Research On Track State Analysis Based On Cloud Platform And Deep Learning

The rapid development of China railway industry has put forward higher requirements for the stability,safety and durability of the track.With the more increase in the speed of high-speed railways,the interaction between the carbody and the track is more intense,which makes carbody damage the track more seriously.At the same time,deterioration of the track status in turn threatens the running safety of the train.Therefore,it is of great research value to combine the synthetic track state analysis of the geometric irregularity of the track and the dynamic response of the carbody.With the rapid development of new-generation Internet technologies such as cloud computing and big data,people can use high-efficiency and low-cost computing resources to analyze large amounts of data.This has made it possible to analyze massive orbital irregularity data accumulated by railway works departments.This thesis starts from the measured track irregularity data and constructs the track irregularity-train vibration state prediction analysis model by using the powerful computing ability of the cloud platform,the machine learning and deep learning theory.The main work of the thesis includes the following aspects:Firtly,in order to speed up the track irregularity feature extraction,this thesis constructs a time-series parallel decomposition model based on Spark.In the beginning the time series is divided and redundant,and then the redundant time subsequences are distributed to the computing nodes of the Spark cluster for sequence decomposition.Finally,the intermediate results of the computing nodes are combined to obtain the result.Experiments show that the proposed model can efficiently decompose large-scale time series.What's more,to solve the problem of weak generalization ability of traditional vibration state prediction model,this thesis designs a prediction method based on stacking ensemble model for track-carbody vibration state.In the beginning,wavelet decomposition is used to filter out the high-frequency noise in the track irregularity data.Then,the track irregularity features are extracted from two angles: time domain and frequency domain.Then the sampling technique is used to alleviate the imbalance of the sample.The feature selection is based on the tree model to remove the redundant features.Finally,a two-tier stacking ensemble model is constructed.Experiments show that the ensemble method can well predict the response of track irregularity to carbody vibrationIn the end,the CNN-based track-carbody vibration state prediction method is studied.In the beginning,the track irregularity data is used to construct the data input layer of the CNN network and then it is mapped as an RGB picture as the input of CNN.Then,the network structure of CNN applicable to the track-carbody vibration state is designed.This method can avoid complicated feature extraction and selection process.Experiments show that the prediction accuracy of this method is higher than the existing vibration state prediction methods,and the final prediction result has an average F1 value of 96%,which is 2% higher than the stacking ensemble prediction model.This thesis mainly proposes two kinds of track-carbody vibration state prediction models,e.g.,Stacking model and CNN model.From the model prediction accuracy,the CNN model is superior to the Stacking model.From the perspective of model efficiency,the Stacking model takes advantage of the parallel computing of the cloud platform and is more efficient than the CNN model.In terms of modeling complexity,the CNN model automatically learns features,so the modeling complexity is lower than the Stacking model.Therefore,focusing on efficiency in practical projects can consider the use of Stacking models,and more emphasis on accuracy can be considered using the CNN model.