Data-Driven Method For Multi-fault Diagnosis In Industrial Process

The production environment of modern manufacturing industry is usually very complex,and the consequences of accidents are very serious.Fault diagnosis technology can effectively improve the safety of the system,reduce the accident risk and improve the economic benefits.Data-driven fault diagnosis technology can avoid the difficulty of mechanism modeling.It is based on the massive data collected in the industrial process to analyze the fault,and has become one of the hot spots in business and academia.The method based on multivariate statistical analysis is widely used in industrial process monitoring.After fault is detected,it is often necessary to trace the source of the fault and locate fault variables.Traditional fault diagnosis methods include Contribution Plot method,Reconstruction-based Contribution method(RBC)and Bayesian Decision method,but these methods have following problems to varying degrees: firstly,there are "smearing effect" between variables,which means that the normal variable is affected by the fault variable and misjudged as a fault;Secondly,under the complex situation of multifault and micro-fault,the existing methods are unable to extract effective fault features,resulting in misdiagnosis and missed diagnosis;Thirdly,the computational complexity of some methods is too high and has no practical application value.To solve these problems,this thesis proposes a fault diagnosis method based on improved Fisher Discriminant Analysis.Firstly,the classification idea of Fisher Discriminant Analysis is analyzed geometrically,and then the classification principle is adjusted to extract the fault features.Then the traditional RBC is extended to multidimensional level,and the fault variables are located based on the extracted fault features.Finally,the effectiveness of the proposed method is verified by the industrial process simulation.This method solves the multi-fault diagnosis problem well and can be applied to smaller faults,and greatly reduces the diagnosis complexity.Next,a fault diagnosis method based on bayesian theory and multi-dimensional RBC is proposed.Firstly,the normal data is trained to obtain the probability density function of variables,and the concept of "deviation factor" is defined,and its probability model is derived.Then the posterior probability is calculated by bayesian theory,and then the fault variable is determined by combining with multi-dimensional RBC.Finally,the superiority of the method is proved by industrial process example.Compared with the improved Fisher Discriminant Analysis method,this method has better interpretation ability,more instructive posterior probability and further improves the adaptability to minor faults.