Machine Learning Case Study: Mining - Hari Menon of Symphony Industrial AI - ARC Industry Forum

Thank You Ellen before I get into the specific case study here I'm representing I'm doing this on behalf of one of our customers who implemented a optimization nai driven a machine-learning driven optimization system and one of the largest gold mining companies in the world and one of the largest coal mines in Nevada so I'll go through that case study to give you a flavor of how machine learning is being applied and can be applied in the real world so before I start I know I know Allen went through quite a few of previews slides on what machine learning is what supervised and unsupervised learning and all that I want to bring it down to a very simple level so at the end of it it is all math and stats so I know how many of you are familiar with type 1 and type 2 errors and false positives and false negatives from your engineering school and or even business school days right so there's a an interesting anecdote here so there is a discussion around machine learning algorithms to detect terrorists among people who enter an airport and if I tagged every person as not a terrorist the accuracy of that algorithm is 99.9999% now I cannot sell that to anybody as a solution because let's say it is accurate having said that in machine learning what is important is what we call as precision and recall recall and precision are the things that we worry about in choosing algorithms or choosing what techniques you use in whether it is optimization or anomaly detection so when I go through this keep that in mind the the big challenge in what I've seen in machine learning so far as a practitioner as an implementer is obviously data is a challenge but assume you have the data then it is tying the domain to the sari tying the signs to the domain problem so you need people who understand at least a little bit of the science and enough of the domain to be able to apply this problem and the third one is what I what I called as a precision versus recall or precision and recall problem and you can pick algorithms that give you a lot of will be high precision but the recall might be low and I'll define what we call in precision is it is before the notion of false positives and false negatives and you can get into that so as I go through this I have about twenty minutes is that oh okay so let me start so this particular use case is someone of using machine learning to do a milling circuit a grinding circuit in in a mine in AG or mine and optimizing the grinding circuit so so the gold production and throughput so I'll define what a what a milling so what a mining complete operations look like essentially is one of the largest like I said one of the largest mine mines in North America and one of the top five mines in the world which produce about a million ounces of gold so you can do the math based on gold prices what the revenues coming out of that mine is etcetera and you have to process about 6,000 to 7,000 tons of ore to generate an ounce of gold so think about it you're processing huge amount of material through a complete set of operational processes to get a few ounces of gold at the end any any kind of optimization you can do in throughput can have significant impact on your downstream revenue one of the these equipment and these operations are have been around for a long time so these are equipment don't get replaced very often these are when I get into the the type of equipment here you will you'll get a sense of that so the to stay the situation was to improve the throughput and yield from this mining operations now there's a lot of dependencies on how much throughput that you can drive through these my these mining operations it's depending on your quality the granularity of the or the the temperatures the there's a whole bunch of file when I walk through the variables it is play that 3040 it's a variable set you have to consider in optimizing the operations the traditional approach was using advanced process control with a my model MPC model predictive controller to do settings on these milling operations what we the company was able to do with machine learning words to increase the throughput by a percentage and that's pretty big deal it's in the tens of millions in terms of revenue for them so this kind of a bird's-eye view of the process you have these are open pit mines you bring in the or crush them create a pile and then there is a grinding circuit the grinding circuit has two big operation says sag mill and a ball mill and a sack mill is has also a lot of large metal balls and it rotates and then a combination of the balls crushing against your or and the gravity is what kind of crusher from about nine inches eight to nine inches in diameter all the way to 150 microns through these two steps so you have the sack mil and then the order is then taken into the ball mill and it is purely crushed with the the balls rotating in that now as I said the traditional operations of you have operators who are looking at screens you have a control center with operators looking at the screens and they are looking at conditions and adjusting the thresholds of how much how much to put can how many tonnage can go through this and the feed is fed in based on what the operators control now the model predictive controller is a one which is adjusting the variables in there whether it is the motor current or the changeable variables in this thing so the opportunity here is to maximize the mill throughput and the challenges are the variability in or properties the process uncertainties and the operator to operator variability now these operators have been working at the mines for decades some even 50 years started when they were right out of high school or even Community College and were trained to operate these machines and then over a period of time they have kind of they have back off the envelope approaches they have things that they have learned learnt over time so they make adjustments to these variables based on what they know they've known from the experience as well as looking at some of the variables on the screen now as you know it's very hard for a human being how are smart they are to look at 20 or 30 variables know the interactions between them know what is optimal when you change three of them in a certain direction etcetera etcetera so so it's not a it's not something an operator can do so so they use some heuristic to edges and what the company will had noticed and we had noticed was that they were always operating in a safe zone so think about a mill like your washing machine if you want to get the maximum throughput out of your washing machine you could load it with as much clothes as possible but after a certain threshold the machine either breaks your clothes are not clean and all of that and the milling is a similar process you can increase the throughput you can increase the amount of feed that goes through but after a certain stage depending on the