See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258653323

Spectra Analysis in Detection of Traces of Explosives

Article in International Journal of Modern Physics B · October 2012

DOI: 10.1142/S0217979212460137

CITATIONS READS

4 509

4 authors, including:

Héctor Diez-Machío M. T. Trobajo

Universidad de León Universidad de León
9 PUBLICATIONS 6 CITATIONS 18 PUBLICATIONS 64 CITATIONS

SEE PROFILE SEE PROFILE

M.V. Carriegos
Universidad de León
88 PUBLICATIONS 235 CITATIONS

SEE PROFILE

All content following this page was uploaded by M.V. Carriegos on 06 July 2015.

The user has requested enhancement of the downloaded file.

 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

International Journal of Modern Physics B

c World Scientific Publishing Company

SPECTRA ANALYSIS IN DETECTION

OF TRACES OF EXPLOSIVES

M.M. LÓPEZ-CABECEIRA, H. DIEZ-MACHIO, M.T. TROBAJO & M.V. CARRIEGOS

Departamento de Matemáticas, Universidad de León,
24071-León, SPAIN
[email protected]

Received Day Month Year

Revised Day Month Year

We focus the problem of classification of Raman spectra by using different techniques:

Statistic models, neural networks and expert systems. We point out some results for the
concrete problem of identification of explosive substances.

Keywords: Pattern classification; Raman spectrography; statistic models; neural net-

works; expert systems.

1. The Problem: Introduction

Our goal is to develop a integrated system of detection of substances via the anal-
ysis of their LIBS or Raman spectra. We focus on the detection of explosives via
the analysis of traces. Samples are used in order perform both supervised and au-
tonomous training of a network.
Sample data used in the study have been provided by Indra Sistemas, León.
Single objects (patterns) are spectra obtained by means of Raman LASER
technology. This technology is used in a wide range of practical studies and
applications4,6,12,14,17 . We consider about 20 explosives and 20 blank substances
and about 10 samples of each substance (taken by Raman LASER in different
environments). The total amount of information consists of 212 spectra of blank
substances and 204 spectra of explosives. We also simulate noise and ill-posed en-
vironments by means of considering some kind of artificial mixtures of samples.

1.1. The data

Our study is performed upon digital hash in the range of 150nm to 1905nm but
getting a fundamental improvement: Each continuous spectrum is reduced to a
digital hash of 850 parameters (intensities). Of course the choice of this intensities
(that is to say, the discretization of the spectrum) is one of the critical parts of this
work.

1
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

2 López-Cabeceira & Al.

14000

13000

12000

11000

10000

9000

8000

7000

6000

5000

4000

3000
0 500 1000 1500 2000 2500 3000 3500

Fig. 1. A sample spectrum of C4 explosive. A real case.

Our methods are focused on the analysis of perturbed samples (see Fig. 1)
because one rarely find a perfect trace (see Fig. 2) of a well known certified explosive.

3000

2500

2000

1500

1000

500

0
0 500 1000 1500 2000 2500 3000 3500

Fig. 2. The spectrum of C4 explosive. Ideal case.

On the other hand our samples show minor calibration uncertainties which may
be assumed and corrected by our classification systems.

2. Classification by means of statistic techniques

2.1. The procedure
Now we deal with the problem of spectrum classification by using statistical tech-
niques: We consider the discretization of spectra; that is to say, we consider 850
observations at well defined wavelengths of intensity of each sample ( 404 explosives
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

Spectra analysis in detection of traces of explosives 3

or mixtures and 212 blank substances). These observations are highly correlated.
Thus our first objective is to filter redundant information by means of selection
of a subset of non-correlated data but describing almost all information and thus
allowing to classify the sample into the set of spectra. This is performed by using
Principal Component Analysis (PCA).
Next we determine if new variables obtained from PCA analysis will be able to
discriminate between explosive and blank patterns; that is to say, if new variables
are enough to simulate the characteristic function of the set of explosives. We per-
form parametric tests for independent samples to determine which components have
significative different means. Of course this step depends upon the set of samples
of explosives and the set of samples of blank substances. Hence the previous choice
of both sets is critical.
Once significant principal components are selected for both sets (explosives and
blanks), a logistic regression is performed. Hence the probability of successful clas-
sification is obtained as well as the probability of type I and type II error.

