PAC-GAN: A Physics-Aware Contrastive Generative Adversarial Network for Few-Shot Rolling Bearing Fault Diagnosis under Noisy Environments

Abstract As core components of rotating machinery, the reliable diagnosis of rolling bearings is crucial for safe equipment operation and maintenance. In real industrial scenarios, the scarcity of fault samples and the interference of strong background noise in vibration signals pose challenges for traditional data-driven methods, such as difficulty in feature ex-traction and insufficient physical authenticity of generated samples. To address this, this paper proposes a Physically Aware Contrastive Generative Adversarial Network (PAC-GAN) to achieve highly robust intelligent diagnostics under conditions of sparse samples and strong noise. Based on the Auxiliary Classification Generative Adversarial Network (ACGAN) framework, this method synergistically enhances diagnostic performance through three mechanisms: First, it employs a complex morlet wavelet continuous wavelet transform to convert one-dimensional vibration signals into high-fidelity time-frequency images, preserving transient fault features; Second, a spectral consistency loss is introduced in the generator, embedding strong constraints from frequency-domain physical priors into the adversarial training to ensure synthetic samples align with fault evolution mechanisms. Finally, the discriminator integrates a Convolutional Block Attention Module (CBAM) with a supervised contrastive learning mechanism, enabling adaptive focusing on fault-sensitive regions and structural optimization of the feature space, respectively. This enhances intra-class compactness and inter-class separability of features. 10-shot few-shot experiments on the CWRU and JNU bearing datasets demonstrate that the proposed method achieves an average diagnostic accuracy of 99.75%. It significantly outperforms the SVM baseline, yields classification accuracy on par with ResNet18 with only 29% of its parameters, and exhibits marked advantages in noise robustness, cross-domain generalization capability and inference efficiency. Under strong noise interference and cross-condition transfer tests, PAC-GAN demonstrates outstanding robustness, generalization capability, and low computational complexity, possessing clear engineering application value. It provides a reliable solution for condition monitoring of rotating machinery under data-constrained conditions.

Authors

Institutions

Publication Details

Journal
Engineering Research Express
Published
2026-11-17
DOI
https://doi.org/10.1088/2631-8695/aec375
Primary Topic
Machine Fault Diagnosis Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

PAC-GAN: A Physics-Aware Contrastive Generative Adversarial Network for Few-Shot Rolling Bearing Fault Diagnosis under Noisy Environments

Nanmu Hui, Binbin Tu, Yan Huo, Keming Chao et al.
3 links
Engineering Research Express
Machine Fault Diagnosis Techniques
article

PAC-GAN: A Physics-Aware Contrastive Generative Adversarial Network for Few-Shot Rolling Bearing Fault Diagnosis under Noisy Environments

Nanmu Hui, Binbin Tu, Yan Huo, Keming Chao, Yinghong Xie, Binglin Wen, Xuesheng Zhao
article en

Abstract

Abstract As core components of rotating machinery, the reliable diagnosis of rolling bearings is crucial for safe equipment operation and maintenance. In real industrial scenarios, the scarcity of fault samples and the interference of strong background noise in vibration signals pose challenges for traditional data-driven methods, such as difficulty in feature ex-traction and insufficient physical authenticity of generated samples. To address this, this paper proposes a Physically Aware Contrastive Generative Adversarial Network (PAC-GAN) to achieve highly robust intelligent diagnostics under conditions of sparse samples and strong noise. Based on the Auxiliary Classification Generative Adversarial Network (ACGAN) framework, this method synergistically enhances diagnostic performance through three mechanisms: First, it employs a complex morlet wavelet continuous wavelet transform to convert one-dimensional vibration signals into high-fidelity time-frequency images, preserving transient fault features; Second, a spectral consistency loss is introduced in the generator, embedding strong constraints from frequency-domain physical priors into the adversarial training to ensure synthetic samples align with fault evolution mechanisms. Finally, the discriminator integrates a Convolutional Block Attention Module (CBAM) with a supervised contrastive learning mechanism, enabling adaptive focusing on fault-sensitive regions and structural optimization of the feature space, respectively. This enhances intra-class compactness and inter-class separability of features. 10-shot few-shot experiments on the CWRU and JNU bearing datasets demonstrate that the proposed method achieves an average diagnostic accuracy of 99.75%. It significantly outperforms the SVM baseline, yields classification accuracy on par with ResNet18 with only 29% of its parameters, and exhibits marked advantages in noise robustness, cross-domain generalization capability and inference efficiency. Under strong noise interference and cross-condition transfer tests, PAC-GAN demonstrates outstanding robustness, generalization capability, and low computational complexity, possessing clear engineering application value. It provides a reliable solution for condition monitoring of rotating machinery under data-constrained conditions.

Engineering Research Express
Shenyang University (CN)
Machine Fault Diagnosis Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon