Intelligent system for increasing the efficiency of iron ore processing
Robots from Redmadrobot Data Lab have trained artificial intelligence to economize resources. Our experts spent several months setting up algorithms that are now analyzing the grain-size composition of iron ore on the conveyor belt at the Sokolov-Sarybai Mining Production Association (SSGPO) in Kazakhstan. With this data, engineers are able to fine-tune the work of the mill that processes raw materials. The intelligent system operates using computer vision and neural networks.
Rod mill settings determine the cost of ore processing.
The longer the processing period, the more energy and water you need. Large, heavier rock pieces on the conveyor belt prolong processing time. To deal with this problem in the past, SSGPO operators would analyze the grain size twice a day manually, like everyone else in the industry.
But there weren’t enough of these manual checks to produce the results needed.
«The business unit tasked us with developing an optimization model based on six key rod mill performance factors including grain-size composition. To solve this problem, we hired the Redmadrobot team, who did a great job»
Dmitry Karbasov,ERG, Head of Industrial AI Department
Redmadrobot and ERG considered many project implementation options, including the use of ultrasound or x-ray sensors. In the end, the engineers settled on using industrial cameras and computer vision algorithms. The system works like this: the AI analyzes the ore on the conveyor, and production engineers configure the rod mill. The system uses cameras to 'look' at the ore, and its algorithms process the images to transform the data into metrics and APIs for mill settings adjustment.
We looked into a number of options to solve our problem. Solutions based on computer vision proved to be the best fit, since other options required either modernizing the entire production line or buying expensive equipment. To estimate the grain-size composition of the lower layers of raw material, we needed to install predictive models. Machine learning is often used for processes involving computer vision and predictive model training.
Our first step was to prepare very high-quality images for algorithm training. Because the speed of the conveyor belt is about two meters per second, we used bright lights and Basler cameras with a shutter speed of 1/2000.
Machine learning is resource-greedy, and computer vision requires servers with video cards. We brought a server right into the plant so that it could control the cameras. At this stage, we had up to 170,000 ore images of the proper quality.
To teach the algorithm to distinguish large rock pieces from the background, we started mapping the images. Each ore fragment larger than 16mm in diameter is considered large; these fragments had to be detected and marked. Sand and smaller fragments were set aside. We fed the information to the algorithm and watched it learn the lesson. Then we tested. At this stage, its ability to detect large rock pieces was at 80 percent. We used ERG data on the correlation between the size and weight of ore fragments to train a predictive model that transformed the data received from computer vision algorithms into data on weight and grain size of the entire raw material layer, from top to bottom. This information helped calculate the overall grain-size composition of the ore and configure the mill settings accordingly. All the data is updated in real time and displayed on a dashboard.
In August 2019, we launched a pilot version of the system on one of the conveyors. The system was able to determine ore composition with a precision of 98 percent. In 2020 ERG plans to roll out the technology in all the facilities with this kind of production process.
Right now the intelligent system determines grain-size composition and contributes to more efficient resource outlay on one conveyor. But soon it will let SSGPO cut production costs, increase ore processing efficiency and labor productivity, and lessen the impact of the human factor on production output. We have recently released the first operational version of the system. Once the technology is fine-tuned, it will be installed and used at other production facilities.
The project made it to the finals of the main industrial award in the CIS, the 2019 OEE Awards, held under the auspices of the Efficient Production 4.0 conference (Moscow, Dec 5-6, 2019), and received an award for its nomination in the category of AI in the Manufacturing Industries.
«In 2020 we expect to see an additional end-product output of up to 200,000 tons a year with a production cost decrease of up to 5 percent»