In today’s challenging economics for steelmaking, the need for enhanced efficiency and better quality is more than ever a daily reality. Among others, pot chemistry and pot level are two key parameters that need precision. The last 10 years has seen a laser based technology making its place in the world of pot chemistry control. Now with nearly thirty sites globally operating a 24/7 LIBS sensor, Tecnar is proud to introduce the next generation of its now established technology. The aim of the new product is to remove the need for lab analysis, usually required to calibrate the first generation instruments, thus giving full autonomy to the process personnel. Furthermore, the new generation sensor delivers a pot level measurement that is unaffected by skimming activities or dross accumulation at the surface. In this paper, we present the scientific and industrial performances of the calibration feature based on solid standards. The results show a relative standard deviation below 2%. Furthermore, thanks to the laser’s micro-sampling capabilities, the soluble contents can be measured directly on the solidified pot sample, without the need to a calculus based on Fe-Al content and pot temperature.
LIBS is now a generally accepted technology for measuring in continuous manner the composition of the zinc and aluminum pots. 1-5) The measurement system is based on the spectral analysis of a laser-generated plasma which identifies emission lines associated with specific elements – Al and Fe in this case – to a percentage content inside the matrix. The pulsed laser used to perform the plasma generation is flash-lamp-based: for every laser shot fired, a high voltage is applied for a very short time between the ends of a quartz lamp filled with inert gas, which discharges light that is then used by the amplification chamber to generate the laser pulse. With time, the efficiency of the flash-lamp reduced and in turn reduces the energy coming out of the laser head. After some research we have found the calibration curves to be dependent on the energy sent on the sample, which forces the user to re-build the calibration curve every month to ensure a continuously accurate reading.
To refresh the calibration curve, users have to extract a pot sample from the bath, let it cool, send it to a laboratory and have its chemistry measured. The calibration curve is then corrected depending on the value found at the time of the extraction and the chemistry measured in laboratory.
Some plants already possess their laboratory since it was their only way to measurement the content of their baths before the adoption of LIBS technology, but others to not have this instrument and have to send the pot sample at an external laboratory, which sends back the results in a matter of days. It is an expensive and time-lagging process for the user, which is the source of some reticence in the industry.
It is our belief at Tecnar that the product is mature enough to emancipate and never have to rely on a laboratory to be entirely reliable. To explore this idea, we have worked to understand thoroughly the ability of the LIBS to calibrate itself without the need of an external laboratory. By using two reference samples with different compositions, it would be possible to interpolate the content of a sample of unknown composition. Thus using the LIBS itself like an external laboratory.
In this paper, we present the analysis corroborating our certainty that it is possible to use a LIBS system as a laboratory on solid samples to calibrate for liquid baths operation.
The LIBS calibration unit is a system that complements the LIBS online chemistry monitor. Its purpose is to completely and independently validate and correct, if needed, the calibration of the pot-monitoring device through a comparison of user-provided solidified pot samples with Certified Reference Material (CRM) samples. In this approach, pot samples are still required but the LIBS itself is used as a laboratory instrument to determine the composition of the pot samples. There is no need for an external laboratory.
When a validation is required, as with the previous method, a procedure is triggered on the LIBS software to launch a data acquisition sequence for calibration. Detailed spectral data is then stored on the LIBS computer for further processing. At the same time, three (3) pot samples are taken at a location near the output of the LIBS probe (known as the lance). These samples are cooled and machined to have a surface texture suitable for LIBS analysis.
After the sampling, the LIBS probe is removed from the pot, the lance is detached from the probe, and replaced by the calibration unit, as shown in Figure 3. The calibration unit consists mainly of a tray designed to hold six (6) samples. Three of those samples are the Certified Reference Material (CRM) that have been qualified by fifteen internationally certified laboratories, and the other three are the pot samples taken by the operator.
Once activated, the unit will automatically acquire data from the three reference samples and from the three pot samples. The data from the three reference samples is used to validate the linear relationship between the light intensity ratio of a spectral line associated with an element, such as aluminum, to a reference spectral line of the zinc (background environment). The results from the CRMs allow the unit to determine the concentration of dissolved aluminum and iron in the three pot samples, using the measured linear relationship. The concentration results of the samples are then supplied to the LIBS software, in the same manners as if the data had been obtained from an external laboratory. It should be noted that the procedure of using CRMs to set a relation and then obtain the concentration of unknown samples is the same as used for laboratory analytical tools. There is no invention at that level.
