Weighted ensembles of artificial neural networks based on Gaussian mixture modeling for truck productivity prediction at open‑pit mines - Mining, Metallurgy & Exploration (2023)

The truck haulage data from open-pit mine sites are usually massive and multidimensional with multi-peak Gaussian distributions. Artificial neural networks (ANNs) are well-known machine learning algorithms to handle massive and multidimensional data for building models. Moreover, Gaussian mixture modeling (GMM) is a suitable option for processing the data under multi-peak Gaussian distributions and improving model accuracy. For the first time, this study used a back propagation neural network (BPNN), an extreme learning machine (ELM), and a Bayesian regularized neural network (BRNN) coupled with GMM to deal with the complex truck haulage data and build three weighted ensemble models (WE-BPNN, WE-ELM, and WE-BRNN models) to predict truck productivity at mine sites. Decision tree (DT), random forest (RF), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost) were used to build models to be compared with the weighted ensemble models. The results showed that the WE-BRNN had a higher accuracy than the WE-BPNN and WE-ELM models. The proposed weighted ensemble models performed better than the benchmark models in predicting truck productivity, indicating that a weighted ensemble approach based on the GMM analysis significantly improved the model accuracy. Based on the relative importance analysis, haul distance was the most crucial input variable for predicting truck productivity. This study provides a new approach to predicting truck productivity, which will help mining companies make sound budget decisions and improve mine planning.