Electric Motor Bearing Fault Noise Detection with Mel-Transformer Model and Multi-Timescale Feature Extraction

Bearings are often found in various industrial systems such as electric motors, and their failure often results in industrial losses and personal danger. This paper proposes a transformer-based method to classify the bearing noise signal efficiently and accurately. Firstly, the vibration noise sample signals are used to be extract the fault feature information by multi-timescale Mel-spectrograms. Secondly, this paper proposes Mel-transformer architecture which is the first to apply the vision Transformer-based algorithm model to fault detection task. This method has a powerful ability to automatically extract vibration noise feature information from the Mel-spectrogram feature map and can distinguish various fault types effectively. Compared with convolutional neural network (CNN) based model, the proposed method is more suitable for processing large-scale bearing data in the industry and requires lower computing resources. It also overcomes the problem that the model cannot be processed in parallel on the vibration sequence due to the limitation of the RNN structure. The effectiveness and feasibility of the proposed method are verified by CWRU dataset.