Browsing by Author "Groenewald, Jacobus Willem De Villiers"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    A process performance monitoring methodology for mineral processing plants
    (Stellenbosch : Stellenbosch University, 2014-04) Groenewald, Jacobus Willem De Villiers; Aldrich, C.; Bradshaw, S. M.; Akdogan, G.; Stellenbosch University. Faculty of Engineering. Dept. of Process Engineering.
    ENGLISH ABSTRACT: Key to remaining competitive within the mineral industry is ensuring that all processes are always being operated optimally. Process performance monitoring is an ideal initiative with which to accomplish this. Not only can it be used to ensure fault free process operation, but it can also be applied for plant performance improvement through a better understanding of the contributors to the success or failure of the process operation. Critical to the success of any proposed monitoring approach would be its ability to cater for the fact that these mineral processes are typically highly complex, dynamic and non-linear. The purpose of this study was to propose and evaluate a methodical approach to plant-wide process performance monitoring for mineral processing plants. Crucial to this approach is the concept of integrating process causality maps with data-based systems for event detection and diagnosis. To this end, process causality maps were developed to provide a means of structuring process data through the use of fundamental process knowledge. Statistical data-based fault detection techniques, being especially powerful with regards to data compression and dimensionality reduction, were employed to allow huge data sets to be analysed more easily. Change point detection techniques allowed for the identification of stationary segments of data in otherwise non-stationary data sets. Variable importance analysis was used to identify and interpret the variable(s) responsible for the event conditions. Using simulated data sets, different techniques were evaluated in order to acquire an appreciation for their effectiveness and reliability. While it was found that no single technique significantly outperformed any other, it was confirmed that for data having different structures and characteristics, none of the techniques were effective in analysing all potential event conditions. It was suggested that all available techniques be run in parallel, with expert interpretation of the results, ensuring a more comprehensive analysis to be performed. Furthermore, given that only process measurements being monitored could be used to detect events and be analysed for importance, the consequences of monitoring too few process measurements were highlighted. A generic analytical methodology for multivariate process performance monitoring was defined, ensuring the use of appropriate techniques and interpretations. The methodology was subsequently successfully applied to a mineral processing concentrator case study. The application of process causality maps was found to significantly simplify the challenge of monitoring the process, not only improving the ability of the techniques applied through a better focussed application, but also the interpretability of the results due to the reduction in complexity. Extreme learning machine, a robust and computationally inexpensive algorithm, was identified as a potential core algorithm for the data analysis techniques forming part of a process performance monitoring solution. With different drivers, at different times, having different effects on the process, visual representation of the data through canonical variate analysis biplots, combined with a sound understanding of the process under investigation, contributed significantly to a better understanding of the important variables for each event condition. From an implementation perspective, adoption of the methodology remains the biggest barrier to success, requiring the most attention in the immediate future.