Abstract: The residual life prediction of mining belt conveyor motor is one of the key technologies to ensure the safety of mine production. In view of the shortcomings of the existing prediction models in feature extraction,time series dependence modeling and computational complexity,the multi-source sensor system on the frequency conversion machine is used to collect the operation data of the mine belt conveyor motor,and the residual life prediction is carried out based on the MSC-ECA-Transformer model. The model embeds a multi-scale causal dilated convolution (MSC) and an efficient channel attention (ECA) module in the Transformer backbone network. The multi-level temporal feature extraction is constructed by MSC to solve the problem of insufficient multi-scale feature capture of the progressive degradation mode of the device by the traditional self-attention mechanism. At the same time,the ECA module is introduced to realize the dynamic weight distribution of the feature channel and enhance the saliency expression of the fault sensitive features. The experimental results show that the MSC-ECATransformer model has excellent prediction accuracy and stability. The mean absolute error (MAE) and root mean square error (RMSE) of the improved model are 0. 085 1 and 0. 091 8,respectively. Compared with the Transformer model,they are reduced by 34. 0% and 36. 2%,respectively,which provides technical support for the residual life prediction of mining motors. 

Key words: belt conveyer,electrical machine,residual life prediction,MSC-ECA-Transformer,multiscale causal dilation convolution,time series