Predicting Extraction Excursions Using Explainable AI

Kearl oil sands is a large open pit mine in Northern Alberta leveraging a fleet of shovels and autonomous haul trucks feeding two processing facilities. Water based gravity separation is used as a primary method of oil extraction. The ore varies in geologic properties and has different processibility characteristics. In this work we build a model to predict likelihood of elevated extraction excursions based on the ore characteristics and plant processing variables. Previous attempts at this problem have been challenged due to highly noisy data and lack of model transparency. We address the first of these challenges by leveraging process surveillance data to isolate to relevant excursion periods where other factors are negligible. To address the second challenge we employ SHAP to time series to shed light on what is driving the model predictions leading up and during events. Our production model evaluation is promising and has uncovered other potential uses of the approach to not only predict but also highlight process control adjustments or post event root cause analysis