Material Identification Using Laser Spectroscopy ... - Semantic Scholar
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms 1
1
2
2
Ota Samek , 9ODGLVODY.U]\åiQHN , David C.S. Beddows , Helmut H. Telle , 1 1 Josef Kaiser , and Miroslav Liška 1
Institute of Physical Engineering, University of Technology, Technicka 2, Brno, 616 69, Czech Republic [email protected] 2 Department of Physics, University of Wales Swansea, Singleton Park, Swansea, UK [email protected]
Abstract. We report on pattern recognition algorithms in discriminant analysis, which were used on Laser Induced Breakdown Spectroscopy (LIBS) spectra (intensity of signal against wavelength) for metal identification and sorting purposes. In instances where accurate elemental concentrations are not needed, discriminant analysis can be applied, to compare and match spectra of “unknown“ samples to library spectra of calibration samples. This type of “qualitative“ pattern recognition analysis has been used here for material identification and sorting. Materials of different matrix materials (e.g. Al, Cu, Pb, Zn, vitrification glass, steels, etc.) could be identified with 100% certainty, using Principle Component Analysis and the Mahalanobis Distance algorithms. The limits within which the Mahalanobis Distance indicate a match status of Yes, Possible or No were investigated. The factors, which dictate these limits in LIBS analysis, were identified - (i) spectrum reproducibility and (ii) the sample-to-sample homogeneity. If correctly applied the combination of pattern recognition algorithms and LIBS provide a useful tool for remote and in-situ material identification problems, which are of a more "identify-and-sort" nature (for example those in the nuclear industry). Keywords: pattern recognition, material identyfication, laser spectroscopy
1
Introduction
One of the spectroscopic technique based on Laser Ablation (LA) spectroscopy is Laser-Induced Breakdown Spectroscopy (LIBS). This technique offers simple, fast and real-time spectrochemical analysis, with little need for sample preparation. In the technique one utilizes the high power densities obtained by focusing the radiation from a pulsed, fixed frequency laser, to generate a luminous micro plasma from an analyte (solid, liquid and gaseous samples). To a good approximation, the plasma composition is representative of the analyte’s elemental composition. In the thirty years or so since its inception the potential of LIBS as an analytical tool has been realized, leading to an ever increasing list of applications, both for analysis in the laboratory and industrial environments [1,2]. W. Skarbek (Ed.): CAIP 2001, LNCS 2124, pp. 443–450, 2001. © Springer-Verlag Berlin Heidelberg 2001
444
O. Samek et al.
Mostly qualitative and quantitative LIBS analysis has been applied to the analysis of solid samples. Extensive studies have been carried out for a wide range of solid samples under different operating conditions to determine parameters such as the electron density, the plasma temperature and spectral line shapes, and their relationship to the validity of analytical outcomes has been studied [3]. Detection limits for solid samples typically are in the range of a few hundreds parts per million, less in a few specific cases. When deciding on a method for elemental analysis, major advantages of LIBS over the more conventional methods are – (i) no or very little sample preparation; (ii) analysis can be carried out equally on all three physical states of matter (solids, liquids and gases); (iii) the analysis is performed in real time (approximately a few seconds when using lasers with a 10-20 Hz repetition rate); and (iv) only a small amount in the order of a few mg is ablated from the surface of solid samples, and hence the method is virtually non-destructive. By using Discriminant Analysis, the LIBS system can be trained to recognise spectra from different samples, regardless of spectrum quality and reproducibility. Ironically, instead of collecting spectra under set ablation parameters the spectra required for the generation of these Discriminant Analysis models need to reflect all measurement conditions. This then implies that the precision and accuracy of the LIBS measurement is no longer a major issue. This is because the system is trained to evaluate the similarity of unknown spectra to its bank of data relating to the samples it is sorting. With diminished importance of spectrum quality, the ablation parameters (e.g. lens-to-sample separation) become less critical thus making a commercial LIBS system that can "point-shoot-and-identify" feasible.
2
Experimental
LIBS systems have become quite common in recent years, and full descriptions of typical systems have been given elsewhere, see e.g. in [4]. Here we only summarise the characteristic features of the laboratory system used in this study. The laser system - a standard Nd:YAG laser was used to generate the LIBS plasma probe beam (at wavelength of 1064nm, at a repetition rate of 10Hz). Individual laser pulses had a pulse length of about 10ns; these could be adjusted for pulse energies of 10-100mJ, using a Glan polariser. The light delivery system - mostly, a laser beam delivery system based on lens / mirror optics was used. Light from the plasma is collected by a lens/lenses, which focused the plasma light emission onto an optical fibre bundle, connected to the analysis spectrometer or in some arrangements directly to the spectrometer. The system for spectral analysis - the system used for spectral analysis consisted of a standard spectrograph (ACR500, Acton) with a gateable, intensified photodiode array detector (IPDA, Princeton Instruments) attached to it. The gating of the detector and the timing for spectral data accumulation are controlled by a PC via a pulse delay generator (PG200, Princeton Instruments).
