Research On Fault Monitoring Method In Industrial Process Based On Block-division Strategy

With the constant expansion of industrial production,an industrial process usually consists of many manufacturing cells and numerous measured variables Meanwhile,industrial process data usually present the characteristics of nonlinear correlationship,non-Gaussian distribution and dynamic,which renders fault monitoring methods based on a global monitoring model difficult to meet the needs of current industrial production.Multiblock fault monitoring methods based on block-division strategy can effectively handle the above issues,and they have been widely investigated.Aiming at the issues existing in modern industrial process fault monitoring,such as difficulty in extracting local information,single monitoring information,problem with determining the division threshold of variable blocks,this paper conducts research on industrial process fault monitoring methods based on block-division strategy.The main contents of this paper are as follows:(1)The existing multiblock fault monitoring methods only rely on the linear relationships between variables for variable block division,ignoring the nonlinear relationships and other high-order correlations between variables,and this manner of variable division is not reasonable.To address this problem,a multiblock PCA fault monitoring method based on JS divergence is investigated.Firstly,JS divergence,a probabilistic statistical method,is used to divide variable blocks,which improves the rationality of variable block division and highlights the process local characteristics.Then,a PCA monitoring model is established in each variable block.Finally,the Bayesian inference method is used to fuse the monitoring results of each variable block to obtain a global monitoring indicator,so as to determine more intuitively whether a fault occurs.Simulation experiment on the TE process shows that the proposed algorithm achieves better monitoring performances than some traditional fault monitoring methods.(2)Given that the existing multiblock fault monitoring methods merely utilize the original observation information of process data,the diversity of monitoring information is insufficient.Under this circumstance,it is hard to detect small or continuous-oscillating faults effectively.In order to solve this problem,a multiblock PCA fault monitoring method based on multiple features extraction is investigated.Firstly,JS divergence is uesd to divide process variables into blocks for extracting process local features.Secondly,latent features such as cumulative error and first-order difference information are extracted from each variable block to complete sub-block expansion.Then,a PCA monitoring model is built in each information sub-block Lastly,a global monitoring index is generated by the Bayesian inference method.Simulation experiments on a numerical example and the TE process verify the effectiveness of the proposed algorithm.(3)To overcome the problems that traditional multiblock PCA fault monitoring methods cannot handle non-Gaussian data and the division threshold of variable blocks is difficult to determine,a multiblock PCA-ICA fault monitoring method based on J-B test and weighted JS divergence is investigated.Since PCA fault monitoring methods are based on the condition that process data ought to satisfy the Gaussian distribution,the proposed method applys the method of Jarque-Bera test to recognize the Gaussianity of process variables one by one and divides process data into the Gaussian and non-Gaussian block.Considering the difficulty in determining the division threshold of variable blocks,the proposed method introduces the block-division strategy of weighted JS divergence.Through this strategy,the Gaussian and non-Gaussian block are expanded into the Gaussian and non-Gaussian weighted sub-blocks,in which PCA and ICA monitoring models are established respectively.Finally,the monitoring results of each weighted sub-block are merged together by the Bayesian inference method for the sake of the global monitoring result.The feasibility of the proposed algorithm is proved by the simulation experiment on the TE process and the practical application of the blast furnace ironmaking process.