This paper is based on the study of machine side bearing temperature overrun faults in wind turbines. By analysing the characteristics of the data collected by the SCADA system, conventional data completion and deletion processes for missing data, outlier data and discrete anomalous data are carried out based on a priori knowledge. In order to reduce the memory and computation cost and to store and process spatial information at a lower spatial resolution, a combination of octave convolution (OctConv) and temporal convolutional network (TCN) is proposed. The sampling method is improved for the asymmetry of information that can be caused by the asymmetry of the OctConv upsampling and downsampling methods, and the Oct-TCN deep learning network is proposed. Application of Oct-TCN deep learning network algorithms to analyse wind turbine generator SCADA data with temporal characteristics is proposed. Based on the identification of fault feature variables, the fault prediction network is trained to make predictions. The system achieved a prediction accuracy of 98.65%.