Controlled generation of unseen faults for Partial and Open-Partial domain adaptation

Katharina Rombach, Gabriel Michau, Olga Fink

Abstract

New operating conditions can result in a significant performance drop of fault diagnostics models due to the domain shift between training and testing data distributions. While several domain adaptation (DA) approaches have been proposed to overcome such shifts, their application is limited if the fault classes represented in the two domains are not identical. To enable better transferability, particularly in "extreme" setups where only the healthy data class is shared, we propose FaultSignatureGAN. This framework utilizes a Wasserstein GAN to generate distinct, controlled synthetic fault signatures. By disentangling fault-specific signatures from domain-specific background noise, the model enables high-accuracy diagnostics for unobserved fault types in a target domain. Our experiments on two bearing fault case studies demonstrate superior results compared to existing methods, especially in scenarios with large domain gaps.

Problem Definition: Label Space Discrepancies

In industrial settings, faults are rare, and machines often start operating under new conditions (a "domain shift") where no faulty data yet exists. Most standard DA techniques assume the source and target domains share the same set of fault classes. However, real-world scenarios often involve label space discrepancies:

  1. Partial DA: The target domain contains only a subset of the source faults.
  2. Open-Partial DA: Both domains have unique "private" classes.
  3. The Extreme Case: Only healthy data is available for the target machine, requiring the model to predict faults that haven't occurred in that specific environment yet.
Overview of training and test data distributions for different domain adaptation settings.

The FaultSignatureGAN Framework

The core innovation of FaultSignatureGAN is the hypothesis that a faulty signal's Fourier spectrum is the sum of a domain-specific component (the machine's "healthy" baseline) and a fault-specific signature. The framework, based on a WGAN-GP, operates in three stages:

  • Generation: A generator creates domain-independent fault signatures in the Fourier domain, conditioned on a specific fault class.
  • Plausibility: A discriminator ensures these signatures, when added to source healthy data, resemble real source faults.
  • Consistency: A cooperative classifier ensures the generated signatures accurately represent the intended fault class.

To "transplant" a fault to a new target domain, the model takes a generated signature, scales it by a factor $w$ (the ratio of healthy signal means between domains), and adds it to the target's healthy data. This allows for the creation of a labeled synthetic training set for the target environment.

The proposed methodology for generating synthetic target data to train a diagnostic classifier.

Experimental Evaluation

The method was evaluated on two benchmark bearing datasets: CWRU (Case Western Reserve University) and the Paderborn University Dataset. The latter is particularly challenging as it involves real, non-artificial damages and significant domain gaps in rotational speed and load.

Key Results

  • Handling Large Gaps: On the difficult Paderborn "Domain 1" shift, FaultSignatureGAN achieved a 23.76% relative improvement over baselines. While standard adversarial alignment methods struggled to reach 50% accuracy, this approach maintained high performance.
  • High Accuracy: On the CWRU dataset, the method achieved an average accuracy of 97.66% across all domain shifts, outperforming state-of-the-art methods like BA3US.
  • Hyperparameter Tuning: A major advantage is that the synthetic faults are realistic enough to serve as a validation set for tuning the final diagnostic classifier, solving a common "black box" problem in unsupervised DA.

Conclusion and Future Work

FaultSignatureGAN demonstrates that fault signatures can be treated as additive components independent of a machine's baseline state. While the current model assumes a linear relationship between operating conditions and fault signatures, it provides a robust solution for extreme domain adaptation. Future research will explore the transferability of signatures between different types of mechanical components and the integration of unlabeled target data to further refine synthetic samples.

Summary updated

Abstract

New operating conditions can result in a significant performance drop of fault diagnostics models due to the domain shift between the training and the testing data distributions. While several domain adaptation approaches have been proposed to overcome such domain shifts, their application is limited if the fault classes represented in the two domains are not the same. To enable a better transferability between two different domains, particularly in setups where only the healthy data class is shared between the two domains, we propose a new framework for Partial and Open-Partial domain adaptation based on generating distinct fault signatures with a Wasserstein GAN. The main contribution of the proposed framework is the controlled data generation with two characteristics. Firstly, previously unobserved target faults can be generated by having only access to healthy target and faulty source samples. Secondly, distinct fault types and severity levels can be generated precisely. The proposed method is especially suited for extreme domain adaption settings that are particularly relevant in the context of complex and safety-critical systems, where only one class is shared between the two domains. We evaluate the proposed framework on Partial as well as Open-Partial domain adaptation tasks on two bearing fault diagnostics case studies. In the evaluated case studies the proposed methodology demonstrated superior results compared to other methods, particularly in the presence of large domain gaps. The experiments conducted in different label space settings (Partial and Open-Partial) showcase the versatility of the proposed framework.
BibTeX
@inproceedings{rombach2022controlled,
  author = {Katharina Rombach and Gabriel Michau and Olga Fink},
  title = {{Controlled generation of unseen faults for Partial and Open-Partial domain adaptation}},
  booktitle = {Reliability},
  year = {2022},
  doi = {10.1016/j.ress.2022.108857},
}
Figure