CN116796179A - Method for detecting wine quality using electronic nose based on noise filtering framework and fusion model - Google Patents

Abstract

The invention relates to the technical field of odor detection and identification. It discloses a method for detecting wine quality using an electronic nose based on a noise filtering framework and a fusion model. A sensor array composed of 9 MOS gas sensors is used to collect resistances responsive to different types of wine odors. data and convert it into an output digital signal; filter the collected data using a noise filtering framework; use principal component analysis method PCA to extract data features (dimensionality reduction); establish a fusion classification model and combine the deep neural network DNN , Support vector machine SVM and decision tree DT are fused into a hybrid model, and the Adaboost algorithm is used to adjust the classifier weights and combine these classifiers to generate the final output; the fusion classification model is used to output the final classification result. Compared with the existing technology, the present invention processes the collected odor data through noise filtering and feature extraction, and uses a fusion model for odor recognition, which improves classification accuracy and can solve complex nonlinear problems.

Description

Electronic nose detection of wine quality based on noise filtering framework and fusion model method

Technical field

The invention relates to the technical field of odor detection and recognition, and specifically relates to a method for detecting wine quality using an electronic nose based on a noise filtering framework and a fusion model.

Background technique

Odor recognition is closely related to human life. Odor recognition is involved in food quality identification, industrial production, environmental monitoring, safety monitoring, disease diagnosis, etc. Further substance identification is carried out through odor recognition. At present, the methods used to identify complex odor samples mainly rely on gas chromatography analysis methods and GC-MS analysis technology. However, these methods often require complex pre-processing steps when analyzing complex odor samples, and the sample analysis cycle is long, and the operation and maintenance costs of the instrument are high. Therefore, these methods have the disadvantages of low analysis efficiency and high analysis cost.

The electronic nose system is a new type of odor analysis equipment that has been developed rapidly since 1982. The electronic nose system is mainly composed of a sensor array and a pattern recognition algorithm. Compared with odor sample analysis equipment such as gas chromatography, the electronic nose system has the advantages of simple sample preprocessing, sensitive response, fast analysis speed, and low analysis cost. Therefore, it is used in many fields. Application to odor recognition.

When existing electronic nose systems perform odor identification, they usually use odor identification methods such as principal component analysis, discriminant factor analysis, cluster independent soft pattern classification, statistical quality control analysis, etc. However, the current electronic nose system During identification, it is easily affected by many external factors and cannot be accurately identified, such as interference from other odors. Moreover, it is difficult for human sense of smell to distinguish different mixtures such as wine or coffee. With the continuous development of electronic nose, based on electronic nose technology, wine types of different properties can be distinguished based on different odors.

However, the sensor array of the electronic nose system is exposed to various gases for a long time, which may cause the sensor to be saturated. Changes in the ambient air are also likely to affect the stability of the gas sensor. Noise is inevitable in the electronic nose signal. This has the disadvantage of low accuracy in identifying different odors. Generally, it can only identify some simple odor samples or odor samples with large differences. In addition, for traditional odor recognition classifiers, traditional machine learning mostly uses a classification model for odor recognition and discrimination, which also has the disadvantage of low accuracy of odor recognition.

Contents of the invention

Purpose of the invention: In view of the problems existing in the prior art, the present invention provides a method for detecting wine quality with an electronic nose based on a noise filtering framework and a fusion model. The collected odor data is processed through noise filtering and feature extraction, and the fusion model is used. Smell recognition improves classification accuracy and can solve complex nonlinear problems.

Technical solution: The present invention provides a method for detecting wine quality with an electronic nose based on a noise filtering framework and a fusion model, which includes the following steps:

Step 1: Use a sensor array composed of 9 MOS gas sensors to collect resistance data responsive to different types of wine odors, and convert the resistance data into an output digital signal;

Step 2: Filter the data collected in step 1 using the noise filtering framework;

Step 3: Use principal component analysis (PCA) to reduce the dimensionality of the data;

Step 4: Establish a fusion classification model. The fusion classification model is based on the deep neural network DNN, support vector machine SVM and decision tree DT, and is fused into a hybrid model. The Adaboost algorithm is used to adjust the classifier weights and combine these classifiers to generate final output;

Step 5: Use the fusion classification model to output the final classification result for the dimensionally reduced data in Step 3.

Further, the nine MOS gas sensors in step 1 are smoke sensor, alcohol sensor, methane natural gas sensor, coal gas sensor, liquefied petroleum gas sensor, carbon monoxide sensor, hydrogen sensor, combustible gas sensor, and air quality sensor.

Further, in step 2, noise filtering is performed on the collected data, as follows:

(1) Set the appropriate wavelet decomposition level and determine the scaling parameter (v);

(2) IQR calculation between original x(t) signal and reconstructed y(t) signal;

(3) Establish the IQR matrix IQR _m×n ;

(4) Based on the IQR matrix, use the maximum independent variable of each signal to determine the most suitable scaled mother wavelet MWT;

(5) Use the most suitable scaled MWT for signal reconstruction.

Furthermore, the discrete wavelet transform DWT represents the data output by the sensor. The mathematical expression of the DWT of the original signal x(t) containing noise is as follows:

Among them, v and w are the scaling parameters and translation parameters respectively; the scaling parameters establish the time and frequency resolution of the scaled mother wavelet MWT; the scaled MWT is given by Indicates that the value of the scaling parameter is inversely proportional to the frequency; moving the parameter w makes the scaled MWT move along the time axis; when performing DWT, the decomposition level and wavelet function type are the two parameters that have the most influence on the change of the signal structure;

The scaling parameter corresponds to the decomposition level in signal reconstruction, which is determined by the following rules:

According to the frequency F _sample of the sampling signal and the level of decomposition, the frequency characteristics F _char are determined;

Find the most suitable mother wavelet for the required signal by calculating the information quality ratio IQR. The IQR value between the original signal x(t) and the reconstructed signal y(t) is given by the following formula:

By comparing the information quality ratio of several mother wavelets, the maximum value is the required signal; in sensor array signal processing, the IQR value of m signals reconstructed by n different MWTs is represented by the following matrix:

Select the best scaled mother wavelet of the required signal based on the maximum IQR value:

According to the selection of the most appropriate mother wavelet for signal reconstruction, the amplitude (y(t)) equation of the specific value of the reconstructed signal is as follows:

where y′ and c correspond to the new scaled value of y and the number of possible class labels respectively.

Furthermore, the goal of using the principal component analysis method for data feature extraction in step 3 is to reduce n groups of vectors to k dimensions, 0<k<n, and under orthogonal constraints, the maximum k variance is used as the new Based on the variables, the goals after feature extraction of the data through principal component analysis are as follows:

(1) Can multiple measurement values of the same odor be combined into one group?

(2) Whether two different odors appear as two separate groups;

(3) How large is the variance of the same odor measurement values within the same group.

Further, in step 4, the Adaboost algorithm is used to adjust the classifier weight as follows:

Ensemble learning resamples the original samples to obtain several data sets. Each data set trains a basic classifier separately. For each set of data, the classifier will obtain a corresponding number of prediction results, and the results are fused using an ensemble method. The decision is made to obtain the final result. In the classification stage, the stored odor data set is divided into two types of data, namely training data 70% and test data 30%. The training data is used to train the model, and the test data is used to test the performance of the model. ;

ε is defined as the weighted error, δ is defined as the function, α is defined as the weight of the weak classifier, y is defined as the output class, and the initial weight is set to 1/m, Let k be the number of boosting iterations, let i loop from 1 to k, train a base classifier C _i , train the deep neural network model DNN, base classifier C ₁ for all samples in the original training set, apply C ₁ in the original training set Go to all examples, calculate the weighted error ε, if ε>0.5, reset the weights of m samples, and then start training the base classifier C ₂ , loop; weak classifier weight/> Use the AdaBoost algorithm to adjust the weight of the sample and the weight of the current classifier. By changing the base classifier, repeat the above steps to the support vector machine SVM model and the decision tree DT model respectively, and finally predict the output results.

