Fast wheelset wear-stage recognition based on multi-domain entropy features and an attention-enhanced lightweight 1D-CNN

This study proposes an intelligent diagnostic framework for high-accuracy, low-latency monitoring of heavy-haul railway wheelset wear under limited computational resources. The framework combines Variational Mode Decomposition (VMD)-based signal decomposition, multi-domain entropy features, and a lightweight attention-based classifier. The Grey Wolf Optimizer (GWO) is used offline to determine the key VMD parameters, which are then fixed for narrowband decomposition of wheel–rail contact acoustic signals during online diagnosis. Adaptive Time-Domain Discrete Entropy (ATDE) and Frequency-Domain Entropy (FDE) are extracted from each modal component to characterize dynamic complexity and energy distribution, thereby enhancing feature robustness and class separability. A One-Dimensional Convolutional Neural Network (1D-CNN) with an attention mechanism then reweights the features and identifies the wear stage. In the four-stage wheelset wear task, the method achieves 99.88% test accuracy and 99.713% mean accuracy over ten independent runs. Compared with time–frequency image-based deep models, the classifier reduces forward inference time by about 99.74% and uses only 13,848 parameters. These results demonstrate improved computational efficiency without sacrificing accuracy, supporting real-time deployment in resource-constrained conditions. Validation on the HUST bearing dataset further suggests its applicability to different mechanical diagnostic tasks.