RU2659712C1 - Method for the identification of small concentrations of explosive and narcotic substances in the air based on analysis of bioelectric potentials of rat olfactory analyzer - Google Patents

Abstract

FIELD: security facilities.

SUBSTANCE: invention relates to the field of safety and gas analyzers, namely to methods for detecting explosive and/or narcotic substances in the air. Invention is based on the analysis of ECoG of signals taken by electrodes implanted in the rat brain. At the first stage, the multilayer neural networks used are trained. This is ensured by the delivery of a certain odor, by removing the ECoG signal, by appropriately converting the signal and recording it. Result of training is the generation of an array of vectors consisting of an odor and the corresponding elements of the ECoG signal. Immediately during the measurement process, ECoG signals taken by electrodes are compared with the use of training systems and neural networks with the previously obtained array of vectors and a detectable odor is determined.

EFFECT: invention provides an improvement in accuracy, while eliminating the need to stimulate the brain of the rat in the learning process.

6 cl, 21 dwg, 2 tbl, 1 ex

Description

Technical field

The invention relates to the field of secretive detection with high sensitivity of narcotic substances with a concentration of not more than 0.1 μg / m ³ and explosives of the order of 0.01 μg / m ³ and can be used for security purposes at airports, railway stations, stadiums and other crowded places .

State of the art

Ensuring security is one of the priorities of each state, and timely identification of places with potentially hazardous substances such as explosives or drugs is a necessary element of security. There are various methods for detecting such substances by technical devices that are based on both chemical and physical methods of determination. The most widely used instruments are those that use analysis methods such as gas chromatography, ion-drift spectrometry, mass spectrometry. As indicated in the article by A.V. Kikhtenko and K.V. Eliseeva (Kikhtenko A.V., Eliseev K.V. Detection of explosive objects: hardware of anti-terrorist services // Russian Chemical Journal. 2005. T. XLIX. No. 4. p. 132-137) [1], most of these devices have sufficiently high sensitivity and the limit of their detection of the specified substances is 1000-0.01 μg / m ³ and the analysis time of the sample takes from several to hundreds of seconds. In the case of analyzers based on ion mobility spectrometry (RU 2471179 IPC G01N 27/64, publication date 2012-12-27) [2], (RU 2360242 IPC G01N 30/96, publication date 2006-01-26) [3] , (RU 2231781 IPC G01N 27/62, publication date, 2004-06-27) [4], (RU 2398309 IPC H01J 49/40, G01N 27/64, publication date 2009-07-16) [5], ( RU 2354963 IPC G01N 27/62, G01N 30/64, publication date 2007-11-08) [6], (RU 2329563 IPC H01J 49/40, G01N 27/64, publication date 2006-12-25) [7] , (RU 2460067 IPC G01N 27/62, publication date 2011-04-20) [8] (application WO 2008049488 A1 IPC G01N 27 / 62A, H01J 49 / 10B, H01J 49/16, G01N 27/64, publication date 2012 -06-20) [9] the analysis time is 1-15 s, and the sensitivity is up to 1000 μg / m ³ . When using gas chromatographs (RU 2395076 IPC G01N 27/64, publication date 2009-03-23) [10], (RU 2431212 IPC H01J 49/40, publication date 2010-07-09) [11], (CN 2622705 Y , IPC G01N 30/72, H01J 49/26, publication date 2003-05-20) [12], (US 6470730 B1, IPC G01N 1/00, G01N 1/02, G01N 1/22, G01N 1/28, G01N33 / 00, publication date 2000-08-18) [13] the sensitivity increases to 0.1 μg / m ³ for the substances to be determined, but the analysis time increases to 20-200 s, and preliminary sample preparation is also required. It should also be noted that significant disadvantages of gas analyzers and gas chromatographs are the dependence of the accuracy of measurements on temperature and humidity.

Along with technical devices, animals are actively used to detect explosive and narcotic substances (Frederickx C., Verheggen FJ, Haubruge E. Biosensors in forensic sciences // Biotechnol. Agron. Soc. Environ. 2011. Vol. 15. No. 4. P. 449-458) [14]. Most often, service dogs are used for these purposes, which are trained to signal the presence of the smell of the desired substance in the air by certain behavioral reactions in the form of barking, stance, signaling behavior (RU 2426141 IPC G01S 1/02, publication date 2010-03-09) [15] , (RU 2338175 IPC G01N 1/22, А01К 15/00, publication date 2006-09-29) [16]. However, despite the fact that these animals have proven themselves in detection, their use is associated with significant economic and organizational costs associated with the conditions of detention and training, which requires a specialist canine specialist and does not provide secretive detection. Therefore, the scope of their application is limited and depends on the state of the animal and its current needs. Given this, at present, more and more actively a search is being made for other macromatical animals that can be used for this purpose.

The most famous examples are:

- use by a Belgian non-governmental organization (AROPO Anti-Persoonsmijnen Ontmijnende Product Ontwikkeling “Development of anti-personnel mine detection products”, https://www.apopo.org/en/componentyone/summarv?show=265, publication date 2014-09-19) [ 17] hamster (Cricetomys gambianus) rats for search and detection of mines in Africa, Thailand and Cambodia;

- the use of rats as biological sensors for explosives by the TAMAR Group (Israel, Web: http://news.sky.com/story/1010634/sniffer-mice-used-to-detect-bombs-and-drugs, publication date 2012-11-13) [18];

The disadvantage of these methods is the use of conditioned reflex training of animals, which requires time and trained specialists.

Along with mammals, the use of insects for the detection of narcotic and explosive substances is known. For example, bees are used to detect TNT, RDX, plastic explosives based on it, as well as drugs - methamphetamine and cocaine (application US 2005009444 A1, NKI 449/27, 119/420, 119/65, 119/421, IPC G01N33 / 00, publication date 2002-12-20) [19], (Rains GC, Tomberlin JK, Kulasiri D. Using insect sniffing devices for detection // Trends Biotechnol. 2008. Vol. 26. No. 6. P. 288-294 ) [twenty].

