Zero-Shot Industrial Fault Diagnosis via Physics-Informed Generative Adversarial Networks (PI-GANs): A Cross-Domain Validation Study

Published February 23, 2026 | Version v1

Journal Open

The scarcity of available run-to-failure dataset

hampers the implementation of deep learning for predictive

maintenance (PdM) in industry. This work proposes a Physics

Informed Generative Adversarial Network (PI-GAN) specifically

tailored for zero-shot fault detection, which can diagnose faults

not seen during training. The key aspect of the PI-GAN is the

integration of the differentiable physics engine, based on Ohm’s

and Newton’s laws, in the generator loss function, using

Adversarial Loss and Physics Residual Loss. The training was

performed for 50 epochs on artificially generated electrical fault

datasets based on a direct current motor (R=2.0 Ω, B=0.001).

Testing was performed on the Case Western Reserve University

(CWRU) bearing dataset. During inference, the value for the

resistance parameter was set to 0.5 Ω, simulating short circuiting.

For 235 real-world signal windows, the PI-GAN offered zero-shot

accuracy of 98.72% on new Inner Race faults, as well as

interdomain accuracy of 77.22% on Outer Race faults. The

Physics loss function reduced to 78.4%, which signifies that the

underlying electromechanical principle is reflected by the

generator, which is efficient and can be applied uniformly for fault

detection, paving the way for future intelligent and self

sustainable maintenance strategies for Industry 4.0.

Files

Name	Size	Download all
Zero-Shot Industrial Fault Diagnosis via Physics-Informed Generative Adversarial Networks (PI-GANs) A Cross-Domain Validation Study.pdf md5:c216bd83a6557dd4672c8be6cc412c77	511.2 kB	Preview Download