Electronic nose smell analysis system and method applied to multiple scenes

Abstract

The application discloses a kind of electronic nose smell analysis system and method applied to multiple scenes, it is characterized in that, including industrial computer, analyzer and sampler composition, wherein, sampler selects pin sampler or multichannel sampler according to detection scene;Pin sampler or multichannel sampler is connected analyzer, analyzer connects industrial computer;External power supply is powered for analyzer and industrial computer, and analyzer is powered for pin sampler or multichannel sampler.The application proposes a kind of machine olfactory system applicable to multiple scenes, with trace sample detection and continuous process volatile detection function, can be applied to chemical or biological laboratory rapid sample detection, or to continuous production process product or by-product monitoring.The machine olfactory system of the application has the advantages of small size, low energy consumption, high constant temperature precision, high degree of automation, fast detection speed and simple use process.

Description

TECHNICAL FIELD

[0001] The present application relates to an electronic nose smell analysis system and method, and relates to the fields of computers, automatic control, data analysis, etc. BACKGROUND

[0002] The present application relates to an electronic nose smell analysis system and method, and relates to the fields of computers, automatic control, data analysis, etc.

[0003] In the technical solution disclosed in the existing patent one: one of the characteristics is that the gas sensor array, the gas sensor array constant temperature working room, the multi-channel fermentation tail gas precision automatic sampling system, the computer control and analysis system are integrated in a test box. The second characteristic is that multiple two-position two-way electromagnetic valves realize automatic switching between multiple fermentation tail gases, between the fermentation tail gas, clean air and the environment air, and between 6000 milliliters / minute and 500 milliliters / minute flow. The third characteristic is to realize simultaneous online detection and analysis of up to five biological fermentation processes, i.e. five fermentation tanks. The fourth characteristic is that the internal precision throttle valve and pressure stabilizing valve of the olfactory simulation instrument have stable pressure and stable flow capacity for gas. The fifth characteristic is that the olfactory simulation instrument and multiple fermentation tanks are plug and play, easy to operate, and the detection and analysis process of multiple biological fermentation processes is automated, networked and visualized.

[0004] In the technical solution disclosed in the existing patent two: one of the characteristics is that the instrument is composed of a test box, a headspace gas generating device, a computer, and a clean air bottle. The second characteristic is that the gas sensor array working room, the measured sample and the headspace volatile gas constant temperature are 55±0.1℃, the constant temperature process only heats, does not cool, and the headspace volatile gas volume is 180 milliliters. The third characteristic is that the gas headspace sampling is automatically completed by the test box micro vacuum pump, electromagnetic valve, throttle valve, sampling needle, automatic sampling lifting device and computer within 5 seconds. The fourth characteristic is that the sampling needle hole directly communicates with the gas sensor array working room gas inlet hole. The fifth characteristic is that the multi-gas smell qualitative and quantitative simultaneous analysis method is realized by a structured neural network module.

[0005] However, the existing patent one and the existing patent two have the following problems:

[0006] 1) The electronic nose instruments mentioned in the existing patent one and the existing patent two are respectively applied to different application scenarios: on-site single detection and continuous online cyclic detection, and cannot solve the demand of being applied to multiple scenarios at the same time;

[0007] 2) The electronic nose instruments in the existing patent one and the existing patent two are too large, and are inconvenient to carry, which limits the use scenarios;

[0008] 3) The automatic sampling device mentioned in the existing patent two is fixed on the host, and cannot be flexibly applied;

[0009] 4) The existing patent one and the existing patent two do not realize effective control of the sensor constant temperature cavity;

[0010] 5) The existing patent one and the existing patent two do not consider the influence of constant pressure on automatic sampling. SUMMARY

[0011] The purpose of the present application is to design a more portable and small host, an automatic offline collector and a multi-channel continuous sampler, so as to realize the universality for different scenes.

[0012] In order to achieve the above purpose, the technical scheme of the present application provides an electronic nose smell analysis system applied to multiple scenes, characterized in that it comprises an industrial computer, an analyzer and a sampler, wherein the sampler selects a pin-type sampler or a multi-channel sampler according to the detection scene; the pin-type sampler or the multi-channel sampler is connected with the analyzer, and the analyzer is connected with the industrial computer; an external power supply supplies power for the analyzer and the industrial computer, and the analyzer supplies power for the pin-type sampler or the multi-channel sampler;

[0013] The analyzer comprises an analyzer shell, a constant temperature cavity arranged in the analyzer shell, and a sensor working cavity arranged in the constant temperature cavity; a wall of the constant temperature cavity comprises an outer wall and an inner wall, and a heat preservation material layer is formed by filling a heat preservation material between the outer wall and the inner wall; a constant temperature adjusting component is arranged in the constant temperature cavity and connected with an integrated control circuit board arranged in the analyzer shell; the sensor working cavity comprises a sensor working cavity base, a plurality of sensor mounting sockets are arranged on the sensor working cavity base, the sensor mounting sockets are connected with the integrated control circuit board, and a gas sensitive sensor array is formed by inserting the gas sensitive sensors into the sensor mounting sockets; a sensor working cavity cover is fastened to the sensor working cavity base, so that the gas sensitive sensor array is located in a cavity between the sensor working cavity cover and the sensor working cavity base; an exhaust outlet, an external sampler socket, a sampling gas path inlet, a calibration gas path inlet and a cleaning gas path inlet are arranged on the analyzer shell, the external sampler socket is connected with the integrated control circuit board, the analyzer is connected with a pin-type sampler or a multi-channel sampler through the external sampler socket; the cleaning gas path inlet and the sampling gas path inlet are connected to an air inlet of the sensor working cavity through a Y-shaped three-way joint; an air outlet of the sensor working cavity is connected with an air inlet of a diaphragm air pump, and an air outlet of the diaphragm air pump is directly connected with the exhaust outlet; the air inlet of the sensor working cavity is also connected with the calibration gas path inlet; one electromagnetic valve is arranged on a pipeline between the Y-shaped three-way joint and the air inlet of the sensor working cavity, on a pipeline between the calibration gas path inlet and the air inlet of the sensor working cavity, and on a pipeline between the exhaust outlet and the air outlet of the sensor working cavity, respectively, and the electromagnetic valves and the diaphragm air pump are connected with the integrated control circuit board.