Also known devices, which are a biotechnological system, a component of which is a bee and a technical device that allows you to evaluate its behavior (Detection of explosives using bees Web: http://www.pergam.ru/articles/biogasanalizer.htm, publication date 11/28/2013 ) [20], (application US 20120264353 A1, NKI 449/2, 449/1, IPC A01K 51/00, A01K 55/00, publication date 2011-04-04) [22]. However, the usefulness of use is doubtful in view of the complexity of training, the dependence of the behavior of bees on the time of year.

М.М., Manso A.G., Orellana C.G.O.,

Velasco H.M., Caballero R.G. and Peguero Chamizo J.C. Acetic Acid Detection Threshold in Synthetic Wine Samples of a Portable Electronic Nose//Sensors. 2013. Vol. 13. P. 208-220, doi:10.3390/s130100208) [23] моделирующий обоняние собаки (Dog-inspired scent detfector sniffs out explosives and narcotics Web:http://www.gizmag.com/electronic-nose/25128/, дата публикации 21.11.2012) [24].It is known to use devices of the "electronic nose" type (

M.M., Manso AG, Orellana CGO,

Velasco HM, Caballero RG and Peguero Chamizo JC Acetic Acid Detection Threshold in Synthetic Wine Samples of a Portable Electronic Nose // Sensors. 2013. Vol. 13. P. 208-220, doi: 10.3390 / s130100208) [23] simulating the dog's sense of smell (Dog-inspired scent detfector sniffs out explosives and narcotics Web: http: //www.gizmag.com/electronic-nose/25128/, publication date 11/21/2012) [24].

The disadvantages of devices of this type are the technical complexity of their creation, the dependence of sensitivity on changes in temperature and humidity, as well as the "aging" of sensitive elements of the system.

At the present stage, methods for detecting low concentrations of explosive and narcotic substances in the air based on the analysis of the bioelectric potentials of the olfactory rat analyzer are promising. It is known that animal macromatics, in particular rats, have an extremely high olfactory sensitivity. Rats determined trinitrotoluene pairs at a concentration of 2.44 μg / m ³ with a probability of 100%, at a concentration of 1.32 μg / m ³ - 95%, and at a concentration of 1.07 μg / m ³ - 60% with virtually no false positives (Yinon J., Zitrin S. Modern methods and applications in analysis of explosives. Chichester, UK; New York, USA: Wiley. 1996, 316 c.) [25].

Existing technologies include remotely controlled animals that can move through difficult terrain; unlike robots, they do not need to recharge their batteries. The animal is supplied with devices with electronic communication devices and sensors, and is controlled using the remote brain stimulation (application US 2003/0199944, IPC, MPI publication date)) [26], application US 2013098310 A1 IPC A01K 1/03, A01K 15/02 , A61N 1/36, A61N 1/372, B62D 63/02, G06N 3/06, H02H 1/00, publication date 2013-04-25) [27], in which a remotely controlled animal learns to search for target odors and can autonomously carry electronic sensors to inaccessible or dangerous places for various missions, including searching for people buried in a pile of rubble, right -enforcement operations and the detection of explosives, chemicals and other hazardous materials. This invention is aimed at solving the above and other problems by creating a method and device for remote control, guidance and training to detect the smell of a freely moving animal through stimulation of its brain. In one possible design, sixteen electrodes implanted in the brain of an animal are used. Electrodes are implanted in the ventral region of the tire or other area of the lateral hypothalamus to stimulate the pleasure center when learning odor recognition.

However, the known method is associated with the development of conditioned reflexes of the animal to receive rewards.

A known measurement system and a unit for recording an electroencephalogram (EEG) in small animals (rats) using a plurality of electrodes that cover either the scalp or the surface of the brain (WO 2011132756 A1, IPC A61B 5/0478, A61B 5/0476, A01K 67/00 , А61В 5/055, А61В 5/0408, АВВ 5/0492, publication date 2011-10-27) [28]. The system is used for non-invasive registration of EEG simultaneously with functional magnetic resonance imaging (fMRI), as well as an electrocardiogram (ECG) and infrared spectroscopy.

A device for screening for the content of a drug / drug component is known, in which the olfactory mucosa of the body is stimulated in such a way that physiological functions are corrected through directed activation or suppression of brain functions (EP 1234540 A1, IPC A01K 1/03, A01K 29/00, A61B 5 / 0478, AB 5/0484, G01N 33/50, publication date 2002-08-28) [29].

The screening apparatus measures the stimulation pattern of the olfactory bulb, which is formed when the olfactory mucosa is stimulated by a substance that is a candidate component of the drug and is introduced to the olfactory mucosa of the body. The screening device then analyzes the stimulation pattern in such a way as to check the correlation between the stimulation pattern and the physiological response caused in the body by a substance that activates or inhibits the brain through stimulation of the olfactory mucosa.

The electrode matrix is mounted on the olfactory bulb of the experimental animal.

The measuring electrode contains a substrate formed of an insulating film and sixteen microelectrodes, which are located on the surface of the substrate in the form of a 4 × 4 matrix. The substrate has a thickness of 1 μm to 100 μm and is formed in the form of a square, each side of which is about 2 mm. The step between a pair of adjacent microelectrodes is about 500 microns. A method of treating decreased olfactory sensitivity includes the steps of: administering a component of a substance that stimulates the olfactory bulb; measurement of the electrical signal generated in the olfactory bulb of the experimental animal when the component of the substance acting on the olfactory mucosa controls the olfactory lining of the experimental animal to obtain an electric signal pattern: determining the correlation between the electric signal pattern and the type and level of physiological response induced in the experimental animal electric signal pattern; and supplying a pattern of an electrical signal that is sufficient to generate the expected physiological response to the olfactory bulb of the test animal in the form of a stimulation pattern. The device is used to test pharmaceuticals.

