Seestrasse 14B l CH-5432 Neuenhof l Tel. +41 56 406 12 12 l email. firstname.lastname@example.org
SPECTRAFLOW COPPER APPLICATION
The SpectraFlow Analyzer can be used to measure copper in several locations of the mine or in the production process. An extensive Feasibility Study was done regarding the possibility and accuracy of the SpectraFlow measurement of Copper.
In order to check whether SpectraFlow is suitable to analyze On Line copper raw materials 4 typical samples were received by SFA for testing with SpectraFlow.
The samples were:
Secundary Cu/Ni Concentrate 6.0 kg
B Cleaner Concentrate 8.0 kg
Rougher Tail 7.5 kg
Primary Cu/Ni Concentrate 4.5 kg
For the samples Sec Cu/Ni Con, B Cleaner Con, Rougher Tail, the analysis has been sent early enough to use them as training samples for the calibration of the initial SpectraFlow test model. As the chemical analysis for the sample Primary Cu/Ni Con was unknown it was not used as part of the training set but used how well the model based on the samples Sec Cu/Ni Con, B Cleaner Con and Rougher Tail can predict the concentration of the Primary Cu / Ni Con sample. So Primary Cu/Ni Con was an independent sample for the validation of the model used.
In reality the 4 samples each represent a different step in the processing of the Copper / Nickel ore. So a rigorous treatment would mean that for each of these locations an independent model should be made. However this is not possible with only three primary samples. Therefore this primary investigation has the purpose of checking whether or not SpectraFlow can be used for On Line Analysis of the copper concentrates.
Therefore all samples are used to compile one single model for all concentrates received.
As general practice for raw material we always integrate moisture in the model building and we also take potential differences of the reflection into account due to different positions of the material, which is analysed.
With such a small number of samples it was also difficult to define an independent validation set to check how well the model predicts statistically “similar” samples that have not been used in building the model. The difficulty was tried to overcome in two approaches:
- The samples Sec Cu/Ni Con, B Cleaner Con, Rougher Tail were split in two. One half of them was used to build the model, the other half was used as “test data” to check how well the model works.
- As mentioned in the introduction the sample Primary Cu/Ni Con was used only to check how well the model built on the other 3 samples can predict its concentration
The training samples were first put into a drying chamber and stayed there for 8 hours at a temperature of 105 ° C. This was done to make sure the moisture is taken out of them.
Once the drying was done the weight of the sample was taken and the first set of training spectra. In order to account for potential differences in reflection the samples were mixed five times. After each mixing a spectrum was taken.
After the 5 spectra for the dry samples have been taken 2 % weight of water was added. Once the water has been added the samples were again mixed 5 times and after each mixing a spectrum was taken. This was done 7 times so that a total of 14 % weight of water was added. With this much water the samples still were reasonably granular.
After each addition of water the weight was recorded, which gave the final moisture concentration of the test sample at each step of taking the training spectra. The concentration of the constituents of interest was corrected by the thus defined moisture for each step.
By doing this for the 3 training samples a total of 105 training spectra were acquired. This is what I called inflating the training samples in my initial description what I plan to do with the samples. The “validation samples” were all left dry respectively at the ambient level. For validation each of the validation samples was also mixed 5 times and the predicted value was averaged over the thus acquired 5 spectra.
The following 3 plots show the training sample spectra dry and with maximum moisture. As explained in the first plot the step change at 1000 Nanometers is due to the instrument and has physical meaning. It is taken out for the model building
Plot 1: Dry and moist spectra of Sec Cu/Ni Conc. Plot 2: Dry and moist spectra of B Cleaner Conc.
Plot 3: Dry and moist spectra of Rougher Tail Sample
The reflectance spectra show what is to be expected. Due to the fact that the samples are of dark color the reflectance is really low. The next plot shows the dry spectra of all 4 samples on one plot with the same scale for all of them.
Plot 4: Dry Spectra of all 4 samples
Comparing the spectrum for Sec Cu / Ni Con with Primary Cu/Ni Con shows a high similarity so the use as “Blind Test Sample” is justified. A reason for concern is the fact that the original spectra as shown in plot 4 do not show an impressive structure. In Plot 5 therefore the first derivative is shown. It was calculated using Savitzky Golay with 25 points and 2nd order polynomial.
The plot shows a much better structure and also confirms the similarity of Sec Cu/Ni Con with Primary Cu/Ni Con.
With the structure shown in Plot 5 there is reason to expect that SpectraFlow can be used to analyze the concentrates
For each of the constituents that were communicated to SFA and of which a laboratory analysis is available the spectra for Sec Cu/Ni Conc., B Cleaner Conc. and Rougher Tails was built using PLS1 (Partial Least Square 1) in the standard software GRAMS. Each of these individual models was then used to predict the concentrations of the constituents of all 4 samples. For each constituent PLS 1 converged and the models used between 7 to 10 Principle Components, which is mathematically speaking quite satisfactory. The key however is the performance of the models in predicting the constituents. Table 1 shows the results of the prediction based on the model that was derived from the spectra, which have been acquired as described in this report
Table 1: Comparison of Concentrations predicted by SpectraFlow with the laboratory results
The values for Sec Cu/Ni Conc., B Cleaner Conc, Copper Rougher Tails were predicted by using the spectra that were derived from the split sample that has not been used to build the model. As it is however expected that the sample itself is quite homogeneous the results are encouraging. It has to be accepted however that we may have only predicted the training samples.
The last 2 columns show the prediction of a “Blind Sample” i.e. a sample that has not been used for model building. Taking into account the limitations (only 3 training samples) the results are really encouraging.
On a limited set of training samples it was possible to build a model for each of the constituents. By integrating moisture into the model it was possible to inflate the number of available training samples and achieve convergence for each model. Using a “Blind Sample” in addition to the split samples that were used for calibration shows very encouraging results in the analysis of the concentrates.