6.10 Neural Network Chromatogram Retrieval System: A Case Study

6.10 Neural Network Chromatogram Retrieval System: A Case Study

During the course of researching this book, many ingenious and dramatic uses of data mining technologies were found. The following case study (presented in its original version) is a very important and new application of neural networks in the analysis of physical evidence as part of a forensic search for the signature of chemicals in an arson investigation. The criminalist that developed this quantitative analysis readily admits that it could not have been performed without the assistance of neural network technology. It represents a unique use of advanced-pattern recognition technology in the context of traditional forensics analysis. In this case study, a neural network is used to detect and discover from a clinical chemical forensic analysis the distinct profile of ignitable kerosene.

NNCRS

Developed by Matt Vona, Criminalist, Bureau of Forensic Service, California Department of Justice

Summary

This research will develop a Neural Network Chromatogram Retrieval System (NNCRS), an automated means of chromatogram profile (or chemical fingerprint) comparison and retrieval unlike any that currently exists. This technology will be used in the comparison and retrieval of ignitable liquid profiles. Ultimately, this technology could be applied to more than arson analysis. The technological advances made here could lay the groundwork for applications used for explosives, other chemical profiles, impression evidence, infrared spectrometry, near infrared, and even biological protein or macromolecular signature analysis. The ability of neural networks to judge degrees of similarity based on pattern recognition while remaining independent of scaling, rotation, or symmetry will allow us to comprehensively reference an unlimited database of chemical signatures. Neural networks have already been applied and tested elsewhere in mainstream industry. It is now time that forensic applications are able to gain from these advances.

This white paper details a research proposal for exploring the application of NNCRS to volatile liquids analysis. We will develop this neural-network—based system of chromatographic signature comparison and retrieval. Simultaneously, through a co-operative, we will develop a system of collecting and redistributing a standards database from a large number of laboratories for use in case work comparison. Ultimately, we will have developed a deliverable NNCRS product, capable of being plugged into many applications across many forensic disciplines.

Background

The number and variety of ignitable liquids encountered in case work is increasing, as new and varied petroleum products enter the market place. The matrixes, or background materials, are often manufactured using the same volatile/ignitable liquids analysts detect. New adhesives, plastics, flooring, and other products are among many being introduced to the market place, which contain trace amounts of ignitable liquids. Recently a study presented at the Technical Working Group for Fire and Explosives (TWGFEX) symposium in Orlando, Florida demonstrated that some forms of newsprint contain ignitable liquid components.

The significance of an ignitable liquid being present in any fire debris sample is becoming increasingly unclear to the forensic analyst. In volatile liquid proficiency tests administered to 44 arson analysts located around the nation by the Proficiency Test Program of the International Forensic Research Institute at Florida International University in May of 2001, 18 of 44 analysts reported the presence of an ignitable liquid when none was present. These errors occurred largely because many analysts detected trace components of a miscellaneous ignitable liquid in an unadulterated sample of linoleum. While the matrix present in the sample may have contained these components, the analysts were not able to differentiate between a background contaminant and the presence of an additional ignitable liquid.

An NNCRS would allow analysts to compare the ignitable liquid signature produced by their unknown to an exhaustive external reference database for comparison. The NNCRS would be capable of sorting through thousands of standard submissions collected on any number of analytical systems from around the world. The neural network could then alert analysts to the possibility that their unknown profile is similar to the pyrolysis products of a particular type of matrix or even the low weight profile of a little known mixture. Analysts taking the proficiency mentioned above would have been alerted to the potential contribution of the linoleum matrix to their results and, thus, avoided writing a misleading report.

