Time Series Anomaly Detection Method And Application Based On Autoencoder And HMM

Anomaly detection is an important issue in the study of time series data mining.Hidden Markov Model(HMM)can be used to establish a parameterized model of nonstationary signals,which has been introduced into the research of time series anomaly detection.The existing symbolization methods of HMM-based time series anomaly detection models cannot well characterize timing characteristics.Flight Operation Quality Assurance(FOQA)finds exceedance from Quick Access Recorder(QAR)data according to predetermined detection standards to ensure the safety of civil aviation flight.Long touchdown exceedance is a type of unsafe events that increases the risk of runway excursion.The current FOQA uses a single parameter——the landing distance,to define the long touchdown exceedance,which cannot be combined with multiple parameters to detect and explain the cause of the exceedance.In response to the above problems,this article has carried out the following work:(1)An Autoencoder and HMM-based time series anomaly detection method(Autoencoder and HMM-based Anomaly Detection,AHMM-AD)is proposed.The method first segments the time series samples through a sliding window,and trains each segmented autoencoder from the segmented sample sets at different positions on the normal time series.Then use the Autoencoder to get the low-dimensional feature representation of each segmented time series sample.Through the K-means clustering processing of the low-dimensional feature representation vector set,the time series sample set is symbolized.Finally,the HMM model is generated from the symbol sequence set of the normal time series,and the anomaly detection is performed on the output probability value of the sample to be tested in the established HMM model.The experimental results show that the AHMM-AD method has significantly improved accuracy,recall rate and F1 value compared with existing HMM-based time series anomaly detection models and Autoencoder-based time series anomaly detection models.(2)An improved detection method for the long touchdown exceedance based on improved AHMM-AD is proposed.According to the characteristics of QAR data,this method divides the QAR samples of each flight into data segments that can extract the characteristics of the same flight stage,and uses the symbolization method based on LSTM Autoencoder to characterize the intercepted data segments.The HMM model in(1)is used to detect flights where happened the long touchdown exceedance,and then the Viterbi algorithm is used to determine the data segment corresponding to the exceedance.The experimental results on real QAR data show that the method proposed can effectively detect the long touchdown exceedance,and can combine multiple QAR parameters to assist field experts to explain the cause of the exceedance.