Light-weight INCN flame combustion state recognition method

Abstract

The invention relates to the technical field of image processing, in particular to a lightweight INCN flame combustion state recognition method, which comprises an image preprocessing module, a feature extraction network and a classification head, and is characterized in that in the feature extraction network of an InceptionNeXt model, Demo block in StarNet is used for replacing original MetaNeXtBlock; designing a three-stage feature extraction structure; the feature extraction network comprises a first down-sampling unit, a first satellite operation unit, a second satellite operation unit, a third satellite operation unit, a second down-sampling unit, a first normalization layer, a fourth satellite operation unit, a fifth satellite operation unit, a sixth satellite operation unit, a third down-sampling unit, a second normalization layer, a seventh satellite operation unit, an eighth satellite operation unit and a ninth satellite operation unit which are cascaded. According to the method, the problems that the weight of the InceptionNeXt model is low and the identification accuracy is basically not influenced are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a flame combustion state recognition method of a lightweight INCN. Background Art

[0002] Countercurrent burner is a combustion equipment commonly used in the industrial and energy fields. Accurately identifying the combustion state in the countercurrent burner can not only optimize the combustion process and improve combustion efficiency, but also reduce the emission of harmful gases and reduce energy waste.

[0003] The InceptionNeXt model is often used for flame recognition and classification, and can effectively identify flame states. However, the InceptionNext model has many parameters and a large amount of computation, which results in high requirements for hardware resources during actual deployment, increasing hardware costs. In addition, due to the increase in computing costs, it cannot meet real-time requirements in actual industrial applications.

[0004] The application number is 202411777558.0, which is based on the improved InceptionNeXt combustion state recognition method. By introducing the TripletAttention mechanism, adding the SimC feature focusing module after the Inception deep convolution, embedding the multi-scale void parallel fusion module after the feature extraction part of the model, and adding the ShuffleCS module based on channel shuffling to the classification part, these improvements effectively improve the recognition accuracy of the InceptionNeXt model. However, this method does not consider the lightweight design of the InceptionNeXt model and cannot meet the application requirements under the condition of limited hardware resources. How to achieve the lightweight design of the InceptionNeXt model without affecting the recognition accuracy is an urgent problem to be solved. Summary of the invention

[0005] In view of the shortcomings of the existing methods, the present invention solves the problem of lightweight InceptionNeXt model without affecting the recognition accuracy.

[0006] The technical solution adopted by the present invention is: a method for identifying the flame combustion state of a lightweight INCN comprises the following steps:

[0007] Step 1, collecting flame images of the combustion state of the counter-flow burner;

[0008] As a preferred embodiment of the present invention, the combustion states include: high oxygen combustion, nitrogen diluted high oxygen combustion, oxygen-rich stable combustion, high nitrogen diluted medium combustion, balanced combustion and low oxygen and low methane combustion.

[0009] As a preferred implementation of the present invention, noise enhancement is performed on the flame image.

[0010] Step 2: Build an improved InceptionNeNt network, including: image preprocessing module, feature extraction network and classification head. In the feature extraction network of the InceptionNeXt model, use the Demo block in StarNet to replace the original MetaNeXtBlock; and design a three-stage feature extraction structure;

[0011] As a preferred implementation of the present invention, the image preprocessing module includes: a 4x4 convolution layer and a normalization layer.

[0012] As a preferred embodiment of the present invention, the feature extraction network includes: a first downsampling, a first star operation, a second star operation, a third star operation, a second downsampling, a first normalization layer, a fourth star operation, a fifth star operation, a sixth star operation, a third downsampling, a second normalization layer, a seventh star operation, an eighth star operation and a ninth star operation cascaded.

[0013] As a preferred embodiment of the present invention, the classification head includes: a global average pooling layer, a linear transformation and activation layer, a tensor reshaping four-dimensional layer, a rectangular self-calibration module, a flattening and normalization layer and a cascade of fully connected layers.

