Convolutional neural network regression model flame equivalence ratio detection method and device based on attention mechanism

Abstract

The invention discloses a flame equivalence ratio detection method based on an attention mechanism in the field of image processing and machine learning, and the method comprises the steps: firstly constructing a flame equivalence ratio detection model based on the attention mechanism, and achieving the end-to-end detection of a flame equivalence ratio; according to the model, through automatic feature extraction and regression prediction functions, a tedious manual feature extraction step in a traditional method is avoided, so that the measurement efficiency and accuracy are improved. The equivalence ratio detection model constructed by using deep learning can automatically extract features from the color camera RGB-2D image and carry out regression prediction, so that a complex manual feature extraction project is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of image processing and machine learning, and specifically relates to a flame equivalence ratio detection method and device based on a convolutional neural network regression model of an attention mechanism. Background Art

[0002] Methane-dominated natural gas is widely used in many fields such as power generation, heating and transportation due to its high efficiency and low emissions. In specific equipment such as gas turbines and industrial furnaces, accurate combustion control is essential to improve efficiency and reduce emissions. This increases the demand for real-time and accurate detection of flame combustion status, especially flame equivalence ratio.

[0003] Common non-invasive detection techniques for flame equivalence ratio include spectral analysis and color modeling. Spectral analysis indirectly obtains the flame equivalence ratio by measuring the chemiluminescence ratio of specific free radicals through optical fibers, but this method requires expensive optical equipment and is limited by on-site layout. In contrast, using an RGB color camera to measure the pixel intensity ratio (IB / IG) of the blue and green channels (B and G) in the image to characterize the equivalence ratio is not only cost-effective, but also applicable to environments with limited optical access, providing an effective alternative.

[0004] However, traditional color modeling methods require tedious preprocessing when processing flame images, such as denoising and adjusting parameters to eliminate the red and yellow background radiation under rich combustion conditions. These steps rely on manual feature extraction, which is inefficient and lacks flexibility. Summary of the invention

[0005] The present invention aims to provide a flame equivalence ratio detection method and device based on a convolutional neural network regression model of an attention mechanism, so as to solve the problems in the prior art of processing flame images, such as cumbersome preprocessing process, low efficiency and lack of flexibility in manual feature extraction.

[0006] The first aspect of the present application provides a flame equivalence ratio detection method based on a convolutional neural network regression model of an attention mechanism, comprising the following steps:

[0007] S1. Build an image dataset: Establish a flame RGB-2D image dataset containing different equivalence ratio combustion states. Each entry in the dataset includes an image path and a corresponding equivalence ratio label.

[0008] S2. Dividing a training set and a test set: dividing the equivalence ratio combustion state into a training set and a test set, wherein the test set contains data of lean combustion and rich combustion;

[0009] S3, data loading and preprocessing: build the model to load the image and perform necessary preprocessing, including resizing and normalization;

[0010] S4. Configure feature extraction network: Design a feature extraction network consisting of convolutional layers, activation functions, and pooling layers, specifically for extracting visual features from images;

[0011] S5. Build an attention mechanism layer: The attention mechanism layer is used to analyze and weigh the extracted feature tensors, highlighting the visual features closely related to the combustion state;

[0012] S6. Construct a fully connected network: Use the weight bias introduced by the neurons in the fully connected layer to fit the regression relationship between the extracted features and the equivalence ratio, and output the predicted value of the flame equivalence ratio;

[0013] S7, model training parameter optimization: use the training set data to train and optimize the model parameters, define the model's loss function and evaluation criteria, and use the Adam optimizer to train the parameters to obtain the optimal detection model;

[0014] S8, generalization performance test, input the test set data after image loading preprocessing into the trained test model framework through the callback mechanism to obtain the detection result output by the detection model.

[0015] The present invention develops a flame equivalence ratio detection model based on the attention mechanism to achieve end-to-end detection of flame equivalence ratio. The model avoids the cumbersome manual feature extraction steps in traditional methods through automated feature extraction and regression prediction functions, thereby improving the efficiency and accuracy of measurement.

[0016] The equivalence ratio detection model built using deep learning can automatically extract features from color camera RGB-2D images and perform regression prediction, avoiding complex manual feature extraction engineering.

[0017] Furthermore, in S1, a CCD or CMOS image sensor is used to collect the flame RGB-2D flame image.