[0014] Preferably, the constant temperature cavity is a circular box-shaped structure, and the overall thickness of the constant temperature cavity is 10 mm, and the thickness of the outer wall and the inner wall is 1 mm.

[0015] Preferably, the constant temperature adjusting component comprises a power-adjustable heating sheet and a fan arranged above the sensor working cavity.

[0016] Preferably, the gas sensitive sensor array is annular or rectangular.

[0017] Preferably, the analyzer shell is further provided with a heat dissipation hole, a data connection socket, a power switch and a power socket, the data connection socket, the power switch and the power socket are connected with the integrated control circuit board, the analyzer is connected with the industrial computer through the data connection socket and connected with the external power supply through the power socket.

[0018] Preferably, the pin sampler comprises a plastic shell, an aviation plug for connecting with the analyzer is arranged on the plastic shell, the plastic shell is open in front for taking and placing the sampling bottle, the sampling bottle is placed on a tooling bottom plate, a heating coupler is arranged between the sampling bottle and the tooling bottom plate, a guide column and a push plate motor are arranged on the tooling bottom plate, a lifting push plate is fixedly connected with a moving end of the push plate motor, the lifting push plate is sleeved with a linear bearing on the guide column, an injection head fixing piece and linear bearings located on both sides of the injection head fixing piece are arranged on the front part of the lifting push plate, a sampling needle is arranged on the injection head fixing piece and connected with a sampling gas path inlet of the analyzer through a gas pipe, the linear bearings are sleeved with spring guide columns, springs are arranged on the spring guide columns below the lifting push rod, and a thrust plate is arranged on the end part of the spring guide column.

[0019] Preferably, the sampling bottle is internally provided with heat preservation material, and a round hole is arranged on the top of the sampling bottle for inserting the sampling needle.

[0020] The application further discloses a data processing method of the electronic nose smell analysis system, and the method comprises the following steps: odor substances in a sample are transported to the gas sensitive sensor array of the electronic nose smell analysis system and react with active materials on the gas sensitive sensor array, so that the conductivity of the gas sensitive sensor array changes, the electronic nose smell analysis system collects the electric signal, pre-processes and pattern recognizes the electric signal, and finally analyzes the attribution of the odor substances.

[0021] Preferably, for non-continuous data obtained by single retrieval, any one of the following data processing methods is used for pattern recognition: a dimension reduction analysis algorithm realized based on principal component analysis (PCA), linear discriminant analysis (LDA) or quadratic discriminant analysis (QDA); a quantitative analysis method realized based on partial least squares regression (PLSR) or neural network BP; a qualitative analysis method realized based on K neighbor analysis (KNN); a clustering algorithm, and the calculation method comprises Euclidean distance, Minkowski distance and Mahalanobis distance, and the classification result is determined according to the classification requirement in the final classification tree diagram.

[0022] Preferably, for continuous data, a steady-state resistance change rate is calculated, whether the biological fermentation process is normal is judged based on the steady-state resistance change rate, and qualitative and quantitative analysis is performed by using a neural network.

[0023] The application provides a machine olfactory system applicable to multiple scenes, which has functions of micro sample detection and continuous process volatile detection, can be applied to rapid sample detection of a chemical or biological laboratory or product or byproduct monitoring of a continuous production process (chemical industry or biological fermentation), and has the advantages of small volume, low energy consumption, high constant temperature precision, high automation degree, fast detection speed and simple use process. Compared with the prior art, the application has the following advantages:

[0024] 1) The present application can simultaneously detect two different scenes, and can even be applied to open space scenes;

[0025] 2) The present application realizes integration, miniaturization and portability;

[0026] 3) The automatic sampling device of the present application can be used independently;

[0027] 4) The constant pressure device is used to ensure that the gas is not affected by constant pressure during sampling;

[0028] 5) The data analysis algorithm of the present application is diverse, and can realize the analysis of test results of various requirements. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is the overall appearance of the analyzer;

[0030] Figure 2 is a schematic view of one side of the analyzer;

[0031] Figure 3 is a schematic view of another side of the analyzer;

[0032] Figure 4 is a C-C cross-sectional view of Figure 3

[0033] Figure 5 illustrates the constant temperature cavity;

[0034] Figure 6 is an assembly drawing of the gas sensitive sensor;

[0035] Figure 7 is the power supply principle diagram of the present application;

[0036] Figures 8 to 10 is a structure schematic view of the pin type sampler;

[0037] Figures 11 to 15 is a structure schematic view of the multi-channel sampler, wherein, Figure 15 is an A-A cross-sectional view of Figure 12

[0038] Figure 16 is a two-dimensional schematic view of LDA, wherein, "+" and "-" represent the positive side and the negative side respectively, the ellipse represents the outer contour of the data cluster, the dotted line represents the projection, and the solid circle and the solid triangle represent the center points of the two types of sample projections respectively. DETAILED DESCRIPTION

[0039] ​​The application will be further described in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate but not to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content of the application, and these equivalent forms also fall within the scope defined by the appended claims.

[0040] The electronic nose smell analysis system applied to multiple scenes disclosed in the embodiment is composed of an industrial computer, an analyzer and a sampler, wherein the sampler can select a pin sampler or a multi-channel sampler according to a detection scene. The analyzer is connected with the industrial computer through a USB data line, the industrial computer transmits a control signal to the analyzer through the USB data line to control the action of the terminal hardware. An analog-digital conversion module in the analyzer converts a sensor voltage value into a digital signal, which is fed back to the industrial computer through the USB data line, and the industrial computer processes and analyzes the feedback data.

[0041] As shown in Figures 1 to 4 , in the embodiment, the analyzer includes an analyzer shell, the analyzer shell has a constant-temperature cavity 1 inside, and the constant-temperature cavity 1 has a sensor working cavity 2 inside. In combination with Figure 5 , in the embodiment, the constant-temperature cavity 1 is a circular box-shaped structure with an inner space slightly larger than that of the sensor working cavity 2. The overall thickness of the constant-temperature cavity 1 is 10 mm, the wall of the constant-temperature cavity 1 is composed of an outer wall, an inner wall and a heat preservation material layer between the outer wall and the inner wall, the thickness of the outer wall and the inner wall is 1 mm, and the heat preservation material layer is formed by filling the heat preservation material between the outer wall and the inner wall. The constant-temperature cavity 1 has a power-adjustable heating sheet and a fan above the sensor working cavity 2 inside, temperature adjustment is performed by using the power-adjustable heating sheet, and air flow is accelerated by the fan, so that the temperature in the constant-temperature cavity 1 is maintained constant. The power-adjustable heating sheet and the fan are connected with an integrated control circuit board 14 arranged in the analyzer shell.