3
Discriminant Analysis
Each spectrum collected using a LIBS instrument is a “finger-print“ of the material being analysed and the conditions under which it was collected. Most of the efforts in
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
445
quantitative LIBS research have been aimed at normalising the spectrum collection conditions and procedures so that the spectrum only characterises the material. If no normalisation routines are carried out, and it is assumed that the sample is homogeneous, then the transient conditions of the LIBS measurement provide the main source of irreproducibility. More commonly known as Discriminant Analysis in spectroscopy, the aim of any pattern recognition algorithm is to unambiguously determine the identity or quality of an unknown sample using a spectrum obtained from the sample. There are two basic applications for spectroscopic discriminant analysis: (i) material purity/quality and (ii) material identification/screening. Material Quality Control - in its capacity for sample checking, discriminant analysis models could in principle replace many quantitative methods currently used. In effect, the algorithm gives an indication, “YES“ or “NO“, of whether the spectrum of the "unknown" sample matches the spectra taken from samples that were known to be of "good" quality. Material Identification - when discriminant analysis is used in a productidentification, or a product-screening mode, the spectrum of the "unknown" is compared against multiple discriminate models. Each model is constructed from the spectra collected from samples representative of various material groups defined by the grade, purity or quality of the sample. An indication of the likelihood of the spectrum matching one of these groups is then made, and the material is therefore classified as the closest match, or no match at all. There are many pattern recognition algorithms that can be used to assess the similarity of a measured spectrum with the training set. Here, the description is restricted to the algorithm used in this study, namely the Mahalanobis Distance Method.
4
Application of Mahalanobis Distance Method of Spectrum Matching to LIBS Spectra
In order to calculate the Mahalanobis Distance (M.Dist), Principle Component Analysis (PCA) is used. This decomposes the training set spectra into a series of mathematical spectra called factors which, when added together, reconstruct the original spectrum. The contribution any factor makes to each spectrum is represented by a scaling coefficient (score) which is calculated for all factors identified from the training set. Thus, by knowing the set of factors for the whole training set, the scores will represent the spectra as accurately as the original responses at all wavelengths. For a detailed description of the Mahalanobis Distance algorithm see [5]. 4.1
Measurements Using the Mahalanobis Distance
The method of measurement and evaluation using the Mahalanobis Distance method will be exemplified for the distinction of three marginally different steels, encountered in the quality control for an assembly of boiler tubes. LIBS spectra were recorded in the range 320-350nm.
446
O. Samek et al.
To fully appreciate the "mechanism" behind the M.Dist measurement, consider first the application of the M.Dist directly to the full spectral data set without using PCA, and in parallel, consider the response of just two major spectral peaks (e.g. Cu (l=344.15nm) and Fe (l=324.84nm)) in multiple spectra collected from a single sample. If the intensities of these two peaks are plotted against each other for numerous measurements, then an elliptical scatter of points would be expected due to the fluctuations in the measurement conditions caused by factors such as spectrometer response, sample handling and sample preparation. It is the scatter of these points which defines the M.Dist about a mean centre, in the same way that the standard deviation s of a one-dimensional measurement x, defines the scatter about the mean measurement. However, the M.Dist applies to all pixels in the spectrum, not just to one or two pixels of a peak, and therefore can be considered to be a multi-dimensional standard deviation that is applied to the whole spectrum. When creating the Discriminant Analysis models a list of the training set spectra was simply entered into the "in-house" / GRAMS PLSplus programs. Choosing the Discriminant Analysis option, the program generated a Discriminant Analysis model for each sample, against which test spectra were matched. The related data for the aforementioned boiler tube samples were all saved into a single calibration file, which - when loaded into the Prediction Code - performed the matching routines for each sample. When checking the identity of the test spectra collected from each sample all were either identified as definite or possible matches to one of the Discriminant Analysis models. Tests carried out on all recorded spectra (in total 48) showed conclusive evidence that by using the M.Dist the spectra could be correctly re-categorised into the three sample groups. However, a few of the spectra did register a NO-NO-POSSIBLE combination when compared with the Discriminant Analysis Models for the three steels. Even though some of the spectra returned only a POSSIBLE match, nevertheless a positive identification could be claimed. This was because of the relatively large M.Dist values calculated for the failing models. For a model returning a “YES“ or a “POSSIBLE“, the M.Dist values were between 0 and 3. For a “NO“ match result the M.Dist were much greater than 3, normally at least an order of magnitude greater than the M.Dist calculated for a “YES“ or POSSIBLE“ result. This point is re-emphasised when using LIBS to grade steels and identify different matrix elements, as outlined in the sections below. From the repeat analysis of the spectra collected from the samples of the three boiler tube steels, it could safely be said that Discriminant Analysis had the potential of providing a superior tool for matching LIBS spectra and identifying “unknown“ materials, when compared to standard semi-quantitative analysis. To emphasise this point Discriminant Analysis has been applied to the identification of materials with different and similar matrices. 4.2
Identification of Materials of Different Matrices
The materials used to test the ability of the LIBS system to identify different matrices were Al, Cu, Pb, Zn, mild steel, stainless steel and simulated vitrification glass. For
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
447
each material, 50 spectra were collected, plus an additional 10 spectra which were treated as “unknowns“. Each spectrum was collected using a 150 lines/mm grating ( centred at l = 450 nm, and accumulated for only 50 laser induced plasma events ). Note that for this part of the study, no optimisation of the plasma generation and recording was afforded but rather a “point-and-shoot“ approach was taken. 30
(a)
25 20 15
30
10
20
0
15
30
10
(b)
25
5
20
0
15
30
(f)
25
5 0 30
(c)
25
Intensity (a.u.)
