Sheng Li
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
Sheng Li has a total of 52 co-authors. Here are the top co-authors by number of shared publications:
- Yadong Xu
The Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hong Kong Special Administrative Region of China, People’s Republic of China
12 shared publications Beibei Sun
7 shared publications- Ke Feng
Xi’an Jiaotong University
6 shared publications Yadong Xu
Department of Industrial and Systems Engineering , The Hong Kong Polytechnic University , Hong Kong 999077 , China , polyu.edu.hk
6 shared publicationsKe Feng
5 shared publicationsQing Ni
5 shared publications
Latest Publications
- Open Access
- 1 January 2026
- Published by Wiley in Shock and Vibration
- Vol. 2026 (1) , 5563009
Intelligent Gearbox Fault Diagnosis Under Variable Load Conditions via Zero‐Shot Learning
Deep learning has demonstrated remarkable potential for intelligent gearbox fault diagnosis; however, its effectiveness remains constrained by two critical challenges: the scarcity of labeled fault samples and the degradation of diagnostic accuracy under nonstationary operating conditions. Conventional supervised approaches rely heavily on abundant labeled data, which are often difficult or costly to obtain due to the rarity of severe fault occurrences. Moreover, variable load and speed fluctuations introduce nonstationary characteristics that impair the reliability of feature extraction and model generalization. To overcome these limitations, this study proposes a novel infrared thermal vision and acoustics‐driven multimodal zero‐shot diagnosis network (IRT‐AMZS). The framework introduces a multireceptive field‐based cross‐modal fusion (MRCF) module that adaptively extracts and integrates complementary fault features from infrared thermal and acoustic signals, enhancing robustness under varying operational conditions. Furthermore, by employing semantic fault prompts, the IRT‐AMZS framework effectively transfers diagnostic knowledge from observed to unobserved fault categories, thereby enabling accurate diagnosis even in the absence of labeled samples for unseen faults. Experimental evaluations on variable‐load gearbox datasets demonstrate that IRT‐AMZS significantly improves diagnostic accuracy and generalization compared to competing methods.
- 1 January 2026
- Published by Wiley in Shock and Vibration
- Vol. 2026 (1) , 5563009
Open AccessIntelligent Gearbox Fault Diagnosis Under Variable Load Conditions via Zero‐Shot Learning
Deep learning has demonstrated remarkable potential for intelligent gearbox fault diagnosis; however, its effectiveness remains constrained by two critical challenges: the scarcity of labeled fault samples and the degradation of diagnostic accuracy under nonstationary operating conditions. Conventional supervised approaches rely heavily on abundant labeled data, which are often difficult or costly to obtain due to the rarity of severe fault occurrences. Moreover, variable load and speed fluctuations introduce nonstationary characteristics that impair the reliability of feature extraction and model generalization. To overcome these limitations, this study proposes a novel infrared thermal vision and acoustics‐driven multimodal zero‐shot diagnosis network (IRT‐AMZS). The framework introduces a multireceptive field‐based cross‐modal fusion (MRCF) module that adaptively extracts and integrates complementary fault features from infrared thermal and acoustic signals, enhancing robustness under varying operational conditions. Furthermore, by employing semantic fault prompts, the IRT‐AMZS framework effectively transfers diagnostic knowledge from observed to unobserved fault categories, thereby enabling accurate diagnosis even in the absence of labeled samples for unseen faults. Experimental evaluations on variable‐load gearbox datasets demonstrate that IRT‐AMZS significantly improves diagnostic accuracy and generalization compared to competing methods.
