and H∞ performance of the estimation error system are proved. Finally, Carsim and Matlab/Simulink co-simulation are used to show the effectiveness of the method.
- Open Access
- 28 October 2022
- Published by PHM Society in Annual Conference of the PHM Society
- Vol. 14 (1)
A Transfer Learning-Based Rolling Bearing Fault Diagnosis Across Machines
Currently, most studies focus on training an excellent deep learning model based on sufficient labeled data collected from machines. However, in real applications, it is costly or impractical to obtain massive labeled data for model training. Therefore, in this paper, a Transfer Learning (TL)-based fault diagnosis method is proposed to transfer the model learnt from one machine (source domain) to another one (target domain). In the training process, labeled source data and unlabeled target data are used, which is very promising for real industrial applications. In this frame of transfer learningbased fault diagnosis, a cyclic spectrum correlation analysis method is firstly introduced to obtain order frequency maps for removing the influence of speed variation and revealing the hidden cyclic frequency of signals. Then, the Dynamic Adversarial Adaptation Network (DAAN) is introduced to transfer label information across machines. The proposed fault diagnosis method across machines is applied on two rolling element bearing datasets collected from two different test rigs. Experimental results demonstrate the effectiveness and superiority of the proposed method compared with stateof-the-art approaches.
- 28 October 2022
- Published by PHM Society in Annual Conference of the PHM Society
- Vol. 14 (1)
Open AccessA Transfer Learning-Based Rolling Bearing Fault Diagnosis Across Machines
Currently, most studies focus on training an excellent deep learning model based on sufficient labeled data collected from machines. However, in real applications, it is costly or impractical to obtain massive labeled data for model training. Therefore, in this paper, a Transfer Learning (TL)-based fault diagnosis method is proposed to transfer the model learnt from one machine (source domain) to another one (target domain). In the training process, labeled source data and unlabeled target data are used, which is very promising for real industrial applications. In this frame of transfer learningbased fault diagnosis, a cyclic spectrum correlation analysis method is firstly introduced to obtain order frequency maps for removing the influence of speed variation and revealing the hidden cyclic frequency of signals. Then, the Dynamic Adversarial Adaptation Network (DAAN) is introduced to transfer label information across machines. The proposed fault diagnosis method across machines is applied on two rolling element bearing datasets collected from two different test rigs. Experimental results demonstrate the effectiveness and superiority of the proposed method compared with stateof-the-art approaches.
- Open Access
- 7 January 2022
- Published by Association for Computing Machinery (ACM)
- p. 357-362
Measuring TIM coverage via Active Learning based YOLOv5 and image processing
Electronic devices, including chips, the foundation of modern information communication technology, have been valued by many governments and companies. In China, the government has published many policies to boost the development of electronic devices’ technology. However, as the performance of the electronic devices improves and the size decreased, so the heat generated by the electronic devices increases, which is a significant challenge for electronic devices to perform better. Then the thermal interface materials (TIM) are introduced to meet this challenge, which is used in electronic devices packaging to enable efficient heat dissipation. In the process of applying TIM to electronic devices packaging, some key parameters are essential for packaging engineers includes TIM coverage, we obtained TIM coverage SAT image via scanning acoustic tomography (SAT) for the purpose of post reliability testing, In this paper, we proposed a method to compute TIM coverage using active learning based YOLOv5, one of the most effective object detection algorithm and image processing for varied TIM coverage SAT images. The results of this method show that it could effectively and accurately compute the TIM coverage with a relatively small number of data, which could be a useful tool for packaging engineers to test the reliability of electronic devices packaging.
- 7 January 2022
- Published by Association for Computing Machinery (ACM)
- p. 357-362
Open AccessMeasuring TIM coverage via Active Learning based YOLOv5 and image processing
Electronic devices, including chips, the foundation of modern information communication technology, have been valued by many governments and companies. In China, the government has published many policies to boost the development of electronic devices’ technology. However, as the performance of the electronic devices improves and the size decreased, so the heat generated by the electronic devices increases, which is a significant challenge for electronic devices to perform better. Then the thermal interface materials (TIM) are introduced to meet this challenge, which is used in electronic devices packaging to enable efficient heat dissipation. In the process of applying TIM to electronic devices packaging, some key parameters are essential for packaging engineers includes TIM coverage, we obtained TIM coverage SAT image via scanning acoustic tomography (SAT) for the purpose of post reliability testing, In this paper, we proposed a method to compute TIM coverage using active learning based YOLOv5, one of the most effective object detection algorithm and image processing for varied TIM coverage SAT images. The results of this method show that it could effectively and accurately compute the TIM coverage with a relatively small number of data, which could be a useful tool for packaging engineers to test the reliability of electronic devices packaging.
- Top Cited Papers
- 12 August 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 18 (9) , 5760-5770
Simulation-Driven Domain Adaptation for Rolling Element Bearing Fault Diagnosis
State-of-the-art deep learning models remain data-intensive, requiring large training datasets to ensure their generalization ability. However, in industry, it is quite expensive or impractical to obtain massive training samples for condition monitoring practitioners. This article proposes a simulation-driven domain adaptation method to circumvent the data deficiency issue using physical-based simulations. A bearing phenomenological model is developed to generate simulated vibration signals. In the frame of domain adaptation transfer learning, a domain adversarial neural network (DANN) is proposed utilizing the simulated data as the source domain. The DANN can align the coarse supervised source domain data and the fine supervised target domain data to conduct adversarial training. Experimental results indicate that the proposed method can reach high classification accuracy using a small amount of real data. Compared to nonadapted and other transfer learning models, the proposed method demonstrates superior performance for bearing fault diagnosis, which is very promising for real industrial applications.
- 12 August 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 18 (9) , 5760-5770
Top Cited PapersSimulation-Driven Domain Adaptation for Rolling Element Bearing Fault Diagnosis
State-of-the-art deep learning models remain data-intensive, requiring large training datasets to ensure their generalization ability. However, in industry, it is quite expensive or impractical to obtain massive training samples for condition monitoring practitioners. This article proposes a simulation-driven domain adaptation method to circumvent the data deficiency issue using physical-based simulations. A bearing phenomenological model is developed to generate simulated vibration signals. In the frame of domain adaptation transfer learning, a domain adversarial neural network (DANN) is proposed utilizing the simulated data as the source domain. The DANN can align the coarse supervised source domain data and the fine supervised target domain data to conduct adversarial training. Experimental results indicate that the proposed method can reach high classification accuracy using a small amount of real data. Compared to nonadapted and other transfer learning models, the proposed method demonstrates superior performance for bearing fault diagnosis, which is very promising for real industrial applications.