Sensing array gas detection data enhancement and concentration identification method based on generative adversarial network

Abstract

The invention discloses a sensor array gas detection data enhancement and concentration identification method based on a generative adversarial network, which comprises the following steps: acquiring original gas detection data, and preprocessing the original gas detection data to obtain preprocessed detection data; constructing a conditional generative adversarial network, and inputting the preprocessed detection data and real data into the conditional generative adversarial network to generate virtual data; and inputting the virtual data and the preprocessed detection data into a concentration-temperature prediction model to obtain concentration and temperature prediction values of each component of the mixed gas. The gas concentration and temperature are used as condition input, sensing response data with physical significance are generated, the model generalization ability under the small sample condition is remarkably improved, and a solution is provided for multi-component gas detection and cross sensitivity suppression.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas identification of sensor arrays, and in particular relates to a sensor array gas detection data enhancement and concentration identification method based on a generative adversarial network. Background Art

[0002] In the field of multi-component gas detection, detection technology based on gas sensor arrays faces multi-dimensional technical challenges. The sensor array consists of four cross-sensitive sensor elements, each of which responds to the four characteristic gases: carbon monoxide, hydrogen, methane, and carbon dioxide. However, in actual detection, each sensor element can only output a single integrated response value. These responses are not only affected by the concentrations of the four gases but also closely related to the ambient temperature, forming a complex multivariable coupled system. Traditional modeling methods have difficulty accurately establishing the complex mapping relationship between multi-dimensional response values ​​and multi-parameter outputs under small sample conditions. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a sensor array gas detection data enhancement and concentration recognition method based on a generative adversarial network. By taking gas concentration and temperature as conditional inputs, it generates physically meaningful sensor response data, significantly improving the model generalization ability under small sample conditions, and providing a solution for multi-component gas detection and cross-sensitivity suppression.

[0004] To achieve the above objectives, the present invention provides a sensor array gas detection data enhancement and concentration recognition method based on a generative adversarial network, comprising:

[0005] Acquiring raw gas detection data, preprocessing the raw gas detection data, and obtaining preprocessed detection data;

[0006] Constructing a conditional generative adversarial network, inputting the preprocessed detection data and real data into the conditional generative adversarial network to generate virtual data;

[0007] The virtual data and the pre-processed detection data are input into a concentration-temperature prediction model to obtain the concentration and temperature prediction values ​​of each component of the mixed gas.

[0008] Optionally, obtaining pre-processed detection data includes:

[0009]

[0010] Among them, x new is the pre-processed detection data; x is the unprocessed original sensor response value; x min is the minimum value of the sensor response value; x max is the maximum value of the sensor response value.

[0011] Optionally, building a conditional generative adversarial network includes:

[0012] S1. Construct a dual network structure of a generator and a discriminator, where both the generator and the discriminator adopt a fully connected network;

[0013] S2, a random noise vector, target gas concentration conditions and ambient temperature conditions are jointly input into the generator, and simulated sensor response data S' is generated through multi-layer perceptron mapping;

[0014] S3. The discriminator determines the authenticity of the input sensor data, receives the real sensor data and the generated data, and uses the corresponding gas concentration condition y and the ambient temperature condition t as joint input conditions to calculate the probability value of the real distribution of the input data through the multi-layer perceptron;

[0015] S4, fixing the generator parameters, and updating the discriminator by maximizing the discrimination score of the real data and minimizing the discrimination score of the generated data;

[0016] S5. Fix the discriminator parameters, and update the generator by minimizing the discriminator's discrimination score on the generated data, while adding concentration condition and temperature condition matching constraints;

[0017] S6. Repeat the training process of the discriminator and the generator from step S2 to step S5 until the discriminator cannot effectively distinguish between real data and generated data, and the generated data simultaneously satisfies the multiple physical constraints of gas concentration-temperature-response;

[0018] S7, verifying the distribution similarity of the generated data;

[0019] S8. After the training is completed, the generator generates virtual data that conforms to the real distribution according to any given target concentration, temperature and random noise.

[0020] Optionally, the network structure of the generator is:

[0021] G(z,y,t)=MLP(z||y||t,θ G );

[0022] Where z is the noise vector, y is the target gas concentration condition vector, t is the ambient temperature condition scalar, || represents vector splicing, θ G The generator parameters.