Intensity (a.u.)
10
20 15 10 5
20
0
15
30
10
(g)
25
5
20
0
15
30
10
(d)
25
5
20
0 370
15 10
420
470
520
570
620
Wavelength (nm) wavelength [nm]
5 0 370
(e)
25
5
420
470
520
570
620
wavelength Wavelength (nm) [nm]
Fig. 1. LIBS spectra collected for samples of (a) aluminium, (b) copper, (c) lead, (d) zinc, (e) mild steel, (f) stainless steel and (g) vitrification glass.
As can be seen from Figure 1, the differences between the spectra collected from each of the samples reflect the completely different matrices. Only the spectra from the mild and stainless steel samples show any similarity, and the ability of LIBS Discriminant Analysis (LIBSDA) to discriminate between these samples was explored further for the identification of the grade of five ferritic steels. To make the test realistic the test spectra were coded, mixed and then sorted using the prediction module. Notice that even though the prediction made on the copper sample resulted in a POSSIBLE verdict, because the M.Dist value was greater than 1, that the spectrum had to be from copper because the other Discriminant Analysis models all gave a M.Dist value of the order 10,000. Clearly, the spectrum was not from any one of
448
O. Samek et al.
these materials. Sorting the spectra in this way a 100% identification result was obtained when using the Discriminant Analysis models to identify the 10 “unknown“ spectra collected from each material. Remembering that the M.Dist is effectively a measure of the similarity of an "unknown" spectrum to a group of training spectra, it can be expected that the M.Dist for the Discriminant Analysis models reporting a “NO“ result to be high. The reasoning behind this is that spectra collected from each element are very different. Hence, when one of the models was presented with a spectrum not belonging to its native set then the M.Dist was large. For models generated using the spectra from similar materials it would be expected that the M.Dist. were smaller for a model giving a “NO“ result. This was indeed the case for grade identification of stainless steels. A typical result is shown in Table 1, where a spectrum taken from a copper sample was compared to the models produced for Al, Cu, Pb, Zn, Stainless Steel, Mild Steel, and Vitrification Glass. Table 1. Typical Prediction Module Result for the identification of materials with totally different spectra, exemplified for a copper sample.
Sample Aluminium Copper Lead Zinc Stainless Steel Mild Steel Vitrification Glass 4.3
Match NO Possible NO NO NO NO NO
M Distance 4,063 1.04 12,760 15,527 3,273 44,805 19,473
Limit Test FAIL PASS FAIL FAIL FAIL FAIL FAIL
Identification of Materials of Similar Matrix (Steel Grade)
Using a spectral segment, once again centred at 450 nm, Discriminant Analysis models where derived from 100 spectra collected from a range of certified stainless steels (SS469 - SS473, for their composition see Ref [5]). Following this, 10 extra spectra were collected from each sample to test the ability of the Prediction Code and the Discriminant Analysis models to identify the samples. Although the spectra collected from the different steels were similar the Prediction Code was still capable of distinguishing between them. The M.Dist values of the (successful) identification procedure for all test-spectra collected from the five steels are given in Table 2. In general we noticed that the M.Dist of the failed models was about one order of magnitude larger compared to the worst-case average value of 1.4 for the SS469 model. This trend was observed for all the test spectra collected for this sample, and was repeatedly seen when sorting other steels. Overall, all “positive“ identification results gave an M.Dist value well below 3. Comparing this with the fact that the M.Dist calculated for the models giving a NO result (i.e. 1 to 2 orders of magnitude greater), a 100% identification result is obtained. The higher than expected M.Dist values for the successful model were more than likely due to the poor reproducibility of the spectra. On this basis the fluctuations in the spectra may be accounted for in the Prediction Module by changing the M.Dist PASS/FAIL limits.
