An electronic nose width learning method for regression prediction of mixed gas concentration

Abstract

The present invention relates to the technical field of electronic noses, and specifically discloses an electronic nose width learning method for regression prediction of mixed gas concentration, comprising the steps of: S1. Selecting different sparse regularization parameters, feature node contraction parameters and initial random mapping matrix assignment methods to construct multiple enhanced width learning models based on the width learning model; S2. Obtaining a sample set and dividing it into a training data set and a test data set according to a preset ratio; S3. Using the training data set to train multiple enhanced width learning models, and recording the corresponding training time; S4. Using the test data set to test the trained multiple enhanced width learning models, and recording the regression prediction error rate; S5. Comprehensively considering the training time and regression prediction error rate, determine the enhanced width learning model with the best performance as the electronic nose regression prediction model. Compared with other regression prediction algorithms, the regression prediction model based on width learning not only shortens the training time, but also reduces the regression prediction error rate.

Description

Technical Field

[0001] The invention relates to the technical field of electronic noses, and in particular to an electronic nose width learning method for performing regression prediction on mixed gas concentration. Background Art

[0002] The electronic nose (E-nose) consists of a gas sensor array, feature extraction and pattern recognition algorithms, and has similar functions to the biological olfactory system. Currently, the electronic nose has been widely used in many fields, such as environmental monitoring, disease detection, agriculture, food storage, etc.

[0003] Since humans spend most of their lives indoors, indoor pollutant gases have a great impact on human health and quality of life. Therefore, it is necessary to design a real-time monitoring method for indoor gas concentrations. Electronic nose is an effective method for analyzing gases and can effectively perform regression prediction analysis of indoor gas concentrations.

[0004] Before applying the electronic nose to gas concentration regression prediction and analysis, a large amount of data samples need to be collected to train the model. However, there is a problem in the actual training of the model: the regression prediction algorithm usually based on the neural network requires a long training time due to the complexity of the network structure, and the final prediction results often cannot meet the actual needs. Summary of the invention

[0005] The present invention provides an electronic nose width learning method for performing regression prediction on the concentration of mixed gases, and solves the technical problem that the regression prediction algorithm model for the electronic nose currently established based on a neural network requires a long training time due to the complexity of the network structure, and the final prediction result often cannot meet the actual needs.

[0006] In order to solve the above technical problems, the present invention provides an electronic nose width learning method for performing regression prediction on the concentration of mixed gas, comprising the steps of:

[0007] S1. Select different sparse regularization parameters, feature node shrinkage parameters and initial random mapping matrix assignment methods to construct multiple enhanced width learning models based on the width learning model;

[0008] S2. Obtain a sample set and divide it into a training data set and a test data set according to a preset ratio;

[0009] S3. Using the training data set to train the plurality of enhanced width learning models, and recording the corresponding training time;

[0010] S4. Testing the trained multiple enhanced width learning models using a test data set, and recording the regression prediction error rate;

[0011] S5. Taking into account the training time and the regression prediction error rate, determine the enhanced width learning model with the best performance as the electronic nose regression prediction model.

[0012] Furthermore, in step S1, the selection range of the sparse regularization parameter is between 2^-30 and 0.0002.

[0013] Preferably, the sparse regularization parameter of the electronic nose regression prediction model is 0.00002.

[0014] Furthermore, in the step S1, the selection range of the characteristic node shrinkage parameter is between 1e-4 and 1e-2.

[0015] Preferably, the characteristic node shrinkage parameter of the electronic nose regression prediction model is 1e-3.

[0016] Furthermore, in the step S1, the options for the initial random mapping matrix assignment method include Gaussian distribution, exponential distribution, Poisson distribution, Rayleigh distribution, lognormal distribution, continuous uniform distribution, and β distribution.

[0017] Preferably, the initial random mapping matrix assignment method of the electronic nose regression prediction model is Gaussian distribution.

[0018] Furthermore, in step S2, the sample set includes at least ten gas concentration value indicators, each of which corresponds to at least 100 samples; the preset ratio is training data set: test data set, and the preset ratio is between 4:1 and 2:1; the gas concentration value indicator is the mixed gas concentration after at least two measurement gases are mixed according to the 1st to Nth preset concentration combinations, N≥10.

