Deep neural network models for improving truck productivity prediction in openpit mines

Accurate prediction of truck productivity plays a pivotal role in improving the efficiency and profitability of openpit mining operations. This paper proposes a deep neural network (DNN) model to overcome the challenge of predicting truck productivity in openpit mines. The prediction model was built using eight variables and was optimized by considering different train-test split ratios, numbers of hidden layers and neurons, and activation functions. The proposed model’s performance was evaluated using various metrics and compared with other commonly used machine learning algorithms. The results showed that the proposed model outperformed traditional machine learning models by achieving higher prediction accuracy. Moreover, a single-variable sensitivity analysis showed that haul distance is the most influential variable for predicting truck productivity. This study marks a pioneering effort in employing DNN to predict truck productivity in openpit mining, signifying a notable advancement in the field. Background Accurate prediction of truck productivity, defined as payload per unit of time during each truck haulage cycle, is essential for enhancing efficiency and profitability in openpit mining, influencing decisions on fleet size, dispatch and route design [1]. Researchers have made efforts to develop prediction models for various mining operations using machine learning techniques [2]. Among these techniques, DNNs can achieve high prediction accuracy by addressing high-dimensional data and nonlinear input-output relationships, typical characteristics of truck haulage data [3]. Nevertheless, prior to this study, there were no existing studies in the literature demonstrating the application of DNNs for predicting truck productivity. Method and data Figure 1 shows the framework employed in the proposed study. Three main steps were included in the study: (1) data processing, including data cleaning and subsampling, (2) model construction, involving data preprocessing, model architecture determination, model training and model evaluation, and (3) model comparison, conducting a single-variable sensitivity analysis to better understand the effect of individual variables on the model’s performance [4]. The proposed model was trained on a massive dataset collected by truck sensors from oil sands mine sites. This dataset contained a total of 540,424 truck cycles. Blank rows resulting from recording errors were removed from the dataset, which was then subsampled to a small dataset (2,702 truck cycles). This dataset was divided into two subsets — a training set and a test set — using a proportional and random split. The splitting ratios used for the training and test sets were 80 percent-20 percent, 75 percent-25 percent and 70 percent- 30 percent. The prediction models were then built using multiple linear regression (MLR), support vector regression (SVR), decision tree (DT), random forest (RF), artificial neural network (ANN) and DNN. Hyperparameter tuning was further performed to avoid overfitting on the test set. Then, in the model evaluation step, the performance of the prediction models was evaluated using the test set and four metrics: mean-squared error (MSE), root-mean-square error (RMSE), mean absolute error (MAE) and R2 [2]. Finally, a single-variable sensitivity analysis was used to assess the relative effect of an input variable on truck productivity. Results and discussion Development of DNN models. Before hyperparameter tuning, the number of hidden layers ranged from two to five; the neurons were changed starting from 14 and increasing in steps of 14, up to a maximum of 84. Moreover, four activation functions were evaluated in predicting truck productivity, including hyperbolic tangent (tanh), exponential linear unit (ELU), rectified linear unit (ReLU) and scaled exponential linear unit (SELU). After predictive modeling, the best-performing DNN model configuration was found to have four hidden layers with 70 neurons per layer and SELU, resulting in the highest R2 and lower errors compared to other configurations. Fig. 1 Brief outline of the study. Fig. 2 Partial rank correlation coefficient for independent variables. Model comparison. In this study, six different machine learning algorithms were applied to predict truck productivity, and their performance was compared using four evaluation metrics: R2, MSE, RMSE and MAE. The algorithms were MLR, SVR, DT, RF, traditional ANN (with a single hidden layer) and DNN (with four hidden layers). Table 1 shows the performance comparison of these six prediction models. From Table 1, DNN, with the highest R2 value of 0.89, outperformed the other models and had a lower MSE of 0.11, a lower RMSE of 0.33 and a lower MAE of 0.22, indicating a strong fit for the data and the highest performance in predicting truck productivity. Sensitivity analysis. In this study, a single-variable sensitivity analysis was conducted using a partial rank correlation coefficient (PRCC) to assess the sensitivity of a dependent variable to an independent variable, as shown in Fig. 2. Haul distance, temperature and destination are found to have the most significant effects on the model output, with PRCC values of minus 0.55, minus 0.33 and 0.26, respectively. In contrast, the impacts of the month, empty speed, haul speed, day and shift are comparatively lower, with PRCC values of 0.19, minus 0.17, 0.08, 0.02 and 0.01, respectively. Conclusions The main findings can be summarized as follows. Firstly, the DNN model, configured with four hidden layers, 70 neurons per layer, and utilizing the SELU activation function, exhibited superior performance compared with other commonly used machine learning models. Secondly, the results of single-variable sensitivity analysis revealed negative correlations between truck productivity and haul distance, destination and temperature. In contrast, month, day, shift, haul speed and empty speed showed positive correlations. Haul distance emerged as the factor with the strongest negative correlation. This study breaks new ground by being the first to employ DNN for truck productivity prediction within the context of openpit mining, contributing to the advancement of the field. Selected references Choi Y, Nguyen H, Bui X-N, Nguyen-Thoi T (2022) Optimization of haulage-truck system performance for ore production in open-pit mines using big data and machine learning-based methods, Resources Policy 75:102522 Fan C, Zhang N, Jiang B, Liu WV (2023) Using deep neural networks coupled with principal component analysis for ore production forecasting at open-pit mines, J Rock Mech Geotechn Eng 16 (2023):727–740 Fan C, Zhang N, Jiang B, Liu WV (2022) Preprocessing large datasets using Gaussian mixture modelling to improve prediction accuracy of truck productivity at mine sites, Arch Mining Sci 67:661–680 Saltelli A (1999) Sensitivity analysis: Could better methods be used?, J Geophys Res: Atmos 104:3789-3793

Deep neural network models for improving truck productivity prediction in openpit mines

Authored by:
Chengkai Fan, Bei Jiang, Omer Faruk Ugurlu, Wei Victor Liu

To view this article you must be a member of SME/UCA.

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

To view this article you must be a member of SME/UCA.

Application of mining-width‑constrained openpit mine production scheduling problem to the medium‑term planning of Radomiro Tomic Mine: A case study

Environmental, social and governance considerations in production scheduling optimization for sublevel stoping mining operations: A review of relevant works and future directions

Related Articles

Comprehensive utilization of iron ore tailings: A review of sustainable practices and technologies

Implementation approaches for digital transformation: Current state of the mining industry

Estimating the extent of comminution needed for optimal downstream processing performance

Rare earth elements from primary, secondary and tertiary sources: Extraction technologies, resource distribution and sustainability

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Use Single Sign On