2.2. Statistical methods. Experimental results

We perform Principal Component Analysis (PCA) from the correlation matrix (no
rotations, no normalized data). The graph of sedimentation (Fig.3) shows that first
principal component gets more that 89% of the variance and the seven principal
components explain more than 95% of the variance.

Fig. 3. The graph of sedimentation of PCA.

On the other hand the correlations between the original variables and these
principal components are nearly zero and are not statistically significant (p > 0.05).
Hence we choose the seven principal components as output of PCA.
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

4 López-Cabeceira & Al.

Our goal is to classify a concrete sample as explosive or not. We reduce the above
set of principal components by means of selecting those variables with significant
p-values for a two sample t-test. Only 4 principal components show statistically
significant different means. This set of principal components are used in a logistic
model as predict variables (p < 0.01). We also reduce sample size and balance the
amount of blank and non-blank (explosive or mixture) samples. This latter reduction
is performed in order to obtain better logistic regression results 18,19 .
We define value 0 for a substance classified as blank (non-explosive) and 1 for a
substance classified as explosive. After performing a logistic regression with critical
classification value p = 0.5 we obtain the following results

Classified Blank Classified explosive

Blank Substance 167 45 (false positive)
Explosive 74 (false negative) 138

The following a priori probabilities are obtained from the above data:
sensibility = p(1|E) = 65.1%, specificity = p(0|B) = 78.8%. Hence the global
probability of successful classification is 71.93%.
Error probabilities are easily obtained. The false positive (type I error) with
probability 0.349 whereas the probability of false negative (type II error) is 0.212.

Fig. 4. ROC curve relating specificity vs sensitivity of procedure for noisy (non-corrected) data
(spectra)

Of course we can modify the critical classification value p = 0.5 in order to

 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

Spectra analysis in detection of traces of explosives 5

decrease the probability of false positive, but this increase the probability of false
negative; this is shown, for noisy case, in Fig. 4 by means of ROC curve relating
specificity on X-axis versus sensitivity on Y -axis.
A compromise is necessary depending on the real application we are trying. Min-
imize false negative (i.e. maximize sensitivity) would apply in our case of detecting
explosive substances. Next we write down explicit pairs (sensitivity, specificity) for
several critical classification values p. Note that this values are obtained from the
COR curve, which is parametrized by p:
critical value (p) sensitivity specifity false positive
0.31 95% 22% 78%
0.36 90% 31% 69%
0.5 65% 79% 21%
0.62 40% 90% 10%
0.8 13% 95% 5%

3. Classification by using neural networks

Once we have performed the PCA analysis we also try neural networks (instead of
logistic regression) in our classification problem. Three different neural networks are
considered: The multi-layer Perceptron (MLP), the radial based network (RBFN)
and self-organized maps (SOM).

3.1. The procedure

The pattern sets are formed by taking 212 samples of blank substances, 204 samples
of explosives, and a set of 190 − 290 mixtures in different experiments which were
random chosen from the 212 × 204 possibilities of simple mixtures blank-explosive.
A set containing 38 explosives, 26 blank samples and 138 mixtures are dropped
as test samples; remaining samples are used for the training of the network. Training
procedure involved a random training set and a random validation set.
First experiments show deceptive results in the training procedure; that is to
say, the network stabilizes on neural weights that were not able to classify effectively
(with acceptable probability of success) even pure explosives. Hence we perform a
previous PCA on the variables.
The network is trained to detect 20 explosive substances giving output NON-
BLANK substance. The detection of blank substances is more difficult because
the objective blank substances in real world is nearly infinite. But we need blank
substances in order to train the network and to avoid super-specialization.
Simulations are performed by using both one and two hidden lawyers. Note that
this is not necessary from the theoretical point of view.
We used the following parameters: Learning constant 0.1 < α < 1 (we include a
simulated annealing for this parameter). Learning error 10−5 < ε < 10−2 . Validation
error 0 < ν < 10−2 . Hidden artificial neurons are set in the range 5 ≤ N ≤ 180.
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

6 López-Cabeceira & Al.

3.2. Neural networks. Experimental results

Comparative results show that best network for our purposes is the Multilayer
Perceptron. As a matter of example we show experimental results on the particular
subset of explosives C4-RDX and G2ECO-Nitrate Ammonia and choosing Berilium
as blank substance.

