Fault prediction method for a nickel flash smelting furnace system based on graph fusion and multi-stage learning

Abstract Considering the characteristics of nickel flash smelting furnace system data, which exhibit both static and dynamic properties, this paper proposes a graph fusion network model combined with a multi-stage learning strategy to efficiently accomplish fault prediction tasks. The model employs a data-knowledge dual-driven approach to construct static and dynamic graphs, and adaptively balances the influence of the two graph structures through a proportional coefficient. Additionally, to improve the model’s execution efficiency, a multi-stage learning strategy and fixed-length sliding window are utilised to focus on recent data trends and reduce computational burden. Experimental results demonstrate that the proposed model significantly enhances the accuracy of fault prediction in nickel flash smelting furnace system, outperforming existing methods and effectively supporting predictive maintenance and other application requirements.