[0023] Optionally, the network structure of the discriminator is:

[0024] D(S,y,t)=σ(MLP(S||y||t,θ D));

[0025] Among them, S is the response value input and σ is the sigmoid activation function.

[0026] Optionally, updating the discriminator by maximizing the discriminant score of the real data and minimizing the discriminant score of the generated data includes:

[0027]

[0028] Among them, θ D is the trainable parameter of the discriminator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ D The gradient of S i is the real sensor response data of the i-th sensor; y i is the true value of the gas concentration of the i-th sample; t i is the true value of the ambient temperature of the i-th sample; logD(S i ,y i ,t i ) is the logarithm of the probability of the discriminator to identify the real sample; z i is a randomly sampled noise vector; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,y i ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator's discrimination probability for the virtual response sample.

[0029] Optionally, the generator is updated by minimizing the discriminant score of the discriminator on the generated data, including:

[0030]

[0031] Among them, θ G is the trainable parameter of the generator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ G The gradient of z i is a randomly sampled noise vector; y i is the true value of the gas concentration of the i-th sample; t iis the true value of the ambient temperature of the i-th sample; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,y i ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator’s probability of distinguishing the virtual response sample; λ is the physical constraint weight coefficient; L phys is the physical constraint loss term.

[0032] Optionally, verifying the distribution similarity of generated data includes:

[0033]

[0034] Where JSD(P||Q) is the JS divergence; P is the real data distribution, Q is the generated data distribution, M is the mixed distribution of P and Q; D KL is the KL divergence; D KL (P||M) is the KL divergence of the true distribution P relative to M; D KL (Q||M) is the KL divergence of the generating distribution Q with respect to M.

[0035] Technical effect of the present invention: The present invention discloses a method for gas detection data enhancement and concentration identification of a sensor array based on a generative adversarial network, which realizes the ability to perceive the ambient temperature through the gas sensor array itself, and completely changes the technical route that traditional gas detection must rely on additional temperature sensors. By combining the data enhancement capability of the conditional generative adversarial network (cGAN), the gas concentration and temperature are used as conditional inputs, which effectively expands the scale of training samples and significantly improves the generalization performance of the model under limited experimental data conditions. The present invention deeply explores the inherent temperature response characteristics of the sensor array, establishes a complete mapping relationship between the response value and the gas concentration and temperature, and only requires the response characteristics of four sensor elements to simultaneously and accurately analyze the four gas concentrations and ambient temperature parameters, creating a new technical paradigm for multi-parameter integrated detection. The high-quality synthetic data generated by cGAN not only retains the cross-sensitivity characteristics and temperature dependence of the sensor array, but also greatly improves the diagnostic accuracy of the concentration identification model. This innovation not only achieves accurate perception of temperature parameters, but also fundamentally solves the inherent defects of traditional methods such as system complexity and measurement asynchrony caused by external temperature sensors. It demonstrates unique advantages such as breakthroughs in the dual functions of the sensor array, improved measurement temporal and spatial consistency, and simplified system architecture. It provides a new technical solution for the field of multi-component gas monitoring and has important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0037] Figure 1 This is a flow chart of a method for gas detection data enhancement and concentration identification based on a generative adversarial network for a sensor array according to an embodiment of the present invention;

[0038] Figure 2 This is a network structure diagram of the conditional generative adversarial network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] like Figure 1As shown, this embodiment provides a sensor array gas detection data enhancement and concentration recognition method based on a generative adversarial network, including:

[0042] S1. Obtain historical data in the gas detection system, including standard concentration ratio data of four target gases (such as CO, CH4, H2, CO2), and the response signals (S1, S2, S3, S4) of the array composed of four sensor elements in a mixed gas environment, and temperature (T) and perform normalization processing;

[0043] S2. After preprocessing the limited sensor data collected, a conditional generative adversarial network (cGAN) is constructed. The concentration combination of the four gases and the ambient temperature are used as dual conditional inputs. The discriminator receives the real sensor data and the generated data and judges their authenticity. Through adversarial training, the model parameters are optimized until the statistical characteristics of the generated data are consistent with the real data, achieving the goal of expanding the training sample size.

