Optimized machine learning methods for identifying the stiffness loss of CRTS-II slab track based on vehicle vibration signals

Abstract

Vehicle loads and environmental actions inevitably cause stiffness loss in CRTS-II slab track. Accurately identifying the stiffness loss of the slab track has been a crucial issue to the operation safety of vehicle-CRTS-II slab track coupled system (VSCS). However, existing identification methods for the slab track conditions which often focus on a single damage condition of track service status are inefficient and laborious. This study proposes optimized machine learning (ML) methods for automatically identifying the stiffness loss of the CRTS-II slab track utilizing vehicle vibration signals. The proposed methods achieve the intelligent identification of fastener and interface damage in the slab track, with high identification efficiency and low manpower cost. Four selected ML methods, i.e., support vector machine (SVM), random forest (RF), light gradient boosting machine (LGBM), and artificial neural network (ANN) with optimized hyperparameters are developed to identify the stiffness loss of the slab track. 2200 cases from the dynamic model of VSCS under different conditions of fastener and interface failure are generated to train and test the ML models. The proposed ML methods perform well in the training and testing process, demonstrating that the presented ML methods can accurately identify the stiffness loss of the slab track. Furthermore, the stacking-ensemble learning framework is presented to optimize the performance of the above four ML methods for identifying the stiffness loss of the slab track. The maximum improvement in accuracy for the four selected ML models, utilizing the acceleration of vehicle body and bogie, is 109.09 % and 31.58 %, respectively. The stacking generation has strong anti-noise robustness and generalization ability, proving the excellent reliability and stability of the proposed optimized ML methods. The feature importance of the ML method based on the vehicle acceleration is also analyzed. The proposed efficient and capable optimized ML methods are expected to be widely adopted to intelligently identify the complex service status of track structure utilizing vehicle vibration signals.