Deep anomaly detection for time series: A survey

The cyberspace environment has evolved into a complex ecosystem, generating vast amounts of diverse time series data from various devices, systems, and software. Detecting anomalies in these massive, multi-source datasets is critical for ensuring system reliability and security. This paper provides a comprehensive review of deep learning approaches for time series anomaly detection. We systematically classify existing methods into six categories based on their objective functions: forecasting models, reconstruction models, generative models, density models, contrastive models, and hybrid models. For each category, we analyze their advantages, disadvantages, and architectural variations to guide researchers in selecting appropriate approaches for specific problems. We further summarize applications across multiple domains including network services, cyber–physical systems, smart grids, smart cities, and healthcare, providing valuable insights into practical implementations. The paper also organizes commonly used public datasets with their key characteristics and examines evaluation metrics ranging from traditional point-level assessments to advanced sequence-adaptive frameworks. Finally, we discuss emerging challenges and promising research directions, including data augmentation strategies, model robustness improvements, generalization capabilities, applications of foundation models and large language models, autoML frameworks, and lightweight model designs. This survey offers a systematic framework for understanding the current landscape of deep time series anomaly detection and provides clear pathways for advancing the field to address real-world challenges.