[0019] Preferably, the preset ratio is 3:1.

[0020] Preferably, the gas concentration value indicator is the concentration of a mixed gas after methane and ethylene are mixed according to the 1st to 38th preset concentration combinations.

[0021] The present invention provides an electronic nose width learning method for regression prediction of mixed gas concentration. Based on a width learning system or model (BLS), the initial random mapping matrix assignment method, sparse regularization parameters, and feature node contraction parameters are changed in the BLS to construct multiple enhanced width learning models (step S1), and each enhanced width learning model is trained using a training data set (step S3) and tested using a test data set (step S4). Finally, the training time and the regression prediction error rate during the test are combined to select the enhanced width learning model with the best performance as the electronic nose regression prediction model (or referred to as enhanced BLS) (step S5). The enhanced BLS not only shortens the training time, but also reduces the regression prediction error rate. Experimental results show that the enhanced BLS is superior to BLS, other enhanced width learning models based on BLS, back propagation neural network (BPNN), least squares support vector machine (LSSVM), extreme learning machine (ELM), linear regression (LR) and other methods in terms of comprehensive training time (the shorter the better) and regression prediction error rate (the lower the better). BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a structural diagram of an experimental device for data collection provided by an embodiment of the present invention;

[0023] Figure 2 is a structural diagram of an existing BLS provided by an embodiment of the present invention;

[0024] Figure 3 is a random number distribution diagram generated by different random number generation methods provided by an embodiment of the present invention;

[0025] Figure 4 is the regression prediction error of different assignment methods in the training data set provided by the embodiment of the present invention;

[0026] Figure 5 is the regression prediction error of different assignment methods in the test data set provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0028] Random Vector Function Link Neural Network (RVFLNN) has a very simple planar network structure and can take mapping features as its input, so it eliminates the disadvantage of long training time and provides powerful function approximation generalization ability. At present, RVFLNN has the ability of approximation generalization and has been widely used in classification and regression.

[0029] The Breadth Learning System (BLS) is a regression model based on the idea of ​​RVFLNN, proposed by CLIPhilip Chen in 2018. Unlike RVFLNN, BLS combines the mapping features and enhanced features generated by the original input data to form a new feature matrix, which is then used as the actual input of the model. The benefits of doing so are twofold. On the one hand, more representative features in the data can be extracted to make the regression prediction results better. On the other hand, high-dimensional (even infinite-dimensional) data can also be processed, so the model has better data processing capabilities. So far, it has been applied to drift compensation, image classification, time series prediction, etc.

[0030] BLS is a simple planar network structure, which is different from complex deep neural networks. The structure of BLS is as follows Figure 1 As shown. BLS first converts the original input data mapping into mapping features through random weights and biases and stores them in feature nodes. Then, the feature nodes are expanded into enhanced nodes through similar random weights and biases. Finally, all feature nodes and enhanced nodes are horizontally connected and input into the network, and the connection weights between input and output can be obtained by the ridge regression algorithm. The detailed derivation process is as follows.

[0031] Assume the input data set is X, the i-th mapping feature Z i (Feature nodes) are generated by a linear mapping function as follows:

[0032] Z i =φ i (XW ei +β ei ) (1)

[0033] Among them, W ei and β ei are randomly generated dimensions with appropriate weights and biases, φ i is a linear mapping function. i =[Z1,Z2,...,Z i ] represents the feature mapping matrix. The j-th group of enhanced nodes is generated as:

[0034] H j =ξ j (Z i W hj +β hj) (2)

[0035] Among them, W hj and β hj are randomly generated dimensions with appropriate weights and biases, ξ j is a nonlinear mapping function. Similar to the previous step, H j =[H1,H2,...,H j ]. The values ​​of i and j are chosen by the complexity of the modeling task.

[0036] Supervised learning tasks, such as regression, usually require good feature inputs to achieve better results. In the process of generating feature nodes, weights and biases are randomly generated. However, randomness is unpredictable. In order to overcome the nature of randomness, a sparse autoencoder is used to fine-tune the randomly generated W. ei and β ei The sparse feature learning model can explore the most basic features of the data. Assuming that there are n sets of feature maps and m sets of enhancement maps in the BLS network, the obtained BLS features are A = [Z m |H n ] Therefore, the output Y of the BLS network can be expressed as:

[0037] Y=AW (3)

[0038] Where W is the connection weight of the network. The calculation formula of W can be expressed as W = A + Y, where A + is the pseudo-inverse matrix of the BLS feature matrix A. Then, the solution of the matrix pseudo-inverse problem is transformed into l2 norm regularized BLS, which can be expressed as the following convex optimization problem:

[0039]

[0040] The second term (min) in the above formula further limits the weight W, and λ is the regularization coefficient.

[0041] Therefore, W can be calculated as:

[0042] W=(λI+AA T ) -1 A T Y (5)

[0043] In particular,

[0044]

[0045] Where I is the identity matrix.

[0046] The basic framework of the BLS method / model can be listed as follows:

[0047] Step 1: Input the training data set of the electronic nose; set the sparse regularization parameter C and the feature node shrinkage parameter S;

[0048] Step 2: Randomly generate weights W ei and bias β ei , and calculate the i-th feature map Z by formula (1) i , set feature map group Z n =[Z1,Z2,...,Z n ];

[0049] Step 3: Randomly generate weights W hj and bias β hj , and calculate the jth enhancement node H by formula (2) j , set the feature enhancement node group H m =[H1,H2,...,H m ];

[0050] Step 4: Let A = [Z m |H n ] Calculate A by formula (6) + ; Calculate the connection weight W through formula (5) and output the connection weight W; Calculate the output Y of the BLS model through formula (3).

[0051] Based on the above BLS model, this embodiment proposes a regression analysis algorithm model with fast training time and high prediction accuracy (called enhanced BLS) to solve the regression analysis problem of the electronic nose. Compared with BLS, enhanced BLS has the following two improvements:

[0052] (1) Change the sparse regularization parameter C and feature node shrinkage parameter S of BLS;

[0053] (2) Change the assignment method of the initial random mapping matrix.

[0054] Based on this, this embodiment proposes an electronic nose width learning method for performing regression prediction on the concentration of mixed gas, comprising the steps of:

[0055] S1. Select different sparse regularization parameters, feature node shrinkage parameters and initial feature mapping matrix assignment methods to construct multiple enhanced width learning models based on the width learning model;

[0056] S2. Obtain a sample set and divide it into a training data set and a test data set according to a preset ratio;

[0057] S3. Using the training data set to train the plurality of enhanced width learning models, and recording the corresponding training time;

[0058] S4. Testing the trained multiple enhanced width learning models using a test data set, and recording the regression prediction error rate;

[0059] S5. Taking into account the training time and the regression prediction error rate, determine the enhanced width learning model with the best performance as the electronic nose regression prediction model.

[0060] (I) Training data set and test data set

[0061] like Figure 1 The data acquisition system shown, the sensor array in the electronic nose device of this embodiment is composed of 16 gas sensors of TGS2600, TGS2602, TGS2610, TGS2620 (four for each sensor), and the sensor array is placed in a closed measurement chamber with a capacity of 60ml. The gas sample is injected into the measurement chamber at a constant flow rate of 300ml / min (controlled by MFC). The chamber is connected in series to a vapor delivery system, which provides versatility for setting concentrations in a highly accurate and highly reproducible manner (in order to make the obtained data more accurate). In addition, the entire experimental process is run in a computer environment to ensure the accuracy of the data.

[0062] The data collection process is:

[0063] A mixed gas of methane (first gas) and ethylene (second gas) generated in the air is injected into the measurement chamber so that the sensor array is in full contact with the gas sample, thereby measuring the conductivity of the sensor array. At the same time, the sensor signal is recorded at the end of the measurement and 16 time series are obtained, which show the current state of the gas. The measurement is performed by continuously collecting signals from the 16 sensor arrays for about 12 hours without interruption.

[0064] A total of about one million data were collected in this embodiment. The concentration values ​​of methane and ethylene were added in the data set to make it a concentration index of the methane and ethylene mixed gas. There are 38 concentration value indicators in the sample (the 1st to 38th preset concentrations). 1000 samples were randomly selected from each concentration value sample to form a data set used in the experiment. At the same time, 75% of each concentration value indicator sample was used as a training sample set or training data set, and the rest was used as a test sample set or test data set. However, the number of samples with a mixed concentration of 170ppm is relatively small, so 101 samples were randomly selected for the experiment, of which 75 samples were training set samples and 26 samples were test set samples. The detailed information of the experimental samples is shown in Table 1.