In order to obtain high accuracy on the CRM and pot samples, the plasma-generation laser is fired several times at the same location, and at several locations. Signal averaging as well as automatic rejection of bad signals (signals on dross or with low signal-to-noise ratio) are used to increase the accuracy of the method.
Once the validation is complete, the calibration unit is easily detached from the probe, the lance is re-installed, and the probe is re-inserted into the galvanizing bath to provide real-time monitoring of the pot level, of the chemistry of the bath, and of the dross level. The complete process of calibration validation takes about three (3) hours.
THOMSON TAU REJECTION METHOD
Presence of outliers in sampling is a main issue when only a fraction of the population is inspected. Since it is practically impossible to inspect the entire pot sample with LIBS, only a limited volume is analyzed to determine the content inside. When unexpected particles such as dross or oxide sites are sampled, they do not contribute coherently to the rest of the matrix. If considered as part of the distribution, they will skew the average value and bias the calibration curve. For this reason, they have to be removed from the calculation.
The rejection method we used is the Thomson Tau method. For every point in the distribution, we calculate , the normalized deviation of data point , as follows:
where the average and standard deviation are calculated from all collected data points. We then compare this value to a threshold rejection value with a confidence level , calculated as
With this value, we compare every with . If , the data point is accepted, while if, the value is rejected and and are calculated again with the new distribution. This recalculation is iterated as long as there are points to be rejected.
Several situations were analyzed to fully understand the best conditions needed to have a good, reliable and accurate calibration curve.
The drilling effect
For this analysis, we performed the usual 1000 shots, 400 buffer experiment but let it run afterwards and recovered the different ratios of Al 308.2, Fe 259.9 and Zn 250.2 every 100 shots. The setup used for this part of the experiment was the standard LIBS configuration.
What is clear from this experiment is that even after surfacing the part, there is a layer to go through before getting to the bulk of the sample. This is clearly demonstrated on the evolution of Zn 250: as the crater gets around 300-shot deep, there is a very clear and important drop in the measured ratio which stabilizes over the following shots which could be explained by the geometry and depth of the crater. For three tests over different sites, the evolution of Zn 250 is roughly the same, which means that the geometry at a specific shot is mostly independent of the sampling site on the surface.
Figure 1 Evolution of aluminum 308.2 nm intensity (left) and ionic to atomic zinc ratio (right) with shot number.
From these results, it appears that there is always a surface effect on the sample. It is still not clear if this effect comes from a crust at the surface of the sample or if it comes from a change in the crater geometry. To analyze the true chemistry of the sample, the calibration process has to first pass through this surface effect.
The micro-sampling approach
The first approach to this problem is to use a high number of measuring sites on a single sample, each site having up to 1000 laser shots to get rid of the top contamination layer and decide, using the Thomson Tau method, which sites are approved to go into the post-averaging of qualified sites. Changes in the system’s operating settings were also required in order to accelerate the measuring rate.
We have reduced the laser lamp voltage, reduced the Q-Switch delay, increased the repetition rate to 10 Hz and increased the argon flow to 0.75 lpm. The prime objective is always to minimize the 95% confidence level error (∆µ) on the average (µ) with at least 15 significant data points (not rejected by the Thomson Tau method) and to minimize the number of data points necessary to have an error under 2%.
The total number of sites analyzed includes the sampling sites that were found as outliers. For this first part, only one series of acquisition was analyzed, meaning that each run corresponds to the average of the same sampling sites. Since this is possible to do so in this part of the optimization, it clears up some variability dependent on the sites samples.
A total of 7 metal samples have been used in this study. Two were coming from an operating CGL line, two more were certified reference material (CRMs) and the last three were home-made synthetic samples produced in Tecnar’s 100 kg capacity crucible. The relative standard deviation expected on these compositions varies from 2% for the CRM to 5% on the synthetic samples made in the Tecnar crucible. Figure 2 shows the calibration curve obtained using all seven samples with the micro-sampling approach.
Figure 2 Calibration curve generated by the micro-sampling method using all samples.
As can be seen from this curve, there is a good fit that can be obtained from the various samples, however we do not have the precision we are looking for yet. The average uncertainty with this calibration curve is 3.3%, which is over the required precision of the method.
