A Novel Fault Diagnosis Framework for Industrial Production Processes Based on Causal Network Inference

With the advent of Industry 4.0, the introduction of data-driven approaches into industrial processes for fault diagnosis has gained substantial attention due to their significant advantage that they mainly rely on the information on process data instead of a priori knowledge. However, the statistically based data-driven methods have difficulty eliminating the "smearing effect" between variables, which affects their effectiveness and interpretability for fault diagnosis, while the previous studies on the causal-based fault diagnosis methods are seriously insufficient. In this study, a novel data-driven fault diagnosis framework based on causal network inference was developed, in which the correlations between variables are explored by employing the partial correlation network (PC-NET) method and the causal propagation direction are determined by a newly developed partial conditional Granger causality (PCGC) method based on the transfer entropy. Subsequently, the occurrences of faults are detected by using a causal-based multivariable sensitivity enhancing transformation (MSET) approach. Finally, a causality-attributing reconstruction-based contribution (RBC) method is developed to isolate and identify the fault variables and to classify the fault grade for taking remedial measures. The effectiveness of the proposed fault diagnosis framework is verified by its application in the actuator system of the industrial sugar production process, and the results demonstrate that the proposed fault diagnosis framework can not only eliminate the smearing effects but also accurately identify fault variables and their contribution rates, exhibiting better interpretability and scalability.