As technology advances, a large number of time series data have emerged in all walks of life. Clustering is a key technique for analysing time series data. However, most of the existing clustering methods calculate the distance of a single discrete data point, but cannot be applied to continuous time-series data with structural distortion (e.g., expansion, contraction, and drift) and noise (e.g., pseudo-event), resulting in low clustering accuracy. In this paper, a novel time series event clustering approach called CBR(Clustering Based on Representative sequences) is proposed. We first introduce a cross-correlation method to measure the distance between sequences with structural distortion, and propose an r-nearest neighbor evaluation system for sequences to construct candidate sets of R-Seqs(Representative sequences) and eliminate pseudo-event interference. Secondly, we formulate composite selection approaches for R-Seqs based on combinatorial optimization and diversifying top-k query to rapidly derive the R-Seqs optimal solution from the candidate sets. Finally, relying on the dynamically constructed distance matrix of R-Seqs and dataset, a matrix clustering method based on K-means is proposed to achieve an efficient division of event classes. Experimental results demonstrate that CBR is superior to the existing approaches in clustering accuracy, efficiency and denoising quality, especially the clustering accuracy is improved by more than 30% on average .