Wentao Zhang

Wentao Zhang

Auto-generated profile, based on publicly available information

Wentao Zhang

Tongji University

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
9 May 2022 in IEEE Access
Verified
Journal Article
Open access
WZWentao Zhang
Wentao Zhang
TZTing Zhang
Ting Zhang
GCGuohua Cui
Guohua Cui
YPYing Pan
Ying Pan
With the development of automated and integrated large-scale industrial systems, accurate and effective fault diagnosis methods are required to ensure the security and reliability of running mechanical equipment. Due to the time consumption and poor generalization performance of conventional machine learning-based methods, deep learning (DL)-based methods have wider application prospects due to their end-to-end architectural properties. However, in the DL models, problems such as a large number of trainable parameters, complicated hyperparameter tuning, and initialization instability increase the difficulty of model training and limit higher performance. To address these disadvantages of the DL method, we proposed a novel DL framework by applying convolutional neural networks (CNNs) based on the optimization of transfer learning (TL). TL can help the model achieve higher precision with less computational cost by transferring low-level features and fine-tuning high-level layers. In addition, data processing was implemented using continuous wavelet transformation (CWT) to convert vibration signals into 2-D images, and support vector machines (SVM) were employed to replace the fully connected layers for better classification. As a result, the proposed method was superior to the classical deep architecture trained from scratch. The performance of the proposed method is analyzed by presenting testing reports, convergence curves, and confusion matrixes. Moreover, experiments comprised of cross-domain diagnosis, simulated composite fault detection, and performance comparison on seven mechanical datasets, including bearings, gearboxes, and rotors, are presented. Based on these results, it can be observed that our method achieved the highest accuracy under various conditions.

14 December 2022 in Machines
Verified
Journal Article
Open access
PDF file
WZWentao Zhang
Wentao Zhang
TZTing Zhang
Ting Zhang
GCGuohua Cui
Guohua Cui
YPYing Pan
Ying Pan
Industrial robots typically perform a variety of tasks and occupy critical positions in modern manufacturing fields. When certain failures occur in the internal structures of robots, it tends to result in significant financial loss and the consumption of human resources. As a result, timely and effective fault diagnosis is critical to ensuring the safe operation of robots. Deep learning-based methods are currently being widely used by researchers to address some shortcomings of traditional methods. However, due to realistic factor limitations, few researchers take the failure pattern of rotating machinery and the location of fault joints into account at the same time, so the fault types of multi-joint robots are not thoroughly investigated. In this case, we proposed a dynamic simulation method that considers the effects of bearing failures at various faulty joint locations and makes it possible to investigate more possible failure cases of multi-joint robots. In addition, we used LSTM and DCNN to perform staged fault diagnosis tasks, allowing us to gradually locate faulty joints and investigate detailed failure forms. According to the experimental results, distinguishable vibration signals corresponding to various fault states are successfully obtained, and our implemented algorithms are validated for their advanced performance in diagnosis tasks via comparative experiments.

29 July 2026 in Engineering Research Express
Verified
Journal Article
Closed access
WZWentao Zhang
Wentao Zhang
XQXianrong QIN
Xianrong QIN
QYQiong Yang
Qiong Yang
QZQing Zhang
Qing Zhang
YSYuantao Sun
Yuantao Sun
In structural health monitoring, vibration measurements are typically collected from spatially distributed sensors to assess structural conditions. However, existing approaches typically treat multi-channel measurements as independent temporal sequences or simply stack them as network inputs, without explicitly encoding the spatial relationships among distributed sensors, leading to suboptimal utilization of measurement information. To address this limitation, the spatial graph is introduced to explicitly represent sensor positions and distances. Accordingly, this study proposes a spatial–temporal representation learning framework for multi-sensor measurements, integrating convolutional neural networks (CNNs) and graph convolutional networks (GCNs) to exploit inter-sensor spatial dependencies while preserving channel-wise temporal characteristics. A depthwise CNN module is employed to extract temporal features from individual sensor signals, followed by a GCN for spatial aggregation. A physically informed adjacency matrix is constructed based on Manhattan distances and a Gaussian weighting scheme to reflect the inter-sensor spatial dependencies. This design preserves channel-wise temporal information while incorporating physically interpretable spatial relationships and relative distance among sensors. The proposed framework is evaluated on two benchmark structures under varying damage severity, data availability, noise levels, and ablation settings. Experimental results demonstrate that the approach achieves improved classification accuracy, enhanced robustness to measurement noise, and stable generalization performance under varying operating conditions. This study provides a measurement-oriented perspective for representing and exploiting spatially distributed sensing data, enhancing the observability and interpretability of structural states.

24 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
XQXianrong QIN
Xianrong QIN
QYQiong Yang
Qiong Yang
WZWentao Zhang
Wentao Zhang
JSJiaxin Shen
Jiaxin Shen
Multi-channel vibration monitoring generates large volumes of data, increasing storage and transmission demands in structural health monitoring (SHM). Compressive sensing (CS) can alleviate this burden, but multi-channel reconstruction should recover individual responses while preserving inter-channel relationships. Accordingly, an Unfolding-based Multi-channel Spectral Attention Compressive Sensing Network (UMSA-CSNet) is developed for joint reconstruction of multi-channel vibration signals. The network follows an ISTA-inspired unfolding structure that combines measurement-consistency updates with learnable prior modeling, with complex Fourier spectra represented by concatenated real and imaginary components. Each stage incorporates a cross-sensor spectral prior module based on self-attention and learnable soft-thresholding to exploit inter-sensor dependencies and spectral sparsity. Evaluation uses a 3-DOF numerical system, the Los Alamos National Laboratory three-story frame dataset, and field-measured wind turbine tower responses. In the numerical example, UMSA-CSNet achieves time-domain reconstruction accuracy comparable to ISTA-1DNet but yields a lower spectral error. On the two measured datasets, UMSA-CSNet produces lower reconstruction and inter-channel correlation errors than the comparison methods; the modal characteristics identified from the reconstructed three-story frame responses are also better preserved. Removing cross-sensor interaction causes the largest degradation in ablation, whereas the effects of the amplitude-spectrum constraint and learnable thresholds are smaller. UMSA-CSNet requires more CPU inference time than ISTA-1DNet, but remains substantially faster than FISTA and MT-BCS under the reported timing conditions. These results indicate that cross-sensor spectral modeling can improve compressed reconstruction of multi-channel vibration measurements while retaining information relevant to subsequent structural analysis.