Purposes Time series data are widely used in fields such as finance, healthcare, industry, and meteorology. Existing methods often focus on instance-level learning, with limited success in modeling the multi-scale and fine-grained characteristics of time series. Moreover, many methods rely on inductive biases from computer vision or natural language processing, which are unsuitable for time series due to their strong time dependencies. This paper proposes a time-series decomposition-based contrast learning framework (Seasonal-Trend Universal Time Seires Representation Learing,STURL) that addresses these challenges by leveraging hierarchical contrast learning of seasonal terms to capture multi-dimensional features. Methods Additionally, a frequency mixing strategy based on Fourier transform enhances the representation of seasonal terms, improve model robustness and stability. Results The experimental results demonstrate that STURL exhibits outstanding performance across various downstream tasks and surpasses baseline methods. These findings highlight the framework’s superiority in improving model accuracy, reducing prediction errors, and enhancing anomaly detection capabilities, underscoring its potential and feasibility for practical applications.
TNC[7]利用时间邻域作为ADF测试的阳性样本,并通过调整样本权重来缓解采样偏差。ELDELE et al[8]提出了一种新的时间对比模块,通过设计硬交叉视图预测任务来学习时间依赖性,强迫模型通过更难的预测任务来学习鲁棒的表示。HAJIMORADLOU et al[9]引入了沿时间和实例维度的相似性蒸馏来进行预训练。YANG et al[10]提出了TimeCLR,通过最小化样本的两个增强视图之间的相似性来学习不变表示。WANG et al[11]结合FC图构造与移动池GNN,能够学习时间序列的空间和时间特征,确保时间一致性并考虑空间相关性,这对下游任务至关重要。MORGADO et al[12]使用视听空间对齐作为360°视频数据的接口任务,涉及空间错位的音频和视频剪辑。HARESH et al[13]则通过联合使用时间正则化项来鼓励不同帧映射到嵌入空间的不同点。CHEN等人[14]提出了序列对比损失,利用重叠子序列进行正负样本采样。最近的TempCLR研究[15]通过DTW技术在视频段落对齐中探索时间动态,并通过基于时间粒度的新型负采样策略更有效地捕捉时间动态。
TimeAutoML[16]通过AutoML框架实现自动配置和超参数优化,利用在实例范围内引入随机噪声来创建负样本。CARL[14]通过对齐增强视频视图之间的序列相似性,采用序列对比损失,保持帧间的时间连贯性和上下文一致性。这使得特征表示在视频长度和内容变化时仍具备鲁棒性,与本研究所关注的时间一致性问题高度相关。TS2Vec[17]利用随机重叠片段捕获多尺度上下文信息,采用时间和实例对比损失,在捕捉时间依赖性方面表现优异。ZHANG et al[18]提出的TS CoT算法通过联合训练增强不同视图表示的全局一致性,对本研究的多维时间依赖性建模具有启发意义。SHIN et al[20]提出的基于重叠窗口的一致性正则化框架,以软标签作为共享优化目标,能在一定程度上提升对时间一致性的捕捉能力。
YUE et al[17]提出了TS2Vec,该方法通过上下文一致性学习,在随机重叠的子序列中,采用两种对比策略来计算时间和实例对比损失。与其他方法不同,TS2Vec通过层次化方式在增强的上下文视图上开展对比学习,为每个时间戳提供鲁棒的上下文表示。该方法能通过在相应时间戳上的最大池化生成全局特征表示,并能够有效捕捉多尺度上下文信息。NGUYEN et al[25]设计了一种融合层次损失和三重损失的新型损失函数,用于增强时间序列表示的多粒度特性。CSL[26]则通过学习多粒度对比和多尺度对齐的时间序列表示来捕捉跨时间范围的依赖信息。COMET[27]结合医学时间序列的四级对比损失,通过对每个损失引入超参数系数,权衡多层次对比损失,从而获得更稳健和全面的特征表示。DUAN et al[28]利用分层结构捕捉不同频率的时间序列特征,并通过强制各组间的一致性约束,生成在不同频率下都有效且鲁棒表示。
FEAT[29]通过分层时间对比、特征对比和重建损失的联合学习,实现特征一致性与时间一致性的协同优化。CHOI和KANG[30]引入了不确定性加权方法,通过建模多个任务的同方差不确定性,权衡多个对比损失,增强时间和转换一致性的建模能力。FOCAL[31]强调模态一致性和变换一致性,学习跨模态共享特征;同时考虑局部性零星偏差,通过松散排名损失对样本对施加约束。ZHANG et al[32]提出时频一致性建模,认为同一时间序列中基于时间和频率的表示应该在时频潜在空间中更接近,从而引入一致性损失以最小化嵌入之间的距离。TimesURL[33]通过构造双优解并引入对比学习联合优化目标,捕捉分段级和实例级信息。LEE et al人[34]提出SoftCLT,通过软对比损失实现细致的学习,基于时间序列实例之间的距离进行实例对比,重点处理时间戳差异。LI et al[35]揭示了时间序列中表示偏差与光谱距离的正相关关系,并通过优化对比损失,减少数据增强中的偏差,提高了嵌入的稳健性和普适性。
UCR 时间序列归档(UCR Time Series Classification Archive)是时间序列分类领域的经典基准数据集,包含128个单变量时间序列数据集,涵盖医疗、工业、运动等多个领域。每个数据集包含训练集和测试集,时间序列长度固定。类别数从2到60不等,适用于二分类和多分类任务。代表性数据集包括ECG200(心电图分类)、GunPoint(手势识别)、Coffee(光谱分类)等。UEA 多变量时间序列分类归档(UEA Multivariate Time Series Classification Archive)是 UCR 的扩展,专注于多变量时间序列分类任务。包含30个多变量数据集,每个时间序列包含多个变量(通道)。应用场景包括动作识别(如BasicMotions)、医疗监测(如Libras)等。由于数据长度和变量数各异,该数据集具有更高的复杂性。
