Research On Intelligent Diagnosis And Maintenance Decision Method Of Train Control On-Board Equipment In High-Speed Railway

As a successful example of China’s independent innovation,the high-speed railway has witnessed a rapid leap in China’s comprehensive national strength.With the train control system and related equipment becoming larger and more complex,its operation and maintenance costs are also increasing.The upgrading of the train control equipment not only ensures the efficient operation of the system,but also increases the risk of hidden failures.Currently,the evaluation of train control on-board equipment by railway site staff still adopts the diagnosis and maintenance mindset of emphasizing experience over science,which is challenging to support the demand for efficient maintenance.As an innovative approach,intelligent technology has developed to a new stage with characteristics such as deep learning,multi-source integration,and knowledge guidance.As an important equipment in high-speed railway train control system,on-board equipment is faced with practical challenges such as weak automatic fault diagnosis ability and lack of scientific and reasonable operation and maintenance guidance.In order to improve the capability of intelligent diagnosis and maintenance of train control on-board equipment,this thesis takes data-driven intelligent diagnosis and knowledge-guided maintenance assistant decision-making as the research context.The research focuses on critical methods such as on-board equipment fault module localization,fault component unit detection,on-board maintenance knowledge extraction and knowledge fusion,and knowledge graph-driven on-board maintenance decision-making.The main work is reflected in the following points.(1)Aiming at the problems of difficulty in automatic fault localization,hard fault feature extraction,and low fault localization precision of train control on-board equipment modules,a module-level fault location method for on-board equipment combining deep features and cost-sensitive learning is proposed.Firstly,a distributed word embedding representation of on-board operational state data is established based on the continuous bag-of-words model to solve the representation sparsity and semantic orthogonality problems in the conversion from data to vector space.Secondly,a multi-scale convolutional neural network deep feature extraction model with batch normalization is constructed to capture the differential features strongly related to the fault module from the operational state data and reduce the impact of internal covariate shift on the network training effect.On the basis of deep fault features,a random forest diagnosis model based on a cost-sensitive learning mechanism is constructed and cascaded with the feature extraction network to improve the fault diagnosis errors cost of the model and enhance the fault location effect of functional modules.Experiments are carried out on the on-board operational state data set.The results show that the model has advantages in fault feature extraction at the module-level of on-board equipment,as well as precision and recall of various fault types location.The F1 value for the overall fault location can reach 90%.(2)The functional module of the on-board equipment includes multiple component units.Aiming at the problem that the fault detection effect of each component unit is affected by the unbalanced distribution of samples,a fault detection model based on an attention capsule network is proposed.Firstly,the on-board operational state data is jointly represented by distributed word embedding and context features.Assigning weights to word embedding representations based on an attention mechanism to improve the ability of the model to focus on more discriminative fault features.Secondly,a fault feature extraction and diagnosis model based on the capsule network is established.By removing the pooling operation and the dynamic routing mechanism between capsules,the bottom-level features of the local position to the whole structure correlation in the on-board data are abstracted to the top-level feature space to improve the fault feature extraction capability and reduce the sequence feature damage caused by information filtering.Then,a multi-class focal loss function is constructed for the sample imbalance problem to dynamically balance the influence of minority and hard samples on the gradient descent in model training and improve the diagnosis precision of the model for on-board component faults.Compared with other fault diagnosis algorithms with outstanding performance on the on-board operational data set,it is proved that the model proposed in this thesis can effectively improve the diagnosis precision and recall of component faults under unbalanced conditions,and the F1 value of fault detection reaches88%.(3)Aiming at the problem that the train control on-board maintenance data are multi-source and heterogeneous,and it is difficult to be organized and expressed,a method of knowledge modeling and knowledge extraction for on-board equipment maintenance is proposed.Firstly,a knowledge modeling method based on knowledge graphs is proposed to achieve an efficient integrated expression from maintenance data to knowledge.Through a two-way collaboration between the top-down conceptual pattern layer and the bottom-up maintenance data layer,maintenance elements and the organizational structure of knowledge are clarified.Secondly,under the guidance of the pattern layer,an entity recognition model based on multi-neural network collaboration is proposed to obtain the core units of maintenance knowledge,namely entities.Based on the representation of maintenance data by composite features of character granularity,a knowledge feature extractor cascaded by a gated convolutional and bi-directional long short-term memory network is established to enhance the extraction effect of complex semantic features of knowledge-intensive data.Combined with the conditional random field under the neighborhood label constraint,the global optimal labels of entities are jointly decoded to achieve structured maintenance knowledge extraction.The knowledge extraction effect of the entity recognition model is verified on the on-board maintenance corpus.The model can improve the distinguishability of the maintenance entity boundary,and the F1 value of knowledge extraction can reach 89%.After adding the composite feature representation to the nine comparison models,the knowledge extraction effect of each model is improved,proving the advantages of composite features in domain knowledge extraction tasks.(4)Aiming at the maintenance guidance requirements of multi-source knowledge association of on-board equipment,the research on multi-source knowledge fusion and knowledge graph-driven maintenance-assisted decision-making is carried out based on the acquired maintenance knowledge.Firstly,to solve the problem of diverse representations and semantic inconsistencies in the extracted maintenance knowledge,an entity alignment model based on the combination of BERT and semi-supervised incremental learning is proposed to fuse the redundant or ambiguous knowledge units.In the absence of context for knowledge units,the deep semantic features of maintenance knowledge are extracted through the BERT model based on transfer learning.Combined with the semi-supervised incremental learning strategy based on pseudo-labels,the model cooperatively uses pseudo-labeled data to improve knowledge fusion precision and introduces embedding-based adversarial training to enhance the robustness of the model.Experiments are carried out on the on-board maintenance knowledge data set,and the precision and recall of this model for knowledge fusion are above98%,which can effectively improve the quality of maintenance knowledge.Then based on the fused standard knowledge resources,a multi-source information association knowledge graph for on-board maintenance is established.Guided by the fault diagnosis results of on-board equipment driven by operational state data,it is mapped to the knowledge graph,and the knowledge-driven maintenance decision-making for on-board equipment is realized through knowledge reasoning.Based on the above research,a prototype system for intelligent diagnosis and maintenance of high-speed railway train control on-board equipment is designed.Through the development of the application system,this thesis proves that the research results can provide practical and effective theoretical and application reference for the intelligent diagnosis and maintenance-assisted decision-making of the functional modules and component units of the high-speed railway train control on-board equipment.