| Time series clustering can transform massive time series information into organized category information.Due to the characteristics of high dimensionality and nonlinearity of time series,most clustering algorithms cannot be directly applied to the original time series data.Existing time series clustering methods usually use representation learning to extract time series features first,and then perform cluster analysis.However,these methods are difficult to intuitively define the similarity of time series,and it is difficult to capture the long-temporal dependence and cross-variable correlation of multivariate time series data,relying heavily on complex feature extraction networks.In order to obtain the time series feature representation suitable for clustering and improve the accuracy of time series clustering,this paper studies the representation of time series based on the contrastive learning method,and proposes a time series contrastive clustering method and a representation learning approach for contrastive clustering of multivariate time series.The specific work content is summarized as follows:(1)A contrastive learning framework is introduced in time series clustering,and a method for time series contrastive clustering is proposed.In order to solve the problem that the time series similarity is difficult to define in deep clustering,the interval similarity of time series is defined from the perspective of positive and negative sample data based on the idea of contrastive learning.At the same time,in order to solve the problem that the existing time series data augmentation methods are difficult to describe the transformation invariance of time series,a data augmentation method suitable for contrastive learning is proposed from the perspective of time series shape stability.Through a large number of experiments on 32 datasets of UCR,the effectiveness of the two methods is verified.On 15 of the data sets,the clustering results are better than the existing methods,and the improvement rate is above 0.01.In the remaining 17 The experimental results on this dataset are basically the same as the baseline.(2)During the research process,it was found that the above-mentioned time series comparative clustering method could not capture the time dependence of data when faced with multivariate time series,and it was difficult to balance the influence of different variables on the overall characteristics.To solve this problem,a representation learning method suitable for multivariate time series contrastive clustering is proposed.This method designs a data mixing augmentation method for multivariate time series,and uses Transformer as an encoder network to extract the long-temporal dependence and crossvariable correlation of data.The effectiveness of the model is verified by comparative experiments on the UEA dataset,and the robustness of the model is verified by experiments on the univariate and hyper multivariate time series datasets.This paper revolves around the problem that the existing time series clustering framework is difficult to define data similarity,and it is difficult to capture the time correlation and cross-variable correlation of multivariate time series data.The comparative learning framework is integrated into the time series clustering problem,which is applicable Research on Time Series Data Augmentation Methods and Representation Learning Networks in Contrastive Learning Framework.The research results show that using the contrastive learning framework for time series clustering can effectively improve the feature extraction effect of the representation network and improve the clustering performance. |