[0042] In combination with Figure 6 , the sensor working cavity 2 includes a sensor working cavity base 2-1, the sensor working cavity base 2-1 has a plurality of annular sensor mounting sockets 2-2 on it, the sensor mounting sockets 2-2 are connected with the integrated control circuit board 14, and in the embodiment, the sensor mounting sockets 2-2 are 16. According to the target detection gas, the corresponding gas-sensitive sensor 2-3 is inserted on the sensor mounting socket 2-2, so as to form an annular gas-sensitive sensor array. After the gas-sensitive sensor 2-3 is installed, the sensor working cavity cover 2-4 is buckled and fixed on the sensor working cavity base 2-1 through the screw 2-5, so that the annular gas-sensitive sensor array is located in the cavity between the sensor working cavity cover 2-4 and the sensor working cavity base 2-1. For the application scene of air detection, the size of the annular gas-sensitive sensor array can be reduced, for example, a rectangular gas-sensitive sensor array can be used, thereby the overall size of the analyzer can be further reduced.

[0043] One side of the analyzer housing is provided with a waste gas exhaust port 4, an external sampler socket one 5, an external sampler socket two 6, a heat dissipation hole 7, a sampling gas path inlet 8, a calibration gas path inlet 9, and a cleaning gas path inlet 10. The other side of the analyzer housing is provided with a USB Type-C socket 11, a power switch 12, and a power socket 13. The external sampler socket one 5, the external sampler socket two 6, the USB Type-C socket 11, the power switch 12, and the power socket 13 are connected to the integrated control circuit board 14. The analyzer is connected to a pin-type sampler or a multi-channel sampler through the external sampler socket one 5 and the external sampler socket two 6. The analyzer is connected to an industrial computer through the USB Type-C socket 11. The analyzer is connected to an external power supply through the power socket 13.

[0044] The cleaning gas path inlet 10 and the sampling gas path inlet 8 are connected to the gas inlet of the sensor working cavity 2 after being connected to one path through a Y-shaped three-way joint. The gas outlet of the sensor working cavity 2 is connected to the gas inlet of a diaphragm air pump, and the gas outlet of the diaphragm air pump is directly connected to the waste gas exhaust port 4. The gas inlet of the sensor working cavity 2 is also connected to the calibration gas path inlet 9. An electromagnetic valve is respectively arranged on the pipeline between the Y-shaped three-way joint and the gas inlet of the sensor working cavity 2, on the pipeline between the calibration gas path inlet 9 and the gas inlet of the sensor working cavity 2, and on the pipeline between the waste gas exhaust port 4 and the gas outlet of the sensor working cavity 2. The flow sequence of the measured gas in the analyzer is: the sampling gas path inlet 8-the sensor working cavity 2-the diaphragm air pump-the waste gas exhaust port 4, which can maximize the reduction of the influence of gas path pollution on the detection effect of the sensor working cavity 2. The electromagnetic valves and the diaphragm air pump are connected to the integrated control circuit board 14.

[0045] The integrated control circuit board 14 has an analog-digital conversion module for converting the voltage signal of the gas sensitive sensor 2-3 into a digital signal and uploading it to the industrial computer. At the same time, the integrated control circuit board has a PWM (Pulse Width Modulation) output function, which can adjust the motor speed of the diaphragm air pump, so that the flow rate of the diaphragm air pump in the normal pressure and no-load state is adjusted within the range of 300-3000 ml / min.

[0046] As shown in FIG. 7, the external power supply independently supplies power to the analyzer, and the external power supply supplies power to the sampler through the integrated control circuit board 14 in the analyzer, and forms a multi-path stabilized power supply through the integrated control circuit board 14, in which 5V is used for power-adjustable heating sheet and fan power supply, 12V is used for gas sensitive sensor 2-3 heating power supply, A and B paths 24V are used for electromagnetic valve power supply, C path 24V is used for diaphragm air pump power supply, and D path 24V can be controlled high-frequency switching on-off state, which is used for adjusting the power of the heating sheet in the pin-type sampler.

[0047] For the sample detection function, a needle sampler can be used, and the air plug of the needle sampler is connected to the air plug (exterior sampler plug one 5 and exterior sampler plug two 6) of the analyzer through a cable to enable the analyzer to supply power to the needle sampler and transmit signals.

[0048] The needle sampler includes a plastic shell, a needle mechanism, and a sampling bottle 3. The plastic shell is provided with an air plug for communication connection with the analyzer. The plastic shell is open at the front for taking and placing the sampling bottle 3. The plastic shell is provided with a rubber pad at the bottom to increase the stability of the needle sampler as a whole.

[0049] The needle mechanism includes a tool base plate, a heating coupler mounted on the tool base plate, and a sampling bottle positioning block 15 mounted on the heating coupler for placing the sampling bottle 3. The sampling bottle 3 is provided with a heat preservation material inside and a round hole at the top for insertion of a sampling needle 21. The tool base plate is provided with a raised block at a position behind the sampling bottle 3, and the raised block is provided with a base, a motor base, and a fixed frame 19. The base is provided with a guide column 17, and the upper part of the guide column 17 is mounted together with the fixed frame 19. The motor base is provided with a push plate motor, and the moving end of the push plate motor is provided with a push plate base mounted on a lifting push plate 18, which is sleeved on the guide column 17 through a linear bearing. The lifting push plate 18 is provided with an injection head fixing piece 20 at the front, and the injection head fixing piece 20 fixes the sampling needle 21 by screwing. The sampling needle 21 is connected to the analyzer through an air pipe 26. The lifting push plate 18 is also provided with a linear bearing 22 at the front end, and the linear bearing 22 is located on both sides of the sampling needle 21. The linear bearing 22 is sleeved on a spring guide column 23, one end of the spring guide column 23 is provided with a limiting pad 24 to prevent the spring guide column 23 from falling out of the linear bearing 22. The spring guide column 23 below the lifting push plate 18 is provided with a spring, and the other end of the spring guide column 23 is provided with a stop plate 25. Before sample detection, the operator covers the sample container with a rubber pad and fixes it with a middle hole bottle cap, and places it in the sampling bottle 3. The sampling bottle 3 is placed in the needle sampler. During the headspace sampling process, the push plate motor retracts, the lifting push plate 18 falls along the guide column 17, the stop plate 25 presses the sampling bottle 3, the sampling needle 21 continues to insert, until the push plate motor completely retracts, at this time, the sampling needle 21 inserts into the sampling bottle to start sampling. After sampling, the push plate motor extends, the sampling needle 21 first separates from the sampling bottle 3, and due to the action of the spring, the stop plate 25 continues to press on the sampling bottle 3 to prevent the sampling bottle 3 from rising with the sampling needle 21. When the sampling needle 21 separates from the sampling bottle 3, the push plate motor continues to extend until the stop plate 25 separates from the sampling bottle 3.

