Maximising mill throughput using machine learning techniques and evolutionary algorithms
Date
2025
Authors
Ghasemi, Zahra
Editors
Advisors
Chen, Lei
Neumann, Frank
Zanin, Max
Neumann, Frank
Zanin, Max
Journal Title
Journal ISSN
Volume Title
Type:
Thesis
Citation
Statement of Responsibility
Conference Name
Abstract
Grinding is a pivotal step in mineral processing plants, accounting for approximately half of all mineral processing costs. Semi-autogenous grinding (SAG) mills are commonly used for grinding, facilitating mineral liberation and preparing ore for subsequent processing steps, such as flotation. SAG mill throughput is one of the key performance indicators of a SAG mill. Maximising SAG mill throughput is an important objective in mineral processing plants, as it can lead to considerable financial benefits. However, achieving this objective is challenging, as numerous governing variables impact mill throughput. The relationships between these factors and mill throughput are highly nonlinear, and the variables interact with each other, making it difficult to determine the independent effect of each variable. In the current era of artificial intelligence (AI) and machine learning (ML), these methods offer promising solutions to the complexities of mineral processing. This thesis addresses the challenge of maximising SAG mill throughput by leveraging these advanced methods, with a focus on the following four key challenges. First, the baseline for maximising SAG mill throughput is accurately modelling SAG mill throughput. This is challenging due to the high multivariable complexity and interactions among input features. Empirical models for this purpose are costly, timeconsuming, and limited to the conditions under which experiments were performed. There has not been a comprehensive exploration of ML models utilising process data for this purpose. Furthermore, the effect of data delays on prediction accuracy has not been examined in the current literature. This challenge will be addressed in the first step of this thesis. Second, grinding is a highly dynamic process, as many impactful features change over time. Capturing these complexities with a single model is challenging. To address this, a new genetic programming method is developed in the second step of this research. This method generates and utilises multiple equations instead of relying on a single model, thereby improving SAG mill throughput prediction. Third, optimising process parameters to maximise mill throughput presents another important challenge. Control process parameters are typically set based on manufacturers’ recommendations or expert knowledge. However, there is a lack of research on utilising ML and AI to identify optimal process parameters. This challenge is addressed in the third step of this research. Finally, similar to most real-world optimisation problems, there are other objectives to consider alongside SAG mill throughput maximisation. Specifically, balancing throughput maximisation with minimising the circulating load is a significant topic that has not been explored in the literature. This challenge will be addressed in the fourth step of this research by developing a multi-objective optimisation framework. It is worth mentioning that one important contribution of this thesis is the use of extensively accumulated operational data, which is rarely utilised for modelling, to develop an accurate mill throughput prediction model that can be adopted for various types of mineral processing plants. The aim of this research is not merely the development of ML or AI techniques. Instead, it focuses on investigating how these advanced methods can be utilised for enhanced decision-making and process control in the grinding step of mineral processing plants. By aligning these techniques with industrial needs, the research aims to maximise mill throughput, contributing to more efficient, reliable, and optimised plant operations. In this research, industrial datasets from SAG mill and ball mill closed-loop processing circuits are utilised as the foundation for analysis and model development. This thesis consists of four peer-reviewed papers, three of which have been published and one that has been submitted for publication. In the first step towards maximising mill throughput, a comprehensive comparative study is performed to identify the most accurate machine learning model for predicting SAG mill throughput. The compared models include genetic programming, recurrent neural networks, support vector regression, regression trees, random forest regression, and linear regression. A real-world data set consisting of 20,161 records is used for this purpose. In this research, for the first time, the delay in data is identified and incorporated into the preprocessing step by applying the cross-correlation method to increase prediction accuracy. As there are different parameters in each model that can impact prediction accuracy, hyperparameter tuning is conducted to optimise the performance of each model. The obtained results revealed that recurrent neural networks are the best-performing models, followed by genetic programming and support vector regression. The recurrent neural network model is then utilised for sensitivity analysis to identify the effect of different input parameters. The analysis indicated that SAG mill throughput is primarily influenced by the mill turning speed and inlet water. SAG mill throughput is expected to increase with higher mill turning speed and lower inlet water, within the limited operating parameters. In the next stage, an enhanced version of genetic programming, named multiequation genetic programming (MEGP), is developed for more accurate prediction of SAG mill throughput. The reason for focusing on GP is its strong performance, having been ranked second in the previous step, as well as its advantage of transparency in providing equations for prediction. MEGP is comprised of two main steps: clustering and predicting. In the clustering step, different categories of data are identified such that each cluster can be accurately modelled with a distinct equation. These equations are then utilised in the prediction step by applying various prediction approaches. MEGP was implemented using four different distance measures, including Euclidean, Manhattan, Chebyshev, and Cosine distance, to assess the impact of these measures on the model’s performance. MEGP, with the best-performing prediction approach, improved prediction accuracy by 10.74% compared with the standard GP. This approach utilises all extracted equations and includes both the number of data points in each data cluster and the distance to clusters as the weighting factor. Furthermore, the Euclidean distance measure resulted in the highest prediction accuracy among the compared methods. In the third stage, an integrated intelligent framework is developed for maximising SAG mill throughput. This framework comprises an ML-based prediction module and an EA-based optimisation module. Local outlier factor (LOF) and recursive feature elimination (RFE) techniques are utilised for outlier detection and feature selection, respectively, to enhance predictive performance. The results revealed that the implementation of the LOF method for outlier detection did not yield substantial improvements, but utilising RFE showed a positive impact, resulting in the removal of five input features. The problem is modelled as a constrained optimisation problem, considering the working limits of input features and particle size distribution requirements as constraints to ensure that the proposed solutions are practically deployable. Various ML models and EAs are compared to identify the best-performing ones for the utilised dataset. The developed framework is able to propose process set points that will lead to maximised SAG mill throughput. Finally the fourth step extends the framework developed in the previous step by introducing a second objective as minimising the circulating load. Circulating load is the quantity of large particles that cannot pass through the sieve at the SAG mill discharge, as they are larger than the sieve aperture size. By adding this objective, the framework ensures that while mill throughput is maximised, the quality of the grinding process is also considered. For this purpose, the problem is modelled as a multi-objective optimisation problem, and various advanced EAs for multi-objective optimisation are utilised. A comparison of hypervolume values revealed that the Nondominated Sorting Genetic Algorithm II (NSGA-II) is the best-performing optimiser. A sensitivity analysis confirmed the robustness of the optimal solutions proposed by NSGA-II, as changing input feature values around the optimal values could not improve objective values. This extended framework is able to propose grinding process set points that maximise mill throughput and minimise circulating load. Furthermore, the optimal solutions are a set of Pareto-optimal solutions rather than a single solution, enabling process experts to select the best settings based on actual process conditions. It is important to note that while the developed models in the research are tailored to the studied system, they can be continually updated with new data inputs, enabling a dynamic and adaptive approach to process modelling. This adaptability allows the framework to be retrained and adjusted for different systems or evolving operational conditions, enhancing its practical value beyond a static solution. It is also worth mentioning that the differences in the number of data points evaluated in each chapter are due to the use of different datasets provided by the industry partner at various stages of the research. While all datasets originate from the same industrial process, newer versions were made available over time. These updated datasets were more complete or included additional relevant features, which allowed for deeper analysis and model refinement in later chapters.
School/Discipline
School of Electrical and Mechanical Engineering
Dissertation Note
Thesis (Ph.D.) -- University of Adelaide, School of Electrical and Mechanical Engineering, 2025
Provenance
This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals