Research
1 Applying Machine Learning to Improving Truck Productivity Prediction Accuracy at Mine Mites
Research Gap: According to the literature review, Gaussian mixture modeling (GMM) may be a more suitable option to improve prediction models because large datasets of truck haulage are usually under multi-peak Gaussian distributions. However, no research has reported the application of GMM to preprocess large datasets obtained from mine sites; it is still unknown if GMM can be used to improve the model predictability of truck productivity at mine sites.
Research Objective: This study was designed to handle large datasets of truck haulage using GMM for developing a novel and accurate prediction model of truck productivity.
Highlights:
A large dataset with multi-peak Gaussians was collected from real mine sites.
Gaussian mixture modelling was applied in preprocessing a large dataset.
A latent variable was added to the prediction model in the form of interaction terms.
Data distributions from oil sands mines (using the haul distance as an example). (a) Haul distance is described by a single Gaussian distribution; (b) Haul distance is described by a multi-peak Gaussian distribution.
Extraction of three latent classes from the training dataset. (a) Boxplots of three classes; (b) Histogram: truck productivity corresponds to three latent classes, which are described by Gaussian distributions.
Research Gap: It is promising to apply tree-based ensemble learning algorithms to building prediction models. However, according to the current literature, no studies have reported the use of tree-based ensemble models to predict truck productivity.
Research Objective: This study was to develop prediction models based on the truck haulage dataset using tree-based ensemble learning algorithms to forecast truck productivity.
Highlights:
This is the first study in which the tree-based ensemble models were applied to predicting truck productivity.
Gaussian mixture modeling extracted a latent variable to improve model predictability.
The proposed tree-based ensemble models outperformed the baseline models with high accuracy.
An illustration of the concept of the RF algorithm.
Relative importance analysis of input variables observed at mine sites.
Scatterplots of the measured truck productivity and predicted truck productivity. The GMM-RF model evaluated by (a) testing dataset and (b) training dataset; the GMM-GBR model evaluated by (c) testing dataset and (d) training dataset; the GMM-MLR model evaluated by (e) testing dataset and (f) training dataset; and the GMM-DT model evaluated by (g) testing dataset and (h) training dataset.
(3) Effect of real-site weather conditions under varying temporary resolutions on truck productivity using interpretable machine learning
Research Gap: The temporal scales (or resolutions) of real-site weather conditions were not taken into account in previous studies. For instance, the maximum precipitation over a week (e.g., 85.30 mm in this study) can have a more substantial impact on road conditions and driving habits than an hour (e.g., 14.10 mm in this study). There is a notable research gap in considering the real-site weather conditions with varying temporal resolutions in predicting mine truck productivity.
Research Objective: The purpose of this research was to construct truck productivity prediction models incorporating real-site weather conditions with varying temporal resolutions.
Highlights:
Knowledge-driven methods are proposed to analyze unique and massive real-site data.
A unified GUI is developed for assessing hourly, daily, weekly, and monthly truck productivity.
The effect of temporal resolutions on truck productivity predictive modeling is investigated.
The impact of extreme weather on truck-shovel allocation is investigated by an interpretable method.
The week-to-week prediction of truck productivity is more accurate for mine planning and design.
SHAP summary plots of ten input variables and their instances’ impacts on the model output.
Relationships between the truck haulage-related input variables and SHAP values represented by one-way partial dependence plots (PDP) from the SHAP analysis based on the hourly-GBR model.
The interaction between the prediction (hourly truck productivity) and the weather-related input variables and truck allocation represented by two-way partial dependence plots.
The graphical user interface (GUI) for assessing hourly, daily, weekly, and monthly truck productivity.
A schematic representation of the study framework for evaluating mining truck productivity. (“###”: the input information is not disclosed as it is the proprietary property of mining companies.)
2 Deep Neural Networks for Forecasting Ore Production at Mine Sites
Research Gap: Studies that predict ore production by incorporating truck haulage information and weather conditions are still scarce due to the high dimensional and nonlinear relationships between these numerous variables that need to be addressed. According to the literature review, there is a lack of research on using the deep neural networks (DNN) algorithm to predict ore production by considering multiple input variables such as truck haulage information and weather conditions.
Research Objective: This study aimed to build a DNN model for predicting ore production at open-pit mine sites using truck haulage information and weather conditions as training data.
Schematic diagrams of the fundamental structures of the best DNN model.
Scatterplots of the normalized actual ore production and normalized predicted ore production. The evaluation results from the (a) DNN model, (b) BPNN model, (c) BRNN model, and (d) QRNN model.
3 Fiber Bragg Grating-based Experimental and Numerical Investigations of CO2-Core Flooding Process
Prior to my Ph.D. program, my research topic was the geological storage of greenhouse gases (CO2). The specifics included using advanced fiber optic sensing technology to monitor the CO2 migration process in the reservoir core and the resulting strain responses accordingly. Because in CO2 geological storage, the mechanical stability of the reservoir and cap layer has always been a subject of concern, which is directly related to the effectiveness of storage and the risk of leakage. Before conducting the experiments, I led the design of the supercritical CO2 core flooding experimental platform, as shown in Figure. This platform mainly includes a gas injection unit, a heating and holding unit, a core holder unit, a pump pressure controller unit, a fiber optic sensing receiver unit, a back pressure control unit, and a data acquisition and analysis unit.
Research Gap: Previous studies focused on using fiber Bragg grating (FBG) to monitor the dynamic strain response and waterfront during water-core flooding experiments. However, there is a research gap in monitoring the supercritical CO2-core flooding process, which can potentially provide technique and theoretical support for real-site monitoring of CO2 storage.
Research Objective: The current study attempted to use the FBG to monitor, in real-time, the process of CO2 flooding in saturated sandstone.
Process mapping of the preparation of the specimen used in this study: (a) Two FBG sensors (each sensor with three axial gratings) axially pasted on the specimen surface; (b) The specimen coated with silicone; (c) 3D positions and distributions of FBG sensors.
Time differences of initial strain responses of the three gratings along the CH1 sensor at a constant effective confining pressure of 8 MPa. The relative strain responses induced by the gaseous CO2, liquid CO2, and supercritical CO2 are shown, respectively, in (a)–(b), (c)–(d), and (e)–(f).