AE-KSVD model for performance degradation assessment of high-speed train gearbox bearings

Rolling bearings, as key components in the gearbox of high-speed train running gear, are crucial for the establishment of degradation assessment models and the formulation of maintenance strategies. Appropriate bearing degradation models can effectively reduce the cascading failures caused by bearing failures. Addressing the issues of delayed judgment and low accuracy in single reconstruction data-driven models, this paper compares single-layer autoencoders with K-SVD, analyzing their advantages and disadvantages. It proposes an AE-KSVD dual-layer data reconstruction model that combines Auto-Encoders (AE) with Dictionary Learning (K-Singular Value Decomposition, K-SVD). The first layer reconstructs the input signal through autoencoding, and the second layer reconstructs the hidden layer of the autoencoder using K-SVD. This model combines the accuracy of dictionary learning models with the efficiency of autoencoder models and introduces the Particle Swarm Optimization algorithm to optimize the hyperparameters of the autoencoder and dictionary learning, achieving optimal feature representation capabilities for automatic extraction of vibration data. The model's adaptability is significantly enhanced while ensuring superior feature learning capabilities. The model's superiority is validated by comparing the PSO optimization results of the AE-KSVD and the dual-layer autoencoder models using bearing data from Xi'an Jiaotong University and full-life fatigue test data of bearings from Cincinnati.