Research On Fault Diagnosis Method Of Urban Rail Transit CBTC System Based On Text Mining

As the urbanization process accelerates,urban rail transit,as a convenient and fast way of travel,is becoming more and more important for the daily lives of residents.As an important control system for urban rail transit,the CBTC system is the key to ensure the safety operation of trains.Therefore,it is significant to resume the operation of trains in order to locate the position of failure in time when the system fails.The fault diagnosis method used in the field is usually that the professional maintenance personnel shall conduct troubleshooting on each device to determine the cause of the fault according to the situation.This method mainly relies on the experience of the maintenance personnel,and the degree of automation is low.The maintenance process generated during the maintenance data contains a large amount of information related to the fault.Currently,the utilization rate of this data is low.Therefore,this article mainly takes the maintenance data generated during the maintenance process of the CBTC system as the research object,and implements fault diagnosis by text mining the data.This paper analyzes the maintenance data of the CBTC system,uses the BERT model to vectorize the text of the data,and implements semantic enhancement through the Bi-GRU-ATT network,and uses the PSO algorithm to perform hyperparameter tuning on the model.A new type of diagnostic architecture to achieve fault diagnosis.The research content mainly includes the following aspects:First,by analyzing the target data,combining the characteristics of the data and the system structure to divide the first-level fault and the second-level fault,and establish a hierarchical relationship between the two-level fault,and use the BERT model to perform semantic extraction and text vector generation of maintenance data.It is compared with the vector representation generated by Word2vec-TF-IDF to verify the effectiveness of BERT in semantic extraction and text vectorization.Secondly,since loss function plays an important role in guiding the training direction of the model,Focal Loss was taken as the loss function of the model proposed in this paper,inspired by the Focal Loss used to solve the problem of data imbalance in image processing,aiming at the characteristics of unbalanced data distribution.In the process of selecting the downstream task network structure,the performance of the data in four different hidden layer structures is compared through experiments,and the appropriate network structure is selected,and then the model is optimized by PSO algorithm.Thirdly,a parallel diagnosis architecture is proposed,and the diagnosis results that do not meet the requirements are fused at the decision layer through error correction strategies.The experiment proves that the final diagnosis result is improved by two percentage points on average compared with the secondary diagnosis result.This method effectively reduces the misdiagnosis rate of the model.Finally,based on the above algorithm,the validity of the model is verified through experiments and the diagnosis platform is built.The experimental results show that for the first-level faults,the accuracy rate can get up to 90.2%,and for the second-level faults,the accuracy rate can get up to 80.1%.There are 49 pictures,11 tables,55 references.