[0044] S3. The generated virtual data and the measured data are input into a concentration-temperature inversion model (such as a neural network, support vector machine, etc.) to establish a complete mapping relationship from sensor response (S1-S4) to gas concentration (C1-C4) and temperature (T);

[0045] In practical applications, the trained concentration inversion model is deployed in a gas detection system. When the sensor array collects mixed gas response signals (S1-S4), they are directly input into the model for calculation. The model outputs the predicted concentrations (C1-C4) and temperature (T) of the four gases. The prediction error for each gas concentration is then evaluated, and the final gas concentration detection model is output.

[0046] Furthermore, the normalization process in step S1 is implemented by the following formula:

[0047]

[0048] Among them, x new is the pre-processed detection data; x is the unprocessed original sensor response value; x min is the minimum value of the sensor response value; x max is the maximum value of the sensor response value.

[0049] Further, such as Figure 2 As shown in Figure 2, the specific steps of the conditional generation adversarial network in S2 are:

[0050] S21. Construct a dual network structure of generator G and discriminator D. Both the generator and the discriminator use a fully connected network (MLP).

[0051] S22, the generator receives a random noise vector z, target gas concentration condition y and ambient temperature condition t as joint input, and generates simulated sensor response data S' through multi-layer perceptron mapping. The generated response data is required to be close to the response pattern of the real sensor under concentration and temperature y in distribution;

[0052] S23, the discriminator judges the authenticity of the input sensor data, and receives both the real sensor data S and the generated data S', and both are equipped with the corresponding gas concentration condition y and ambient temperature condition t as joint input conditions. The multi-layer perceptron calculates the probability value of the input data coming from the real distribution and outputs a discrimination score between 0 and 1;

[0053] S24. The generator parameters are fixed, and the discriminator is updated by maximizing the discrimination score of the real data and minimizing the discrimination score of the generated data;

[0054] S25, fix the discriminator parameters, and update the generator by minimizing the discriminator's discrimination score on its generated data, while adding concentration condition and temperature condition matching constraints;

[0055] S26. Alternately repeat the discriminator and generator training process of step S22 to step S25 until the discriminator cannot effectively distinguish between real data and generated data, and the generated data simultaneously satisfies the multiple physical constraints of gas concentration-temperature-response;

[0056] S27, generating data to verify distribution similarity;

[0057] S28. After the training is completed, the generator can generate sensor response data S' that conforms to the real distribution based on any given target concentration y, temperature t and random noise z, which is used to expand the original data set.

[0058] Furthermore, the generator and discriminator network structures in S21 are:

[0059] The network structure formula of the generator is:

[0060] G(z,y,t)=MLP(z||y||t,θ G );

[0061] Where z is the noise vector, y is the target gas concentration condition vector, t is the ambient temperature condition scalar, || represents vector splicing, and θ G is the generator parameter;

[0062] The network structure formula of the discriminator is:

[0063] D(S,y,t)=σ(MLP(S||y||t,θ D ));

[0064] Where y is the target gas concentration condition vector, t is the ambient temperature condition scalar, S represents the response value input, and σ is the sigmoid activation function.

[0065] Furthermore, the objective functions in S24 and S25 are:

[0066]

[0067] Where λ is the physical constraint weight (it is recommended to be 0.5 initially and gradually increased to 1.0);

[0068] Physical constraints:

[0069]

[0070] Among them, f y (·) and f t (·) is the branch head in the shared network that outputs concentration and temperature respectively, and adjusts the strength of the temperature constraint (usually 0.3-0.5 because the temperature fluctuation range is smaller than the concentration).

[0071] Furthermore, the discriminator in S24 is updated:

[0072] Update the discriminator parameters θ through the Adam optimizer D ,

[0073]

[0074] Among them, θ D is the trainable parameter of the discriminator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ D The gradient of S i is the real sensor response data of the i-th sensor; y i is the true value of the gas concentration of the i-th sample; t i is the true value of the ambient temperature of the i-th sample; logD(S i ,y i ,t i ) is the logarithm of the probability of the discriminator to identify the real sample; z i is a randomly sampled noise vector; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,yi ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator's discrimination probability for the virtual response sample.

