Morlet wavelet pooling with frequency-domain feature selection for fault diagnosis in high-speed train

High-speed train bogies operate within complex dynamic environments and are subject to continually fluctuating noise interference. Existing diagnostic methods are challenged by intense noise and intricate operational regimes, particularly as convolutional neural network pooling layers are prone to omitting critical information and blurring subtle fault-related signal features. In this study, we propose a Morlet-wavelet-based pooling layer that leverages frequency-domain information filtering. By integrating the superior time–frequency localization properties of Morlet wavelets into convolutional neural networks (CNNs) pooling design, this approach facilitates a transition from spatial-domain to frequency-domain feature selection. Building upon this, we developed a multi-scale convolutional neural network incorporating Morlet wavelet pooling (MW-CNN), which synergizes multi-scale feature extraction with residual learning to form a deep architecture specifically tailored for high-speed train bogie fault diagnosis. Experiments conducted on a CRH380A high-speed train simulation dataset demonstrate that MW-CNN achieves 99.94% accuracy on the source domain and maintains robust diagnostic performance across varying noise levels and operational conditions. Furthermore, visualization via Eigen-CAM confirms that MW-CNN effectively highlights critical regions within multi-channel signals, providing reliable attention guidance for fault detection and validating the method’s practical efficacy and robustness.