Zhiheng Su

Zhiheng Su

Auto-generated profile, based on publicly available information

Zhiheng Su

University of Electronic Science and Technology of China, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

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

Representative Publications

View All Publications

19 July 2024
Verified
Conference Paper
Open access
ZSZhiheng Su
Zhiheng Su
JZJiyang Zhang
Jiyang Zhang
PLPenglong Lian
Penglong Lian
JZJianxiao Zou
Jianxiao Zou
SFShicai Fan
Shicai Fan
Domain generalization aims to address the problem of model transfer to the unknown target domain. However, existing domain generalization based fault diagnosis methods still suffer from insufficient exploration of domain-invariant features and lack consideration of the correlation between target data and different source data when solving the multi-source domain generalization problem. To address these issues, a weighted ensemble domain generalization (WEDG) method via multi-level invariant feature learning is proposed in this paper. By leveraging domain-level domain-invariant feature learning based on higher-order moment matching to adapt the overall data distribution of different source domains, sample-level domain-invariant feature learning is introduced to deeply learn domain-invariant features of different class data across domains through random cross-domain sample mixing. Additionally, considering the different similarities between target data and different source domains, the proposed method separately trains extra classifiers for each source domain and integrates the prediction results of all classifiers based on the similarity between target data and source domains to enhance the final generalization ability of the model. Experimental results on the CWRU and Gearbox datasets demonstrate that the WEDG outperforms other methods on domain generalization fault diagnosis problem in both bearings and gears.

25 July 2024
Verified
Conference Paper
Open access
JZJiyang Zhang
Jiyang Zhang
JZJianxiao Zou
Jianxiao Zou
ZSZhiheng Su
Zhiheng Su
PLPenglong Lian
Penglong Lian
SFShicai Fan
Shicai Fan
Deep learning-based intelligent fault diagnosis under noise environments often experiences significant performance deterioration, since noises can obscure useful fault pattern information within feature space. Recently, several deep neural networks with denoising abilities have performed well in fault diagnosis under the single noise environment with known distribution (i.e., the additive Gaussian white noise). However, since above methods ignore the estimation of complex noise and the denoising capability of the single branch network used are insufficient, applying these methods in complex noise environments is difficult, where noise does not simply follow a single known distribution. To address this issue, a novel dual temporal convolutional neural network with noise estimation and attention mechanism (DTCNA-Net) is proposed. Specifically, DTCNA-Net consists of a noise estimation subnetwork and a dual temporal convolutional network (Dual-TCN) with attention mechanism. The noise estimation subnetwork is firstly utilized to estimate complex noise, which can be used to guide denoising and feature extraction. Then, raw data and the estimated noise are combined as the input of Dual-TCN. Dual-TCN includes two parallel TCN branches with channel and spatial attention mechanisms, respectively. By the effective denoising capability of attention mechanism and the discriminative feature extraction ability of TCN, Dual-TCN can effectively eliminate noises within feature spaces from channel and spatial dimensions, and extract complementary discriminative features. Finally, by integrating features extracted by the two network branches, DTCNA-Net can achieve good diagnosis performance under complex noise environment. Experiments on two fault diagnosis datasets show the effectiveness of DTCNA-Net.