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
449
These limits are somewhat arbitrary anyway, reflecting the tolerance s provided by the operator. Analysts often use values greater than 1-2 and 2-3 for the YES and POSSIBLE results, e.g. 1-5 and 5-15, respectively. Therefore, these limits have to be determined. For example, for our specific case, the limit values for the identification modules could be changed from 0 < MDist < 1 to
PASS; 1 < MDist < 3
POSSIBLE; and MDist > 3
FAIL ,
0 < MDist < 3 PASS; 3 < MDist < 6 POSSIBLE; and MDist > 6 FAIL , and thus return a PASS for all the successful matches. The factors which dictate these limits in LIBS analysis are (i) spectrum reproducibility and (ii) the sample-to-sample homogeneity. Table 2. M.Dist identification results from the „successful“ models, out of the test against training reference spectra from all five steels.
Sample No 1 2 3 4 5 6 7 8 9 10 Average MD YES POSSIBLE NO
5
SS469 1.63 1.61 2.12 1.33 1.50 1.30 1.38 1.03 0.87 1.29
Mahalanobis Distance SS470 SS471 SS472 0.72 0.71 0.85 0.84 0.68 0.99 0.95 0.58 1.92 1.36 0.93 1.13 0.59 2.02 0.82 1.72 0.73 0.99 0.57 0.81 0.94 0.86 0.86 1.13 0.58 0.90 0.96 0.81 0.75 1.13
SS473 0.75 0.46 0.53 1.06 0.76 0.69 0.83 0.91 0.90 0.60
1.4
0.9
0.9
1.1
0.7
1 9 0
8 2 0
9 1 0
6 4 0
9 1 0
Summary
In summary the pattern recognition algorithm in discriminant analysis can be used on LIBS spectra for metal identification and sorting purposes. It is easy to apply, and the results obtained indicate a 100% identification rate for materials both of common and non-common matrices. However, as with all Multivariate Quantitative Analysis methods, careful application is required if the technique is to be applied both correctly and successfully. For example, the limits within which the M.Dist indicate a match status
450
O. Samek et al.
of YES, POSSIBLE or NO can be changed (see section 4.3 above). By testing the models produced with randomly collected spectra from different samples of the material it represents, the range of M.Dist values which give a positive identification needs to be found. If this is not done then the model might incorrectly miss-identify materials. Furthermore, by adjusting the M.Dist limits, the poor reproducibility can in principle be accounted for, provided there are significant elemental differences in the samples being sorted, such that clear changes in the spectral responses can be observed. Therefore, if correctly applied the combination of discriminant analysis and LIBS provides a useful tool for remote and in-situ material identification problems. We would like to note that the examples reported here represent only a few of the material identification applications identified by LIBS researchers. With the correct marketing, further industrially motivated applications may soon appear which require remote and in-situ analysis. Remote-LIBS technology exists and has been demonstrated in numerous applications (see e.g. [6]). In this work an attempt was made in an application which exploits LIBS to satisfy the commercial niche identified in the introduction, i.e. that of remote and in-situ material identification. Acknowledgement. O. Samek gratefully acknowledges the financial support by Grants GACR 101/98/P282 and CEZ:J22/98:262100002. D.C.S. Beddows acknowledges the sponsorship for his Ph.D. research by BNFL plc, Sellafield.
References 1. 2. 3. 4. 5. 6.
Majidi, V., Joseph, R.: Spectroscopic applications of laser-induced plasmas. Crit. Rev. Anal. Chem. 23 (1992) 143-162. Radziemski, L.: Review of selected analytical applications of laser plasmas and laser ablation 1987-1994. Microchem. J. 50 (1994) 218-243. Leis, F., Sdorra, W., Ko, J.B., Niemax, K.: Basic investigation for laser microanalysis: I. Optical emission spectroscopy of laser produced sample plumes. Mikrochim. Acta II (1989) 185-199. Samek, O., Beddows, D.C.S., Kaiser, J., Kukhlevsky, S., Liška, M., Telle, H.H., Young, J.: The application of laser induced breakdown spectroscopy to in situ analysis of liquid samples. Opt. Eng. 39 (2000) 2248-2262. Beddows, D.C.S.: Industrial application of remote and in situ laser induced breakdown spectroscopy. Ph.D. Thesis, University of Wales Swansea (2000). Davies, C.M., Telle, H.H., Montgomery, D.J., Corbett, R.E.: Quantitative analysis using remote laser-induced breakdown spectroscopy (LIBS). Spectrochim. Acta B 50 (1995) 1059-1075.
