A local dynamic neural network method for quantitative analysis of mixed gases
By analyzing the quantitative analysis of local dynamic neural network hybrid gas, the response data of the gas sensor array is used for interpolation and feature extraction, which solves the problem of inaccurate analysis of the gas sensor array in a variety of gas environments. Concentration prediction.
Patent Information
- Application Number
- CN202310407607.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-17
- Publication Date
- 2025-01-21
- Estimated Expiration
- 2043-04-17
AI Technical Summary
It is difficult for existing gas sensor arrays to accurately identify the concentration of mixed gases in the coexistence environment of multiple gases, and there is a nonlinear response phenomenon, resulting in inaccurate quantitative analysis.
Using local dynamic neural network hybrid gas quantitative analysis methods, collect response data through the gas sensor array, perform interpolation processing to expand the sample set, build a self -encoder extraction response data feature, and use local dynamic neural network models High accuracy prediction of mixed gas concentration.
It effectively reduces training costs, improves prediction accuracy, simplifies hardware implementation difficulty, and is suitable for quantitative analysis of multiple gases.
Smart Images
Figure CN116448957B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of mixed gas detection based on a gas sensor array, and in particular relates to a local dynamic neural network mixed gas quantitative analysis method. Background Art
[0002] Gas sensors have the advantages of simplicity, small size, and real-time detection. They play an important role in gas detection fields such as civil, industrial, and environmental monitoring. Gas sensors have selectivity problems and will respond to gases other than the target gas. In an environment where multiple gases exist, a single gas sensor cannot accurately identify the concentration of the target gas, which greatly limits its application. Faced with an environment where two or more gases coexist, there is a development trend from a single gas sensor to a gas sensor array. The gas sensor array has a high accuracy in identifying a single gas, but due to the cross-interference of the gas of a single gas sensor, the array has a nonlinear response to the mixed gas, resulting in inaccurate identification of the concentration of the mixed gas. Generally, the accurate analysis of the mixed gas by the gas sensor array is achieved by combining a quantitative analysis algorithm.
[0003] Mixed gas quantitative analysis methods are mainly divided into two categories: multiple regression analysis and artificial neural network. For small sample data, multiple regression analysis methods such as principal component regression, support vector regression and partial least squares regression can obtain more stable quantitative analysis results than neural networks. However, the multiple regression method is not applicable when there are many types of gases. Due to the interaction between gases, there are different variable relationships in the response of the sensor. The establishment of a multiple regression model greatly increases the complexity and is not conducive to simple implementation. In this case, neural networks are a better choice. The differences between the sensor response characteristics can better realize the quantitative analysis of multiple gases. However, the performance of the neural network algorithm is very dependent on the capacity of the training data. The more complex the network model, the stronger its ability, and more training data is required. Therefore, to obtain a high-performance model for practical applications, increasing the data sample size is an indispensable step. However, since gas sensors are susceptible to environmental interference, it is difficult to maintain the consistency and reliability of the test data sample during the experiment through manual operation. In addition, the practical application of mixed gas quantitative analysis methods involves the hardwareization of algorithms, but traditional neural network models such as CNN (convolutional neural network) have complex neural network structures and high training costs, which are not conducive to hardware implementation. Summary of the invention
[0004] In view of the technical problems existing in the above-mentioned prior art, the present invention proposes a local dynamic neural network mixed gas quantitative analysis method to achieve high-accuracy quantitative analysis of mixed gases.
[0005] The technical solution adopted by the present invention is as follows:
[0006] A local dynamic neural network mixed gas quantitative analysis method, characterized by comprising the following steps:
[0007] S1. Testing a mixed gas with known gas concentrations through a gas sensor array. Each gas sensor in the gas sensor array collects corresponding response data to form a gas sensor array response vector. The concentrations of each gas in the mixed gas form a mixed gas concentration vector. A sample point is formed by the mixed gas concentration vector and the corresponding gas sensor array response vector. Multiple groups of mixed gases with different concentrations are introduced to obtain a sample set S consisting of N sample points. The number of gas types in the mixed gas is m.
