High-speed train small-amplitude hunting recognition based on multi-source discrepancy adversarial network
The hunting motion of a high-speed train can increase the load on all parts of the vehicle.The vigorous hunting motion can cause a greater risk of impact or even derailment between the wheels and rails,which seriously threatens the safety of driving.Therefore,it is necessary to recognize the hunting state of the vehicle,especially the small-amplitude hunting state before the onset of the hunting instability.The majority of current research is centered around deep learning and single-source transfer learning with single-condition data.However,high-speed trains face complex and changing conditions during operation.Taking into account the different distributions of multi-source data under various working conditions,it is difficult to satisfy the recognition accuracy requirements across different working conditions with a train hunting recognition model built using only a single working condition data.The paper proposed a high-speed train small-amplitude hunting state recognition method based on multi-source two-layer discrepancy adversarial network.The method used real hunting motion data with different distributions under multiple working conditions in the training process,and adopted the two-layer discrepancy adversarial training strategy.The initial layer of disparity confrontation combined a moment matching module and a domain confrontation module.This layer model reduced both the distributional differences between the source and target domains as well as the distributional differences between the source and source domains.On the basis of the primary layer of difference confrontation,a secondary layer of difference confrontation training method was used.This layer model was able to further align the marginal and conditional distributions of the data to better learn distinguishable features and improve the accuracy of the diagnostic task.After the feasibility of the model was verified by public bearing data,it was used in the high-speed train hunting state recognition study.The experimental results demonstrate that the method accurately recognizes different states of hunting motion.It can achieve over 99％ accuracy in various recognition tasks and outperforms both single-source and other multi-source models in diagnostic effectiveness,proving its reliability.The method has certain engineering application value in the intelligent monitoring of the state of hunting of high-speed trains.
high-speed trainsmall-amplitude huntingmulti-source domain transfer learningtwo-layer discrepancy adversarialfault recognition
万方数据
维普