1
2
2
Ota Samek , 9ODGLVODY.U]\åiQHN , David C.S. Beddows , Helmut H. Telle , 1 1 Josef Kaiser , and Miroslav Liška 1
Institute of Physical Engineering, University of Technology, Technicka 2, Brno, 616 69, Czech Republic [email protected] 2 Department of Physics, University of Wales Swansea, Singleton Park, Swansea, UK [email protected]
Abstract. We report on pattern recognition algorithms in discriminant analysis, which were used on Laser Induced Breakdown Spectroscopy (LIBS) spectra (intensity of signal against wavelength) for metal identification and sorting purposes. In instances where accurate elemental concentrations are not needed, discriminant analysis can be applied, to compare and match spectra of “unknown“ samples to library spectra of calibration samples. This type of “qualitative“ pattern recognition analysis has been used here for material identification and sorting. Materials of different matrix materials (e.g. Al, Cu, Pb, Zn, vitrification glass, steels, etc.) could be identified with 100% certainty, using Principle Component Analysis and the Mahalanobis Distance algorithms. The limits within which the Mahalanobis Distance indicate a match status of Yes, Possible or No were investigated. The factors, which dictate these limits in LIBS analysis, were identified - (i) spectrum reproducibility and (ii) the sample-to-sample homogeneity. If correctly applied the combination of pattern recognition algorithms and LIBS provide a useful tool for remote and in-situ material identification problems, which are of a more "identify-and-sort" nature (for example those in the nuclear industry). Keywords: pattern recognition, material identyfication, laser spectroscopy
1
Introduction
One of the spectroscopic technique based on Laser Ablation (LA) spectroscopy is Laser-Induced Breakdown Spectroscopy (LIBS). This technique offers simple, fast and real-time spectrochemical analysis, with little need for sample preparation. In the technique one utilizes the high power densities obtained by focusing the radiation from a pulsed, fixed frequency laser, to generate a luminous micro plasma from an analyte (solid, liquid and gaseous samples). To a good approximation, the plasma composition is representative of the analyte’s elemental composition. In the thirty years or so since its inception the potential of LIBS as an analytical tool has been realized, leading to an ever increasing list of applications, both for analysis in the laboratory and industrial environments [1,2]. W. Skarbek (Ed.): CAIP 2001, LNCS 2124, pp. 443–450, 2001. © Springer-Verlag Berlin Heidelberg 2001
444
O. Samek et al.
Mostly qualitative and quantitative LIBS analysis has been applied to the analysis of solid samples. Extensive studies have been carried out for a wide range of solid samples under different operating conditions to determine parameters such as the electron density, the plasma temperature and spectral line shapes, and their relationship to the validity of analytical outcomes has been studied [3]. Detection limits for solid samples typically are in the range of a few hundreds parts per million, less in a few specific cases. When deciding on a method for elemental analysis, major advantages of LIBS over the more conventional methods are – (i) no or very little sample preparation; (ii) analysis can be carried out equally on all three physical states of matter (solids, liquids and gases); (iii) the analysis is performed in real time (approximately a few seconds when using lasers with a 10-20 Hz repetition rate); and (iv) only a small amount in the order of a few mg is ablated from the surface of solid samples, and hence the method is virtually non-destructive. By using Discriminant Analysis, the LIBS system can be trained to recognise spectra from different samples, regardless of spectrum quality and reproducibility. Ironically, instead of collecting spectra under set ablation parameters the spectra required for the generation of these Discriminant Analysis models need to reflect all measurement conditions. This then implies that the precision and accuracy of the LIBS measurement is no longer a major issue. This is because the system is trained to evaluate the similarity of unknown spectra to its bank of data relating to the samples it is sorting. With diminished importance of spectrum quality, the ablation parameters (e.g. lens-to-sample separation) become less critical thus making a commercial LIBS system that can "point-shoot-and-identify" feasible.
2
Experimental
LIBS systems have become quite common in recent years, and full descriptions of typical systems have been given elsewhere, see e.g. in [4]. Here we only summarise the characteristic features of the laboratory system used in this study. The laser system - a standard Nd:YAG laser was used to generate the LIBS plasma probe beam (at wavelength of 1064nm, at a repetition rate of 10Hz). Individual laser pulses had a pulse length of about 10ns; these could be adjusted for pulse energies of 10-100mJ, using a Glan polariser. The light delivery system - mostly, a laser beam delivery system based on lens / mirror optics was used. Light from the plasma is collected by a lens/lenses, which focused the plasma light emission onto an optical fibre bundle, connected to the analysis spectrometer or in some arrangements directly to the spectrometer. The system for spectral analysis - the system used for spectral analysis consisted of a standard spectrograph (ACR500, Acton) with a gateable, intensified photodiode array detector (IPDA, Princeton Instruments) attached to it. The gating of the detector and the timing for spectral data accumulation are controlled by a PC via a pulse delay generator (PG200, Princeton Instruments).
