Generalised Few-shot Learning for Rotor System Diagnosis

Current methods of rotor system condition monitoring require a substantial amount of manual work for large fleets of machines. Data-driven automated fault diagnosis models based on deep learning (DL) have the potential to drastically reduce the amount of time spent on manual analysis. However, there is often a lack of data from the entire range of possible operating conditions of a system. The poor generalisation of most DL-based models to operating conditions not present in the training data decreases their usefulness in industry applications. This paper investigates the generalisation ability of a few-shot learning method using only a single example to learn each new class. A prototypical network with a modified Deep Convolutional Neural Networks with Wide First-layer Kernels (WDCNN) architecture was used as the few-shot model. The generalisation of the model was studied from sensor to sensor and across operating speeds. The results indicate that the model is robust to changes in sensor orientation and relatively robust against changes in sensor location. Additionally, the model showed promising results when tested on operating speeds many times higher or lower than it was trained on. The results show that few-shot learning methods have the potential to work in industry applications where limited training data is available. This research also gives an excellent baseline for future research on the generalisation of few-shot learning methods on rotor system fault diagnosis over large changes in operating conditions.