[0014] As a preferred embodiment of the present invention, the rectangular self-calibration module includes: the input feature passes through a rectangular self-calibration attention region layer, a batch normalization layer and a multi-layer perceptron, and then is element-wise multiplied with the input feature rectangle.

[0015] As a preferred embodiment of the present invention, a flame combustion state recognition system of a lightweight INCN includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a flame combustion state recognition method of a lightweight INCN.

[0016] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and when the computer program code is executed by a processor, a method for identifying a flame combustion state of a lightweight INCN is implemented.

[0017] Beneficial effects of the present invention:

[0018] 1. In the feature extraction network of the InceptionNeXt model, the Demo block in StarNet is used to replace the original MetaNeXtBlock, and the structure of the three-branch feature extraction network is redesigned; the star operation is used to perform element-by-element multiplication to fuse the features of different subspaces, and the input is mapped to a high-dimensional nonlinear feature space without increasing the computational cost, which greatly reduces the number of parameters of the baseline model;

[0019] 2. Embed SAFM into the feature extraction network end of the InceptionNeXt model to make the model pay more attention to the important features in the image;

[0020] 3. Aiming at the problem that the model is easily affected by the image background when identifying the combustion state, a rectangular self-calibration module (RCM) is introduced in the classification head. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the lightweight INCN network structure;

[0022] Figure 2 This is a schematic diagram of the star operation (Demo block) structure in StarNet;

[0023] Figure 3 It is a schematic diagram of the structure of the SAFM module;

[0024] Figure 4 It is a schematic diagram of the structure of the RCM module;

[0025] Figure 5 It is the structural principle diagram of the counter-flow burner;

[0026] Figure 6 It is the ROC curve diagram of the improved model proposed in the present invention and other comparative models. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0028] like Figure 1 As shown, a flame combustion state recognition method of a lightweight INCN includes the following steps:

[0029] Step 1, collecting flame images of the combustion state of the counter-flow burner;

[0030] The combustion states include: high oxygen combustion, nitrogen diluted high oxygen combustion, oxygen-rich stable combustion, high nitrogen diluted medium combustion, balanced combustion and low oxygen and low methane combustion.

[0031] The dataset of visible light images of burning flames is enhanced. Noise is selected and applied to the images. The dataset is expanded to 5820 images by reusing existing images and applying different noise combinations multiple times. The ratio of the number of training, validation and test images for each category in the dataset is 70:20:7.

[0032] Step 2: Build an improved InceptionNeNt network (lightweight INCN network); including: image preprocessing module, feature extraction network and classification head;

[0033] To address the problem of large number of parameters in the InceptionNeNt benchmark model, the MetaNeXtBlock in the feature extraction (Stages) network of InceptionNeXt is replaced with the Demoblock in StarNet (star operation), and the structure of the feature extraction (Stages) network is redesigned;

[0034] Image preprocessing includes: 4x4 convolution layer and normalization layer;

[0035] The feature extraction network includes: a first downsampling, a first star operation, a second star operation, a third star operation, a second downsampling, a first normalization layer, a fourth star operation, a fifth star operation, a sixth star operation, a third downsampling, a second normalization layer, a seventh star operation, an eighth star operation and a ninth star operation cascaded;

[0036] Among them, the first downsampling is 1 times, the second downsampling is 2 times, and the third downsampling is 2 times.

[0037] In order to solve the problem of large number of parameters in the InceptionNeXt model, the Demo block in StarNet is used to replace the original MetaNeXtBlock in the feature extraction (Stages) network of the InceptionNeXt model, and the structure of the feature extraction (Stages) network is redesigned; the Demo block in StarNet uses the star operation (Star operation), that is, element-by-element multiplication to fuse the features of different subspaces, and maps the input to a high-dimensional nonlinear feature space without increasing the computational cost; the improved structure is as follows: Stages contains 3 Stages parts, each Stage contains 3 Demoblocks, namely Figure 1 Three downsampling branches of the feature extraction network;

[0038] like Figure 2 As shown in the figure, the core of the Demo block is the star operation, and the expression of the star operation is shown in formula (1):

[0039]

[0040] In the formula, X is the input data, and its dimension is The matrix representing the input features, d is the number of input channels, n is the number of feature elements; W 1 and W 2 is a weight matrix, which is used to perform linear transformation on the input data X to improve the network's ability to extract features from the input data; B 1 and B 2It is the bias vector, which works together with the weight matrix in the linear transformation process of the input data to adjust the output position after the linear transformation, thereby increasing the expressiveness and flexibility of the model; * represents element-by-element multiplication.

