Integrated Classification Method for Vehicle Wheel-set Condition Based on Imbalanced Datasets

The wheel-rail vibration signal of a metro vehicle has the characteristics of being non-linear, non-stationary, and a low value of signal-to-noise ratio. The paper proposes an integrated classifier model based on the imbalanced datasets to achieve the disassembly detection and fault diagnosis of wheel flats. Four typical wheel flat faults, namely tread normal, tread scratch, tread peeling, and circumference wear were studied in the research. Feature extraction of signals was performed by incorporating the variational mode decomposition and fuzzy entropy. Fault datasets were constructed and support vector samples were filtered by employing the bias support vector machines. SMOTE (synthetic minority oversampling technique) oversampling was applied and non-support vector samples were combined. An integrated classifier was then built by incorporating the directed acyclic graph for the fault identification. Finally, the study analyzed the precision and recall ratios to evaluate the classification performance of an integrated classifier the. Experiments were conducted at a depot under no load running test. The experimental results show that the proposed method can achieve over 96% of accuracy for the given 4 fault models, which can be effectively employed in the wheel flat faults diagnosis for the metro vehicles.