Research And Application Of Knowledge Extraction Method For Subway Onboard Equipment Based On Deep Learning

With the advancement of urbanization construction,the subway has been widely used in China and has become one of the most convenient and efficient modes of transportation for residents.As one of the most important control equipment in urban rail transit,subway onboard equipment is the key guarantee for the safe operation of subway trains.A significant amount of unstructured maintenance data regarding subway on-board equipment has been produced as a result of the accumulation of subway operating distance and duration;this data contains a sizable amount of text information concerning problems.At present,the utilization rate of such text data is low,so it is necessary to use and rely on knowledge extraction technology to mine the value of text data in subway on-board equipment,comprehensively understand the safety status,rules and influencing factors of subway train operation,and be able to put the value of fault data into many intelligent subway applications such as fault diagnosis,intelligent recommendation and intelligent services,as well as lay the foundation for building a knowledge map in this field.In this dissertation,we study text classification,named entity recognition,and entity relationship extraction algorithms based on deep learning,on the basis of the maintenance data recorded in the vehicle section of a metro company from 2016 to2021.We also combine the existing fault data and knowledge extraction algorithms to form a dataset for training fault diagnosis models and realize the initial application in fault diagnosis.The main work of this dissertation is as follows:(1)A BERT(Bidirectional Encoder Representation from Transformers)and Bi GRU(Bi directional Gate Recurrent Unit)fault text classification model based on key layer fusion is proposed to address the problem of low classification accuracy and incomplete classification performance in unstructured fault data automatic classification tasks.The model processes the text data into word vectors with location information in the word embedding layer and inputs them to the BERT layer firstly.Based on the conventional 12-layer BERT model,the semantic information is fully obtained for fusion and dimensionality reduction by encoding the bidirectional Transformer Encoder in layers 2,4,6,8 and 12,and is then input to the Bi GRU layer to extract contextual information.The final fault classification results are obtained in the output layer.The experimental results show that the model performs better in three evaluation metrics and converges faster in the training process.(2)A deep neural network-based named entity recognition and entity relationship extraction model for subway on-board equipment has been proposed.Firstly,Bi LSTM(Bi directional Long Short Term Memory)and CNN(Convolutional Neural Network)parallel networks are used to extract contextual and local feature information;Secondly,in the MHA(Multi Headed Attention)layer,information from different dimensions in the data is fused and the dependency relationships between each unit in the sequence are captured to obtain structural information within the feature sequence;the inner connection between the tags is learned by CRF(Conditional Random Field)to ensure the sequential nature of the tag sequences.Finally,the Transformer model is used to extract knowledge triplets.The experimental findings demonstrate that the Transformer model has superior recognition performance for all five of the entity connection categories specified in this study,whereas the named entity recognition model proposed in this dissertation has higher recognition accuracy.(3)For subway on-board equipment,a problem diagnosis model based on CIWOA(Continuous Improvement Whale Optimization Algorithm)-BP(Back Propagation)is suggested.The findings demonstrate that the CIWOA method put forth in this dissertation is superior to other conventional optimization algorithms for optimizing the fault diagnosis model based on BP neural network.The experimental results demonstrate that the CIWOA method described in this research is more efficient in optimizing the fault diagnosis model based on BP neural network than other conventional optimization techniques.