A method for analyzing machine olfactory enhancement or machine taste inhibition

Abstract

The present invention discloses a method for analyzing machine olfactory enhancement or machine taste inhibition, comprising: a machine olfactory model, a machine taste model, a sensory evaluation model, a variable projection importance model, an olfactory-taste synesthesia model, a convolutional neural network, a grid search-support vector machine, and an olfactory effect analysis model. The method proposed by the present invention reflects the machine olfactory enhancement or machine taste inhibition effect that occurs during flavor substance analysis, and provides a tool for explaining the interaction of olfactory-taste sensory organs.

Description

Technical Field

[0001] The present invention relates to the technical field of taste processing, and in particular to a method for analyzing machine olfactory enhancement or machine taste inhibition. Background Art

[0002] Human perception of flavor substances mainly comes from the sensory mixing phenomenon between smell and taste. There is a strong correlation between free flavor substances and soluble volatile substances, so people usually mistake the sensation of some volatile substances for "taste". Machine smell and machine taste can obtain olfactory and taste information of flavor substances and achieve excellent results in the field of flavor substance analysis, but machine smell and machine taste only analyze flavor substances from the perspective of data and algorithms, and cannot truly reflect the human sensory mixing phenomenon; some papers propose olfactory perception models, and some patents propose taste perception models for taste identification and olfactory-taste synesthesia perception models for flavor identification, but they do not analyze the effect of machine smell on enhancing or inhibiting machine taste, which affects the understanding of the interaction between smell and taste in the tasting of flavor substances. Summary of the invention

[0003] The main purpose of the present invention is to provide a method for analyzing machine olfactory enhancement or machine taste inhibition.

[0004] The technical solution adopted by the present invention is: a method for analyzing the enhancement of machine olfaction or the inhibition of machine taste, comprising:

[0005] Machine olfactory model, machine taste model, sensory evaluation model, variable projection importance model, smell-taste synesthesia model, convolutional neural network, grid search-support vector machine and olfactory action analysis model;

[0006] The machine smell and machine taste detection results are sent to a variable projection importance model, the machine taste detection results and the analysis results of the variable projection importance model are input into an olfactory-taste synesthesia model, the output of the olfactory-taste synesthesia model is sent to a convolutional neural network, the convolutional neural network output and the sensory evaluation results are input into a grid search-support vector machine, the grid search-support vector machine generates a prediction output and sends it to an olfactory effect analysis model, and the enhancement or inhibition of the machine smell on the machine taste is qualitatively reflected according to the model calculation results;

[0007] The machine olfactory model and the machine taste model are used to obtain the olfactory and taste information of the flavor substances, and send the olfactory and taste information to the variable projection importance model, and send the taste information to the olfactory-taste synesthesia model;

[0008] The variable projection importance model integrates the olfactory and taste information of the substance and determines the optimal combination relationship of the flavor information according to the magnitude of the olfactory-taste sensor variable importance function, and the determined flavor information is input into the olfactory-taste synesthesia model;

[0009] The olfactory-taste synesthesia model obtains machine taste information and flavor information, and adjusts the synesthesia model channel structure according to the input information form, including:

[0010] Under taste information input, the number of taste channels is the number of taste information sensor variables, and the number of olfactory channels does not change and has no input;

[0011] Under the input of flavor information, the number of taste channels is the number of taste sensor variables contained in the flavor information, and the number of olfactory channels is the number of olfactory sensor variables contained in the flavor information. Finally, the one-dimensional output time series of the model connection node is obtained and sent to the convolutional neural network;

[0012] The convolutional neural network converts the one-dimensional output time series of the connection nodes of the synesthesia model into two-dimensional perception data, and obtains perception features through convolution and pooling, and the perception features are input into the grid search-support vector machine;

[0013] The grid search-support vector machine uses the convolutional neural network perception features as input data and the real sensory evaluation results as output labels to generate predicted output labels, which are then input into the olfactory action analysis model.

[0014] Furthermore, the olfactory effect analysis model is formula (1) and (2):

[0015] (1)

[0016] (2)

[0017] in, For the The calculation results of the analysis model under the input of taste and flavor characteristics data of the group olfactory-taste synesthesia model, =1,2,… n,n is the total number of taste (or flavor) characteristic data;

[0018] For the The prediction output of grid search-support vector machine under the input of flavor feature data of group olfactory-taste synesthesia model;

[0019] For the The predicted output of grid search-support vector machine under the input of taste feature data of group olfactory-taste synesthesia model, the model olfactory and flavor features refer to the input of material olfactory and taste information into the synesthesia model, and the features extracted by convolutional neural network from the synesthesia model output;

[0020] is the final calculation result. If If the calculated result is greater than 0, the machine's sense of smell has an enhanced effect on the machine's sense of taste. If the calculated result is less than 0, the machine sense of smell has an inhibitory effect on the machine sense of taste.

[0021] Advantages of the present invention:

[0022] The method for analyzing the enhancement or suppression of machine olfaction or machine taste proposed in the present invention reveals the interaction between machine olfaction and machine taste, and provides a tool for understanding the phenomenon of sensory mixing.

[0023] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1 is a technical flow chart of a method for analyzing machine olfactory enhancement or machine taste inhibition of the present invention;

[0026] Figure 2 It is a radar chart of the olfactory information of a strawberry-sucrose mixed solution and a bitter melon-sucrose mixed solution for analyzing a method for enhancing or inhibiting machine olfactory sense of the present invention;

[0027] Figure 3 It is a radar chart of taste information of a strawberry-sucrose mixed solution and a bitter melon-sucrose mixed solution for analyzing a method for enhancing or inhibiting machine olfactory sense or taste sense of a machine according to the present invention;

[0028] Figure 4 It is the result of the variable projection importance analysis of a strawberry-sucrose mixed solution for a method of analyzing machine olfactory enhancement or machine taste inhibition of the present invention;

[0029] Figure 5 It is the result of variable projection importance analysis of bitter melon-sucrose mixed solution in a method of analyzing machine olfactory enhancement or machine taste inhibition of the present invention;

[0030] Figure 6 It is a structural diagram of a convolutional neural network for analyzing a method of enhancing or suppressing machine olfactory sense or taste of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0032] Table 1 is a calculation result of the analysis model under the input of taste and flavor information of a strawberry-sucrose mixed solution of a method for analyzing machine olfactory enhancement or machine taste inhibition of the present invention;

[0033] Table 2 is the calculation results of the analysis model under the input of taste and flavor information of bitter melon-sucrose mixed solution in a method of analyzing machine olfactory enhancement or machine taste inhibition of the present invention.

