A Data-Driven Convolutional Regression Scheme For On-Board and Quantitative Detection of Rail Corrugation Roughness

Accurate and fast measurement of rail corrugation is an important prerequisite for formulating rail grinding strategy. This paper presents a novel data-driven convolutional regression scheme for on-board and quantitative detection of rail corrugation roughness. In this scheme, the manner of data pre-processing and standardization is first defined. Then, a novel application-oriented model named RCNet is proposed to quantitatively detect the rail corrugation roughness by taking the processed axle box acceleration signals as input. The RCNet possesses the merit of automatic calculation and end-to-end data flow. Finally, the axle box accelerations obtained by the vehicle–track coupled dynamics simulation and field experiment are employed to detect rail roughness using the RCNet, and the performance of the RCNet is evaluated from aspects of qualitative and quantitative detection of rail roughness as well as efficient computation. Results show that the RCNet maintains high robustness in processing field measurement signals containing many interferences. The quantitative fitting degree of rail corrugation roughness is about 95.7%, and the average time cost of a single sample is only 0.22 ms. Comparative analysis indicates that the developed convolutional-based RCNet outperforms three reported machine learning algorithms in terms of fitting degree, loss level and timeliness of rail corrugation roughness detection. Moreover, the generalization ability of the RCNet in different application scenarios is further assessed.