Adaptive Data Mining-Based Deep Detection for Streaming Time-Series Anomaly Detection

Abstract

Addressing the challenges of anomaly sparsity, insufficient labeling, time-varying data distribution, and diverse anomaly patterns in streaming time-series scenarios, this paper proposes an adaptive data mining and energy model-based deep detection framework for anomaly detection. This method maps the input sequence to a latent representation space using a deep encoder. Self-supervised contrastive Learning enhances the robustness of the representation to noise and perturbations, and an online memory mechanism maintains a set of recent low-energy samples to characterize the current normal structure. Based on this, an energy scoring function based on normal prototypes is constructed, transforming the consistency between samples and dynamic normal centers into a continuously comparable anomaly metric. Simultaneously, a smoothly updated quantile threshold achieves stable online discrimination. This mechanism unifies structure maintenance, scoring scale, and alarm decision-making within a single process, enabling the model to maintain adaptability to changes in normal patterns and sensitivity to anomaly deviations in a continuously arriving data stream. Comparative experimental results demonstrate that the proposed method achieves superior overall performance across multiple evaluation metrics, validating the effectiveness and practical value of combining adaptive structure maintenance and energy measurement in time-series anomaly detection tasks.

Cover Image
Downloads
Published
2025-06-01
Section
Articles