Anomaly Detection in Nuclear Power Production Based on Neural Normal Stochastic Process

Area of Science:

• Nuclear Engineering
• Data Science
• Artificial Intelligence

Background:

• Nuclear power plants utilize extensive sensor networks for safety monitoring and anomaly detection.
• Sensor data integrity is crucial for effective anomaly detection but is often compromised by the harsh operating environment.
• Incomplete sensor data poses a significant challenge for traditional anomaly detection methods.

Purpose of the Study:

• To develop a novel anomaly detection method for nuclear power data that can handle missing sensor readings.
• To introduce the Neural Normal Stochastic Process (NNSP) model, designed to process incomplete time-series data directly.
• To evaluate the performance of NNSP against established anomaly detection techniques on real-world nuclear power data.

Main Methods:

• Proposed the Neural Normal Stochastic Process (NNSP), a deep learning approach utilizing a sequentialization structure.
• Encoded incomplete sensor data into continuous latent representations within a neural network.
• Trained the model to identify anomaly patterns by specifying supervisory signals at missing or future time points.

Main Results:

• NNSP demonstrated superior performance in anomaly detection tasks involving incomplete time-series data.
• On the Power Generation System (PGS) dataset with 15% missing data, NNSP achieved an F1 score of 83.72%.
• The model outperformed five mainstream baseline methods, including ARMA, Isolation Forest, LSTM-AD, VAE, and NeutraL AD.

Conclusions:

• The NNSP method effectively addresses the challenge of missing sensor data in nuclear power anomaly detection.
• NNSP offers a robust alternative to data imputation, directly learning from incomplete monitoring data.
• The model shows significant potential for enhancing the safety and reliability of nuclear power operations.