[0041] and Respectively represent two different linear transformation operations on the input data, and It is the transposed form of the weight matrix, which is multiplied by X to achieve linear transformation, and then the corresponding bias vector B is added. 1 and B 2 ; The results of these two linear transformations are multiplied element by element in the star operation, that is, the corresponding elements are multiplied to obtain the final output result.

[0042] To simplify the analysis process, the weight matrix and bias are combined, that is, W = [WB], X = [X 1], then the star operation can be rewritten as For a single channel, W 1 ,W 2 Defined as At this time, the form of star operation is Expand this formula, as shown in formula (2):

[0043]

[0044]

[0045] In the formula, items refers to feature items, that is, the new feature combination obtained by the star operation; i and j are variables used for index summation, x i and x j represents the elements of the input feature vector in the and dimensions, α (i,j) To determine each term x i x j The weight in the final result,

[0046] After the calculation is performed in d-dimensional space through the star operation, the result is expressed as The star operation is a combination of different items, which represent the product relationship of features of different dimensions, forming an implicit high-dimensional feature space. Although the calculation is still performed on the basis of the original dimensional input data, from the perspective of feature representation, the mapping to the high-dimensional feature space has been realized, and only the multiplication and addition combination of input features and weights is used without introducing additional complex calculations. The star operation cleverly realizes the expansion of dimensions at a low computational cost and mines higher-dimensional feature representations. The Demoblock containing the star operation is used to replace the MetaNeXtBlock in the feature extraction (Stages) network, which greatly reduces the number of parameters of the baseline model and ensures the feature extraction capability of the lightweight model.

[0047] The classification head and feature extraction network are connected through a spatial adaptive modulation module (SAFM);

[0048] The premise of accurately identifying the combustion state is to pay attention to the important features of the flame contained in the flame image; however, due to the special physical properties of the flame, the lightweight InceptionNeXt benchmark model is difficult to capture the important features of the flame in the image; therefore, the present invention embeds SAFM into the feature extraction (Stages) network end of the InceptionNeXt model, so that the model pays more attention to the important features in the image.

[0049] like Figure 3 , SAFM first performs channel segmentation on the normalized input features, as shown in formula (3):

[0050] [X 0 ,X 1 ,X 2 ,X 3 ]=Split(X) (3)

[0051] In the formula, X is the input feature and Split() is the channel splitting operation.

[0052] For the input feature X 0 Perform convolution operation as shown in formula (4):

[0053]

[0054] In the formula, DW-Conv 3×3 (X 0 ) is a depthwise convolution operation with a kernel size of 3×3, used to extract X 0 The local characteristics of It is the input feature after the convolution operation.

[0055] For the input feature X i(1≤i≤3), first downsample and then perform depth convolution, and then upsample to the original resolution, as shown in formula (5):

[0056]

[0057] In the formula, ↑ p represents upsampling the feature to the original resolution p by nearest neighbor interpolation, is a downsampling operation.

[0058] The multi-scale features processed above Aggregation is performed through convolution 1×1, as shown in formula (6):

[0059]

[0060] In the formula, Conv 1×1 () is a 1×1 convolution operation, and Concat() is a concatenation operation along the channel dimension.

[0061] Get the aggregated features After that, it is normalized by the GELU activation function, as shown in formula (7):

[0062]

[0063] In the formula, φ() represents the GELU activation function. The calculation formula of GELU is CELU(x)=xΦ(x), where xΦ(x) is the cumulative distribution function of the standard normal distribution; ⊙ is the element-by-element multiplication operation. Adaptively modulate the original input feature X, and in this way adjust the input feature according to the learned feature importance to provide the model with more useful feature information, so that the model can focus on the features that are more important for the recognition task.

