29.1   Introduction

This chapter provides background information on the challenges and issues related to detection and identification of biological pathogens and chemical agents in the environment. Although laboratory-based methods can be used for effective identification, these techniques are generally unusable for first responders in a field environment, because they require significant time and expertise of a laboratory technician. The chapter introduces the issues and challenges in detection and identification, and also the concept of using strategies based on biological entities. In particular, an approach using multiple sensors and a fuzzy logic inference procedure is used to mimic how insects detect and characterize biochemical agents. The resulting data fusion technique provides effective detection and identification of complex chemical and biological agents in real-world environments.

29.2   Characterizing the Complexity of Detecting Chemical Agents and Biological Pathogens

Detection of harmful chemical vapors or biological pathogens has obvious value to society but also presents a number of problems. First, automated sensors are needed to provide a detect-to-warn capability. There exists a broad range of laboratory techniques that can separate and uniquely identify harmful agents, but these techniques take too much time and require rare human skills for analysis to be useful in an emergency. To be of use to first responders or those on the battlefield, chemical and biological sensors need to be both real time and completely automated. Second, although it is desirable to have stand-off sensors that can detect an agent without physical contact from a distant position (say hundreds of meters or even a kilometer), these often-active sensors are expensive, may require the use of non-eye-safe lasers, and also require rare human analytical skills to interpret the measured data. Third, we live in an environment with literally millions of chemical vapor mixtures making the false alarm problem, not the detection problem, which is the major concern. Fourth, for biological sensors, the false alarm problem is extremely difficult to detect since all bacteria and viruses are composed of the same amino acids and nucleic acids, but are arranged in different order. However, for both chemical and biological agents, we are most interested in a relatively short list of chemicals and pathogens (in the hundreds) compared to the millions of interferants naturally found in our environment.

To understand a useful strategy for data fusion to improve the performance of automatic detection with a low false alarm rate, one should take advantage of a diverse array of sensor technologies such that each sensor uniquely responds to the presence of, and concentration of, the agents of interest. This does not suggest orthogonal sensors, that is, sensors that only respond to one thing. Even if orthogonality could be guaranteed, it would lead to the need for too many sensors for a practical system. Therefore, our data fusion approach is based on having a diverse array of sensors where for the agents of interest we have a subset of sensors responding with unique sensitivities. Sensor responses will be categorized as qualitative for true-false types of detection and quantitative where the amount of response is functionally related to chemical concentration or the number of colony-forming units (NCU) for biological spores. We can use both kinds of sensor information, although the quantitative ones are more useful because they can be used for corroboration of vapor concentration (or NCU) and rejection of false alarms. This is because the diverse sensors each respond uniquely to a given chemical or biological agent.

Consider how nature has solved the chemical detection and identification problem with olfaction, and the biological detection and identification problem with the immune response system. Olfaction in humans and mammals is quite complex to study, and there are ethical barriers to neurological response experiments on live animals. However, insects have the most sensitive chemical detection capability of all animals and can demonstrate robustness to experimentation. The brain and antenna neurons will function for hours after the head has been removed from the body, and there has been more than 50 years of detailed scientific study to support the development of safer and more effective pesticides.

