A novel knowledge enhanced graph neural networks for fault diagnosis with application to blast furnace process safety

Abstract

With the advent of industry 4.0, many traditional industries are moving toward automation, intelligence, and large-scale. The continuous expansion of production scale also means that the structure of industrial processes and the interactions between subsystems are becoming increas- ingly complex, which also introduced potential safety risks into the actual production. Due to the complexity and danger of the production process, the process safety of the modern blast furnace (BF) ironmaking process is a prominent problem. In this paper, a novel fault detection and di- agnosis (FDD) framework is proposed, which can detect the abnormality and infer the root cause with no need of building an accurate mechanism model as the traditional methods. After a fault is detected, to better learn fault propagation mechanisms of the BF ironmaking process, a process knowledge model was proposed and integrated with the data-driven approaches for fault diagnosis. Then, to discover the causalities matrix of the faulty data generating procedure, a novel root cause analysis method based on Graph Neural Networks (GNN) was developed. Moreover, to accurately describe the causal interactions of faulty variables and solve the problem of the redundant edges in the causal discovery, the process knowledge constraint item was added to the GNN model to guarantee the discovered causal graph matches with the practical domain knowledge to strengthen its application in the practical industry. The experimental results on real BF data set and a bench- mark data set (The RT 580 data set) demonstrate that our algorithm not only obtains a significant improvement over other methods but also has a favorable application in the industrial process.