Session: 15-06

Paper Number: 134685

134685 - Neural Network-Driven Methodology for Predictive Health Monitoring and Aging Management in Nuclear Power Plant Operations 

Abstract: 

In the nuclear power plant field, ensuring the safety and longevity of critical components like Class I/II components is paramount. Long-term operation (LTO) and aging management of nuclear power plants (NPPs) are critical for ensuring safe, reliable, and sustainable energy production. As NPPs age, components and systems may degrade, requiring comprehensive aging management strategies to maintain safety and performance. LTO involves systematic evaluation and maintenance to extend a plant’s operational life beyond its initial design period. This includes regular safety upgrades, predictive maintenance, and rigorous monitoring to mitigate the effects of aging, thereby ensuring the continued safe operation of NPPs in their extended service life.

The proposed study introduces a novel approach leveraging neural networks, for feature extraction in predicting the health status of Class I pipe. The methodology uses a digital twin of the pipe, which provides a synthetic dataset crucial for the training of the neural network. This digital twin mimics the real-world conditions and behaviors of the  pipe, generating sensor readings that serve as inputs for the neural network. By utilizing this synthetic dataset, the model can be trained and validated under various simulated conditions, ensuring robustness and reliability in its predictions.

The digital twin, represented as a Finite Element (FE) model of a Class I pipe, simulates progressive degradation caused by thinning, taking into account the aging effects on the components. This model effectively mirrors the gradual wear and tear experienced by the pipe over time, providing a realistic representation of its aging process.

The architecture of the neural network efficiently condenses the input data (sensors readings) into a single, comprehensive signal. This output signal can be considered as health status of the pipe, offering a precise yet informative summary of its condition. The ability of the ANN to distill complex sensor data into a single health indicator is crucial for effective monitoring and safety of nuclear components.

The output of ANN which provides valuable insights into the temporal aspects of the pipe's condition. This time series is then used to forecast the Residual Useful Life (RUL) of the component, employing advanced forecasting techniques such as Long Short-Term Memory (LSTM) networks and Prophet. These methods are adept at handling time series data, allowing for accurate predictions of the pipe's future health status.

The forecasting of the RUL is a critical aspect in the framework of long-term operation of nuclear power plants. This proactive approach to maintenance not only enhances the safety of the nuclear plant but also leads to significant cost savings.

This methodology represents a significant step forward in the field of predictive maintenance for nuclear power plants. The combination of a digital twin, neural network-based feature extraction, and advanced forecasting techniques provides a comprehensive and reliable tool for assessing the health of nuclear components.

Presenting Author: Salvatore Angelo Cancemi University of Pisa

Presenting Author Biography: Hi I'm Salvatore Cancemi, determined, adaptive and fast learning person with a broad and acute interest on Industrial engineering in order to develop new skills and solve new challenges. I'm working on R&D international projects at University of Pisa in Mechanical and Structural Engineering Division focusing on piping/pressure Vessel thermomechanical phenomena, safety aspects and predictive maintenance based on artificial intelligence (AI) algorithm

Authors:

Salvatore Angelo Cancemi University of Pisa
Michela Angelucci University of Pisa
Rosa Lo Frano University of Pisa
Sandro Paci University of Pisa