Research On Track State Prediction And Anomaly Detection Of Track Dynamic Inspection Data

The development of high-speed railway has accelerated,not only making transportation conveniently,but also driving economic development.The track dynamic inspection vehicle plays an important role of security on the safe operation of high-speed railway.it is able to judge whether the track state is safe or not by analyzing track dynamic inspection data collected by track inspection vehicle.At present,the main data mining research on track dynamic inspection data is still in the stage of simple statistical analysis,so it is of great significance to carry out in-depth mining of the relevant data.But there are many challenges for research on track dynamic inspection data,such as a large number of noises,complex causality between different detection variables and data imbalance.In response to these challenges,from the perspective of the intersection of rail transit and artificial intelligence,this thesis studies the methods of track state prediction and anomaly detection of track dynamic inspection data.The main work of this thesis is summarized as follows:(1)Aiming at the challenges that a large number of noise and complex causality between different detection variables in track dynamic inspection data,we propose a multiattention neural network for the state prediction of track dynamic inspection data.The model constructs the multi-attention block and the dilated convolutional neural networklong short term memory network block by dilated convolution and long short term memory network.The multi-attention block includes the temporal attention mechanism and the dimensional attention mechanism,which can capture different pattern on temporal and dimension.The dilated convolutional neural network-long short term memory network block learns the dilated convolution results of the various dilation rates by long short term memory network,so as to capture the temporal feature under different views.In addition,the gated recurrent unit-autoregression block in the model learns the temporal features of the original data,improving the feature extraction ability of the model,finally the multi-layer feedforward neural network is used to integrate the results of different blocks to produce the prediction result.In the experiments of multiple prediction target tasks in track dynamic inspection data of a certain section of a high-speed railway,the prediction errors of the model are all lower than those of the comparison methods,which shows the effectiveness of the model proposed in this thesis.In addition,all blocks are proved to improve the prediction accuracy of the model.(2)Aiming at the challenges that data imbalance in dynamic inspection data,we propose an anomaly detection model based on autoencoder for track dynamic inspection data.The model is an denoising autoencoder which combines temporal convolutional network and long short-term memory network.The model can effectively reduce the influence of noise in the data by denoising processing,the encoder is constructed by combination of temporal convolutional network and stacked long short term memory network,while the decoder is constructed by stacked long short term memory network.The model makes anomaly judgment by reconstruction error for the anomaly detection of track dynamic inspection data.On the anomaly detection experiments of four typical anomaly waveforms in track dynamic inspection data,the model is better than the comparison model in all indicators,which shows the effectiveness of the model proposed in this thesis,which proves the effectiveness of the model.In addition,the effectiveness of each component and the overall structure of the model is proven by ablation experiments,and the optimal parameters are determined by hyperparameter selection experiments.