[0050] For the multi-channel sampling detection function, the present application uses a needle sampler as shown in Figures 11 to 13The multi-channel sampler is shown. The aviation socket 27 of the multi-channel sampler is connected to the corresponding aviation socket (external sampler socket one 5 and external sampler socket two 6) of the analyzer through a cable, so that the analyzer supplies power to the multi-channel sampler and transmits signals. The multi-channel sampler has five gas input interfaces 28, which are connected to the detection gas source using a gas pipe. The gas output interface of the multi-channel sampler is connected to the sampling gas path inlet 8 of the analyzer through a gas pipe. During the multi-channel cyclic detection process, the multi-channel sampler receives the control signal given by the analyzer, and sequentially opens the specified channel for detection.

[0051] The electronic nose smell analysis system disclosed in the embodiment mainly has the following two functions:

[0052] First, single sample detection, the complete process is as follows:

[0053] (1) Start: connect the analyzer and industrial computer with USB-TypeC data line, connect the pin type sampler with the corresponding interface of the analyzer, press the start button on the analyzer, and preheat the equipment for about 20 minutes, so that the temperature in the sample containing cavity of the sampler (offline) and the temperature of the sensor working cavity in the host are increased to the preset constant temperature;

[0054] (2) Sample preparation: place the measured solid or liquid sample in a beaker of appropriate size, place the beaker in the constant temperature tank of the sampler (offline), cover the sealing gasket and lid, align the connector at the bottom of the constant temperature tank with the connector of the sampler, and then click the confirmation button on the operation interface. Preheat and constant temperature for 10 minutes;

[0055] (3) Sampling process: after the temperature stabilizes to the set constant temperature, click the "start sampling button" button, the top rod retracts, the sampling needle descends until it pierces the sealing gasket and enters the sample space, the electromagnetic valve is controlled to open and close in turn, the calibration and sampling process is completed, and the sensor response data is recorded by the upper computer. After sampling is completed, the cleaning gas path is opened, and the motor works at the maximum speed to flush the sensor working cavity with environmental air;

[0056] (4) Data processing: the data analysis software in the upper computer analyzes the category, intensity or key component concentration estimate value of the gas, obtains the analysis result and displays it on the upper computer screen.

[0057] Second, multi-channel cyclic detection, the complete process is as follows:

[0058] (1) Start: connect the analyzer and industrial computer with USB-TypeC data line, connect the multi-channel sampler with the corresponding interface of the host, press the start button on the analyzer, and preheat the equipment for about 20 minutes, so that the temperature of the sensor working cavity in the analyzer is increased to the preset constant temperature, and the sampling interfaces with serial numbers 1-5 are connected to the gas environments to be detected;

[0059] (2) Configure channel parameters: double-click each interface icon to configure sampling parameters for each gas source, including channel selection, calibration time, sampling time, etc.

[0060] (3) Sampling process: click the "Start cyclic sampling" button to make the device cycle through the "calibration-sampling-flushing" process in the 5 gas sources, and record the sensor response data by the host computer.

[0061] (4) Data processing: the data analysis software in the host computer analyzes the category, intensity or key component concentration estimate of the gas, and displays the analysis results on the host computer screen.

[0062] The data processing method of the electronic nose smell analysis system applied to multiple scenes disclosed in the embodiment includes the following steps:

[0063] The odor substances in the sample are transported to the gas sensitive sensor array of the electronic nose smell analysis system and react with the active material thereon, causing the electrical conductivity of the gas sensitive sensor array to change, the electronic nose smell analysis system collects the electrical signal, pre-processes and pattern recognizes the electrical signal, and finally analyzes the attribution of the odor substances.

[0064] In the analysis of experimental data, due to the need for noise reduction, drift compensation and information compression, the collected data is often pre-processed. Data preprocessing has a significant impact on the performance of the electronic nose system. A good signal preprocessing method can reduce recognition error and complexity, thereby improving system recognition ability. In the electronic nose system, when selecting a data preprocessing method, data conversion, noise removal, baseline drift elimination, response amplitude normalization and feature dimension reduction are often considered. Some common data preprocessing techniques for electronic nose systems are as follows:

[0065] 1.1 Difference method

[0066] The difference method is the simplest data preprocessing technique, and the mathematical expression is:

[0067] R ij =x ij -x 1j (1)

[0068] Here R ij represents the response signal of the jth sensor at the ith time collection point after data preprocessing, x ij represents the instantaneous response signal of the jth sensor at the ith time collection point, and x 1j represents the initial response signal of the jth sensor. The difference method can remove additive noise and drift.

[0069] 1.2 Relative method

[0070] The relative method includes the following algorithms:

[0071]

[0072]

[0073]

[0074]

[0075] R in the formula ij represents the response signal of the jth sensor at the ith time collection point after data preprocessing; x ij represents the instantaneous response signal of the jth sensor at the ith time collection point, x 1j represents the initial response signal of the jth sensor. The algorithm max represents the maximum value. In formula (2), x 1j represents the initial response signal of the jth sensor. X in formula 3 represents a matrix composed of the response data of the measured sample at all sensors and each collection time. For an electronic nose system containing p sensors, if the data of each sample is collected at n time points, X is an n*p matrix; x j in formula 4 represents a vector composed of the response data of the measured sample at each collection time of the jth sensor; the case of formula 5 is to select a vector x r as a reference response for scaling, x i in formula 6 represents a vector composed of all response data of the measured sample at the ith time collection point. ||x i || represents the modulus of the vector x i . The relative method can remove the multiplicative noise and drift.