[0018] Furthermore, in S2, an equivalence ratio between 0.73 and 1.05 is a lean fuel, and an equivalence ratio between 1.18 and 1.47 is a rich fuel.

[0019] Further, configuring the feature extraction network in S4 includes the following steps:

[0020] S4-1, Convolution layer: The convolution operation is performed by performing the cross-correlation calculation of formula (1) between the input image and the convolution kernel:

[0021]

[0022] S4-2, Activation function: ReLU in formula (2) is selected as the activation function;

[0023] f(x)=max(0,x) (2);

[0024] S4-3, pooling layer: The maximum pooling of formula (3) is used to represent the area by selecting the maximum value in the area, highlighting the main features of the image and being tolerant to small changes in the feature position;

[0025]

[0026] Furthermore, in S4-1, the model construction sets 128 convolution kernels, the convolution kernel size is set to 3x3, and the sliding step is 1; in S4-3, the pooling matrix S is set to 2x2, and the sliding step is 2.

[0027] Furthermore, the construction of the attention mechanism layer in S5 includes the following steps:

[0028] S5-1. Feature processing and score calculation: Receive the multi-dimensional feature tensor output from the feature extraction network after convolution, activation, and pooling. This tensor is calculated through two branches to generate the attention score:

[0029] The first branch: Apply the tanh activation function to perform nonlinear transformation on the input tensor, and combine the weight W0 and bias B0 to calculate the first set of attention scores Score1;

[0030] The second branch: Calculate the row-wise feature deviation matrix of the input tensor. The specific operation is to calculate the deviation Diff between the pixel value at each position in the input tensor and its row average value. The deviation is processed by the tanh function, weight W2 and bias B2 to obtain the second set of attention scores Score2;

[0031] S5-2, score fusion and weight normalization: Score1 and Score2 are added and multiplied with the introduced weight tensor V, and normalized using the softmax function to generate the final attention weight tensor Weight;

[0032] S5-3, output generation: Multiply the calculated attention weight tensor Weight with a mask Mask designed with a multi-channel attention principle to obtain the final output Output.

[0033] Furthermore, in S7, the root mean square error MSE is used as the loss function, and the Adam optimizer calculates the gradient of each parameter based on the loss function, and uses the back-propagation algorithm to make dynamic adjustments based on the differences between the outputs;

[0034]

[0035] Furthermore, the determination coefficient R is used in S7 2 And mean absolute error MAE as model evaluation parameters:

[0036]

[0037] Where y represents the estimated value of the data, represents the true value, represents the average value, and N is the total number of elements of the calculation variable.

[0038] The second aspect of the present application provides a flame equivalence ratio measuring device, which is used to implement any one of the methods described in the first aspect, and includes the following units:

[0039] Image acquisition unit: using one or more RGB color image sensors with high-resolution and high-speed video capture functions to capture the rapid dynamic changes of flames and obtain flame RGB-2D images;

[0040] Data storage unit: used to store the flame RGB-2D images captured by the image acquisition unit and build a data set, and configure network attached storage NAS or other external storage solutions to achieve long-term storage and backup of data;

[0041] Program processing unit: used for training detection models, including multi-core processors and graphics processors;

[0042] Software and framework selection: Use Python or C++ programming language, combined with TensorFlow, PyTorch, Keras deep learning framework to achieve flexible and efficient program coding.

[0043] Furthermore, the image sensor is selected from CMOS or CCD technology; the data storage unit preferably selects SSD with fast data access speed and low response time. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Flowchart for constructing a convolutional neural network flame equivalence ratio detection model based on the attention mechanism in an embodiment of the present invention

[0045] Figure 2 A workflow diagram for constructing a model attention mechanism for an embodiment of the present invention;

[0046] Figure 3 This is a MAE (mean absolute error) histogram of the equivalent ratio test of 5 test sets using the grouped K-Fold cross-validation method in the model generalization ability test in the embodiment of the present invention;

[0047] Figure 4 It is a structural framework diagram of a convolutional neural network flame equivalent ratio measurement application device based on an attention mechanism, which is a preferred embodiment provided by the present invention. DETAILED DESCRIPTION

[0048] The following is further described in detail through specific implementation methods:

[0049] Example 1: A flame equivalence ratio detection model based on the attention mechanism. The model is constructed using the TensorFlow and Keras frameworks. Figure 1 The model building flow chart shown further describes the technical details of the embodiment of the present invention:

[0050] Step 1: Image dataset construction: Collect flame RGB-2D images of different equivalence ratio combustion states and establish a dataset containing image paths and corresponding equivalence ratio labels.

