Controlled Generation of Unseen Faults for Partial and Open-Partial Domain Adaptation

TL;DR

This work tackles fault diagnostics under extreme domain shifts where only the healthy class is shared across domains, by introducing FaultSignatureGAN, a framework that learns domain-independent fault signatures in the Fourier domain and transfers them to a target domain through a Wasserstein GAN with gradient penalty paired with a semantic classifier. The method enables controlled generation of unseen target faults, allowing Partial and Open-Partial Domain Adaptation without target fault data and without relying on extrapolation of generative models. Empirical results on bearing datasets (CWRU and Paderborn) demonstrate that generating synthetic target faults substantially improves classification performance, especially when domain gaps are large, often outperforming strong baselines and remaining competitive with or superior to Unilateral methods. The approach also supports hyperparameter tuning with synthetic data and speeds up data acquisition for new assets, offering a practical pathway for reliable fault diagnostics in safety-critical systems.

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 of the trained models 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 synthetic fault data generation with two main distinct characteristics. Firstly, the proposed methodology enables to generate unobserved fault types in the target domain by having only access to the healthy samples in the target domain and faulty samples in the source domain. Secondly, the fault generation can be controlled to precisely generate distinct fault types and fault severity levels. The proposed method is especially suited in 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. Our experiments conducted in different label space settings showcase the versatility of the proposed framework. The proposed methodology provided superior results compared to other methods given large domain gaps.

Figures (9)

• Figure 1: Four DA configurations according to label space discrepancies: (a) ClosedSet ; (b) Partial, (c) OpenSet,(d) both Open-Partialboris2021universal.
• Figure 2: Illustration of the source and target alignment challenge when only one class is shared between the domains on the example of the Partial DA setting: The source and the target datasets are shown in Fig. \ref{['fig:challenge_1']} whereby only one class (green class) is represented in the target domain. The alignment step based on one class only is shown in Fig. \ref{['fig:challenge_2']}, whereby the challenge of finding the optimal alignment is indicated. The quality of chosen alignment method can only be tested during the a-posteriori evaluation, when the target classes have been observed ( see Fig. \ref{['fig:challenge_3']}).
• Figure 3: FaultSignatureGAN in the Partial DA settings: the original data setting is depicted in Fig. \ref{['fig:partialsubs_1']}; the missing target classes are generated in Fig. \ref{['fig:partialsubs_2']}; the target dataset is augmented with synthetically generated data in Fig. \ref{['fig:partialsubs_3']} and a classifier is trained on the augmented dataset in Fig. \ref{['fig:partialsubs_4']}.
• Figure 4: FaultSignatureGAN in the Open-Partial DA settings: the original data setting is depicted in Fig. \ref{['fig:OSPAsubs_1']}; the missing source and target classes are generated in Fig. \ref{['fig:OSPAsubs_2']}; the source and target dataset is augmented with synthetically generated data in Fig. \ref{['fig:OSPAsubs_3']} and a classifier is trained on the augmented dataset in Fig. \ref{['fig:OSPAsubs_4']}.
• Figure 5: FaultSignatureGAN: Training Phase: Training the A) generative model to generate domain independent fault characteristics while imposing B) plausibility with the discriminator in the source domain and C) semantic consistency with the classifier. Execution Phase: The generation of unseen target data.