[0034] A method for analyzing machine olfactory enhancement or machine taste inhibition, comprising:

[0035] Machine olfactory model, machine taste model, sensory evaluation model, variable projection importance model, smell-taste synesthesia model, convolutional neural network, grid search-support vector machine and olfactory action analysis model;

[0036] The machine smell and machine taste detection results are sent to a variable projection importance model, the machine taste detection results and the analysis results of the variable projection importance model are input into an olfactory-taste synesthesia model, the output of the olfactory-taste synesthesia model is sent to a convolutional neural network, the convolutional neural network output and the sensory evaluation results are input into a grid search-support vector machine, the grid search-support vector machine generates a prediction output and sends it to an olfactory effect analysis model, and the enhancement or inhibition of the machine smell on the machine taste is qualitatively reflected according to the model calculation results;

[0037] The machine olfactory model and the machine taste model are used to obtain the olfactory and taste information of the flavor substances, and send the olfactory and taste information to the variable projection importance model, and send the taste information to the olfactory-taste synesthesia model;

[0038] The variable projection importance model integrates the olfactory and taste information of the substance and determines the optimal combination relationship of the flavor information according to the magnitude of the olfactory-taste sensor variable importance function, and the determined flavor information is input into the olfactory-taste synesthesia model;

[0039] The olfactory-taste synesthesia model obtains machine taste information and flavor information, and adjusts the synesthesia model channel structure according to the input information form, including:

[0040] Under taste information input, the number of taste channels is the number of taste information sensor variables, and the number of olfactory channels does not change and has no input;

[0041] Under the input of flavor information, the number of taste channels is the number of taste sensor variables contained in the flavor information, and the number of olfactory channels is the number of olfactory sensor variables contained in the flavor information. Finally, the one-dimensional output time series of the model connection node is obtained and sent to the convolutional neural network;

[0042] The convolutional neural network converts the one-dimensional output time series of the connection nodes of the synesthesia model into two-dimensional perception data, and obtains perception features through convolution and pooling, and the perception features are input into the grid search-support vector machine;

[0043] The grid search-support vector machine uses the convolutional neural network perception features as input data and the real sensory evaluation results as output labels to generate predicted output labels, which are then input into the olfactory action analysis model.

[0044] The olfactory effect analysis model is formula (1) and (2):

[0045] (1)

[0046] (2)

[0047] in, For the The calculation results of the analysis model under the input of taste and flavor characteristics data of the group olfactory-taste synesthesia model, =1,2,… n,n is the total number of taste (or flavor) characteristic data;

[0048] For the The prediction output of grid search-support vector machine under the input of flavor feature data of group olfactory-taste synesthesia model;

[0049] For the The predicted output of grid search-support vector machine under the input of taste feature data of group olfactory-taste synesthesia model, the model olfactory and flavor features refer to the input of material olfactory and taste information into the synesthesia model, and the features extracted by convolutional neural network from the synesthesia model output;

[0050] is the final calculation result. If If the calculated result is greater than 0, the machine's sense of smell has an enhanced effect on the machine's sense of taste. If the calculated result is less than 0, the machine sense of smell has an inhibitory effect on the machine sense of taste.

[0051] refer to Figures 1 to 6A method for analyzing whether machine olfaction enhances or inhibits machine taste, characterized in that it includes: machine olfaction, machine taste, sensory evaluation technology, variable projection importance model, olfactory-taste synesthesia model, convolutional neural network, grid search-support vector machine and olfactory effect analysis model, the machine olfaction and machine taste detection results are sent to the variable projection importance model, the machine taste detection result and the variable projection importance model analysis result are input into the olfactory-taste synesthesia model, the output of the olfactory-taste synesthesia model is sent to the convolutional neural network, the convolutional neural network output and the sensory evaluation result are input into the grid search-support vector machine, the grid search-support vector machine generates a predicted output and sends it to the olfactory effect analysis model, and the enhancement or inhibition of the machine olfaction on the machine taste is qualitatively reflected according to the model calculation result.

[0052] In this embodiment, machine smell and machine taste are used to obtain olfactory and taste information of flavor substances, and the olfactory and taste information is sent to the variable projection importance model, and the taste information is sent to the smell-taste synesthesia model.

[0053] The olfactory and taste information of flavor substances are recorded as and , l =1,…, m ; i =1,…, n , m and n are the number of olfactory and taste model channels in the olfactory-taste synesthesia model. Substituting into the taste channel equations (1)-(3), the olfactory channel formulas are (4)-(5)

[0054] （ 1)

[0055] (2)

[0056] (3)

[0057] (4)

[0058] (5)

[0059] Among them, E 1i (t), E 2i (t), E 3i (t) are respectively i Potential status of nodes E1, E2, and E3 of channels i =1,2,…, n , n is the number of channels of the taste perception model, n for Number of sensor variables; Wr1 , W r2 , W r3 is the connection coefficient between nodes E1, E2 and E3 and RG in the same channel, and the connection coefficient values ​​of all channels are consistent; W e3e2 It is the connection coefficient between nodes E2 and E3 of the same channel. The connection coefficient value of all channels is the same; NG p ( t ) is the potential state of the peripheral noise in the taste perception model; a and b1 are the coefficients of the differential equation; RG is the external stimulus in the taste perception model. 1l (t) The potential state of the node P1 of the lth channel (l=1,…,m), where m is the number of channels in the olfactory perception model, and m is Contains the number of sensor variables; a and b2 are differential equation coefficients; Q(P 2l (t)) is the output of node P2 of the first channel; Q(P 1j (t)) is the output of the jth channel P1 node (j=1,…,m); D2(t) is the potential state of the feedback link from the anterior olfactory nucleus to the periglomerular cells in the olfactory nerve conduction pathway; W pp is the connection coefficient between all channel P1 nodes; W p1p2 is the connection coefficient of the P1 and P2 nodes of the same channel, and the connection coefficient values ​​of the P1 and P2 nodes of all channels are consistent. 1l (t), M 2l (t), G 1l (t), G 2l (t) are the potential states of the nodes M1, M2, G1, and G2 of the lth channel (l=1,…,m); Q(M 1l (t)), Q(M 2l (t)), Q(G 1l (t)), Q(G 2l (t)), Q(P 1l (t)) is the output of the nodes M1, M2, G1, G2, and P1 of the first channel, where the P1 node is the periglomerular cell module node in the olfactory perception model; Q(M 1j (t)) is the output of the node M1 of the jth channel (j=1,…,m); D1(t) is the potential state of the feedback link from the anterior olfactory nucleus to the olfactory bulb in the olfactory nerve conduction pathway; The connection coefficient between all channel M1 nodes; W m1g1 is the connection coefficient of M1 and G1 nodes; W m1g2 is the connection coefficient between M1 and G2 nodes; W m1p1 is the connection coefficient of the M1 and P1 nodes of the same channel, and the connection coefficient values ​​of all channels M1 and P1 nodes are consistent; W m1r For the same channel M1 and The connection coefficient W m1m2 is the connection coefficient of nodes M1 and M2 in the same channel.

[0060] and The sensor independent variable is fed into formula (6) to calculate the marginal contribution of the independent variable to the principal component.

[0061] (6)

[0062] in the formula For the j The marginal contribution of the independent variables to the principal components, j =1,…, m , m is the total number of independent variables; Indicates h principal components, Indicates m principal components, h =1,…, m ,; It is the main axis No. j Quantity, yes right Y The ability to explain, Y is the output vector; for and right Y The ability to explain, represents the first principal component.

