Method for detecting target gas in mixed volatile organic compound gas

Abstract

The invention discloses a method for detecting target gas in mixed volatile organic compound gas, which comprises the following steps: periodically acquiring one-dimensional gas response intensity data of gas to be detected by a sensor array, and converting the one-dimensional gas response intensity data into a two-dimensional color image; and inputting the gas detection model, detecting whether target volatile organic compound gas exists in the gas to be detected, if so, extracting steady-state data Xw in the one-dimensional gas response intensity data, inputting the steady-state data Xw into the gas quantitative model after normalization processing, and outputting the concentration of the target volatile organic compound gas. A Transformer encoder processes an input sequence by using a multi-head attention mechanism and dynamically pays attention to different parts of the input sequence, so that a vector of each position can interact with all positions of the whole input sequence, the capability of capturing a long-distance dependency relationship is stronger, and the self-attention mechanism is global operation and is more prone to capturing low-frequency information, so that the self-attention mechanism is more suitable for being used in a large scale. A CNN encoder extracts local features in a two-dimensional color image, and the local features are fused with global features, so that the preference of Transform to low-frequency gas information is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas detection. More specifically, the present invention relates to a method for detecting target gases in a mixed volatile organic compound gas. Background Art

[0002] With the rapid growth of the economy and the increasing emissions of various volatile organic compounds, long-term exposure to air containing volatile organic compounds will have a great impact on human health. For example, acetone, ethanol, and isopropanol, these three gases are chemically similar and all have toxicity, teratogenicity, and carcinogenicity, and are the main pollutants in the air; when inhaling acetone with a concentration higher than 173 ppm, it will seriously affect the central nervous system and damage important organs of the human body; long-term exposure to ethanol gas will not only cause headaches, liver and kidney function disorders, but also paralyze the central nervous system; low-concentration inhalation of isopropanol will irritate facial organs, and exposure to excessive isopropanol gas may cause symptoms such as dizziness, nausea, and coma, and in severe cases, it may even lead to death. Therefore, it is of great significance to introduce a high-precision method for detecting volatile organic compounds.

[0003] An electronic nose (sensor array) is a bionic device that has the function of mimicking the mammalian olfactory system to distinguish gases. As a device for quickly, efficiently, sensitively, and low-costly identifying the types of target gases and predicting the concentration of target gases, it has been widely used in the detection of toxic and harmful gases. It uses a multi-sensor array to replace the olfactory receptor and a pattern recognition algorithm to replace the biological nerve transmission process. Due to the disadvantage of cross-sensitivity of a single sensor, the electronic nose system uses gas sensors with cross-sensitivity to form a sensor array to obtain gas fingerprint information, and cooperates with a pattern recognition algorithm to weaken the influence of cross-sensitivity to a certain extent, so as to achieve the purpose of identifying the target gas signal in the mixed gas.

[0004] Traditional machine learning algorithms have been widely used in electronic nose gas detection. The convolutional neural network uses convolutional operations to deeply decode gas information, and can integrate the processes of feature extraction, feature processing, and pattern recognition. Existing research has achieved remarkable results in identifying gas types by using CNN. However, the receptive field of CNN is small, the range of information that can be obtained is limited, and it cannot capture the global features in gas information. The research results show that the sensor data recorded by the electronic nose system at different time points can comprehensively reflect the overall characteristics of the gas, which reflects the global correlation of the data. In addition, the response curve of the electronic nose changes rapidly in the dynamic response interval and dynamic recovery interval containing high-frequency information, and changes slowly in the steady-state interval containing low-frequency information. And CNN only tends to capture the high-frequency components of the signal, and thus will capture more high-frequency information in gas information, thereby ignoring the low-frequency information containing a large number of gas steady-state characteristics. Summary of the Invention

[0005] The present invention provides a method for detecting target gases in a mixed volatile organic compound gas, aiming to improve the above problems.

