Abnormal Change Detection In The Long Running Process Of A Dynamic System

During the long time operation of complex structure dynamic systems,a series of measurement data is generated.These time-varying measurement data is called time series.However,in the actual operation of the system,the operating state of the system may have abnormal changes due to external interference or self failure.These abnormal changes can affect the stability and safety reliability of the system.In order to detect the abnormal changes of the system in time and make effective preventive measures,it is of practical significance to study the anomaly detection of time series.The paper summarizes three common anomalies in time series.They are point anomalies,pattern anomalies and sequence anomalies.This paper mainly focuses on the research of time series pattern anomaly detection.The time series pattern anomaly detection is researched from one dimension time series pattern anomaly detection and multidimensional time series pattern anomaly detection.In one dimensional pattern anomaly detection,data is preprocessed first and piecewise linear representation and pattern representation of time series are carried out.Based on the similarity measurement and classification of time series,a time series pattern representation method based on morphological change is proposed.In one dimension time series data pattern anomaly detection,the improved DTW algorithm uses the similarity coefficient to embody the pattern anomaly,the anomaly detection algorithm based on the pattern feature of the k-nearest neighbor uses the anomaly factor to measure the pattern anomaly.The multivariate dynamic system measurement data anomaly detection algorithm based on Q-filtering and LSTM neural network are proposed.The above two algorithms are anomaly detection from the perspective of filtering and prediction respectively.A set of dynamic system measurement time series data anomaly detection system is developed independently.In the system,the momentum wheel speed data of the actual aircraft is used to verify the algorithm.The results show that the algorithm has good feasibility and applicability in anomaly detection,and can improve the accuracy of anomaly detection.The innovations of this paper include the use of combined field method to preprocess the time series data,the improvement of the pattern representation method to realize the morphological representation of time series data,the implementation of MW-DTW algorithm and pattern feature based on k-nearest neighbor algorithm to achieve anomaly detection of one dimensional time series data,the optimization of Q-filtering detection algorithm and LSTM neural network algorithm to achieve anomaly detection of multidimensional time series data.The detection of abnormal changes in time series data is the core part of real-time monitoring system with fault application,such as equipment monitoring,fraud detection,network intrusion detection and other fields.This study provides practical methods and technical tools to detect abnormal changes in time series data of dynamic system.This paper is supported by the National Natural Science Foundation of China(Grant No.61473222,91646108).