Multivariate Time Series Classification Method And Application Based On Segmentation

Multivariate time series is a sequence composed of multiple time-varying variables,which widely exists in various fields.Due to the massive,high-dimensional and real-time characteristics of multivariate time series,traditional static data mining methods cannot effectively process this type of data.Therefore,how to effectively mine useful information from multivariate time series is a valuable research goal in the field of time series data processing.Among them,the multivariate time series classification problem is a practical task.This paper mainly studies the multivariate time series feature representation and classification,reduces the time series dimension through feature representation,thereby improving the classification efficiency.The main work and innovations are as follows:(1)In this paper,a correlation-based multivariate time series segmentation algorithm(MVC-seg)is proposed.The algorithm fully considers the correlation among multivariate time series variables and between the variables in different time periods.Combined with the idea of sliding window,real-time data processing can be realized.By observing the degree of correlation change in different time periods to determine whether to segment,if the degree of change is greater than the set threshold,then segment.The experimental results show that the proposed method effectively reduces the dimension of time series.Compared with other segmentation algorithms,it not only improves the operation efficiency of classification data mining,but also improves the classification accuracy.(2)A multivariate time series classification method based on improved DTW is proposed.In traditional time series similarity measures,since the DTW algorithm only considers the value of data points on the Y-axis,tries to account for the variability of the Y-axis by wrapping the X-axis,which can lead to over-matching.According to the idea that the trend of the same state time point is basically the same,the trend measurement(that is,the degree of data correlation of the local time series segment)is used to alleviate the dynamic matching problem caused by the drift of the y value.A new distance metric based on the degree of sequence trend consistency that can be extended to multivariate time series is proposed.According to this metric,an improved time series similarity measurement method RDTW is proposed.Combined with the MVC-seg segmentation algorithm as a preprocessing step,a multivariate time series classification method NN-RDTW is designed.Experiments show that this method achieves better measurement results on multiple datasets.(3)On the basis of the above research content,a time series classification prototype system is designed and implemented for the multivariate time series data of ECG in the medical field.The system mainly includes functional modules of importing data,data preprocessing,multivariate time series segmentation and time series classification,which provides a feasible and convenient method for researchers in related fields.The main function modules and operation instructions of the system are introduced in detail.The test results of the system prove the correctness and reliability of the system,and provide a feasible method for time series classification.