[0051] In step 1, flame images of different equivalence ratio combustion state data are collected, premixed flame burners are deployed, fuel and air flow rates are adjusted to configure different equivalence ratio combustion conditions, and CCD or CMOS image sensors are used to collect flame RGB-2D flame images. The flame images are saved in a specific folder on the computer, and a data set is constructed containing the image sequence path and the corresponding regression value label Φ. In this example, flame images of 10 equivalence ratio combustion states are collected, and 100 RGB flame images are collected for each condition.

[0052] Step 2: Divide the training set and test set: In order to analyze the performance of the model under different combustion conditions, a grouped K-Fold cross validation was implemented, and the data set was divided into two main categories according to the equivalence ratio: one category contains data with equivalence ratios of 0.73, 0.79, 0.86, 0.94 and 1.05, and the other category contains data with equivalence ratios of 1.18, 1.26, 1.35, 1.43 and 1.47. Through five rounds of validation, data from these two categories were selected as test sets in each round to ensure that each group contains lean and rich flame images, so as to comprehensively evaluate the prediction performance of the model under different combustion conditions. The specific grouping strategy is detailed in Table 1.

[0053] Table 1 Grouping table of K-Fold cross validation

[0054]

[0055] Step 3: Data loading and preprocessing, load the image from the constructed dataset, perform necessary preprocessing, including resizing and normalization, and step 2 processing is designed to ensure that the model can effectively process the input image. For the collected flame image, the processing includes the following sub-steps:

[0056] 3-1. Resizing: The input network model image needs to have consistent dimensions. All collected flame images are resized to the expected input size of 256x256 pixels.

[0057] 3-2. Normalization processing: The collected RGB-2D color images are normalized by channel, and the integer pixel values ​​from 0 to 255 of each color channel (red, green, and blue) are independently converted to floating-point numbers between 0 and 1 to improve the stability and convergence speed of model training.

[0058] Step 4: Feature extraction network configuration, design a feature extraction network consisting of convolutional layers, activation functions, and pooling layers, specifically for extracting visual features from images. The steps are listed as follows:

[0059] 4-1 Convolution: The convolution operation is performed by performing the cross-correlation calculation of Equation 1 between the input image and the convolution kernel. This process first performs an element-by-element multiplication operation between the convolution kernel and the local area of ​​the image (a matrix matching the size of the convolution kernel) in each red, green, and blue channel of the input image. The convolution kernel slides across the entire image domain according to the set step size to ensure full coverage.

[0060] Finally, the calculation results of each channel are added together to form a feature map containing the color information of the RGB image.

[0061]

[0062] Preferably, the model is constructed with 128 convolution kernels, the convolution kernel K size is set to 3x3, and the sliding step size is 1.

[0063] 4-2. Activation: The convolution operation itself has linear characteristics. In order to enable the network to parse complex feature maps, the activation function is introduced into the network to add nonlinear characteristics.

[0064] f(x)=max(0,x) (2)

[0065] Preferably, ReLU (Rectified Linear Unit) in Formula 2 is selected as the activation function to alleviate the gradient vanishing problem. Although the insensitivity of the ReLU function to negative values ​​may be considered a disadvantage, this feature becomes an advantage when analyzing flame images. By setting negative activation values ​​to zero, ReLU effectively ignores the darker pixel values ​​in the image.

[0066] 4-3 Max Pooling: The dimension of the feature map matrix generated after activation is still high. Therefore, the pooling step is used to downsample and reduce the dimension of the feature map.

[0067]

[0068] Preferably, the maximum pooling of Formula 3 is selected to represent the region by selecting the maximum value in the region, highlighting the main features of the image and maintaining tolerance for small changes in the feature position. The pooling matrix S is set to 2x2, and the sliding step size is 2.

[0069] Step 5: Construction of the attention mechanism layer. After the convolution, activation, and pooling processing in step 3, the visual feature information of the flame state of the RGB-2D image is effectively extracted. The attention mechanism layer is designed to analyze and weigh the extracted feature tensors, dynamically adjust the network focus according to the differences and importance of the features, highlight the visual features closely related to the combustion state, and effectively exclude background noise and irrelevant information.