[0064] The classification head includes: global average pooling layer, linear transformation and activation layer, tensor reshaping 4D layer, rectangular self-calibration module, flattening and normalization layer and fully connected layer cascade;

[0065] The four-dimensional layer of tensor reshaping is an existing module. The previous layer: the linear transformation and activation layer outputs a three-dimensional tensor, but the rectangular self-calibration module only accepts four-dimensional tensor input, so the three-dimensional tensor needs to be converted into a four-dimensional tensor. Therefore, tensor reshaping is used before the rectangular self-calibration module.

[0066] The image consists of a foreground object and a background. The foreground object is the object closer to the lens. The foreground image in the flame image is the flame. In order to solve the problem that the model is easily affected by the image background when identifying the combustion state, a rectangular self-calibration module RCM is introduced in the classification head (MlpHead).

[0067] like Figure 4 As shown, RCM is achieved by horizontal pooling H p and vertical pooling V p The axial global context is captured and two axis vectors are generated, which are then added together to model a rectangular region, as shown in Equation (8):

[0068]

[0069] Where RCA stands for rectangular self-calibration attention, H p represents horizontal pooling, V p represents vertical pooling, represents broadcast addition, i.e. element-wise addition, and x is the input feature.

[0070] The shape self-calibration function is used to adjust the RCA rectangular attention area to make it closer to the foreground object in the image, as shown in formula (9):

[0071]

[0072] In the formula, ξ C () represents the shape self-calibration function, Represents input features, δ represents Sigmoid function, ψ represents large kernel strip convolution, k represents the kernel size of strip convolution, and φ represents batch normalization followed by ReLU function.

[0073] The feature fusion function is used to fuse the attention features and input features, as shown in formula (10):

[0074] ξ F (x,y)=ψ 3×3 (x)⊙y (10)

[0075] In the formula, ξ F () represents the feature fusion function, ψ 3×3 represents a 3×3 depthwise convolution, ⊙ represents the Hadamard product, and ⊙ weights the calibrated attention feature y onto the refined input feature.

[0076] Combining equations (8) to (10), the overall formula of RCM can be obtained as follows:

[0077]

[0078] In the formula, F out represents the output features after processing by the rectangular self-calibration module (RCM), ρ represents batch normalization and multi-layer perceptron;

[0079] The RCM module achieves the focus on the foreground object, improves the model’s extraction quality of flame features in the image, and further improves the model’s performance.

[0080] The above is an improvement to the lightweight combustion state recognition model of the improved InceptionNeXt, which solves the problem of its large number of parameters and takes into account the computational complexity and accuracy of the model.

[0081] Experimental process:

[0082] Train the model and set the training hyperparameters; the number of training rounds is 300, the number of samples in each iteration is 16, the learning rate is 0.0001, and the coefficient of the exponential moving average is 0.9998; in each training round, verify on the validation set and retain the best performing model.

[0083] Test the model, test the best performing model on the test set, and output the final flame image recognition results.

[0084] The training and testing of the model of the present invention were carried out in a 64-bit Ubuntu 18.04 operating system equipped with a 13th Gen Intel(R) Core(TM) i5-13490F CPU and an NVIDIA GeForce RTX4060Ti GPU. The CUDA version in the deep learning environment was 11.8, the PyTorch version was 2.3.1, and the Python version was 3.8.

[0085] like Figure 5 The dataset used in this paper is derived from the work of Kang et al. (2022), and contains visible light images of the flame of a counterflow burner under six different combustion states captured by a FASTEC TS-5 high-speed CMOS camera; the camera is sensitive to light with a wavelength range of 350 to 950nm, covering the entire visible light band, and has a high sampling rate of 100 frames per second, thereby ensuring the clarity and quality of the captured images; in order to simulate a variety of combustion states, the researchers successfully created six different combustion conditions by adjusting the gas ratio in the counterflow burner, and collected video data of the flame under each state. After the original video data was processed by frame extraction, a total of 2,640 visible light images of flames under six types of combustion states were finally screened out, with a resolution of 640×480 for each image.