Further, the specific structure of the deep neural network model DNN is as follows:

The deep neural network model DNN consists of 9 input layers, 7 hidden layers and 4 output layers; the input of the 9 input layers is the collected data of 9 MOS gas sensors, and the activation function of the hidden layer is the ELU function, which The mathematical expression is as follows:

x<0: ELU(x)＝e ^x -1

x≥0：ELU(x)＝x

The activation function of the output layer is the normalized function Softmax. The Softmax function formula is as follows: All z _i are elements of the input vector,/> A probability distribution is formed, normalized to ensure that the sum of the output values is equal to 1, and each output result is between (0, 1), K is the number of categories of the multi-category classifier.

Further, the support vector machine SVM kernel function is set to a linear kernel, and the number of classifiers is 600. The linear kernel is a function that uses linear division of data; the decision tree DT, the number of classifiers is set to 600, The random state is set to 0, the maximum depth is set to 2, and the learning rate is set to 1.

Beneficial effects:

1. This invention aims at the possibility of saturation of the sensor caused by long-term exposure to various gases. Changes in the ambient air are also likely to affect the stability of the gas sensor. Noise is inevitably present in the electronic nose signal, and the use of a noise filtering framework is the most effective. To minimize the noise in the electronic nose detection process, establish a scaling mother wavelet, establish an appropriate signal decomposition level, and calculate the information quality ratio to select the most appropriate mother wavelet for signal reconstruction.

2. The present invention uses the principal component analysis method to extract features of the data, find out the distribution of the data, observe whether the samples can be accurately classified, and estimate the accuracy of statistical calculations when processing the data. Using dimensionality reduction processing technology, by converting multiple original indicators into multiple comprehensive indicators and using several indicators to replace the original variables, the goal of optimizing dimensionality reduction is to reduce n groups of vectors to k dimensions.

3. The present invention uses a machine learning algorithm to classify calculations performed on data after PCA dimensionality reduction, with the purpose of estimating the accuracy of statistical calculations when processing data. In addition, the present invention designs a hybrid classifier that combines traditional machine learning and deep learning models, including deep neural network (DNN), support vector machine (SVM) and decision tree (DT), and merges it into a hybrid model. The Adaboost algorithm is used to adjust the weights of the classifiers to combine these classifiers to generate the final output. The support vector machine kernel function can map samples from the original space to a higher-dimensional feature space, thereby making the samples linearly separable and improving the accuracy of classification. However, support vector machines (SVM) have disadvantages. Only using SVM can obtain good accuracy in the training set, but the accuracy in the test set is poor. Essentially, decision trees (DT) classify data through a series of rules, so good accuracy can be obtained in classification. Deep neural network (DNN) has multiple nonlinear mapping feature transformations and can fit highly complex functions. The model is relatively simple, and learning is faster and less time-consuming. Compared with shallow modeling methods, deep modeling can represent actual complex nonlinear problems in a more detailed and efficient manner. Therefore, this method uses deep neural network (DNN), support vector machine (SVM) and decision tree (DT) to construct a fusion model, which can not only obtain better classification accuracy, but also solve complex nonlinear problems.

4. The Adaboost algorithm of the present invention adopts a serial method when training base classifiers, and there are dependencies between each base classifier. The basic idea is to stack base classifiers layer by layer. During training, each layer gives a higher weight to the samples that were misclassified by the previous layer's base classifier. During testing, the results of each layer of classifiers are weighted to obtain the final result. Integrated learning is used to adjust the weights and the calculations are repeated to obtain better accuracy.

Description of the drawings

Figure 1 is a flow chart of the wine odor identification process of the present invention;

Figure 2 shows the specific steps of noise processing in the present invention;

Figure 3 is a structural framework diagram of the system of the present invention;

Figure 4 is a connection diagram of the sensor of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to more clearly illustrate the technical solutions of the present invention, but cannot be used to limit the scope of the present invention.

