Zhiyi He
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
Zhiyi He has a total of 55 co-authors. Here are the top co-authors by number of shared publications:
- Haidong Shao
Hunan University
8 shared publications Junsheng Cheng
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, College of Mechanical and Vehicle Engineering, Hunan University, Changsha, China
6 shared publicationsYu Yang
5 shared publicationsHaidong Shao
4 shared publications- Hongwei Hu
College of Mechanical and Vehicle Engineering, Changsha University of Science and Technology, Changsha 410114, China; Hunan Province University Key Laboratory of Intelligent Testing and Control Technology for Engineering Equipment, Changsha 410114, China
3 shared publications - Ming Zeng
School of Mechanical Engineering and Electronic Information, China University of Geosciences, Wuhan, China
3 shared publications
Latest Publications
- 22 September 2026
- Published by SAGE Publications in Structural Health Monitoring
An interpretable graph neural network inspired by graph signal processing for mechanical fault diagnosis
As mechanical equipment grows more complex, multi-sensor monitoring and data fusion are essential for accurate fault diagnosis. While graph neural networks (GNNs) are popularly used in this field, their performance evaluations remain largely empirical, lacking a unified theoretical framework to explain and assess their underlying propagation mechanisms. To address this, we propose a unified GNN framework from the perspective of graph signal processing (GSP). We mathematically demonstrate that each propagation layer in representative GNNs (including graph convolutional networks, graph attention networks, and GraphSAGE) is equivalent or approximately equivalent to performing a single-step proximal gradient descent to solve a specific instance within this framework. The framework also provides an interpretable foundation, revealing the commonalities, differences, and limitations of representative GNNs. To overcome the limitations, we design a more principled GSP optimization objective by removing the activation regularizer and adopting a multi-step gradient descent strategy. Using algorithm unrolling, we derive a novel architecture named gradient descent unrolled GNN (GDUGNN), which adopts a decoupled architectural paradigm where feature transformation is performed by a convolution and max-pooling module, and feature aggregation is implemented by unrolling the gradient descent algorithm into multiple propagation layers with learnable step sizes. This design naturally introduces skip connections from both the current and the first-layer inputs, effectively mitigating over-smoothing while ensuring inherent interpretability. A fault diagnosis model based on GDUGNN is developed and evaluated on a wind turbine simulator and a rock drill. Experiments demonstrate that GDUGNN achieves superior diagnostic performance with stronger resistance to overfitting and over-smoothing compared to mainstream GNNs and their variants. Our code will be available at https://github.com/CUG-FDGroup .
- 22 September 2026
- Published by SAGE Publications in Structural Health Monitoring
An interpretable graph neural network inspired by graph signal processing for mechanical fault diagnosis
As mechanical equipment grows more complex, multi-sensor monitoring and data fusion are essential for accurate fault diagnosis. While graph neural networks (GNNs) are popularly used in this field, their performance evaluations remain largely empirical, lacking a unified theoretical framework to explain and assess their underlying propagation mechanisms. To address this, we propose a unified GNN framework from the perspective of graph signal processing (GSP). We mathematically demonstrate that each propagation layer in representative GNNs (including graph convolutional networks, graph attention networks, and GraphSAGE) is equivalent or approximately equivalent to performing a single-step proximal gradient descent to solve a specific instance within this framework. The framework also provides an interpretable foundation, revealing the commonalities, differences, and limitations of representative GNNs. To overcome the limitations, we design a more principled GSP optimization objective by removing the activation regularizer and adopting a multi-step gradient descent strategy. Using algorithm unrolling, we derive a novel architecture named gradient descent unrolled GNN (GDUGNN), which adopts a decoupled architectural paradigm where feature transformation is performed by a convolution and max-pooling module, and feature aggregation is implemented by unrolling the gradient descent algorithm into multiple propagation layers with learnable step sizes. This design naturally introduces skip connections from both the current and the first-layer inputs, effectively mitigating over-smoothing while ensuring inherent interpretability. A fault diagnosis model based on GDUGNN is developed and evaluated on a wind turbine simulator and a rock drill. Experiments demonstrate that GDUGNN achieves superior diagnostic performance with stronger resistance to overfitting and over-smoothing compared to mainstream GNNs and their variants. Our code will be available at https://github.com/CUG-FDGroup .