An excellent reference detailing the biology of insect olfaction can be found in Kaissling.¹ In short, insects have around 60 classes of sensilla, or olfactory sensors, totaling approximately 100,000 sensilla per antenna. Two antennae are used to better associate the chemical plume detected with wind direction to navigate to the source. Each sensilla is visually different but all have one or more pores with chemical coatings to control the types of chemical vapors that can pass on to the neural receptor beneath the pore. The biochemistry involved is truly remarkable, but for our purposes, it is sufficient to note that each sensilla responds a little bit to any large molecule that can pass through the coating. Air comprises mostly small, diatomic (N₂, O₂), and triatomic (H₂O, CO₂, O₃, and so on) molecules due to solar radiation; larger molecules precipitate out from reactions with water and sunlight. The large molecules make their way into the sensilla pore, contact the nerve dendrite, and are then broken down by enzymes and removed. To protect against too much stimulation, there are structures on the antenna and in the pore itself to reduce uptake in much the same way a human iris closes in response to bright light. The sensilla self-cleaning process takes a few seconds whereas the detection process takes only a few milliseconds. In this way, all olfaction sensing has a similar process: fast change detection, slow adaptation to changing backgrounds, and self-cleaning capability. Furthermore, the chemical functionality of each sensilla is not orthogonal, but has an overlapped response based on parameters like solubility, ion mobility, diffusivity, and even mechanical filtering by lipids and proteins. Each sensilla class is neurologically wired to a specific area of the insect’s brain, called the mushroom body, which receives all of the corresponding sensilla neuron pulses in an avalanche when a vapor passes by the antenna. The mushroom body structures on the brain for all the sensilla classes look like a bunch of grapes. Several individual mushroom bodies respond to an odor stimulus with electrical activity and, when the insect recognizes the odor, a projected neural response is seen throughout the rest of the brain. This is often further observed as an associated behavioral response by the insect. This general response is seen for pure chemicals as well as chemical mixtures.

The biological process of detecting chemicals is quite different from the way we approach chemical identification. Generally, one first separates chemical mixtures into the constitutive molecules, and then one systematically identifies each molecule based on properties such as molecular weight, number of carbon atoms, mass spectroscopy, and so on. This is very reasonable for skilled analysts in a laboratory setting, but not easily automated in a real-time detect-to-warn sensor. So what is nature teaching us by the way olfaction sensing happens in insects? It is basically the same process in all life. To mimic the biological system for chemical detection and identification, we can summarize a few requirements:

• Use a diverse set of sensors with overlapping, but unique responses

• Treat pure (neat) chemicals and chemical mixtures similarly in terms of sensor response patterns

• Allow the responding classes of sensors to vary with concentration

• Ensure the sensors are self-cleaning

• Associate the presence and concentration of each specific chemical or chemical mixture with the pattern of responses from the diverse sensor array

Consider the task of pathogen detection.² There are literally millions of bacteria and viruses all around us. Most are harmless to humans and animals, and we have acquired immunity to many so long as the pathogen is of low concentration (low NCU). Our immune response system can give us guidance in terms of developing a sensor and data fusion strategy analogous to the biomimicry of the olfaction system for chemical detection.

When a harmful virus, bacteria, or foreign protein pass through the cell walls of our skin into the bloodstream, it is called a pathogen. All vertebrates have a defense system called the immune response system that produces substances called antigens that support the killing and removal of pathogens. To simplify, there are generally two types of cells involved: B-cells that originate in the bone marrow and T-cells that originate in the thymus gland. The B-cells mark the pathogen, whereas the T-cells destroy it. The B-cells have materials called immunoglobulin molecules (antibodies) that bind to a specific pathogen. When this binding occurs, helper T-cells stimulate the particular B-cell variant to reproduce. Some of these cloned B-cells produce soluble antibodies that freely move about marking the corresponding pathogens they encounter. Other cloned B-cells provide memory for the immune response to facilitate a rapid counterattack should the same pathogen appear in the future. The key fact here is that the natural antigens in our bodies can form almost infinite combinations of amino acids that bind to unique sites on invading pathogens. Much of this natural immunity is part of our DNA but still these antigens can mutate to acquire immunity to a new pathogen. The memory B-cells retain the biological code to fight off frequent reinfections of the same pathogen at a later time.

If one approaches biological agent detection the same way as chemical detection, there will be great difficultly identifying a specific pathogen by means of molecular mobility, solubility, Raman spectrum, or even fluorescence spectrum.^3,* This is because the characteristic protein materials associated with a given pathogen are very large (often thousands of amino acids long), the molecule has many folds and a generally complex structure, and there are large numbers of similar proteins associated with harmless bacteria, viruses, and foreign proteins. These complexities lead to a far more difficult challenge of specificity in biological detection. Given a sample of a suspected pathogen, one can visually identify it either from microscope images or by separating and identifying pieces of the DNA or unique proteins.

