Reliable monitoring of mineral process systems is key to more efficient plant operation. Multivariate statistical process control based on principal component analysis is well-established in industry, but may not be effective when dealing with nonlinear or transient processes, where process behaviour can change rapidly over time. Although many advances have been made in unsupervised process monitoring, the detection of incipient process drift or nascent abnormal process conditions remains an open research problem.In this investigation, monitoring of mineral process operations based on the use of random forest classifiers is considered. When the random forests are used, data representing normal operating conditions are split into subclasses with K-means clustering, after which an ensemble of random forests is trained as binary classifiers on the subclasses in a one-against-the-rest approach. These hyperensemble models are subsequently used to classify new operational data as in-control or out-of-control.One simulated and one real world case study are considered. In each case, the models were trained on data representing normal operating conditions and then tested on new process data that were generally different from the training data to test their ability to identify these data as out-of-control. In the real-world case study, the approach is compared with traditional multivariate statistical process monitoring based on a principal component model, as well as a recently proposed isolation forest model. The random forest-based models performed markedly better than the principal component model and slightly better than the isolation forest model as far as identification of out-of-control data were concerned, but this came at the expensive of a significantly larger false positive rate than these two models.
Read full abstract