[0063] In this embodiment, the variable projection importance model integrates the olfactory and taste information of the substance and determines the optimal combination relationship of the flavor information according to the size of the variable importance function of the olfactory-taste sensor, and the determined flavor information is input into the olfactory-taste synesthesia model.

[0064] Through formula (6), we can calculate the variable projection importance model and The VIP score of each sensor variable in is denoted as [VIP1,…,VIP t ], t for and The total number of sensor variables in the medium determines the flavor information components of the substance according to the VIP scores from large to small [ , ]in Represents the variable projection importance model determined The components that make up the flavor information; express The components that make up the flavor information and Substitute them into formulas (7)-(8) and (9)-(11) respectively:

[0065] (7)

[0066] (8)

[0067] (9)

[0068] (10)

[0069] (11)

[0070] Among them, E 1i (t), E 2i (t), E 3i (t) are respectively i Potential status of nodes E1, E2, and E3 of channels i =1,2,…, n , n is the number of channels of the taste perception model, n for Number of sensor variables; W r1 , W r2 , W r3 is the connection coefficient between nodes E1, E2 and E3 and RG in the same channel, and the connection coefficient values ​​of all channels are consistent; W e3e2 It is the connection coefficient between nodes E2 and E3 of the same channel. The connection coefficient value of all channels is the same; NG p ( t ) is the potential state of the peripheral noise in the taste perception model; a and b1 are the coefficients of the differential equation; RG is the external stimulus in the taste perception model. 1l (t) The potential state of the node P1 of the lth channel (l=1,…,m), where m is the number of channels in the olfactory perception model, and m is Contains the number of sensor variables; a and b2 are differential equation coefficients; Q(P 2l (t)) is the output of node P2 of the first channel; Q(P 1j (t)) is the output of the jth channel P1 node (j=1,…,m); D2(t) is the potential state of the feedback link from the anterior olfactory nucleus to the periglomerular cells in the olfactory nerve conduction pathway; W pp is the connection coefficient between all channel P1 nodes; W p1p2 is the connection coefficient of the P1 and P2 nodes of the same channel, and the connection coefficient values ​​of the P1 and P2 nodes of all channels are consistent. 1l (t), M 2l(t), G 1l (t), G 2l (t) are the potential states of the nodes M1, M2, G1, and G2 of the lth channel (l=1,…,m); Q(M 1l (t)), Q(M 2l (t)), Q(G 1l (t)), Q(G 2l (t)), Q(P 1l (t)) is the output of the nodes M1, M2, G1, G2, and P1 of the first channel, where the P1 node is the periglomerular cell module node in the olfactory perception model; Q(M 1j (t)) is the output of the node M1 of the jth channel (j=1,…,m); D1(t) is the potential state of the feedback link from the anterior olfactory nucleus to the olfactory bulb in the olfactory nerve conduction pathway; The connection coefficient between all channel M1 nodes; W m1g1 is the connection coefficient of M1 and G1 nodes; W m1g2 is the connection coefficient between M1 and G2 nodes; W m1p1 is the connection coefficient of the M1 and P1 nodes of the same channel, and the connection coefficient values ​​of all channels M1 and P1 nodes are consistent; W m1r For the same channel M1 and The connection coefficient W m1m2 is the connection coefficient of nodes M1 and M2 in the same channel.

[0071] In this embodiment, the olfactory-taste synesthesia model obtains machine taste information and flavor information, and adjusts the channel structure of the synesthesia model according to the form of input information: (1) Under taste information input, the number of taste channels is the number of taste information sensor variables, and the number of olfactory channels does not change and there is no input; (2) Under flavor information input, the number of taste channels is the number of taste sensor variables contained in the flavor information, and the number of olfactory channels is the number of olfactory sensor variables contained in the flavor information. Finally, the one-dimensional output time series of the model connection node is obtained and sent to the convolutional neural network.

[0072] In this embodiment, the convolutional neural network converts the one-dimensional output time series of the synesthesia model connection node into two-dimensional perception data, and obtains the perception features through convolution and pooling. The convolution and pooling operations are shown in formula (12), and the perception features are input into the grid search-support vector machine.

[0073] (12)

[0074] In the formula, , . Input matrix X It is the two-dimensional perception data transformed from the one-dimensional output time series of the connection nodes of the smell-taste synesthesia model, and the dimension is ( C ,D ), with taste information and flavor information input, the one-dimensional output time series dimension of the synesthesia model is (1,900). C =30, D =30, kernel matrix size Y yes( A , B ), A =3, B =3, E ( i , j ) is the calculation output, and the final output is 1×128.

[0075] In this embodiment, the grid search-support vector machine uses the convolutional neural network perception features as input data and the real sensory evaluation results as output labels to generate predicted output labels, which are then input into the olfactory action analysis model.

[0076] The support vector machine data processing process is as follows:

[0077] Assume that the feature data (data is the final output feature of the convolutional neural network) is n Dimension, total L Group data, that is .

[0078] The decision surface can be expressed as

[0079] (13)

[0080] In the formula is the weight coefficient of the decision surface, is a nonlinear mapping function, b Threshold

[0081] In order to minimize the structural risk, the optimal classification hyperplane should satisfy the following conditions

[0082] (14)

[0083] Introducing non-negative slack variables , so that the classification error is within a specified range. Therefore, the optimization problem is transformed into

[0084] (15)

[0085] In the formula c —Penalty factor, which controls the complexity and generalization ability of the model

[0086] By introducing the Lagrangian algorithm, the optimization problem is transformed into a dual form.

[0087] (16)

[0088] in a i and a j is the Lagrangian algorithm condition, i =1,2,…, n ; j =1,2,…, n , n Input data dimension for support vector machine

[0089] (17)

[0090] Introducing RBF kernel function

[0091] (18)

[0092] Where g is the kernel function parameter, which controls the range of the input space. x i and x j is a data sample.

[0093] The above optimization problem is transformed into

[0094] (19)

[0095] It can be seen that the optimization problem depends on two important parameters c and g , these two parameters will affect the performance of the support vector machine. The present invention uses grid search to optimize the parameters c and g The principle is to set each parameter c and g The possible values ​​are permuted and combined, and all possible combinations are listed to form a "grid". Each combination is then used to train the support vector machine, and the performance is evaluated using cross-validation. After the fit function tries all parameter combinations, it returns a suitable classifier that automatically adjusts to the best parameter combination. The best parameter combination is used to train the support vector machine, and the trained support vector machine produces the predicted output label.

[0096] In this embodiment, the olfactory effect analysis model is formula (20) and (21):

[0097] (20)

[0098] (twenty one)

[0099] in, For the The calculation results of the analysis model under the input of taste and flavor characteristics data of the group olfactory-taste synesthesia model, =1,2,… n,n is the total number of taste (or flavor) characteristic data; For the The prediction output of grid search-support vector machine under the input of flavor feature data of group olfactory-taste synesthesia model; For the The predicted output of grid search-support vector machine under the input of taste feature data of group olfactory-taste synesthesia model, the model olfactory and flavor features refer to the input of material olfactory and taste information into the synesthesia model, and the features extracted by convolutional neural network from the synesthesia model output; is the final calculation result, if If the calculated result is greater than 0, the machine's sense of smell has an enhanced effect on the machine's sense of taste. If the calculated result is less than 0, the machine sense of smell has an inhibitory effect on the machine sense of taste.

