Physics-Informed LSTM hyperparameters selection for gearbox fault detection

A situation often encountered in the condition monitoring (CM) and health management of gearboxes is that a large volume of CM data (e.g., vibration signal) collected from a healthy state is available but CM data from a faulty state unavailable. Fault detection under such a situation is usually tackled by modeling the baseline CM data and then detect the fault by examining any deviation of the baseline model versus newly monitored data. Given that the CM data is mostly time series, the long-short term memory (LSTM) neural network can be employed for baseline CM data modeling. The LSTM is free from the choice of the number of lagged input time series and can also store both long-term and short-term time series dependency information. However, we found that an LSTM with its hyperparameters selected whilst minimizing validation mean squared error (VAMSE) does not differentiate the faulty and healthy states well. There is still room for detectability improvement. In this paper, we propose a physics-informed hyperparameters selection strategy for the LSTM identification and subsequently the fault detection of gearboxes. The key idea of the proposed strategy is to select hyperparameters based on maximizing the discrepancy between healthy and physics-informed faulty states, as opposed to minimizing VAMSE. Case studies have been conducted to detect the gear tooth crack and tooth wear using laboratory test rigs. Results have shown that the proposed physics-informed hyperparameters selection strategy returns an LSTM that can better detect these faults than the LSTM returned from minimizing VAMSE.