[0076]

[0077] 1.3 Difference quotient method

[0078] The difference quotient method is a combination of the difference method and the relative method, and its mathematical expression is to replace x ij in formulas 2-6 with (x 1j -x ij ). Therefore, the difference quotient method has the effects of the difference method and the relative method, and can remove the additive and multiplicative noise and drift.

[0079] 1.4 Normalization

[0080] In the case of not considering the concentration influence, the egg needs to be identified in detail, and the normalization processing is a very common data preprocessing method. The normalization of the electronic nose data has two kinds, one is to normalize the response of a single sensor, and the other is to normalize the entire response data image, which are represented by formulas 7 and 8 respectively.

[0081]

[0082] In formulas 7 and 8, the algorithm max represents the maximum value, and the algorithm min represents the minimum value. The normalization processing converts the response value into a dimensionless number, and the value falls in the interval [0.1], which can effectively reduce the calculation error, but when the signal is weak, the normalization will amplify the noise.

[0083] The pattern analysis based on the preprocessed data includes the following aspects:

[0084] 1) Dimensionality reduction analysis, mainly using principal component analysis PCA, linear discriminant analysis LDA or quadratic discriminant analysis QDA.

[0085] 1. Principal component analysis PCA: find the principal component in high-dimensional data, and use the "principal component" data to represent the original data, so as to achieve the purpose of dimensionality reduction, which focuses on the comprehensive evaluation of information contribution influence.

[0086] Principal component analysis PCA is to recombine a group of new independent comprehensive indicators to replace the original indicators, which have certain correlation (such as P indicators).

[0087] The purpose of principal component analysis PCA is to reduce the dimension of the data set in the experiment (reduce the number of variables: because some variables have correlation, the information reflected has certain overlap, and the opposite variance contribution in the data set is maintained).

[0088] The way to perform principal component analysis PCA is to retain low-order principal components and ignore high-order principal components after considering the correlation; usually the mathematical processing is to make linear combination of the original P indicators as new comprehensive indicators.

[0089] The most classic way of performing principal component analysis (PCA) is to use the variance of F1 (the first selected linear combination, i.e., the first composite index) to express it, i.e., the greater the Var(F1), the more information F1 contains. Therefore, the F1 selected from all linear combinations should be the one with the largest variance, hence the name F1 as the first principal component. If the first principal component is not sufficient to represent the information of the original P indicators, consider selecting F2, i.e., the second linear combination. In order to effectively reflect the original information, the information already in F1 should not appear in F2. In mathematical language, this means that Cov(F1, F2) = 0 (representing that F1 and F2 are not related), then F2 is called the second principal component, and so on to construct the third, fourth, …, and Pth principal components.

[0090] Therefore, through principal component analysis (PCA), the original variables are recombined into a new set of several independent composite variables, and according to actual needs, several fewer composite variables can be taken out to reflect as much information as possible from the original variables. This statistical method is called principal component analysis or principal component analysis, which is also a mathematical method for dimensionality reduction.

[0091] The specific algorithm flow of principal component analysis (PCA) is as follows:

[0092] (1) Standardize the original data, if the sample data matrix is as follows

[0093]

[0094] Standardize the original data:

[0095]

[0096] where,

[0097] (2) Calculate the sample correlation coefficient matrix (note that here it is assumed that the original data is standardized and still represented by X. The Xi and Xj mentioned in the method of calculating the element value of the coefficient matrix given later are not the original data matrix X, but the standardized matrix elements. This is because X is still used here.)

[0098]

[0099] where,

[0100] (3) Calculate the eigenvalues and corresponding eigenvectors of the correlation coefficient matrix R:

[0101] Eigenvalues: λ1, λ2, …, λ p

[0102] Eigenvectors: a i = (ai1 , a i2 , …a ip ), i = 1, 2, …, p

[0103] (4) Select important principal components

[0104] P principal components can be obtained by principal component analysis. According to the foregoing description, the information contained in F1 is greater than that in F2, and this can be extended. Therefore, the variance of each principal component is also decreasing, and the amount of information contained is also decreasing. Therefore, when using this method, not all P principal components are selected, but the first K principal components are selected according to the size of the cumulative contribution of each principal component. The contribution rate here refers to the proportion of the variance of a principal component in the total variance of principal components, that is, the proportion of a certain eigenvalue in the sum of all eigenvalues, that is:

[0105]

[0106] The greater the contribution rate of a principal component, the greater the amount of original information contained in the principal component. The selection of the principal component K value is mainly determined by the cumulative contribution rate of the principal component. Generally, when the cumulative contribution rate reaches more than 85%, it can be considered that the K principal components contain most of the information of the original information.

[0107] (5) Calculate the principal component score, which is in the following form:

[0108]

[0109] f ij = a j1 x i1 +a j2 a i2 +…·+a jp a ip , i = 1, 2, …, n; j = 1, 2, …, k

[0110] (6) According to the data of the principal component score, subsequent analysis and modeling of the problem are carried out.

[0111] 2. Linear discriminant analysis LDA (Latent Dirichlet Allocation): The analysis maximizes the inter-class distance and minimizes the intra-class distance, which is a relatively classic and popular supervised algorithm in the field of data mining at present.

[0112] Linear Discriminant Analysis (LDA), also known as Fisher Linear Discriminant (FLD), is a classic algorithm for pattern recognition. It was introduced into the fields of pattern recognition and artificial intelligence by Belhumeur in 1996.

[0113] The basic idea of ​​linear discriminant analysis (LDA) is to project high-dimensional pattern samples onto an optimal discriminant vector space to extract classification information and compress the feature space dimensionality. After projection, the pattern samples are guaranteed to have the maximum inter-class distance and the minimum intra-class distance in the new subspace, meaning the patterns have optimal separability in that space. Therefore, it is an effective feature extraction method. Using this method, the inter-class scatter matrix of the projected pattern samples can be maximized while the intra-class scatter matrix is ​​minimized.

[0114] Linear discriminant analysis (LDA): The original points (PCA) are clustered together as much as possible among similar points and separated as much as possible among different points. A line is then found and projected onto this line to make them clearly distinguishable.

[0115] so Figure 16 Of the two projections shown, ① did not achieve good differentiation, while ② satisfied the principle of "grouping similar objects together as much as possible and separating different objects as much as possible". We will take ② as the target value.