rest of the variables it kind of clogs it chokes all of that so to be so traditionally the operators have been operating it in very safe zones because they have only so much ability to know how much they can push it so they they take the consumer conservative approach what machine learning was able to do is so the good news is like the advanced process controller the NPC worked in the sense that it reduces reduce the process variability now but it reduces a process variability but it still operated within very safe zones what the machine learning algorithm and the continuous learning was able to do is push the limits of how much feed can be pushed through the link to the right or increase the amount of feed that can that can go through this process and this kind of a a kind of a block diagram view of what how it works the the traditional approach the APC objective was using a certain constraint maximize the throughput and then the the APC worked to set the variables to maximize the throughput but the challenges were multiple interacting constraints across the middle circuit not at each individual mill but across the middle circuit there were interaction can can strange and then you had reactive management operators had to react to the changes in our properties now there are other opportunities in this mind like the or properties were determined based on a lab test so there was a lag in when the the lab test results came out now you could use video camera based approaches we are thinking of a whole bunch of approaches to improve how those or properties can be characterized even without a lab test so you can get ahead of the curve here so APC op operated or automated the optimizer the limits were set by the operator so the operator would set the limits and then let the APC operate within those limits so that's how this was being run what the machine learning vai driven APC was AI would provide the guidance your prices to set the min and Max because now you're providing an advisory you're not directly controlling it you're providing an advisory today say what's a min and Max and that naturally push the min and Max to the right-hand side and increase the throughput so this is kind of the a cognitive learning based approach we used an algorithm called a LS TM auto encoder based algorithm which can take in a large number of variables and do a predictive modeling of the predictive optimization of the this circuit the the goal here from a from a change management perspective the idea here was to provide the data the advisory and the corresponding data visualization because one of the things you have to be those of you probably already know about this because you're in the middle of your machine learning projects is you don't want to put a black box algorithm and say okay these are your predictive results and then first of all you're not going to get acceptance out of very experienced practitioners and the second thing is they wouldn't know why it is recommending a certain thing so the explain ability when you select algorithms one of the things is for industrial applications the explained ability of those algorithms need to be there now those algorithms can be explained through data visualization through other approaches as a kind of text-based explanations etcetera and we approach it to data visualization so when it shows a certain results it will show which variables have moved and which have moved in certain direction and what the envelope is so that over a period of time the operators have a good feel for okay why is it recommending that I push the limits forward so I think there's a lot of technical details here around what was done in kind of setting the I think I have a video here but I don't think I'm able to play that video the it essentially dynamically adjusts constraints and then let the model predictive controller to run the run tomorrow yeah it is so essentially the ability to kind of move the throughput to a higher level because if you see on the left hand side when they were operating at a very safe zone and as soon as they ran into an issue they will lower the threshold so that the throughput drops drastically now the algorithm was able to tell them that there is no need to lower this or you could even keep it higher because this was a temporary temporary issue and in effect drove the throughput to a much higher value so a little bit about the Algar the approach you essentially are creating a I think digital twin is used very kind of vary widely these days a dynamical dynamic lead constructed digital twin right and we use a recurrent neural network and be able to optimize see the production optimize the throughput on this and there's a lot of literature published around LS TM auto encoder models and recurrent neural nets that that's available online so the annual revenue impact of this was about five million for mine and the reduction in operator induced variability you had an advisory system that allowed you to kind of rely on to optimize the production yeah I don't have a slide on the challenges I can go through some of them so data like I said data integration was a challenge we had to start with what was already available through historians a lot of data being collected in data historians for other reasons for analytics and all of that so the machine learning models used time series data that was collected in historians additional sensors had to be put in at a later stage in the project to get the model much much more kind of fine-tuned the second challenge like I said was you cannot do this from a lab you have to really go into the mind and sit with the operators and walk them through the results and educate them on this thing so there is a change management in any of these machine learning approaches if you want really implement matching learning in factory floors of mines or oil fields or all of that you have to take a you have to account for change management that is needed for folks you adopt this solution then it is another one is sustaining the solution any machine learning model depending on how drastic the conditions vary can have something called model crust and you have to have some sort of a support mechanism to say okay the model has drifted we need to relearn from the new set of data historical data but the new set of data and we tweak the the models depending on what kind of models you are using so all of them have to be taken into consideration from if you're implementing a solution like this so that's that's all I had [Applause]

Channel: ARC Advisory Group

Views: 843

Rating: 5 out of 5

Keywords: ARC Advisory, Technology Evaluation, Supplier Selection, Digital Transformation, Digitalization, Industry Manufacturing, Market Research, Market Studies, ARC Industry Forum, Orlando Forum, Connected Machines, M2M, Industrial IoT, Internet of Things, Mining, Hari Menon, Symphony Industrial AI, Machine Learning, Artificial Intelligence

Id: dmCZg_p80Cs

Length: 18min 31sec (1111 seconds)

Published: Sun May 19 2019