At this point the method seems very promising. We think standardizing rigorously the acquisition method will reduce the error on the measurement further. Nonetheless, we allowed ourselves to explore a new strategy for data filtering which will follow.
Precision through reproducible conditions
As it is clear from the previous experiments that changing the number of cleaning shots, the buffer size or the integration delay does not affect the uncertainty on the average value, we have built an eye-safe prototype that would minimize the uncertainty on the average LIBS operation value by letting the sample be the only source of variation.
With this prototype in hand, seen in Figure 3, the flow of argon is more constant since the trajectory remains the same on all samples, the angle and plane of analysis is independent of the sample surfacing declination or sample.
Figure 3 Six positions carrousel prototype that connects to the LIBS system.
With this prototype, we acquired used the wavelengths of Al 308.3 nm and Fe 259.9 nm normalized on Zn 303.2 nm over 25 shots on 4 samples and filtered the values using the Thomson Tau rejection method.
To test the calibration, we built a calibration curve based on the values obtained on the reference samples and then compared the calculated value from the calibration curve to the reported aluminum content of the two pot samples. For these two samples, we report an average error of 1.26%, which is under the target error.
A different approach was studied in order to have a more reliable measurement of the effective aluminum content in the sample in the presence of embedded dross particle in the pot sample. The basis of the new approach was to combine every shot performed by the system into one histogram. While some samples present a normally distributed histogram chart, others present heavily skewed distributions. On the left of Figure 4, we see a symmetric distribution of aluminum results coming from a real pot sample with low iron content. On the right side, we see a skewed distribution obtained on a pot sample containing high iron content.
Figure 4 Normally (medium dross sample) and skewed (high dross sample) distributed histogram charts.
A normal distribution will have a better reproducibility and a lower error on the average value than a skewed distribution. Skewed distributions could induce an additional error on the content measurement of aluminum and iron and have to dealt with a different strategy.
We found that two Gaussian distributions could be adjusted to the histogram and give an excellent correlation. The reason as to why this fit works is that when we perform a measurement on a site, we shoot on a site containing either 0 or 1+ dross/oxide particles. If the site contains only dissolved aluminum, the measurement is , while if the site contains a particle, the measurement is , the sign of the equality explained by the higher aluminum content in particles than in the zinc matrix. The correct average value to be extracted from the Double-Gaussian method should be from the distribution with the lowest average value.
Direct measurement of soluble content on solid samples
A complete set of measurement was then produced using the two CRM samples for establishing a calibration curve and then the two real pot samples, labeled CGL1 and CGL2, were also measured. Figure 5 through Figure 8 display the data. For both aluminum and iron, the D-Gauss and the micro-sampling processing were applied and compared. The charts have the LIBS SIGNAL on the x-axis. The y-axis show the composition measured by the certified laboratories for the CRM and by atomic absorption using the DEAL software for calculating the soluble fractions.
The objective of this final trial was to determine which data processing is the most accurate and also to confirm the potential of the LIBS on solids for directly characterizing the soluble content. This last item is a breakthrough is the field of laboratory analysis because clashes strongly with all other laboratory instruments currently available. Indeed, the current laboratory practices always analyze a fairly large volume of material, which in turn produces a measurement that is closer to the total content of aluminum or iron. In contrast to this, we believe that the very small sampling that a laser shot combined with a spectrum per spectrum processing algorithm can enable the potential to sort out the outlying shots on dross particles and directly isolate the result that is proportional to the soluble composition of aluminum or iron.
By looking at Figure 5 through Figure 8, we see in black diamonds shape dots the results on the CRM samples that have produced the grey dashed calibration curve. The CRM samples are considered perfectly homogenous, equivalent to a pot sample completely free of dross. We also see the light grey with black outline that represent the soluble fractions of aluminum and iron in the CGL samples. Finally we have the dark grey triangles that show the total content of aluminum and iron in the CGL samples.
If the LIBS technique used here was to produce signals proportional to the total content of each element, because a vast majority of readings were to have dross content in the analyzed volume, then the dark grey triangles would be found right on the dashed calibration curves produced by the certified reference material. But in the contrary the dark triangles are found way above the line. CGL1’s total content is not very high above for aluminum because it was obviously taken in a clean galvanneal pot. But for all other total content dots, it is very clear that they do not fit the calibration curve. This is a first confirmation that the LIBS signal of this experiment does not measure the total content.
