Research On Track Dynamic Inspection Data Mining Algorithm For Track State Evaluation

The stability of operation of high-speed railway is related to the safety of people’s life closely.Track dynamic inspection data(TDID)is collected by comprehensive highspeed track dynamic inspection vehicle via a range of sensors thus can reflect the track conditions precisely.Therefore,based on machine learning theory,studying how to mine the evolution of high-speed railways from a large amount of TDID,and how to quickly and accurately judge and eliminate anomalies in TDID has theoretical significance in the field of spatio-temporal data mining.At the same time,it is also of important practical value to make predictions and early warnings for possible future out-of-limit situations,and to provide maintenance personnel with advice on assisting decision-making.As an important infrastructure for high-speed railway,the creep camber of bridge section is susceptible to various factors such as seasons,climate,and physics.The prediction of development of trend is very important for the evaluation of the track condition,and current research is difficult to meet the requirements of accurate prediction demand.At the same time,in TDID,there are few anomaly samples,and the correlation patterns are complicated,and it is inevitable that there are some structural diseases that even inspectors may ignore.Therefore,in response to the above existing problems,based on spatio-temporal data mining method,this article takes into account the characteristics of the multi-dimensionality,spatiotemporal,and sample imbalance of the TDID,and expands the prediction of creep camber of bridge and the anomaly detection of track.The main work and research contents of this paper include:(1)Aiming at the prediction problem of creep camber of bridge,a warning algorithm based on time series components decomposition is proposed.First,use the spatial periodicity presented by the bridge data in the TDID to identify,segment and complete the construction of the value of camber of time series.Then considered the trend and seasonality of the bridge creep sequence,decomposition experiments are carried out through the component decomposition algorithm,which not only removes the noise,but also ensures that the trend and seasonal characteristics are extracted correctly.Finally,based on the prediction algorithm of machine learning,each component is predicted in parallel and then combined to obtain the final predicted value of the camber of bridge.Quarter-long forecast experiments on 101 32m-bridges and 72 24m-bridges of a highspeed railway in my country shows that,compared with the Linear Regression(LR),the prediction algorithm of creep camber of bridge proposed in this paper are improved by31% and 54%,respectively in the best case.(2)Aiming at the problem that the size of TDID is unbalanced,there are few anomaly samples,and the common single-scale auto-encoder model is greatly affected by the window,a method for anomaly detection of TDID through correlation autoencoding and multi-scale fusion is proposed,that is,UADTA-CAMF,An Unsupervised Anomaly Detection Method for Track Dynamic Inspection Data via Correlation Auto Encoding and Multi-scale Fusion.First,the calculation of the correlation matrix is realized and the construction of a multi-scale auto-encoder is carried out.Then,in view of the uncertainty in the multi-scale,the DS evidence reasoning theory is used to fuse the anomaly probabilities from different scales to improve the accuracy of anomaly detection.UADTA-CAMF follows the basic idea of unsupervised and learns the performance pattern of normal data from a large amount of normal data.Finally,the training set based on normal data and the testing set embedded with five typical simulation anomalies are carried out at different times,different types of lines,and different sections.The experimental results show that the UADTA-CAMF model outperforms other multivariate sequence anomaly detection and comparison networks in all indicators for the task of detecting anomaly data in TDID,especially in the experiments on the same type of lines,reaching an accuracy of 0.806 and 0.853 in precision and recall.(3)In the current research on multivariate time series,the relationship between multiple detection items is dynamic and non-linear,which common models are difficult to capture.Based on that,RFADGNN,Region Fusion Anomaly Detection Graph Neural Network for Track Dynamic Inspection Data,is proposed.Firstly,the original TDID was embedded as the graph nodes to construct a graph structure to capture the complex dynamic correlation between the dimensions of the TDID,and then based on the graph attention,the detection items with higher correlation of the current detection items were given higher weights.At the same time,combined with the characteristics of TDID,the regional information is embedded and integrated into the network.In particular,the bridge and the curve are two factors.In the ablation experiment,it was verified that the precision and recall of RFADGNN increased by 6% and 9%,and the TAR(time-series aware recall)increased by 2% after the introduction of regional features.The results of the comparison experiment also show that RFADGNN performs better than other models in the anomaly detection task,especially in the comparison experiment of mixed anomalies,compared with the prediction-based comparison model LSTM(Long ShortTerm Memory),RFADGNN has improved precision and recall by 7% and 4%respectively.