- Top Cited Papers
- 29 October 2025
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 69
Time–frequency fully-connected graph neural network: An effective multiscale spatiotemporal dependency learning method for multisource machine fault diagnosis
- 29 October 2025
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 69
Top Cited PapersTime–frequency fully-connected graph neural network: An effective multiscale spatiotemporal dependency learning method for multisource machine fault diagnosis
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Multiview Contrastive Shapelet Learning: A Novel Semi-Supervised Approach for Explainable Machine Fault Diagnosis With Insufficient Annotated Data
Rotating machinery plays a vital role in modern industry, whose failures may cause sudden damage to the equipment and affect the reliability and safety of the whole mechanical system. Although numerous deep learning-based methods have emerged in industrial fault diagnosis, most of them suffer from two key limitations. Firstly, the majority of these techniques are predicated on the assumption of abundant data availability. In practical industrial settings, however, labeled samples are limited, rendering these methods ineffective under such constraints. Secondly, a significant limitation of these intelligent methods lies in their lack of interpretability, which hampers their applicability in high-reliability fault diagnosis systems. To address these problems, this paper proposes a multiview contrastive shapelet learning (MCSL) framework for semi-supervised fault diagnosis of rotating machinery. MCSL leverages both a supervised contrastive learning module (SCL) and a self-supervised contrastive learning module (SSCL) to comprehensively exploit labeled and unlabeled vibration signals. In SCL, a Shapelet learner block is used to extract key explainable patterns from labeled vibration signals. Subsequently, the SCL algorithm is employed to minimize the feature distance between the original sequence and the extracted shapelets. In SSCL, several data augmentation techniques are first applied. Then, the augmented data is fed into the Shapelet learner block. Furthermore, an interactive convolutional block is employed to extract multi-scale features. The parameters of the MCSL model are updated within an integrated training framework. Through experimental validation utilizing both public and self-collected datasets, it is evident that MCSL not only outperforms state-of-the-art methods in diagnostic accuracy but also demonstrates enhanced interpretability, underscoring its significant potential for industrial applications.
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Multiview Contrastive Shapelet Learning: A Novel Semi-Supervised Approach for Explainable Machine Fault Diagnosis With Insufficient Annotated Data
Rotating machinery plays a vital role in modern industry, whose failures may cause sudden damage to the equipment and affect the reliability and safety of the whole mechanical system. Although numerous deep learning-based methods have emerged in industrial fault diagnosis, most of them suffer from two key limitations. Firstly, the majority of these techniques are predicated on the assumption of abundant data availability. In practical industrial settings, however, labeled samples are limited, rendering these methods ineffective under such constraints. Secondly, a significant limitation of these intelligent methods lies in their lack of interpretability, which hampers their applicability in high-reliability fault diagnosis systems. To address these problems, this paper proposes a multiview contrastive shapelet learning (MCSL) framework for semi-supervised fault diagnosis of rotating machinery. MCSL leverages both a supervised contrastive learning module (SCL) and a self-supervised contrastive learning module (SSCL) to comprehensively exploit labeled and unlabeled vibration signals. In SCL, a Shapelet learner block is used to extract key explainable patterns from labeled vibration signals. Subsequently, the SCL algorithm is employed to minimize the feature distance between the original sequence and the extracted shapelets. In SSCL, several data augmentation techniques are first applied. Then, the augmented data is fed into the Shapelet learner block. Furthermore, an interactive convolutional block is employed to extract multi-scale features. The parameters of the MCSL model are updated within an integrated training framework. Through experimental validation utilizing both public and self-collected datasets, it is evident that MCSL not only outperforms state-of-the-art methods in diagnostic accuracy but also demonstrates enhanced interpretability, underscoring its significant potential for industrial applications.
- Top Cited Papers
- 8 April 2025
- Published by Elsevier in Journal of Manufacturing Systems
- Vol. 80, 563-577
Domain constrained cascadic multireceptive learning networks for machine health monitoring in complex manufacturing systems
- 8 April 2025
- Published by Elsevier in Journal of Manufacturing Systems
- Vol. 80, 563-577
Top Cited PapersDomain constrained cascadic multireceptive learning networks for machine health monitoring in complex manufacturing systems
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
Multiperspective Temporal Pooling Convolutional Neural Networks for Fault Diagnosis of Mechanical Transmission Systems
The rapid development of convolutional neural networks (CNN) has significantly contributed to the progress of intelligent fault diagnosis of mechanical transmission systems. Nevertheless, a significant number of prevailing CNN-based diagnostic models may suffer from two notable constraints. First, the existing models often employ fixed temporal pooling for feature extraction, which restrict their ability to effectively capture and analyze a comprehensive range of temporal information. Second, these models may struggle to precisely forecast the operational state of the monitored machinery amidst nonstationary circumstances, such as time-varying or disturbed environments. These challenges limit their feature extraction capabilities and hinder their practical implementation and utilization. To tackle the aforementioned issues, this study develops a multiperspective temporal pooling convolutional neural network (MTPCNN). The main contributions encompass: 1) A multikernel feature perception module (MFPM) and a balanced attention module (BAM) are established for multilevel information exploration and optimal feature selection; and 2) An innovative multiperspective temporal pooling learning (MTPL) strategy is introduced to aid the model in dynamically selecting the optimal temporal pooling method for the input data. A laboratory dataset collected from a gearbox fault simulator and an industrial dataset collected from a high-speed rail are used for the validation of the proposed approach. The extensive experimental results validate the superiority of the developed MTPCNN model over seven competitive approaches.
