Abstract: With the widespread application of Internet of Things (IoT) technology, precise anomaly detection in the massive time series data collected by IoT systems holds significant academic and practical value. However, existing methods generally face many challenges, such as the difficulty in accurately modeling the complex spatio-temporal dependencies within time series and poor noise resistance. To address these limitation, we propose MaStGT-UAD(Memory-augmented Spatio-temporal Graph Transformer based Unsupervised Anomaly Detection ), an unsupervised anomaly detection method. MaStGT-UAD utilizes a memory-augmented transformer and a dynamic graph structure learning method to separately extract spatio-temporal features from the input time series, where the memory module is designed to constrain the transformer’s ability to reconstruct anomalies. Then, the extracted spatiotemporal features are fused using Graph NeuralNetwork (GNN). To address the prevalent issue of concurrent noise in input data, this study adopts a novel window-based graph structure learning method, ensuring that MaStGT-UAD can accurately distinguish between anomalies and noise even in complex environments.. MaStGT-UAD achieves an average F1 score of 93.26% for anomaly detection, surpassing other latest deep learning methods by 17.49%.

Key words: Multivariate Time Series, Unsupervised Anomaly Detection, Memory module, Dynamic Graph Learning, Concurrent Noise

摘要： 随着物联网技术的广泛应用，对其采集的海量物联网时间序列数据进行精准异常检测具有非常重要的学术与应用价值。然而，现有方法普遍面临着诸多挑战，例如：难以精确建模海量、非平稳、高维、高噪物联网多元时间序列的复杂时空依赖关系、抗噪能力不足等。针对这些局限性，本研究提出了名为记忆增强型时空图 Transformer 的无监督异常检测方法(MaStGT-UAD)。首先，MaStGT-UAD 通过记忆增强型 Transformer 和动态图结构学习，有效提取时间序列数据中的时空特征，其中记忆模块旨在改进 Transformer 对异常的重构能力；然后，利用图神经网络(GNN)对提取出的时空特征进行融合；针对输入数据中普遍存在的并发噪声问题，本研究采用了一种新颖的基于窗口的图结构学习方法，以确保 MaStGT-UAD 在复杂环境下仍能有效区分异常和噪声。通过在三个物联网多元时间序列公开数据集上的对比实验验证了 MaStGT-UAD 的有效性与先进性，其异常检测平均 F1 分数达到了 93.26%，相较于其他最新地深度学习方法，提升了 17.49 个百分点。

关键词: 多元时间序列, 无监督异常检测, 记忆模块, 动态图学习, 并发噪声