[0075] Furthermore, the generator in S25 is updated:

[0076] Update the generator parameters θ through the Adam optimizer G ,

[0077]

[0078] Among them, θ G is the trainable parameter of the generator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ G The gradient of z i is a randomly sampled noise vector; y i is the true value of the gas concentration of the i-th sample; t i is the true value of the ambient temperature of the i-th sample; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,y i ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator’s probability of distinguishing the virtual response sample; λ is the physical constraint weight coefficient; L phys is the physical constraint loss term.

[0079] Further, the specific steps of the Adam optimizer are:

[0080] S1. Calculate the gradient:

[0081] For the parameter θ t , the gradient of the current batch is:

[0082] g t =▽ θ L(θ t );

[0083] S2. Update the first-order moment (momentum):

[0084] m t=β1m t-1 +(1-β1)g t ;

[0085] Among them, m t : exponential moving average of the first moment of gradient; β1 (default 0.9): momentum decay rate;

[0086] S3. Update the second-order moment (squared gradient):

[0087]

[0088] Among them, v t : exponential moving average of the second-order moment of the gradient; β1 (default 0.999): square gradient decay rate;

[0089] S4. Bias correction:

[0090]

[0091] S5. Parameter update:

[0092]

[0093] Among them, η is the initial learning rate; ε (default 1e-8) is a small constant to prevent division by zero.

[0094] Furthermore, the distribution similarity of S27 is calculated by calculating the Jensen-Shannon divergence between the generated data and the real data. The formula is as follows:

[0095]

[0096] Where JSD(P||Q) is the JS divergence; P is the real data distribution, Q is the generated data distribution, M is the mixed distribution of P and Q; D KL () is KL divergence; D KL (P||M) is the KL divergence of the true distribution P relative to M; D KL (Q||M) is the KL divergence of the generating distribution Q with respect to M; D is the Kullback-Leibler divergence (KL divergence).

[0097] Further, Jensen-Shannon divergence, the specific steps are as follows:

[0098] S1, real data S and generated data Calculate the kernel density estimate (KDE) for each sensor channel separately:

[0099]

[0100] Where K is a Gaussian function.

[0101] S2, for the discretized response value space:

[0102]

[0103] The value range of S3 and JSD is [0, log2], and the normalized values ​​are usually reported:

[0104]

[0105] In a specific application example of the present invention, in step S1, for this embodiment, the composition of the four gas combination ratio data set of CO, CH4, H2, and CO2 is shown in Table 1 below:

[0106] Table 1

[0107]

[0108] In the gas concentration prediction task, the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) are three core evaluation indicators, which quantify the prediction performance of the model from different perspectives:

[0109] (1) Root mean square error (RMSE):

[0110] Reflects the overall deviation between the predicted value and the true value. It is more sensitive to larger errors. The calculation formula is:

[0111]

[0112] Among them, y i is the true concentration value, is the predicted concentration value. The unit of this value is consistent with the concentration (such as ppm). The smaller the value, the higher the model accuracy.

[0113] (2) Mean absolute error (MAE):

[0114] Measuring the absolute average of the prediction errors is more robust:

[0115]

[0116] (3) Coefficient of determination (R 2 ):

[0117] Evaluate the model's ability to explain the data variance, ranging from 0 to 1:

[0118]

[0119] The closer this indicator is to 1, the better the model fit is; a negative value indicates that the model performance is worse than the mean prediction.

[0120] After training with the original samples, the evaluation indicators of different gas samples and temperatures are shown in Table 2 below:

[0121] Table 2

[0122]

[0123] After using the cGAN model to expand the sample, the evaluation indicators of different gas samples and temperatures are shown in Table 3 below:

[0124] Table 3

[0125]

[0126] Experimental results analysis:

[0127] By comparing Table 2 and Table 3, we can clearly see the significant effect of the cGAN model in gas detection data enhancement. When the generated data enhancement is not used (Table 2), the prediction error of each gas concentration is relatively high, among which the RMSE of C2 reaches 2.0ppm, MAE is 1.3ppm, and R 2 The RMSE of temperature prediction is 0.82, and the RMSE of temperature prediction is 0.8℃, indicating that the prediction performance of the basic model has certain limitations. After the cGAN generated data expansion (Table 3), all indicators are significantly improved: the RMSE of C1 is reduced from 1.5ppm to 0.8ppm, a decrease of 46.7%; the R 2 The accuracy of the model improved from 0.88 to 0.98, which is close to perfect fit. The MAE of temperature prediction was significantly reduced from 0.5℃ to 0.2℃, with an accuracy improvement of 60%. This comprehensive improvement verifies that the data generated by cGAN not only expands the number of samples, but more importantly, fully maintains the statistical characteristics and physical laws of the original data, making the generated data able to effectively improve the generalization ability of the model. It is particularly noteworthy that the R 2 The values ​​all exceeded 0.94, indicating a high degree of consistency between the generated data and the real data, providing a reliable data foundation for accurate monitoring of gas detection systems in complex environments. In practical applications, this improvement means more accurate gas leak warnings, more reliable industrial process monitoring, and more precise environmental quality assessments.

[0128] The JSD results of the samples generated in this example are shown in Table 4 below:

[0129] Table 4

[0130]

[0131]

[0132] Among them, 0-0.15: excellent, 0.15-0.3: acceptable, and greater than 0.3: needs improvement.

[0133] Based on the evaluation results, the cGAN model demonstrates significant advantages in gas sensor data generation. Through a conditional adversarial training mechanism, the model successfully achieves high-fidelity data generation for multiple sensor channels, with the normalized JSD for the S2 and S4 channels reaching excellent levels of 0.12 and 0.07, respectively, indicating a high degree of consistency between the generated data and the real data distribution. Crucially, the joint distribution JSD is an acceptable level of 0.26, demonstrating that the model effectively captures the cross-sensitivity characteristics and cooperative response patterns between sensors. By incorporating gas concentration and temperature as joint conditional inputs and combining them with a physically constrained loss function, cGAN ensures the accuracy of single-sensor responses while maintaining the physical rationality of multi-dimensional responses, such as strictly adhering to inherent sensor characteristics such as range limits and concentration-response monotonicity. Although there is still room for optimization in the S3 channel and joint distribution indicators, compared with other generation methods, cGAN has demonstrated excellent data enhancement capabilities under small sample conditions (only 500 sets of real data). The generated data used for downstream tasks can improve the accuracy of the concentration prediction model by more than 20%, while maintaining key physical properties such as temperature adaptability, providing an efficient solution to the data scarcity problem of gas detection systems.

[0134] In response to the technical challenges in the prior art, the present invention innovatively uses conditional generative adversarial networks (cGAN) for data enhancement. By generating a model, the response rules of the sensor array under multi-gas environments and temperature changes can be learned. This data enhancement method can not only effectively expand the scale of training samples, but also maintain the cross-sensitivity and temperature response characteristics of the sensor elements, providing more reliable data support for subsequent multi-parameter joint inversion. Specifically, the gas concentration and temperature are input into the generator as conditional variables, and the response characteristics of the sensor array under specific conditions are learned through adversarial training. This method can not only effectively expand the training data set, but also ensure that the generated data strictly follows the given concentration-temperature conditions, thereby maintaining the cross-sensitivity and temperature dependence of the sensor array. Compared with traditional GAN, the introduction of cGAN makes the generation process more controllable and physically meaningful, providing a more reliable data basis for subsequent multi-parameter inversion.

[0135] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A sensor array gas detection data enhancement and concentration recognition method based on generative adversarial network, characterized in that: include: Acquiring raw gas detection data, preprocessing the raw gas detection data, and obtaining preprocessed detection data; Constructing a conditional generative adversarial network, inputting the preprocessed detection data and real data into the conditional generative adversarial network to generate virtual data; The virtual data and the pre-processed detection data are input into a concentration-temperature prediction model to obtain the concentration and temperature prediction values ​​of each component of the mixed gas.

2. The sensor array gas detection data enhancement and concentration identification method based on generative adversarial network according to claim 1 is characterized in that: The pre-processed detection data includes: Among them, x new is the pre-processed detection data; x is the unprocessed original sensor response value; x min is the minimum value of the sensor response value; x max is the maximum value of the sensor response value.