The invention discloses a method for detecting wine quality using an electronic nose based on a noise filtering framework and a fusion model. See Figure 1, which includes the following steps:

Step 1: Use a sensor array composed of 9 MOS gas sensors to collect resistance data responsive to different types of wine odors, and convert the resistance data into an output digital signal.

The electronic nose system includes a sensor array and a data processing module. The sensor array consists of a smoke sensor, an alcohol sensor, a methane gas sensor, a methane natural gas sensor, a liquefied petroleum gas sensor, a carbon monoxide sensor, a hydrogen sensor, a combustible gas sensor, and an air quality sensor. The sensors form a sensor array. The sensor responds specifically to different volatile organic compounds. Interactions with different gas molecules change the resistance of each channel in different ways. Each gas sensor measures a specific pattern, known as an odor fingerprint. The output is in the form of an analog sensor resistance value, and the connection diagram is shown in Figure 4. The sensor models used in the embodiments of the present invention are shown in Table 1 below:

Table 1 Sensor model

MQ-2 smoke sensor MQ-3 alcohol sensor MQ-4 Methane natural gas sensor MQ-5 gas sensor MQ-6 Liquefied gas gas sensor MQ-7 carbon monoxide sensor MQ-8 hydrogen sensor MQ-9 Combustible gas sensor MQ-135 air quality sensor

The data processing module is designed using Raspberry Pi. In the smell recognition system, Raspberry Pi's 8-channel A-D converter: Raspberry Pi does not have analog ports provided by each sensor. On the contrary, to read data from each sensor, we need to analog these The data is converted to 17-bit digital data for the Raspberry Pi.

The ADC Pi is an 8-channel 17-bit analog-to-digital converter designed to work with the Raspberry Pi and other compatible single-board computers. The ADC Pi is based on two Microchip MCP3424 A/D converters, each containing 4 analog inputs. The MCP3424 is a delta-sigma a/D converter with low noise differential inputs.

The original resistance data passes through an analog-to-digital converter, that is, an A/D converter, which converts the analog signal into an output digital signal. The data is then communicated with a single-board computer (Raspberry Pi) for subsequent data processing.

Step 2: Filter the data collected in step 1 using the noise filtering framework.

(1) Set the appropriate wavelet decomposition level and determine the scaling parameter (v).

(2) IQR calculation between original x(t) signal and reconstructed y(t) signal.

(3) Establish the IQR matrix IQR _m×n .

(4) Based on the IQR matrix, use the maximum independent variable of each signal to determine the most suitable scaled mother wavelet MWT.

(5) Use the most suitable scaled MWT for signal reconstruction.

details as follows:

Noise filtering is the key to improving data quality. Discrete wavelet transform DWT represents the data output by the sensor. The mathematical expression of DWT of the original signal x(t) containing noise is as follows:

Among them, v and w are the scaling parameters and translation parameters respectively; the scaling parameters establish the time and frequency resolution of the scaled mother wavelet MWT; the scaled MWT is given by means that the value of the scaling parameter is inversely proportional to the frequency; higher values of v mean lower frequencies, and vice versa. The moving parameter w moves the scaled MWT along the time axis; when performing DWT, the decomposition level and wavelet function type are the two parameters that have the most influence on changes in signal structure.

The scaling parameter corresponds to the level of decomposition in signal reconstruction. Lower frequencies result in more decomposition, so setting the appropriate decomposition level is especially important. The level of decomposition used in this method can be determined by the following rules:

The frequency characteristics F _char are determined based on the frequency F _sample of the sampling signal and the level of decomposition, thereby effectively avoiding signal defects caused by mismatch between the signal frequency characteristics and the reference range.

