Root cause diagnosis in process industry via Bayesian network enhanced by prior knowledge and randomized optimization

Abstract

This study presents a novel root cause diagnosis algorithm for steel production that combines the Peter Clark (PC) algorithm and the Discrete Random Genetic-Particle Swarm Optimization (DRGAPSO) algorithm. This synthesized approach combines prior knowledge and preserves some coupling between variables to more accurately reflect real-world production scenarios. The prior knowledge is coded into the PC algorithm, while the DRGAPSO algorithm partially breaks through the limitations of the causal relationships obtained by the PC algorithm due to the addition of stochastic operators, refining these causal relationships to create a complete Bayesian network containing correlations. The propagation probabilities between variables are then calculated to trace the fault propagation path. The method was validated using real-world data from Huaxi Iron and Steel Co. to generate visualized fault tracking paths to demonstrate its effectiveness. The proposed method significantly outperforms other similar schemes in terms of structural scoring, and the comparison of the visualization results further highlights the reliability of the proposed method in root cause analysis of faults, making it an important tool for improving the quality of steel production products.