[0006] The present invention is implemented as follows. A method for detecting target gases in a mixed volatile organic compound gas, the method specifically includes the following steps:

[0007] Step (1): Periodically collect one-dimensional gas response intensity data of the gas to be measured through a sensor array, and convert the collected one-dimensional gas response intensity data into a two-dimensional color image;

[0008] Step (2): Input the two-dimensional color image into a gas detection model. The gas detection model detects whether there is a target volatile organic compound gas in the gas to be measured. If the detection result is yes, then execute Step (3);

[0009] Step (3) Extract the steady-state data X from the one-dimensional gas response intensity data w , perform normalization processing on the steady-state data X w to obtain the data X′ w , input the data X′ w into a gas quantification model, and the gas quantification model outputs the concentration of the target volatile organic compound gas.

[0010] Furthermore, the target volatile organic compound gases are acetone, ethanol, and isopropanol.

[0011] Furthermore, the response intensity data within a set time period before reaching the sensor response peak is used as the steady-state data X w .

[0012] Furthermore, the gas detection model consists of a Transformer encoder, a CNN encoder, a convolutional attention fusion module, and a linear mapping layer; among them, the Transformer encoder is used to extract global features in the two-dimensional color image, the CNN encoder is used to extract local features in the two-dimensional color image, the convolutional attention fusion module fuses the global features and local features, and the fused features are output through the linear mapping layer to obtain the detection result.

[0013] Furthermore, the Transformer encoder includes: a first linear layer, a flattening layer, and a normalization processing layer connected in sequence, an average pooling layer and a second linear layer connected to the normalization processing layer, the average pooling layer is connected to the second linear layer, the query Q, key K, and value V output by the second linear layer, the query Q and the key K are multiplied, the attention weight is obtained through Softmax, the attention weight and the value V are multiplied, and then the result is added to the value V after a 3*3 depthwise separable convolution operation through residual addition to obtain the attention output, and the attention output is added to the two-dimensional color image through residual addition.

[0014] Furthermore, the CNN encoder consists of three layers. The first layer consists of a 3×3 convolutional CNN, batch normalization BN, and the activation function h-swish. The second layer is composed of three sub-layers connected in sequence. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and the activation function ReLU. The second sub-layer consists of a 3×3 depthwise separable convolutional layer, batch normalization BN, and the activation function ReLU. The third sub-layer consists of a 1×1 convolutional layer and batch normalization BN. The third layer is composed of three sub-layers connected in sequence. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and the activation function ReLU. The second sub-layer consists of a 5×5 depthwise separable convolutional layer, batch normalization BN, and the activation function ReLU. The third sub-layer consists of a 1×1 convolutional layer and batch normalization BN.

[0015] Furthermore, the convolutional attention fusion module includes: the activation function Sigmoid; the Transformer encoder performs Hadamard dot product with the output of the CNN encoder through the activation function Sigmoid, and the CNN encoder performs Hadamard dot product with the output of the Transformer encoder through the activation function Sigmoid, and the feature maps formed by the two dot products are again subjected to Hadamard dot product.

[0016] Furthermore, the gas quantification model consists of a one-dimensional convolutional layer, a Transformer encoder, and a fully connected layer.

[0017] Furthermore, the conversion process of the one-dimensional gas response intensity data is as follows:

[0018] Normalize the currently collected one-dimensional gas response intensity data; calculate the arccosine value of the normalized one-dimensional gas response intensity data, and for each sensor in the sensor array at each sampling point

[0019] The Transformer encoder uses the multi-head attention mechanism to process the input sequence, dynamically focusing on different parts of the input sequence, enabling the vectors at each position to interact with all positions of the entire input sequence. Therefore, compared with traditional CNNs, it has a stronger ability to capture long-range dependencies, making up for the problem of insufficient receptive fields of CNNs. And precisely because the self-attention mechanism of the Transformer is a global operation, it is more inclined to capture low-frequency information and is not good at capturing high-frequency information. The CNN encoder is used to extract local features in two-dimensional color images, and the convolutional attention fusion module fuses global features with local features. This fusion effect improves the preference of CNNs for low-frequency gas information. Finally, the fused features are output as detection results through the linear mapping layer. Description of the Drawings

[0020] Figure 1 Flow chart of the method for detecting target gas in the mixed volatile organic compound gas provided by the embodiment of the present invention;

[0021] Figure 2 Schematic diagram of the conversion process of one-dimensional gas response intensity data into two-dimensional color image provided by the embodiment of the present invention;

[0022] Figure 3 Schematic diagram of the interception of steady-state data provided by the embodiment of the present invention;