- 4 March 2026
- Published by SAGE Publications in Structural Health Monitoring
Replay-free evolutionary classifier for few-shot continuous fault diagnosis of wind turbine
In industrial settings, it is important to diagnose new faults from the updated data stream. The model that learns incrementally by taking in new data can meet this requirement. However, the problems of time-consuming iterative updates and catastrophic forgetting in traditional incremental learning methods greatly reduce the diagnostic efficiency. To solve these problems, this paper introduces a replay-free evolutionary classifier (RFEC) architecture. The core of this strategy is to optimize the classification decision process by reconstructing the association between classification weights and feature vectors, with effective constraints on the objective function. Also, a nonparametric prototype classifier is introduced to ensure timely online fault diagnosis. This classifier uses classification prototypes and initial weights to allow quick updates during later incremental phases. The experimental results of the wind turbine simulation experimental platform show that RFEC has better accuracy and real-time update efficiency than other popular incremental learning methods. Specifically, it achieves an average accuracy of 90.57% and a dramatically reduced total training time of approximately 0.06 seconds during the incremental update stage.
- 4 March 2026
- Published by SAGE Publications in Structural Health Monitoring
Replay-free evolutionary classifier for few-shot continuous fault diagnosis of wind turbine
In industrial settings, it is important to diagnose new faults from the updated data stream. The model that learns incrementally by taking in new data can meet this requirement. However, the problems of time-consuming iterative updates and catastrophic forgetting in traditional incremental learning methods greatly reduce the diagnostic efficiency. To solve these problems, this paper introduces a replay-free evolutionary classifier (RFEC) architecture. The core of this strategy is to optimize the classification decision process by reconstructing the association between classification weights and feature vectors, with effective constraints on the objective function. Also, a nonparametric prototype classifier is introduced to ensure timely online fault diagnosis. This classifier uses classification prototypes and initial weights to allow quick updates during later incremental phases. The experimental results of the wind turbine simulation experimental platform show that RFEC has better accuracy and real-time update efficiency than other popular incremental learning methods. Specifically, it achieves an average accuracy of 90.57% and a dramatically reduced total training time of approximately 0.06 seconds during the incremental update stage.
- 23 January 2026
- Published by SAGE Publications in Structural Health Monitoring
Multivariate multiscale diversity entropy and its application to rotating machinery fault diagnosis
Entropy-based methods effectively assess the irregularity of time series to obtain underlying fault features from vibration signals. Multiscale diversity entropy (MDE) has been widely applied in feature extraction due to the high consistency between the entropy value and sequence complexity. However, its sequential processing of multivariate signals inherently ignores the valuable intersignal relationships, leading to a loss of critical information embedded in their correlations. To address this drawback, considering the dynamic and complex changes occurring both within and between channels, the multivariate MDE (MMDE) is proposed as a new complexity measurement method, which can capture their dynamic relationships and coupling feature of multivariate signals. Based on this, a fault diagnosis method for rotating machinery is proposed, which integrates MMDE and support vector machine. Both simulation and experimental results validate that the proposed MMDE excels at detecting dynamic changes in signals. Furthermore, the presented diagnostic method demonstrates superior capability over other feature extraction techniques.
- 23 January 2026
- Published by SAGE Publications in Structural Health Monitoring
Multivariate multiscale diversity entropy and its application to rotating machinery fault diagnosis
Entropy-based methods effectively assess the irregularity of time series to obtain underlying fault features from vibration signals. Multiscale diversity entropy (MDE) has been widely applied in feature extraction due to the high consistency between the entropy value and sequence complexity. However, its sequential processing of multivariate signals inherently ignores the valuable intersignal relationships, leading to a loss of critical information embedded in their correlations. To address this drawback, considering the dynamic and complex changes occurring both within and between channels, the multivariate MDE (MMDE) is proposed as a new complexity measurement method, which can capture their dynamic relationships and coupling feature of multivariate signals. Based on this, a fault diagnosis method for rotating machinery is proposed, which integrates MMDE and support vector machine. Both simulation and experimental results validate that the proposed MMDE excels at detecting dynamic changes in signals. Furthermore, the presented diagnostic method demonstrates superior capability over other feature extraction techniques.