[0008] S2. Interpolate the sample set S to obtain the interpolated sample set S*. All the mixed gas concentration vectors in the interpolated sample set S* constitute a concentration sample set, and the corresponding gas sensor array response vectors constitute a response sample set; specifically:
[0009] According to the distance between the mixed gas concentration vector of each sample point in the sample set S and the zero vector of the same dimension, the sample points are sorted in ascending order of distance; for the first N-1 sample points in the sorted sample set S', the concentration distances between the i, i = 1, 2, ..., N-1 sample points and all subsequent sample points are calculated in turn, and the mixed gas concentration vector is interpolated between the i, i = 1, 2, ..., N-1 sample points and the sample points with a distance less than the average concentration, and the k sample points nearest to the interpolation point are found as neighbor points, based on which the response vector of the interpolation point is calculated, and then the interpolation point is added to the sorted sample set S'; until all the interpolation points of the first N-1 sample points in the sorted sample set S' are added to the sorted sample set S', and the interpolated sample set S* is obtained;
[0010] S3, construct an autoencoder consisting of an encoder and a decoder, the encoder and the decoder share a bottleneck layer; the response sample set is randomly divided into a response sample training set and a response sample test set, the response sample training set is used as the input of the encoder and the output of the decoder, the autoencoder is trained to obtain a trained autoencoder; the response sample set is input into the trained autoencoder to obtain a trained response sample set, and the neural network structure and weights of the encoder including the bottleneck layer in the autoencoder are extracted; the concentration sample set and the trained response sample set form a new sample set D;
[0011] S4, passing the mixed gas to be tested into the gas sensor array, collecting response data in real time, and obtaining the response vector of the mixed gas to be tested; traversing all sample points in the new sample set D, calculating the distance between the response vector of the mixed gas to be tested and the response vector of the sample point, and finding the l sample points in the new sample set D that are the nearest neighbors to the response vector of the mixed gas to be tested as local points, and then obtaining the response sample set and concentration sample set of the local points;
[0012] S5, constructing a local dynamic neural network model (i.e., a measurement model); wherein the hidden layer of the local dynamic neural network model is composed of an encoder including a bottleneck layer in the autoencoder, and uses the weights extracted by S3, and the bottleneck layer is followed by a fully connected layer as an output layer;
[0013] S6. Taking the response sample set of the local point as input and the concentration of a gas in the corresponding concentration sample set as the training target, the local dynamic neural network model is trained to obtain m trained local dynamic neural network models corresponding to various gases; the response vectors of the mixed gas to be tested are respectively input into the m trained local dynamic neural network models, and the concentrations of each gas of the mixed gas to be tested are output.
[0014] Furthermore, the sample set S in S1 = {(X1, Y1), (X2, Y2), ..., (X N ,Y N )}; where X i =(x i (1) ,x i (2) ,...,x i (m) ), i = 1, 2, ..., N is the mixed gas concentration vector of the i-th sample point in the sample set S, where x i (j) ,j=1,2,...,m is the concentration of the jth gas in the i-th sample point in the sample set S; Y i =(y i (1) ,y i (2) ,...,y i (n) ), i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sample set S, where n is the number of gas sensors in the gas sensor array, y i (v) ,v=1,2,...,n is the response value of the vth gas sensor in the i-th sample point in the sample set S.
[0015] Furthermore, the total number N of sample points is determined according to the concentration range of the mixed gas to be tested.
