Machine-Learning Analysis of the Canadian Royalties Grinding Circuit

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

I have spotted a few typographical errors and my annotations in the attached manuscript will help.

I have also included some comments on Figure 1 and 2; both require improvement to get better clarity of what is being conveyed.

My main overall concern however, is that the authors achieve correlation by manipulating 27 parameters without relating these to any scientific theory. With so many parameters to play with it is possible to fit any data even for the wrong reasons. At the very least the authors  also discuss how the correlations relate to know theory. Especially for the parameters of significance. They need to discuss the important theoretical factors that affect P₈₀ and then show how the modal behavior compares with theoretical expectations.

It will also be useful to see how well the model predicts other outputs apart from just the P₈₀. Are the solids density of the streams correct as well. We can only half confidence in the model's prediction of the P₈₀ if intermediate data is also correct.

To summarize, it is not enough that model worked successfully but can only contribute to our knowledge if the trends are clearly related to scientific theory. The other alternative is to also use a circuit simulator and compare with the machine learning model.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The English is generally okay, except for a few mistakes.

Author Response

The PCA is not about fitting data.  It is used to find trends in the data.  Also the principal components are not correlated.  PCA is not used to predict output; it is an unsupervised technique.

I was told by Canadian Royalties Inc personnel that the %solids of the streams is correct.

I have made changes to Figures 1 and 2.

The number of parameters was reduced to 16.  I have made links to how some of the parameters are linked to theory.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The authors seemed to have forgotten a very basic principle in mineral processing:  understanding the mineralogy of the system is paramount in understanding much of what happens in a mineral processing plant.  The mineralogy yields information and the hardness of the ore, the liberation size to grind to, the potential grades and recoveries that can be achieved and the pulp chemistry of the system.  In this case, the authors are trying to maintain a consistent P₈₀ for the flotation feed.

In the first instance, crushing and grinding are about the gradual size reduction of the ore in order to liberate the valuable minerals for the gangue.  Primary grinding is about liberating the gangue, while regrinding is about liberating the valuable minerals from the sulphide gangue.  The second point is that while plants usually specify a target P₈₀, the reality is that the P₈₀ will vary as the ore entering the plant varies.  That is, the ore is heterogeneous with changes in mineral content, hardness, grain size and locking characteristics with time.  Therefore, it is unsurprising that the P₈₀ will vary.

As the mineralogy of the feed is not measured the next best parameter to examine is the elemental composition of the feed, in particular the sulphur feed grade.  It is highly likely that as the sulphur feed grade increases (i.e. the percentage of nickel sulphide, copper sulphide and pyrrhotite) the percentage of gangue minerals decreases, which suggests that the ore is potentially softer and the P₈₀ will become finer.  Conversely, the lower the sulphur feed grade the more gangue there is int he feed and the ore will become harder, which means the P₈₀ is likely to coarsen.

If the throughput is maintained at a constant rate, these subtle variations in hardness will manifest themselves are changes in circulating load in the crushing and grinding circuits.  Harder ore, for a constant feed rate will see an increase in the circulating load of the cone crusher and around the primary and secondary ball mills.  As this load moves up and down the P₈₀ will increase and decrease.  Changes in circulating load in the crusher and grinding circuit will manifest themselves as changes in the amps pulled by the cone crusher and variations in the bearing pressure of the primary and secondary ball mills.

Some of the parameters listed are probably redundant or do not provide any useful information, for example the primary crusher amps, and the cyclone feed percent solids.

The ore silo level is an interesting parameter.  Simply put, its effect will depend on how the ore passes through the silo.  Broadly speaking, as ore enters the silo the coarse particles will tend to roll to the outer edge of the silo, while the finer particles will tend to stay in the centre of the silo.  If the ore is pulled from the centre of the silo it is likely to be made up of finer material, but as the level of the silo drops more coarse rock will be pulled into the ball mill feed.  And, as the silo level continues to drop the proportion of coarse material will increase.  It is probably reasonable to assume that the coarse rock is going to be harder (i.e. lower sulphide grade), so coarsen the ball mill feed will have two effects:  the feed size has increased as has the ore hardness.  Both will result in a increase in the circulating load and a coarsening of the P₈₀.

So, I think the authors need to reconsider their approach, and think about what changes in the mineralogy do to the process.  From this determine which parameters are redundant as they show the same information as other parameters . . . that is, rather than using all the parameters only use these that are directly relevant to the process.  For example, if the primary ball mill power increases it is a reasonable indication that the mill is filling up with ore (i.e. higher circulating load), and it is highly likely that the secondary mill power will also increase shortly thereafter.  The two are not independent, so you only need to use one of them in the analysis.

I would also suggest that the authors examine their raw data.  The data should be cleansed prior to completing their analysis.  That is, data sets with critical pieces of information should be removed.  Also, data sets that have less than 80 percent plant availability or an interruption in normal operation should also be removed.  For example, maintenance shutdown is not normal operation, and this data should be removed from the data set.

It might be useful for the authors to plot specific set of data as either time series or cusum plots to understand how they interact.  For example, primary mill power, fresh feed, sulphur grade and silo level.  From these observations a more reasoned approach might be possible.

In setting up machine learning algorithms it is necessary to understand the data set and how the various parameters fit together.

Finally, there are several typos in the manuscript that need to be attended to.

Comments on the Quality of English Language

The English is of a reasonable standard.

Some of the description is a little laborious, but I think this is related to the authors not narrowing down the parameters that are important.

Author Response

The mineralogy was not available.  However, we used head assays to infer changes in the ore/mineralogy. The correlation between hardness tests (Bond Work Index) and head assays was done.  Do the head assays impact hardness?  This will indirectly tell us whether changes in mineralogy occurred.

The elemental composition (nickel, copper, cobalt, iron and sulphur) of the feed was considered in the analysis (principal component analysis).  These were also correlated with Bond Work Index measurements to determine whether the ore got harder or softer.

A recommendation was made to Canadian Royalties concentrator to measure the circulating load in the grinding circuit.

The effect of silo height was discussed in the article.  

The total number of variables was reduced to 16.  Also, the head assays and hardness were used to indirectly determine the effects on P80.  Unfortunately, the mineralogy was not available.

The raw data was cleaned to ensure that the shutdowns were excluded for the data used in the analysis.

Plotting graphs of two variables resulted in low R-squared values.  Thus, did not reveal any significant trends.  As a result these were not included in the manuscript.  This shows the power of principal component analysis.

 

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All the outstanding issues have been more or less addressed, just 3 issues I would like to point out:

 

Page 9 line 260, replace le with the

Page 14 line 343 sentence not making sense, cross out "developed".

Please provide at least one reference to support your bin segregation hypothesis.

 

Comments on the Quality of English Language

English okay, apart from the minor typos that have been highlighted for the authors to address

Author Response

Page 9 line 260 Correction was made

Page 14 line 343 Correction made

Please provide at least one reference to support your bin segragation hypothesis  I could find one reference only.  This was somewhat related to silo height heterogeneity, but the work was done for wheat particles.  The concept is somewhat related to mineral processing.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

It is a pity that the authors don't have any mineralogy, however all is not lost.  A discussion with the geologists will allow the authors to determine if the main minerals were chalcopyrite, pentlandite and pyrrhotite.  If this is the case it should be possible (using a few assumptions) to convert the elemental assays to minerals (e.g. 2.888 x Cu assay = % chalcopyrite, and the sulphur associated with chalcopyrite is this number multiplied by 0.3494).  The authors can then estimate the chalcopyrite and pentlandite grades, then subtract the sulphur associated with these two minerals from the total sulphur and assume the remainder is present as pyrrhotite.  It is highly likely that cobalt is present in ppm levels and is not really going to impact what happens in the mill.  Once the elements have been converted to minerals, they can be subtracted from 100 to yield a non-sulphide gangue assay.  Why do this?  Each of these minerals tends to have a different hardness.  Pyrrhotite tends to be slightly harder than chalcopyrite and pentlandite, and the non-sulphide gangue is harder again.  Further, plotting sulphur does not adequately show the split between the different sulphide minerals.

It is also likely that when the silo level is low the percentage of non-sulphide gangue will increase.  If this happens the ore is harder, the circulating load will increase and the particle size distribution should coarsen.  Perhaps plotting grade against silo level might reveal something interesting.

Perhaps the biggest influence on P₈₀ is the cyclone feed pressure, if all other parameters are kept the same.  Generally, the ball mill liner, ball charge percent solids are "fixed" and do not vary by very much.  The cyclone configuration (inlet diameter, vertex finder diameter, spigot diameter, cone angle) is also fixed.  So, what is left to impact the particle size distribution?

The ore will.  If the ore becomes harder the circulating load will increase, and eventually this will result in a coarsening of the grind.  If the ore is softer the load will decrease and the P₈₀ will become finer.  So, the composition of the ore and the size of the rocks will have an influence.

However, if the ore is stable, the only thing that impacts the particle size distribution of the cyclone overflow is the feed pressure.  This is achieved by closing a cyclone (if there are multiple cyclones) or adding more water.  Increasing the feed pressure leads to the cyclone cutting at a finer size.

So, your results are not unexpected.

The paper is an improvement on the earlier draft.  I still feel that there is space to use cusum plots to show how the P₈₀ varies with some of these parameters.  For example, plotting P80 and cyclone feed pressure, or silo level.  A cusum plot might help identify the real influencing parameters, and how they relate.