[0100] Experimental verification:

[0101] In the present invention, strawberry-sucrose mixed solution and bitter melon-sucrose mixed solution of different concentrations are used as examples to explain in detail the machine olfactory enhancement or machine taste inhibition analysis method of the present invention. However, in practical applications, it is not limited to strawberry or bitter melon-sucrose mixed solution, and can also be widely applied to other substances with machine olfactory enhancement or machine taste inhibition effects.

[0102] 1. Experimental equipment and materials

[0103] The samples may be strawberry-sucrose mixed solution, bitter melon-sucrose mixed solution and sucrose solution of different concentrations, and their specific parameters are as follows:

[0104] Strawberry-sucrose mixed solution 1: a mixed solution of 180 mmol / L sucrose solution and 2 g strawberry extract;

[0105] Strawberry-sucrose mixed solution 2: a mixed solution of 395.757 mmol / L sucrose solution and 4 g strawberry extract;

[0106] Bitter melon-sucrose mixed solution 1: a mixed solution of 395.757 mmol / L sucrose solution and 6 g of bitter melon extract;

[0107] Bitter melon-sucrose mixed solution 2: a mixed solution of 561.371 mmol / L sucrose solution and 6 g of bitter melon extract;

[0108] Sucrose solution 1: molar concentration is 180mmol / L;

[0109] Sucrose solution 2: molar concentration is 395.757mmol / L;

[0110] Sucrose solution 3: molar concentration is 561.371mmol / L;

[0111] The machine olfactory detection platform uses the PEN3 electronic nose, whose sensor array contains 10 metal oxide sensors that can realize cross-sensitive detection of odor information.

[0112] The machine taste detection platform uses the SA-402B electronic tongue, whose sensor array consists of 2 reference electrodes and 5 basic taste sensors; the basic taste sensors can detect the sensory information of the sample to be tested, including the five basic tastes of sour, umami, salty, bitter, and astringent, as well as the aftertaste.

[0113] (II) Specific process of obtaining olfactory information of strawberry or bitter melon-sucrose mixed solution:

[0114] (1) The experimenter puts 5 ml of solution sample into a 50 ml sample bottle, tightens the bottle cap and keeps it for 10 minutes to ensure that the gas at the top of the sealed bottle reaches saturation;

[0115] (2) Before gas detection, the experimenter needs to clean and calibrate the sensor air chamber. The specific operation is: introduce clean gas dried with activated carbon into the air chamber at a flow rate of 300ml / min and maintain it for 60s;

[0116] (3) After calibration, the test begins. Each set of samples is tested for 100 seconds to allow the sensor response value to reach a stable state. The sensor response value is G / G0 (G0 / G), where G is the conductivity of the sensor when the measured gas enters the gas chamber, and G0 is the conductivity of the sensor when pure air enters the gas chamber.

[0117] (4) Prepare 30 sets of parallel samples for each solution, and obtain 30×7=210 sets of data for strawberry or bitter melon-sucrose mixed solution (4 kinds in total) and sucrose solution (3 kinds). The conductivity value at the 60th second of the sensor response curve was taken as the characteristic value for data analysis. The experimental conditions for obtaining beer olfactory-taste information were consistent: room temperature 20±0.5℃, relative humidity 65±2%RH. Figure 2 The radar charts of olfactory information of four mixed solutions are given.

[0118] (III) Specific process of obtaining taste information of strawberry or bitter melon-sucrose mixed solution and sucrose solution:

[0119] (1) Place solution sample, reference solution and positive and negative electrode cleaning solutions;

[0120] The reference solution is a solution containing 0.3 mmol / L tartaric acid and 30 mmol / L potassium chloride; the preparation process of the positive electrode cleaning solution is: add 300 mL of 95% ethanol to about 500 mL of distilled water, stir thoroughly, add 100 mL of 1M hydrochloric acid solution, transfer the solution to a 1000 mL volumetric flask for constant volume to obtain a positive electrode cleaning solution; the preparation process of the negative electrode cleaning solution is: add 7.46 g of potassium chloride and 500 mL of 95% ethanol to about 500 mL of distilled water, stir evenly, add 10 mL of 1M potassium hydroxide solution, and transfer the solution to a 1000 mL volumetric flask for constant volume to obtain a negative electrode cleaning solution.

[0121] (2) Before the test begins, place the positive sensor array in the positive cleaning solution and the negative sensor array in the negative cleaning solution for 90 seconds. After the test, place the positive and negative sensor arrays in two containers containing reference solutions for cleaning for 120 seconds, replace the reference solution and continue cleaning for 120 seconds. Replace the reference solution again and allow the sensor to return to zero balance for 30 seconds to ensure a stable output signal.

[0122] (3) After the sensor response output reaches equilibrium, the taste information acquisition begins. The detection time for each solution sample is 30 seconds. After the measurement, it is quickly washed twice in the reference solution, and then the aftertaste value of the basic taste information is detected in the replaced reference solution. Once the measurement is completed, step (2) is repeated to clean and calibrate the sensor.

[0123] (4) Five parallel samples of strawberry or bitter melon-sucrose mixed solution (four kinds in total) and sucrose solution (three kinds) were prepared. By setting the system parameters, each group of samples was tested six times, that is, 30 sets of taste information data were obtained for each solution. After the experiment, a total of 30 × 7 = 210 sets of taste data were obtained. The voltage value of the sensor response curve at the 30th second was taken as the characteristic value for data analysis. Figure 3 A radar chart of taste information of four mixed solutions is given.

[0124] (IV) Specific process of obtaining flavor information of strawberry or bitter melon-sucrose mixed solution:

[0125] Obtaining flavor information of strawberry-sucrose mixed solution:

[0126] The olfactory information and taste information of strawberry-sucrose mixed solutions 1 and 2 are fused, and their flavor information is obtained through the variable projection importance model. Formula (22) gives its expression:

[0127] (twenty two)

[0128] in the formula For the j The marginal contribution of the independent variables to the principal components,j =1,…, m , m is the total number of independent variables; Indicates h principal components, Indicates m principal components, h =1,…, m ,; It is the main axis No. j Quantity, yes right Y The ability to explain, Y is the output vector; for and right Y The ability to explain, represents the first principal component.

[0129] Figure 4 The analysis results of the variable projection importance model are given. The flavor information set includes cpa (AAE): umami sensor aftertaste value detection; cpa (CAO): sour sensor aftertaste value detection; cpa (C0O): bitter sensor aftertaste value detection; AE1: astringency sensor; CA0: sourness sensor; CT0: saltiness sensor; cpa (AE1): astringency sensor aftertaste value detection; AAE: umami sensor; W1W, W2W, W2S, W5S, W1S are metal oxide sensors in the electronic nose).