[0023] Figure 4 Schematic diagram of the structure of the gas detection model provided by the embodiment of the present invention;

[0024] Figure 5 Schematic diagram of the structure of the Transformer encoder provided by the embodiment of the present invention;

[0025] Figure 6 Schematic diagram of the structure of the CNN encoder provided by the embodiment of the present invention;

[0026] Figure 7 Schematic diagram of the structure of the convolutional attention fusion module provided by the embodiment of the present invention;

[0027] Figure 8 Schematic diagram of the structure of the gas quantification model provided by the embodiment of the present invention. Detailed implementation manners

[0028] The following further describes in detail the specific implementation manners of the present invention by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art have a more complete, accurate and in-depth understanding of the inventive concept and technical solutions of the present invention.

[0029] Figure 1 Flow chart of the method for detecting target gas in the mixed volatile organic compound gas provided by the embodiment of the present invention. The method is as follows:

[0030] Step (1): Periodically collect one-dimensional gas response intensity data of the gas to be detected through a sensor array, and convert the collected one-dimensional gas response intensity data into a two-dimensional color image;

[0031] Step (2): Input the two-dimensional color image into the gas detection model. The gas detection model detects whether there is a target volatile organic compound gas in the gas to be detected. If the detection result is yes, then execute step (3);

[0032] Step (3): Extract the steady-state data X in the one-dimensional gas response intensity data w , and after normalizing the steady data X w obtain the data X'.w , input the data X' w into a gas quantification model, and the gas quantification model outputs the concentration of the target volatile organic compound gas, where the target volatile organic compound gas is acetone, ethanol, and isopropanol.

[0033] Take the response intensity data within a set time period before reaching the sensor response peak as the steady-state data X w , in the present invention, the response intensity data within 45 seconds before the sensor response peak is intercepted as the stable data, as Figure 3 shown.

[0034] Combine Figure 2 to illustrate the conversion process of the two-dimensional color image. The conversion process of the two-dimensional color image is specifically as follows:

[0035] Since the one-dimensional gas response intensity data usually cannot be used as an effective input for a convolutional neural network, therefore, first convert the collected one-dimensional gas response intensity data into a two-dimensional color image as an effective input for the mixed gas detection model. The conversion method is specifically as follows:

[0036] Use data conversion to convert the gas response intensity data collected each time into a two-dimensional color image. The image is in JPG format, with a size of 470×470, 100 dpi, and 24 bits. The process of converting the one-dimensional gas response intensity data into a two-dimensional image using data conversion is specifically as follows:

[0037] Normalize the one-dimensional gas response intensity data X = {x 1 ,..., x i ,…, x m} of a certain sampling point currently collected, where x i represents the response intensity data currently collected by the i-th sensor, m represents the number of sensors in the sensor array, and the sensor uses a semiconductor metal oxide sensor. Normalize the response intensity data X to the interval [-1, 1] using formula (1). Formula (1) is specifically as follows:

[0038]

[0039] where X is the response intensity data currently collected by the sensor array, max(X) and min(X) respectively represent the maximum and minimum values in the response intensity data currently collected by the sensor array; represents the normalized response intensity data x i .

[0040] Since the response intensity data currently collected by the sensor array does not involve the time axis, there is no need to encode the timestamp as the radius. Instead, the arccosine value of the currently normalized one-dimensional gas response intensity data is directly calculated using Equation (2), which is as follows:

[0041]

[0042] After scaling the and transforming it into the polar coordinate system, the correlation between the response intensity data of different sensors is captured through the triangular summation relationship between each point, resulting in a matrix composed of data, as shown in Equation (3). The matrix is defined by the following equation:

[0043]

[0044] where Φ 1 and Φ n respectively represent the response intensity data after normalization and arccosine of the response intensity data of the first sensor and the nth sensor in the sensor array.

[0045] In the embodiment of the present invention, the gas detection model consists of a Transformer encoder, a CNN encoder, a convolutional attention fusion module, and a linear mapping layer. Among them, the Transformer encoder is used to extract global features in the two-dimensional color image, the CNN encoder is used to extract local features in the two-dimensional color image, the convolutional attention fusion module fuses the global features and local features, and the fused features are output through the linear mapping layer to obtain the detection result, which is the presence or absence of the target organic gas, as Figure 4 shown.

