Abstract: Time series anomaly detection aims to identify rare patterns or deviations that significantly differ from normal behavior. Existing research focuses on designing more advanced network architectures or optimizing objectives to accurately capture the normal patterns in the data. However, training data contaminated with anomalies and the lack of anomalous information can lead to biases in the normal decision boundary learned by traditional methods, resulting in suboptimal detection performance. This paper proposes an unsupervised anomaly detection method based on restricted distribution mapping and pseudo anomaly calibration. Specifically, the input data features are first extracted using a temporal convolutional network. Then, to learn a more accurate normal decision boundary, this method optimizes the normal representation learning process from two aspects. Firstly, a restricted distribution mapping method is proposed to map the unknown distribution of the training data to a Gaussian distribution through feature normalization, and to minimize the one-class distance between the normal data and the center of the hypersphere, to make the data more compact in the feature space. Secondly, a pseudo anomaly calibration mechanism is proposed to generate multiple types of pseudo-anomaly data through data perturbation operations applied to the original data. This mechanism incorporates anomaly knowledge into the model by distinguishing between normal and pseudo-anomaly data, thereby refining the decision boundary. Meanwhile, the class separability between normal and abnormal data is expanded by maximizing the one-class distance between the pseudo-anomalous data and the center of the hypersphere. In the testing phase, anomalies are detected by measuring the one-class distance between the input data and the hyper-sphere center. Extensive experiments conducted on five real-world datasets demonstrate the superior performance of this method compared to current mainstream methods, as well as its higher robustness under different levels of training data contamination with anomalies.

Key words: time series, anomaly detection, restricted distribution mapping, pseudo anomaly calibration

摘要： 时间序列异常检测旨在识别与正常行为显著不同的稀有模式或偏差。现有的研究侧重于设计更先进的网络架构或优化目标来尽可能准确地捕获数据的正常模式。然而，异常污染下的训练数据和异常信息的缺失会致使传统方法学习到的正态决策边界产生偏差，从而导致次优级的检测性能。提出一种受限分布映射和伪异常校准引导下的无监督异常检测方法。具体地，利用时间卷积网络提取输入数据特征。为了学习更精确的正态决策边界，从两个方面优化正态表征学习过程：第一，提出一种受限分布映射方法，通过特征归一化将训练数据的未知分布映射至高斯分布，并最小化正常数据与超球中心的单类距离，使数据在特征空间中更加紧凑；第二，提出一种伪异常校准机制，利用数据扰动操作从原始数据生成多种类型的伪异常数据，并通过区分正常数据和伪异常数据向模型注入异常知识，矫正决策边界。同时，通过最大化伪异常数据与超球中心的单类距离，扩大正常与异常数据之间的类可分离性。在测试阶段，通过测量输入数据与超球中心的单类距离来检测异常。在5个真实数据集上进行的大量实验表明，该方法实现了相比目前主流方法的更优性能，并且在不同程度的异常污染训练数据下具备更高的鲁棒性。

关键词: 时间序列, 异常检测, 受限分布映射, 伪异常校准