[0130] Flavour information acquisition of bitter melon-sucrose mixed solution:

[0131] The olfactory information and taste information of bitter melon-sucrose mixed solutions 1 and 2 were fused, and their flavor information was obtained through the variable projection importance model. Figure 5 The analysis results of the variable projection importance model are given. The flavor information set includes cpa (CT0): salty sensor aftertaste value detection; cpa (AE1): astringent sensor aftertaste value detection; cpa (CAO): sour sensor aftertaste value detection; CT0: salty sensor; cpa (C0O): bitter sensor aftertaste value detection; CA0: sour sensor; cpa (AAE): umami sensor aftertaste value detection; AE1: astringent sensor; W2S, W3S, W1S, W6S, W2W, W5S, W3C, W1W are metal oxide sensors in the electronic nose).

[0132] (V) Process of obtaining sweetness perception scores of strawberry or bitter melon-sucrose mixed solution:

[0133] (1) Participants

[0134] 20. Before the experiment, all subjects were healthy (they had no cold, rhinitis, taste disorder, influenza or COVID) and had strong right hands.

[0135] (2) Experimental process

[0136] ① Training

[0137] Each subject received 3 training sessions to rate the sweetness of the stimulus. The subject was told: 1 point means no sweetness, 12 points means very sweet. We gave the subject 5 ml of sucrose solution and told the corresponding sweetness to feel the difference between different sweetness.

[0138] ②Sweetness evaluation of substances with olfactory enhancement and taste effects

[0139] Before the experiment, we ventilated the room to avoid odor interference. Sucrose solution (molar concentration of 180 mmol / L, volume of 5 ml) was identified by a random 2-digit number. Mixed solutions (concentration of 180 mmol / L sucrose solution-2 g strawberry extract, volume of 5 ml) were labeled by a random 3-digit number. All mixed solutions had the same color. Solutions were prepared at least 1 day before the experiment and refrigerated. Solutions were taken out of the refrigerator and placed in a dry bath incubator 1 hour before the test to maintain room temperature. We knew all participants not to eat or smoke at least 1 hour before the experiment. All participants were evaluated in separate rooms. We invited four participants at a time.

[0140] At the beginning of the experiment, we asked the participants to sit on a chair, wear a blindfold (to avoid color interference from the solution or other objects), and keep quiet. We randomly selected 2-digit 5 ​​ml solutions and instructed the participants to taste with a straw in the first round. Participants gave sweetness perception scores. After tasting, participants rinsed their mouths with pure water 3 times and relaxed for three minutes (after each tasting, participants performed the same treatment). We randomly selected 3-digit 5 ​​ml solutions in the second round and instructed the participants to taste. Then, participants gave scores. We randomly selected 3-digit 5 ​​ml solutions in the third round, but we instructed the participants to wear a medical cotton swab to plug their noses. Participants gave sweetness evaluation scores. The ratio of the solution was still 180 mmol / L sucrose solution-2 g strawberry extract. In each round, we gave four participants solution samples to balance the delivery order and instructed the participants to wear a blindfold during the experiment to avoid visual interference. After three rounds, one experiment was completed. Each participant repeated the experiment 3 times, and we took the average of the 3 results.

[0141] Similar to the above experimental process, we used sucrose solution samples with a molar concentration of 395.757 mmol / L and mixed solutions (the ratio was 395.757 mmol / L sucrose solution-4 g strawberry extract) to perceive sweetness and obtain evaluation scores.

[0142] ③Sweetness evaluation of substances with olfactory inhibitory taste effects

[0143] The sweetness evaluation process was consistent with the above experiment. It is worth noting that the samples were replaced by sucrose solutions (molar concentrations of 395.757mmol / L and 561.371mmol / L) and mixed solutions (ratios of 395.757mmol / L sucrose solution-6g bitter melon extract and 561.371mmol / L sucrose solution-6g bitter melon extract). All mixed solutions had the same color.

[0144] (VI) The taste and flavor information of strawberry or bitter melon-sucrose mixed solution and the taste information of sucrose solution are input into the olfactory-taste synesthesia model, and the model output features are extracted through the convolutional neural network:

[0145] ① The taste information of strawberry-sucrose mixed solutions 1 and 2 (60 groups in total) was input into the taste channel of the olfactory-taste synesthesia model, and the input of the olfactory channel was 0; the flavor information of strawberry-sucrose mixed solutions 1 and 2 (60 groups in total) was input into the olfactory-taste synesthesia model, among which cpa (AAE), cpa (CAO), cpa (C0O), AE1, CA0, CT0, cpa (AE1), AAE were input into the taste channel of the olfactory-taste synesthesia model, and W1W, W2W, W2S, W5S, W1S were input into its olfactory channel. The taste information of sucrose solution (molar concentration of 180mmol / L and 395.757mmol / L) (60 groups in total) was input into the taste channel of the olfactory-taste synesthesia model, and its olfactory channel had no input. The dynamic characteristic equation of the smell-taste synesthesia model is as follows, where formulas (23)-(25) are the dynamic characteristic equations of the smell-taste synesthesia model for the cranial neural modules (E1, E2, E3):

[0146] (twenty three)

[0147] (twenty four)

[0148] (25)

[0149] E 1i (t), E 2i (t), E 3i (t) are respectively i Potential status of nodes E1, E2, and E3 of channels i =1,2,…,n , n is the number of channels in the taste perception model; W r1 , W r2 , W r3 is the connection coefficient between nodes E1, E2 and E3 and RG in the same channel, and the connection coefficient values ​​of all channels are consistent; W e3e2 It is the connection coefficient between nodes E2 and E3 of the same channel. The connection coefficient value of all channels is the same; NG p ( t ) is the potential state of the peripheral noise in the taste perception model; a and b1 are the coefficients of the differential equation; RG is the external stimulus in the taste perception model; RG i ( t ) is the i The potential state of external stimuli in the channel taste perception model.

[0150] Formulas (26)-(29) are the dynamic characteristic equations of the nucleus tractus solitarius module in the olfactory-taste synesthesia model:

[0151] (26) (27)

[0152] (28)

[0153] (29)

