Minerals Flotation Process Optimization

Building PCA Models for Fault Detection in an Industrial

Rougher Flotation Circuit.

Luis G. Bergh and Felipe Niada.

Automation and Supervision Center for Mineral Industry, CASIM.

Chemical Engineering Department, Santa Maria University, Valparaiso, Chile.

ABSTRACT

On line fault detection, for instrumentation and process operation, has become important part of industrial programs leading to improve process operation and therefore product quality over time. Today, great amount of process variables are routinely collected at high frequency by Distributed Control Systems (DCS). Also, many variables, mainly related to the quality of a product, such as the concentrate grade and process recovery in the flotation processes are infrequently available. High problem dimensionality, highly correlated process input variables, rather low signal/noise ratios and missing data are some of the main difficulties found in modeling the process for monitoring and diagnosis purposes. Multivariate statistical projection methods, such as Principal Component Analysis (PCA), have been proposed to effectively deal with these situations. In this work, an industrial rougher flotation circuit is operated under distributed control of froth depth, and chemical reactive dosages, to experimentally collect operation data at steady state, to build a PCA model. The rougher circuit is formed by 5 banks with an array of 2-3-3-3-3 cells of 3000 ft3. Each bank has a froth depth measurement and control and a Metso camera to send froth surface images to be processed. Feed, final concentrate and final tailings are obtained by on stream analysis, using Courier system from Outokumpu. Feed flow rate and its characteristics are also measured. The dosage of chemical reactive are measured and controlled as a ratiowith feed tonnage. The image processing system from Metso produces almost 30 variables characterizing the flotation froth. Some years ago, several studies were conducted to correlate these froth characteristics with concentrate grade. The results were most of the time inconsistent and no industrial application using the full set of information provided by cameras is known until today. Froth velocity is commonly used to modify the set point of froth depth controllers. The hypothesis that froth characteristics captured by image analysis correspond to the so called top of froth, and that the grade on the froth is not homogeneous, but distributed along the depth, will lead to the conclusion that better models can be found only if feed characteristics, chemical reagent dosages and cell operating conditions (froth depth and sometimes air flow rate) are included in the data. However, in one bank more than fifty variables are involved now, posing difficulties to obtain representative models. Therefore, the use of multivariate statistical projection methods is the appropriate tool to model the correlation existing in the data. In this work, the first part of building alternative PCA models is discussed.

CONCLUSIONS.

Flotation control quality is strongly depending on the accuracy of measurements and estimations. The flotation process is complex and it is a real challenge to decide which variables are to be changed in order to drive back the process to a normal operation. The application of multivariate statistical methods, and particularly PCA, is a powerful tool to build linear models containing the essential of the process phenomena with the minimum number of latent variables. The application of PCA models to monitoring flotation cells, based on the combination of froth characteristics and operating variables has been demonstrated.

These PCA models can be effectively used as part of a supervisory control strategy, especially when control decisions are infrequently made.