Generalization analysis and improvement of CNN-based nuclear power plant fault diagnosis model under varying power levels

The domain discrepancy caused by power level gap between training set (source domain) and test set (target domain) limits the generalization of data-driven models in practical nuclear power plant fault diagnosis applications. In this study, a highly fine-grained quantitative generalization description of Convolutional Neural Network (CNN) model is conducted with high-dimensional and strong-nonlinear complex nuclear power plant simulation data. Results show that with source domain of single power level, CNN suffers from over-fitting and poor domain discrepancy generalization; with source domain of multiple power levels, the sensitivity of CNN to minor domain discrepancy is greatly reduced and its generalization is significantly boosted. Besides, a novel Artificial Disturbance Method (ADM) based domain discrepancy generalization promotion framework is proposed in this study, which alleviates the over-fitting of CNN by adding disturbed training data to its training set. The feasibility and superiority of the framework is proven when Gaussian noise disturbance or uniformly distributed noise disturbance is adopted to generate disturbed training data. The ADM-based framework reduces the requirement for the number of power levels contained in source domain, this advantage endows it with strong practicability in actual nuclear power plant fault diagnosis tasks where the available source domain data are extremely limited.