A Three-Dimensional ResNet and Transformer-Based Approach to Anomaly Detection in Multivariate Temporal-Spatial Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Anomaly detection in multivariate time series is crucial across various domains.
• Existing methods often lack parallel processing capabilities and struggle to fuse temporal and spatial features effectively.

Purpose of the Study:

• To propose TDRT, a novel three-dimensional ResNet and transformer-based model for anomaly detection.
• To enhance the accuracy of anomaly detection by automatically learning multi-dimensional temporal-spatial features.

Main Methods:

• Developed TDRT, a hybrid model integrating ResNet and transformer architectures.
• Applied TDRT to multi-dimensional industrial control time-series data to capture temporal-spatial correlations and long-term dependencies.

Main Results:

• TDRT demonstrated superior performance in identifying anomalies within temporal-spatial datasets.
• Achieved an average anomaly detection F1 score exceeding 0.98 and a recall of 0.98 on benchmark datasets (SWaT, WADI, BATADAL).
• Significantly outperformed five state-of-the-art anomaly detection algorithms.

Conclusions:

• TDRT offers a highly parallel and accurate approach for anomaly detection in multivariate time series.
• The method effectively extracts temporal-spatial correlations, enabling efficient mining of long-term dependencies in complex data.