For example, the well-known flesh-eating bacteria, Escherichia coli, has a unique protein that protrudes out into the surrounding area of the cell and attacks other cells. Figure 29.1 shows a rendering of the Colicin Ia transmembrane protein that does this dastardly work.

The various ribbon-like sections (peptides) of the protein are drawn with colors and shapes to specifically call attention to biochemical properties, such as binding (or sticking) to other well-known reagents. The protein in Figure 29.1 is a long chain of amino acids, of which there are 20 types. Each type is given a letter symbol (A–V) and the entire sequence for a known protein can be found in many Web sites referred within protein data banks. Table 29.1 provides the sequence for Colicin Ia.

To gain specificity, biologists generally use an array of antibodies where each antibody is attached to a quenched fluorescing molecule, which shows fluorescence only when the antibody binds to the specific pathogen. Dyes are also used where the stain sticks if the pathogen is present and washes away if not. This sticking is actually a weak atomic force where a section of the protein is compatible with the sticking section of the antibody. An array of wells containing the antibody markers is called an immunoassay sensor but is really more like a microscope slide than an electronic chemical detector. One separates the proteins of the pathogen and applies a series of markers to identify the protein. Usually, a bioassay will have a large number of markers such that specific combinations would indicate one pathogen or another. This works well in the laboratory. But in the field, there is a broad spectrum of pathogens along with harmless bacteria in the air. So the real-world test is: can one identify a particular pathogen in a mixture of other biological material? This is why data fusion techniques are of high importance to biological detection.

Another biological identification approach is to isolate the pathogen’s DNA and identify the DNA sequence. This is very often used in criminology or forensic science and is a useful technique to identify a captured pathogen. Unlike proteins, DNA is composed of sequences of only four different nucleic acids and serves the purpose of information storage only. A triplet of any three nucleic acids makes an amino acid. There are two basic steps to identifying a pathogen by DNA. First, one needs to replicate (amplify) the DNA once separated from the cell proteins, using polymerase chain reactions (PCR), to clone a unique section of the DNA for the pathogen of interest. The next step is to identify the sections of interest by measuring the mobility (using electrophoresis) of the DNA sections relative to known DNA. Only a few sections need to match to have high confidence and specificity. However, one drawback is that the PCR process can also amplify DNA from contaminants, leading to false positives.

In short, we have a fairly broad selection of chemical detectors which, if used in combination with data fusion, can provide both low false negatives (false rest, or no response to the actual agent) and low false positives (different sensing modalities will not corroborate false alarms among them). This is consistent with biological olfaction systems. For biological pathogen sensing, there are many more complex challenges even beyond the mechanics of wet biochemistry and pathogen isolation. This is because all biological materials are made of the same chemicals in long chains. The difference between a killer pathogen and bread yeast is only in the order of the amino and nucleic acids. Data fusion will be a very important tool in the automated analysis of biological sensing in real-world environments.

29.3   Chemical Sensors

Our purpose here is to explain briefly the typical types of real-time chemical detectors available to illustrate the challenges for false alarm reduction and automation. It should be clear that a skilled technician could use a host of laboratory equipment to quickly isolate and identify chemicals and mixtures. However, that is not the point here. We need to automate the detection and identification decision using specific concepts in data fusion based on science. Barring any new fantastic sensor technologies such as Mr. Spock’s tri-corder,* we will briefly discuss a collection of typical chemical sensors based on first principles of physics, each of which provides an aspect of the full atomic signature of the chemical. Each sensor alone has vulnerabilities to false alarms, but since they respond differently to the same molecule, each offers important information for fusion into a robust automated detection and classification system. Most of these technologies have been around for more than 50 years and are not likely to ever be replaced. We make no distinction between point sensors, which must physically capture the chemical, or stand-off sensors, which optically interrogate a parcel of air from a distance. The important strategy of our discussion here is to fuse information from a variety of physically different sensors, expecting that a good data fusion algorithm will preserve the best detection performance while suppressing false alarms from each sensor individually.

