A Data-driven Soft Sensor for Predicting Grade and Recovery in a Platinum Concentrator

Abstract

The timely and accurate prediction of the grade and recovery in flotation circuits is critical for optimising metallurgical operations, where even a 1% increase in throughput can yield significant profit improvements. Traditional laboratory analyses often cause significant delays that limit the ability to adapt to the rapid and dynamic changes inherent in these processes. This study investigated the feasibility of a data-driven soft sensor that uses machine learning to infer the grade and recovery in real time. A comprehensive dataset collected from the Anglo-American archive, consisting of key flotation process variables, was pre-processed using robust feature selection techniques. This ensured that only the most influential parameters, such as the reagent concentrations, sump conditions, and cell-level measurements, were retained. The performance of various machine learning models, linear regression, neural networks, random forest and extreme gradient boosting was evaluated to predict the grade of platinum group metals (PGMs) in a flotation circuit. Although linear regression provides a useful baseline understanding of the process, its poor performance reinforces the inherent nonlinearity of the process. Among nonlinear models, Random Forest achieved the best generalisation performance, demonstrating robust predictive accuracy on unseen data. Extreme gradient boosting (XGBoost), while exceptionally well capturing complex dynamics on the training set with a coefficient of determination (R2) near 99%, exhibited signs of overfitting, owing to its lower performance on the unseen data. Neural networks have also demonstrated strong generalisation underscoring the potential of deep learning in this application. Moreover, hyperparameter tuning is essential for optimising non-linear models for optimum performance. Random forest is a reliable model for the development of data-driven soft sensors to improve real-time process control in flotation circuits. Real-time predictions were achieved with minimal computational overhead, offering a viable alternative to the conventional time-consuming laboratory methods. This integrated soft sensor not only enhances process control but also provides operators and metallurgists with actionable insights that ultimately contribute to more efficient and responsive flotation operations.