- 21 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Reliability
- Vol. 74 (4) , 5544-5553
Uncertainty NVAE-SVDD Based on Monte Carlo Dropout for Trustworthy Detection of Mechanical Failures
Most mechanical fault detection methods rely on deterministic feature extraction and fail to model prediction uncertainty, often leading to overconfident errors when facing unknown or borderline faults in complex conditions. This compromises the reliability of detection outcomes. To address these issues, this article proposes a trustworthy mechanical fault detection method based on uncertainty modeling using Monte Carlo (MC) dropout within the nouveau variational autoencoder support vector data description (NVAE-SVDD) framework. The method employs a NVAE with a hierarchical latent variable structure as the feature extractor. By explicitly incorporating the MC dropout strategy, it enables effective quantification of uncertainty during the prediction process. Meanwhile, SVDD is used to construct a compact boundary around normal samples in the NVAE latent space, enabling efficient fault identification. The proposed method not only accurately identifies normal conditions but also fully leverages predictive uncertainty to distinguish boundary anomalies and unknown faults. Experimental results show that the proposed method outperforms traditional one-class models and uncertainty-free baselines in accuracy and robustness, greatly improving fault detection reliability.
- 21 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Reliability
- Vol. 74 (4) , 5544-5553
Uncertainty NVAE-SVDD Based on Monte Carlo Dropout for Trustworthy Detection of Mechanical Failures
Most mechanical fault detection methods rely on deterministic feature extraction and fail to model prediction uncertainty, often leading to overconfident errors when facing unknown or borderline faults in complex conditions. This compromises the reliability of detection outcomes. To address these issues, this article proposes a trustworthy mechanical fault detection method based on uncertainty modeling using Monte Carlo (MC) dropout within the nouveau variational autoencoder support vector data description (NVAE-SVDD) framework. The method employs a NVAE with a hierarchical latent variable structure as the feature extractor. By explicitly incorporating the MC dropout strategy, it enables effective quantification of uncertainty during the prediction process. Meanwhile, SVDD is used to construct a compact boundary around normal samples in the NVAE latent space, enabling efficient fault identification. The proposed method not only accurately identifies normal conditions but also fully leverages predictive uncertainty to distinguish boundary anomalies and unknown faults. Experimental results show that the proposed method outperforms traditional one-class models and uncertainty-free baselines in accuracy and robustness, greatly improving fault detection reliability.
- 13 September 2025
- Published by SAGE Publications in Structural Health Monitoring
An interpretable algorithm unrolling network inspired by group-based sparse coding for mechanical fault diagnosis
Intelligent fault diagnosis (IFD) can provide reliable maintenance decision-making for mechanical systems. Although generic deep learning models have been applied successfully in IFD, most are purely data-driven and lack interpretability in architectures and features. To this end, an interpretable IFD model is proposed by using algorithm unrolling from a signal processing perspective. For a group-based sparse coding (GSC) denoising problem, we first propose a new denoising algorithm called EGISTA, which combines the e xtragradient and the g roup-based i terative s hrinkage- t hresholding a lgorithm (EGISTA). Then, an algorithm unrolling network called EGISTA-Net is derived from EGISTA. EGISTA-Net has an interpretable encoder–decoder architecture since the architecture is constructed under the guidance of the GSC denoising algorithm. We apply EGISTA-Net to two IFD tasks. In ablation studies, we show that the extragradient can speed up EGISTA-Net’s convergence and improve noise robustness, while the decoder can enhance the interpretability of extracted features. We also interpret the feature extraction process of EGISTA-Net by visualizing the signals reconstructed using layerwise features. Compared with state-of-the-art IFD models, EGISTA-Net achieves higher diagnostic accuracy with limited training set sizes and shows stronger noise robustness. Our experiment results demonstrate that EGISTA-Net is a promising interpretable IFD model.
- 13 September 2025
- Published by SAGE Publications in Structural Health Monitoring
An interpretable algorithm unrolling network inspired by group-based sparse coding for mechanical fault diagnosis
Intelligent fault diagnosis (IFD) can provide reliable maintenance decision-making for mechanical systems. Although generic deep learning models have been applied successfully in IFD, most are purely data-driven and lack interpretability in architectures and features. To this end, an interpretable IFD model is proposed by using algorithm unrolling from a signal processing perspective. For a group-based sparse coding (GSC) denoising problem, we first propose a new denoising algorithm called EGISTA, which combines the e xtragradient and the g roup-based i terative s hrinkage- t hresholding a lgorithm (EGISTA). Then, an algorithm unrolling network called EGISTA-Net is derived from EGISTA. EGISTA-Net has an interpretable encoder–decoder architecture since the architecture is constructed under the guidance of the GSC denoising algorithm. We apply EGISTA-Net to two IFD tasks. In ablation studies, we show that the extragradient can speed up EGISTA-Net’s convergence and improve noise robustness, while the decoder can enhance the interpretability of extracted features. We also interpret the feature extraction process of EGISTA-Net by visualizing the signals reconstructed using layerwise features. Compared with state-of-the-art IFD models, EGISTA-Net achieves higher diagnostic accuracy with limited training set sizes and shows stronger noise robustness. Our experiment results demonstrate that EGISTA-Net is a promising interpretable IFD model.
- 5 August 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (8) , 086117
Noise-robust rotating machinery fault diagnosis method based on enhanced graph isomorphism network with multi-sensor signals
Graph neural network-based multi-sensor fault diagnosis approaches have found widespread application in rotating machinery. However, most existing methods overlook inter-graph relationships and are susceptible to sensor noise, failing to effectively capture the similarities or differences between samples of different fault types. To address the above issues, a novel noise-robust rotating machinery fault diagnosis scheme based on an enhanced graph isomorphism network with multi-sensor signals is proposed. The proposed method employs a graph feature extractor consisting of multiple layers of GINs to generate the initial graph representation of multi-source signals from rotating machinery. Based on this, a new relational feature extractor is designed to capture the relationships between different fault samples. Furthermore, a novel dynamic feedback training strategy is introduced to optimize the raw graph representation continuously by fully leveraging information from existing fault samples. Extensive experimental results demonstrate the effectiveness of this method, with the results indicating that it outperforms several benchmark methods across various noise conditions.
- 5 August 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (8) , 086117
Noise-robust rotating machinery fault diagnosis method based on enhanced graph isomorphism network with multi-sensor signals
Graph neural network-based multi-sensor fault diagnosis approaches have found widespread application in rotating machinery. However, most existing methods overlook inter-graph relationships and are susceptible to sensor noise, failing to effectively capture the similarities or differences between samples of different fault types. To address the above issues, a novel noise-robust rotating machinery fault diagnosis scheme based on an enhanced graph isomorphism network with multi-sensor signals is proposed. The proposed method employs a graph feature extractor consisting of multiple layers of GINs to generate the initial graph representation of multi-source signals from rotating machinery. Based on this, a new relational feature extractor is designed to capture the relationships between different fault samples. Furthermore, a novel dynamic feedback training strategy is introduced to optimize the raw graph representation continuously by fully leveraging information from existing fault samples. Extensive experimental results demonstrate the effectiveness of this method, with the results indicating that it outperforms several benchmark methods across various noise conditions.
- 29 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (12) , 22057-22067
A Few-Shot Domain Class Incremental Motor Fault Diagnosis Method With Dynamic Anchor Prototyping
Domain incremental class learning has been studied in the fault diagnosis of motors and other mechanical equipment. However, in domain-incremental few-shot scenarios, existing incremental methods, troubled by feature confusion and overfitting, lead to low fault-classification accuracy of mechanical systems under dynamic conditions and fail to meet the need for real-time monitoring of operating-condition switches. Therefore, this paper proposes a few-shot domain class incremental motor fault diagnosis method with dynamic anchor prototyping (DAP). In the initial training stage, the anchor loss constrains intra-class feature compactness. Meanwhile, a domain-decoupled and class-separable feature space is built to suppress catastrophic forgetting from old-new knowledge confusion in the incremental stage. In the incremental stage, a dynamic class prototype classifier is designed to adaptively align few-shot features based on the initial domain pre-training parameters to achieve efficient extraction of cross-domain key discriminative features without back-propagation iterations, which significantly reduces the training time. The motor fault diagnosis experimental results demonstrate that the proposed method achieves higher accuracy with the lower forgetting rate in comparison to traditional incremental learning methods.