[0154] Where a and b1 are the coefficients of the differential equation, n is the number of channels of the taste perception model; J 1i (t), J 2i (t), K 1i (t), K 2i (t) are the potential states of nodes J1, J2, K1, and K2 of the i-th channel (i=1,…,n); Q(J 1i (t)), Q(J 2i (t)), Q(K 1i (t)), Q(K 2i (t)), Q(E 1i (t)), Q(E 2i (t)), Q(E 3i (t)) and Q(J 5i (t)) is the output of the nodes J1, J2, K1, K2, E1, E2, E3, and J5 of the i-th channel; E1, E2, and E3 are the nodes of the facial nerve module, the glossopharyngeal nerve module, and the vagus nerve module in the cranial nerve module in the taste perception model respectively; Q(J 1j (t)) is the output of node J1 of the jth channel (j=1,…,n); Q(K 1j(t)) is the output of the node K1 of the jth channel; D6(t) is the potential state of the feedback link from the insular cortex to the solitary nucleus in the taste nerve conduction pathway; W jj is the connection coefficient between all channel J1 nodes; W j1j2 , W j2j1 is the connection coefficient of the J1 and J2 nodes of the same channel, and the connection coefficient values ​​of the J1 and J2 nodes of all channels are consistent; W j1k1 , W k1j1 is the connection coefficient of the J1 and K1 nodes of the same channel, and the connection coefficient values ​​of the J1 and K1 nodes of all channels are consistent; W j1k2 , W k2j1 is the connection coefficient of the J1 and K2 nodes of the same channel, and the connection coefficient values ​​of the J1 and K2 nodes of all channels are consistent; W j1e1 is the connection coefficient of the same channel J1 and E1 node, and the connection coefficient value of all channels J1 and E1 nodes is consistent; W j1e2 is the connection coefficient of the same channel J1 and E2 nodes, and the connection coefficient values ​​of all channels J1 and E2 nodes are consistent; W j1e3 is the connection coefficient of the same channel J1 and E3 node, and the connection coefficient value of all channels J1 and E3 nodes is consistent; W j1j5 is the connection coefficient of the J1 and J5 nodes of the same channel, and the connection coefficient values ​​of the J1 and J5 nodes of all channels are consistent; W j2k1 , W k1j2 is the connection coefficient of the J2 and K1 nodes of the same channel, and the connection coefficient values ​​of the J2 and K1 nodes of all channels are consistent; W kk is the connection coefficient between all channel K1 nodes; W k1k2 , W k2k1 is the connection coefficient of the K1 and K2 nodes of the same channel. The connection coefficient values ​​of the K1 and K2 nodes of all channels are consistent. d6 is the connection coefficient between the J1 node and the feedback link D6; i, j are both from 1 to n, i≠j, and n is the number of channels of the taste perception model.

[0155] Formulas (30)-(33) are the dynamic characteristic equations of the ventroposterior medial nucleus of the thalamus module in the olfactory-taste synesthesia model:

[0156] (30)

[0157] (31)

[0158] (32)

[0159] (33)

[0160] Wherein, a and b1 are coefficients of differential equations, n is the number of channels of the taste perception model; J3(t), J4(t), K3(t), K4(t) are the potential states of the J3, J4, K3, K4 nodes; Q(J3(t)), Q(J4(t)), Q(K3(t)), Q(K4(t)), Q(B3(t)) are the outputs of the J3, J4, K3, K4, B3 nodes; Q(J 1j (t)) is the output of the node J1 of the jth channel, where j ranges from 1 to n; D5(t) is the potential state of the feedback link from the insular cortex to the ventroposterior medial nucleus of the thalamus in the taste conduction pathway; W j3j1 is the connection coefficient of the J1 node and the J3 node of all channels; W j3j4 , W j4j3 is the connection coefficient of nodes J3 and J4; W j3k3 , W k3j3 is the connection coefficient of J3 and K3 nodes; W j3k4 , W k4j3 is the connection coefficient of nodes J3 and K4; W j4k3 , W k3j4 is the connection coefficient of nodes J4 and K3; W k3k4 , W k4k3 is the connection coefficient of K3 and K4 nodes; K d5 is the connection coefficient between the J3 node and the feedback link D5; The potential state of the central noise of the taste perception model.

[0161] Formula (34) is the dynamic characteristic equation of the insular cortex module of the olfactory-taste synesthesia model:

[0162] (34)

[0163] M(t) represents The potential state of the node, a and are coefficients of differential equations, j =1,…, n , n is the number of channels of the taste perception model; W mj1 and W mk3 All are connection coefficients; Q(J 1j Q(K3(t)) is the output of the node K3.

[0164] Formulas (35)-(36) are the dynamic characteristic equations of the periglomerular cell module of the olfactory-taste synesthesia model:

[0165] (35)

[0166] (36)

[0167] Among them, P 1l (t), P 2l (t) is the potential state of the nodes P1 and P2 of the lth channel (l=1,…,m), m is the number of channels of the olfactory perception model; a and b2 are the coefficients of the differential equation; Q(P 1l (t)) is the output of node P1 of the first channel; Q(P 1j (t)) is the output of the jth channel P1 node (j=1,…,m); D2(t) is the potential state of the feedback link from the anterior olfactory nucleus to the periglomerular cells in the olfactory nerve conduction pathway; W pp is the connection coefficient between all channel P1 nodes; W p1p2 It is the connection coefficient of the P1 and P2 nodes of the same channel. The connection coefficient values ​​of the P1 and P2 nodes of all channels are consistent.

[0168] Formulas (37)-(40) are the dynamic characteristic equations of the olfactory bulb module of the smell-taste synesthesia model:

[0169] (37)

[0170] (38)

[0171] (39)

[0172] (40)

[0173] Among them, M 1l (t), M 2l (t), G 1l (t), G 2l (t) are the potential states of the nodes M1, M2, G1, and G2 of the lth channel (l=1,…,m), and m is the number of channels in the olfactory perception model; a and b2 are the coefficients of the differential equation; Q(M 1l (t)), Q(M 2l (t)), Q(G 1l (t)), Q(G 2l (t)), Q(P 1l (t)), is the output of the nodes M1, M2, G1, G2, and P1 of the first channel, where the P1 node is the periglomerular cell module node in the olfactory perception model; Q(M 1j (t)) is the output of node M1 of the jth channel (j=1,…,m); Q(G 1j(t)) is the output of the node G1 of the jth channel; D1(t) is the potential state of the feedback link from the anterior olfactory nucleus to the olfactory bulb in the olfactory nerve conduction pathway; D4(t) is the potential state of the feedback link from the outer cortex to the olfactory bulb in the olfactory nerve conduction pathway; and is the connection coefficient between all channels M1 and G1 nodes; W m1g1 , W g1m1 is the connection coefficient of the M1 and G1 nodes of the same channel, and the connection coefficient values ​​of all channels M1 and G1 nodes are consistent; W m1g2 , W g2m1 is the connection coefficient of the M1 and G2 nodes of the same channel, and the connection coefficient values ​​of all channels M1 and G2 nodes are consistent; W m1p1 is the connection coefficient of the M1 and P1 nodes of the same channel, and the connection coefficient values ​​of all channels M1 and P1 nodes are consistent; W m1r For the same channel M1 and RO l Connection coefficients for all channels M1 and RO l The connection coefficient value of RO is consistent; l is the external stimulus of the lth channel, and the external stimulus of all channels is the same; RO l (t) is the potential state of the external stimulus of the first channel; W m2g1 , W g1m2 is the connection coefficient of the M2 and G1 nodes of the same channel, and the connection coefficient values ​​of all channels M2 and G1 nodes are consistent; W g1g2 is the connection coefficient of the G1 and G2 nodes of the same channel, and the connection coefficient values ​​of all channels G1 and G2 nodes are consistent; W m1m2 is the connection coefficient of the M1 and M2 nodes of the same channel, and the connection coefficient values ​​of the M1 and M2 nodes of all channels are consistent.