CAILLAULTÉ P, BIGANDA.Comparative study on univariate forecasting methods for meteorological time series[C]//2018 26th European Signal Processing Conference (EUSIPCO).Rome,Italy,IEEE,2018:2380-2384.
[2]
BASUS, MUKHERJEEA, KLIVANSKYS.Time series models for internet traffic[C]//Proceedings of IEEE INFOCOM’96.Conference on Computer Communications.San Franciso,CA,USA,IEEE,1996,2:611-620.
[3]
MAKRIDISG, KYRIAZISD, PLITSOSS.Predictive maintenance leveraging machine learning for time-series forecasting in the maritime industry[C]//2020 IEEE 23rd international conference on intelligent transportation systems (ITSC).Khodes,Greece,IEEE,2020:1-8.
[4]
SEZERO B, GUDELEKM U, OZBAYOGLUA M.Financial time series forecasting with deep learning:A systematic literature review:2005—2019[J].Applied soft computing,2020,90:106181.
[5]
SHUMWAYR H, STOFFERD S, SHUMWAYR H,et al.Time series regression and ARIMA models[J].Time Series Analysis and Its Applications,2000:89-212.
TONEKABONIS, EYTAND, GOLDENBERGA.Unsupervised representation learning for time series with temporal neighborhood coding[J].arXiv preprint arXiv:2021.
[8]
ELDELEE, RAGABM, CHENZ,et al.Time-series representation learning via temporal and contextual contrasting[J].arXiv preprint arXiv:2021.
[9]
HAJIMORADLOUA, PISHDADL, TUNGF,et al.Self-supervised time series representation learning with temporal-instance similarity distillation[C]//First Workshop on Pre-training:Perspectives,Pitfalls,and Paths Forward at ICML,2022.
[10]
YANGX, ZHANGZ, CUIR.Timeclr:A self-supervised contrastive learning framework for univariate time series representation[J].Knowledge-Based Systems,2022,245:108606.
[11]
WANGY, XUY, YANGJ,et al.Fully-connected spatial-temporal graph for multivariate time-series data[C]//Proceedings of the AAAI conference on artificial intelligence.Vancouver,British Columbic,Canada,2024,38(14):15715-15724.
[12]
MORGADOP, LIY, NVASCONCELOSN.Learning representations from audio-visual spatial alignment[J].Advances in Neural Information Processing Systems,2020,33:4733-4744.
[13]
HARESHS, KUMARS, COSKUNH,et al.Learning by aligning videos in time[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.IEEE/CVF,2021:5548-5558.
[14]
CHENM, WEIF, LIC,et al.Frame-wise action representations for long videos via sequence contrastive learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.IEEE/CVF,2022:13801-13810.
JIAOY, YANGK, DOUS,et al.Timeautoml:autonomous representation learning for multivariate irregularly sampled time series[J].arXiv preprint arXiv:2020.
[17]
YUEZ, WANGY, DUANJ,et al.Ts2vec:Towards universal representation of time series[C]//Proceedings of the AAAI conference on artificial intelligence.AAAI,2022,36(8):8980-8987.
[18]
ZHANGK, WENQ, ZHANGC,et al.Self-supervised learning for time series analysis:Taxonomy,progress,and prospects[J].IEEE transactions on pattern analysis and machine intelligence,2024,46(10):6775-6794.
[19]
XUM A, MORENOA, WEIH,et al.Retrieval-based reconstruction for time-series contrastive learning[J].Arxivabs231100519,2023.
