A wheel polygon identification algorithm based on rail vibration using Empirical Mode Decomposition and Singular Value Decomposition (EMD-SVD) and a neural network

With the advancement of high-speed and heavy-haul railway transportation, wheel polygonization has become a critical factor affecting operational safety. Addressing the limitations of manual inspection and the lack of real-time monitoring, this study investigates the dynamic recognition and online detection of wheel polygonization for HXD1 electric locomotives. A vehicle-track rigid-flexible coupling dynamic model considering flexible wheelsets was developed to analyze rail vibration responses under various harmonic orders and wear depths. Furthermore, a feature extraction method based on Empirical Mode Decomposition and Singular Value Decomposition (EMD-SVD) was proposed, and a 9-5-1 structured BP neural network was designed using a 9-dimensional feature vector. Simulation results demonstrate that the model effectively identifies faults for polygonization of the 9th order or higher with wear depths exceeding 0.1 mm. For engineering application, a trackside vibration-based online detection system was implemented at the Huaihua Locomotive Depot, collecting 720 sets of wheel-mapped vibration data. The BP neural network successfully identified three suspected wheels. Static verification using an M/Wheel instrument showed that the measurement results were highly consistent with the network’s predictions, with radial runout rankings aligning well with the diagnostic scores. This research validates the proposed methodology and provides a potential technical approach for the dynamic monitoring and early warning of locomotive wheel out-of-roundness.