Title: ChronosAD: Enhancing Anomaly Detection Precision via Time Series Foundation Models

Abstract:

Accurate anomaly detection is essential across numerous sectors, such as healthcare, finance, and industrial operations. Nevertheless, current techniques frequently face challenges in generalizing effectively across diverse datasets, particularly when dealing with subtle or context-specific irregularities. To address these limitations, we present ChronosAD, a new architecture that utilizes a time series foundation model as a feature extraction engine. The proposed method operates through a two-phase process: initially, it leverages the foundation model to generate zero-shot embeddings for each time series. Subsequently, a specially designed Temporal Block—integrating Bidirectional Long Short-Term Memory (BiLSTM) units with Multi-Head Attention mechanisms—refines these embeddings to better capture temporal relationships and emphasize significant patterns. In contrast to prior methods, ChronosAD demands very little task-specific adjustment while maintaining strong generalization capabilities across various fields, including automotive, cyber-physical, medical, and industrial systems. Our extensive evaluation across 11 benchmark datasets reveals that ChronosAD surpasses current state-of-the-art methods by an average of 4.72% in AUC and 6.60% in AP. The implementation code is publicly accessible at https://github.com/intelligolabs/ChronosAD.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC