Anomaly Pattern Detection Of Electric Power Sensor Data

With the development of the electric power Internet of Things technology,the electric power system has generated a large amount of electric power sensing data,which reflects the operating status of the entire electric power Internet of Things system.However,because of the complex environment and transmission process of electric power sensing data,abnormalities are common.Electric power sensor data is a type of typical time-series data.This type of abnormality not only refers to data point abnormalities,but also includes continuous data segment abnormalities,that is,pattern abnormalities.Therefore,this paper makes a specific study on the pattern anomaly of electric power sensor data.First,the main characteristics of the electric power sensor data and the existing anomaly problems in the electric power sensor data are analyzed.Through the analysis of the existing exception handling methods,the advantages and disadvantages are found,and the improvement direction is proposed.Secondly,aiming at the problem of pattern anomaly of electric power sensor data,data characteristics of electric power sensor data is fully considered,and three anomaly pattern detection methods for electric power sensor data are proposed.They respectively aim at the anomaly changing trend,changing amplitude and density distribution in hyper-dimensional space of sample data values that actually exist in the process of electric power sensor data collection,in order to achieve low false alarm rate,high accuracy rate and high recall rate under the premise of low time overhead.The main research contents of this article are as follows:First,aiming at the problem of anomaly changing trend in the values of electric power sensor data samples,an anomaly detection method based on the frequency domain characteristics of electric power sensor data is proposed.The method extracts the frequency domain features of the data by Fast Fourier Transform,and uses statistical values to describe the frequency domain features to reduce the dimension of the feature space,and then completes the fast processing of high-dimensional data based on the Isolation Forest algorithm to obtain the anomaly score of electric power sensor data.Based on this,we can judge whether the data is anomaly or not and the anomaly degree of the data.At the same time,in order to meet the needs of real-timeanomaly detection,this paper proposes an anomaly pattern detection method for streaming power sensing data,which is based on the sliding window mechanism.Second,aiming at the problem of anomaly changing amplitude in the values of electric power sensor data samples,an anomaly detection method based on the motion characteristics of electric power sensor data is proposed.The method considers the motion characteristics of the power data,and uses the speed,acceleration and other characteristics of the electric power sensor data value to calculate the motion characteristics,and then uses the Isolation Forest algorithm to determine the anomaly situation.Third,aiming at the problem that the values of the electric power sensor data samples show anomaly density distribution in hyper-dimensional space,an anomaly pattern detection based on the spatial density of the electric power sensor data is proposed.The spatial distribution of the data is used to perform data modeling of the density model.Then,the values of each sensor are used to match the density interval in the model,and the density values of each sensor are obtained,and anomaly detection is performed based on the density values.In order to verify the feasibility and effectiveness of the method,an anomaly pattern detection system for power sensor data is designed and implemented,and experiments were performed based on real grid data.The experimental results show that the anomaly detection method based on the frequency domain characteristics has a recall rate of above 82%,a false alarm rate of less than 1%,and an accuracy rate of98%-99% when detecting samples with anomaly changing trend in the values,and the anomaly detection method based on motion characteristics a recall rate of about 85%,a false alarm rate of 0.6%-1.0%,and an accuracy rate of 98%-99% when detecting samples with anomaly changing amplitude in the values,and the spatial density-based anomaly detection method has a recall rate of about 80%,a false alarm rate of 1%-2%,and an accuracy rate of about 97% when detecting samples with anomaly density distribution in hyper-dimensional space.The experimental results of the anomaly detection method for streaming electric power sensor data show that the method can accurately judge the anomaly of more than 98% of electric power sensor devices,and can meet the needs of real-time performance.In addition,compared with DBSCAN and One Class SVM,the anomaly pattern detection method based on frequency domain features has higher accuracy and lower time overhead.