[0070] Preferably, the attention mechanism workflow is constructed as follows Figure 2 ,This attention mechanism involves numerical calculation of multiple multi-dimensional tensors. The specific tensor dimensions are shown in Table 2, including tensor dot multiplication, cross multiplication, activation processing, normalization processing, etc.

[0071] Table 2 Tensors and their dimensions in the attention mechanism

[0072]

[0073] 5-1: The input tensor received by the attention mechanism is a 127x127x128 tensor Inputs obtained after processing by the feature extraction network. Next, this input tensor is used to calculate the attention score through two branches.

[0074] 5-2: The first branch generates a set of attention scores (Score1) by applying a nonlinear transformation using the tanh function and processing the input tensor with weights W0 and bias B0.

[0075] 5-3: The second branch calculates the row-wise feature deviation matrix of the input tensor. The specific operation includes calculating the absolute value deviation between the pixel value at each position in the input tensor and the average value (Mean) of the pixel value in its row, and generating a deviation tensor (Diff). This deviation tensor is transformed by the tanh function through the nonlinear transformation of the weight W2 and the bias B2 to generate the second set of attention scores (Score2).

[0076] 5-4: By adding Score1 and Score2, multiplying it with the introduced weight tensor V, and normalizing it using the softmax function, we get the attention weight tensor Weight.

[0077] 5-5: Considering the possible random noise in the image, multiply the Weight by a mask to generate the final output. This mask is designed according to the principle of multi-channel attention mechanism. By performing maximum pooling on the multi-channel corresponding pixels of the input tensor Inputs, the highest weight is assigned to the most important pixels, effectively eliminating background noise and information irrelevant to the prediction target, and optimizing the use of computing resources.

[0078] Step 6: Fully connected network construction, using the weight bias introduced by the fully connected layer neurons, the model fits the regression relationship between the extracted features and the equivalent ratio, and outputs the flame equivalent ratio prediction value. This stage mainly includes two key steps, flattening and full connection, as described below:

[0079] 6-1 Flattening: This step involves converting the multi-dimensional feature tensor Output into a one-dimensional array. The original spatial feature information is retained while simplifying the data structure.

[0080] 6-2 Full connection: After flattening, the one-dimensional data is introduced into the neurons with weights (W) and biases (B) to construct the mapping relationship between the extracted features and the equivalent ratio.

[0081] Preferably, in this example, a single neuron can construct a regression relationship between an equivalent ratio label Φ and a one-dimensional array X converted from a feature tensor Output.

[0082] Φ=W*X+B (4)

[0083] Step 7: Construct model parameter optimization, input the model training set, define the model loss function, use the Adam optimizer for parameter training, and use the model evaluation parameters for effect analysis. Specifically, the convolution kernel parameters of the feature extraction network, the weights and biases introduced in the attention mechanism, and the weights and bias parameters introduced by neurons in the regression prediction stage jointly affect the output quality of the model, and need to be continuously optimized throughout the training cycle. The steps are described as follows.

[0084] 7-1: Preferably, the root mean square error (MSE) is selected as the loss function. It is defined as follows:

[0085]

[0086] Specifically, the root mean square error (MSE) provides a stable error measure that is highly sensitive to large errors by quantifying the squared difference between the actual output and the predicted output of Φ. The differentiability and continuous derivatives of MSE are crucial to maintaining model stability in the calculation of partial derivative gradients.

[0087] 7-2: The Adam optimizer calculates the gradient of each parameter based on the loss function and uses the back-propagation algorithm to dynamically adjust according to the difference between the outputs. The operation of the Adam optimization algorithm can be described in detail as follows:

[0088] 1. Compute the first-order moment estimate (mean) and the second-order moment estimate (variance) of the gradient: This is done by taking a weighted average of the gradients, where the weights are given by the decay factors β1 and β2.

[0089] First moment estimate:

[0090] m t =β 1 m t-1 +(1-β 1 )·g t (6)

[0091] Second moment estimate:

[0092] v t =β 2 v t-1 +(1-β 2 )·g t 2 (7)

[0093] 2. Perform bias correction on the first-order moment and second-order moment estimates: In the early stages of training, mt and vt may be underestimated due to initialization to zero, and bias correction is required:

[0094] Bias-corrected first moment estimate:

[0095]

[0096] Bias-corrected second moment estimate:

[0097]

[0098] 3. Update parameters using bias-corrected first- and second-order moment estimates: Parameter update is performed using the bias-corrected estimates using the following formula:

[0099]

[0100] Specifically, η is the learning rate, and ε is a very small value added to avoid division by zero.