[0174] Formulas (41)-(44) are the dynamic characteristic equations of the anterior olfactory nucleus module of the olfactory-taste synesthesia model:

[0175] (41)

[0176] (42)

[0177] (43)

[0178] (44)

[0179] Where a and b2 are the coefficients of the differential equation, m is the number of channels of the olfactory perception model; F1(t), F2(t), I1(t), I2(t) are the potential states of the F1, F2, I1, I2 nodes respectively; Q(F1(t)), Q(F2(t)), Q(I1(t)), Q(I2(t)), Q(A3(t)) are the outputs of the F1, F2, I1, I2, A3 nodes; Q(M 1k (t)) is the output of the M1 node of the kth channel, and the M1 node is the key node of the olfactory bulb in the olfactory conduction model (k=1,…,m); W i1a3 is the connection coefficient between I1 and A3 nodes; W f1m is the connection coefficient between the F1 node and the A1 node of all channels; W f1f2 , W f2f1 is the connection coefficient of nodes F1 and F2; W f1i1 , W i1f1 is the connection coefficient of nodes F1 and I1; W f1i2 , W i2f1 is the connection coefficient of F1 and I2 nodes; W f2i1 and W i1f2 is the connection coefficient between F2 and I1 nodes; W i1i2 , W i2i1 is the connection coefficient of nodes I1 and I2.

[0180] Formulas (45)-(48) are the dynamic characteristic equations of the anterior piriform cortex module of the olfactory-taste synesthesia model:

[0181] (45)

[0182] (46)

[0183] (47)

[0184] (48)

[0185] Where a and b2 are differential equation coefficients, m is the number of channels in the olfactory perception model; A1(t), A2(t), B3(t), B4(t) are the potential states of nodes A1, A2, B3, B4 respectively; Q(A1(t)), Q(A2(t)), Q(B3(t)), Q(B4(t)), Q(C(t)) are the outputs of nodes A1, A2, B3, B4, C, where node C is the topological form of the outer cortex module in the olfactory perception model; Q(M 1k (t)) is the output of the M1 node of the kth channel, and the M1 node is the key node of the olfactory bulb in the olfactory conduction model (k=1,…,m); W b3c is the connection coefficient between nodes B3 and C; Wa1m is the connection coefficient between the M1 node and the A1 node of all channels; W a1a2 , W a2a1 is the connection coefficient of nodes A1 and A2; W a1b3 , W b3a1 is the connection coefficient of nodes A1 and B3; W a1b4 , W b4a1 is the connection coefficient of nodes A1 and B4; W a2b3 and W b3a2 is the connection coefficient of nodes A2 and B3; W b3b4 , W b4b3 is the connection coefficient of nodes B3 and B4.

[0186] Formula (49) is the dynamic characteristic equation of the connection module of the olfactory-taste synesthesia model:

[0187] (49)

[0188] Where a and b1 are the coefficients of the differential equation, s is the number of channels of the taste perception model minus 1; N l (t) is the potential state of the Nth node of the lth channel, l = 1, 2, ..., n; n is the number of channels in the taste perception model; Q(K3(t)), Q(M(t)), Q(C(t)) are the outputs of the K3, M, C nodes; Q(N j (t)) is the output of the N-node of the j-th channel, where l ranges from 1 to s; W nn is the connection coefficient between all channel N nodes; W nk3 is the connection coefficient between K3 and N nodes; W nm is the connection coefficient between N and M nodes; W nc is the connection coefficient between N and C nodes; Q(C(t)) is the output of the outer cortex module, and Q(M(t)) is the output of the insular cortex module.

[0189] Formula (50) is the dynamic characteristic equation of the feedback link of the olfactory-taste synesthesia model:

[0190] (50)

[0191] and is the feedback parameter, in the olfactory model i From 1 to 4, the taste model i From 5 to 6.

[0192] Next, we extracted the N-node output time series of the orbitofrontal cortex of the olfactory-taste synesthesia model (data dimension was 1×900, 180 groups in total), and the convolutional neural network ( Figure 6) extracts the features of the N-node output time series. The convolutional neural network processing process is as follows: the input dimension is 30×30, and its shape is 32×32 by "same" padding. The outputs of other convolutional layers are processed by "valid" padding. 'Ci' represents the i-th convolution operation, and the convolution operation is shown in formula (51). 'Pj' represents the j-th pooling operation (i=1,2,3,4; j=1,2,3). In all convolution operations, the size of the convolution kernel is 3×3 and the stride is 1. In all downsampling operations, the stride is 2 and the filter size is 2×2. We use average pooling, and the formula for calculating the output size after pooling is shown in (52)-(53).

[0193] (51)

[0194] In the formula, , . Input matrix X The dimension is ( C , D ), kernel matrix size Y yes( A , B ), E ( i , j ) is the calculation output.

[0195] (52)

[0196] (53)

[0197] × is the input data size, F × F is the convolution kernel size, S is the step size, × is the output data size.

[0198] Finally, after two layers of convolution calculation, 8 feature matrices of size 28×28 are obtained, and average pooling operations are performed on feature selection and information filtering. The 8 feature matrices of size 28×28 become 14×14. After the remaining convolution and pooling processes, 32 feature matrices of size 2×2 are finally obtained and shaped into a feature matrix of 1×128 as features.

[0199] Finally, we obtained two feature sets: strawberry-sucrose mixed solution feature set 1 (the total data dimension is 90×128, including the taste and flavor information of strawberry-sucrose mixed solution 1 and the feature set under the taste information input of sucrose solution 1) and strawberry-sucrose mixed solution feature set 2 (the total data dimension is 90×128, including the taste and flavor information of strawberry-sucrose mixed solution 2 and the feature set under the taste information input of sucrose solution 2).

[0200] ② The taste information of bitter melon-sucrose mixed solutions 1 and 2 (60 groups in total) was input into the taste channel of the olfactory-taste synesthesia model, and the input of the olfactory channel was 0; the flavor information of bitter melon-sucrose mixed solutions 1 and 2 (60 groups in total) was input into the olfactory-taste synesthesia model, among which cpa (CT0), cpa (AE1), cpa (CAO), CT0, cpa (C0O), CA0, cpa (AAE), AE1 were input into the taste channel of the olfactory-taste synesthesia model, and W2S, W3S, W1S, W6S, W2W, W5S, W3C, W1W were input into its olfactory channel. The taste information of sucrose solution (molar concentration of 395.757mmol / L and 561.371.757mmol / L) (60 groups in total) was input into the taste channel of the olfactory-taste synesthesia model, and its olfactory channel had no input.

[0201] Similarly, next, the N-node output time series of the orbitofrontal cortex of the olfactory-taste synesthesia model was extracted (data dimension was 1×900, 180 groups in total), and the convolutional neural network extracted the features of the N-node output time series. Finally, two feature sets were obtained: bitter melon-sucrose mixed solution feature set 1 (total data dimension was 90×128, including the taste and flavor information of bitter melon-sucrose mixed solution 1 and the feature set under the taste information input of sucrose solution 2) and bitter melon-sucrose mixed solution feature set 2 (total data dimension was 90×128, including the taste and flavor information of bitter melon-sucrose mixed solution 2 and the feature set under the taste information input of sucrose solution 3).

