A reliable data-driven method for condition monitoring in nuclear power plants

Condition monitoring, which is the basis of condition-based maintenance (CBM) and fault tolerant control (FTC) in nuclear power plants (NPPs), can accurately estimate the sensor measurement of the complex system, and then determine whether the system or equipment is abnormal. Data-driven method, such as Auto-Associative kernel regression (AAKR) and Auto-Associative neural network (AANN), has many advantages compare to model-based method and there are many application scenarios. But due to the uncertainty of measurement data, the incompleteness of training data and the uncertainty of model hyper-parameters, data-driven model has inherent uncertainty. That is, the predictions and residuals generated by data-driven model is not the true values. When the condition monitoring system determines that there is an abnormality in NPPs, we hope that the anomaly detection result has high reliability. Otherwise, the uncertainty of data-driven model will have a significant impact on decision-making, and even on the safety and economics of NPPs. This paper proposes a novel condition monitoring technology named Auto-Associative Kernel Ridge Regression (AAKRR) and then analyse model uncertainty using Monte Carlo method. The test results in dataset from NPPs show that AAKRR model has a small uncertainty interval and has strong anomaly detection capabilities.