[0065] Table 1. Composition of the experimental dataset

[0066]

[0067] Note: A concentration of 0 means that pure air (strictly configured in a nitrogen: oxygen ratio of 4:1) is injected into the measurement chamber.

[0068] In other embodiments, the sample set, gas concentration value index, the number of samples corresponding to each gas concentration value index, the sample ratio of the training data set and the test data set, and the specific concentration gas of the gas concentration value index can be set according to different requirements or conditions.

[0069] In other embodiments:

[0070] The sample set generally includes at least ten gas concentration value indicators, that is, N≥10, but generally N≤50. Too few indicators will affect the training effect, and too many indicators will increase the workload of sample collection.

[0071] Each gas concentration value indicator corresponds to at least 100 samples (except for some special concentration value indicators). This embodiment selects 1000 samples, but generally no more than 2000 to avoid too much workload; the sample ratio of the training data set and the test data set is between 4:1 and 2:1. In principle, the samples of the training data set should be larger than the test data set;

[0072] The gas concentration value indicator is the concentration of a mixed gas after at least two measurement gases (such as methane, ethylene, CO, etc.) are mixed according to the first to Nth preset concentration combinations.

[0073] (II) BLS sparse regularization parameter C and feature node shrinkage parameter S

[0074] In the BLS model, both the feature extraction of the original data and the sparsity of the feature mapping matrix require parameter settings. At the same time, the parameter settings indirectly affect the performance of the BLS model, and thus also affect the regression prediction performance of the electronic nose system. Therefore, it is necessary to select the optimal parameter values ​​to improve the performance of the electronic nose system.

[0075] As mentioned above, the number of feature nodes and enhanced nodes is selected according to the complexity of the modeling task. It is difficult to evaluate the complexity of the modeling task in an artificial way during model training. Therefore, the present embodiment pre-sets the search range of the number of nodes, and adopts the traversal method to determine the number of feature nodes and enhanced nodes. At the same time, the sparse regularization parameter C in the model and the shrinkage parameter S of the feature node are optimized, and the training time of the training data set and the regression prediction error rate of the test data set are calculated to evaluate the performance of the algorithm. Each algorithm is repeated 10 times, and the average value of the 10 results is used as the final regression prediction error rate. The results of the optimized parameters are shown in Table 2.

[0076] Table 2. Results after parameter changes

[0077] Parameter settings ERR_train ERR_test Optimal number of nodes C = 2^-30; S = 1e-3 0.0027 0.0060 NumFea=3; NumWin=11; NumEnhan=181; C=0.00002;S=1e-3 0.0030 0.0077 NumFea=3; NumWin=11; NumEnhan=189; C=0.00002;S=1e-3 0.0023 0.0047 NumFea=5; NumWin=12; NumEnhan=292;

[0078] Note: NumFea represents the number of feature nodes in each feature window, NumWin represents the number of feature windows, and NumEnhan represents the number of enhancement nodes. ERR_train is the regression prediction error of the training set of the BLS model, and ERR_test is the regression prediction error of the test set of the BLS model.

[0079] It can be seen from Table 2 that in the process of optimizing parameters S and C, the system works more ideally and reduces the regression prediction error of the system. When the optimal result is obtained, the number of nodes in the model increases greatly, because with the increase in the number of nodes, more features are extracted from the original data to a certain extent, thus obtaining better results.

[0080] It should be noted that in the process of finding the optimal sparse regularization parameter and enhanced node shrinkage parameter, the sparse regularization parameter is selected from multiple values ​​between 2^-30 and 0.0002, and the feature node shrinkage parameter is also selected from multiple values ​​between 1e-4 and 1e-2. Multiple groups of sparse regularization parameters and feature node shrinkage parameter values ​​are formed through permutations and combinations, and then they are trained and tested separately. Finally, the sparse regularization parameter C = 0.00002 and the enhanced node shrinkage parameter S = 1e-3 are selected as the optimal choices based on the comprehensive regression prediction error rate, as shown in Table 2.

