Twenty Second Annual Meeting 20192020 Grade control with
- Slides: 19
Twenty Second Annual Meeting: 2019/2020 Grade control with ensembled ML: A comparative case study at Carmen de Andacollo copper mine Camilla Zacché da Silva Jed Nisenson Jeff Boisvert
Presentation outline • Context • Proposed workflow • Dataset • Case Study • Results • Conclusions 2
Context • Grade control is a procedure that provide selectivity for the excavation of ore/waste types with the final goal of maximizing profit or reducing loss in the mining operation (Verly, 2005). Ordinary kriging Inverse distance Nearest neighbors GOAL: This study aims to use of machine learning algorithms to enhance accuracy on the predicted model and, as consequence enhance ore/waste decision making, maximizing the profit of the mining project. 3
Proposed workflow for the grade control model Step 1: collect all data available 4 Step 2: train ML algorithm on available data Step 3: generate spatial grade model for variables of interest Step 4: apply economical cutoff grade to determine material destination Step 5: generate minable plan using available dig limit algorithms
Case study • 5 estimated grade models • Two of which use machine learning: • collocated cokriging with ERBFN model as auxiliary information; • collocated cokriging with SVR model as auxiliary information. • Inverse distance (power=2) • Ordinary kriging • Intelligent grade control • Apply economical cutoff grade to determine material destination; • Based on destination, generate minable plan (using available algorithms) 5
Data set • 10 blast areas • The cutoff considered is 0. 2% Blast Number 1 2 3 4 5 6 7 8 9 10 6 Number of samples 133 106 115 123 170 143 96 117 120 124 Mean 0. 31 0. 26 0. 50 0. 41 0. 21 0. 22 0. 24 0. 27 Std. Deviation 0. 17 0. 08 0. 2 0. 14 0. 11 0. 24 0. 15 0. 45 CV Maximum Minimum 0. 54 0. 3 0. 40 0. 34 0. 28 0. 52 0. 47 1. 08 0. 61 1. 64 1. 06 0. 53 0. 94 0. 74 0. 81 1. 04 0. 86 1. 54 1. 09 4. 76 0. 03 0. 05 0. 13 0. 15 0. 07 0. 02 0. 07 0. 04
Estimated grade models • Inverse distance • Ordinary kriging • Intelligent Grade Control • Unsupervised spatial prediction algorithm, based on local simulation (Vasylchuk and Deutsch, 2018; 2019) 7
ERBFN training • ERBFN (Samson, 2019) as auxiliary information • Parameter setting: • Learning rate: 0. 01 • Number of hidden units: 0. 3, 0. 35, 0. 40 and 0. 45 number of data samples. 8 Data Blast 1 Blast 2 Blast 3 Blast 4 Blast 5 Blast 6 Blast 7 Blast 8 Blast 9 Blast 10 correlation 0. 89 0. 93 0. 94 0. 6 0. 7 0. 84 0. 6 0. 7
Stacked SVR training • Data correlation 9 Blast 1 Blast 2 Blast 3 Blast 4 Blast 5 Blast 6 Blast 7 Blast 8 Blast 9 Blast 10 0. 63 0. 74 0. 86 0. 76 0. 96 0. 57 0. 78 0. 77 0. 52 0. 62
Trends 10
Estimated models MSER ID Average OK 0. 0088 0. 0078 SVR + Ccok 0. 0070 ERBFN + Ccok IGC 0. 0075 0. 0079 0. 014 Avergae 5 -Fold MSER 0. 012 0. 01 Average 0. 008 0. 006 0. 004 0. 002 1 ID 11 2 OK 3 CCOK+SVR 4 CCOK+ERBFN 5 IGC
Material Destination • Cutoff is 0. 2% of Total Copper. • Given the estimated models the destination is determines for the material in each blast. Optimal minable dig limits • Dig limits are obtained through an automated algorithm (Vasylchuk, 2018). Reference models • One realization is randomly selected from a set of 50 for each blast area. Algorithm: Sequential Gaussian Simulation 12
True Ore/waste Destination – Blast #4 94% of ore and 6% of waste TRUE Waste 13 Ore
Dig Limits § Dig delimitation based on the ore/waste decision. The dig limits were processed through an automated algorithm. TRUE Waste 14 Ore
Ore incorrectly mined over 10 blasts Number of blocks per blast incorrectly mined as waste relative to true values TRUE Average ID 196 OK 200 SVR 156 ERBFN 181 IGC 247 400 Total of blocks incorrectly mined 350 300 250 200 150 100 15 1 ID OK 2 3 CCok +SVR 4 CCok +ERBFN IGC 5
Waste incorrectly mined over 10 blasts Number of blocks per blast incorrectly mined as ore relative to true values True ID OK SVR ERBFN IGC Average 254 218 257 214 171 Total of blocks incorrectly mined 400 350 300 250 200 150 100 16 ID 1 2 OK CCok +SVR 3 CCok +ERBFN 4 IGC 5
Total of blocks misclassified 4600. 0 4500. 0 4400. 0 Total of blocks 4300. 0 4200. 0 4100. 0 4000. 0 3900. 0 3800. 0 3700. 0 3600. 0 17 ID OK CCok+SVR CCok+ERBFN IGC
Conclusions • The use of trend models obtained through ML algorithms, have in both cases enhanced classification of the extracted material; • Even though the models did not have parameter set as problem specific, the trend models still demonstrate gain over the other methodologies; • In ML hypermeters can challenging, but in this framework the setting is straightforward. • ERBFN in this case study outperformed SVR 18
This presentation and the associated papers and software for the sole use of CCG members. Selected papers from this report may be distributed to non-members for the purpose of promoting CCG research. The report may be circulated within CCG member organizations; however, the following copyright notice must be adhered to. Copyright, 2020, Centre for Computational Geostatistics All rights reserved. No part of this presentation may be used or reproduced without written permission, except for members of the Centre for Computational Geostatistics.
- Cinderella dressed in yellow rhyme
- Informs annual meeting
- Aupha annual meeting
- American psychiatric association annual meeting 2020
- Grand lodge of kentucky
- Annual theory meeting
- Cwemf
- How to run an annual general meeting
- Nrg oncology meeting
- Aashto annual meeting 2015
- Scts annual meeting
- Nrg oncology meeting 2017
- American epilepsy society annual meeting 2017
- Grade 11 annual teaching plan
- Ems grade 8 term 1 assignment
- 27 miles per gallon into kilometers per liter
- For today's meeting
- Today meeting or today's meeting
- What is meeting and types of meeting
- What is meeting and types of meeting