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
Multiperspective Temporal Pooling Convolutional Neural Networks for Fault Diagnosis of Mechanical Transmission Systems
The rapid development of convolutional neural networks (CNN) has significantly contributed to the progress of intelligent fault diagnosis of mechanical transmission systems. Nevertheless, a significant number of prevailing CNN-based diagnostic models may suffer from two notable constraints. First, the existing models often employ fixed temporal pooling for feature extraction, which restrict their ability to effectively capture and analyze a comprehensive range of temporal information. Second, these models may struggle to precisely forecast the operational state of the monitored machinery amidst nonstationary circumstances, such as time-varying or disturbed environments. These challenges limit their feature extraction capabilities and hinder their practical implementation and utilization. To tackle the aforementioned issues, this study develops a multiperspective temporal pooling convolutional neural network (MTPCNN). The main contributions encompass: 1) A multikernel feature perception module (MFPM) and a balanced attention module (BAM) are established for multilevel information exploration and optimal feature selection; and 2) An innovative multiperspective temporal pooling learning (MTPL) strategy is introduced to aid the model in dynamically selecting the optimal temporal pooling method for the input data. A laboratory dataset collected from a gearbox fault simulator and an industrial dataset collected from a high-speed rail are used for the validation of the proposed approach. The extensive experimental results validate the superiority of the developed MTPCNN model over seven competitive approaches.
- 22 February 2025
- Published by Elsevier in Energy and Buildings
- Vol. 333
A novel temporal domain adaptation framework for residential electricity consumption forecasting under incomplete information
- 22 February 2025
- Published by Elsevier in Energy and Buildings
- Vol. 333
A novel temporal domain adaptation framework for residential electricity consumption forecasting under incomplete information
- Top Cited Papers
- 1 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 33 (1) , 302-313
Composite Neuro-Fuzzy System-Guided Cross-Modal Zero-Sample Diagnostic Framework Using Multisource Heterogeneous Noncontact Sensing Data
Zero-sample diagnostic methods have gained recognition in addressing the scarcity of gearbox fault samples, thereby being regarded as a promising technique to guarantee gearbox safety. However, historical zero-sample approaches typically neglect the use of multi-modal non-contact sensing data and rarely consider the interpretability of the diagnostic process. This oversight limits their application in industrial environments that require high reliability or operate under extreme conditions. Therefore, this paper presents a composite neuro-fuzzy system-guided cross-modal zero-sample diagnostic framework, termed FCZD-IA, which employs infrared thermography and acoustic data to monitor gearbox conditions. Specifically, FCZD-IA uses a proposed composite neural system as a decision-maker in the diagnostic task, while integrating a deep backbone network to discriminatively learn high-level fault features from multi-modal data. Moreover, a specific training strategy is designed to guide the learning process of the FCZD-IA to promote robust and interpretable zero-sample diagnostics. Comprehensive experimental results validate the effectiveness of the proposed framework and its superiority over other competitive methods.
- 1 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 33 (1) , 302-313
Top Cited PapersComposite Neuro-Fuzzy System-Guided Cross-Modal Zero-Sample Diagnostic Framework Using Multisource Heterogeneous Noncontact Sensing Data
Zero-sample diagnostic methods have gained recognition in addressing the scarcity of gearbox fault samples, thereby being regarded as a promising technique to guarantee gearbox safety. However, historical zero-sample approaches typically neglect the use of multi-modal non-contact sensing data and rarely consider the interpretability of the diagnostic process. This oversight limits their application in industrial environments that require high reliability or operate under extreme conditions. Therefore, this paper presents a composite neuro-fuzzy system-guided cross-modal zero-sample diagnostic framework, termed FCZD-IA, which employs infrared thermography and acoustic data to monitor gearbox conditions. Specifically, FCZD-IA uses a proposed composite neural system as a decision-maker in the diagnostic task, while integrating a deep backbone network to discriminatively learn high-level fault features from multi-modal data. Moreover, a specific training strategy is designed to guide the learning process of the FCZD-IA to promote robust and interpretable zero-sample diagnostics. Comprehensive experimental results validate the effectiveness of the proposed framework and its superiority over other competitive methods.
