Abstract
Unsupervised domain adaptation (UDA) has achieved remarkable success in fault diagnosis, bringing significant benefits to diverse industrial applications. While most UDA methods focus on cross-working condition scenarios where the source and target domains are notably similar, real-world applications often grapple with severe domain shifts. We coin the term 'distant domain adaptation problem' to describe the challenge of adapting from a labeled source domain to a significantly disparate unlabeled target domain. This problem exhibits the risk of negative transfer, where extraneous knowledge from the source domain adversely affects the target domain performance. In response, we propose a novel Online Selective Adversarial Alignment (OSAA) approach. Central to OSAA is its ability to dynamically identify and exclude distant source samples via an online gradient masking approach. Furthermore, we construct an intermediate domain to act as a transitional domain and ease the adaptation process. Experimental results on two real-world datasets validate the superior performance of OSAA over state-of-the-art methods.
Introduction: The Challenge of Distant Domain Adaptation
In the field of industrial AI, models are often trained on clean, labeled data from laboratory machines (the source domain) and then deployed on noisy, complex machines in actual factory environments (the target domain). This process, known as Unsupervised Domain Adaptation (UDA), typically assumes a degree of similarity between the two domains.
However, real-world "distant" scenarios—such as moving from a lab-controlled bearing test to a high-speed industrial gearbox—present severe domain shifts. In these cases, standard UDA algorithms often attempt to force the model to learn features from source samples that are irrelevant to the target. This leads to negative transfer, where the adaptation process actually degrades the model's performance on the target data compared to a simple source-only baseline.
Proposed Methodology: Online Selective Adversarial Alignment (OSAA)
To bridge the gap in distant domain adaptation, we introduce OSAA, which incorporates three primary innovations:
- Online Selection Module: This core component calculates the "relevance" of each source sample in real-time. By combining reconstruction loss (from an autoencoder) and classification loss, the model identifies "distant" or "unreliable" samples. Their gradients are then masked during backpropagation, ensuring that only the most helpful data guides the adaptation.
- Intermediate Domain Construction: To mitigate the impact of severe domain shifts, we create a transitional domain by sampling equally from the source and target domains. This acts as a stepping stone, easing the alignment process.
- Conditional Adversarial Alignment: The model utilizes a discriminator to learn domain-agnostic features. By incorporating predicted label information, the alignment process respects underlying classes, ensuring that similar faults across domains are correctly mapped.
Experimental Evaluation
The OSAA method was rigorously tested on two gold-standard benchmarks in the fault diagnosis field:
- Paderborn University (PU) Dataset: Evaluates "Artificial-to-Real" scenarios, where source data comes from man-made damages and target data from actual fatigue-induced failures.
- Case Western Reserve University (CWRU) Dataset: Tests cross-speed and cross-location adaptation, such as training on drive-end motor data and testing on fan-end data.
Key Results and Insights
OSAA significantly outperformed state-of-the-art methods across multiple adaptation scenarios:
- Superior Accuracy: In the challenging PU "Artificial-to-Real" tasks, OSAA achieved an average F1-score of 60.94%, compared to 45.12% for the next best method (AdvSKM).
- The Power of Selection: Ablation studies revealed that removing the selection mechanism caused performance to drop by over 25%, highlighting it as the primary driver of success.
- Robustness: Sensitivity analysis showed the model remains stable across various hyperparameters, making it a viable candidate for practical industrial deployment.
Conclusion and Future Directions
The OSAA approach demonstrates that rejecting irrelevant data can be more beneficial than attempting to align all available source data in distant domain scenarios. By using gradient masking for online selection, OSAA avoids the need for expensive pre-processing.
Future research will focus on automatically determining the optimal selection threshold and exploring "Open-Set" distant domain adaptation, where the target domain may contain fault classes not present in the source training data.
Summary updated