Wrong Identification Correct Identification

C4 0 (false negative) 10 (RDX-C4)
mixture C4+blank 9 (false negative) 6 (RDX-C4)
G2ECO 0 (false negative) 10 (G2ECO-NA)
mixture G2ECO+blank 4 (false negative) 10 (G2ECO-NA)
Berilium (blank) 0 (false positive) 3

Standard statistical techniques lead to the following quantitative results: A priori

probabilities are: sensibility = p(1|E) = 0.8462, specificity = p(0|B) = 0.9231.
Hence the global probability of successful classification is 0.8846. On the other
hand the probability of false positive (type I error) is p(1|B) = 0.0769 while the
probability of false negative (type II error) is p(0|E) = 0.1538.
The same compromise of minimizing false negative would be taken because if the
network will be used as software of a explosive detection device it will be critical to
control type II error even at the cost of increase type I error. Our work is intended
to perform this improvement.

4. Classification by means of expert systems based on fuzzy logic

The same set of samples are used in the development of a fuzzy expert system. We
consider 212 samples of blank substances, 204 samples of explosive substances and
a set of mixtures.

4.1. The system

Our system is based in two ideas: The localization of significative peaks and the
relationship of these peaks with homologous peaks in known explosive substances.
In order to identify peaks we focus on the spectrum to distinguish which high
intensities in the received spectrum are real peaks and which ones are due to signal
noise. We will not assign a digital crisp value YES/NO to each spectrum: A fuzzy
value α (or probability) of being a peak is assigned in order to simulate human
reasoning. But note that we do not try to simulate the reasoning process of a
human agent, we try to simulate the reasoning of a human expert because we know
that some expert agents are able to distinguish many spectra by a simple look at a
concrete spectrum. Of course these experts will perform the training of our system
by solving some quiz. As a matter of example, the expert may identify the following
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

Spectra analysis in detection of traces of explosives 7

spectrum of a trace of Berilium in Fig 5. It is clear that automated identification

may be impossible by means of absence of peaks.

11000

10000

9000

8000

7000

6000

5000

4000

3000
0 500 1000 1500 2000 2500 3000 3500

Fig. 5. A sample trace of Berilium

Many different parameters were tried in order to know which are the critical
properties of the signal are focused by the expert agent. Best results are obtained
by taking into account:
(1) Where are located the maxima of the slope of the signal? µ
(2) Which is the maximum of the intensity? ι
(3) Once above points (1) and (2) are solved, the significance of the solutions. σ
Once again we assign a fuzzy value µ, ι, σ ∈ (0, 1) to each of the above solutions.
We combine the values α, µ, ι, σ in order to obtain a fuzzy value (or probability)
π = π(α, µ, ι, σ) ∈ (0, 1) of being a peak. Finally we set a critical value π0 in order
to decide a peak if π(α, µ, ι, σ) ≥ π0 and to decide not a peak otherwise.
As usual in fuzzy reasoning it is necessary to state several constants in order to
traslate to fuzzy variables which we understand as big slope, equal intensities, close
peaks, most peaks agree and so on. By means of modifying these constants we get
more or less plasticity for our system. Hence more (or less) samples are classified
and therefore more (or less) false positives/negatives are obtained. Once again we
need a compromise fluency/accuracy for our system.
A nice property of this method is that system is proved to be robust; that is to
say, small perturbations of the parameters give small differences in the results.

