An Intelligent Diagnosis Method for Metro Vehicle Wheelset Out-of-Roundness Based on Feature Mode Decomposition Combined With Deep Learning

To address the complex noise characteristics and nonlinear coupling present in vibration signals of metro vehicles under real operating conditions, this study proposes an intelligent diagnosis framework that combines signal decomposition with deep learning. Feature mode decomposition (FMD) is used for signal preprocessing, with its parameters optimized via the football team training algorithm (FTTA), using divergence entropy as the fitness function. The envelope spectrum peak factor is subsequently applied to select optimal mode components, which are reconstructed to yield an effective fault signal. This reconstructed signal is then input into a Bidirectional Long Short-Term Memory (BiLSTM) network for fault classification. Experimental validation using real axle-end vibration data from metro vehicles confirms that the proposed method can accurately identify four typical wheelset out-of-roundness conditions—smooth, slight, moderate, and severe—in the indicators of diameter run-out and polygons with accuracy exceeding 97%. This approach provides a reliable technique for quantitatively evaluating the health condition of metro vehicle wheelsets and demonstrates significant potential for practical applications.