Application of Bayesian networks based on Sequential Monte Carlo simulation and physical model in fault diagnosis of horizontal three-phase separator system

Abstract

In offshore production platforms, horizontal three-phase separator is common process equipment. Its main function is to complete the dehydration and degassing of crude oil. Separator system is more complex, and its failure may cause significant economic losses and disastrous consequences. Therefore, it is critical to accurately and quickly identify where and why faults occur in the separator system. In this study, separator fault diagnosis model based on Bayesian networks is developed. Moreover, Sequential Monte Carlo simulation and physical model are introduced to overcome field problems such as missing separator failure data and the inability of experts to provide accurate empirical knowledge. Using this model, 13 faults in a separator in an offshore crude oil processing system are successfully diagnosed. Meanwhile, the proposed model is compared with deep neural network, convolutional neural network, and deep residual network, with accuracy rates of 100%, 91.34%, 87.99%, and 94.62%, respectively. Then, the diagnostic accuracy of each model for different faults is also compared in this paper under various signal-to-noise ratios. The results show that the method proposed in this paper has better noise immunity compared to the other three models. Therefore, the accuracy and robustness of the proposed model is further demonstrated. Finally, to analyze the fault-tolerance of the proposed model, 2–3 error evidence is randomly entered. The results show that the proposed model has better fault tolerance compared to data-driven Bayesian networks.