Adaptive residual CNN-based fault detection and diagnosis system of small modular reactors

With the development of Industry 4.0 technology, it is a popular trend to reduce maintenance costs and ensure the safety of novel nuclear systems combined with deep learning (DL) technology. In this paper, an intelligent fault detection and diagnosis system (IFDDS) based on designed adaptive residual convolutional neural networks (ARCNNs) for small modular reactors (SMRs) is proposed. The features under different noise levels are learned as the residual and passed through the designed networks. Additionally, the learning efficiency is enhanced by the soft threshold (ST) method assembled in the adaptive residual processing (ARP) module. The Bayesian optimization (BO) method is adopted to improve the learning decay rate (LDR) of designed networks for better diagnosis performance. A total of 1,760 experimental data points under 11 different operation scenarios at three different noise levels are collected from the established Chinese lead-based nuclear reactor (CLEAR) platform to verify the effectiveness of the proposed IFDDS. The comparisons with the traditional RCNNs and CNNs adopted in previous works highlight the proposed diagnosis method's superiority. The performance of IFDDS is further improved by using the BO method. The proposed method, as a maiden attempt of intelligence research for SMRs, will provide remote decision-making support for nuclear operators in unattended conditions. Moreover, the universal method can also be applied to other diagnosis systems under a noise environment.