Unsupervised Process Fault Detection with Random Forests

Process monitoring technology plays a vital role in the automation of mineral processing plants, where there is an increased emphasis on safe, cost-effective, and environmentally responsible operation. Members of an important class of advanced diagnostic systems are data-driven and deal with potentially large numbers of variables at any given time by generating diagnostic sequences in lower-dimensional spaces. Despite rapid development in this field, nonlinear process systems remain challenging, and in this investigation, a novel approach to the monitoring of complex systems based on the use of random forest models is proposed. Random forest models consist of ensembles of classification and regression trees in which the model response is determined by voting committees of independent binary decision trees. In this study, a framework for diagnosing steady- and unsteady-state faults with random forests is proposed and demonstrated with simulated and real-world case studies.