[0046] The two-dimensional color image is input in parallel to the Transformer encoder and the CNN encoder for global and local feature extraction of gas information. Subsequently, the gas information tensors output in parallel interact in the convolutional attention fusion module to compensate each other's receptive fields and alleviate the preference for low-frequency and high-frequency signals of both sides.

[0047] In the embodiment of the present invention, the Transformer encoder includes: a first linear layer, a flattening layer, and a normalization layer connected in sequence, an average pooling layer and a second linear layer connected to the normalization layer, the average pooling layer is connected to the second linear layer, the query Q, key K, and value V output by the second linear layer, the query Q and key K are multiplied and the attention weights are obtained through Softmax, and the attention weights are multiplied by the value V and then added to the value V after a 3*3 depthwise separable convolution operation through residual addition to obtain the attention output. The attention output is added to the two-dimensional color image through residual addition, as Figure 5 shown.

[0048] The two-dimensional color image is linearly mapped into gas tensor information in a high-dimensional space through a first linear layer. The two-dimensional gas tensor is flattened into a series of smaller patch sequences X′ by a flattening layer. After the normalization processing layer normalizes each patch sequence X′, downsampling is performed through an average pooling operation, and then it is mapped into query Q, key K, and value V through a second linear projection. A positional embedding is added to key K, and the attention weights are obtained by calculating the dot product of Q and K and applying Softmax. The attention weights are used to perform weighted summation on the V values to obtain the global correlation of gas features. However, the detailed feature information extracted after the downsampling attention calculation will be lost, affecting the final attention distribution and model performance. Therefore, DW convolution is performed on the feature map corresponding to value V, and independent operations are performed on each channel to effectively extract local features, realizing the refinement and reconstruction of the feature information of V. The reconstructed value V is then added to the output of the initial downsampling attention calculation by residual addition to obtain the final attention output.

[0049] The CNN encoder consists of three layers. The first layer consists of a 3×3 convolutional CNN, batch normalization BN, and an activation function H-swish. The second layer is sequentially connected by three sub-layers. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and an activation function ReLU. The second sub-layer consists of a 3×3 depthwise separable convolutional layer, batch normalization BN, and an activation function ReLU. The third sub-layer consists of a 1×1 convolutional layer and batch normalization BN. The third layer is sequentially connected by three sub-layers. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and an activation function ReLU. The second sub-layer consists of a 5×5 depthwise separable convolutional layer, batch normalization BN, and an activation function ReLU. The third sub-layer consists of a 1×1 convolutional layer and batch normalization BN, as Figure 6 shown;

[0050] In the first layer, a 3×3 convolution is used to capture the local spatial relationships of gas image data. Then, a batch normalization layer (BN) is used to stabilize the network training. Finally, a non-linear activation function, h-swish, is used to enable the model to represent more complex gas sensing relationships. In the second and third layers, depthwise separable convolutions with different receptive fields are used to construct the second and third convolutional blocks. The DW convolution reduces the computational burden while also being able to extract spatial information. This step can effectively capture the spatial distribution characteristics of gas concentration changes. The 3×3 DW convolution is used to capture the local information in the input features. The smaller convolution kernel is suitable for extracting fine gas features, such as small amplitude changes in gas responses. The 5×5 DW convolution can capture a larger range of features and is thus suitable for capturing large-scale features in gas data, enabling a more comprehensive identification of complex patterns in gas data. Before the DW convolution, a pointwise amplification using 1×1 pointwise convolution is performed, which can increase the number of channels in the feature map and enable the model to capture more feature information. This step helps to better describe the feature differences between different gases in the sensor response. After the DW convolution, a 1×1 pointwise convolution is used for further dimensionality reduction, which can reduce the number of channels and computational complexity, while retaining the key information in the sensor response and reducing redundant features.

[0051] The convolutional attention fusion module includes: the activation function Sigmoid; the Transformer encoder performs a Hadamard dot product with the output of the CNN encoder through the activation function Sigmoid, and the CNN encoder performs a Hadamard dot product with the output of the Transformer encoder through the activation function Sigmoid. The feature maps formed by the two dot products are again subjected to a Hadamard dot product, as Figure 7 shown.

