Health State Recognition of Bearing based on Time-Frequency Spectrogram and Deep Learning

Health state recognition of bearing lays the cornerstone of long-term safe operation of electromechanical systems. Vibration signals of bearing contain a lot of characteristic information of its health state that needs careful signal processing and feature extraction traditionally. In response to the high expert knowledge and experience reliance as well as labor work, a bearing health state recognition method based on MobileViT lightweight deep learning network is proposed, which models and analyzes the problem of large intra-class differences but small inter-class differences. First, the original one-dimensional vibration signal of the rolling bearing is transformed into a two-dimensional time-frequency spectrogram using Short Time Fourier Transform (STFT), which is used as the model input. Secondly, a MobileViT network based on the Transformer framework is proposed to realize the extraction and encoding of image features by introducing an attention mechanism in convolutional neural networks, and automatic recognition of the health state of bearing. The early stopping method is used during model training to avoid the problem of overfitting caused by continued training of the model and to improve its convergence speed. The experimental results show that the average classification accuracy reaches 99.7%, which is highly robust and stable compared to other methods.