[0016] Furthermore, the specific steps of S2 include:
[0017] S21, traverse all sample points in the sample set S, and calculate the mixed gas concentration vector X of the sample point i The distance from the zero vector O of the same dimension Sort the sample points in ascending order of distance to obtain the sorted sample set S'={(X1',Y1'),(X2',Y2'),...,(X N ',Y N ')}; where X i ',i=1,2,...,N is the mixed gas concentration vector of the i-th sample point in the sorted sample set S'; Y i ', i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sorted sample set S';
[0018] S22, let h=1, and the sorted sample set S' is used as the initial interpolated sample set S*;
[0019] S23, for the hth sample point (X h ',Y h '), calculate its difference with all subsequent sample points (X u ',Y u '),u=h+1,...,N concentration distance Then calculate the hth sample point (X h ',Y h ') corresponds to the average concentration distance
[0020]
[0021] S24, at the hth sample point (X h ',Y h ') and less than the average concentration distance Take the average value of the mixed gas concentration vector between the sample points to do the mixed gas concentration vector interpolation, and get the mixed gas concentration vector X at the interpolation point new =(x new (1) ,x new (2) ,...,x new (m) ), where x new (j) ,j=1,2,...,m is the concentration of the jth gas at the interpolation point; traverse all sample points in the interpolated sample set S* and calculate the concentration distance between the interpolation point and the sample point Where N' is the total number of sample points in the interpolated sample set S*, X p *, p = 1, 2, ..., N' is the mixed gas concentration vector of the pth sample point in the interpolated sample set S*; based on the concentration distance Find the k sample points {(X1”,Y1”),(X2”,Y2”),...,(Xk ”,Y k ”)}, as neighbor points, and then calculate the response vector Y of the interpolation point new =(y new (1) ,y new (2) ,...,y new (n) ):
[0022]
[0023] Among them, X q ”, q=1,2,...,k is the mixed gas concentration vector of the qth neighbor point nearest to the interpolation point; Y q ”, q=1,2,...,k is the response vector of the qth neighbor point nearest to the interpolation point; y new (v) ,v=1,2,...,n is the response value of the vth gas sensor corresponding to the interpolation point;
[0024] S25, interpolation point (X new ,Y new ) is added to the interpolated sample set S*, and it is determined whether h is equal to N-1. If so, the interpolation is completed to obtain the final interpolated sample set S*; otherwise, h=h+1 is set and the process returns to step S23.
[0025] Furthermore, the value range of k in S2 is 2-10.
[0026] Furthermore, the number of neurons in the input layer and the output layer of the autoencoder in S3 is the number of gas sensors, the number of neurons in the bottleneck layer ranges from 10 to 20, the number of hidden layers of the encoder and the decoder is 1, and the number of neurons ranges from 20 to 40.
[0027] Furthermore, the value range of l in S4 is 25 to 50.
[0028] Furthermore, the Euclidean metric is used to calculate the distance.
[0029] The beneficial effects of the present invention are:
[0030] 1. The present invention proposes a local dynamic neural network mixed gas quantitative analysis method, which first expands the small sample data by interpolation, then realizes the feature extraction of the response data through the autoencoder to reduce the cross-interference of different gases, and finally realizes the prediction of the mixed gas concentration by a local dynamic neural network model composed of an encoder network structure and fixed weights to learn the response data more finely;
[0031] 2. Traditional neural network models often input all sample data for training, which takes a long time to train. The present invention effectively reduces the training cost by training the local dynamic neural network model that only inputs a small sample set (local points). In addition, the local points can better reflect the characteristics of the points to be tested, achieving a more accurate prediction effect.
[0032] 3. In the hardware implementation of the algorithm, the present invention uses an autoencoder to extract data features, fixes the encoder and its weights in the autoencoder, and uses local neighbor data to train the fully connected layer. The network structure is simple, which greatly shortens the training process, saves training time, and simplifies the difficulty of hardware implementation. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 The structure diagram of the autoencoder constructed for Example 1;
[0034] Figure 2 A structural diagram of a local dynamic neural network model constructed in Example 1;
[0035] Figure 3 The error mean values and 95% confidence intervals of the predicted concentrations of the four gases in Example 1; wherein (a) is the absolute error and (b) is the relative error. DETAILED DESCRIPTION
[0036] 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 implementations. 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.
[0037] Example 1
[0038] This embodiment provides a local dynamic neural network mixed gas quantitative analysis method, which specifically includes the following steps:
[0039] S1. Test the mixed gas with known gas concentrations through the gas sensor array, collect the corresponding gas sensor array basic resistance and response resistance data, and calculate the ratio of the two to obtain response data. The response data corresponding to each gas sensor in the gas sensor array constitutes a gas sensor array response vector, and the concentration of each gas in the mixed gas constitutes a mixed gas concentration vector; a sample point is formed by the mixed gas concentration vector and the corresponding gas sensor array response vector, and multiple groups of mixed gases with different concentrations are introduced to obtain a sample set S={(X1,Y1),(X2,Y2),...,(X N ,Y N )};
[0040] Among them, X i =(x i(1) ,x i (2) ,...,x i (m) ), i = 1, 2, ..., N is the mixed gas concentration vector of the i-th sample point in the sample set S, where m is the number of gas types in the mixed gas. In this embodiment, m is 4, x i (j) ,j=1,2,...,m is the concentration of the jth gas in the i-th sample point in the sample set S; Y i =(y i (1) ,y i (2) ,...,y i (n) ), i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sample set S, where n is the number of gas sensors in the gas sensor array. In this embodiment, n is 4, y i (v) ,v=1,2,...,n is the response value of the vth gas sensor in the i-th sample point in the sample set S.