[0116]

[0117] LDA and PCA are both commonly used dimensionality reduction techniques. PCA mainly focuses on finding a better projection method from the perspective of feature covariance. LDA considers the annotation more, aiming to increase the distance between data points of different categories and make data points of the same category more compact after projection. LDA is essentially a projection, projecting a high-dimensional point onto a high-dimensional line (in this projection process, trying to maximize the distance between data points of different categories and make data points of the same category more compact).

[0118] During the analysis, if the repeatability is good, PCA can be used; if the repeatability is poor, further optimization is needed, and LDA should be used instead (LDA is an optimization based on PCA).

[0119] The LDA algorithm process is as follows:

[0120] Input: Dataset D = {(x1, y1), (x2, y2), ..., (x...} m y m )}, where any sample x i Let y be an n-dimensional vector. i∈ {c1, c2,.., c k}, dimension d.

[0121] Output: the reduced sample set D'

[0122] 1) Calculate the within-class scatter matrix S w

[0123] 2) Calculate the between-class scatter matrix S b

[0124] 3) Calculate the matrix

[0125] 4) Calculate S w -1 S b The largest d eigenvalues and the corresponding d eigenvectors (w1, w2,.., wd) of S d ) get the projection matrix w

[0126] 5) For each sample feature x i in the sample set, convert it to a new sample z i = W T x i

[0127] Get the output sample set D' = {(z1, y1), (z2, y2),..., (z m , y m )}

[0128] 3. Quadratic Discriminant Analysis QDA

[0129] QDA can distinguish between classification tasks by constructing a curve, which is better than LDA which can only perform linear discrimination in general classification problems, and can also produce contour effects. LDA can perform classification and dimensionality reduction at the same time (such as the first figure in the vertical direction), while QDA can only perform classification tasks.

[0130] Points of similarity and difference between linear discriminant analysis and quadratic discriminant analysis:

[0131] When the covariance matrices of different classification samples are the same, use linear discriminant analysis; when the covariance matrices of different classification samples are different, use quadratic discriminant.

[0132] General form of quadratic discriminant function:

[0133]

[0134] Where W is a d-order symmetric matrix, and w is a d-dimensional weight vector.

[0135] Assume that each sample satisfies Gaussian distribution, so the discriminant function of each class can be defined as the comparison between the square of Mahalanobis distance from the sample to the mean and a given threshold:

[0136]

[0137] where mk is the sample mean of each sample class, ∑k is the covariance matrix of each class, and the first term on the right side is a discriminant threshold, which depends on the prior probability and ∑k;

[0138] Estimation of mean and covariance:

[0139]

[0140]

[0141] For a two-class problem, the decision surface is:

[0142] g1(x)-g2(x)=0 (5)

[0143] The decision rule is: if the left side is greater than the right side, decide as w1 class, and if the right side is greater than the left side, decide as w2 class. If the decision is wrong, the threshold of the two classes can be adjusted to reduce the error rate.

[0144] In particular, if one of the classes is distributed relatively close to the Gaussian distribution, i.e. the distribution is clumpy, and the other class is uniformly distributed near the first class, for this case, no decision surface like formula (5) is needed, only the discriminant function of the first class needs to be calculated. As can be seen in formula (2), if the second term (Mahalanobis distance square) is less than the threshold of the first term, then g(x)>0, so it is decided as w1 class, and vice versa w2 class.

[0145] Second) Quantitative analysis

[0146] 1. Partial Least Squares Regression (PLSR): It is more superior for data processing with more independent variables or dependent variables, and the regression coefficients can show the correlation and influence of each independent variable on the dependent variable.

[0147] PLS: According to the different weights of variables, the regression coefficients of each variable are calculated, and the regression variance is established. The software provides three PLS algorithms, which are used to calculate the concentration or content, predict the sensory score and predict the shelf life.

[0148] The principle of PLS is that: partial least squares regression PLS is an extension of multiple linear regression model, in its simplest form, only a linear model is used to describe the relationship between the independent variables Y and the predictor variable group X: Y = b0 + b1X1 + b2X2 +... + bpXp: In the equation, b0 is the intercept, and the value of bi is the regression coefficient of data points 1 to p. First, the first principal component is extracted from the independent variable matrix X, called t1. t1 must satisfy: t1 is a linear combination of independent variables x1, x2,..., xp; t1 reflects the original information of the independent variable matrix X as much as possible; t1 makes the largest possible explanation to the dependent variable y.

[0149] The algorithm flow of PLS is as follows:

[0150] Assume that P independent variables {x1,..., xp} and q dependent variables {y1,..., yq} constitute the data table X = {x1,..., xp} and Y = {y1,..., yq} of independent variables and dependent variables. Components t1 and u1 are extracted from X and F, and when t1 and u1 are extracted, t1 and u1 should carry as much variation information in their respective data tables as possible, and the correlation between t1 and u1 should be maximized. After the first components t1 and u1 are extracted, regression of X on t1 and Y on u1 is performed. If the regression equation has reached a satisfactory accuracy at this time, the components are determined; otherwise, the residual information of X explained by t1 and F explained by u1 will be used to extract the second round of components t2 and u2, and the regression of X and Y on t2 and u2 will be continued. Iterate the above process until the accuracy meets the requirements. If m components t1,..., tm are finally extracted from X, and regression of F on t1,..., tm is performed, finally all can be transformed into the regression equation of Y on the original variables x1,..., xp, completing the partial least squares regression modeling.

[0151] The modeling steps of PLS are as follows:

[0152] step1: data standardization

[0153] step2: calculate the correlation coefficient matrix

[0154] step3: extract components of the independent variable group and the dependent variable group respectively, such as when the ratio of the current k components to explain the independent variables reaches 90%, take the first k components

[0155] step4: calculate the regression equation of k components on the standardized index variable and the component variable

[0156] step5: calculate the regression equation between the dependent variable and the independent variable group, that is, bring the components in step3 to the regression equation obtained in step4 to obtain the regression equation between the standardized index variables, and then restore the standardized regression variables to the original variables.

[0157] 2. Back Propagation Neural Network (BPNN): It includes two processes, signal forward propagation and error backward propagation, through matrix multiplication.

[0158] Back propagation artificial neural network is another algorithm most applied in the field of electronic nose. This algorithm is powerful, easy to understand and simple to train. The simplicity in concept and algorithm of BP-ANN, and its successful application in many practical problems make it currently occupy the mainstream position in the pattern recognition of electronic nose.