3
Discriminant Analysis
Each spectrum collected using a LIBS instrument is a “finger-print“ of the material being analysed and the conditions under which it was collected. Most of the efforts in
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
445
quantitative LIBS research have been aimed at normalising the spectrum collection conditions and procedures so that the spectrum only characterises the material. If no normalisation routines are carried out, and it is assumed that the sample is homogeneous, then the transient conditions of the LIBS measurement provide the main source of irreproducibility. More commonly known as Discriminant Analysis in spectroscopy, the aim of any pattern recognition algorithm is to unambiguously determine the identity or quality of an unknown sample using a spectrum obtained from the sample. There are two basic applications for spectroscopic discriminant analysis: (i) material purity/quality and (ii) material identification/screening. Material Quality Control - in its capacity for sample checking, discriminant analysis models could in principle replace many quantitative methods currently used. In effect, the algorithm gives an indication, “YES“ or “NO“, of whether the spectrum of the "unknown" sample matches the spectra taken from samples that were known to be of "good" quality. Material Identification - when discriminant analysis is used in a productidentification, or a product-screening mode, the spectrum of the "unknown" is compared against multiple discriminate models. Each model is constructed from the spectra collected from samples representative of various material groups defined by the grade, purity or quality of the sample. An indication of the likelihood of the spectrum matching one of these groups is then made, and the material is therefore classified as the closest match, or no match at all. There are many pattern recognition algorithms that can be used to assess the similarity of a measured spectrum with the training set. Here, the description is restricted to the algorithm used in this study, namely the Mahalanobis Distance Method.
4
Application of Mahalanobis Distance Method of Spectrum Matching to LIBS Spectra
In order to calculate the Mahalanobis Distance (M.Dist), Principle Component Analysis (PCA) is used. This decomposes the training set spectra into a series of mathematical spectra called factors which, when added together, reconstruct the original spectrum. The contribution any factor makes to each spectrum is represented by a scaling coefficient (score) which is calculated for all factors identified from the training set. Thus, by knowing the set of factors for the whole training set, the scores will represent the spectra as accurately as the original responses at all wavelengths. For a detailed description of the Mahalanobis Distance algorithm see [5]. 4.1
Measurements Using the Mahalanobis Distance
The method of measurement and evaluation using the Mahalanobis Distance method will be exemplified for the distinction of three marginally different steels, encountered in the quality control for an assembly of boiler tubes. LIBS spectra were recorded in the range 320-350nm.
446
O. Samek et al.
To fully appreciate the "mechanism" behind the M.Dist measurement, consider first the application of the M.Dist directly to the full spectral data set without using PCA, and in parallel, consider the response of just two major spectral peaks (e.g. Cu (l=344.15nm) and Fe (l=324.84nm)) in multiple spectra collected from a single sample. If the intensities of these two peaks are plotted against each other for numerous measurements, then an elliptical scatter of points would be expected due to the fluctuations in the measurement conditions caused by factors such as spectrometer response, sample handling and sample preparation. It is the scatter of these points which defines the M.Dist about a mean centre, in the same way that the standard deviation s of a one-dimensional measurement x, defines the scatter about the mean measurement. However, the M.Dist applies to all pixels in the spectrum, not just to one or two pixels of a peak, and therefore can be considered to be a multi-dimensional standard deviation that is applied to the whole spectrum. When creating the Discriminant Analysis models a list of the training set spectra was simply entered into the "in-house" / GRAMS PLSplus programs. Choosing the Discriminant Analysis option, the program generated a Discriminant Analysis model for each sample, against which test spectra were matched. The related data for the aforementioned boiler tube samples were all saved into a single calibration file, which - when loaded into the Prediction Code - performed the matching routines for each sample. When checking the identity of the test spectra collected from each sample all were either identified as definite or possible matches to one of the Discriminant Analysis models. Tests carried out on all recorded spectra (in total 48) showed conclusive evidence that by using the M.Dist the spectra could be correctly re-categorised into the three sample groups. However, a few of the spectra did register a NO-NO-POSSIBLE combination when compared with the Discriminant Analysis Models for the three steels. Even though some of the spectra returned only a POSSIBLE match, nevertheless a positive identification could be claimed. This was because of the relatively large M.Dist values calculated for the failing models. For a model returning a “YES“ or a “POSSIBLE“, the M.Dist values were between 0 and 3. For a “NO“ match result the M.Dist were much greater than 3, normally at least an order of magnitude greater than the M.Dist calculated for a “YES“ or POSSIBLE“ result. This point is re-emphasised when using LIBS to grade steels and identify different matrix elements, as outlined in the sections below. From the repeat analysis of the spectra collected from the samples of the three boiler tube steels, it could safely be said that Discriminant Analysis had the potential of providing a superior tool for matching LIBS spectra and identifying “unknown“ materials, when compared to standard semi-quantitative analysis. To emphasise this point Discriminant Analysis has been applied to the identification of materials with different and similar matrices. 4.2
Identification of Materials of Different Matrices
The materials used to test the ability of the LIBS system to identify different matrices were Al, Cu, Pb, Zn, mild steel, stainless steel and simulated vitrification glass. For
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
447
each material, 50 spectra were collected, plus an additional 10 spectra which were treated as “unknowns“. Each spectrum was collected using a 150 lines/mm grating ( centred at l = 450 nm, and accumulated for only 50 laser induced plasma events ). Note that for this part of the study, no optimisation of the plasma generation and recording was afforded but rather a “point-and-shoot“ approach was taken. 30
(a)
25 20 15
30
10
20
0
15
30
10
(b)
25
5
20
0
15
30
(f)
25
5 0 30
(c)
25
Intensity (a.u.)