A known system and method for detecting brain activity in response to a chemical stimulus involves implanting an electrode into a subject’s brain and measuring the amplitude of the corresponding brain waves before and after the introduction of a chemical stimulus (WO 2005037100 A1, IPC A61B 5/0484, A61B 5/04847, A61B 5 / 4011, publication date 2005-04-28) [30]. A chemical stimulus is a compound or mixture of compounds of microbial, bacterial or phytochemical origin. A chemical stimulus that produces the desired change can be added to pharmaceutical, food, cosmetic or other industrial products consumed by humans or animals, or used to increase brain sensitivity. This is especially beneficial for taste and odor dysfunctions. The system uses at least two measuring electrodes: one in the structures of the limbic system (hippocampus or amygdala), and the other in the orbitofrontal cortex or region of the ventral casing, the reference electrode is implanted in the pyrifiform cortex or olfactory bulb of the brain of a rat or mouse. The amplitudes of alpha, beta, gamma, delta, or theta brain waves are measured. The test substances are administered orally or nasally using a special module. The system comprises means for processing and analyzing the correlation between the measured electrical signal and brain activity. The chemical stimuli presented are smell or taste. The system is used for therapeutic purposes.

The closest analogue to the implementation and the achieved result to the present invention is a drug detector based on the sense of smell of an animal (CN 1865996, IPC AB 5/00, G01N 33/00, publication date 2006-11-22) [31], taken as a prototype, which includes: a CCD camera, a micro electrode matrix, a reference electrode attached by screws to the animal’s skull, a preamplifier, a filter, a four-sleeve labyrinth for smelling rats and a computer control and analysis system. The electrical activity of the neurons of the olfactory cortex of the animal produces specific memory for a volatile substance located in small areas of the labyrinth. At the training stage, the animal is placed in the central area of the labyrinth, the test substance is placed in the chamber at the end of one of the arms; if the camera is correctly selected, the rat receives stimulation of the pleasure center with simultaneous feeding of food. Monitoring the behavior of the animal is a CCD - camera connected to a PC. The electrical activity of the rat is recorded when the location of the volatile substance is correctly selected on the data acquisition card. Under real conditions, rat EEG is compared with the EEG database and, based on a coincidence analysis, concludes that there is a drug in the test object.

The disadvantages of this method are the complexity and complexity of its implementation, due to the need for preliminary training of the animal with food reinforcement, visual observation, additional stimulation of the pleasure center and the lack of data on the sensitivity of the method for detecting a given substance, which depends on the concentration of vapor of the substance and the distance to the olfactory rat analyzer.

The objective of the present invention is to develop a simpler method for detecting low concentrations of 0.01 μg / m ^{3 of} explosive and narcotic substances in a concentration of 0.1 μg / m ³ based on the analysis of bioelectric potentials of the olfactory rat analyzer.

The simplification of the method is achieved by eliminating the preliminary training of the animal with food reinforcement, visual observation and additional stimulation of the pleasure center.

Increasing the sensitivity of the method is ensured by eliminating the analysis of conditioned reflex activity when registering bioelectrical activity from the surface of the olfactory rat bulbs, stimulated by air with a low concentration of narcotic or explosive and an original algorithm for classifying patterns of induced brain activity.

Also, the increase in sensitivity is provided by a special, claimed method of analysis of signal components and the use of machine learning systems.

From the prior art, it is possible to detect volatile matter in rat rat olfactory receptors in air with concentrations of more than 1 μg / m ³ (Gupta P., Albeanu DF, & Bhalla US Olfactory bulb coding of odors, mixtures and sniffs is a linear sum of odor time profiles / / Nature Neuroscience. 2015. doi: 10.1038 / nn. 3913) [32], which confirms the achievement of a new technical result by the claimed invention.

List of Figures

The essence of the method for detecting low concentrations of explosive and narcotic substances in the air based on the analysis of bioelectric potentials of the olfactory rat analyzer is illustrated by the figures of the drawings.

FIG. 1 is a flowchart that implements the method.

FIG. 2 is a block diagram of a processor for processing rat ECOG signals.

FIG. 3 - Block diagram of the training of neural network classifiers.

FIG. 4 - The block diagram of the storage of time periods of ECoG for each smell.

Figure 5. - An illustration of the selection of a time interval of an ECoG with a breathing signal of duration τ with a delay Δt after the start of inspiration t ₀ .

FIG. 6 - A block diagram of the formation of training arrays of input vectors {a _j } and output string variables {b _j }.

FIG. 7 is a block diagram of a preprocessing array of input vectors {α _j }.

FIG. 8 is a block diagram of a preprocessing array of output string variables.

FIG. 9 is a flowchart of a backpropagation algorithm.

FIG. 10 - Block diagram of a multilayer neural network.

FIG. 11 - Functional diagram of the nodal element of the hidden layer and the output layer of the multilayer neural network.

FIG. 12 is a flowchart for performing straight-through calculations of the output vectors θ.

FIG. 13 is a flowchart for performing back-flow error calculations Δw of the weight coefficients w and their modifications w = w + Δw.

FIG. 14 - Illustration of the results of the linear and non-linear classifier, where: a - non-linear arrangement of examples of two odors depicted by points of light and dark color, b - use of a linear classifier that is not able to separate the areas of belonging of examples of different odors, c - use of a non-linear classifier that allows to separate areas of belonging are examples of different odors.

FIG. 15 is a block diagram of a neural network classification.

FIG. 16 - The block diagram of the memorization of the ECOG time interval for recognition.

FIG. 17 is a block diagram of the averaging of output vectors of all multilayer neural networks.

FIG. 18 is a flowchart for determining rat odor.

FIG. 19 - Graphs of electrical activity generated in a rat in one session of presentation of trinitrotoluene vapor (0.1 μg / m ³ ).

Figure 20 is an arrangement of trepanation holes, screws and ligatures for mounting a microelectrode block on a rat head.

FIG. 21 is a diagram of an embodiment of the method.