[0041] S2. Interpolate the sample set S to obtain the interpolated sample set S*. All the mixed gas concentration vectors in the interpolated sample set S* constitute a concentration sample set, and the corresponding gas sensor array response vectors constitute a response sample set; specifically:
[0042] S21, traverse all sample points in the sample set S, and calculate the mixed gas concentration vector X of the sample point i The distance from the zero vector O of the same dimension Sort the sample points in ascending order of distance to obtain the sorted sample set S'={(X1',Y1'),(X2',Y2'),...,(X N ',Y N ')}; where X i ',i=1,2,...,N is the mixed gas concentration vector of the i-th sample point in the sorted sample set S'; Y i ', i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sorted sample set S';
[0043] S22, let h=1, and the sorted sample set S' is used as the initial interpolated sample set S*;
[0044] S23, for the hth sample point (X h ',Y h '), calculate its difference with all subsequent sample points (X u ',Yu '),u=h+1,...,N concentration distance Then calculate the hth sample point (X h ',Y h ') corresponds to the average concentration distance
[0045]
[0046] S24, at the hth sample point (X h ',Y h ') and less than the average concentration distance Take the average of the mixed gas concentration vector between the sample points to perform mixed gas concentration vector interpolation, and get the mixed gas concentration vector X at the interpolation point new =(x new (1) ,x new (2) ,...,x new (m) ), where x new (j) ,j=1,2,...,m is the concentration of the jth gas at the interpolation point; traverse all sample points in the interpolated sample set S* and calculate the concentration distance between the interpolation point and the sample point Where N' is the total number of sample points in the interpolated sample set S*, X p *, p = 1, 2, ..., N' is the mixed gas concentration vector of the pth sample point in the interpolated sample set S*; based on the concentration distance Find the k=5 sample points {(X1”,Y1”),(X2”,Y2”),...,(X k ”,Y k ”)}, as neighbor points, and then calculate the response vector Y of the interpolation point new =(y new (1) ,y new (2) ,...,y new (n) ):
[0047]
[0048] Among them, X q ”, q=1,2,...,k is the mixed gas concentration vector of the qth neighbor point nearest to the interpolation point; Y q ”, q=1,2,...,k is the response vector of the qth neighbor point nearest to the interpolation point; y new (v),v=1,2,...,n is the response value of the vth gas sensor corresponding to the interpolation point;
[0049] S25, interpolation point (X new ,Y new ) is added to the interpolated sample set S*, and it is determined whether h is equal to N-1=109. If so, the interpolation is completed to obtain the final interpolated sample set S*; otherwise, h=h+1 is set and the process returns to step S23.
[0050] S3. Construct an autoencoder consisting of an encoder and a decoder. The encoder and the decoder share the bottleneck layer. The structure of the autoencoder is as follows: Figure 1 As shown, including:
[0051] 1 input layer, the number of neurons is 4;
[0052] 3 hidden layers, with 35, 15, and 35 neurons respectively;
[0053] 1 output layer, the number of neurons is 4;
[0054] The activation functions of the input layer, hidden layer, and output layer are all Leaky ReLU (leaky linear rectification) functions, and alpha (hyperparameter) is set to 0.2;
[0055] The response sample set is randomly divided into a response sample training set and a response sample test set. The response sample training set is used as the input of the encoder and the output of the decoder respectively, and the autoencoder is trained to obtain a trained autoencoder. The response sample set is input into the trained autoencoder to obtain a trained response sample set, and the neural network structure and weights of the encoder including the bottleneck layer in the autoencoder are extracted. The concentration sample set and the trained response sample set form a new sample set D.
