Data augmentation-based fault diagnosis framework for nuclear power plants with robust anti-data interference capability
Jiangkuan Li, Shu-xin Zheng, Jiaoshen Xu, Meng Lin, Zhenfeng Niu, Jiahao Cheng, Chenxin Zhu, Ruifeng Tian, Sichao Tan
Abstract
The robustness of deep learning-based fault diagnosis methods for nuclear power plants may be deteriorated by inevitable data interference, including random noise, measurement drift, and partial parameters missing. To address this issue, a data augmentation-based robustness enhancement framework is proposed for Convolutional Neural Network (CNN), which introduces three augmentation strategies: Train With Noisy Data (TWND), Train With Missed Data (TWMD), and Train With Noisy and Missed Data (TWNMD), each designed to counteract specific data taint scenarios. Experimental validation, performed using high-dimensional operating data from a simulated million-kilowatt pressurized water reactor, demonstrates the proposed methods significantly improve CNN’s robustness under various taint modes. Notably, CNN trained with TWNMD achieves an average accuracy improvement of 29.77% in mixed taint conditions, alongside an 8.18% reduction in standard deviation. These findings highlight the feasibility of data augmentation for enhancing the applicability and reliability of deep learning models in nuclear power plant fault diagnosis tasks.