Anomaly detection using spatial and temporal information in multivariate time series.

Area of Science:

• Industrial IoT
• Machine Learning
• Time Series Analysis

Background:

• Industrial systems generate vast multivariate time series data.
• Anomaly detection is crucial for system safety and reliability.
• Limited labeled data hinders supervised learning for anomaly detection.

Purpose of the Study:

• To propose a novel anomaly detection network, STADN, that leverages both spatial and temporal dependencies.
• To address the challenge of limited labeled data in industrial anomaly detection.
• To improve the accuracy and robustness of anomaly detection in complex systems.

Main Methods:

• STADN utilizes a graph attention network to model spatial dependencies between variables.
• A long short-term memory network is employed to capture temporal dependencies.
• The model predicts future sensor behavior and detects anomalies based on prediction errors, with error reconstruction for improved discrimination.

Main Results:

• STADN effectively integrates spatial and temporal information for anomaly detection.
• The model demonstrates superior performance compared to existing methods on real-world datasets.
• Reconstruction of prediction errors enhances the model's ability to distinguish anomalies from normal behavior.

Conclusions:

• STADN offers a powerful solution for anomaly detection in industrial systems with multivariate time series data.
• The proposed approach effectively handles the challenge of unlabeled data by focusing on prediction errors.
• Experimental validation confirms STADN's state-of-the-art performance in identifying system anomalies.