CN108981800B

CN108981800B - Method for visualizing machine smell-taste perception effect by using neuro-dynamics system model

Info

Publication number: CN108981800B
Application number: CN201810695541.9A
Authority: CN
Inventors: 门洪; 石岩; 焦亚楠; 巩芙榕; 房海瑞; 刘晶晶
Original assignee: Northeast Dianli University
Current assignee: Northeast Electric Power University
Priority date: 2018-06-25
Filing date: 2018-06-25
Publication date: 2023-09-08
Anticipated expiration: 2038-06-25
Also published as: CN108981800A

Abstract

The invention discloses a method for visualizing a machine sniffing-taste-perception effect by using a neuro-dynamics system model, which comprises the following steps: s1, acquiring taste and smell information of a beer machine by utilizing an SA-402B type electronic tongue and a PEN3 type electronic nose; s2, extracting information characteristics affecting beer flavor based on variable projection importance evaluation indexes and multi-mode recognition technologies (a support vector machine, a random forest and an extreme learning machine); s3, based on the KIII nonlinear neural dynamics system model, inputting the time dimension for influencing the characteristics of the beer flavor information, and visualizing the action rule of the smell information to dynamically present. The invention relates to a sensing process of machine smell and machine taste based on an action mode of processing external stimulus by a nonlinear dynamics nervous system of a nerve conduction mechanism, which reflects a smell-taste information action rule in the nonlinear dynamics system and has important significance in realizing the sensing research of machine smell-taste.

Description

Method for visualizing machine smell-taste perception effect by using neuro-dynamics system model

Technical Field

The invention relates to the technical field of machine perception intelligent evaluation, in particular to a method for visualizing a machine smell-taste effect by using a neural dynamics system model.

Background

The machine perception technology is an important research direction in the automation field based on artificial intelligence driving, wherein the machine olfactory perception (electronic nose) and the machine gustatory perception (electronic tongue) are used for realizing intelligent acquisition of sensory information by simulating the working principles of biological olfactory sense and gustatory sense. However, based on direct fusion of sensory information of the electronic nose/tongue multisensor, the simultaneous perception phenomenon of processing olfactory and gustatory information in the inside of a neuromotor system is not considered, and the asynchronism of the olfactory and gustatory in the time dimension of sensory experience is not reflected.

Human perception of flavoring substances results mainly from the synergistic effect of smell and taste, resulting in sensory mixing. Because of the strong correlation between free tasting substances and soluble volatile substances, the perception of some volatile substances is often mistaken for a "taste sensation". The sniff-to-taste response time dyssynchrony can be divided into "anterior nasal effects" and "posterior nasal effects". The "front nose effect" refers to smell information firstly smelling food, and the "rear nose effect" refers to obvious aftertaste information after swallowing food, and all have relevant influence on sensory evaluation quality.

The model framework structure is constructed from the mathematical level based on the artificial intelligence mode recognition method, the statistical principle and the numerical analysis method are applied to obtain classification results, such as K-means clustering, support vector machines, extreme learning machines and the like, which do not have the time dimension input mode of sensing information, and multi-sensing characteristic information from the electronic tongue and the electronic nose are input in parallel to obtain a comprehensive information output result. Such parallel input mechanisms tend to be unable to embody the sensory alliance phenomenon in the actual sensory evaluation process.

Disclosure of Invention

In order to solve the problems, the invention provides a method for visualizing the machine smell-taste combined effect by using a neural system model, which is based on the action mode of processing external stimulus by a nonlinear dynamic neural system of a nerve conduction mechanism, correlates machine smell with the perception process of machine taste, reflects the action rule of smell-taste information in the nonlinear dynamic system, and realizes the perception research of machine smell-taste combined effect.

In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:

a method for machine sniffing-taste-perception effect visualization using a thermodynamic system model, comprising the steps of:

s1, acquiring taste and smell information of a beer machine by utilizing an SA-402B type electronic tongue and a PEN3 type electronic nose;

s2, extracting information features affecting beer flavor based on variable projection importance evaluation index (VIP) and multi-mode recognition technology (support vector machine, random forest and extreme learning machine);

s3, based on the KIII nonlinear neural dynamics system model, inputting the time dimension for influencing the characteristics of the beer flavor information, and visualizing the action rule of the smell information to dynamically present.

In the step S1, the machine taste and machine smell sensing detection obtains 20-dimensional simultaneous sense data, wherein the machine taste information takes the voltage value of the sensor at the 30 th S as a characteristic value, the 5 basic taste sensors obtain 10-dimensional characteristic data, the machine smell information takes the conductivity value of the sensor at the 60 th S as a characteristic value, and the 10 sensors obtain 10-dimensional characteristic data.

Wherein, the time dimension input in the step S3 comprises three stages of 'front nose effect', 'smell-taste coexistence', 'rear nose effect'; the simultaneous data are entered in the model sequentially in three different time periods.