[0086] According to the different gas proportions in the countercurrent burner, the six combustion states contained in the data set are named as: Oxygen-Enriched Stable Combustion (OESC), High Oxygen Combustion (HOC), Nitrogen-Diluted High Oxygen Combustion (ND-HOC), Low Oxygen and Low Methane Combustion (LO-LMC), High Nitrogen-Diluted Moderate Combustion (HNMC) and Balanced Combustion (BC); these combustion states cover the main working conditions of the countercurrent burner. Selecting this data set for experiments can effectively verify the performance and reliability of the model of the present invention.

[0087] In order to further evaluate the effect of the model in the combustion state recognition task, precision, recall, F1-score, overall accuracy, TOP-3 accuracy, log loss, parameter number and number of samples processed per second (FPS) are used. FPS is one of the key indicators to measure the efficiency of model reasoning. A higher FPS value means that the model can process more input samples per unit time, reflecting the real-time and high efficiency of the model.

[0088] In order to verify the superiority of the proposed model in the combustion state recognition task, a comprehensive evaluation was conducted on its performance, and a detailed comparative experimental analysis was conducted with other four advanced deep learning models, NextViT, ConvNeXtV2, MobileNetV4 and CAS-ViT. The proposed model and the other four models all used the same hyperparameters during training and were experimented on the same data set. The comparative experimental results are shown in Table 1: The macro average (Macroavg) indicators of Precision, Recall and F1-score of the proposed model reached 85.81%, 83.45% and 93.81% respectively. 83.98%, which is 4.87%, 7.80% and 7.32% higher than the second-best model NextViT (80.94%, 75.65%, 76.66%) respectively; the weighted average F1-score is 84.11%, which is also ahead of NextViT (76.91%) and MobileNetV4 (72.63%), verifying that the star operation can efficiently map the input features to the high-dimensional feature space, and enhance the model's ability to express features while reducing the number of parameters; in terms of overall accuracy (Acc), the model of the present invention leads with 83.92% , while NextViT is 76.19%, MobileNetV4 is 71.90%, ConvNeXtV2 is 39.76%, and CAS-ViT is 44.29%; this verifies the effectiveness of the lightweight improvement of the model of the present invention, and strikes a balance between the number of model parameters and performance; in terms of TOP-3 accuracy (TOP-3Acc), the model of the present invention also performs well, reaching 83.92%; the higher TOP-3 accuracy in the experiment shows that the improved model of the present invention can still provide more reliable prediction results when the flame image is difficult to distinguish; this verifies the spatial adaptive modulation module (SAFM) for attention The effectiveness of important features; the logarithmic loss (LogLoss) of the model of the present invention is 0.5455, which is significantly lower than all the comparison models, indicating that its combustion state prediction confidence is higher and the uncertainty is lower; this verifies the effectiveness of the rectangular self-calibration module (RCM) in reducing the background impact of the flame image on performance improvement; in terms of parameter amount (Params), the model of the present invention is only 80.03MB, which is smaller than NextViT (117.28MB) and MobileNetV4 (119.47MB), which verifies the simplicity and compactness of the star operation and shows that it has a greater contribution to reducing the amount of model parameters. This not only helps to reduce the consumption of computing resources, but also makes the model easier to deploy and optimize in practical applications.In terms of the number of samples processed per second (FPS), the FPS value of the model proposed in the present invention is 434.26, which is significantly better than other comparison models. This verifies that the lightweight architecture proposed in the present invention can more effectively reduce the computational burden and achieve higher throughput during the reasoning process. The lower FPS of other comparison models will affect the performance in practical applications that require fast response, so the lightweight improved model proposed in the present invention achieves a good balance between reasoning speed and computational efficiency.