[0202] 7. Analysis of the effects of smell on taste

[0203] Next, the present invention inputs the taste and flavor feature set in the strawberry-sucrose mixed solution feature set 1 (or 2) into the grid search-support vector machine, and inputs the output of the grid search-support vector machine into formulas (20) and (21).

[0204] The support vector machine data processing process is as follows:

[0205] Assume that the feature data (data is the final output feature of the convolutional neural network) is n Dimension, total L Group data, that is .

[0206] The decision surface can be expressed as

[0207] (54)

[0208] In the formula is the weight coefficient of the decision surface, is a nonlinear mapping function, b Threshold

[0209] In order to minimize the structural risk, the optimal classification hyperplane should satisfy the following conditions

[0210] (55)

[0211] Introducing non-negative slack variables , so that the classification error is within a specified range. Therefore, the optimization problem is transformed into

[0212] (56)

[0213] In the formula c —Penalty factor, which controls the complexity and generalization ability of the model

[0214] By introducing the Lagrangian algorithm, the optimization problem is transformed into a dual form.

[0215] (57)

[0216] in a i and a j is the Lagrangian algorithm condition, i =1,2,…, n ; j =1,2,…, n , n Input data dimension for support vector machine

[0217] (58)

[0218] Introducing RBF kernel function

[0219] (59)

[0220] Where g is the kernel function parameter, which controls the range of the input space. x i and x j is a data sample.

[0221] The above optimization problem is transformed into

[0222] (60)

[0223] It can be seen that the optimization problem depends on two important parameters c and g , these two parameters will affect the performance of the support vector machine. The present invention uses grid search to optimize the parameters c and g The principle is to set each parameter c and g The possible values ​​are permuted and combined, and all possible combinations are listed to form a "grid". Each combination is then used to train the support vector machine, and the performance is evaluated using cross-validation. After the fit function tries all parameter combinations, it returns a suitable classifier that automatically adjusts to the best parameter combination. The best parameter combination is used to train the support vector machine, and the trained support vector machine produces the predicted output label.

[0224] The calculation results are shown in Table 1. The calculated result is 0.1584 (0.1587), where the values ​​in brackets represent the calculated results under the input of taste and flavor feature sets in feature set 2 of the strawberry-sucrose mixed solution. It can be seen that the calculated results are all greater than 0, which indicates that the calculated result under the input of flavor features is greater than the taste features, that is, the synesthesia model's sweetness perception intensity of the flavor information of the mixed solution (including olfactory and taste information) is higher than that of the taste information, that is, the introduction of smell can enhance the perception of sweetness, that is, when people taste the strawberry-sucrose mixed solution, the machine sense of smell will enhance the machine taste perception.

[0225] Similarly, Table 2 shows the calculation results of the taste and flavor feature set input in bitter melon-sucrose mixed solution feature set 1 (or 2). All the calculation results are less than 0, and it is finally concluded that the machine olfactory sense will inhibit the machine taste perception when people taste the bitter melon-sucrose mixed solution.

[0226] The working principle of the method for analyzing machine olfactory enhancement or machine taste inhibition of the present invention is as follows: through an olfactory-taste sensor, which can be an electronic nose-electronic tongue, the taste and flavor information of a substance with olfactory enhancement or taste inhibition effect is obtained and input into an olfactory-taste synesthesia model, and in addition, a human sweetness evaluation score of the substance is obtained through sensory evaluation; the output feature of the olfactory-taste synesthesia model under information input is extracted through a convolutional neural network, and the feature is used as input data, and the sweetness evaluation score is sent as an output label to a grid search-support vector machine model, and the grid search-support vector machine model generates a prediction label; the prediction label is input into the olfactory effect analysis model, and the model calculation result qualitatively points out that the human sense of smell will enhance the taste perception when tasting a strawberry-sucrose mixed solution, and the human sense of smell will inhibit the taste perception when tasting a bitter melon-sucrose mixed solution.

[0227] Table 1 Calculation results of the analytical model under the input of taste and flavor information of strawberry-sucrose mixed solution

[0228]

[0229] Table 2 Calculation results of the analysis model under the input of taste and flavor information of bitter melon-sucrose mixed solution

[0230]

[0231] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for analyzing the enhancement of machine olfaction or the inhibition of machine taste, characterized in that: include: Machine olfactory model, machine taste model, sensory evaluation model, variable projection importance model, smell-taste synesthesia model, convolutional neural network, grid search-support vector machine and olfactory action analysis model; The machine smell and machine taste detection results are sent to a variable projection importance model, the machine taste detection results and the analysis results of the variable projection importance model are input into an olfactory-taste synesthesia model, the output of the olfactory-taste synesthesia model is sent to a convolutional neural network, the convolutional neural network output and the sensory evaluation results are input into a grid search-support vector machine, the grid search-support vector machine generates a prediction output and sends it to an olfactory effect analysis model, and the enhancement or inhibition of the machine smell on the machine taste is qualitatively reflected according to the model calculation results; The machine olfactory model and the machine taste model are used to obtain the olfactory and taste information of the flavor substances, and send the olfactory and taste information to the variable projection importance model, and send the taste information to the olfactory-taste synesthesia model; The variable projection importance model integrates the olfactory and taste information of the substance and determines the optimal combination relationship of the flavor information according to the magnitude of the olfactory-taste sensor variable importance function, and the determined flavor information is input into the olfactory-taste synesthesia model; The olfactory-taste synesthesia model obtains machine taste information and flavor information, and adjusts the synesthesia model channel structure according to the input information form, including: Under taste information input, the number of taste channels is the number of taste information sensor variables, and the number of olfactory channels does not change and has no input; Under the input of flavor information, the number of taste channels is the number of taste sensor variables contained in the flavor information, and the number of olfactory channels is the number of olfactory sensor variables contained in the flavor information. Finally, the one-dimensional output time series of the model connection node is obtained and sent to the convolutional neural network; The convolutional neural network converts the one-dimensional output time series of the connection nodes of the synesthesia model into two-dimensional perception data, and obtains perception features through convolution and pooling, and the perception features are input into the grid search-support vector machine; The grid search-support vector machine uses the convolutional neural network perception features as input data and the real sensory evaluation results as output labels to generate predicted output labels, which are input into the olfactory action analysis model; The olfactory effect analysis model is formula (1) and (2): Among them, S a (i) is the calculation result of the analysis model under the input of the taste and flavor characteristic data of the i-th group of olfactory-taste synesthesia model, i=1, 2, ...n, n is the total number of taste characteristic data; S f (i) is the prediction output of the grid search-support vector machine under the input of the flavor feature data of the i-th group of olfactory-taste synesthesia model; S t (i) is the predicted output of the grid search-support vector machine under the input of the taste feature data of the ith group of olfactory-taste synesthesia model, where the model olfactory and flavor features refer to the input of the material olfactory and taste information into the synesthesia model, and the features extracted by the convolutional neural network from the synesthesia model output; S a is the final calculation result. If S a If the calculated result is greater than 0, the machine's sense of smell has an enhanced effect on the machine's sense of taste. a If the calculated result is less than 0, the machine sense of smell has an inhibitory effect on the machine sense of taste.

Images