Clustering Algorithm For Multivariate Time Series And Its Application

Multivariate Time Series(MTS)is a collection of observations of multidimensional variables recorded in time order.Clustering technology is one of the important methods to analyze MTS.The existing model-based MTS clustering algorithm has the problem of high time complexity when processing the MTS data set with unequal-length samples.The Quick Access Recorder(QAR)records a series of parameters during the flight of an aircraft and the QAR data is a typical MTS sample.The Flight Operation Quality Assurance(FOQA)ensures the flight safety of civil aviation by analyzing QAR data.However,the FOQA can only detect exceedance according to monitoring items and cannot discover potential outliers outside the monitoring items.Therefore the following research is carried out around the multivariate time series clustering and its application in anomaly detection on QAR data:(1)An MTS clustering algorithm based on Lift Ratio(LR)component extraction(MUTSCA <LRCE>)is proposed.The algorithm uses the equal frequency discretization to symbolize the MTS and computes LR vectors used to express the temporal pattern between time series of the MTS sample.It extracts a fixed number of different key components from both ends of each LR vectors after sorting.All the extracted key components are spliced to form a feature vector representing the MTS sample.This process converts the unequal MTS sample set to a set of equal-length feature vectors.Finally,a k-means algorithm is used to cluster the set of isometric feature vectors.Experiments on multiple common data sets show that compared with the existing methods,the proposed algorithm can significantly improve the clustering speed of MTS data sets with unequal length on the premise of ensuring the clustering effect.(2)An anomaly detection method on QAR data set based on the Hierarchical Clustering based Principal Component Analysis(HC-PCA)is proposed.The method firstly uses the MUTSCA<LRCE> clustering algorithm in(1)to perform cluster analysis on the QAR data set of flights to find outliers,and then converts the QAR data of outlier flights to a single-dimensional time series represented by cosine angles.The sliding window is usedto extract equal-length features of sub-sequence from the cosine-angle sequence,and this features are used to generate a matrix,then the PCA is used to reduce the dimension of this matrix.The top-down hierarchical clustering is performed on row vectors of the reduced matrix according to the amount of information within each column.The anomaly nodes are detected according to the number of vectors contained in the clustering tree nodes;The abnormal data segments of outlier flight data are generated by merging the positions of vectors generated in the cosine angle sequence in the abnormal nodes.The experimental results show that not only can the method detect abnormal flights,but also locate the abnormal data within QAR samples.