self-supervised anomaly detection

Traditional unsupervised methods for time-series anomaly detection suffer from poor generalizability and struggle to adapt to dynamically evolving normal patterns. Method: This paper systematically surveys recent advances in self-supervised learning (SSL) for time-series anomaly detection, introducing the first comprehensive taxonomy tailored to this domain. It unifies modeling paradigms—including masked reconstruction, contrastive learning, temporal discrimination, and prediction consistency—and task construction strategies, while integrating mainstream architectures such as temporal convolutional networks, Transformers, and graph neural networks. Contribution/Results: We propose a structured knowledge graph and open-source a continuously maintained GitHub repository (Awesome-Self-Supervised-Time-Series-Anomaly-Detection). The work explicitly identifies key open challenges and future research directions, establishing the first authoritative classification framework and benchmark reference for SSL-based time-series anomaly detection.

This study addresses the challenges in time series anomaly detection posed by extreme class imbalance and scarce labeled data, which hinder supervised approaches and lead to high false positive rates in unsupervised methods. To overcome these limitations, the authors propose an unsupervised detection framework that integrates Haar discrete wavelet transform with a tailored t-test. By decomposing the signal across multiple scales and applying statistically grounded significance testing, the method effectively identifies anomalies without requiring labeled data. This work is the first to synergistically combine Haar wavelets with theoretically justified t-tests, substantially reducing false positives while enhancing detection accuracy. Extensive experiments on 343 real-world datasets demonstrate that the proposed approach outperforms current state-of-the-art unsupervised and self-supervised methods in both detection speed and accuracy.

This study addresses the persistent challenge of simultaneously achieving accuracy, efficiency, and interpretability in time series anomaly detection. We propose an autonomous program search framework driven by large language models (LLMs). Departing from paradigms that employ LLMs directly as detectors, this work pioneers an approach wherein the LLM autonomously generates NumPy code to construct compact and transparent detectors. By integrating local spectral feature analysis with covariance-aware distance algorithms, the proposed method operates entirely without neural network training. Evaluated on the TSB-AD benchmark, our approach comprehensively outperforms existing baselines, delivering high accuracy, computational efficiency, and GPU-free deployment. This research establishes a novel paradigm for interpretable anomaly detection.

This work addresses the challenge that existing unsupervised methods struggle to reliably detect subtle and noisy anomalies in complex time series, often being misled by noise in normal samples and missing near-normal anomalies. To overcome this limitation, we propose a novel unsupervised anomaly detection framework that integrates active learning: it enhances temporal dependency modeling through a masked time series reconstruction feedback mechanism and employs a minimax optimization strategy to differentially treat normal and anomalous samples, thereby improving robustness against noise and weak anomalies. Extensive experiments across four multivariate time series datasets and seven backbone models demonstrate that our method achieves an average AUC improvement of 12.39%, significantly outperforming current unsupervised approaches.

Existing one-class anomaly detection methods suffer from poor generalization due to geometric bias in benchmark datasets, while prevailing pseudo-anomaly generation techniques fail to faithfully model the intrinsic structure of normal data and over-rely on image-domain operations. Method: This paper proposes a novel paradigm that synthesizes pseudo-anomalies exclusively within a frozen pre-trained feature space—bypassing image-level augmentation entirely. Contribution/Results: Key innovations include (1) an adaptive linear feature perturbation mechanism that dynamically tailors noise distribution per sample, and (2) a contrastive learning objective explicitly decoupling geometric bias from semantic anomaly modeling. Evaluated on both standard and geometric-bias-mitigated benchmarks, our method consistently outperforms state-of-the-art approaches, demonstrating superior generalization and robustness. The implementation is publicly available.

研究通过几何特性预测异常检测性能，提出伪异常探针以改善无异常数据时的模型选择。

This work addresses the challenge of unsupervised anomaly detection in multivariate time series, where modeling normal patterns remains difficult due to the absence of a unified and robust representation of normality. To this end, the paper proposes the U²AD framework, which uniquely integrates time-varying score networks with a unified training objective. By leveraging score-based generative modeling, U²AD learns the manifold structure of normal data across both local and global temporal contexts and enables deterministic reconstruction through an ordinary differential equation solver. Without requiring labeled data, the method precisely characterizes the distribution of normality and facilitates early anomaly identification. Extensive experiments demonstrate that U²AD significantly outperforms state-of-the-art approaches across multiple benchmarks, achieving notable improvements in both detection accuracy and early warning capability.

Detecting anomalies in images and video is an essential task for multiple real-world problems, including industrial inspection, computer-assisted diagnosis, and environmental monitoring. Anomaly detection is typically formulated as a one-class classification problem, where the training data consists solely of nominal values, leaving methods built on this assumption susceptible to training label noise. We present a dataset folding method that transforms an arbitrary one-class classifier-based anomaly detector into a fully unsupervised method. This is achieved by making a set of key weak assumptions: that anomalies are uncommon in the training dataset and generally heterogeneous. These assumptions enable us to utilize multiple independently trained instances of a one-class classifier to filter the training dataset for anomalies. This transformation requires no modifications to the underlying anomaly detector; the only changes are algorithmically selected data subsets used for training. We demonstrate that our method can transform a wide variety of one-class classifier anomaly detectors for both images and videos into unsupervised ones. Our method creates the first unsupervised logical anomaly detectors by transforming existing methods. We also demonstrate that our method achieves state-of-the-art performance for unsupervised anomaly detection on the MVTec AD, ViSA, and MVTec Loco AD datasets. As improvements to one-class classifiers are made, our method directly transfers those improvements to the unsupervised domain, linking the domains.

This work addresses the challenge of effectively detecting unknown anomalies in autonomous driving systems under the absence of anomaly labels. To this end, we propose a self-supervised online anomaly detection framework based on the Joint Embedding Predictive Architecture (JEPA), which, to the best of our knowledge, is the first to apply JEPA to temporal object state data in autonomous driving. By learning latent representations from unlabeled data, the method enables effective identification of previously unseen anomalies. Integrating classical anomaly detection techniques, our approach demonstrates high detection performance for object state anomalies in unsupervised settings, as validated on the real-world nuScenes dataset. This significantly enhances the online robustness of autonomous driving systems against unexpected and anomalous scenarios.

本文提出了一种非参数框架，使用伪隔离异常分数来定义和检测异常值，并将其应用于K-means聚类中以识别特定于簇的异常值。