Author Response

I have used the Pn, Cp, Po and NSG in the PCA.  However, I am not sure about the CUSUM plots.  I always thought that CUSUM charts were used for control purposes and not to compare variables or parameters.  I have asked Minerals to send you the file with the CUSUM charts.  I did not see any consistent trends.  However, the reviewer is welcome to share his/her thoughts.  I am open to discuss and include them.

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

Firstly, CUSUM plots can be used to detect changes in the process.  If the CUSUM plot is parallel to the x-axis there is no change in the parameter of interest.  However, if the CUSUM plot has a positive slope, it means that the parameter is increasing, and a negative slope indicates the parameter is decreasing.  For example, in the CUSUM plot below the P₈₀ initially increases, then decreases before trending up again.  Broadly, the primary ball mill feed follows the same trend.  Initially it increases, then decreases before increasing again.  By comparing the two CUSUM plots you can say with reasonable certainty that feed rate has an influence on P₈₀.  The is comforting as we know from experience that this is true, because reducing the throughput invariably leads to a finer particle size distribution and an increase in throughput results in a coarsening of the grind.

Cyclone feed pressure also influences the P₈₀.  We know from experience that a decrease in cyclone feed pressure will result in a coarser particle size distribution, and if you want the cyclone to cut at a finer size you increase the cyclone feed pressure.   In this case, the CUSUM plot shows an inverse relationship.  That is, as the cyclone feed pressure trends down, the P₈₀ trends up, and if the cyclone feed pressure trends up the P₈₀ tends to become finer.

Interestingly, the CUSUM plot of the NSG against the primary ball mill feed you will see that the throughput tends to decrease as the NSG feed grade increases (the ore is harder), then as the NSG feed grade trends down the throughput trends up.  This also corresponds with the P₈₀, in that as the NSG feed grade trends up the P₈₀ increases, and with a decrease in NSG grade the P₈₀ becomes finer.

The correlation with silo height is a little more convoluted.  

My view is that the authors should include the CUSUM plots at the beginning of their data analysis.  It might also help to show the time series plots of P₈₀ and throughput to show how the data varies with time.  Then based on these observations think about their analysis.

For example, throughout and NSG are probably not independent variables.  So, throughput multiplied by NSG might be a more appropriate parameter to use in your analysis of P₈₀.  Cyclone feed pressure should be independent and should have a direct impact on P₈₀.

Ultimately, to control the P₈₀ it is not a simple process of fixing one parameter.  It is complex needing an understanding of the ore, the upstream processes, the ball mills and cyclones.  Machine learning can do these things if the programming behind it is completed with a sound knowledge of the process.  Giving all the parameters to a data scientist and hoping that they will come up with something that works is, as the saying goes "garbage in garbage out".  It is extremely unfair on the data scientist as they have no context or understanding of the process.  I think that the CUSUM plots start to give the data some context to show what parameters should be used in developing a machine learning algorithm that will help improve the operation of the process.

Author Response

Firstly, I have worked in a concentrator in the past.  I knew all the information you mentioned in your reviews.  There were well received and thank you for that. 

In the concentrator where I worked as soon as the silo height decreased, the particle size distribution of the SAG feed increased.  We observed this on the sensor output that measured rock diameter.  The KPA in the SAG increased, so the throughput decreased and vice versa.  As far as the calculation of Pn, Cp, Po  and NSG, I know how to do these.  I did not use these because I thought using Ni, Cu, Po and S was enough to discuss the effect of head grade on hardness.

I have used mineralogy to design flowsheets in the past, so I know and appreciate its importance. 

I agree that cyclone feed pressure impacts P80.  However, when I worked in operations, the %solids to cyclone feed had a tremendous impact on P80 as well.  I can tell you that when the %solids was greater than 67%, regardless of pressure, the P80 coarsened and, at times, the flotation feed pipe would get plugged.  Thus, yes cyclone feed pressure is important, but the %solids in the cyclone feed is very important as well.  When I worked in operations, the grinding circuit could handle a certain amount of slurry.  When the supervisors asked to increase production, what happened is that the slurry flowrate would increase but there would be more solids in the pulp than water du to capacity limitations.  This was due to the grinding circuit not being able to handle higher slurry flowrates.  Production came first, cut water to accommodate the solids.  The higher %solids to cyclone feed caused the P80 to increase regardless of the pressure.  Assuming all the variables are kept the same, yes pressure will significantly impact P80. 

As far as the CUSUM charts, we use PCA to avoid analyzing two factors at a time.  That is the purpose of PCA, AI and machine learning. However, I have included CUSUM charts in the paper to support/explaining the findings.

In this study only the parameters that are important in the grinding circuit were considered.

Thank you very much for your input.  It is greater appreciated and valued. If you have further comments, I would be glad to discuss. 

   

Author Response File: Author Response.pdf