[0081] (III) Enhanced BLS initial random mapping matrix assignment method

[0082] As mentioned above, randomness is unpredictable, and random numbers are randomly generated according to certain rules. Before the random number is generated, its specific value cannot be obtained, which makes the random mapping matrix generated by the random number in the network unpredictable, resulting in the quality of the feature extraction result. Therefore, when the BLS model selects the random mapping method to perform feature mapping on the input data, this may cause the loss of important feature information, thereby increasing the final regression prediction error rate. Therefore, this embodiment selects other random number generator methods to assign the generated random numbers to the initial random mapping matrix to reduce the regression prediction error rate. In this embodiment, seven random number generation methods such as Gaussian distribution, exponential distribution, Poisson distribution, Rayleigh distribution, lognormal distribution, continuous uniform distribution, and beta distribution are compared to select the best random number allocation method. The expressions of the seven distribution methods are shown in Table 3 below.

[0083] Table 3. Expressions of distribution methods

[0084]

[0085]

[0086] like Figure 3As shown, each method of generating random numbers randomly generates 100 points. It can be found that the random numbers generated by all random number generation methods are in the range of [-2,6], but the distribution of random numbers is different. The random numbers of the BLS model are uniformly distributed between [-1,1]. Since the value of the generated random number cannot be controlled, this embodiment chooses to expand the range of random numbers to make the distribution of random numbers larger or smaller, and then selects a random number allocation method that is more suitable for BLS.

[0087] from Figure 3 It can be seen that the random numbers that conform to the Gaussian distribution are distributed in [-2, 4], the random numbers that conform to the exponential distribution are distributed in [0, 4], the random numbers that conform to the Rayleigh distribution are distributed in [0, 2], the random numbers that conform to the lognormal distribution are distributed in [0, 6], the random numbers that conform to the continuous uniform distribution are distributed in [0, 1], and the random numbers that conform to the beta distribution are distributed in [0, 1]. Since the random numbers generated by these methods are continuous, in order to compare with these continuous random numbers, this embodiment selects Poisson distribution to generate discrete random numbers in [0, 4].

[0088] Since the range of the generated random numbers is different, better data features can be extracted during feature mapping to achieve better regression prediction accuracy.

[0089] In this embodiment, when the sparse regularization parameter C and the feature node shrinkage parameter S are at the optimal level, the initial random matrix is ​​changed by using the unpredictability of randomness to explore an optimal assignment method. The detailed results are shown in Table 4.

[0090] Table 4. Regression prediction error rates for different random weight assignment methods

[0091] Weight matrix assignment method ERR_train ERR_test Gaussian distribution 0.0023 0.0036 Exponential distribution 0.0057 0.0082 Poisson distribution 0.0042 0.0084 Rayleigh distribution 0.0067 0.0100 Lognormal distribution 0.0063 0.0070 Continuous uniform distribution 0.0153 0.0130 Beta distribution 0.0129 0.0155

[0092] It is obvious that different assignment methods will change the results of regression analysis. Among all the methods, only the initial random mapping matrix obtained by Gaussian distribution can obtain better results than the original method.

[0093] Figure 4 and Figure 5 The regression prediction errors of different random weight assignment methods in the training data set and the test data set are shown respectively. It can be seen that compared with other distributions, Gaussian distribution has obvious advantages in the initial matrix assignment of the BLS model. At the same time, compared with the original random distribution method of BLS, the regression prediction error results obtained on the test data set using Gaussian distribution to generate the initial random mapping matrix have been significantly improved. Therefore, this embodiment selects Gaussian distribution as the initial random mapping matrix assignment method.

[0094] (IV) Comparison of enhanced BLS and other regression algorithms

[0095] In order to prove the effectiveness of enhanced BLS, this example will compare the regression prediction ability of this method with existing methods (including BLS, BPNN, LSSVM, ELM, LR). In order to ensure the correctness of the experimental results, each program is repeated 10 times, and then the average value is taken as the final result. The results are shown in Table 5.

[0096] Table 5. Regression prediction error rate of different algorithms (%)

[0097] BLS BPNN LSSVM ELM LR Enhanced BLS Training Dataset 0.0023 0.0018 0.0035 1.7807 0.0375 0.0023 Test Dataset 0.0047 0.0124 0.0175 2.2927 0.0425 0.0036

[0098] It can be found from Table 5 that although the enhanced BLS is higher than BPNN in the regression prediction error of the training data set, the regression prediction error rate from the test data set is better than BLS, BPNN, LSSVM, ELM, and LR.