[0056] S4, passing the mixed gas to be tested into the gas sensor array, collecting response data in real time, and obtaining the response vector of the mixed gas to be tested; traversing all sample points in the new sample set D, calculating the distance between the response vector of the mixed gas to be tested and the response vector of the sample point, and finding l=30 sample points in the new sample set D that are the nearest neighbors to the response vector of the mixed gas to be tested as local points, and then obtaining the response sample set and concentration sample set of the local points;
[0057] S5, construct a local dynamic neural network model (i.e., measurement model); wherein, the hidden layer of the local dynamic neural network model is composed of an encoder including a bottleneck layer in the autoencoder, and the weights extracted in S3 are used, and the bottleneck layer is connected to a fully connected layer as the output layer; the structure of the local dynamic neural network model is as follows Figure 2 As shown, including:
[0058] 1 input layer, the number of neurons is 4;
[0059] 2 hidden layers, with 35 and 15 neurons respectively;
[0060] 1 output layer, the number of neurons is 1;
[0061] Among them, the activation function of the output layer is a Linear function;
[0062] S6. Taking the response sample set of the local point as input and the concentration of a gas in the corresponding concentration sample set as the training target, the local dynamic neural network model is trained to obtain m trained local dynamic neural network models corresponding to various gases; the response vectors of the mixed gas to be tested are respectively input into the m trained local dynamic neural network models, and the concentrations of each gas of the mixed gas to be tested are output.
[0063] This embodiment takes a mixed gas composed of nitrogen dioxide, ammonia, methane and carbon dioxide as an example, and uses an adaptive moment estimation (Adam) optimizer to adjust the parameters of the dynamic local neural network to minimize the loss function. The optimal hyperparameters for training the local dynamic neural network model are shown in Table 1. The gas concentration prediction results of the mixed gas are shown by the average values of the absolute error and relative error and the confidence intervals at 95% confidence level, respectively. Figure 3 As shown in (a) and 3(b), the relative errors of gas concentration prediction of the four gases are all less than 4% at 95% confidence level, with high accuracy and good prediction effect.
[0064] Table 1 Optimal hyperparameters for dynamic local neural network training
[0065]
[0066] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other alternative features that are equivalent or have similar purposes; all the disclosed features, or all the steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.
Claims
1. A local dynamic neural network mixed gas quantitative analysis method, characterized in that: The following steps are involved: S1. The gas sensor array is used to test the mixed gas with known gas concentrations. Each gas sensor in the gas sensor array collects corresponding response data to form a gas sensor array response vector. Each gas concentration constitutes a mixed gas concentration vector. A sample point is formed by the mixed gas concentration vector and the corresponding gas sensor array response vector. Multiple groups of mixed gases with different concentrations are introduced to obtain a sample set S consisting of N sample points = {(X1, Y1), (X2, Y2), ..., (X N ,Y N )}; where the number of gas types in the mixed gas is m; X i =(x i (1) ,x i (2) ,...,x i (m) ), i = 1, 2, ..., N is the mixed gas concentration vector of the i-th sample point in the sample set S, where x i (j) ,j=1,2,...,m is the concentration of the jth gas in the i-th sample point in the sample set S; Y i =(y i (1) ,y i (2) ,...,y i (n) ), i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sample set S, where n is the number of gas sensors in the gas sensor array, y i (v) ,v=1,2,...,n is the response value of the vth gas sensor in the i-th sample point in the sample set S; S2. Interpolate the sample set S to obtain the interpolated sample set S*, where all the mixed gas concentration vectors constitute the concentration sample set, and the corresponding gas sensor array response vectors constitute the response sample set; specifically: According to the distance between the mixed gas concentration vector of each sample point in the sample set S and the zero vector of the same dimension, the sample points are sorted in ascending order of distance; for the first N-1 sample points in the sorted sample set S', the concentration distances between the i, i = 1, 2, ..., N-1 sample points and all subsequent sample points are calculated in turn, and the mixed gas concentration vector is interpolated between the i, i = 1, 2, ..., N-1 sample points and the sample points with a distance less than the average concentration, and the k sample points nearest to the interpolation point are found as neighbor points, based on which the response vector of the interpolation point is calculated, and then the interpolation point is added to the sorted sample set S'; until all the interpolation points of the first N-1 sample points in the sorted sample set S' are added to the sorted sample set S', and the interpolated sample set S* is obtained; The specific steps include: S21, traverse all sample points in the sample set S, and calculate the mixed gas concentration vector