Data driven deep learning fault diagnosis method based on vision transformer and multi-head attention for different working condition

Abstract Data-driven deep learning methods have shown excellent performance in bearing fault diagnosis tasks. In application scenarios, the network relies on data needs to adapt to signals under different working conditions. The data in actual scenarios are continuously monitored and consist of multi-source heterogeneous streams, which makes that data-driven fault diagnosis methods need to improve the quality of data with domain differences. Under data normalization and resampling, this paper combines data pre-processing techniques to integrate prior features into the network to improve the accuracy of identifying non-stationary time series data. The improved multi-head attention mechanism is used to further extract features accurately. In the extended multi-head attention mechanism, the dependencies of LSTM and BiGRU on non-stationary time series sequences are combined. By employing the encoder structure of the network, the signal is transformed into STFT images, which are modelled with global dependencies to enable accurate feature extraction. Case studies show that the method has advantages over traditional neural networks under different working conditions and noise levels.