Mass Balancing Extended to Machine Learning

Mass Balancing Extended to Machine Learning

The methodology MIDAS Tech International has been developing has migrated from mass balancing to information theory to machine learning.

It appears these methods are the only ones available that allow unit processes to be modelled in detail; entirely from the data.
This is ideal for a simulation system that can be used for both data-driven models and imposed models (i.e. if a new unit is to be added to the circuit; and therefore there is no plant data).

However in summary: here are key concepts:

  1. Mass balancing is generally applied via least squares minimisation.
  2. A different approach is to use information theory.
  3. If one does use information theory, one can also infer missing variables – in particular go to a higher level of depth of understanding the ore. (multimineral particle distributions)
  4. A concept to improve the uncertainty of the missing variables is to use ore variability to advantage (patented method). The basic idea is to consider changes in ore processing for different ores to identify how units are behaving (with respect to the ore).
  5. If one knows how the units are processing the ore; then it is only necessary to model these changes if operating conditions change.
  6. The information theory is consistent with logistic regression; so it is possible to estimate unit changes due to changed operating conditions if changes operating conditions are also recorded. Hence this approach is data-driven; which is similar to machine learning approaches.
  7. The method is therefore most suitable to methods involving regulator plant audits.

Link is:
MIDAS Tech International gave the course as well this year, but it was obvious where I was now working at was so different to status quo, people were getting confused. MIDAS Tech International also had the issue that the participants were more diverse. Some had limited software skills and had never seen a simulator; and others were Professors.

The course was slip into ‘Fundamentals’ and ‘Advanced’ and I gave this at Rio Tinto.  The idea was to discuss non-contentious issues in Fundamentals, and in Advanced discuss concepts that were beyond status quo. (i.e. much further than current mainstream research )

There were about 20 who attended; and in general all attended both courses (two weeks/ 3 days per week)

MIDAS Tech International has developed an Exercise for participants to deduce the density partition curve for a dense medium cyclone given only the flows (within each size) of the products and the full density distributions of two feeds.

In the LV presentation MIDAS Tech International provide a more complex variation with multi-mineral particles.

As a teacher there is nothing more satisfying than when a participant is able to deduce for themselves a key concept. For example, I provide lead-in on how to deal with the multi-mineral liberation comminution model.