Anomaly Detection For Vehicle Maintenance Electronic Health Records Data

With the rapid growth of the car ownership in China,vehicle maintenance has become an important livelihood service industry that is related to the quality of people living.Vehicle maintenance electronic health records are the core data of the vehicle maintenance industry.During the collection and application of vehicle maintenance electronic health records,a great number of abnormalities inevitably exist within the vehicle maintenance record data.Correctly distinguishing between normal and abnormal record data can protect the legitimate rights and interests of consumers,and it has great significance for improving the management efficiency and service quality in the vehicle maintenance industry.The right detection of data can promote the transformation and upgrading of the vehicle maintenance industry and improve the car living standards of citizens.For these practical problems of abnormal vehicle maintenance record data,the main following works are addressed in this paper.(1)A model for abnormal detection of vehicle maintenance electronic health record data based on a deep sparse auto-encoder is proposed to accurately detect the abnormal record data.The Doc2 vec and Glove models are used to convert the natural language description fields in the original data into vectors.In order to improve the reconstruction coding ability of the proposed model,the one-dimensional regular sparse suppression is added to the first layer of the encoder in the deep auto-encoder,which only allows a small number of neurons to be active.The added layer improves the model by learning more representative properties of the data.Anomaly discrimination is realized by the reconstruction error between normal and abnormal data records.(2)For the situation that the vehicle maintenance record data contains rich technical knowledge about vehicle maintenance,a vehicle maintenance technology knowledge graph based on vehicle maintenance record data is further constructed.A classification system of entities and relationships in the vehicle maintenance field is developed under the guidance of vehicle maintenance field experts in the first time,and build a vehicle maintenance knowledge graph based on the classification system.The entity and relationship analysis,network quantitative analysis and complex network analysis are carried out for the knowledge graph,which prove that the scale of the constructed knowledge graph will not increase infinitely with the expanded vehicle maintenance record data and meet the characteristics of complex networks.(3)To detect the abnormal graph data within the constructed vehicle maintenance technical knowledge graph,a deep-deviation-network-based anomaly detection model is proposed for vehicle maintenance technical graph data.The Graph2 vec is applied to convert the initial graph data into vector data.In order to improve the anomaly detection effect,the deep deviation networks combining both feature representation learning and anomaly detection are used to detect anomalies in the vehicle maintenance technical graph data.Anomaly discrimination is achieved by using the anomaly scores of normal data and anomalous data.