A noise robust and distribution-adaptive framework for multivariate time series anomaly detection.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Unsupervised anomaly detection for multivariate time series (MTS) faces challenges with noisy data and non-stationarity.
• Existing methods often assume noise-free data and struggle with distribution shifts, limiting real-world applicability.

Purpose of the Study:

• To propose NORDA, a novel framework for robust unsupervised anomaly detection in MTS.
• To address limitations of existing methods regarding data noise and non-stationarity.

Main Methods:

• NORDA integrates a multi-order difference mechanism to mitigate noise during representation learning.
• A mixed reversible normalization module dynamically models non-stationary variations and adapts to distribution shifts.
• A Transformer-based encoder extracts robust latent representations by modeling inter-channel dependencies.

Main Results:

• NORDA significantly outperforms sixteen baseline methods across seven benchmark datasets.
• The framework demonstrates high robustness against noise contamination in MTS data.
• NORDA enhances dynamic adaptability to distribution shifts through its reversible architecture.

Conclusions:

• NORDA offers a robust and effective solution for unsupervised anomaly detection in noisy and non-stationary MTS.
• The proposed multi-order difference and reversible normalization mechanisms are key to its superior performance.
• This framework advances the state-of-the-art in anomaly detection for real-world time series data.