Looking at the dissolved content, graphically imaged by the light gray squares, it is acceptable to say that they are in good alignment with the dashed calibration line, both for iron and aluminum.
We therefore have confirmation that the LIBS reading on the CGL samples produce measurements proportional to the soluble fraction, thus enabling a direct soluble measurement technology without the need for a calculus based on solubility curves.
Figure 5 Aluminum results using micro-sampling
Figure 6 Aluminum results using D-Gauss.
Figure 7 Iron results using micro-sampling.
Figure 8 Iron results using D-Gauss
The next step was to determine which processing produced the best results. Table 1 presents the aluminum measurements that were produced by using the calibration curves that were generated with the two CRM samples. Both the D-Gauss and the Micro-sampling methods were used. The true content of the samples is also displayed to ease the comparison.
Table 2 LIBS evaluation of soluble aluminum and iron, using the D-Gauss and Micro-sampling, based on a two-point calibration curve generated by CRM material.
Table 3 Error on calibrated results for aluminum and iron using the D-Gauss and the Micro-sampling methods.
Table 3 lists the errors that were produced by measuring the composition of CGL1 and CGL2 using the calibration curves generated by the CRM samples GVL1 and GVL2. Both processing methods were used. The error on iron is relatively high and inconsistent. This can be attributed to the fact that the high CRM point for iron was very far from the operating range of CGL1 and CGL2. Furthermore, since the samples contained a high amount of excess iron in both cases, it is possible that there is a larger error on what is considered the true soluble composition.
Nonetheless, as this work is more focused on aluminum since it is the active element in the reaction, it can be said that the result is highly acceptable. The micro-sampling technique has generated an averaged relative error of 1.25%. Since the goal of the LIBS calibration on solid samples was to have a measurement error below 2%, the experiment is considered highly successful.
Finally, it seems that the micro-sampling method is producing better results than the D-Gauss method. It is believed that the cause of this observation is that the skewed distribution obtained by considering all laser shots is still too dense and narrow for a double Gaussian fit to adequately filter the shots on dross containing laser shots.
In conclusion, this paper has presented the past year of efforts aimed at making the LIBS pot monitoring technology completely independent from an external laboratory for its routine calibration. The goal of this effort was to remove the need for a lab, give better autonomy to the end user and enhance the repeatability of the calibration itself. Two methods were studied. The first method, called micro-sampling, consisted of measuring on many different sites and qualifying each site using the Thompson-Tau algorithm. The second method, called the D-Gauss, consisted of processing all the laser shots of all analyzed spots in a general distribution where two Gaussian curves would be fit, the lowest one then revealing the soluble content of the sample. Calibration curves using two certified reference material samples were generated and consequently used to guess the composition of real pot samples of GA and GI pot chemistry. The results have proven that LIBS is capable of directly evaluating the soluble fraction of the elements on solidified samples without the need of a solubility curve model. Furthermore, the relative error obtained on aluminum, the most important element in continuous galvanizing, was only 1.25%, well below the 2% target of the project, which makes this new technology a very competitive alternative to the traditional laboratory practice.
1) Novel method for on-line chemical analysis of continuous galvanizing baths, E. Baril, L. St-Onge, M. Sabsabi, J. Lucas, GALVATECH 2004 Proceedings, Chicago (IL), USA.
2) US PATENT no 6909505, Method and apparatus for molten metal analysis by laser induced breakdown spectroscopy, John M. Lucas, Mohamad Sabsabi, René Héon, National Research Council of Canada / Noranda Inc..
3) Online monitoring of effective aluminum and total iron, Alexandre Nadeau, François Nadeau, Craig Dewey, GALVATECH 2011 Proceedings, Genoa, Italy.
4) Online monitoring of dross and effective aluminum , Alexandre Nadeau, Lutfu Ozcan, François Nadeau, Mohamad Sabsabi, Paul Bouchard, Craig Dewey , The Asia-Pacific Galvanizing Conference 2009 Proceedings, Jeju Island, South Korea.
5) Lab free pot chemistry monitoring: LIBS brought to the next level, Alexandre Nadeau, Jonathan Roy, François Nadeau, Chris Warnement, Jeffrey Stechschulte, GALVATECH 2015 Proceedings, Toronto, Canada.
Alexandre Nadeau, Jean-David Grenon, Marc Choquet,
Tecnar Automation Ltée,
1321 Hocquart, Saint-Bruno-de-Montarville, QC, Canada