The main principle of noise filtering is to reconstruct the signal without losing the basic information. This method finds the most suitable mother wavelet for the required signal by calculating the information quality ratio IQR. The IQR value between the original signal x(t) and the reconstructed signal y(t) is given by the following formula:

By comparing the information quality ratio of several mother wavelets, the maximum value is the required signal; in sensor array signal processing, the IQR value of m signals reconstructed by n different MWTs is represented by the following matrix:

Select the best scaled mother wavelet of the required signal based on the maximum IQR value:

According to the selection of the most appropriate mother wavelet for signal reconstruction, the amplitude (y(t)) equation of the specific value of the reconstructed signal is as follows:

where y′ and c correspond to the new scaled value of y and the number of possible class labels respectively. At this point, the complete noise filtering process has been completed, which can theoretically effectively improve the quality of data.

Step 3: Use principal component analysis (PCA) to reduce the dimensionality of the data.

Use principal component analysis to extract features from the data (also called dimensionality reduction), find out the distribution of the data, and observe whether the samples can be accurately classified. Its purpose is to estimate the accuracy of statistical calculations when processing data.

Principal component analysis (PCA) is a multivariate statistical method that compresses data into fewer dimensions, reducing the complexity of high-dimensional data while retaining patterns and trends. Using dimensionality reduction processing technology, several indicators are used to replace the original variables by converting multiple original indicators into multiple comprehensive indicators. The goal of optimized dimensionality reduction is to reduce n sets of vectors to k dimensions (0<k<n). Under orthogonal constraints, the maximum k-variance is used as the basis for new variables.

The following information was inferred from principal component analysis:

(1) Can multiple measurement values of the same odor be combined into one group.

(2) Whether two different odors can appear as two separate groups.

(3) How large is the variance of the same odor measurement values within the same group.

Step 4: Suggest a fusion classification model. The fusion classification model is based on the deep neural network DNN, support vector machine SVM and decision tree DT, and is fused into a hybrid model. The Adaboost algorithm is used to adjust the classifier weights and combine these classifiers to Generate the final output.

In order to obtain better performance, this method designs a hybrid classifier that combines traditional machine learning and deep learning models, including deep neural network (DNN), support vector machine (SVM) and decision tree (DT), into one Hybrid model, which uses the Adaboost algorithm to adjust the weights of classifiers to combine these classifiers to generate the final output. The details of using the Adaboost algorithm to adjust the classifier weight are as follows:

Ensemble learning resamples the original samples to obtain several data sets. Each data set trains a basic classifier separately. For each set of data, the classifier will obtain a corresponding number of prediction results, and the results are fused using an ensemble method. The decision is made to obtain the final result. In the classification stage, the stored odor data set is divided into two types of data, namely training data 70% and test data 30%. The training data is used to train the model, and the test data is used to test the performance of the model. ;

ε is defined as the weighted error, δ is defined as the function, α is defined as the weight of the weak classifier, y is defined as the output class, and the initial weight is set to 1/m, Let k be the number of boosting iterations, let i loop from 1 to k, train a base classifier C _i , train the deep neural network model DNN, base classifier C ₁ for all samples in the original training set, apply C ₁ in the original training set Go to all examples, calculate the weighted error ε, if ε>0.5, reset the weights of m samples, and then start training the base classifier C2, loop; weak classifier weight/> Use the AdaBoost algorithm to adjust the weight of the sample and the weight of the current classifier. By changing the base classifier, repeat the above steps to the support vector machine SVM model and the decision tree DT model respectively, and finally predict the output results.

The specific structure of the deep neural network model DNN is as follows:

The deep neural network model DNN consists of 9 input layers, 7 hidden layers and 4 output layers; the input of the 9 input layers is the collected data of 9 MOS gas sensors, and the activation function of the hidden layer is the ELU function, which The mathematical expression is as follows:

x<0: ELU(x)＝e ^x -1

x≥0：ELU(x)＝x

The activation function of the output layer is the normalized function Softmax. The Softmax function formula is as follows: All z _i are elements of the input vector,/> A probability distribution is formed, normalized to ensure that the sum of the output values is equal to 1, and each output result is between (0, 1), K is the number of categories of the multi-category classifier.

