Traction motor temperature prediction via variational mode decomposition and multi-scale spatiotemporal feature fusion

Temperature prediction of traction motors plays a crucial role in the health assessment, safe operation, and early-warning maintenance of the traction motors for high-speed trains. In practical operation, traction motor temperature data are affected by varying operating conditions and usually exhibit non-stationary fluctuations, multi-scale characteristics, and coupled spatiotemporal dependencies. Aiming at the problems of insufficient feature extraction depth in traditional time-series prediction models and the failure of prediction results to meet engineering requirements under complex operating conditions, this paper proposes a traction motor temperature prediction framework integrating variational mode decomposition, a stacked bidirectional gated recurrent attention unit, a multi-channel temporal convolutional network, and cross-attention; the framework is termed variational mode decomposition-MTGAC. The model is designed to extract and fuse multi-scale spatiotemporal features in data, thereby achieving accurate prediction of traction motor temperature. First, variational mode decomposition is used to perform multi-scale decomposition on the original monitoring data, which eliminates noise and redundant components and improves data quality. Second, a temporal feature extraction module based on a stacked bidirectional gated recurrent attention unit is designed to capture the long-term and short-term dependencies in the data and extract the features of temperature changes at different timescales. Meanwhile, a spatial feature extraction module based on a stacked multi-channel temporal convolutional network is constructed; by leveraging the parallel computing advantage of multi-channel convolution kernels, this module mines the spatial features among temperature changes from multiple dimensions. Finally, a cross-attention mechanism is adopted to deeply fuse spatiotemporal features and dynamically adjust weights to highlight key information, thus enhancing the model’s feature representation capability. A dataset is constructed using the actual operation data of CR300BF electric multiple units, and comparative experimental tests are conducted under multiple prediction scenarios. The experimental results show that under the two scenarios with prediction output steps of 45 and 90 min, compared with Informer, Bidirectional Gated Recurrent Unit, Convolutional Neural Network, and Transformer models, the variational mode decomposition-MTGAC model achieves an increase of 1.10%–3.95% in R 2 and reductions of 24.90%–73.99% and 42.34%–68.41% in mean squared error and mean absolute error, respectively. In different prediction scenarios, the temperature prediction curve of the proposed model has a higher fitting degree with the actual temperature curve. It not only improves the prediction accuracy and model generalization ability but also reduces the prediction error, providing technical support for traction motor temperature prediction and early-warning applications in high-speed trains.