Research And Application Of Time-Series Anomaly Detection Model For Urban Common Trench

Time-series anomaly detection aims to discover special patterns in the corresponding data features that do not conform to general laws,and is a classical problem in the field of machine learning,which has a great significance in both academia and industry.In intelligent operation and maintenance projects of modern Urban Common Trench,in order to improve the comprehensive management ability and maintenance efficiency of the trench,it is necessary to continuously monitor the involved production process.The monitoring method is to continuously collect key time-series data through sensors,then analyze them in real-time modeling to find abnormalities and make preventive measures in time.However,as the construction of China’s Urban Common Trench gradually tends to automation,large-scale,systematic,the collected time-series data is increasingly complex,presenting a large amount of data,non-linear and other characteristics.Traditional methods are less efficient,cannot capture long time-series correlation,and are not good at modeling high-dimensional data,so it is gradually difficult to cope with such complex industrial scenarios.How to efficiently and accurately perform anomaly detection on these massive time-series data becomes an urgent problem.In this thesis,unsupervised time-series anomaly detection methods are investigated with time-series data of Urban Common Trench,and a prototype system for anomaliy detection in time-series of Urban Common Trench is also constructed,the main work is as follows:1.To address the existing problems of low computational efficiency of time-series anomaly detection and difficulty in mining long time-series remote correlations,an anomaly detection model based on long time-series prediction is proposed.The data redundancy is reduced by Kalman filtering,the local correlation between adjacent time-series is captured using inflated causal convolution,and the remote correlation of long-distance time-series is learned using an improved sparse self-attention mechanism.Compared with traditional anomaly detection models,the use of a network structure entirely consisted of attention mechanisms in the modeling phase can effectively reduce time consumption,and improve anomaly detection accuracy due to the ability to capture both local and remote correlations of time series.2.To address the problems of high maintenance cost of existing time-series anomaly detection models and difficulty in effectively exploiting the dependencies between multidimensional time-series variables,a multidimensional time-series anomaly detection model based on generative adversarial network is proposed.The temporal correlations in the normal state are captured using the proposed long time-series prediction model,which is embedded in the framework of generative adversarial,and the generators and discriminators are fully utilized to detect anomalies by discriminating the results and reconstructing the errors.Instead of dealing with each data stream individually,the proposed multidimensional time-series anomaly detection model considers the entire set of variables simultaneously to capture potential interactions between variables,and the trained model can be adapted to multiple corridor subsystems,making it scalable and reducing model maintenance overhead.3.Combining the above two time series anomaly detection models,a prototype system of time series data anomaly detection for Urban Common Trench is designed and implemented by using front-end and back-end separation techniques.The system mainly uses the proposed anomaly detection model to realize the prediction and anomaly detection functions for time-series of the Urban Common Trench.In order to facilitate the operation and maintenance personnel to understand the state of the trench in real time and take timely countermeasures in case of accidents,the system also provides the functions of operation and maintenance information management,trench data visualization,early warning and alarm,which verifies the feasibility of the proposed models.