[0052] The output of the CNN encoder is transformed into a weight that affects the output of the Transformer encoder through the sigmoid activation function, which means the degree of influence of the gas information in the local area on the global gas distribution. A sudden change in the local gas response will trigger a significant change in the global features. Similarly, the output of the Transformer encoder is also transformed into a weight to correct the output result of the CNN encoder through the sigmoid function, achieving the effect of the global influencing the local; in this way, the low-frequency information and the high-frequency information are compensated for each other by the weights generated by the sigmoid function. Finally, the local features and the global features are further fused through the Hadamard product to obtain a comprehensive feature, taking into account both the local and global gas concentration information. This fusion helps the model to identify local anomalies and capture the overall trend.

[0053] The gas samples containing different target organic gases are divided into training samples and test samples. The gas detection model is trained based on the training samples and tested based on the test samples. When the prediction accuracy of the gas detection model reaches the set standard, the training stops.

[0054] In the embodiment of the present invention, the gas quantification model is composed of a one-dimensional convolutional layer, a Transformer encoder, and a fully connected layer. The mixed gas data first encodes the gas information through a one-dimensional convolutional CNN, and then extracts the long-range dependence of the gas information through the Transformer encoder, as Figure 8 shown.

[0055] The invention uses the CGS-8 intelligent gas sensing analysis system provided by Beijing Elite Technology Co., Ltd. for data collection. Before the gas sensing experiment, the position, type, and optimal working current of the sensor should be considered first. During the establishment of the sensor array, different sensors will cause negative impacts such as cross-sensitivity and mutual interference in data collection. Therefore, researchers need to develop appropriate sensor combinations and optimal working currents through multiple groups of experiments. In the experiment, a sensor array composed of 8 commercial semiconductor metal oxide sensors was screened. According to the requirements of different types of sensor manuals, the optimal working current was adjusted to make the sensor response relatively stable. Among them, TGS2600 and TGS822 are produced by Figaro Engineering Inc., and WSP2110 and MQ3 are produced by Winsen Electronics Co., Ltd. The sensor models and the optimal working currents of the sensors are shown in Table 1.

[0056] Table 1 Sensor models, target gases, detection ranges, and optimal working currents

[0057]

[0058] After setting the heating current, the sensor starts to preheat for 1 hour. Once the baseline is stable, different volumes of liquid are collected using a flat-tip sampler. At the same time, the switches of the liquid evaporator and the fan are turned on. The fan is used to accelerate the diffusion of the gas. Wait for 3 to 5 minutes until the sensor response intensity reaches the stable stage, then open the gas chamber and flush the gas chamber with air until the sensor response intensity returns to the baseline level. In a single experiment, the concentration of each gas is set between 1 and 20 ppm, and the response and recovery curves of the original resistance and the calculated response values are automatically recorded in a CSV file. All gas sensitivity tests are carried out at one atmosphere, with a relative humidity of (50 ± 10)%, an ambient temperature of (26 ± 2)°C, a sampling interval of 1 s, and the response of the sensor is defined as R a / R g where R a and Rg They are the resistance values of the sensor in air and in the target gas, respectively.

[0059] The static gas distribution method was used. A flat-tip sampler was used to extract 95% ethanol, 98% acetone, and 99% isopropanol, respectively, with an extraction volume of 10 μL. Since high-purity reagents were used directly, according to formula (4), the volume of the extracted liquid would be very small, making it difficult to control and inject into the evaporating dish. Therefore, acetone, ethanol, and isopropanol diluted to 10% were used in the sensing experiment. The expressions are as follows:

[0060]

[0061] Among them, Q is the volume of the liquid to be measured (mL), V is the volume of the gas chamber (mL), C is the concentration of the gas to be prepared (ppm), M is the molecular weight of the substance (g / mol), d is the concentration of the liquid to be measured (%), and r is the liquid density (g / cm 3 ), T R is the laboratory ambient temperature (°C), and T B is the gas chamber temperature (°C).

[0062] The present invention adopts the above-mentioned static gas distribution method to form a gas sample containing target volatile organic compound gases with different concentrations. First, target volatile organic compound liquids with different concentrations are configured, and then a variety of target volatile organic compound liquids with different concentrations are mixed and evaporated to form a mixed gas of a variety of target volatile organic compound gases with different concentration ratios as the sample gas. The gas sample is divided into a training sample and a test sample. The gas detection model is trained based on the training sample, and the gas detection model is tested based on the test sample. When the prediction accuracy of the gas detection model reaches the set standard, the training is stopped.

