An Improved Method for PWR Fuel Failure Detection Using Cascade-forward Neural Network With Decision Tree

Abstract

When a fuel rod is damaged, determining the degree of fuel failure makes sense. The operators can decide whether to continue operating the reactor or shut it down based on the severity of the fuel failure. The isotopic ratio of two radioactive fission products (FPs) is a typical technique for evaluating the degree of fuel failure, although this is not applicable in the case of little fuel failure but large tramp uranium mass. The feedforward neural network (FFNN) has been used to identify fuel failures in order to overcome the shortcomings of the isotopic ratio method, although there is still inadequacy in the ability to distinguish between an intact fuel rod and a defective fuel rod with a small defect. In this study, we propose a cascade-forward neural network with a decision tree for fuel failure detection that performs well at classifying the degree of fuel failure and, in particular, at differentiating between an intact fuel rod and a defective fuel rod with a small size defect. The input of the neural network is the specific activity of FPs measured in the coolant. The degree of fuel failure is determined by the neural network’s output, which is labeled using one-hot encoding. The training set is constructed using the Booth-type diffusion model and the first-order kinetic model. The performance of the improved neural network is demonstrated. It is shown that the improved method is more accurate and responsive than the previous neural network when recognizing the onset of fuel failure. Finally, the most important nuclides are determined through the sensitivity analysis, and the neural network is simplified according to the importance of nuclides and the limitation of the radioactive detector in practical application.