29.4   Biological Sensors

Since this book is addressed to those not focused on the biological sciences, we need to discuss several important terms from biology to understand better biological sensors and the real information they produce. We will then proceed to discuss the major categories of biological sensors. This is essential to the systems engineering behind a sensor data fusion algorithm for automatically interpreting biological sensor patterns. Biological organisms can divide every 20 min in a fermentation reactor so that a single colony-forming unit (CFU) can become more than 1 billion cells in about 10 h! Sizes of biological materials⁸ are also important to keep in mind. A good approximation to the size of an atom is around 0.1 nm. The air is mostly small diatomic (O₂, N₂) and triatomic (H₂O, CO₂) molecules, whereas chemical agents with 10–30 atoms are about 1–3 nm in size. Proteins are large collections (dozens to thousands) of amino acids and can be more than 1 µm in length, but weak atomic forces will tend to fold and bunch them up like a loose pile of string of approximately 10 nm in size. Viruses are collections of protein, which can pass into a cell and replicate. Most viruses are harmless whereas some interfere with cellular functions, leading to disease and are between 40 and 100 nm in size. Single-cell bacteria tend to be between 1 and 10 µm in size and exist in the air as aerosols of dry spores that can be in clumps of several CFU per particle. Aerosol particles in the 1–2 µm size range are called respirable because the particles are small enough to make it deep into a human lung but large enough to get trapped in the alveoli sacks of the lung, leading to infection.⁹

The goal of biological detection systems is to detect viruses, bacteria, and toxins before they infect or poison humans. This is generally done by detecting the shape of proteins on the surfaces of the virus or bacteria spore, although other methods can also be used. However, biological detection becomes difficult as earth is full of harmless biological materials made of the same amino acids and very similar proteins.

Biological materials of interest are pathogens, which can be viral or bacterial, and biotoxins, which are chemicals produced by pathogens that are toxic. Biotoxins are a crossover threat. They are a toxic chemical that can poison the environment the same way a toxic industrial chemical or chemical warfare agent can, but the difference is that a life form produces the chemical rather than a reaction plant. Examples of biotoxins are Ricin and Botulinum Toxin (Botox).¹⁰ Ricin is produced as a by-product of processing castor beans for oil, whereas Botox is produced by the rod-like bacteria Clostridium botulinum, which itself is harmless. Ricin is one of the most toxic chemicals known and simply kills any cell it comes in contact with. Botox causes paralysis in muscles and small quantities injected under the skin are used medically to remove facial wrinkles (essentially by paralyzing facial muscles). C. botulinum occurs naturally in uncooked foods (such as shellfish) and if ingested it can live in the intestine and is extremely difficult to get rid of, thus leading to severe paralysis and death. Rapid detection of biotoxins is also extremely difficult, but there are laboratory procedures to isolate and identify these agents. These procedures could be automated using data fusion as a decision aid in the future.

Aerosol collection and concentration is used to capture and concentrate particles in a specific size range that exhibit general properties of interest, such as electrostatic charge, fluorescence, or mobility.¹¹ The standard performance measure for these devices is the concentration of agent-containing particles per liter of air (ACPLA). Although for some pathogens, only a single CFU is needed to eventually infect and kill a human, for most pathogens, a sizable number of CFU are needed to be lethal. For anthrax, approximately 10,000 spores are needed to infect a person’s lungs. Anthrax is quite common on livestock farms but farmers are only exposed to small concentrations and their immune response prevents a serious infection. In general, if someone manufactured a biological agent, they would work very hard to have a particle size in the 1–2 µm range by mixing the spores with materials that would neutralize electrostatic charges that naturally would cause the particles to agglomerate into bigger clumps of particles. This would help the spores easily disperse and eventually be trapped in the lungs and quickly infect a person. Biological aerosols can be distinguished from fine dust or metal aerosols (found in subway systems) in that the biological aerosol will almost certainly fluoresce into the visible light wavelength range when illuminated with UV light. The nature of the fluorescence spectrum will vary somewhat by pathogen species,³ but there are always a large number of things that in various combinations can produce a similar fluorescence spectrum. The challenge is to collect and concentrate biological aerosols with properties consistent with a man-made biological weapon and then package the material for further identification processing.