Disclosure of invention

A method for detecting low concentrations of explosive and narcotic substances in the air based on the analysis of the bioelectric potentials of the olfactory rat analyzer, which consists in implanting a microelectrode matrix with eight working and one reference electrode into the upper surface of the olfactory bulb of the rat, recording the electrocorticographic (ECoG) signal of the olfactory bulb in a given frequency range at the time of inspiration, the extraction of five groups of signs, expressed as mathematical values of mean, variance, asi metrics and excesses calculated for the amplitudes of the ECoG signal of each lead, the first derivative of the amplitudes of the ECoG signal of each lead, the frequency-amplitude spectrum of each lead, the cross-correlation coefficients of the amplitudes of the ECoG signal between the leads, the cross-correlation coefficients of the frequency-amplitude spectrum between the leads, the processing of each groups of signs of a separate multilayer neural network (MNS), while for the training of each MNS, an array of odor indicators presented to the rat is given an additional number A set of times with a given pause duration, which is the output array for MNS training, during training for each MNS, the classification weight coefficients are calculated using the back propagation algorithm and, when identifying a smell, five groups of ECOG signal signs process five MNSs and calculate the probabilities of the rat being presented with one odor from the given odors for each group of signs, calculate the arithmetic average of the probabilities from five MNSs with the choice of smell with a maximum probability as p recognition result.

In special cases, the execution of the method:

- the number of presentations of each smell when teaching MNF is 60 times;

- the duration of the pause between presentations of odors during training MNF is 20 seconds;

- the frequency range of odor detection in the ECOG signal is 50-140 Hz.

The number of presentations of one odor 60 times is a prerequisite for training a neural network to classify an odor.

The pause duration of 20 s corresponds to the recovery time of receptors in the rat nasal cavity between presentations, which is due to the physiology of the rat.

The specified frequency range provides the optimal ratio between interference at the input of the neural network classifier and the preservation of the information content of the data at its input.

Block diagram 1 (Fig. 1) contains a reference electrode 1, a microelectrode array 2, an ECoG signal amplifier 3, an ADC 4, a frequency-band filter 5, and a signal processing processor 6.

The block diagram 3 (Fig. 2) contains a training block for neural network classifiers 7 and a block for neural network classification 8.

Block diagram 7 (Fig. 3) contains blocks for storing time periods F ⁱ ECoG for each i-th odor and smell name b ⁱ 11, forming training arrays of input vectors {a _j } of each j-th MHC 12, preprocessing arrays of input vectors {a _j } 13, forming a training array of output string variables {b} from the textual names of smells 14 common to all MNSs, preprocessing an array of output string variables {b} 15, and executing the back propagation algorithm for each j-th MNS 16.

The block diagram 11 (Fig. 4) contains blocks for supplying the i-th smell 17 to the rat, storing the time interval F ^{i of the} ECoG signal of duration τ with a delay Δt after the start of inspiration t _{0 of the} i-th smell 18 and storing the textual name of the smell b ⁱ 19.

The block diagram 12 (Fig. 6) contains blocks for filtering time periods F ⁱ ECoG of a signal with a band-pass filter in the frequency range f ₁ -f ₂ 20, generating an array of matrices x ⁱ _j from the signal amplitudes of all leads 21, forming an array of matrices x ⁱ _j from the values of the first derivative of the signal amplitudes with respect to time of all leads 22, the formation of an array of vectors x ⁱ _j from the cross-correlation coefficients of the signal amplitudes between leads 23, the calculation of the frequency-amplitude spectrum of time segments F ⁱ ECoG signal in the frequency range f ₁ -f ₂ 24 forming an array of matrices x ⁱ _j from the values of the amplitudes of frequencies of all leads 25, the formation of an array of vectors x ⁱ _j from the coefficients of cross-correlation of the values of the amplitudes of frequencies between leads 26, combining arrays of matrices {x ⁱ _j } of each ith smell into training arrays of input vectors x _j for each j-th MHC 27, combining arrays of vectors {x ⁱ _j } of each i-th odor into training arrays of input vectors x _j for each j-th MHC 28, calculating the mean, variance, asymmetry coefficients and kurtosis for each lead from the matrix x _j and combining them into a vector a _j 29, calculating the mean, variance, asymmetry and kurtosis coefficients for each vector x _j and combining them into vector a _j 30.

Block diagram 13 (Fig. 7) converts for each j-th MNS array of input vectors a _j = (a _j1 , a _j2 , ..., a _jnj ) into an array of vectors a _j = (a _j1 , a _j2 , ..., a _jnj ) by standardizing the values of the array {a _j } for each component of the vector a _j .

The block diagram 15 (FIG. 8) encodes an array of string variables b into an array of output vectors β. Coding is carried out in accordance with table 1.

The block diagram 16 (Fig. 9) contains blocks for performing straight-line calculations of output vectors θ 31, performing back-stream calculations of errors Δw of weight coefficients w and their modifications w = w + Δw 32.

The structural diagram of a multilayer neural network (Fig. 10) contains an input, hidden and output layer.

Functional diagram of the nodal element of the hidden layer and the output layer of the multilayer neural network (Fig. 11) contains an adder 33 that calculates the scalar product of the input signal by the vector of weight coefficients w, and a nonlinear transducer 34.

The block diagram 31 (Fig. 12) performs layer-by-layer transformations over the input vectors α.

The block diagram 32 (Fig. 13) calculates the weighted coefficients w.

Block diagram 8 (Fig. 15) contains blocks for storing the ECoG segment F ⁱ 34, generating arrays of input vectors {a _j } for each j-th MNS 12, preprocessing arrays of input vectors {a _j } 13, and performing direct-flow calculations of output vectors θ _j for each j-th MHC 31, averaging the components of the output vectors θ _j over all MHC 35, and determining the presented smell 36.

The block diagram 34 (Fig. 16) contains blocks for supplying an odor to a rat for recognizing 17 and storing the time interval F ^{i of the} ECoG signal of duration τ with a delay interval Δt after the start of inspiration t ₀ upon presentation of smell 18.

