Prediction of Remaining Useful Life of Railway Tracks Based on DMGDCC-GRU Hybrid Model and Transfer Learning

The effective prediction of the remaining useful lifetime (RUL) of the railway track can provide key information for intelligent operation and maintenance. However, accurate prediction of railway track RUL is still challenging due to the issues such as serious data noises, large sampling data interval, and small samples in the track inspection dataset, which are usually caused by the complex and various operating environment. To address these issues, a new prediction method for railway track RUL, which is the combination of Dynamic Multi-Scale Gated Dilated Causal Convolution (DMGDCC)and Gated Recurrent Unit (GRU) called DMGDCC-GRU, is proposed in this work. Firstly, an adaptive filtering module was constructed based on dynamic convolution to solve the problem of insufficient feature extraction that is caused by time-varying noise interference in the data. Secondly, a feature extraction network with multi-scale gated dilated causal convolution structure was designed to monitor the damage feature at different scales as well as to generate the boundary representation for the degree of damage at each time-step by the gating mechanism. Based on these, a GRU, which can improve the performance of time-series feature extraction, was introduced to learn the feature of damage trend in full life cycle of railway track efficiently. Finally, an innovative pre-training method for transfer learning from bearing to rail was developed to mitigate the problems about the small samples and long sampling interval. Among the many state-of-the-art RUL prediction methods, experimental evaluations with the actual rail degradation datasets showed that the proposed method can achieve 38.9% reduction in RMSE and 13.4% increase in R-2(Degradation due to rail peeling), 31.7% reduction in RMSE and 8.2% increase in R-2(Degradation due to rail corrugation). Significantly, the validation results of rolling bearing datasets also proved the high generalization performance of the proposed method.