Research On Anomaly Detection Method Of Time Series Data Based On Neural Network

With the rapid development of the Internet and sensor-based Io T technologies,applications such as network transactions,security monitoring,telecommunications management,and industrial automation built on such technologies have generated and accumulated massive time-series data to be analyzed.However,various factors cause noise or data loss in these time series data.As a result,the data contains anomalies,which severely affects applications and services based on these data,and also sets new requirements for anomaly detection of time series data.Depending on the state of the data,the anomaly detection of time series data can be divided into anomaly detection of data in the time series database where data collection has been completed and anomaly detection of real-time online time series data(stream data).For the time series data in the database,traditional time series data anomaly detection algorithms can be used for detection.However,anomalies can be divided into random anomalies and event anomalies,and existing algorithms cannot further distinguish these two anomalies.These algorithms either repair all anomalies during data repair to lose event information,or treat random anomalies as event anomalies during event detection to increase the false alarm rate.Therefore,it is necessary to design algorithms to distinguish anomalies.Stream data is a large amount of time-series data that arrives in rapid succession in sequence.There is a concept drift phenomenon in stream data,and the real-time requirements are very strict.If its anomaly detection model and threshold cannot be updated in real time,the accuracy of anomaly detection will continue to decrease with time.Therefore,this paper takes the time series data in the database and the stream data as the research object,and researches and analyzes the anomaly detection of these two types of data.The main work of this paper is:(1)A method of detecting and repairing time series anomaly based on the stacked denoising autoencoders(SDAE)and multi-sensor cooperation is proposed.This method is based on the collected time series data,and is used for abnormal detection of data in the central database of the system.This method divides the data into two cases: single-dimensional and multi-dimensional.For single-dimensional data,SDAE is used to detect and repair abnormal data.For multi-dimensional data,the collaboration between multiple sensors is used to classify random and event anomalies on the basis of anomaly detection.Experiments are performed on simulated data based on Intel lab temperature data and compared with the other two methods.The experimental results verify that the method has a higher accuracy rate of anomaly detection and can distinguish between the two abnormalities more accurately.(2)A method of detecting stream data anomaly based on Long Short-Term Memory Network(LSTM)and sliding window is proposed.This method is aimed at streaming data and is used for online anomaly detection.This method uses LSTM to realize the prediction and difference calculation of stream data,and uses sliding window for distribution modeling to realize the dynamic determination of anomaly scores,so that the threshold of anomaly judgment can be maintained,and the problem of stream data concept drift is solved.Experiments are performed on the simulated data based on the pressure data of the hydraulic state database,and compared with the other two methods.The experimental results verify that the method has a higher accuracy rate for anomaly detection of convection data.