The block diagram 35 (Fig. 17) averages the components of the output vectors θ _j over all MHCs to obtain the final output vector θ.

The block diagram 36 (Fig. 18) converts the calculated output vector θ into the binary vector β = (β ₁ , ..., β _i , ... β ₃ ) by initializing the component β _{i with} one if the ith component of the vector θ is the maximum of all values of the components of the vector θ, and initialization of the remaining components of the vector β by zeros, and the conversion of the binary vector β to the string variable b.

The reference electrode 1 is located in the bone of the skull of the rat, the microelectrode matrix 2 is placed on the surface of the olfactory bulb of the rat, recorded using dental materials on the skull of the rat and records the electrocorticographic (ECoG) signal of the olfactory bulb of the rat, which is amplified by a multi-channel amplifier 3, is converted to digital a signal with a sampling frequency of 1000 Hz in the ADC 4 and is filtered by a frequency-band filter 5 with a passband in the range from 0.5 to 200 Hz to eliminate low acoustic and high-frequency interference (Fig. 1).

The filtered signal is transmitted to processor 6 for training five multilayer neural networks (MNS) 7, if MNS are not trained, or for classification 8, if MNS are trained (Fig. 2). MNS training begins with memorizing the training time periods F ^{i of the} eight-channel ECoG signal and the names of smells b ⁱ for each i-th odor (block 11, Fig. 3).

To smell the rat, the i-th of three odors was introduced from a medical syringe with a volume of 20 ml from a 20 ml syringe into the surrounding area: air, a simulator of trinitrotoluene and a component of the cocaine drug, methyl benzoate (block 17, Fig. 4). During the supply of odor, the time interval F ^{i of the} eight-channel ECoG signal is memorized with a duration of τ = 500 ms with a delay of Δt = 50 ms from the moment of the start of inspiration t ₀ (block 18) and the textual name of the smell b ⁱ (block 19) is memorized. The onset of inspiration corresponds to a local maximum on the graph of the breathing signal (Fig. 5), which was recorded using an infrared range finder located on the rat peritoneum and fixing the breath by the movement of the peritoneum. Each i-th smell is served 60 times. There is a pause of 20 seconds between odor feeds.

Arrays of input vectors {a _j } are formed from the time intervals F ^{i of the} eight-channel ECoG signal for each j-th of five MNSs (block 12). For the first three MNS, the time segments F ^{i of the} ECoG signal are pre-filtered with a band-pass filter for all eight channels in the frequency range from f ₁ = 50 to f2 = 140 Hz (block 20) (Fig. 6). For the first MNS, an array of matrices x ⁱ _j (j = 1) is formed from the filtered signal from the values of the signal amplitudes of all leads (block 21), for the second MNS, an array of matrices x ⁱ _j (j = 2) is formed from the values of the first derivative of the signal amplitudes with respect to time all leads (block 22), for the third MNS an array of vectors x ⁱ _j (j = 3) is formed from the cross-correlation coefficients of the signal amplitudes between the leads (block 23). For the last two MNSs, the frequency-amplitude spectrum of the time intervals F ^{i of the} ECoG signal is calculated in the frequency range from f ₁ = 50 to f2 = 140 Hz in steps of 1 Hz (block 24). For the fourth MNS, based on the calculated amplitude spectrum, an array of matrices x ⁱ _j (j = 4) is formed from the values of the amplitudes of the frequencies of all leads (block 25); for the fifth MNS, an array of vectors x ⁱ _j (j = 5) is formed from the cross-correlation coefficients of the values amplitudes of frequencies between assignments (block 26). Further, for the first, second, and fourth MHC, the array of matrices {x ⁱ _j } is combined (block 27), and for the third and fifth MHC, the array of vectors {x ⁱ _j } (block 28) for each ith smell is combined into a training array {xj} input vectors x _j for each j-th MNS:

{x _j } = {x ¹ _j } ∪ {x ² _j } ∪ {x ³ _j }

After that, for the first, second and fourth MNS, the mean, variance, asymmetry and kurtosis values are calculated for each lead from the matrix x _j and their combination into the vector a _j (block 29), for the third and fifth MNS, the mean, variance is calculated , asymmetry and kurtosis coefficients for each vector x _j and combining them into vector a _j (block 30). The average is calculated by the formula

where m is the number of training examples in the array {x _j }, the variance is calculated by the formula:

,

,

the asymmetry coefficient is calculated by the formula:

,

,

the excess coefficient is calculated by the formula:

.

.

An array of input vectors {α _j } for each j-th MNS undergoes preprocessing (block 13) (Fig. 7), which consists in converting an array of vectors a _j = (a _j1 , a _j2 , ..., a _jn ) into an array of vectors a _j = (a _j1 , a _j2 , _... , a _{j nj} ) by standardizing the values of the array {a _j } for each component of the vector a _j according to the formula:

where a _{j ave} = (a _{j 1 ave} , a _{j 2 ave} , ..., a _{j nj ave} ) is the vector of averaged values of each component of the vector a _j over the entire array {a _j }, σ _j = (σ _{j 1 ave} , σ _{j 2 ave} , ... σ _{j nj ave} ) is the vector of standard deviation for each component of the vector a _j , n _j is the dimension of the input vector of the j-th MNS. The value of n _j for the first, second, and fourth MHC is 32, because Four values are calculated — mean, variance, asymmetry, and kurtosis — for each of the 8 leads. The value of n _j for the third and fifth MHC is 4 by the number of the above calculated values.