4.2. Expert system. Experimental results

We obtain the following results by using the expert system: In the identification of
explosive substances 96%, in the identification of blank substances 87% and in the
identification of mixtures 77%.
 July 2, 2012 14:52 WSPC/INSTRUCTION FILE
MontseHectorMaiteMiguelIJMPB(Rev4a)

8 López-Cabeceira & Al.

Acknowledgements
This research has been developed in collaboration with Indra Sistemas, León within
SEDUCE project framework; partially supported by CDTI (MECC, Spain), MCyT
(Spain), and JCyL (Spain).
1. J. E. Everett et al., in The practical handbook of genetic algorithms. Applications. 2nd
ed., ed. L. Chambers (CHAPMAN & HALL/CRC, Boca Raton, USA, 2001).
2. B. Grard and D. Thierry, “ Neural networks for process control and optimization: Two
industrial applications”, ISA Transactions 42, 2003, pp. 39–51.
3. M. Iglesias, B. Naudts, A. Verschoren and C. Vidal, Foundations of generic opti-
mization. Volume 1: A combinatorial approach to epistasis (Springer, Dordrecht, The
Netherlands, 2005).
4. M. R. Khoshayanda, H. Abdollahib, M. Shariatpanahia, A. Saadatfarda and A. Mo-
hammadi, “Simultaneous spectrophotometric determination of paracetamol, ibuprofen
and caffeine in pharmaceuticals by chemometric methods”, Spectrochim Acta A Mol
Biomol Spectrosc 70, pp. 491–499.
5. T. Kohonen, S. Kaski, K. Lagus, j. Salojärvi, J. Honkela, v. Paatero and A. Saarela,
”Self Organization of a Massive Document Collection”, IEEE Transactions on Neural
Networws 11(3), 2000, pp. 574–585.
6. C. Krafft, G. Steiner, C. Beleites and R. Salzer, “Disease recognition by infrared and
Raman spectroscopy”, J. Biophoton 2(1-2), 2009, pp. 13–28.
7. M. Lee, “Evolution of behaviours in autonomous robot using artificial neural network
and genetic algorithm”, Information Sciences 155, 2003, pp. 43–60.
8. C. G. Looney, Pattern recognition using neural networks. Theory and applications for
engineers and scientists (Oxfors University Press, USA, 1997).
9. Q. Meng, M. Lee, “Error-driven active learning in growing radial basis function net-
works for early robot learning”, Neurocomputing 71(7-9), 2008, pp. 1449-1461.
10. Z. Michalewicz, Genetic Algorithms + Data Stuctures = Evolution programs. 2nd ed.
(Springer-Verlag, New York, 1994).
11. W. F. Pearman and A. W. Fountain III, “Classification of Chemical and Biological
Warfare Agent Simulants by Surface-Enhanced Raman Spectroscopy and Multivariate
Statistical Techniques”, Applied Spectroscopy 60(4), 2006, pp. 356–365.
12. Y. Roggo, P. Chalus, L. Maurer, C. Lema-Martinez, A. Edmond and N. Jent, “A
review of near infrared spectroscopy and chemometrics in pharmaceutical technologies”,
Journal of Pharmaceutical and Biomedical Analysis 44, 2007, pp. 683–700.
13. P. M. Talaván and J. Yáñez, “Parameter setting of the Hopfield networks applied to
TSP”, Journal of Neural Networks 15, 2002, pp. 363–373.
14. M. Urbano-Cuadrado, M. D. Luque-de-Castro and M. A. Gómez-Nieto, “Study of
spectral analytical data using fingerprints and scaled similarity measurements, Anal
Bioanal Chem 381, 2005, pp. 953–963.
15. I. Viuela and I. M. Galván-León, Redes de neuronas artificiales. Un enfoque práctico
(Prentice Hall, Madrid, 2004).
16. J. Weng, “On developmental mental architectures”, Neurocomputing 70, 2007,
pp. 2303–2323.
17. L. Yizeng, W. Hailong, S. Guoli, J. Jianhui, L. Sheng and Y. Ruqin, “Aspects of recent
developments in analytical chemometrics”, Science in China: Series B Chemistry 49(3),
2006, pp. 193–203.
18. A. Agresti. Categorical Data Analysis (Wiley-Interscience, New York, 2002)
19. D.W. Hosmer, S. Lemeshow. Applied Logistic Regression (Wiley, New York, 2000)

View publication stats