Investigating merits of deep self-supervised learning on a fatigue RUL prognostics application
Résumé
This study aims to detect and localize rotor unbalance and shaft crack damage in a fewshot data scenario. It proposes a reinforcement learning-based approach with physics preferences (that we name RLP2). RLP2 is used to guide the physics consistency in the rotor finite element mimetic neural network (RFEMNN). The RFEMNN is first trained in an unsupervised manner using mixed simulation and experimental datasets in the task of reconstructing rotor's vibration signals. Then, the RFEMNN is fine-tuned in the RLP2 framework using a physics preference reward as policy loss to ensure similarity between hidden layer output and rotor system parameters. The output of the RFEMNN is fed into a downstream multi-output convolutional neural network (CNN) for fault diagnostic and localization. The proposed method's effectiveness is demonstrated through experiments on a PT500 platform under zero-shot, one-shot, and few-shot learning scenarios. The obtained results indicate the potential of this method for the predictive maintenance of rotor systems in real-world applications with limited training data.