Total Unwrapped Phase-Based Diagnosis of Wall Thinning in Nuclear Power Plants Secondary Piping Structures

Inspection-based techniques are being widely used for aging management of passive structures in nuclear power plants (NPPs). This paper presents a condition-based maintenance approach to detect corrosion in NPPs secondary piping structures using distributional features extracted from the total unwrapped phase signal of accelerometers. To demonstrate a condition-based maintenance approach, a scaled-down experimental testbed and data-driven methodology are developed and validated for detecting surrogate corrosion processes in a piping structure. The surrogate corrosion process is emulated by removing different levels of mass from the pipe-bend section and measuring the vibration of the pipe bend using five tri-axial accelerometers. The data collected are processed using the Hilbert-Huang transform. Distributional features of the total unwrapped phase from each accelerometer are used to develop a binary classifier to predict a degraded pipe against baseline pipes (healthy state). The binary classification model is extended to a multiclass classification to predict the level of mass removal. Binary and multiclass classification are performed using traditional machine learning algorithms such as logistic regression, neural networks, random forest, and support vector machine. All the algorithms achieved at least 99% prediction accuracy in each direction of the tri-axial accelerometer and significantly outperformed intrinsic mode function-based prediction methods. Performance of classification algorithms were evaluated using data for different number of the tri-axial accelerometers, including a single sensor. The overall prediction accuracy for all these cases was over 99%. Also, the models were interpreted using Shapley additive explanation values to understand the contribution of each input feature in diagnosing the percentage of mass removal. Finally, the performance of a transfer learning model is demonstrated to adopt a machine learning model on to an another elbow, and the prediction performance over 99% is achieved with partial retraining.