[0159] The learning process of BP-ANN consists of two processes, signal forward propagation and error backward propagation. In forward propagation, samples are transmitted from input layer to output layer through processing in each hidden layer (usually only one hidden layer to avoid increasing computational complexity). The neurons in each layer are connected by different weights, the input of a neuron is the weighted sum of the outputs of all neurons in the upper layer, and the output of the neuron is generated by transformation of an activation function. The activation function has step function, linear function, and the most applied is Sigmoid function. If the actual output of the output layer does not match the expected output, the error backward propagation is entered. In error backward propagation, the output error is transmitted to the input layer through the hidden layer, and the error is allocated to each neuron in each layer to obtain the error signal of each neuron, and then the weights of each neuron are adjusted in the direction of gradient descent. The process of adjusting the weights of each layer through forward propagation of signal and backward propagation of error is repeated until the output error of the network is reduced to an acceptable level or the learning time reaches the pre-designed learning time. The trained network can be used for prediction calculation. Recording a BP-ANN requires storing its network structure, weights and activation function, etc.

[0160] Compared with the above-mentioned algorithms based on statistical theory, the biggest advantage of BP-ANN is that it can realize the nonlinear mapping between input and output data. In electronic nose signal processing, the relationship between input and output is often nonlinear, which makes BP-ANN widely applied and good effect. The difficulties in the application of BP-ANN include the design of network structure, the selection of the number of hidden layer neurons, the selection of activation function, the initialization of weights, the setting of network error, etc. There is no uniform standard for the selection of the above-mentioned parameters, and they are usually selected according to the experience of researchers. Some improved algorithms, such as adding momentum term and self-adaptive adjustment of learning rate, can improve the BP-ANN algorithm to some extent.

[0161] The steps of BP algorithm are as follows:

[0162] 1. Initialize parameters W, b

[0163] 2. Calculate the state and activation of each layer using the forward propagation formula

[0164] z (l) = W (l) a (l-1) + b (l)

[0165] a (l) = f(z (l) )

[0166] 3. Calculate

[0167]

[0168]

[0169] 4. Compute the partial derivative of the cost function with respect to the function for the training data using the following formula

[0170]

[0171]

[0172] Third) Qualitative analysis

[0173] K-Nearest Neighbor (KNN): KNN is more suitable for sample sets with more cross or overlap of class domains than other methods.

[0174] KNN (K-Nearest Neighbor) is one of the simplest algorithms in data mining classification techniques, and its guiding idea is "close to red, close to black", that is, your neighbors can infer your category.

[0175] KNN (K Near Neighbor): k nearest neighbors, that is, each sample can be represented by its k nearest neighbors. The principle of KNN is to judge which category x belongs to when predicting a new value x according to the category of its nearest K points. The key is the selection of K value and the calculation of point distance.

[0176] (1) Distance calculation

[0177] The Euclidean distance used in KNN algorithm is calculated as follows:

[0178]

[0179]

[0180] Extended to multi-dimensional space, the formula becomes:

[0181]

[0182] (2) Selection of K value

[0183] A K value that is too large will lead to underfitting, while a K value that is too small will lead to overfitting. Cross-validation is needed to determine the appropriate K value. Cross-validation involves splitting the sample data into training and validation data in a certain ratio, such as a 6:4 ratio. Starting with a small K value, the value of K is gradually increased, and the variance of the validation set is calculated to find a suitable K value.

[0184] Fourth) Clustering Algorithm

[0185] System clustering: The calculation methods include Euclidean distance, Minkowski distance, and Mahalanobis distance, and the classification result is determined in the final classification dendrogram according to the classification requirements.

[0186] The above-mentioned electronic nose odor analysis system includes the following steps in its continuous online data processing method:

[0187] The system displays 16 voltage response curves online. When the steady-state resistance change rate of these 16 curves exceeds 50%, an alarm is issued. If the steady-state resistance change rate of multiple curves exceeds 70%, a warning is issued indicating that the bio-fermentation process is abnormal and may be contaminated.

[0188] The formula for the rate of change of resistance is:

[0189] Rs: The resistance value of the sensor in clean air, which can be regarded as the natural resistance value and is greatly affected by temperature;

[0190] Rb: Value of the series voltage divider resistor;

[0191] V: Total voltage across the sensor and series resistor

[0192] Vb: Voltage across the series resistor

[0193] Vb = V * Rb / (Rs + Rb)

[0194] Vb*Rs+Vb*Rb=V*Rb

[0195] Rs=(V*Rb-Vb*Rb) / Vb=(V-Vb)Rb / Vb

[0196] V is a fixed value (5V in the experiment), Rb is a fixed resistance (10kΩ in the experiment), Vb is measured in the experiment, and the value of Rs can be calculated according to the above formula, Rs1 is calculated according to the average voltage value Vb1 between 5-10s in the calibration, and the maximum voltage value Vb2 of the voltage of the sampling time is divided into the resistance, and the time corresponding to the minimum value Rs2 of the sensor resistance, (Rs1-Rs2) / Rs1 is the maximum change rate of the resistance.

[0197] Continuous online data processing process:

[0198] The present application gives early warning of whether the fermentation process is normal or not, whether it is contaminated or not, etc. after the end of the headspace sampling process of the fermentation tail gas, predicts the components of tail gas O2, CO2, NH3, etc., estimates the parameters of dissolved oxygen, respiratory quotient, etc., estimates the concentration of the bacteria and the product, and analyzes the activity of the bacteria. The test data and analysis results are displayed in real time and transmitted remotely through the network.

[0199] In the present application, neural networks are used for qualitative and quantitative analysis, and multiple neural network modules for multiple odor recognition are established. The structured neural network module is composed of two columns of single-output single-hidden-layer neural networks, which are used for qualitative and quantitative analysis of multiple odors. From the horizontal direction, a pair of neural network modules represents an odor, and the first neural network module with s1 hidden layer nodes is responsible for identifying whether a sample is from the odor represented, and the second neural network module with s2 hidden layer nodes is responsible for estimating the intensity and concentration of the sample. The logarithm of the neural network is equal to the number of odor categories, and they correspond one by one. The basic learning algorithm of the neural network module is the error backpropagation algorithm.

[0200] The learning sample subset of the neural network module responsible for odor qualitative analysis is only generated by the localized "one-to-all" decomposition method, and only the training samples near the decision boundary participate in learning.

