Automated Model Generation for Machinery Fault Diagnosis Based on Reinforcement Learning and Neural Architecture Search

In recent years, deep learning (DL)-based fault diagnosis methods have demonstrated significant success in various domains due to their high accuracy. Similar to other data-driven techniques, DL models are application-specific and strongly depend on the data which they are developed upon. Different DL models need to be designed for different tasks. In addition, manual-tuning of the DL structures and associated parameters as part of the model design is a trial-and-error optimization process, which is time-consuming and sensitive to changes. To address these limitations, this article presents a reinforcement learning (RL) and neural architecture search (NAS)-based automatic modeling framework (RL-NAS) for fault diagnosis of machinery. The RL-NAS method can adaptively and automatically design high-accuracy network models according to the intended diagnosis tasks. A weight-sharing mechanism has been developed to alleviate the bottleneck of NAS where the computation time increases exponentially with the expansion of the search space and the deepening of the model structure. To evaluate its effectiveness, efficiency, and reproducibility, the proposed framework is validated on five datasets from different institutes, spanning the applications of rolling element bearings, gears, and ball screws. The results show that for all five applications, the RL-NAS method successfully searched out diagnostics models with over 97% accuracy within a short time, from 40 s to 30 min depending on the complexity of the diagnosis task, showing good performance of the proposed method.