Applying Machine Learning Techniques to Improving Truck Productivity Prediction Accuracy at Mine Sites

In oil sands mining, off-the-road trucks play a leading role in transporting bulk materials (ores and waste). The productivity of truck haulage (also referred to as truck productivity), defined as the truck payload per unit time in each truck haulage cycle, is of great interest to the mining industry since truck productivity is directly associated with a mine’s overall productivity. Accurate truck productivity prediction is significant for making budget decisions and developing mine planning. However, the current approach used for predicting truck productivity has four major concerns, leading to inaccurate predictions of truck productivity at mine sites. First, the approach (i.e., curve-fitting) is built based on average values. Second, only one input variable (i.e., haul distance) is involved. Third, simple regression method (i.e., least squares) is used to construct prediction models. Fourth, temporal resolutions are not considered in building prediction models. In response to these concerns, this Ph.D. thesis aims to apply machine learning techniques to improving truck productivity prediction accuracy at mine sites. In particular, this thesis focuses on developing a unified toolkit for truck productivity prediction in oil sands mining, which consists of various machine learning models built based on massive truck haulage data at varying temporal resolutions (e.g., per cycle, hour, day, week, and month). These machine learning algorithms were employed to train complex relationships between truck productivity and multiple input variables, analyze the contributions of input variables to the model output, investigate the effect of temporal resolutions on establishing prediction models, and design a unified graphical user interface (GUI). The results showed that Gaussian mixture modeling (GMM) efficiently clustered massive truck haulage data into three subgroups (i.e., low, medium, and high truck productivity) and significantly improved the model accuracy. For example, a multiple linear regression (MLR) model reached a coefficient of determination (R2) of 75% based on GMM analysis, which was much higher than the MLR model (23%) before clustering. After that, nonlinear algorithms were used to build more complex truck productivity prediction models. The results presented that the tree-based ensemble models performed better than single models in predicting truck productivity. Also, the Bayesian regularized neural network (BRNN) model outperformed the back propagation neural network (BPNN) and extreme learning machine (ELM) models. For these machine learning models without considering temporal resolutions, haul distance contributed the most in constructing linear and nonlinear relationships. When involving temporal resolutions, the nonlinear relationship between inputs and truck productivity progressively diminished with decreasing temporal resolutions (i.e., from hourly to monthly). Regardless of the temporal resolutions, the three most influential input variables were haul distance, empty speed, and ambient temperature. In addition, mining engineers can make more accurate predictions of truck productivity at the weekly resolution compared with other resolutions. The feature importance of the four weather-related input variables increased as decreasing temporal resolutions. Extreme weather, such as extreme wind speed, precipitation, and relative humidity, had a certain effect on truck-shovel allocation at mine sites. Finally, a unified GUI was designed and developed for the first time to predict truck productivity at varying temporal resolutions. Overall, this thesis developed a unified toolkit to improve truck productivity prediction in oil sands mining. The findings will help mine management better understand and forecast truck productivity for hauling efficiency improvement, strategic decision-making, and cost reductions in oil sands mining and other mine sites using truck haulage.