APE: A Post-Training Enhancement Framework for Time Series Anomaly Detection based on Agentic Pseudo-Anomaly Generation

Shuqi Zhao, Zijie Chen

Keywords: Deep learning, Time series anomaly detection, Agent

Abstract: Time Series Anomaly Detection (TSAD) is a critical task across various domains, yet the scarcity of anomaly labels in real-world scenarios makes unsupervised reconstruction models the dominant paradigm. However, the normal-only reconstruction paradigm is prone to over-generalization, where high-capacity models reconstruct both normal and anomalous patterns well, leading to poor separability in the representation space. While Pseudo-Anomaly Generation (PAG) has emerged as a promising solution, it suffers from distribution misalignment and backbone coupling. In this proposal, we introduce APE, an Agentic-PAG based Post-Training Enhancement framework. APE leverages an autonomous agent to mine multimodal scenario priors and generate high-fidelity pseudo-anomalies. Combined with a three-branch contrastive learning objective, APE operates as a model-agnostic plug-in to significantly enhance normal-anomaly separability in the representation space for existing TSAD backbones.

Submission Number: 13