Wherein the multi-pattern recognition technique comprises.

The invention has the following beneficial effects:

the action rules of the olfactory and gustatory information are visualized and dynamically presented, and the action mode of the nonlinear dynamic nervous system based on nerve conduction mechanism for treating external stimulus is searched.

Drawings

Fig. 1 is a technical flowchart in an embodiment of the present invention.

Fig. 2 is a 20-dimensional histogram of VIP scores for feature variables.

FIG. 3 is a graph of GA-SVM parameter optimization.

FIG. 4 is a graph of the impact of decision tree number on RF classification performance.

FIG. 5 is a graph showing the effect of hidden layer neuron count on ELM classification performance.

Fig. 6 is a visual diagram of the input of the illusion information in the time dimension of the beer smell-taste information 12 channel.

Detailed Description

The present invention will be described in further detail with reference to examples in order to make the objects and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

Examples

Experimental materials: the parameters of the brands, the alcohol degree, the raw juice wheat concentration and the raw materials of the 5 different brands of beer are shown in the table 1. 10 repeated samples are selected for each brand of beer, and in order to ensure the identity of substances in the detection process, the electronic nose experiment and the electronic tongue experiment of each beer are synchronously carried out.

Table 1 5 beer brands

The bionic taste detection platform adopts an SA-402B type electronic tongue system developed by Japanese Instrument company, and the sensor array of the experiment consists of 2 reference electrodes and 5 basic taste sensors. The basic taste sensor can realize the sensory information detection of 5 basic tastes including acid, fresh, salty, bitter and astringent taste of the sample to be detected. The electronic nose system adopts PEN3 electronic nose developed and developed by Airsense corporation in Germany, and the gas sensor array comprises 10 metal oxide sensors, so that the detection of odor cross-sensitive information can be realized.

1. Data acquisition experimental procedure

Electronic tongue experiment:

(1) Placing beer sample, reference solution and positive and negative electrode cleaning solution.

(2) Before the test starts, the positive and negative electrode sensor arrays are washed in a washing solution for 90s, after the test is finished, the positive and negative electrode sensor arrays are washed in a reference solution for 120s, the positive and negative electrode sensor arrays are washed in another reference solution for 120s, and finally, the reference solution is put into the positive and negative electrode sensor arrays to balance and zero the sensors for 30s, so that the stability of output signals is ensured.

(3) After the response output of the sensor reaches equilibrium, beer taste information acquisition is started, the detection time of each beer is 30s, the measurement is finished, the beer taste information is quickly cleaned in the reference solution for 2 times, the measurement is returned to the reference solution for detecting the aftertaste value of the basic taste information, the measurement is finished once, and the step (2) is repeated to clean and calibrate the sensor.

(4) 3 parallel samples are prepared for beer of different brands, and each group of samples is repeatedly detected for 6 times by setting system parameters, namely 18 groups of taste information data are intelligently acquired for beer of each brand, and 90 groups of taste data are acquired after the experiment is finished.

The experimental condition of the electronic tongue is that the room temperature is 20+/-0.5 ℃ and the relative humidity is 65+/-2% RH. The voltage value of the 30s of the sensor response curve is taken as a characteristic value for data analysis.

Electronic nose experiment:

(1) 5mL of beer is taken and placed into a 50mL sample bottle, and the bottle cap is screwed down and kept still for 10 minutes, so that the gas at the top of the sealed bottle is ensured to reach a saturated state.

(2) Before the detection of the gas starts, the sensor air chamber is cleaned and calibrated, and clean air with the flow rate of 300mL/min after activated carbon treatment is introduced into the air chamber for 60s.

(3) And after calibration, starting detection, and detecting each group of samples for 80 seconds to ensure that the response information of the sensor reaches a stable value. The response of the sensor is defined as G/G0 (G0/G), where G is the conductance of the sensor when the measured gas enters the chamber and G0 is the conductance of the sensor when pure air enters the chamber.

(4) 18 sets of replicates were prepared for each beer, and a total of 90 sets of data were obtained for 5 beers.

The experimental conditions of the electronic nose are that the room temperature is 20+/-0.5 ℃ and the relative humidity is 65+/-2% RH. The conductivity value of the sensor response curve at 60s is taken as a characteristic value for data analysis.

2. Simultaneous sense data feature mining

The 90 groups of data contained in the electronic tongue and electronic nose detection data are randomly divided into two groups, one group contains 72 groups of data serving as a training set for establishing a classification model, and the other group contains 18 groups of data serving as a test set for checking classification performance of the model. In order to accelerate the convergence rate of the training model and simultaneously convert the multi-dimensional data into a unified dimensionless form, the electronic tongue/nose detection data needs to be normalized, and the normalization interval is (-1, 1).