The bioassay is any device used to measure a defining aspect of a biological material. The term has its roots in minting coins, where an assayer certifies the purity of the metal from a series of chemical processes.¹² There are many types of bioassays, the discussion of which is well beyond the scope of this discussion. In general, the bioassay deposits the biological material in whole spores, vegetative bacteria, hosted viruses, enzyme-stripped proteins, or nucleic acids in small wells, each with a binding antigen, stain, or fluorophore* marker that biochemically binds to the pathogen or its components. The concentrated biomaterial must, in general, be preprocessed and distributed onto the assay, which requires subsequent washing to remove unbound markers, and then the assay is evaluated to see which combination of markers were bound to the suspected pathogen. In the laboratory, a homogeneous sample of pathogen distributed in the assay produces a repeatable color pattern from the various markers so that a broad set of markers in a single assay can be used to identify a number of pathogens. But this does not mean that a mix of pathogens can be assayed simultaneously. The assay pattern is based on a single pathogen being present, typically from a cultured sample. However, literally thousands of wells can be placed on a glass slide, each with a specific antigen or marker, and automatically scanned to produce an image for pattern analysis. The intensity of each marker can then be used as qualitative (yes or no) data fusion information or quantitative (gray scale) data fusion information if appropriate.

Techniques such as PCR are used to amplify small amounts of DNA for subsequent analysis. PCR is commonly used to identify individual people through trace amounts of their DNA. Electrophoresis is a DNA detection technique where the mobility of various molecules forced by an electric field allows repeatable separation based on the molecular weight, electric charge of the molecule, and the size of the molecule. This sieving effect allows DNA from one sample to be compared to another, thus allowing a positive identification. Therefore, it is very essential to have a library of DNA patterns for each pathogen of interest and simply compare the unknown sample to see whether there is a match to a pathogen of concern. In a controlled laboratory, this technique for DNA identification works very reliably. However, the challenge is to start with a pure DNA source, because the PDR will amplify everything in the sample. An air sample could contain hundreds or even thousands of DNA sources. One must sort the samples as much as possible to be sure of the results.

The bio-sensor output for data fusion is really the color response from binding markers in a large number of assay wells. These are typically analyzed using neural networks, or simple pattern match scores for the wells that correspond to a particular pathogen class. From a data fusion perspective, these sensor channels can be interpreted qualitatively (true or false) or quantitatively (on a scale). Since there is a broad range of potentially available information from various sensors, the use of fuzzy logic can be employed to combine data such as aerosol properties (particle size, shape, and charge) as well as protein marker tag data. This logically leads one to inferenced-based logic of the form “if this is pathogen X, the set Y of sensor channels should have a matching qualitative response, and the set Z of sensor channels should have corroborating quantitative responses for the given concentration of pathogen X.”

29.5   Developing Quantitative and Qualitative Information

We have discussed very briefly a range of chemical and biological sensor technologies that are necessary to understand to create a scientific basis for sensor fusion. Why do we assume some sort of fusion processing is required? First, we know that no single sensor can respond to everything we are interested in detecting. Not even our own olfactory and immune systems have such a sensor. But rather, biomimicry suggests that a combination of diverse sensors is necessary where a subset of the sensors is all that is needed for any particular chemical or biological agent detection. Figure 29.2 shows a completely general sensor response as a function of concentration.

