Anomalies Detection in Structures, System and Components for Supporting Nuclear Long Term Operation Program

Abstract

Long Term Operation (LTO) of nuclear power plants (NPPs) will play a key role to reach net zero target. Monitoring and predictive approach detecting in advance faulty SCCs conditions may provide a further key tool in LTO framework. Detecting anomalies may allow the transition from time-based to condition-based predictive maintenance of the NNPs. Predictive algorithms could reduce the number of unplanned outages caused by reactor system failures (one-day outage of a 1000-MW NPPs causes losses of about 500 k$), improving the capacity factor, and keeping high safety margin level of NPPs. To this end, innovative approach by unsupervised machine learning technique (ML) is proposed to detect anomalies of SSCs. Based on principal component analysis and mahalanobis distance is possible to detect in advance the failure of the components. To the purpose a 2D digital twin of primary nuclear pipe under nominal conditions (inner temperature of 300° and an internal pressure of 15.5 MPa) is implemented in finite element code to provide a dataset for unsupervised ML code. The algorithm is then tested under anomaly pattern that deviate from nominal conditions. The results show good code prediction capabilities anticipating the pipe failure. Traditional monitoring combined with ML technique may support LTO program increasing the safety and competitiveness of NPPs.