| Time series analysis is an important part of knowledge discovery and data mining,but the uncertainty,high dimension and dynamic characteristics of time series increase the difficulty of data mining.The cloud model combines randomness and fuzziness to deal with the problem of uncertainty.In terms of time series analysis,the cloud model is still in the initial stage of exploration,at present,there are three time series representation methods based on cloud model: the early cloud segmentation approximate representation method does not consider gradient information and neglects the trend of observation values.The recent piecewise two-dimensional normal cloud representation(2D-NCR)method introduces gradient information of adjacent points,which can directly and effectively reflect the data distribution and changing trend of time series,and has good performance in time series classification experiments.The three-dimensional piecewise cloud representation method further introduces the second-order gradient information and proposes an overlapping partition strategy to improve improve the classification effect of time series classification,but the model becomes complex,and the variable length overlapping region makes the symbolic results difficult to understand.The work of this paper is mainly based on the framework of 2D-NCR method,which is improved from the aspects of adaptive segmentation,similarity calculation and overall gradient information.The main research results are as follows:(1)The problem existing in time series representation which based on 2D-NCR is pointed out: for two points that are closely connected and contain critical information,isometric segmentation may cut the connection between them,resulting in information loss.The 2D-ANCR method is proposed,segmentation based on the sliding window,self-adaptive segmentation of time series according to the internal characteristics of the original time series data.The time series classification experiment shows that this method can effectively reduce the classification error rate.(2)Two problems of the time series similarity measurement method based on2D-NCR are pointed out: no reasonable treatment of the special case of zero similarity in the calculation process,which makes the subsequent calculation results of similarity are unstable;Only the role of the numerical characteristic expectation Ex and entropy En of the cloud model is considered,but the role of superentropy He is not considered.A 2D-NCRWP time series similarity measurement method is proposed,which improves the similarity calculation formula of 2D-NCR,and solves the problem of zero diffusion in the process of similarity calculation.A 2D-ISD time series similarity measurement method is proposed,which applies the strategy of combining shape similarity and distance similarity to time series similarity calculation,and proposes an improved distance similarity.Experiments results show that both methods can effectively improve the classification performance.(3)Based on the framework of 2D-NCR,without significantly increasing model complexity,a similarity measurement strategy based on 2D-ENCR is proposed.This method uses all gradient information instead of first order gradient information to construct the second cloud of two-dimensional cloud model,thus,the trend information of time series is applied more comprehensively.The time series classification experiment proves that this method has better classification effect than the existing cloud similarity measurement strategy. |