[0099] Since the training set and test set of each enhanced width learning model in step S1 are the same, the corresponding training time is not much different. Step S5 basically only considers the regression prediction error rate. Of course, when the regression prediction error rate is the same, the shorter the training time, the better.

[0100] The training time is recorded for comparison with other regression prediction algorithms, and the comparison results are shown in Table 6. Obviously, LSSVM has the longest training time to complete the regression analysis, while LR has the shortest training time to complete the regression analysis. Among the six methods, enhanced BLS has a very good performance, and its training time is only slightly longer than LR, but enhanced BLS has good regression prediction accuracy.

[0101] Table 6. Training time of different algorithms

[0102]

[0103]

[0104] Combining Table 5 and Table 6, we can see that the training time of LR is shorter than that of enhanced BLS, but the regression prediction results are worse than those of enhanced BLS. However, the four methods of BLS, BPNN, LSSVM and ELM not only have longer training time than enhanced BLS, but also have worse prediction errors than enhanced BLS.

[0105] (V) Conclusion

[0106] The application of electronic nose in the regression prediction analysis of mixed gases is very important. In view of the shortcomings of the existing regression prediction methods proposed in the background technology, namely, long training time and unsatisfactory prediction accuracy, this embodiment studies BLS and finds that BLS is very suitable for regression analysis of mixed gases. Therefore, this embodiment proposes an electronic nose width learning method (enhanced BLS) to improve the traditional BLS. First, the parameters in the model are optimized to obtain better prediction results for regression analysis of the mixed gas of methane and ethylene. Second, the assignment method of the initial random mapping matrix of the BLS model is changed, and verified by experimental results, and Gaussian distribution is selected as the initial random mapping matrix assignment method. When the enhanced BLS is used for regression prediction of a mixed gas of methane and ethylene, the final result is better than BLS, other enhanced width learning models based on BLS, BPNN, LSSVM, ELM and LR.

[0107] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An electronic nose width learning method for regression prediction of mixed gas concentration, characterized in that: Includes steps: S1. Select different sparse regularization parameters, feature node shrinkage parameters and initial random mapping matrix assignment methods to construct multiple enhanced width learning models based on the width learning model; S2. Obtain a sample set and divide it into a training data set and a test data set according to a preset ratio; the sample set includes at least ten gas concentration value indicators, and the gas concentration value indicator is the mixed gas concentration after at least two measurement gases are mixed according to the 1st to Nth preset concentration combinations, N ≥ 10; S3. Using the training data set to train the plurality of enhanced width learning models, and recording the corresponding training time; S4. Testing the trained multiple enhanced width learning models using a test data set, and recording the regression prediction error rate; S5. Taking into account the training time and the regression prediction error rate, the enhanced width learning model with the best performance is determined as the electronic nose regression prediction model; the sparse regularization parameter of the electronic nose regression prediction model is 0.00002; the feature node shrinkage parameter of the electronic nose regression prediction model is 1e-3; the initial random mapping matrix assignment method of the electronic nose regression prediction model is Gaussian distribution.

2. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 1, characterized in that: In step S1, the selection range of the sparse regularization parameter is between 2^-30 and 0.0002.

3. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 1, characterized in that: In the step S1, the selection range of the characteristic node shrinkage parameter is between 1e-4 and 1e-2.

4. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 1, characterized in that: In the step S1, the options for the initial random mapping matrix assignment method include Gaussian distribution, exponential distribution, Poisson distribution, Rayleigh distribution, lognormal distribution, continuous uniform distribution, and β distribution.

5. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 1, characterized in that: In step S2, each gas concentration value indicator corresponds to at least 100 samples; the preset ratio is training data set: test data set, and the preset ratio is between 4:1 and 2:

1.

6. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 5, characterized in that: The preset ratio is 3:

1.

7. The electronic nose width learning method for performing regression prediction on the concentration of mixed gas according to claim 5, characterized in that: The gas concentration value indicator is the concentration of the mixed gas after methane and ethylene are mixed according to the 1st to 38th preset concentration combinations.

Images