Ensemble learning with diversified base models for fault diagnosis in nuclear power plants

Abstract

The performance and improvement of most currently proposed fault diagnosis models for nuclear power plants are limited due to their dependence on a single classification method. To solve this problem, this study proposes novel ensemble learning models (ELMs) obtained by integrating diversified base models through the plurality voting method and weighted voting method, respectively. Simulation data including one steady-state condition and 12 fault conditions are used to evaluate ELMs. The results show that ELMs achieve the best overall diagnostic performance and exhibit excellent performance for all faults. Compared with other common ensemble classification models, ELMs get better diagnostic performance at a dramatically lower time cost, and show significantly stronger robustness when diagnosing data with random noise. The universality of ELMs is illustrated by analyzing their performance under different input features. In summary, ensemble learning is a promising solution for fault diagnosis of nuclear power plants to improve safety and economy.