After the formation of arrays of input vectors {α}, the training array of output string variables {b} common to all MNSs is formed from the textual names of smells corresponding to each vector α (block 14) and the subsequent preprocessing of this array, which consists in encoding the textual names of smells b ¹ , b ² , b ³ into three-component numerical vectors β ¹ = (1, 0, 0), β ² = (0, 1, 0), β ³ = (0, 0, 1) (block 15) (Fig. 8 ) in accordance with table 1:

After preprocessing, the training arrays of input {α} and output vectors {β} are used for MHC training and for this, the error back propagation algorithm (Rumelhart DE, Hinton GE and Williams RJ. Learning internal representations by error propagation. In Parallel distributed processing, London: MIT Press, vol. 1, 1986, 550 p.) [33], which is intended to adjust the weight coefficients w (block 16) (Fig. 9). The error back-propagation algorithm consists of an iterative procedure for performing direct-flow calculations with input vectors α (block 31) (Fig. 9) to calculate the output vector θ and then performing back-flow calculations of errors Δw of weight coefficients w and their modifications w = w + Δw (block 32 )

The error back propagation algorithm is implemented for a multilayer neural network (MNS), which contains computational nodal elements organized in layers: input layer, hidden layer, and output layer (Fig. 10). The nodes of the input layer of the MHF do not perform any calculations and are used to distribute each component of the input vector α to all the node elements of the hidden layer having a weight matrix w ₁ . Similarly, the nodal elements of the output layer are associated with all elements of the hidden layer having a matrix of weight coefficients w ₂ . The calculations produce nodal elements in the hidden and output layer of the MHC. In direct flow calculation, each nodal element ij of the hidden i = 1 and output i = 2 layers, where index j is the ordinal number of the nodal element in the layer, contains a weight vector w _ij , an adder Σ 33 and a nonlinear converter σ 34 (Fig. 11). The adder 33 calculates the scalar product of the input signal vector of the nodal element of the MHF with the vector of weights w _ij . Nonlinear transducer 34 calculates a sigmoid function:

As a result of layer-by-layer calculations carried out by the nodal elements of the hidden and output layer above the input vector α, we obtain the calculated output vector θ at the output of the MHF (Fig. 12). In the reverse flow calculation (Fig. 13), each nodal element of the output layer of the MHC calculates the function δ ₂ according to the formula:

δ ₂ = (θ-β) ⋅σ ₂ (1-σ ₂ )

where θ is the calculated output vector of the neural network, β is the output vector, σ ₂ is the sigmoid function at the output of the nodal element of the output layer of the neural network,

and modifies its weighting factors w ₂ by the error Δw ₂ = δ ₂ ⋅σ ₁ according to the formula:

w ₂ = w ₂ + Δw ₂

Next, each nodal element of the hidden layer of the MHC calculates the function δ ₁ according to the formula:

δ ₁ = Σδ ₂ ω ₂ ⋅σ ₁ (1-σ ₁ )

where σ ₁ is the sigmoid function at the output of the nodal element of the hidden layer of the neural network, and modifies its weight coefficients w ₁ by the error Δw ₁ = δ ₁ ⋅σ ₀ according to the formula:

w ₁ = w ₁ + Δw ₁

where σ ₀ = α. Indexes 0, 1, and 2 relate to the input, hidden, and output layers, respectively. The iterative procedure of direct and reverse flow calculations continues until the calculation limit is reached, consisting in changing Δw weight coefficients w: Δw <Δw _min , where Δw _min = 0.001.

The error back propagation algorithm used in teaching the MNS provides the construction of nonlinear boundaries separating different odors represented by string variables b ⁱ . In FIG. 14 shows a case of a non-linear arrangement of examples of two odors depicted by dots of light and dark color (FIG. 14a). The use of a linear classifier constructing linear dividing boundaries, such as discriminant analysis, does not allow us to separate the areas of belonging of examples of different odors (Fig. 14b). The use of a nonlinear classifier that constructs nonlinear dividing boundaries, such as the MHF with the error back propagation algorithm, allows us to separate the domain of examples of different odors (Fig. 14c).

After training five MHCs, processor 6 classifies odors 8 (FIG. 15). For this i-th three odors placing the rat in accordance with the previously described unit 17, and storing the time interval F ⁱ eight-ECoG signal similar to block 18. From the time interval F ⁱ ECoG signal eight-shaped array of input vectors {a _j} for each j- one of the five MHCs is similar to the previously described block 12. The preprocessing of arrays {a _j } for each j-th of the five MHCs is similar to the previously described block 13. Performing straight-through calculations of the output vectors θ _j for each j-th MHC is similar to the previously described block 31. Next the three components of the output vectors θ _j are averaged over all five MHCs to obtain the final output vector θ (block 35) (Fig. 17) by the formula

,

,

where k = (1,2,3) is the index of the component of the output vector θ. The odor is determined by converting the output vector θ into the binary vector β = (β ₁ , ..., β ₁ , ..., β ₃ ) by initializing the component β _{i with} one if the ith component of the vector θ is the maximum of all the values of the components of the vector θ, and initialization of the remaining components of the vector β with zeros:

and converting the binary vector β to the string variable b in accordance with table 1 (block 36) (Fig. 18). The value of the string variable b is the result of the identification of the smell.

When classifying odors in twenty series, each of which consists of the sequential supply of three predetermined odors, for each odor in one or two out of twenty cases, the rat odor was incorrectly identified. Thus, for a total of sixty odor presentations, the odor recognition accuracy was at least 90%.

Odor recognition experiments were performed on eight sexually mature rats, and the odor recognition accuracy results for eight rats are shown in Table 2.

An example of a specific implementation of the method.

Rat preparation.

In FIG. 20 shows a layout of trepanation holes, screws and ligatures for mounting a microelectrode block on a rat’s head, where:

37 - fixing screws;

38 - trepanation holes for shoe mounting;

39 - trepanation hole above the olfactory bulb;

40 - steel surgical ligatures;

41 - the area of contact of the microelectrode pads to the bone.

To form a microelectrode block, holes are made in the lateral crests of the parietal bones on the right and left at the level of the parietal bone, and two steel surgical ligatures with a diameter of 100 μm are made for them. Ligatures are inserted into holes and filled with cold-cured plastic so that their ends fit into the thickness of the microelectrode block.