[0101] 7-3: Preferably, small batch training is adopted in this example, and the training is divided into 10 batches, and 32 samples are processed in each batch.

[0102] 7-4: Using the coefficient of determination (R 2 ) and mean absolute error (MAE) as model evaluation parameters. Specifically, a higher R 2 A high MAE value indicates that the model regression fit is good and can explain the data variability more accurately. A low MAE value indicates the accuracy of the prediction.

[0103]

[0104] Formulas 11 and 12 give R 2 , MAE, and parameter variance calculation formula, where y represents the data estimate, represents the true value, represents the average value, and N is the total number of elements of the calculation variable.

[0105] Preferably, the model R in all training sets of the embodiment of the present invention 2 The mean is 0.9958, the MAE mean is 0.0122, and the model training effect is good.

[0106] Example 2: A method for detecting the flame equivalence ratio using the model of Example 1.

[0107] In the generalization performance test, the test set data is input into the model framework trained in Example 1 through a callback mechanism, and the mean absolute error (MAE) is used as an evaluation indicator to test the detection performance of the model for flame images of unknown equivalence ratio. Figure 3 The model prediction results obtained by K-Fold cross validation using 5 groups are shown. In the figure, G1, G2, G3, G4, and G5 represent the MAE values ​​of the test data with different equivalent ratios under each group, and the straight line graph with inverted triangle marks shows the maximum difference (Gap) between each equivalent ratio and its adjacent equivalent ratio.

[0108] The results show that the proposed model can effectively detect flame images of unknown equivalence ratios and accurately distinguish between lean and rich combustion states. The model has a small measurement error on each equivalence ratio flame image and can clearly distinguish adjacent equivalence ratio values ​​in the data set, showing good detection ability.

[0109] Example 3 details the main components of the measuring device for realizing the equivalent ratio detection model in Example 1. It should be noted that this example does not cover all assembly contents. Figure 4 The main system units are described in:

[0110] 1. Image acquisition unit: This unit uses a high-resolution image sensor as its core, and its high-speed video capture function can cope with the rapid changes in flame dynamics.

[0111] Preferably, in this embodiment, a Photron SA4 high-speed color CMOS camera is selected, which is suitable for capturing rapidly changing scenes because it can provide a high frame rate.

[0112] 2. Data storage unit: to store the data set containing ten types of equivalent ratio flame images and save the parameters in the model building process.

[0113] In this embodiment, SSD is preferably selected as the data storage unit. SSD has fast data access speed and short response time, which meets the requirements of real-time data processing.

[0114] 3. Program processing unit: Model training relies on a program processing unit equipped with high-performance processors and graphics processors to support complex data processing tasks.

[0115] In this embodiment, a GPU image processor is preferably used, and a multi-core mode is adopted. The processor is specifically used to execute all the processes and steps of the above-mentioned equivalent ratio measurement method based on deep learning. The specific content is recorded in Example 1, and this application will not repeat it again.

[0116] 4. Software and framework selection: In this embodiment, Python is selected as the programming language, and TensorFlow and Keras frameworks are used to build the equivalent ratio model code. These two frameworks provide the required support and flexibility for the model.

[0117] In this application, Figure 4 Only a schematic layout of the system units is shown, and not all possible electronic components are shown in detail. The embodiments covered by this patent application include methods, systems or computer program products. Therefore, the implementation methods of this patent combine hardware and software. The above embodiments are exemplary descriptions of the present invention and are not limitations on the scope of the invention. Under the premise of following the basic spirit and principles of the present invention, technicians in this field can make various reasonable modifications, changes and improvements. All these changes should be deemed to belong to the scope of protection of the present invention and are included in the claims of this patent.