To understand the difference between quantitative and qualitative information, Figure 29.2 shows four general response ranges. If a sensor has a very narrow concentration range of linear response, we call it a qualitative sensor because its information is really just a true-false indicator. If a sensor has a fairly broad range of linear response, we call it a quantitative sensor because its output over the linear range can be used to accurately estimate concentration or the quantity of the chemical or biological agent present. When fusing multiple sensors, both types of information are very useful, but for different reasons that are explained in the next section.

The lower trace detection limit in Figure 29.2 is the region where the sensor started to respond to the agent. In this region, one generally sets a detection threshold and corresponding false alarm rate for deciding whether the sensor output is signal (agent present) or noise. For a given signal level, one can evaluate a number of detection thresholds (and the corresponding false alarm rates) and detection probabilities to trace out a curve called the receiver operating characteristic, or ROC curve. An ROC curve can be generated for a wide range of signal levels, allowing one to better evaluate the detection threshold for overall system performance. The ROC curve comes from automated detection systems in radar and sonar, which are traditionally low clutter environments (meaning that there is no need for specificity, just detection). There is a lot of interest in ROC curves for chemical and biological sensors because of false alarms, but the real cause for false alarms in chemical and biological sensors is not noise, but a lack of specificity in the detection output of the sensors. However, the lower trace detection limit is an important boundary for the data fusion algorithm to consider using the sensor output, or lack of output, as information.

The upper nonlinear and saturation ranges are less useful for quantitative information because the conversion into concentration is more complicated and much more prone to error. This is particularly true for the saturation region because the flatter slope means a small difference in sensor output that corresponds to a large difference in estimated concentration. Furthermore, in the saturation region, the sensor could become coated with agent and require cleaning before readings would return to normal after the agent has gone. Generally, we use known limits for detection and the linear range to define where the sensor information is reliable. This can be done generally, or for each agent of interest, to give the user an idea of the data quality range. For some chemicals, a given sensor would not respond whereas others may cause saturation and even contamination (permanent saturation). The latter can be used to turn off sensors that might be damaged by high concentrations of certain agents. A good example is smoke detectors that are ruined by volatile chemicals in paint. A data fusion algorithm could only expose such sensors for short periods, if at all, to dangerous chemicals to prevent contamination. Since diverse (heterogeneous) sensors will be affected by contamination differently, data fusion algorithms could be used to identify and mitigate potential contamination and are very important to maintaining sensor health.

29.6   Inferencing Networks for Heterogeneous Sensor Fusion

We have presented a very brief discussion of biomimetics, olfaction, chemical and biological sensing, and qualitative and quantitative sensor information for data fusion. In this section, we focus on the data fusion algorithms directly, based on the science of the sensing. Sensor data fusion can use any operationally appropriate algorithm that combines different information from various sources to enhance confidence in the estimated pattern. For example, a pattern recognition algorithm such as a neural network could use inputs from multiple sensors, and one could train the network to associate various sensor patterns to corresponding situation outputs. This can work quite well when the sensor inputs are very well defined. The problem with chemical and biological sensor fusion is that we are looking for very specific sensor patterns among an approximately infinite number of chemical or biological backgrounds. So in our case, the inputs are not well defined in general, but the inputs for the situations of interest are so. One might consider a statistical classifier, or Bayesian algorithm, where the sensor outputs are part of a feature vector and one attempts to detect the situation pattern by a statistical measure of the feature vector relative to the mean and covariance of a corresponding training set. This, too, is problematic because the sensor amplitude responses can be nonlinear with concentration, as well as vary by agent or agent mixture. The third approach to pattern recognition is called a syntactic algorithm,¹³ or rule-based approach. Although neural networks and Bayesian classifiers are rich in mathematics and very satisfying to derive and evaluate, the rule-based approach is often overlooked, primarily because one generally does not know the rules or syntax needed. In our case, for chemical and biological data fusion, we are seeking a consistent sensor response for physically diverse sensors on a given agent. Since the sensors are each measuring an aspect of the agent that is unique, we fully expect each sensor output to be uniquely representative of the agent observed. We will give ourselves a little wiggle room in the rules for the syntax by using fuzzy, rather than hard, logic for the situation pattern recognition. This will allow variations in sensor quantitative calibration and low levels of interference from other chemical or biological materials that get into the sensor signals.