- 29 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (12) , 22057-22067
A Few-Shot Domain Class Incremental Motor Fault Diagnosis Method With Dynamic Anchor Prototyping
Domain incremental class learning has been studied in the fault diagnosis of motors and other mechanical equipment. However, in domain-incremental few-shot scenarios, existing incremental methods, troubled by feature confusion and overfitting, lead to low fault-classification accuracy of mechanical systems under dynamic conditions and fail to meet the need for real-time monitoring of operating-condition switches. Therefore, this paper proposes a few-shot domain class incremental motor fault diagnosis method with dynamic anchor prototyping (DAP). In the initial training stage, the anchor loss constrains intra-class feature compactness. Meanwhile, a domain-decoupled and class-separable feature space is built to suppress catastrophic forgetting from old-new knowledge confusion in the incremental stage. In the incremental stage, a dynamic class prototype classifier is designed to adaptively align few-shot features based on the initial domain pre-training parameters to achieve efficient extraction of cross-domain key discriminative features without back-propagation iterations, which significantly reduces the training time. The motor fault diagnosis experimental results demonstrate that the proposed method achieves higher accuracy with the lower forgetting rate in comparison to traditional incremental learning methods.
- 24 April 2025
- Published by Elsevier in Digital Signal Processing
- Vol. 164
A new gear weak feature extraction method based on modified symplectic geometry mode decomposition
- 24 April 2025
- Published by Elsevier in Digital Signal Processing
- Vol. 164
A new gear weak feature extraction method based on modified symplectic geometry mode decomposition
- 16 May 2024
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 35 (8) , 086010
Motor online novelty detection scheme based on one-class hyperdisk
Online novelty detection is of great importance in the series production of motors. This study developed an online novelty detection scheme for motors based on a one-class hyperdisk (OCHD) model. In the OCHD approach, the decision boundary is estimated using a hyperdisk (HD), which is derived from the training sample set. The HD model addresses the underestimation issue commonly associated with convex-hull-based methods by providing a more accurate estimation of the class region. Furthermore, an optimal separating hyperplane is constructed at the nearest point on the HD by solving a quadratically constrained quadratic program problem. Statistical features refined by the Laplacian score are employed in the proposed novelty detection scheme. This study introduces an online novelty detection scheme for assessing motor quality in actual series production. The test results from the offline experiment demonstrate the superiority of the OCHD method. Datasets collected at the end of the production line were evaluated using the proposed novelty detection scheme. The inspection results for motor components confirm that the proposed method effectively identifies faulty motors during the series production process.
- 16 May 2024
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 35 (8) , 086010
Motor online novelty detection scheme based on one-class hyperdisk
Online novelty detection is of great importance in the series production of motors. This study developed an online novelty detection scheme for motors based on a one-class hyperdisk (OCHD) model. In the OCHD approach, the decision boundary is estimated using a hyperdisk (HD), which is derived from the training sample set. The HD model addresses the underestimation issue commonly associated with convex-hull-based methods by providing a more accurate estimation of the class region. Furthermore, an optimal separating hyperplane is constructed at the nearest point on the HD by solving a quadratically constrained quadratic program problem. Statistical features refined by the Laplacian score are employed in the proposed novelty detection scheme. This study introduces an online novelty detection scheme for assessing motor quality in actual series production. The test results from the offline experiment demonstrate the superiority of the OCHD method. Datasets collected at the end of the production line were evaluated using the proposed novelty detection scheme. The inspection results for motor components confirm that the proposed method effectively identifies faulty motors during the series production process.