After this, a trepanation hole is created above the right or left olfactory bulb, without damaging the dura mater, by cutting with a dental bur with a head diameter of 0.5 mm. The hole is located along the length of 6.5-9 mm rostrally from the bregma, along the width of 0.5-1.7 mm from the sagittal suture.

Using a dental drill in the frontal bone, 2 trepanation holes for surgical steel hooks and two holes for fixation screws are created, followed by screwing screws of the corresponding diameter into them. When screwing in the screws, it is necessary to control the screwing depth (not more than 1-1.5 mm) so as not to damage the dura mater.

A layer of cold-hardened plastic (Protacryl-M, Acrodent) with a thickness of about 1 mm is applied to the prepared bones of the skull to create a layer of strong adhesion between the bone and the base of the microelectrode block. The front edge of the microelectrode pad made of plastic is located 4 mm rostral from bregma (the junction of the frontal and parietal bones), the posterior edge reaches the lambda (the intersection point of the lambdoid and sagittal sutures), the pad does not reach 1-1.5 mm to the right and left to the crest. The location and dimensions of the pads are calculated in this way to avoid contact with the cerebrospinal fluid and interstitial fluid released from the muscles running along the crest, as well as to ensure maximum contact area with the bones of the skull.

Using a manipulator, a “gluing” of microelectrodes is immersed in a trepanation hole above the olfactory bulb. Depending on the subsequent tasks, the layout of the microelectrode matrix, the place and depth of its immersion can vary. As a rule, gluing consists of two adjacent rows of 4 tungsten electrodes each.

After immersion of the microelectrode matrix, the trepanation hole is filled with light-cured plastic, locally, within 1-1.5 mm around the edges of the trepanation hole.

Then, holding the spring electrodes, form a block of light-cured plastic so that the connector is located above the paired parietal bones of the animal.

A harness vest 42 was put on the rat (Fig. 21), in which a module for recording and transmitting bioelectric potentials 43 from the implanted microelectrode array 44 was mounted on top of the rat. A displacement sensor 45 was placed on the rat peritoneum during breathing. The microelectrode array connector via the WiFi transmitter is connected to the WiFi receiver, which is connected to PC 46.

The connector of the implanted microelectrode matrix, the reference electrode and the output of the displacement sensor were connected to an eight-channel amplifier 3.

Preparation of gas mixtures of substances for the presentation of odors to rats.

The concentration of saturated vapors of simulators of two substances: trinitrotoluene and cocaine was calculated by the Mendeleev - Clapeyron formula for 25 ° C. VP = RT (M / m), where V is the volume of the container, l, R = 8.31 J \ mol * K (universal gas constant), T is in Kelvin, M is the mass of the substance, kg, m is the molar mass of the substance , kg / mol, P is the vapor pressure of the substance, Pa.

To create saturated vapors of methyl benzoate, 1 ml of the substance was placed in a 0.4 L glass container, kept for 2 hours. For first-order dilution with a 5 ml glass syringe, 4 ml of the resulting gas mixture was taken and transferred to a 15 l glass container. The resulting gas mixture was mixed for 10 minutes with a rubber bulb sealed to the container. To obtain the final dilutions with concentrations of 0.1 μg / m ³ and 0.01 μg / m ^3, 1 ml of the gas mixture was taken from the first-order dilution and placed in a 15 L glass container.

By a similar method, saturated pairs of cocaine simulators and trinitrotoluene simulators with the same concentrations were obtained.

Pairs of substances of the indicated concentration were delivered by a syringe to the nose of the rat on inspiration, which was recorded by a respiration sensor.

Molecules of substances after dissolution in the mucus lining the nasal cavity interact with specific receptor proteins responsible for odorant binding located on the membranes of the olfactory olfactory receptor epithelium. The interaction of the odorant molecule with a specific protein receptor leads to the formation of a receptor potential in the local area of the membrane due to the opening of ion channels. The simultaneous binding of several odorant molecules on the membrane leads to the summation and more pronounced depolarization of the receptor membrane, which, upon reaching a critical level, forms a generator action potential, which is transmitted along the axon to the overlying structures of the olfactory analyzer. The axons of the olfactory cells are combined into groups of 20-100 fibers and, as part of the olfactory nerve, go to the olfactory bulb, while receptors of the same type converge on one glomerulus of the olfactory bulb.

The arrival of nerve impulses from the olfactory epithelium to the glomeruli of the olfactory bulbs leads to the formation of a specific spatial pattern of neural activity, which, when summed up, forms focal potentials. Registration of focal potentials with arrays of microelectrodes implanted in the olfactory bulb allows detecting activity patterns specific to each smell or their mixtures by comparing them with reference patterns recorded in response to the presentation of reference odors or their mixtures. If the parameters of the current and reference patterns of bioelectric activity coincide, the conclusion is made about the presence of the target substance in the supplied air sample.

The method provides the detection of narcotic substances in a concentration of not more than 0.1 μg / m ³ and explosive - 0.01 μg / m ³ . The accuracy is improved through the use of machine learning systems and the claimed method of signal analysis. The method also provides a simplification of the process of preparation for use, by eliminating the need for stimulation of the brain with electrical signals or other training.

Information sources:

1. Kikhtenko A.V., Eliseev K.V. Detection of explosive objects: hardware of anti-terrorist services // Russian Chemical Journal. 2005.T. XLIX. Number 4. from. 132-137.