The Inference or question for our syntax rules to answer is simply:

If we are observing agent X at concentration Y, are the diverse sensor signals consistent and at what concentration?

This straightforward question makes it relatively easy to design fuzzy logic to fuse information from a variety of sensor sources to reject false alarms of agents that are not the chemicals or pathogens of interest. This rule applies for situations where at least some of the sensors respond to the agent in question. It is not an algorithm that enhances SNR by integrating responses across sensors. We will use both quantitative and qualitative sensor information, and combine it in a way that will only permit false alarms to slip through at very low concentrations, where only the qualitative information is present. As the concentration is increased, more sensor channels report, and the quantitative information then dominates the false alarm rejection because the sensors must report a consistent concentration to pass the syntax. This is an effective means to reject false alarms without trading off detection sensitivity because the criterion narrows as the concentration increases. A reported false alarm at a very low concentration is much less of a nuisance than a false alarm at a higher concentration, which requires immediate action.

The inference approach to fuzzy logic is not an obvious biomimicry of our own thought process. We are taught by the scientific method to collect information, apply known reduction methods, and evaluate results for a conclusion. This is actually a very complicated task to make software perform. However, if we turn the process around and say “if conclusion X is true, what must the sensor inputs be?” the algorithm amazingly forms a well-defined logic network that is generally quite simple to implement in software.¹⁴ We are focused on this approach because both chemical and biological sensors suffer from very high false alarm rates and are exposed to an approximately infinite range of background clutter, but can detect a limited number of agents of interest at concentrations low enough to be useful as a warning system. We will discuss chemical and biological sensors in the same general terms of either qualitative or quantitative information and present a general fuzzy continuous inference network logic algorithm (CINet), that actually can extend generally to a broader range of sensors.

To describe our CINet in the most general terms, we will use hypothetical data from three different sensors for a range of calibrated exposures to two different chemicals. Figure 29.3 shows the measured responses versus concentration, normalized to a common scale of 0–1.

Figure 29.3 shows good qualitative information from sensor 1 and good quantitative information from sensors 2 and 3. The fact that each chemical will produce a unique response for physically different sensing methods (IMS, PID, SAW, spectroscopy, etc.) is what we are exploiting to recognize a target chemical and reject false alarms from clutter chemicals that may fool some or all of the sensors.

Figure 29.4 arranges the sensor data by chemical. Now a number of simple possibilities can be easily defined. For example, if only sensor 1 is responding, it could be a low concentration of chemical A, but not B. If only sensors 1 and 2 are responding, it could be a medium concentration of chemical A, but not B at any concentration. If only sensor 3 is responding, it could be chemical B at a low concentration, but not A, and so on. At higher concentrations where all three sensors are responding, the estimated concentration from the two quantitative sensors (sensors 2 and 3) should be used to compare consistency. Since their concentration slopes are quite different, they will only agree over a range of concentrations when the target chemical is present. It should be noted that for a small set of sensors and chemicals, one might find a few specific concentrations where an ambiguity exists. However, these ambiguities would be known in advance from analysis and the actual concentration for a real measurement would naturally fluctuate significantly over time, making this false alarm a rare and quantified event.

