Process Monitoring Methods Based On Statistical Machine Learning For Process Industries

With the application of automatic system and the development of information technology,process industries such as steel,petrochemical,nonferrous metals and building materials are developing in the direction of large scale,complexity and integration.The production process is composed of many sections from raw materials to final products,which forms a plant-wide process.At the same time,customized production makes its varieties and specifications show diversity,coupled with the uncertainty of raw materials,different equipment status,external environment and process technology,which makes its working modes changeable.In addition,process industries have multi-level characteristics,mainly including equipment level,real-time control level,process control level and manufacturing execution level.Each level has a clear division of labor and is related to each other.This makes the process industry present complexity from three dimensions(process,working condition and level).Safety and quality control of product play a key role to process industries.No matter which dimension is abnormal,it will propagate along three-dimension and lead to fault evolution.Incipent faults in the process will most likely cause serious consequences such as loss of life and property,damage to the ecological environment.This study aims to comprehensively improve the safety and reliability of process industry.From the three dimensions,this study focuses on the theoretical research and application verification of fault detection and operating performance assessment in the process industry,striving to effectively reduce or avoid the occurrence of faults,ensure product quality and improve the economic benefits of enterprises.The innovations mainly include the following aspects:To deal with incipent faults in plant-wide process industries,an incipent fault detection and identification strategy is proposed.First,the normal and fault samples are introduced into the support vector data description(SVDD),the radius is improved,and the robust monitoring model is established to realize the incipent fault detection.Then,the fault information is extracted by stacked restricted Boltzmann machine(RBM).Finally,the extracted information is used to train the probabilistic neural network to realize the incipent fault identification.To solve the problem of many variables and coupling correlations in the process industry,an adaptive process monitoring framework based on hybrid similarity measure is proposed.First,the mixed similarity measure is established by using the maximal information coefficient and general Jaccard coefficient to realize the block division.Then,the gap metric and SVDD are fused to establish the monitoring model.Finally,the adaptive radius is calculated to realize the whole process adaptive monitoring.In order to solve the problem of product specifications changing frequently and switching between on load and idle conditions,a common-individual subspace modeling and full condition monitoring strategy are proposed.First,the minimum error minimax probability machine is improved to realize multimode identification.Second,the common-individual subspace model is constructed by parametric t-distribution stochastic neighborhood embedding(t-SNE)and quality data.Then,SVDD is used to monitor the idle condition.Finally,based on t-SNE,slow feature analysis and co-integration analysis,the dynamic and static monitoring of nonlinear process under load condition is realized.Aiming at the coupling and correlation of process variables,the imbalance of samples,and the outlier problem,a lifecycle operating performance assessment is proposed for multi-level plant-wide process with partial communication.First,according to the process knowledge and communication situation,the data form upstream,midstream,downstream and process control level are processed.Then,the deep belief network,the hybrid sampling boosting,Partial robust M-regression,and other algorithms are organically combined to realize the identification of operating conditions and the effective assessment of different operating conditions,and trace the causes of non-optimization status.The above research results provide new ideas and solutions for fault detection and operating performance assessment in the process industry.The proposed methods are verified in the Tennessee Eastman process and the hot strip mill process.Compared with the traditional methods,the experimental results show that the proposed methods are effective and practical.