The laboratory analysis of fire debris typically involves the following steps:

1. Obtain a sample of the fire debris, which is believed to contain residues of an ignitable liquid.

2. Use one of several methods to extract the ignitable liquid from the debris.

3. Analyze the extracted ignitable liquid on a gas chromatograph, mass spectrometer.

4. Interpret and classify the unknown ignitable liquid through comparison of the results to a comparable standard.

Because of variations that could occur in steps 1 to 3, the data produced from any one standard or unknown is never identical from one laboratory, analyst, or instrument to another. Neural networks, like analysts, are able to recognize similarities. This unique ability allows them to recognize patterns independent of slight changes in instrumentation or despite the presence of pyrolysis products. Because they are automated, neural networks are simultaneously able to comb through a very large number of possible references and return a small number of likely matches to the analyst.

Research History

Our research began as part of the California Methamphetamine Signature Program. Chromatograms, unlike mass or infrared spectra, cannot easily be searched using conventional technology. Because of retention time changes, impurities, and variations in column length, two chromatograms of the same mixture of substances will often differ. This difference confuses conventional technology. Thus, a computer program was developed which was capable of searching a large database of chromatograms versus an unknown chromatogram.

Arson was chosen as the field with which to initiate this research because of the large amount of readily available data. Neural networks were explored because of their ability to match patterns despite a large amount of actual difference and signal noise. A neural network application was developed, specifically to look for matching chromatograms. The neural network was tested by giving it an exemplar chromatogram and having it search a reference database of over 400 reference chromatograms. The preliminary results were very exciting. Figure 6.8 illustrates the exemplar chromatogram, Figures 6.9 and 6.10 illustrate the matching reference chromatograms, and Figure 6.11 illustrates the next closest nonmatching reference chromatogram.

Figure 6.8: Example given to the neural network. The C-12 denotes the position of dodecane.

Figure 6.9: One of the two matches found by the neural network. The C-12 denotes the position of dodecane.

Figure 6.10: A second match found by the neural network. The C-12 denotes the position of dodecane.

Figure 6.11: The closest non-match found by the neural network. The C-12 denotes the position of dodecane.

The neural network found an identical chromatogram and a very "closely related" chromatogram as possible matches out of 400 other unrelated chromatograms. By "closely related," we mean that the unknown's profile was very similar to the example. This is important, because other variations due to different instrumentation and sampling techniques from lab to lab will not be masked from the computer. The neural network can identify all "closely related" chromatograms and give the analyst a chance to further evaluate a smaller list of possible matches more thoroughly.

Notice that the second matched chromatogram (Figure 6.10) only extends out to C14. One may see this variation in different samples, but also in the same sample analyzed at two different labs. The neural network has demonstrated the ability to recognize similarity. Most computer programs could only match a chromatogram to an identical chromatogram. Further, the network's next closest nonmatch (Figure 6.11) was also a kerosene petroleum distillate type, but it was sufficiently different to be excluded from any further serious consideration.

The positive impact of the neural network's ability to make subtle distinctions without relying on identically formatted digital information is far reaching.

One of the first logical consequences for arson analysis is that the neural network could recognize a weathered sample as being similar to a nonweathered sample. For instance, the "unknown" chromatogram on which the network was trained (Figure 6.8) shows a distillate distribution with the highest peak of C11. The second matched chromatogram (Figure 6.10) shows a distillate with the highest peak at C12. Such variations can occur with weathering and other variations and could be readily explained or excluded by a trained analyst. It is important to note that while this technology will assist an analyst in comparing the sample to a large database, it can never operate independently of an experienced analyst.

A second consequence could be that the neural network could recognize that two samples are similar, even though different techniques were employed to collect data from one volatile liquid. As an example, the first matched chromatogram (Figure 6.9) shows a slightly left-tilted Gaussian peak distribution with a hydrocarbon peak distribution from about C9 to peak C17, and the highest peak at C11. The second matched distillate (Figure 6.10) doesn't extend out quite this far. Such variations can occur with differences in sampling techniques. Remember that the neural network's purpose is merely to narrow down a much larger list of chromatograms for the analyst to review, so it would preferably include too many exemplars, as opposed to too few.

This preliminary assessment has demonstrated a basic proof of concept. A neural network can recognize the subtleties in patterns generated by volatile substance analysis. While this test is not in any form a complete study, it has demonstrated a capability, which must be further evaluated by this research project.