Optimization of a Dynamic Fault Diagnosis Model Based on Machine Learning

A delay dynamic coupled fault diagnosis (DDCFD) model is a kind of effective fault diagnosis model based on probability graph for complex systems. Previous research assumes that its structure and parameters are all known before application. Since the initial model is always constructed by experts, it is possible that it is inconsistent with practical systems, which may impact the diagnostic accuracy. In order to improve the model by applying the practical data, this paper develops a machine learning method for the DDCFD model. A parameter learning method and three structure learning methods are studied. Using this approach, prior knowledge and test data can be combined together to get a better diagnostic model. The methods are tested and compared on simulated examples and an application case. The result indicates that diagnostic accuracy can be improved by means of parameter learning compared to the initial model, and it can be further improved by structure learning. The score-based algorithm is the best choice for learning the structure of the DDCFD model.