Fault Detection Based On Robust Principal Component Analysis In A Blast Furnace Process

The blast furnace ironmaking is one of the most important parts of national economic system.The occurrence of anomalous event in the blast furnace processes results in enormous economic losses and safety hazard.To ensure the safety and reliability of the blast furnace processes,fault detection is necessary.Generally speaking,in the blast furnace processes,three primary fault detection methods are utilized:analytic model methods,knowledge-based methods and data-driven methods.Recently,the distributed control system is widely used,which means a large amount of process data could be collected.Moreover,compared with the other two methods,the data-driven methods do not need exact process models and a rich fund of expert knowledge.Therefore,the data-driven methods are widely utilized in the blast furnace processes.The blast furnaces always work under high temperature;high pressure and high dust degree environments,and exist complex physical and chemical reactions all the time.Moreover,the blast furnaces are semi-automation production equipment,which need to be controlled by operators and adjusted parameters.Therefore,the collected data matrix may contain some process noise,outliers and minor faults,which would lead to poor fault detection performance by utilizing traditional data-driven methods.Aiming at these issues,in this paper,the author proposes three methods and presents simulations based on numerical simulations,the TE process and a real blast furnace process accordingly:(1)Aiming at the process noise contained in the collected data matrix of the blast furnace processes,the principal component pursuit method is used for fault detection and a correlation-based statistic is proposed.This method is an on-line fault detection method,and divides the training and new testing matrix based on principal component pursuit method respectively.Then,the low rank matrix and sparse matrix are obtained.The low rank matrix of training matrix contains the important information of processes,while the sparse matrix of training matrix may contain the process noise.Then,the correlation coefficient of variables in the low rank matrix of training matrix are calculated.Utilizing the Hotelling's T2 statistic in the low rank matrix and correlation-based statistic in the sparse matrix could detect the faults,and judge the current conditions of blast furnaces.Compared with the traditional fault detection methods,the principal component pursuit could reduce the influences of process noise and there is no restricted conditions of data.The types of data could be nonlinearity,nongaussianity and so on.(2)Aiming at the outliers contained in the collected data matrix of the blast furnace processes,the improved principal component pursuit method is proposed for fault detection and a novel statistic is proposed.The improved principal component pursuit method solves a convex optimization problem to obtain a low rank coefficient matrix and a sparse matrix.The low rank coefficient matrix contains the important relations between variables,which is obtained under the outlier-free condition and reduces the influences of outliers.The sparse matrix may contain the outliers.Then the low rank coefficient matrix and the correlation coefficient between variables are used for constructing the statistic.The proposed method could obtain the low rank coefficient matrix after removing the outliers.Not only reduces the influences of outliers,but also obtains the essence relations of variables,which could improve the performance of fault detection.(3)Aiming at the minor faults contained in the collected data matrix of the blast furnace processes,the robust principal component pursuit method is proposed for fault detection.The small changes in temperature and parts raw material hanging may lead to minor faults.A minor fault is an abnormal condition occurring at some variables for a period time,which is different from outliers.Therefore,a novel fault detection method is needed.The proposed method divides the data matrix into two parts:a low rank matrix containing the process important information and a sparse matrix containing the minor faults by solving a convex optimization function.The minor faults are divided from rows and columns of data matrix simultaneously.Then the Hotelling's T2 statistic is used for fault detection in the low rank matrix,which is the same as in the principal component analysis method.The final parts concludes the main points of paper,and the future research directions are discussed.