And obtaining the importance ranking of the 20-dimensional original smell-taste sensory information according to the variable projection importance evaluation index (VIP score), wherein the VIP score ranking of the 20-dimensional characteristic variable is shown in figure 2. And then the VIP scores are sequentially overlapped to form 20 variable subsets, and feature variable screening is carried out by combining a Support Vector Machine (SVM), a Random Forest (RF) and an Extreme Learning Machine (ELM) multi-mode identification technology.

Table 2 shows the prediction accuracy under the SVM, RF, ELM classification model of 20 variable subsets formed by sequentially superimposing based on VIP score order. As the number of variables increases, the classification accuracy tends to rise gradually. The recognition accuracy of the SVM and ELM at the 7 th variable subset and the recognition accuracy of the RF at the 9 th variable subset are respectively equal to that of the original fusion data set, which shows that the original fusion data contain a large amount of redundant information in terms of beer flavor information recognition. With continued superposition of variables, the SVM exhibited a highest recognition accuracy of 96.67% under the 12 th variable subset, the RF exhibited a highest recognition accuracy of 94.44% under the 11 th variable subset, and the ELM exhibited a highest recognition accuracy of 98.33% under the 12 th variable subset. The number of variables is continuously increased, and the recognition accuracy under each model does not exceed the maximum value. Thus, by combining the action behaviors among the variables with the multi-mode identification method, the optimal variable combination is obtained by dynamic characterization and is the 12 th variable subset. Wherein, 3, 4, 5, 6, 7, 10 and 12 are olfactory information, 1, 2, 8 and 11 are basic information of taste sense, and 9 is information of aftertaste sense.

FIG. 3 is a graph of the parameter optimization of the GA-SVM under the 12 th variable subset five-fold cross-validation method. Fig. 4 is a plot of the change in the influence of the variable subset decision tree number 11 on the RF classification performance. FIG. 5 is a graph showing the effect of the number of hidden layer neurons in the 12 th variable subset on ELM classification performance.

TABLE 2 prediction accuracy for subsets of variables based on multimodal recognition techniques

3. Action law visual dynamic presentation of smell information

And (3) establishing a KIII nerve dynamics perception model of 12 channels, wherein the operation time is 700ms, and the operation step length is 1ms. In the first stage of the "anterior nasal effect", it is prescribed that olfactory information is input between 51-200ms, i.e., 3, 4, 5, 6, 7, 10 and 12 channels, in the second stage of the "olfactory-taste coexistence", olfactory-taste coexistence information other than gustatory aftertaste, i.e., 1, 2, 3, 4, 5, 6, 7, 8, 10, 11 and 12 channels, is input between 201-400ms, and in the third stage of the "posterior nasal effect", olfactory and gustatory aftertaste information is input between 401ms-550ms, i.e., 3, 4, 5, 6, 7, 9, 10 and 12 channels.

Fig. 6 is a diagram showing the output of beer smell-taste sensory information in the time dimension. As can be seen from the time dimension output graph, on the basis of the chaotic characteristic of the nonlinear dynamics nervous system, in the first stage of the 'front nose effect', although the olfactory information is input in 7 channels, obvious fluctuation exists in 1, 2, 8 and 9 channels, so that the input of the olfactory information can influence other channels. Meanwhile, the neural system based on nonlinear dynamics also embodies a dynamic output mode of multi-channel simultaneous perception.

And respectively calculating the respective perception results of the three stages. The perception accuracy of the perception model in the second stage of 'smell-taste coexistence' is 86.11%, and the perception accuracy in the third stage of 'rear nose effect' is 79.45% higher than that in the first stage of 'front nose effect' by 72.78%, so that the aftertaste information of taste sense has an important effect on whether the flavor information of beer can be correctly perceived.

Therefore, the action rule of the smell and taste information can be reflected by utilizing the neuro-dynamics system model, and the action result is visualized and dynamically presented. The method can obtain the simultaneous perception results of the front nose effect, the smell-taste coexistence and the rear nose effect, so as to realize the acquisition of bionic sensory experience information in the true sense, and the research results can promote the electronic nose and the electronic tongue instrument to replace the assessment personnel in sensory evaluation, so that the highly bionic application is realized.

The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.

Claims

1. A method for visualizing a machine sniffing-taste-perception effect by using a neuro-dynamics system model, which is characterized in that: the method comprises the following steps:

s2, extracting information characteristics affecting beer flavor based on variable projection importance evaluation index (VIP) and a multi-mode identification technology;

s3, inputting the characteristics of the flavor information of the influence beer in a time dimension based on a KIII nonlinear neural dynamics system model, and visualizing and dynamically presenting the action rule of the smell information;

the time dimension input in the step S3 comprises three stages of 'front nose effect', 'smell-smell coexistence', 'rear nose effect'; and inputting the simultaneous sense data in the model in three different time periods successively, and visualizing and dynamically presenting the action rule of the smell-taste information based on the neuro-dynamics system model.