3 présentations

• 
  10h30 - 10h55

  Applying Policy Gradient to the Adaptive Short-Term Mine Planning of Industrial Mining Complexes under Processing Plant Uncertainty

  • Joao Pedro De Carvalho, prés., Cosmo Lab - McGill University
  • Roussos Dimitrakopoulos, COSMO Stochastic Mine Planning Laboratory, Université McGill

  This work presents a policy gradient approach that allocates shovels to benches in the mines throughout an industrial mining complex, given uncertainties associated with geological attributes, equipment performance and processing plant responses. The method considers the interaction between the decision maker and the mining complex environment that simulates the performance of the processing plant. This considers high-order simulations of time-series-based processing plant attributes based on geometallurgical attributes. Additionally, a clustering algorithm provides feasible allocations where shovel equipment can be allocated. The material quantity that can be transported from the mining benches to the processor is forecasted based on discrete event simulation that considers equipment performance uncertainty. This provides experiences for the policy gradient method to train the agent to perform shovel allocation decisions in order to maximize the mining complex’s expected cash flow. A case study at a gold mining complex composed of two open pit mines highlights the benefits of assessing the uncertainty associated with processing plant responses and improves the cash flow forecast by 10% compared to the current practice in place.
• 
  10h55 - 11h20

  Simulation and optimization of rail operation in the mining industry

  • Nicolas Blais, prés., Université Laval

  Rail transportation is a key/bottle-neck operation in many mines' chains value. It is a complex environment because we need to manage train drivers, maintenance of the rail/trains/locomotives, and the fluidity of the rail. The current presentation discusses about a simulation model to help planners to take tactic and strategic decisions such as planning the maintenance, how to balance the driver in the networks, etc. Also, an optimization model has been made to help the planner daily by choosing when a train is going to depart and when to schedule the maintenance. We are going to talk about all the technical and non-technical challenges of the case study.
• 
  11h20 - 11h45

  Backpropagation-based Inference for spatial interpolation to estimate the Blastability Index in an open pit mine

  • Yakin Hajlaoui, prés., Polytechnique Montréal
  • Michel Gamache, Polytechnique Montréal
  • Richard Labib, École polytechnique de Montréal
  • Jean-François Plante, GERAD - HEC Montréal

  Blastability index (BI) is a measure that indicate the resistance of a rock to fragmentation when blasting. With novel
  technologies, miners are now able to collect and calculate BI while drilling. In this research, we study the capability
  to estimate the BI for non-drilled holes using only the spatial locations and observations of BI measurements in drilled
  holes. Since it is a spatial prediction problem, spatial interpolation techniques are investigated. This study proposes a
  back-propagation inference treatment combined with traditional geostatistical treatments like variography to enhance
  spatial interpolation models such as Gaussian Processes (GPs) and Inverse Distance Weighting (IDW). The proposed
  treatment improved GP and IDW and provided IDW variants that are very similar to a single-layer neural networks
  in so many aspects such as training, regularization, and prediction. Unlike traditional neural networks that are based
  on the assumption that data observations are independent, our models account for the spatial dependency structure
  of our data.
  Keywords: Machine learning, spatial interpolation, Gaussian Process, Inverse Distance Weighting, Back-propagation,
  neural networks.