Anomaly Detection for Asynchronous Multivariate Time Series of Nuclear Power Plants Using a Temporal-Spatial...

Area of Science:

• Industrial process monitoring
• Data science
• Nuclear engineering

Background:

• Multivariate time-series (MTS) anomaly detection is vital for industrial process monitoring, particularly in safety-critical nuclear power plants (NPPs).
• Existing data-driven methods often struggle to fully leverage temporal-spatial correlations in operational MTS data.
• Asynchronous time-lagged correlations in NPP data present a significant challenge for current anomaly detection techniques.

Purpose of the Study:

• To propose a novel reconstruction-based MTS anomaly detection approach using a temporal-spatial transformer.
• To effectively learn dependencies and extract temporal-spatial correlations from asynchronous MTS data at various scales.
• To enhance the accuracy and timeliness of anomaly detection in industrial processes, especially NPPs.

Main Methods:

• Development of a temporal-spatial transformer model for MTS anomaly detection.
• Implementation of a two-stage temporal-spatial attention mechanism.
• Integration of a multi-scale strategy to capture dependencies at different data scales.

Main Results:

• The proposed model demonstrates superior feature learning capabilities for asynchronous MTS data.
• Experiments show improved performance in signal reconstruction and anomaly detection accuracy.
• The model enables earlier detection of anomalous events compared to existing methods.

Conclusions:

• The temporal-spatial transformer effectively addresses the challenge of asynchronous time-lagged correlations in MTS data.
• The proposed approach enhances operational safety and reduces potential losses in NPPs through early anomaly detection.
• This method offers a significant advancement in data-driven industrial process monitoring.