Intelligent identification for vertical track irregularity based on multi-level evidential reasoning rule model

Vertical track irregularity is one of the most significant indicators to evaluate track health. Accurate identification of vertical track irregularity is beneficial to achieve precise maintenance of the track and thus avoid accidents. However, the continuous variation of the track irregularity and the imbalance of the abnormal/normal data samples make it difficult to guarantee the accuracy of identification models. Therefore, by considering the interaction between train and track, a multi-level evidential reasoning (M-ER) rule model is proposed to build the nonlinear causal relationship of vibration signals and vertical track irregularity. In the modeling process of M-ER, the referential evidence matrix (REM) and fusion parameters (i.e., reliability factors and importance weights) are determined and optimized. In the model, the reliability factor of evidence is determined through trend analysis, while the importance weights of evidence and REM are optimized by sequential quadratic programming (SQP). In the inference process of M-ER, sample expansion strategy and two-level evidence fusion mechanism are designed. Specifically, in the first level, samples on each vibration signal are fused with their nearest neighboring historical samples obtained by K-Nearest Neighbor(K-NN) method. In the second-level, the results generated in the first-level are integrated by ER rule. We evaluate the M-ER rule model with an actual data set from China railway. The experimental results show that the model can identify the vertical track irregularity more accurately compared with the single-level ER rule model and other typical machine learning based models.