2. RU 2471179 IPC G01N 27/64, publication date 2012-12-27.

3. RU 2360242 IPC G01N 30/96, publication date 2006-01-26.

4. RU 2231781 IPC G01N 27/62, publication date, 2004-06-27.

5. RU 2398309 IPC H01J 49/40 G01N 27/64, publication date 2009-07-16.

6. RU 2354963 IPC G01N 27/62, G01N 30/64, publication date 2007-11-08.

7. RU 2329563 IPC H01J 49/40, G01N 27/64, publication date 2006-12-25.

8. RU 2460067 IPC G01N 27/62, publication date 2011-04-20.

9. WO 2008049488 A1 NKI G01N 27 / 62A, H01J 49 / 10B, IPC H01J 49/16, G01N 27/64, publication date 2012-06-20.

10. RU 2395076 IPC G01N 27/64, publication date 2009-03-23.

11. RU 2431212 IPC H01J 49/40, publication date 2010-07-09.

12. CN 2622705 Y, IPC H01J 49/26, publication date 2003-05-20.

13. US 6470730 B1, NKI, IPC G01N 1/00, G01N 1/02, G01N 1/22, G01N 1/28, G01N 33/00, publication date 2000-08-18.

14. Frederickx C., Verheggen F.J., Haubruge E. Biosensors in forensic sciences // Biotechnol. Agron Soc. Environ. 2011. Vol. 15. No. 4. P. 449-458.

15. RU 2426141 IPC G01S 1/02, publication date 2010-03-09.

16. RU 2338175 IPC G01N 1/22, A01K 15/00, publication date 2006-09-29.

17. APOPO Anti-Persoonsmijnen Ontmijnende Product Ontwikkeling, Web: https://www.apopo.org/en/component/one/summary?show=265, publication date 2014-09-19.

18. TAMAR Group (Israel) (http://news.sky.com/story/1010634/sniffer-mice-used-to-detect-bombs-and-drugs, publication date 2012-11-13).

19. US 2005009444 A1, NKI 449/27, 119/420, 119 / 6.5, 119/421, IPC G01N 33/00, publication date 2002-12-20.

20. Rains G.C., Tomberlin J.K., Kulasiri D. Using insect sniffing devices for detection // Trends Biotechnol. 2008. Vol. 26. No. 6. P. 288-294.

21. Detection of explosives with bees Web: http: //www.pergam.ru/articles/biogasanalizer.htm, publication date 11/28/2013.

22. US 20120264353 A1, NKI 449/2, 449/1, IPC A01K 51/00, A01K 55/00, publication date 2011-04-04.

23. Macias M.M., Manso A.G., Orellana C.G.O., Gonzalez Velasco H.M., Caballero R.G. and Peguero Chamizo J.C. Acetic Acid Detection Threshold in Synthetic Wine Samples of a Portable Electronic Nose // Sensors. 2013. Vol. 13. P. 208-220; doi: 10.3390 / sl30100208.

24. Dog-inspired scent detector sniffs out explosives and narcotics Web: http://www.gizmag.com/electronic-nose/25128/, published date 11/21/2012.

25. Yinon J., Zitrin S. Modern methods and applications in analysis of explosives. Chichester, UK; New York, USA: Wiley. 1996, 316 c.

26. Application US 2003/0199944 A1, IPC, MPI publication date 2003-20-23.

27. Application US 2013098310 A1 IPC A01K 1/03, A01K 15/02, A61N 1/36, A61N 1/372, B62D 63/02, G06N 3/06, H02H 1/00, publication date 2013-04-25.

28. WO 2011132756 A1, IPC A61B 5/0478, A61B 5/0476, A01K 67/00, A61B 5/055, A61B 5/0408, A61B 5/0492, publication date 2011-10-27.

29. EP 1234540 A1, IPC A01K 1/03, A01K 29/00, A61B 5/0478, A61B 5/0484, G01N 33/50, publication date 2002-08-28.

30. WO 2005037100 A1, IPC A61B 5/0484, A61B 5/04847, A61B 5/4011, publication date 2005-04-28.

31. CN 1865996, IPC A61B 5/00, G01N 33/00, publication date 2006-11-22.

32. Gupta P., Albeanu D.F., & Bhalla U.S. Olfactory bulb coding of odors, mixtures and sniffs is a linear sum of odor time profiles // Nature Neuroscience. 2015.doi: 10.1038 / nn. 3913.

33. Rumelhart D.E., Hinton G.E. and Williams R.J. Learning internal representations by error propagation. In Parallel Distributed Processing, London: MIT Press, Vol. 1, 1986, 550 p.) [33].

Claims (6)

1. A method for detecting explosive and narcotic substances in the air based on an analysis of the bioelectric potentials of the olfactory rat analyzer, which consists in implanting a microelectrode matrix from working and at least one reference electrodes into the upper surface of the olfactory bulb of the rat, recording the electrocorticographic signal of the olfactory bulb in a given frequency range in the moment of inspiration, the extraction of five groups of signs expressed in the form of mathematical values of mean, variance, asymmetry and exce CCA calculated for the amplitudes of the electrocorticographic signal of each lead, the first derivative of the amplitudes of the electrocorticographic signal of each lead, the frequency-amplitude spectrum of each lead, the cross-correlation coefficients of the amplitudes of the electro-corticographic signal between the leads, the cross-correlation coefficients of the frequency-amplitude spectrum between the leads, the processing of each group of signs separate multilayer neural network, while for training each multilayer neural network will complement an array of odor indicators presented to the rat is generated for a specified number of times with a given pause duration, which is the output array for training a multilayer neural network; when training for each multilayer neural network, weight classification coefficients are calculated and, when identifying an odor, five groups of signs of an electrocorticographic signal process five multilayer neural networks and calculate the probability of belonging to the rat odor to one of the given odors for each group without signs, calculate the arithmetic mean of the probabilities of five multilayer neural networks with the choice of smell with maximum probability as the result of recognition. 2. The method according to p. 1, characterized in that eight working electrodes are used. 3. The method according to p. 1, characterized in that the weighting classification coefficients are calculated according to the algorithm of the back propagation of the error. 4. The method according to p. 1, characterized in that the number of presentations of each smell in training a multilayer neural network is 60 times. 5. The method according to p. 1, characterized in that the duration of the pause between presentations of odors during training of a multilayer neural network is 20 seconds. 6. The method according to p. 1, characterized in that the frequency range of detection of odors in the electrocorticographic signal is 50-140 Hz.

Images