X of the sample point i The distance d from the zero vector O of the same dimension (Xi,O) , i=1,2,...N, sort the sample points in ascending order of distance, and get the sorted sample set S'={(X1',Y1'),(X2',Y2'),...,(X N ',Y N ')}; where X i ',i=1,2,...,N is the mixed gas concentration vector of the i-th sample point in the sorted sample set S'; Y i ', i = 1, 2, ..., N is the gas sensor array response vector of the i-th sample point in the sorted sample set S'; S22, let h=1, and the sorted sample set S' is used as the initial interpolated sample set S*; S23, for the hth sample point (X h ',Y h '), calculate its difference with all subsequent sample points (X u ',Y u '),u=h+1,...,N concentration distance Then calculate the hth sample point (X h ',Y h ') corresponds to the average concentration distance S24, at the hth sample point (X h ',Y h ') and less than the average concentration distance Take the average value of the mixed gas concentration vector between the sample points to do the mixed gas concentration vector interpolation, and get the mixed gas concentration vector X at the interpolation point new =(x new (1) ,x new (2) ,...,x new (m) ), where x new (j) ,j=1,2,...,m is the concentration of the jth gas at the interpolation point; traverse all sample points in the interpolated sample set S* and calculate the concentration distance between the interpolation point and the sample point Where N' is the total number of sample points in the interpolated sample set S*, X p *, p = 1, 2, ..., N' is the mixed gas concentration vector of the pth sample point in the interpolated sample set S*; based on the concentration distance Find the k sample points {(X1”,Y1”),(X2”,Y2”),...,(X k ”,Y k ”)}, as neighbor points, and then calculate the response vector Y of the interpolation point new =(y new (1) ,y new (2) ,...,y new (n) ): Among them, X q ”, q=1,2,...,k is the mixed gas concentration vector of the qth neighbor point nearest to the interpolation point; Y q ”, q=1,2,...,k is the response vector of the qth neighbor point nearest to the interpolation point; y new (v) ,v=1,2,...,n is the response value of the vth gas sensor corresponding to the interpolation point; S25, interpolation point (X new ,Y new ) is added to the interpolated sample set S*, and it is determined whether h is equal to N-1. If so, the interpolation is completed to obtain the final interpolated sample set S*; otherwise, h=h+1 is set and the process returns to step S23; S3, construct an autoencoder consisting of an encoder and a decoder, the encoder and the decoder share a bottleneck layer; the response sample set is randomly divided into a response sample training set and a response sample test set, the response sample training set is used as the input of the encoder and the output of the decoder, the autoencoder is trained to obtain a trained autoencoder; the response sample set is input into the trained autoencoder to obtain a trained response sample set, and the neural network structure and weights of the encoder including the bottleneck layer in the autoencoder are extracted; the concentration sample set and the trained response sample set form a new sample set D; S4, passing the mixed gas to be tested into the gas sensor array, collecting response data in real time, and obtaining the response vector of the mixed gas to be tested; traversing all sample points in the new sample set D, calculating the distance between the response vector of the mixed gas to be tested and the response vector of the sample point, and finding the l sample points in the new sample set D that are the nearest neighbors to the response vector of the mixed gas to be tested as local points, and then obtaining the response sample set and concentration sample set of the local points; S5, constructing a local dynamic neural network model; wherein the hidden layer of the local dynamic neural network model is composed of an encoder including a bottleneck layer in the autoencoder, and uses the weights extracted by S3, and the bottleneck layer is followed by a fully connected layer as an output layer; S6. Taking the response sample set of the local point as input and the concentration of a gas in the corresponding concentration sample set as the training target, the local dynamic neural network model is trained to obtain m trained local dynamic neural network models corresponding to various gases; the response vectors of the mixed gas to be tested are respectively input into the m trained local dynamic neural network models, and the concentrations of each gas of the mixed gas to be tested are output.
2. According to the local dynamic neural network mixed gas quantitative analysis method of claim 1, it is characterized in that: The total number of sample points N is determined according to the concentration range of the mixed gas to be tested.
3. The local dynamic neural network mixed gas quantitative analysis method according to claim 1 is characterized in that: The value range of k in S2 is 2 to 10.
4. The local dynamic neural network mixed gas quantitative analysis method according to claim 1, characterized in that: The value range of l in S4 is 25 to 50.
5. The local dynamic neural network mixed gas quantitative analysis method according to claim 1, characterized in that: In S3, the number of neurons in the input layer and output layer of the autoencoder is the number of gas sensors, the number of neurons in the bottleneck layer ranges from 10 to 20, the number of hidden layers of the encoder and the decoder is 1, and the number of neurons ranges from 20 to 40.