Adversarial Adaptive 1-D CNNs for Robust Bearing Fault Diagnosis
Abstract: Intelligent fault diagnosis models often struggle when deployment conditions differ from training environments, a phenomenon known as domain shift. This paper presents A2CNN (Adversarial Adaptive 1-D Convolutional Neural Network), a framework that leverages unsupervised domain adaptation to maintain high diagnostic accuracy across varying motor loads and speeds. By employing an adversarial training strategy, A2CNN learns domain-invariant features, achieving an average accuracy of 99.21% on the CWRU dataset without requiring labeled data from the target domain.
Introduction
Rolling element bearings are vital components in industrial machinery. While deep learning models like standard CNNs have shown promise in fault diagnosis, they typically assume that training (source) and testing (target) data share the same statistical distribution. In practice, changes in operating conditions—such as a motor shifting from 1hp to 3hp—alter vibration signals, leading to significant performance degradation. Traditionally, addressing this "domain shift" required expensive and time-consuming manual labeling of data for every new condition. A2CNN addresses this challenge through Unsupervised Domain Adaptation (UDA).
The A2CNN Framework
The proposed A2CNN architecture processes raw 1-D vibration signals, first transforming them via Fast Fourier Transform (FFT) to extract frequency-domain features. The model consists of four primary components: a Source Feature Extractor ($M_S$), a Target Feature Extractor ($M_T$), a Label Classifier ($C$), and a Domain Discriminator ($D$).
The core of the methodology is an adversarial "minimax" game. While the Label Classifier is trained on labeled source data, the Target Feature Extractor ($M_T$) is trained to "fool" the Domain Discriminator. $M_T$ attempts to generate features so similar to the source domain that the discriminator cannot distinguish between them. This process forces the network to ignore condition-specific noise and focus on the fundamental signatures of bearing faults.
A key architectural innovation is partial weight untying. Rather than forcing the source and target extractors to share all parameters, A2CNN ties the earlier layers (to maintain efficiency) while allowing later "adaptation layers" to evolve independently. This provides the flexibility necessary to bridge significant distribution gaps.
Experimental Evaluation
The model was evaluated using the Case Western Reserve University (CWRU) Bearing Dataset, focusing on four states: normal, inner race fault, outer race fault, and ball fault. Experiments involved six cross-domain tasks across three different motor loads (1hp, 2hp, and 3hp).
A2CNN demonstrated exceptional robustness, achieving an average accuracy of 99.21%. This significantly outperforms previous state-of-the-art methods like WDCNN with Adaptive Batch Normalization (95.95%). Notably, in difficult transfer tasks where other models faltered, A2CNN maintained high precision and recall, indicating it is less prone to false alarms or missed detections.
Key Insights
- 1-D Convolutional Layers: The authors advocate for 1-D CNNs over 2-D image-based approaches. 1-D architectures are computationally lighter and better suited to the periodic nature of vibration time-series data.
- Wide Kernels: Using wide kernels in the initial convolutional layer helps the network capture low-frequency information while suppressing high-frequency industrial noise.
- Feature Visualization: t-SNE visualizations confirm that after adversarial training, fault clusters from different domains overlap almost perfectly, demonstrating the successful learning of domain-invariant features.
Conclusion
A2CNN provides a powerful solution for industrial fault diagnosis under varying working conditions. By combining adversarial training with partial weight untying, the model eliminates the need for labeled target data while maintaining near-perfect accuracy. Future work will explore adapting the model to noisier real-world factory environments and multi-source domain adaptation scenarios.
Summary updated