A bearing fault diagnosis method combining the Gramian angular difference field and a dual-channel convolutional neural network

Regarding the problems of weak noise resistance, poor generalisation performance and low diagnostic accuracy under small-sample conditions in traditional deep learning-based bearing fault diagnosis methods, a rolling bearing fault diagnosis method based on a dual-channel fusion convolutional neural network and bidirectional gated recurrent unit (DCFCNN-BiGRU) is proposed. It uses the Gramian angular difference field (GADF) to encode temporal signals as 2D features. The features on the timing and coding graphs are extracted by two parallel convolutional channels of different dimensions and, after fusion of the features of two different dimensions, a bidirectional gated recurrent unit (BiGRU) is introduced to extract the temporal features, while an attention mechanism is introduced to suppress unnecessary features. Finally, feature mapping to a vector space combined with online label smoothing regularisation (OLSR) is used to construct a more reasonable probability distribution, thus achieving end-to-end fault diagnosis through a distance function. The Case Western Reserve University (CWRU) and Jiangnan University bearing datasets are used to verify the proposed model and comparison experiments are conducted with other models. The experimental results show that DCFCNN-BiGRU performs well in different working conditions and performs better in noisy environments and small-sample conditions compared with other models.