Claims

1. A flame equivalence ratio detection method based on a convolutional neural network regression model with an attention mechanism, characterized in that The following steps are involved: S1. Build an image dataset: Establish a flame RGB-2D image dataset containing different equivalence ratio combustion states. Each entry in the dataset includes an image path and a corresponding equivalence ratio label. S2. Dividing a training set and a test set: dividing the equivalence ratio combustion state into a training set and a test set, wherein the test set contains data of lean combustion and rich combustion; S3, data loading and preprocessing: build the model to load the image and perform necessary preprocessing, including resizing and normalization; S4. Configure feature extraction network: Design a feature extraction network consisting of convolutional layers, activation functions, and pooling layers, specifically for extracting visual features from images; S5. Build an attention mechanism layer: The attention mechanism layer is used to analyze and weigh the extracted feature tensors, highlighting the visual features closely related to the combustion state; S6. Construct a fully connected network: Use the weight bias introduced by the neurons in the fully connected layer to fit the regression relationship between the extracted features and the equivalence ratio, and output the predicted value of the flame equivalence ratio; S7, model training parameter optimization: use the training set data to train and optimize the model parameters, define the model's loss function and evaluation criteria, and use the Adam optimizer to train the parameters to obtain the optimal detection model; S8, generalization performance test, input the test set data after image loading preprocessing into the trained test model framework through the callback mechanism to obtain the detection result output by the detection model.

2. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 1 is characterized in that: In the S1, a CCD or CMOS image sensor is used to collect the flame RGB-2D flame image.

3. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 2 is characterized in that: In the S2, an equivalence ratio between 0.73 and 1.05 is a lean fuel, and an equivalence ratio between 1.18 and 1.47 is a rich fuel.

4. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 3 is characterized in that: Configuring the feature extraction network in S4 includes the following steps: S4-1, Convolution layer: The convolution operation is performed by performing the cross-correlation calculation of formula (1) between the input image and the convolution kernel: S4-2, Activation function: ReLU in formula (2) is selected as the activation function; f(x)=max(0,x) (2); S4-3, pooling layer: The maximum pooling of formula (3) is used to represent the area by selecting the maximum value in the area, highlighting the main features of the image and being tolerant to small changes in the feature position; 5. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 4 is characterized in that: In S4-1, the model construction sets 128 convolution kernels, the convolution kernel size is set to 3x3, and the sliding step is 1; in S4-3, the pooling matrix S is set to 2x2, and the sliding step is 2.

6. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 5 is characterized in that: The construction of the attention mechanism layer in S5 includes the following steps: S5-1. Feature processing and score calculation: Receive the multi-dimensional feature tensor output from the feature extraction network after convolution, activation, and pooling. This tensor is calculated through two branches to generate the attention score: The first branch: Apply the tanh activation function to perform nonlinear transformation on the input tensor, and combine the weight W0 and bias B0 to calculate the first set of attention scores Score1; The second branch: Calculate the row-wise feature deviation matrix of the input tensor. The specific operation is to calculate the deviation Diff between the pixel value at each position in the input tensor and its row average value. The deviation is processed by the tanh function, weight W2 and bias B2 to obtain the second set of attention scores Score2; S5-2, score fusion and weight normalization: Score1 and Score2 are added and multiplied with the introduced weight tensor V, and normalized using the softmax function to generate the final attention weight tensor Weight; S5-3, output generation: Multiply the calculated attention weight tensor WeigScoreht with a mask Mask designed with a multi-channel attention principle to obtain the final output Output.

7. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 6 is characterized in that: In S7, the root mean square error (MSE) is used as the loss function. The Adam optimizer calculates the gradient of each parameter based on the loss function and uses the back propagation algorithm to dynamically adjust according to the difference between the outputs.

8. The flame equivalence ratio detection method based on the convolutional neural network regression model of the attention mechanism according to claim 7 is characterized in that: The coefficient of determination R is used in S7 2 And mean absolute error MAE as model evaluation parameters: Where y represents the estimated value of the data, represents the true value, represents the average value, and N is the total number of elements of the calculation variable.

9. A flame equivalence ratio detection device, used to implement the method according to any one of claims 1 to 8, characterized in that The following units are included: Image acquisition unit: using one or more RGB color image sensors with high-resolution and high-speed video capture functions to capture the rapid dynamic changes of flames and obtain flame RGB-2D images; Data storage unit: used to store the flame RGB-2D images captured by the image acquisition unit and build a data set, and configure network attached storage NAS or other external storage solutions to achieve long-term storage and backup of data; Program processing unit: used for training detection models, including multi-core processors and graphics processors; Software and framework selection: Use Python or C++ programming language, combined with TensorFlow, PyTorch, Keras deep learning framework to achieve flexible and efficient program coding.

10. The flame equivalence ratio detection device according to claim 9, characterized in that: The image sensor is selected from CMOS or CCD technology; the data storage unit preferably selects SSD with fast data access speed and low response time.

Images