Fault diagnosis of main shaft bearings in mining drills for transfer imputation of missing data

In response to the problems of excessive noise, large drift, and numerous missing values in the monitoring data under complex underground working conditions of mining drilling machines, a bidirectional temporal convolutional joint generative adversarial interpolation network with embedded spatio-temporal attention (BiTCGAIN-STA) was designed. The bidirectional temporal convolutional network (BiTCN) is used to capture the temporal dependency between previous and subsequent time sequences, and the spatio-temporal attention (STA) mechanism is employed to adaptively allocate time and channel weights. Through generative adversarial training, the distribution consistency and diversity of the interpolated samples are improved. At the same time, real data fine-tuning is performed on the target domain to enhance the transfer robustness. A bearing fault diagnosis model based on adaptive weighted fusion and the Informer network is proposed. The Informer long sequence feature extraction network is used to deeply represent the fused signals, thereby improving the ability to identify weak fault features. Experimental results show that, under different missing rates, the root mean square error (RMSE) of the BiTCGAIN-STA model is significantly higher than that of mainstream models such as Mean, MICE, and GAIN, achieving high-quality data reconstruction. The bearing fault diagnosis model has an identification accuracy of 99.87% for weak faults, significantly higher than models such as Transformer and graph neural network (GNN).