Mines are being exhausted and ore grades for many minerals are falling around the world.
One way to manage this is to use technology that makes the most of existing ores. Optimising the sorting process helps to accomplish this.
Challenges in ore sorting
Traditional ore sorting, while a crucial part of mining, has its drawbacks.
Detection of low-grade material and subtle mineralisation can be challenging. Real-time adaptation to fine-tune the balance between grade and recovery is often not possible, resulting in lower concentrate quality and more gangue entering the product stream.
Tomra's solution
As AI, ML and deep learning have entered the picture, so have opportunities for more flexible and precise sorting methods.
Sensor-based sorting specialist Tomra Mining has developed an AI-powered ore sorting technology called Contain, designed to enhance the recovery of inclusion-type ores that are hard to detect with traditional methods.
“There are limits to what humans can do, and machines now complement our abilities quite effectively,” says Tomra software team lead Stefan Jürgensen.
Contain uses convolutional neural networks to analyse x-ray imagery in real time, classifying rocks based on the probability of subsurface ore mineral inclusions.
Field trials at Austrian mining company Wolfram Bergbau showed that when Contain was integrated with Tomra’s existing sorting technologies COM XRT and Obtain, customers could either boost plant throughput by 8% with no loss in recovery or cut ore losses by 33% while maintaining throughput. This is possible through Contain’s improved detection accuracy, which allows operators to fine-tune the balance between recovery and throughput based on their priorities.
The technology works for a wide spectrum of ore grades and complex mineralisations that traditionally result in high misclassification or excessive product loss, such as in tungsten, nickel and tin ores.
Jürgensen likens using Contain to a human carefully examining all sides of a rock and taking the time to compare it to others. “Our machine classifies 1000s of rocks per second, but on the same quality level. That’s the difference,” he says.
Training the system
The system has been trained on tens of thousands of inclusion-type ore samples, like tungsten, nickel and tin, which form distinct patterns on x-rays. Tomra is currently testing on further, similar inclusion-type commodities and is exploring expansion into other minerals.
Training with different types of rock allows the system to adapt to different deposits, creating complex, deep neural networks to accommodate diverse environments.
“Humans are very good at detecting patterns everywhere in the world. But the machine has no clue about the world, and therefore we have to present more of the world to the system,” says Jürgensen.
“But at the same time, because it’s able to understand and analyse so many more samples, it’s actually able to see some things that we miss, because we can’t really take the images of 10,000 rocks into our head.”
What does it mean?
Contain helps high-volume processing plants to maximise concentrate grade, minimise valuable material loss and accommodate cost constraints.
As the mining industry evolves, deep learning can preserve both mineral quantity and quality.
For more information, visit: www.tomra.com
