Sample-Evaluation-Enhanced Machine Learning Approach for Fault Diagnosis of Hybrid Systems

Industrial processes are hybrid systems, when they involve both discrete and continuous variables. Discrete variables contain valuable and easily comprehensible information, particularly crucial information about fault occurrences. To use this information, this study proposes a hybrid neural network model that integrates knowledge and data. A knowledge-based expert system is developed using discrete variables to assist the neural network in decision-making tasks. This study proposes a fusing expert system-based neural network optimization algorithm for hybrid systems fault diagnosis in industrial processes. Continuous variables are used as learning objects for the neural network, whereas guidance for the learning process is provided by discrete variable information. The mixed integer linear programming (MILP) algorithm is used to extract logic rules from discrete variables, and the influence of samples in the network learning algorithm is evaluated on the basis of the attributes of these logic rules. The effectiveness of the proposed algorithm is validated through experimental verification in a three-tank simulation system and a practical industrial case of coal gasification processes. Experimental results demonstrate that the proposed algorithm exhibits higher fault diagnostic performance, better generalization ability, and stronger adaptability to sparse samples than baseline methods.