Bayesian Inference Based Industrial-process Monitoring Methods

Process monitoring technology is the effective way to ensure process safety and toimprove product quality for industrial processes. With the rapid development of scienceand technology, especially the computer technology, massive real-time data informationhas been one of the most distinct charactors of modern industrial processes.In thebeginning of21stcentury, data-driven based process monitoring methods were generatedat the right moment and gained wide attention in both academy and industry. Among them,multivariate statistical process monitoring (MSPM) is a research hotspot in current processmonitoring area,Fault detection and fault diagnosis are two main research directions for processmonitoring area. Most process monitoring studies based on MSPM methods have focusedon fault detection, principal component analysis (PCA) and partial least squares (PLS) asthe core methods have developed relatively well, by contrast MSPM performs lesssatisfactorily in fault diagnosis and identification areas. The commonly used unsupervisedbased methods of fault diagnosis include completely decomposed contribution (CDC)method and reconstruction-based contribution (RBC) method. Both methods have thesmearing effect, that is, faulty variables will “pollute” their impacts on the fault detectionindex to the non-faulty variables, resulting in false fault diagnosis.This paper proposed a new unsupervised data-driven fault diagnosis method bycombining reconstruction-based contribution and Bayesian inference, together with thePCA-based fault detection method. This approach, which applyed previous diagnosisresults into the fault diagnosis of currect sample, can differentiate the major fault variablesand minor fault variables, and ultimately eliminate the smearing effect in fault diagnosis,based on the accumulation effect of fault information. During the really complex industrialprocesses, process faults rarely show a random behavior, on the contrary, they will begradually propagated to varing variables due to the actions of process controllers and thecorrelations between variables, which is called fault propagation. Both CDC and RBC cannot well track the fault propagation process, but the method proposed in this paper could analyze different faulty variables during fault propagation and locate the ultimate rootfaulty variables in the end.Finally, simple numerical simulations and complex Tennessee Eastman (TE) industrialprocess were used to verify the effectiveness and applicability of the proposed method interms of solving those two problems including smearing effect and fault propagation.Conclusion can be drawed that this method provides a new thought and solution for thefault diagnosis in practically complex industrial processes.