[0201] The neural network module column responsible for odor quantitative analysis is composed of a first column and multiple neural network module columns for odor concentration quantification - a second column. In order to determine the category and concentration of a to-be-determined odor sample, all neural network modules in the first column for odor recognition need to participate in decision-making, and the neural network module with the maximum output determines the category of the sample. However, only the neural network module in the second column corresponding to the neural network module with the maximum output in the first column is needed to quantify the concentration and intensity of the odor sample.

Claims

1. An electronic nose odor analysis system applicable to multiple scenarios, characterized in that, The system comprises an industrial control computer, an analyzer, and a sampler. The sampler can be either a pin-type sampler or a multi-channel sampler, depending on the testing scenario. The pin-type sampler is used for single-sample testing, while the multi-channel sampler is used for cyclic sample testing. The multi-channel sampler has multiple gas input interfaces and connects to the detection gas source via a gas tube. The gas output interface of the multi-channel sampler is connected to the sampling gas inlet of the analyzer via a gas tube. The pin-type or multi-channel sampler connects to the analyzer, which in turn connects to the industrial control computer. An external power supply powers both the analyzer and the industrial control computer, while the analyzer powers the pin-type or multi-channel sampler. The analyzer includes an analyzer housing, within which a constant temperature chamber is located, and within the constant temperature chamber is a sensor working chamber. The walls of the constant temperature chamber include an outer wall and an inner wall, with insulation material filling the space between the outer and inner walls to form an insulation layer. A constant temperature regulating component is located within the constant temperature chamber and is connected to an integrated control circuit board located within the analyzer housing. The sensor working chamber includes a sensor working chamber base with multiple sensor mounting sockets connected to the integrated control circuit board. Gas sensors are inserted into these sensor mounting sockets to form a gas sensor array. A sensor working chamber cover is fastened and fixed to the sensor working chamber base, positioning the gas sensor array within the cavity between the sensor working chamber cover and the sensor working chamber base. The analyzer housing has an exhaust gas outlet and an external sampling port. The analyzer includes a sampler connector, a sampling gas inlet, a calibration gas inlet, and a cleaning gas inlet. An external sampler connector connects to the integrated control circuit board. The analyzer connects to a pin-type sampler or a multi-channel sampler via the external sampler connector. The cleaning gas inlet and the sampling gas inlet are combined using a Y-shaped tee connector and then connected to the air inlet of the sensor working chamber. The exhaust port of the sensor working chamber connects to the air inlet of the diaphragm pump, and the exhaust port of the diaphragm pump is directly connected to the waste gas outlet. The air inlet of the sensor working chamber also connects to the calibration gas inlet. A solenoid valve is installed on the pipeline between the Y-shaped tee connector and the air inlet of the sensor working chamber, on the pipeline between the calibration gas inlet and the air inlet of the sensor working chamber, and on the pipeline between the waste gas outlet and the exhaust port of the sensor working chamber. The solenoid valves and the diaphragm pump are connected to the integrated control circuit board.

2. The electronic nose odor analysis system for multiple scenarios as described in claim 1, characterized in that, The constant temperature cavity is a circular box-shaped structure with an overall thickness of 10mm, and the thickness of both the outer wall and the inner wall is 1mm.

3. The electronic nose odor analysis system for multiple scenarios as described in claim 1, characterized in that, The temperature control component includes a power-adjustable heating element and a fan located above the sensor working chamber.

4. The electronic nose odor analysis system for multiple scenarios as described in claim 1, characterized in that, The gas sensor array is either ring-shaped or rectangular.

5. The electronic nose odor analysis system for multiple scenarios as described in claim 1, characterized in that, The analyzer housing is also provided with heat dissipation holes, a data connection port, a power switch, and a power socket. The data connection port, power switch, and power socket are connected to the integrated control circuit board. The analyzer is connected to the industrial control computer through the data connection port and to the external power supply through the power socket.

6. The electronic nose odor analysis system for multiple scenarios as described in claim 1, characterized in that, The pin-type sampler includes a plastic housing with an aviation plug for connecting to the analyzer. The plastic outer shell has an opening at the front for placing and removing sampling bottles. The sampling bottles are placed on the fixture base plate, and a heating coupler is provided between the fixture base plate and the sampling bottles. The fixture base plate is equipped with a guide post and a push plate motor. The lifting push plate is fixedly connected to the moving end of the push plate motor, and the lifting push plate is sleeved on the guide post through a linear bearing. The front of the lifting push plate is equipped with an injection head fixing component and linear bearings located on both sides of the injection head fixing component. The sampling needle is located on the injection head fixing component and is connected to the sampling gas inlet of the analyzer through a gas tube. The linear bearing is sleeved on the spring guide post, and a spring is installed on the spring guide post located below the lifting push rod. A thrust plate is provided at the end of the spring guide post.

7. The electronic nose odor analysis system for multiple scenarios as described in claim 6, characterized in that, The sampling bottle is lined with insulating material, and has a round hole at the top for inserting the sampling needle.

8. A data processing method for the electronic nose odor analysis system as described in claim 1, characterized in that, Includes the following steps: Odor substances in the sample are delivered to the gas-sensitive sensor array of the electronic nose odor analysis system as described in claim 1 and react with the active material thereon, causing a change in the conductivity of the gas-sensitive sensor array to generate an electrical signal. The electronic nose odor analysis system as described in claim 1 acquires the electrical signal, performs preprocessing and pattern recognition on the electrical signal, and finally analyzes the attribution of the odor substances.

9. The data processing method of the electronic nose odor analysis system as described in claim 8, characterized in that, For discontinuous data obtained from a single retrieval, any of the following data processing methods can be used for pattern recognition: dimensionality reduction analysis algorithms based on principal component analysis (PCA), linear discriminant analysis (LDA), or quadratic discriminant analysis (QDA). Quantitative analysis methods based on partial least squares regression (PLSR) or backpropagation (BP) neural networks; qualitative analysis methods based on k-nearest neighbor analysis (KNN). Clustering algorithms use methods such as Euclidean distance, Minkowski distance, and Mahalanobis distance to calculate the classification results in the final classification tree diagram based on the classification requirements.

10. The data processing method of the electronic nose odor analysis system as described in claim 8, characterized in that, For continuous data, the steady-state resistance change rate is calculated, and the bio-fermentation process is judged to determine whether it is normal based on the steady-state resistance change rate. A neural network is then used for qualitative and quantitative analysis.

Images