Enhancement of anomaly detection system using artificial intelligence for sensor signals in nuclear power plants

Abstract

Nuclear facilities are considered one of the most critical infrastructures, which require more reliable and efficient operating systems, especially nuclear power plants. Multiple sensors are usually used in nuclear power plants to monitor the equipment's efficiency and the operational parameters of the systems. Sensors have an important role in monitoring and controlling physical parameters, like pressure, coolant flow, temperature, and radiation levels. These sensors are susceptible to malfunctions due to harsh working conditions or outside issues. Anomaly detection systems have a vital role in sensors' signal processing to avoid dangerous accidents during wrong readings and compromised safety systems. Traditional fault detection scenarios are always difficult to catch the complex, non-linear relationships and failures in sensor data. In this work, a hybrid deep learning and ensemble-based approach is proposed. It integrates Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRUs) to considerably improve anomaly detection and classification abilities in critical systems. They are combined through a stacking ensemble with machine learning models (XGBoost, CatBoost, and LightGBM). This combination is useful for features extra-y capturing temporal and spatial dependencies in sensor data by exploiting the strengths of both deep learning and ensemble strategies. The proposed approach is investigated in a recent dataset that was collected from sensors in a nuclear power plant with high precision and minimum false alarms. The obtained results of our hybrid approach show significant enhancement of evaluation metrics like accuracy, precision, recall, and F1score compared to standalone models. These evaluation metrics outperform the proposed architecture to accurately identify anomalies with approximately 100% accuracy for binary anomaly detection and 99% accuracy for multi-class fault classification. It introduces a suitable solution for robustness systems in harsh environments such as nuclear facilities to improve safety and operational efficiency.