Elevator equipment faces multiple risk factors such as mechanical wear and tear and electrical aging during long-term operation, which leads to frequent failures. The safety of elevator operation is related to the safety of public life and property, and accurate prediction of elevator failures is of great significance in preventing accidents. Aiming at the problems of feature redundancy and low prediction accuracy in existing elevator failure prediction methods, this study proposes a TimesNet elevator operation accident prediction model that integrates DLinear and deformable convolution. Firstly, MIC correlation analysis is used to eliminate feature redundancy, and TimeGAN technique is used to enhance the fault data to balance the sample distribution; then MS-TimesNet model is constructed for feature extraction, and the complex change patterns of the time series data are captured by the dynamic convolution module and the TimesNet module; finally, the DLinear method is applied to reconstruct the features from the two dimensions, namely, the trend and residuals to improve the prediction accuracy. The experiments are validated using the operation data of 30 elevators distributed in 24 different areas, and the results show that the proposed model achieves the accuracy of 0.98, 0.97 and 0.94 on the training, validation and test sets, respectively, which is better than the comparative models of BiLSTM and RNN. The study proves that TimesNet fusing DLinear and deformable convolution can effectively improve the performance of elevator fault prediction and provide reliable technical support for the safe operation of elevators.
