An Evolutionary One-Shot Neural Architecture Search Method Based on Single-Path Cells Toward Physics-Informed Fault Diagnosis

Abstract Fault diagnosis is important for the complex equipment, and physics-informed fault diagnosis has become an emerging trend. While physics-informed fault diagnosis is hard to realize, unless several problems are addressed, one of the limitations is searching for the best architecture, which influences the performance greatly. Neural architecture search (NAS) has been a research hotspot. However, limited by computing resources and the deviation of supernet prediction, NAS might miss the best architecture, and impedes the application of NAS in physics-informed fault diagnosis greatly. Thus, this article proposes an evolutionary one-shot NAS method based on single-path cells (SPC-NAS) for physics-informed fault diagnosis. The proposed method develops a new supernet based on single-path cells, to reduce the computing resources and improve the reusability. An improved supernet training method is introduced to reduce the deviation between the one-shot model prediction and the stand-alone model accuracy. Finally, an evolutionary search strategy with constraint is developed to find the best architecture. The experimental results show that the proposed method can automatically find the best architecture for different tasks, which achieved an accuracy of 100% in Case Western Reserve University (CWRU) dataset with only 0.292 M parameters. All the results indicate that the proposed method can address the limitation to search the best architecture and provides a foundation for future integration with physics-informed methods.