MixerFormer-Covariance Metric Neural Network: A New Few-shot Learning Model for Bearing Fault Diagnosis

Manh-Hung Vu, Van-Truong Pham

Abstract: Deep inside the electrical machine, bearings are a delicate and easily injured component. Consequently, identifying bearing failures is a highly informative and crucial area. Instead of using manual methods, deep learning models were applied to diagnose bearing failures from abnormal signals returned from electrical machines. However, the weakness of modern end-to-end models is that they require a large amount of data to train in order to give highly accurate diagnostic results. Some studies have shown that Few-shot models are very suitable for diagnosis under limited data conditions, but few-shot models require many steps to make the diagnosis. Therefore, to overcome those disadvantages, the Mixer-former Covariance Metric Neural Network model is proposed. This is a few-shot training model but only needs 1 stage to diagnose the fault instead of more than 2 stages like other few-shot models. On the Case Western Reverse University (CWRU) and HUST-Bearing datasets, two open-access datasets, several experiments and ablation investigations were performed. The results demonstrate that the proposed model outperforms other modern deep learning models when there is limited data available. In addition, this model still satisfies the requirements and produces excellent results when trained with a significant quantity of data.