29.7   The Path Forward

This brief look at a methodology to perform heterogeneous sensor fusion for chemical and biological agents is based on using a diversity of sensors that are well calibrated for the target agents of interest. It can be seen that a sensor manifest or agent library is then needed containing the calibration SRM tables and response thresholds for each agent in the library. This would allow a remote process to fuse the information from an array of sensors exposed to the agent in question. The inference-based logic can then employ only those sensors and sensor channels with useful responses for a given agent inference, and each sensor’s contribution to the overall fusion confidence would be independent of the other sensors. If the sensor has a response to the agent, it contributes to the inference logic by providing context for the sensor signal output. As signals from multiple sensors are fused, this context looks for consistency using fuzzy logic metrics. For small concentrations of a given agent, there is a higher possibility of a false alarm, but since the concentration is small, this is much less of a concern compared to higher concentrations that carry greater danger to people exposed. Higher concentrations would naturally bring in more sensors and sensor channels with higher SNR signals. The CINet logic captures this by making it more and more difficult for a false alarm to persist because of the ability to look for consistency across the sensor information.

This approach to false alarm rejection may be most interesting in the medical applications area, where qualitative signals from bioassay wells can be combined with more quantitative metrics such as volatile chemicals in the blood and breath. The CINet logic can automate many metabolic states based on highly objective metrics and calibrated sensor response libraries. The final diagnosis/prognosis would still require a medical doctor or pathologist, but the degree of sensor automation will likely produce very good decision aids. For example, current technology in 2007 can measure more than 1000 blood chemicals. Bioassays can have more than thousands of wells with binding proteins that fluoresce to varying degrees. It would not be difficult to also execute thousands of well-thought-out CINets to quickly produce a short list of likely situations. This methodology is only an automation of information processing that parallels human knowledge from the scientific method. But, as with human knowledge, we need diversity of information, repeatable calibration, and a reasoned syntax to fuse the information into a situation recognition.

References

1. K.E. Kaissling, Physiology of pheromone reception in insects, ANIR-AVNP, 6(2), 2004, pp. 73–91.

2. C.K. Mathews, K.E. van Holde, and K.G. Ahern, Biochemistry (San Francisco: Addison Wesley Longman, 1999) 3rd ed., pp. 241–244.

3. P.J. Hargis et al., UV detection of biological species, Edgewood Chemical Biological Center, ECBC Report Number CR-028, August 2000.

4. G.A. Eiceman and Z. Karpas, Ion Mobility Spectroscopy (Boca Raton: Taylor and Francis, 2005).

5. Y. Sun and K.Y Ong, Detection Technologies for Chemical Warfare Agents and Toxic Vapors (New York: CRC Press, 2005), pp. 184–194.

6. Q. Cui and I. Bahar, ed., Normal Mode Analysis: Theory and Applications to Biological and Chemical Systems (Boca Raton: Chapman and Hall/CRC, 2006).

7. E.B. Hanlon et al., Prospects for in vivo Raman spectroscopy, Phys. Med. Biol., 45, 2000, R1–R59.

8. C.K. Mathews, K.E. van Holde, and K.G. Ahern, Biochemistry (San Francisco: Addison Wesley Longman, 1999) 3rd ed., p. 14.

9. W.C. Hinds, Aerosol Technology (New York: Wiley, 1999).

10. J. Ali, L. Rodrigues, and M. Moodie, Jane’s US Chemical-Biological Guidebook (Coulsdon, Surrey, UK: Jane’s Information Group, 1998), Chapter 3.

11. C.S. Cox and C.M. Wathes, Bioaerosols Handbook (Boca Raton: CRC Press, 1995).

12. http://en.wikipedia.org/wiki/Assay “Assay” definition.

13. R. Schalkoff, Pattern Recognition: Statistical, Structural, and Neural Approaches (New York: Wiley, 1992).

14. J.A. Stover and R.E. Gibson, Modeling confusion in automated systems, SPIE, vol. 1710, Science of Artificial Neural Networks, 1992, pp. 547–555.

15. D. Swanson, Signal Processing for Intelligent Sensor Systems (New York: Marcel-Dekker, 2000), pp. 478–489.

* These are discussed in detail in Section 29.3.

* The science officer from Star Trek could detect anything with this device under any false alarms. With great prejudice we interpret the tri in tri-corder to imply automated data fusion with at least three sensor channels.

* A fluorophore is a material that fluoresces only when it is successfully bound to a target protein.