[20]
SHINY, YOONS, SONGH,et al.Context consistency regularization for label sparsity in time series[C]//International Conference on Machine Learning.Honolulu,Havaii,USA,PMLR,2023:31579-31595.
[21]
ZHONGS, SONGS, ZHUOW,et al.A multi-scale decomposition mlp-mixer for time series analysis[J].arXiv preprint arXiv:2023.
[22]
CHOWDHURYR R, LIJ, ZHANGX,et al.Primenet:Pre-training for irregular multivariate time series[C]//Proceedings of the AAAI Conference on Artificial Intelligence.Washington D C,USA,2023,37(6):7184-7192.
[23]
CHENP, HUANGD, HED,et al.Rspnet:Relative speed perception for unsupervised video representation learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence.Attanta,Georgia,USA,2021,35(2):1045-1053.
[24]
JENNIS, JINH.Time-equivariant contrastive video representation learning[C]//Proceedings of the IEEE/CVF international conference on computer vision.Shen Zhen,IEEE,2021:9970-9980.
[25]
NGUYEN ADUY, TRANT H, PHAMH H,et al.Learning Robust and Consistent Time Series Representations:A Dilated Inception-Based Approach[J].arXiv e-prints arXiv:2023.
[26]
LIANGZ, ZHANGJ, LIANGC,et al.A shapelet-based framework for unsupervised multivariate time series representation learning[J].arXiv preprint arXiv:2023.
[27]
WANGY H, YUH, WANGH S,et al.2023.Contrast everything:A hierarchical contrastive framework for medical time-series.In NeurIPS[C]//Proceedings of the 37th Conference on Neural Information Proceeding Systems.New Orleans,LA:Curran Associates Inc,2023,36:35694-55717.
[28]
WANGY, HANY, WANGH,et al.Contrast everything:A hierarchical contrastive framework for medical time-series[J].Advances in Neural Information Processing Systems,2023,36:55694-55717.
[29]
KIMS, CHUNGE, KANGP.FEAT:A general framework for feature-aware multivariate time-series representation learning[J].Knowledge-Based Systems,2023,277:110790.
LIUS, KIMURAT, LIUD,et al.Focal:Contrastive learning for multimodal time-series sensing signals in factorized orthogonal latent space[J].Advances in Neural Information Processing Systems,2023,36:47309-47338.
[32]
ZHANGX, ZHAOZ, TSILIGKARIDIST,et al.Self-supervised contrastive pre-training for time series via time-frequency consistency[J].Advances in neural information processing systems,2022,35:3988-4003.
[33]
LIUJ, CHENS.Timesurl:Self-supervised contrastive learning for universal time series representation learning[C]//Proceedings of the AAAI conference on artificial intelligence.Vancauver,BC:AAAI Press,2024,38(12):13918-13926.
[34]
LEESEUNGHAN,PARKTAEYOUNG,LEEKIBOK.Soft Contrastive Learning for Time Series in ICLR[C]//Procedings of the 12th International Conference on Learning Representations.kigali:Openview,2024.
[35]
LEES, PARKT, LEEK.Soft contrastive learning for time series[J].arXiv preprint arXiv:2023.
[36]
DAUH A, BAGNALLA, KAMGARK,et al.The UCR time series archive[J].IEEE/CAA Journal of Automatica Sinica,2019,6(6):1293-1305.
[37]
BAGNALLA, DAUH A, LINESJ,et al.The UEA multivariate time series classification archive[J].arXiv preprint arXiv:2018.
[38]
ZHOUH, ZHANGS, PENGJ,et al.Informer:Beyond efficient transformer for long sequence time-series forecasting[C]//Proceedings of the AAAI conference on artificial intelligence.Virtual Event:AAAI Press,2021,35(12):11106-11115.
[39]
RENH, XUB, WANGY,et al.Time-series anomaly detection service at microsoft[C]//Proceedings of the 25th ACM SIGKDD International Conference On Knowledge Discovery & Data Mining.Anchorage,AK,USA:ACM,2019:3009-3017.
[40]
WUH, HUT, LIUY,et al.Timesnet:Temporal 2d-variation modeling for general time series analysis[J].arXiv preprint arXiv:2022.
[41]
CHOIH,SOM A, TURAGAP.AMC-Loss:Angular Margin Contrastive Loss for Improved Explainability in Image Classification[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).Seattle,WA,USA:IEEE,2020:3659-3666.
[42]
BOUHSINET, AAROUSSI IEL, FAYSALA,et al.SimO loss:anchor-free contrastive loss for fine-grained supervised contrastive learning[J].arXiv e-prints arXiv:2024.