Bubble tracking system in industrial mixture
Computer vision
Backstory
EORA loves hackathons: they help us keep in shape and create new things. At the hackathon, many of the EORA's projects were invented, and then finalized within the framework of business tasks. This project is one of them.
Customer
One of the largest mining and metallurgical companies in Russia.
Task
Create a service that allows to evaluate the characteristics of industrial foam during flotation process. The data is transmitted to the operator in a convenient form.
What is flotation
This is the extraction of non-ferrous metals (e.g. copper, nickel) from the rock. The rock is crushed, mixed with water and reagents and foamed. The rock settles, and non-ferrous metals remain in the foam.
Solution
Bubble tracking program based on computer vision technology.
Information panel (generation and display)
Data analysis inside the program
Computer (running the program)
Video camera (receiving video stream)
How it works
foam color
the velocity of bubble formation
frequency of bubble disappearance
number of bubbles
average size of bubbles
direction and speed of foam movement
size of bubbles (small, medium, large)
Knowing the location and direction of movement of bubbles, we can determine more than 10 characteristics of industrial foam:
Technologies
classical methods of computer vision (CV)
классические методы компьютерного зрения (CV)
advanced methods of deep learning (DL)
Development process
классические методы компьютерного зрения (CV)
Step 01
Step 04
Data markup
Computer model training
Step 02
Step 05
Selection of the best methods for solving the task
Testing and debugging the solution
Step 03
Interface design
Development time
классические методы компьютерного зрения (CV)
April 16-18, 2021
3 days
interesting fact
We took the first place in our track. In total, there were 23 teams from 65 regions at the hackathon
reduce reagent and water costs
increase the metal recovery rate
Advantage of the service
Using a simple linear algorithm, we can predict the trajectory of the foam flow for a few seconds ahead. The prediction accuracy is very high.
Before developing, we studied similar products that exist on the market. None of them reads as many foam parameters as ours.
This allows to:
ML engineer
Sergey Solovyov
Designer, fullstack developer
Valery Shlyapnikov
CV/DS engineer
Nikita Buzanov
CV/DS engineer
Andrey Ragimov
Project team
Results
The hackathon organizer successfully uses an analog of our service in production
The EORA program gives an increase in the extraction of useful components by 0.3-0.4% (that's a lot)
The effectiveness of the training will increase significantly if synthetic data is generated and a neural network is trained on them.
Individual tracking of each foam bubble is possible
You can significantly speed up the service due to high-quality video processing
If you install depth cameras, you can determine the height of the foam without sensors
