Inversion and identification of vertical track irregularities considering the differential subgrade settlement based on fully convolutional encoder-decoder network

Abstract

Differential subgrade settlement plays a key role in the formation of track geometry and significantly affects the running performance of a moving vehicle. This work therefore contributes to the on-line inversion of the track irregularities and the differential subgrade settlement hidden in the track irregularities is further excavated. To achieve the above goal, a vertically vehicle-track-subgrade coupled dynamics model with high accuracy and efficiency is first established by introducing the Green’s function method. The track irregularities are then generated by a probabilistic model and the rail deflections caused by the settlement are also accounted for. The vertical accelerations of the vehicle excited by track irregularities and various vehicle speeds are subsequently derived. On this basis, a 1-D fully convolutional encoder-decoder network is constructed to predict the track irregularities by treating the acceleration data as the network inputs. A total of seven scenarios involving different input variables are investigated and the results show that when the wheelset, bogie and car body accelerations are simultaneously considered in network training, the prediction performance achieves the optimum. Meanwhile, the network robustness with respect to various vehicle speeds and degradation levels of track irregularities is also demonstrated. Finally, a time–frequency unification method is employed to identify the settlement locations and wavelengths from the predicted track irregularities. Two cases are conducted to further illustrate the effectiveness of the presented settlement identification method.