【摘要】 时间序列分类（TSC）是将时序数据按其动态模式划分到预定类别的任务。现实世界的时间序列通常包含趋势项、季节性分量、异常值及噪声的复杂耦合，其精准分解对分类性能提升至关重要。因此，提出一种时间序列分类方法 TSSEA，其通过FFT和STL将时间序列分解为3个视图：时间、频率和季节。基于这些视图，通过编码器间的对比学习实现迭代学习。对3个现实世界的数据集进行大量实验，结果表明，所提出的TS-SEA方法在处理多样化的时间序列应用时，相较现有方法表现出最佳的性能。更多还原

【Abstract】 Time series classification(TSC) is the task of categorizing sequential data into predefined classes according to their temporal patterns. Real-world time series usually contain complex coupling of trend terms, seasonal components, outliers, and noise, and its accurate decomposition is crucial to improve classification performance. Therefore, a time series classification method, TS-SEA, is proposed, which decomposes the time series into three views: temporal, frequency, and seasonal by means of FFT and STL. Based on these views, iterative learning is realized by means of contrast learning between encoders. The results indicate that in comparison with existing methods, the proposed TS-SEA method can exhibit the better performance when dealing with diverse time series applications.更多还原