Diagnostics of sintering processes on the basis of PCA and two-level neural network model

The application of chemometrics methods to continuous monitoring and diagnostics of sintering process faults for improving iron-ore sinter quality is considered in the article. The sintering process is among complex multivariate processes. A number of agglomeration process faults have often similar symptoms, resulting in late fault detection by an operator and as a consequence, wrong process control decisions. To support the efficient operative decision making, it is proposed to use the process fault monitoring and diagnostics system. The proposed system uses a two-level neural network (NN) diagnostic model. The high-level neural network is used to localize the process faults whereas their reasons are determined by the low-level neural networks. To reduce essentially the time of HL-NN training and retraining, the task dimension is preliminarily reduced with the principal component analysis method so that the scores obtained from initial data are fed into high-level neural network inputs. The use of principal component analysis allowed detection of sintering process faults with T-2 and Q statistics. Only upon detecting the fault, a NN diagnostic model starts working to determine the fault reason. The system algorithm provides for special measures to prevent the NN from possible loss of the identified fault due to operator's inactivity. To increase the diagnosis depth for controlling fault symptoms that are evident on the sinter cake surface, optical digital cameras are installed and images from them are processed with proposed algorithms on the basis of fuzzy clusterization to take into account uncertainties in the initial information. In the article, the process fault monitoring and diagnostics system of agglomeration process is considered. The proposed system is based on PCA and the two-level artificial neural network model. The high-level network is used for localization of the sintering process faults, while low-level networks identify the causes of these faults. The use of PCA and hierarchical neural network model allowed to reduce significantly the dimension of the networks and to cut the time needed for their retraining.