Intensity (a.u.)
10
20 15 10 5
20
0
15
30
10
(g)
25
5
20
0
15
30
10
(d)
25
5
20
0 370
15 10
420
470
520
570
620
Wavelength (nm) wavelength [nm]
5 0 370
(e)
25
5
420
470
520
570
620
wavelength Wavelength (nm) [nm]
Fig. 1. LIBS spectra collected for samples of (a) aluminium, (b) copper, (c) lead, (d) zinc, (e) mild steel, (f) stainless steel and (g) vitrification glass.
As can be seen from Figure 1, the differences between the spectra collected from each of the samples reflect the completely different matrices. Only the spectra from the mild and stainless steel samples show any similarity, and the ability of LIBS Discriminant Analysis (LIBSDA) to discriminate between these samples was explored further for the identification of the grade of five ferritic steels. To make the test realistic the test spectra were coded, mixed and then sorted using the prediction module. Notice that even though the prediction made on the copper sample resulted in a POSSIBLE verdict, because the M.Dist value was greater than 1, that the spectrum had to be from copper because the other Discriminant Analysis models all gave a M.Dist value of the order 10,000. Clearly, the spectrum was not from any one of
448
O. Samek et al.
these materials. Sorting the spectra in this way a 100% identification result was obtained when using the Discriminant Analysis models to identify the 10 “unknown“ spectra collected from each material. Remembering that the M.Dist is effectively a measure of the similarity of an "unknown" spectrum to a group of training spectra, it can be expected that the M.Dist for the Discriminant Analysis models reporting a “NO“ result to be high. The reasoning behind this is that spectra collected from each element are very different. Hence, when one of the models was presented with a spectrum not belonging to its native set then the M.Dist was large. For models generated using the spectra from similar materials it would be expected that the M.Dist. were smaller for a model giving a “NO“ result. This was indeed the case for grade identification of stainless steels. A typical result is shown in Table 1, where a spectrum taken from a copper sample was compared to the models produced for Al, Cu, Pb, Zn, Stainless Steel, Mild Steel, and Vitrification Glass. Table 1. Typical Prediction Module Result for the identification of materials with totally different spectra, exemplified for a copper sample.
Sample Aluminium Copper Lead Zinc Stainless Steel Mild Steel Vitrification Glass 4.3
Match NO Possible NO NO NO NO NO
M Distance 4,063 1.04 12,760 15,527 3,273 44,805 19,473
Limit Test FAIL PASS FAIL FAIL FAIL FAIL FAIL
Identification of Materials of Similar Matrix (Steel Grade)
Using a spectral segment, once again centred at 450 nm, Discriminant Analysis models where derived from 100 spectra collected from a range of certified stainless steels (SS469 - SS473, for their composition see Ref [5]). Following this, 10 extra spectra were collected from each sample to test the ability of the Prediction Code and the Discriminant Analysis models to identify the samples. Although the spectra collected from the different steels were similar the Prediction Code was still capable of distinguishing between them. The M.Dist values of the (successful) identification procedure for all test-spectra collected from the five steels are given in Table 2. In general we noticed that the M.Dist of the failed models was about one order of magnitude larger compared to the worst-case average value of 1.4 for the SS469 model. This trend was observed for all the test spectra collected for this sample, and was repeatedly seen when sorting other steels. Overall, all “positive“ identification results gave an M.Dist value well below 3. Comparing this with the fact that the M.Dist calculated for the models giving a NO result (i.e. 1 to 2 orders of magnitude greater), a 100% identification result is obtained. The higher than expected M.Dist values for the successful model were more than likely due to the poor reproducibility of the spectra. On this basis the fluctuations in the spectra may be accounted for in the Prediction Module by changing the M.Dist PASS/FAIL limits.
