Uncertainty-aware nuclear power turbine vibration fault diagnosis method integrating machine learning and heuristic algorithm

Abstract

Nuclear power turbine fault diagnosis is an important issue in the field of nuclear power safety. The numerous state parameters in the operation and maintenance of nuclear power turbines are collected, forming a complex high-dimensional feature space. These high-dimensional feature spaces contain redundant information, which increases the training cost and reduces the recognition accuracy and efficiency of the fault diagnosis model. To address the aforementioned challenges, a vibration fault diagnosis algorithm in nuclear power turbines is proposed. First, a long short-term memory-based denoising autoencoder (LDAE) is designed to enhance the capability of uncertainty awareness. Then, a feature extraction method integrating variational mode decomposition (VMD), L-cliffs-based effective mode selection, and sample entropy is devised to extract the latent features from the complex high-dimensional feature space. Furthermore, using extreme gradient boosting (XGBoost) as the classifier, LDAE-VMD-XGBoost model is constructed for fault diagnosis of nuclear power turbines. Considering the impact of multiple hyperparameters of LDAE-VMD-XGBoost model on the performance, the pathfinder algorithm is used to optimise the model hyperparameter settings and improve the fault diagnosis accuracy. Experimental results demonstrate the performance of the proposed improved LDAE-VMD-XGBoost in accurate nuclear power turbine vibration fault diagnosis.