Fault diagnosing for train suspensions using automated operational modal analysis

Monitoring and fault diagnosis of suspension systems are crucial for ensuring the safe operation of railway vehicles. However, the structural complexity of suspension systems, non-ideal excitation signals, and high noise levels in vibration data pose significant challenges to conventional methods. To address these issues, this paper presents a Convolutional Neural Network-based Automated Operational Modal Analysis method (CNN-AOMA) that integrates the Stochastic Subspace Identification-Covariance (SSI-COV) algorithm with deep learning. A large set of control parameter samples is first generated through Monte Carlo (MC) simulation to construct overlapping stabilisation diagrams. Physical modal parameters are then extracted using Kernel Density Estimation (KDE), while a CNN architecture intelligently classifies and predicts parameter combinations, substantially enhancing the automation level and computational efficiency of modal identification. The feasibility of the proposed method was verified experimentally on a 3-DOF simplified suspension test rig. Further validation was conducted through a 1/5-scale bogie roller rig test, which demonstrated the accuracy and effectiveness of the CNN-AOMA method in identifying modal parameters under both normal and fault conditions.