Multivariate Data Analysis for Process Evaluation, Prediction and Monitoring at INCO's Copper Cliff Smelting and Refining Operations

Abstract

Industrial processes generate large quantities of process and product quality data. Most of this data is stored and is analyzed in a univariate fashion. However important information may be lost through the implementation of univariate analysis methods. This information is contained in the correlation structure amongst the process and product quality variables and between these two types of variables. Through multivariate analysis this information is retained. As a result process evaluation, prediction and monitoring are more effectively performed. Multivariate data analysis techniques were therefore applied to data sets that summarized three of INCO's Copper Cliff smelting and refining processes. In the first instance the analysis of historical data pertaining to a batch leaching process was undertaken and the time required for leaching was predicted. In the second a multivariate soft sensor was developed in order to predict the concentration of nitric oxide contained in the feed gas to the smelter acid plant. The final project involved the analysis and monitoring of a continuous nickel carbonyl process. The resulting models were evaluated and significant variables with respect to the variation in the process and product quality data and the correlation between them were identified. The product quality data was also well predicted using new process data only that was input to the models. Finally new data was input to the models and the process was monitored using a reduced number of latent variables. Contribution plots were used to identify the original variables that contributed most to the observations that exceeded the established control limits.