[0063] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited by the above methods. As long as various non-substantive improvements are made by adopting the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for detecting a target gas in a mixed volatile organic compound gas, characterized in that: The method specifically comprises the following steps: Step (1): periodically collecting one-dimensional gas response intensity data of the gas to be measured through a sensor array, and converting the collected one-dimensional gas response intensity data into a two-dimensional color image; Step (2): input the two-dimensional color image into the gas detection model, and the gas detection model detects whether the target volatile organic compound gas exists in the gas to be tested. If the detection result is yes, execute step (3); Step (3) extracting the steady-state data X from the one-dimensional gas response intensity data w , for stable data X w After normalization, we get the data X′ w , the data X′ w The gas quantitative model is input, and the gas quantitative model outputs the concentration of the target volatile organic compound gas.

2. The method for detecting target gas in mixed volatile organic gas according to claim 1, characterized in that: The target VOC gases are acetone, ethanol and isopropanol.

3. The method for detecting target gas in mixed volatile organic gas according to claim 1, characterized in that: The response intensity data within the set time before reaching the sensor response peak is taken as the steady-state data X w .

4. The method for detecting target gas in mixed volatile organic gas according to claim 1, characterized in that: The gas detection model consists of a Transformer encoder, a CNN encoder, a convolutional attention fusion module and a linear mapping layer; among them, the Transformer encoder is used to extract global features in two-dimensional color images, the CNN encoder is used to extract local features in two-dimensional color images, and the convolutional attention fusion module fuses global features with local features, and outputs the detection results through the linear mapping layer.

5. The method for detecting target gas in mixed volatile organic gas according to claim 4, characterized in that: The Transformer encoder includes: a first linear layer, a flattening layer, and a normalization processing layer connected in sequence, an average pooling layer and a second linear layer connected to the normalization processing layer, the average pooling layer is connected to the second linear layer, the query Q output by the second linear layer is output with the key K and the value V, the query Q and the key K are dot-multiplied by Softmax to obtain the attention weight, the attention weight and the value V are dot-multiplied, and the residual is added to the value V after the 3*3 depth-separable convolution operation to obtain the attention output, and the attention output is residually added with the two-dimensional color image.

6. The method for detecting target gas in mixed volatile organic gas according to claim 4, characterized in that: The CNN encoder consists of three layers. The first layer consists of a 3×3 convolutional CNN, batch normalization BN, and activation function h-swish; the second layer consists of three sub-layers connected in sequence. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and activation function ReLU, the second sub-layer consists of a 3×3 depth-separable convolutional layer, batch normalization BN, and activation function ReLU, and the third sub-layer consists of a 1×1 convolutional layer and batch normalization BN; the third layer consists of three sub-layers connected in sequence. The first sub-layer consists of a 1×1 convolutional layer, batch normalization BN, and activation function ReLU, the second sub-layer consists of a 5×5 depth-separable convolutional layer, batch normalization BN, and activation function ReLU, and the third sub-layer consists of a 1×1 convolutional layer and batch normalization BN.

7. The method for detecting target gas in mixed volatile organic gas according to claim 4, characterized in that: The convolutional attention fusion module includes: activation function Sigmoid; the Transformer encoder performs Hadamard point multiplication with the output of the CNN encoder through the activation function Sigmoid, and the CNN encoder performs Hadamard point multiplication with the output of the Transformer encoder through the activation function Sigmoid. The feature map formed by the two point multiplications is Hadamard point multiplication again.

8. The method for detecting target gas in mixed volatile organic gas according to claim 1, characterized in that: The gas quantification model consists of a one-dimensional convolutional layer, a Transformer encoder and a fully connected layer.

9. The method for detecting target gas in mixed volatile organic gas according to claim 1, characterized in that: The conversion process of one-dimensional gas response intensity data is as follows: The currently collected one-dimensional gas response intensity data is normalized; the arccosine value of the normalized one-dimensional gas response intensity data is calculated, and the correlation between the response intensity data collected by different sensors is captured by calculating the triangular sum relationship between the response intensity data of each sensor in the sensor array at each sampling point.

Images