Material Identification Using Laser Spectroscopy and Pattern Recognition Algorithms
449
These limits are somewhat arbitrary anyway, reflecting the tolerance s provided by the operator. Analysts often use values greater than 1-2 and 2-3 for the YES and POSSIBLE results, e.g. 1-5 and 5-15, respectively. Therefore, these limits have to be determined. For example, for our specific case, the limit values for the identification modules could be changed from 0 < MDist < 1 to
PASS; 1 < MDist < 3
POSSIBLE; and MDist > 3
FAIL ,
0 < MDist < 3 PASS; 3 < MDist < 6 POSSIBLE; and MDist > 6 FAIL , and thus return a PASS for all the successful matches. The factors which dictate these limits in LIBS analysis are (i) spectrum reproducibility and (ii) the sample-to-sample homogeneity. Table 2. M.Dist identification results from the „successful“ models, out of the test against training reference spectra from all five steels.
Sample No 1 2 3 4 5 6 7 8 9 10 Average MD YES POSSIBLE NO
5
SS469 1.63 1.61 2.12 1.33 1.50 1.30 1.38 1.03 0.87 1.29
Mahalanobis Distance SS470 SS471 SS472 0.72 0.71 0.85 0.84 0.68 0.99 0.95 0.58 1.92 1.36 0.93 1.13 0.59 2.02 0.82 1.72 0.73 0.99 0.57 0.81 0.94 0.86 0.86 1.13 0.58 0.90 0.96 0.81 0.75 1.13
SS473 0.75 0.46 0.53 1.06 0.76 0.69 0.83 0.91 0.90 0.60
1.4
0.9
0.9
1.1
0.7
1 9 0
8 2 0
9 1 0
6 4 0
9 1 0
Summary
In summary the pattern recognition algorithm in discriminant analysis can be used on LIBS spectra for metal identification and sorting purposes. It is easy to apply, and the results obtained indicate a 100% identification rate for materials both of common and non-common matrices. However, as with all Multivariate Quantitative Analysis methods, careful application is required if the technique is to be applied both correctly and successfully. For example, the limits within which the M.Dist indicate a match status
450
O. Samek et al.
of YES, POSSIBLE or NO can be changed (see section 4.3 above). By testing the models produced with randomly collected spectra from different samples of the material it represents, the range of M.Dist values which give a positive identification needs to be found. If this is not done then the model might incorrectly miss-identify materials. Furthermore, by adjusting the M.Dist limits, the poor reproducibility can in principle be accounted for, provided there are significant elemental differences in the samples being sorted, such that clear changes in the spectral responses can be observed. Therefore, if correctly applied the combination of discriminant analysis and LIBS provides a useful tool for remote and in-situ material identification problems. We would like to note that the examples reported here represent only a few of the material identification applications identified by LIBS researchers. With the correct marketing, further industrially motivated applications may soon appear which require remote and in-situ analysis. Remote-LIBS technology exists and has been demonstrated in numerous applications (see e.g. [6]). In this work an attempt was made in an application which exploits LIBS to satisfy the commercial niche identified in the introduction, i.e. that of remote and in-situ material identification. Acknowledgement. O. Samek gratefully acknowledges the financial support by Grants GACR 101/98/P282 and CEZ:J22/98:262100002. D.C.S. Beddows acknowledges the sponsorship for his Ph.D. research by BNFL plc, Sellafield.
References 1. 2. 3. 4. 5. 6.
Majidi, V., Joseph, R.: Spectroscopic applications of laser-induced plasmas. Crit. Rev. Anal. Chem. 23 (1992) 143-162. Radziemski, L.: Review of selected analytical applications of laser plasmas and laser ablation 1987-1994. Microchem. J. 50 (1994) 218-243. Leis, F., Sdorra, W., Ko, J.B., Niemax, K.: Basic investigation for laser microanalysis: I. Optical emission spectroscopy of laser produced sample plumes. Mikrochim. Acta II (1989) 185-199. Samek, O., Beddows, D.C.S., Kaiser, J., Kukhlevsky, S., Liška, M., Telle, H.H., Young, J.: The application of laser induced breakdown spectroscopy to in situ analysis of liquid samples. Opt. Eng. 39 (2000) 2248-2262. Beddows, D.C.S.: Industrial application of remote and in situ laser induced breakdown spectroscopy. Ph.D. Thesis, University of Wales Swansea (2000). Davies, C.M., Telle, H.H., Montgomery, D.J., Corbett, R.E.: Quantitative analysis using remote laser-induced breakdown spectroscopy (LIBS). Spectrochim. Acta B 50 (1995) 1059-1075.