- 26 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-11
Imbalanced Learning for Gearbox Fault Detection via Attention-Based Multireceptive Field Convolutional Neural Networks With an Adaptive Label Regulation Loss
Accurate gearbox fault identification is paramount for industrial production. In practice, gearboxes typically operate under normal conditions (rarely under faulty conditions), resulting in a long-tailed distribution of monitoring data. However, the majority of current algorithms are crafted based on the assumption of balanced sample distributions, which don’t correspond with the prevalent conditions encountered in actual industrial settings. To cope with this challenge, an attention-based multireceptive field convolutional neural network (AMFCN) is established in this paper. This study’s main contributions can be summarized as follows: 1) We introduce a global contextual attention module (GCAM) to instruct the model to focus on learning ample features; 2) We establish a hierarchical receptive field module (HRFM) to incorporate powerful multi-level learning capabilities into the AMFCN model; 3) We devise an adaptive label regulation loss (ALRL) to facilitate the model to obtain accurate fault identification results, particularly in situations with imbalanced data distributions. Two case studies show that the AMFCN model achieves 83.72% and 81.63% accuracy on two extremely imbalanced gearbox datasets, outperforming seven competitive algorithms.
- 26 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-11
Imbalanced Learning for Gearbox Fault Detection via Attention-Based Multireceptive Field Convolutional Neural Networks With an Adaptive Label Regulation Loss
Accurate gearbox fault identification is paramount for industrial production. In practice, gearboxes typically operate under normal conditions (rarely under faulty conditions), resulting in a long-tailed distribution of monitoring data. However, the majority of current algorithms are crafted based on the assumption of balanced sample distributions, which don’t correspond with the prevalent conditions encountered in actual industrial settings. To cope with this challenge, an attention-based multireceptive field convolutional neural network (AMFCN) is established in this paper. This study’s main contributions can be summarized as follows: 1) We introduce a global contextual attention module (GCAM) to instruct the model to focus on learning ample features; 2) We establish a hierarchical receptive field module (HRFM) to incorporate powerful multi-level learning capabilities into the AMFCN model; 3) We devise an adaptive label regulation loss (ALRL) to facilitate the model to obtain accurate fault identification results, particularly in situations with imbalanced data distributions. Two case studies show that the AMFCN model achieves 83.72% and 81.63% accuracy on two extremely imbalanced gearbox datasets, outperforming seven competitive algorithms.
- 31 July 2024
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture
- Vol. 240 (12) , 2458-2469
Digital twin-driven attention-guided convolutional networks for intelligent fault diagnosis across different domains
Fault diagnosis of rolling bearings is crucial in industries to ensure the reliability, safety, and economic feasibility of mechanical systems. Conventional data-driven methods for fault diagnosis depend on pre-existing datasets containing various failure modes for training. However, obtaining such datasets can be challenging, especially in critical industrial environments, hindering the applicability of these methods across different scenarios. To address this challenge, this research introduces a novel approach called a digital twin-driven attention-guided convolutional network (DTACN) for bearing fault diagnostics with limited data. This study offers two significant enhancements: (1) Development of an intricate digital twin model for bearings, which includes multi-scale Kinetic simulations of the bearing’s operational parameters. This model utilizes the architectural characteristics of the bearing and the severity of the failure to predict the vibratory system’s response accurately; and (2) Establishment of a domain adaptation-based framework employing deep convolutional models and attention mechanism to transfer knowledge from simulated signals to real signals. This framework facilitates efficient fault recognition even in scenarios with limited prior knowledge. The effectiveness of the proposed approach has been rigorously validated through comprehensive experiments.
- 31 July 2024
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture
- Vol. 240 (12) , 2458-2469
Digital twin-driven attention-guided convolutional networks for intelligent fault diagnosis across different domains
Fault diagnosis of rolling bearings is crucial in industries to ensure the reliability, safety, and economic feasibility of mechanical systems. Conventional data-driven methods for fault diagnosis depend on pre-existing datasets containing various failure modes for training. However, obtaining such datasets can be challenging, especially in critical industrial environments, hindering the applicability of these methods across different scenarios. To address this challenge, this research introduces a novel approach called a digital twin-driven attention-guided convolutional network (DTACN) for bearing fault diagnostics with limited data. This study offers two significant enhancements: (1) Development of an intricate digital twin model for bearings, which includes multi-scale Kinetic simulations of the bearing’s operational parameters. This model utilizes the architectural characteristics of the bearing and the severity of the failure to predict the vibratory system’s response accurately; and (2) Establishment of a domain adaptation-based framework employing deep convolutional models and attention mechanism to transfer knowledge from simulated signals to real signals. This framework facilitates efficient fault recognition even in scenarios with limited prior knowledge. The effectiveness of the proposed approach has been rigorously validated through comprehensive experiments.