Deep learning application for monitoring degradation in nuclear safety systems

Abstract

The safe operation of nuclear power plants relies on maintaining the structural integrity of various systems and components, such as equipment-piping systems. Ageing and degradation from flow-accelerated erosion and corrosion can lead to cracks and leakages, posing risks like loss of coolant accidents (LOCA). To prevent such incidents, regular monitoring and maintenance are vital. Recent efforts to implement Artificial Intelligence (AI)-driven condition monitoring aim to enhance safety and efficiency. However, the effectiveness of simulation-based degradation detection models needs to be validated using experimental or real-time data from nuclear power plants. This research explores an AI-based monitoring framework’s validity for nuclear equipment-piping systems through experimentation. A piping system is designed and subjected to harmonic excitations representing typical pump-induced vibrations in nuclear plants. Degradation levels are classified based on wall thickness loss as minor, moderate or severe. Non-uniform degradation is implemented at structural discontinuities such as the elbows. Sensor response is collected from accelerometers installed on the experimental system and its corresponding digital twin. Deep neural networks such as multilayer perceptron and convolutional neural networks are developed to detect the degraded locations and their severity. The results from the experimental data, as well as the simulated data, are compared for accuracy.