3. The sensor array gas detection data enhancement and concentration identification method based on generative adversarial network according to claim 1 is characterized in that: Building a conditional generative adversarial network involves: S1. Construct a dual network structure of a generator and a discriminator, where both the generator and the discriminator adopt a fully connected network; S2, a random noise vector, target gas concentration conditions and ambient temperature conditions are jointly input into the generator, and simulated sensor response data S' is generated through multi-layer perceptron mapping; S3. The discriminator determines the authenticity of the input sensor data, receives the real sensor data and the generated data, and uses the corresponding gas concentration condition y and the ambient temperature condition t as joint input conditions to calculate the probability value of the real distribution of the input data through the multi-layer perceptron; S4, fixing the generator parameters, and updating the discriminator by maximizing the discrimination score of the real data and minimizing the discrimination score of the generated data; S5. Fix the discriminator parameters, and update the generator by minimizing the discriminator's discrimination score on the generated data, while adding concentration condition and temperature condition matching constraints; S6. Repeat the training process of the discriminator and the generator from step S2 to step S5 until the discriminator cannot effectively distinguish between real data and generated data, and the generated data simultaneously satisfies the multiple physical constraints of gas concentration-temperature-response; S7, verifying the distribution similarity of the generated data; S8. After the training is completed, the generator generates virtual data that conforms to the real distribution according to any given target concentration, temperature and random noise.

4. The sensor array gas detection data enhancement and concentration identification method based on generative adversarial network according to claim 3 is characterized in that: The network structure of the generator is: G(z,y,t)=MLP(z||y||t,θ G ); Where z is the noise vector, y is the target gas concentration condition vector, t is the ambient temperature condition scalar, || represents vector splicing, θ G The generator parameters.

5. The method for gas detection data enhancement and concentration identification based on a generative adversarial network for sensor arrays according to claim 3, wherein: The network structure of the discriminator is: D(S,y,t)=σ(MLP(S||y||t,θ D )); Among them, S is the response value input and σ is the sigmoid activation function.

6. The method for gas detection data enhancement and concentration identification based on a generative adversarial network for sensor arrays according to claim 3, wherein: The discriminator is updated by maximizing the discriminant score of the real data and minimizing the discriminant score of the generated data, including: Among them, θ D is the trainable parameter of the discriminator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ D The gradient of S i is the real sensor response data of the i-th sensor; y i is the true value of the gas concentration of the i-th sample; t i is the true value of the ambient temperature of the i-th sample; logD(S i ,y i ,t i ) is the logarithm of the probability of the discriminator to identify the real sample; z i is a randomly sampled noise vector; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,y i ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator's discrimination probability for the virtual response sample.

7. The sensor array gas detection data enhancement and concentration identification method based on generative adversarial network according to claim 3 is characterized in that: The generator is updated by minimizing the discriminant score of the generated data: Among them, θ G is the trainable parameter of the generator; Adam() is the adaptive moment estimation optimization algorithm; is the loss function for θ G The gradient of z i is a randomly sampled noise vector; y i is the true value of the gas concentration of the i-th sample; t i is the true value of the ambient temperature of the i-th sample; G(z i ,y i ,t i ) is the virtual response data output by the generator; D(G(z i ,y i ,t i ),y i ,t i ) is the probability of the discriminator to distinguish the virtual response data; log(1-D(G(z i ,y i ,t i ),y i ,t i )) is the logarithm of the complement of the discriminator’s probability of distinguishing the virtual response sample; λ is the physical constraint weight coefficient; L phys is the physical constraint loss term.

8. The method for gas detection data enhancement and concentration identification based on a generative adversarial network for sensor arrays according to claim 3, wherein: Verifying the distribution similarity of generated data includes: Where JSD(P||Q) is the JS divergence; P is the real data distribution, Q is the generated data distribution, M is the mixed distribution of P and Q; D KL is the KL divergence; D KL (P‖M) is the KL divergence of the true distribution P relative to M; D KL (Q‖M) is the KL divergence of the generating distribution Q with respect to M.

Images