Machine Learning in Mining Conference: insights and innovative solutions for a ML mining future

Ankita Singh - Rio Tinto

The ML in Mining conference provided a deep dive into how machine learning is being applied in the resources sector.

Perth 10th November

A running theme through the conference was the openness of the speakers in sharing their learnings in a tangible way, their success, and the challenges that they faced in developing machine learning solutions for their organisation. Winners of the Speak at the Machine Learning in Mining Conference challenge were able to present their innovative projects, including:

Sanabel Abu Jwade - Rio Tinto

Brice Gower - Augment Technologies

Russell Menezes - RadiXplore

Ankita Singh - Rio Tinto 

The conference included deep technical presentations on use cases for ML, as well as thought provoking discussions on how we can scale machine learning - rethinking how we share data, how we productionise ML, what skills and operating systems we need and the ethical considerations.

Here's the overview of their projects:

Brice Gower  - CEO and Co-Founder, Augment Technologies:

TOPIC: Ore Movement Policy; meta-reinforcement learning, heuristics, and a physics engine.

In open-cut mining, blasting is used to move hundreds of thousands of cubic meters of rock every day, the fast pace and sheer volume of ore that needs to be moved makes it incredibly difficult to put the time and effort into accurately understanding how the forces of blasting reshape the insitu orebody; but this is extremely important as dilution and mixing can cause significant loss of the revenue-generating ore. Because of the complexity of the problem, acquiring supervised answers for how the ore moved to is not practical, and hence Augment went down the path of building a custom simulated environment to augment the data that was practical to measure and the resulting Reinforcement Learning policy is now delivering accurate models of how the chaotic blasting forces mix the different ore types together, all computed in under 5 minutes after the blast based on data that is practical to measure. 


Sanabel Abu Jwade - Graduate Software Engineer, Rio Tinto

TOPIC: Sign Recognition in Underground Mines - PoC. 

In the data-driven world that we live in today, mines strive to collect quality data to make informed optimization decisions in an increasingly competitive world. One example is the ability to optimize traffic flow in underground mines by collecting truck location data.

Typically, collecting truck locations in underground mines with the absence of GPS and Wifi is a challenge. Some mines rely on a manual process where the drivers log the location based on signs on the wall as they pass them. While manual logging provides useful data, it’s inefficient, distracting, and most of the time, not done promptly. 

To automate the process of logging truck locations, we investigated the viability of the use of computer vision and machine learning to automatically locate, identify and log wall signs in underground mines. The project considered environmental challenges in underground mines and integrated noise into the test data. The test data included signs with blur, occlusion, uneven lighting, and perspective transformations.

This work was done as a group project at UWA to provide a proof of concept for a Perth-based company specialised in developing mining software. The project investigated and compared two approaches utilizing a range of computer vision and machine learning techniques. The project recommended an approach that was able to identify signs correctly 100% of the time when at least 60% of the sign is visible to the camera. Moreover, the recommended approach was demonstrated to be robust against environmental noise and occlusions present in underground mines.


Russell Menezes - CEO, RadiXplore:

TOPIC: Use Machine Learning to Find Unrecognised Mineral Deposits within Legacy Text Reports.

70% of data in mining exploration is unstructured and exists in the form of PDFs, images and text files. Currently, only 2% of companies use this data and those that do barely scrape the surface, because it is time consuming and expensive to search through. For Western Australia alone, there are over 20 million pages of text-based information within the WAMEX exploration reports. This data is extremely valuable as it contains the subject matter expertise of over ten thousand individuals who contributed to them over the last 150 years.

How can we read and understand information from these reports like a geologist would, but at the same time go through all the 20 million pages accurately and quickly?

Computers can perform tasks millions of times faster than humans but are not smart enough to understand geology on its own. Can we teach the computer enough geology so that it could help us read through unstructured data (including scanned and handwritten reports) and find unrecognised deposits within it? 

Introducing Natural Language Processing (NLP) - the technology used to aid computers to understand the human language. The ultimate objective of NLP is to read, decipher, understand, and make sense of the human languages in a manner that is valuable.

In this talk we will introduce some NLP techniques and use it to derive insights from a scanned, handwritten image of a document. We will then explore some real world use cases where we apply these techniques at scale to see how explorers can use NLP to find unrecognised mineral deposits within the Western Australian WAMEX database which contains over 1 billion words.


Ankita Singh, Graduate Data Scientist, PACE Analytics | Rio Tinto

Through the lens of machine learning and texture analysis: What’s inside my Tiramisu?

We all love a delicious Tiramisu! It has those rich, creamy layers alternating with biscuits dipped in coffee. The earth’s subsurface resembles the structure of a Tiramisu from which the energy and resources sector extracts vital minerals, metals, and gas. Specifically, we drill and extract rock cores to accurately characterize, develop, and estimate reserves in the energy industry. These rocks are then imaged using X-rays with a vision to understand what lies beneath us. The x-ray images depict rock features as variations in the grey-level intensities, then transformed into segmented counterparts, i.e., labelled images. However, their characterization using labelled images remains a challenge. These challenges include (1) lack of structural information of rock geometry while creating labelled images and (2) operator-based segmentation techniques which cause rock descriptors to be biased. 
We address the first challenge using an unsupervised machine learning algorithm. This technique is applied to account for geometrical and structural differences in the rock features, especially pores and fractures, which are key for gas flow in the earth’s subsurface. The second challenge is addressed using a texture analysis technique. This method can be directly applied to x-ray images without the need for labelled images which are affected by user bias. 
While these techniques were first developed and used in the 1970s, they provide remarkable insights and approaches to alleviate long-standing challenges in the energy industry. Can we use the same techniques in the mining and metals sector to better understand our resources?