The support vector machine SVM kernel function is set to a linear kernel, the number of classifiers is 600, the linear kernel is a function that uses linear division of data; the decision tree DT, the number of classifiers is set to 600, and the random state is set to 0, the maximum depth is set to 2, and the learning rate is set to 1.

Step 5: Use the fusion classification model to output the final classification result for the dimensionally reduced data in Step 3.

The present invention prepares wine samples for testing, puts the wine samples into the odor collection box, and the closed environment is more conducive to the collection of wine odors. Wine samples are divided into freshly opened, stored for 2 days, and stored for 4 days, with three freshness levels: fresh, semi-fresh, and stale. A sensor array composed of 9 gas sensors was used to collect data on the tested wine samples, and the resistance data of the different responses formed by the sensors when they were just opened, left for 2 days and 4 days were collected, that is, the odor fingerprint.

The original resistance data passes through an analog-to-digital converter, that is, an A/D converter, which converts the analog signal into an output digital signal. The data is then communicated with the single-board computer (Raspberry Pi) for subsequent data processing.

Use the noise filtering framework to perform noise filtering on the collected original odor data. By establishing a scaling mother wavelet MWT on the signal data, determine the appropriate signal decomposition level, and calculate the information quality ratio IQR to select the most appropriate mother wavelet for signal reconstruction. , to achieve the purpose of noise reduction on the original odor data.

The filtered data uses principal component analysis method PCA for feature extraction and becomes feature data that is easy to be trained by the machine learning model. A fusion model consisting of three models, DNN, SVM, and DT, is selected to classify the data. The fusion model uses the Adaboost algorithm to adjust the weights of the classifiers to combine these classifiers.

The F ₁ score is obtained based on the precision and recall of the experiment. The F ₁ score is the average of the accuracy and recall that measures the model's ability. Theoretically, the accuracy of the proposed model will be higher, the standard deviation will be lower, and the F ₁ score will have more values of 1.00, which means that the fusion model has better predictability, indicating that the model is more effective in detecting wine quality. .

The above embodiments are only for illustrating the technical concepts and features of the present invention. Their purpose is to enable those familiar with this technology to understand the content of the present invention and implement it accordingly, and cannot limit the scope of protection of the present invention. All equivalent transformations or modifications made based on the spirit and essence of the present invention shall be included in the protection scope of the present invention.

Claims (8)

