SWDAE: A New Degradation State Evaluation Method for Metro Wheels With Interpretable Health Indicator Construction Based on Unsupervised Deep Learning

Machine learning has shown advantages in assessing wheel degradation in metro vehicles for fault prognostic and health management (PHM). However, practical implementation faces challenges due to disturbances in wheel vibration signals caused by factors like load, road conditions, and temperature, introducing significant noise that masks evaluation tendencies. This article develops a novel unsupervised approach to evaluate wheel degradation under strong noise disturbance. It utilizes a second-generation wavelet transform for time–frequency analysis of noisy signals and introduces a second-generation wavelet deep autoencoder (SWDAE) network to extract adaptive feature representations in different frequency bands. The training algorithm alternately optimizes the wavelet transform and DAE. With frequency-saliency interpretability, health indicators (HIs) for the degradation process are constructed using principal component analysis (PCA) on the obtained features in each frequency band, selecting the most representative frequency component based on the monotonicity of the HIs. State changes are automatically determined using a second-order derivative-based assessment method aligned with the first/second warning strategy. Comparative experiments using Beijing Subway wheel data demonstrate the monotonicity and physical significance of the constructed HIs, with warning locations accurately matching changes in wheel diameter recorded during repairs.