Fault early warning of equipment in nuclear power plant can effectively reduce unplanned forced shutdown and avoid significant safety accidents. This paper presents a Bayesian Long Short-Term Memory (LSTM) neural network method for fault early warning method of nuclear power turbine. The Long Short-Term Memory neural network prediction model is developed to address data uncertainty while taking into account complicated situation of the equipment operation. Quantitative reliability validation method is established based on Bayesian inference. A wavelet packet multi-scale time-frequency analysis is employed for data denoising. A Probabilistic Principal Component Analysis (PPCA) method combined with key factor analysis is proposed for dimension reduction and dealing with the data uncertainty. The principal component inverse search method is developed to identify the critical factors mainly contributing to the turbine fault. Numerical results indicate that the proposed novel model is validated with Bayesian confidence of 92% by using the real-world steam turbine data and the model can provide accurate warning in the early creep stage of the fault. INDEX TERMS Bayesian inference, long short-term memory, discrete wavelet packet transform, nuclear power turbine, probabilistic principal component analysis.
A Bayesian framework-based approach is proposed for the quantitative validation and calibration of the kriging metamodel established by simulation and experimental training samples of the injection mechanism in squeeze casting. The temperature data uncertainty and non-normal distribution are considered in the approach. The normality of the sample data is tested by the Anderson–Darling method. The test results show that the original difference data require transformation for Bayesian testing due to the non-normal distribution. The Box–Cox method is employed for the non-normal transformation. The hypothesis test results of the calibrated kriging model are more reliable after data transformation. The reliability of the kriging metamodel is quantitatively assessed by the calculated Bayes factor and confidence. The Bayesian factor and the confidence level results indicate that the kriging model demonstrates improved accuracy and is acceptable after data transformation. The influence of the threshold ε on both the non-normally and normally distributed data in the model is quantitatively evaluated. The threshold ε has a greater influence and higher sensitivity when applied to the normal data results, based on the rapid increase within a small range of the Bayes factors and confidence levels.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.