1. A method for detecting wine quality using electronic nose based on noise filtering framework and fusion model, which is characterized by including the following steps: Step 1: Use a sensor array composed of 9 MOS gas sensors to collect resistance data responsive to different types of wine odors, and convert the resistance data into an output digital signal; Step 2: Filter the data collected in step 1 using the noise filtering framework; Step 3: Use principal component analysis (PCA) to reduce the dimensionality of the data; Step 4: Establish a fusion classification model. The fusion classification model is based on the deep neural network DNN, support vector machine SVM and decision tree DT, and is fused into a hybrid model. The Adaboost algorithm is used to adjust the classifier weights and combine these classifiers to generate final output; Step 5: Use the fusion classification model to output the final classification result for the dimensionally reduced data in Step 3. 2. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that the gas sensors of the 9 MOS in step 1 are smoke sensors, alcohol sensors, and methane sensors respectively. Natural gas sensor, coal gas sensor, liquefied petroleum gas sensor, carbon monoxide sensor, hydrogen sensor, combustible gas sensor, air quality sensor. 3. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that, in the step 2, noise filtering is performed on the collected data, specifically as follows: (1) Set the appropriate wavelet decomposition level and determine the scaling parameter (v); (2) IQR calculation between original x(t) signal and reconstructed y(t) signal; (3) Establish the IQR matrix IQR _m×n ; (4) Based on the IQR matrix, use the maximum independent variable of each signal to determine the most suitable scaled mother wavelet MWT; (5) Use the most suitable scaled MWT for signal reconstruction. 4. The method of electronic nose detecting wine quality based on noise filtering framework and fusion model according to claim 3, characterized in that discrete wavelet transform DWT represents the data output by the sensor, and the DWT of the original signal x(t) containing noise The mathematical expression of is as follows: Among them, v and w are the scaling parameters and translation parameters respectively; the scaling parameters establish the time and frequency resolution of the scaled mother wavelet MWT; the scaled MWT is given by Indicates that the value of the scaling parameter is inversely proportional to the frequency; moving the parameter w makes the scaled MWT move along the time axis; when performing DWT, the decomposition level and wavelet function type are the two parameters that have the most influence on the change of the signal structure; The scaling parameter corresponds to the decomposition level in signal reconstruction, which is determined by the following rules: According to the frequency F _sample of the sampling signal and the level of decomposition, the frequency characteristics F _char are determined; Find the most suitable mother wavelet for the required signal by calculating the information quality ratio IQR. The IQR value between the original signal x(t) and the reconstructed signal y(t) is given by the following formula: By comparing the information quality ratio of several mother wavelets, the maximum value is the required signal; in sensor array signal processing, the IQR value of m signals reconstructed by n different MWTs is represented by the following matrix: Select the best scaled mother wavelet for the required signal based on the maximum IQR value: According to the selection of the most appropriate mother wavelet for signal reconstruction, the amplitude (y(t)) equation of the specific value of the reconstructed signal is as follows: where y′ and c correspond to the new scaled value of y and the number of possible class labels respectively. 5. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that, in step 3, the goal of using principal component analysis method to extract features of data is to combine n groups The vector is reduced to k dimensions, 0＜k＜n. Under orthogonal constraints, the maximum k variance is used as the basis for new variables. The objectives after feature extraction of data through principal component analysis are as follows: (1) Can multiple measurement values of the same odor be combined into one group? (2) Whether two different odors appear as two separate groups; (3) How large is the variance of the same odor measurement values within the same group. 6. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that in step 4, the Adaboost algorithm is used to adjust the classifier weight as follows: Ensemble learning resamples the original samples to obtain several data sets. Each data set trains a basic classifier separately. For each set of data, the classifier will obtain a corresponding number of prediction results, and the results are fused using an ensemble method. The decision is made to obtain the final result. In the classification stage, the stored odor data set is divided into two types of data, namely training data 70% and test data 30%. The training data is used to train the model, and the test data is used to test the performance of the model. ; ε is defined as the weighted error, δ is defined as the function, α is defined as the weight of the weak classifier, y is defined as the output class, and the initial weight is set to 1/m, Let k be the number of boosting iterations, let i loop from 1 to k, train a base classifier C _i , train the deep neural network model DNN, base classifier C ₁ for all samples in the original training set, apply C ₁ in the original training set Go to all examples, calculate the weighted error ε, if ε>0.5, reset the weights of m samples, and then start training the base classifier C ₂ , loop; weak classifier weight/> Use the AdaBoost algorithm to adjust the weight of the sample and the weight of the current classifier. By changing the base classifier, repeat the above steps to the support vector machine SVM model and the decision tree DT model respectively, and finally predict the output results. 7. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that the specific structure of the deep neural network model DNN is as follows: The deep neural network model DNN consists of 9 input layers, 7 hidden layers and 4 output layers; the input of the 9 input layers is the collected data of 9 MOS gas sensors, and the activation function of the hidden layer is the ELU function, which The mathematical expression is as follows: x<0: ELU(x)＝e ^x -1 x≥0：ELU(x)＝x The activation function of the output layer is the normalized function Softmax. The Softmax function formula is as follows: All z _i are elements of the input vector,/> A probability distribution is formed, normalized to ensure that the sum of the output values is equal to 1, and each output result is between (0, 1), K is the number of categories of the multi-category classifier. 8. The method for detecting wine quality by electronic nose based on noise filtering framework and fusion model according to claim 1, characterized in that the support vector machine SVM kernel function is set to a linear kernel, and the number of classifiers is 600, The linear kernel is a function that divides the data linearly; for the decision tree DT, the number of classifiers is set to 600, the random state is set to 0, the maximum depth is set to 2, and the learning rate is set to 1.