[0089] Table 1 Comparative experiment

[0090]

[0091]

[0092] like Figure 6 The paper verifies the remarkable superiority of the lightweight improved model by comparing the ROC curves and AUC values ​​of the improved model with the current mainstream image classification models (NextViT, ConvNeXtV2, MobileNetV4, CAS-ViT) on six types of combustion state flame images (OESC, HOC, ND-HOC, LO-LMC, HNMC, BC); the experimental results show that the lightweight improved model shows the most comprehensive performance improvement among all the compared models, and the AUC values ​​are significantly higher than NextViT in identifying the six types of combustion states; the AUC of the OESC category of the improved model is 0.27 higher than that of NextViT; compared with ConvNeXtV2, the improved model performs better in the LO-LMC and HNMC categories, with AUC values ​​of 0.97 and 0.98 in the LO-LMC and HNMC categories, respectively, while ConvNeXtV2 has only 0.74 and 0.83; compared with MobileNetV4 In comparison, the improved model is ahead of MobileNetV4 in HOC and LO-LMC. The AUC values ​​of the improved model in HOC and LO-LMC are both 0.97, while MobileNetV4 only has 0.94 and 0.91, and it does not lag behind in any category. Compared with CAS-ViT, the improved model has achieved comprehensive leadership in all six types of combustion state recognition tasks. In terms of the AUC values ​​of ND-HOC and LO-LMC, CAS-ViT only achieved 0.72 and 0.80, while the improved model reached 0.97, with an outstanding advantage. In addition, the improved model shows stronger robustness, and its AUC value fluctuation range (0.93-0.98) is significantly smaller than that of the comparison model (such as 0.69-0.83 of NextViT), indicating that its generalization ability is better. It can be seen from the ROC curve that the overall comprehensive performance of the improved model ranks first. This result verifies the effectiveness of the lightweight improvement strategy of the InceptionNeXt model proposed in this invention.

[0093] Table 2 Ablation experiment

[0094]

[0095] The present invention conducts an ablation experiment to verify the improvement effect of different optimization strategies on model performance. A total of 4 groups of experiments are conducted on the same flame visible light image, all of which maintain the same hyperparameter settings and experimental environment. The ablation experiment results are shown in Table 2. The InceptionNeXt model is used as the benchmark model, and the plus sign indicates that the optimization strategy is adopted. Compared with the benchmark model InceptionNeXt, after adding the star operation alone, the model performance decreases, but the parameter amount is greatly reduced from 179.61MB to 77.69MB; this verifies the contribution of the introduction of the star operation to the lightweight model. After further introducing the SAFM module, the model performance is significantly improved; the macro-average F1-score and weighted average F1-score are increased to 74.08% and 74.21% respectively, and the overall accuracy and TOP-3 accuracy are also increased to 73.81% respectively. and 89.52%, while the logarithmic loss is reduced to 0.7388, and the parameter amount is slightly increased to 79.96MB; this shows that the SAFM module can effectively improve the recognition performance of the model without significantly increasing the parameter amount of the model, verifying that the SAFM module can enhance the model's ability to focus on key features in flame images; after continuing to add the RCM module, the model performance is greatly improved; the macro average of F1-score reaches 83.98, the weighted average reaches 84.11, the overall accuracy reaches 83.92%, the TOP-3 accuracy reaches 93.73%, the logarithmic loss is reduced to 0.5455, and the parameter amount is 80.03MB; the results show that the lightweight improved model surpasses the baseline model in multiple key indicators, and the parameter amount is much lower than the baseline model, achieving a balance between performance improvement and lightweight, which verifies the importance of reducing background influence for improving model recognition.

[0096] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for identifying the flame combustion state of a lightweight INCN, characterized in that: The following steps are involved: Step 1, collecting flame images of the combustion state of the counter-flow burner; Step 2: Build an improved InceptionNeNt network, including: image preprocessing module, feature extraction network and classification head. In the feature extraction network of the InceptionNeXt model, use the Demo block in StarNet to replace the original MetaNeXtBlock; and design a three-stage feature extraction structure.

2. The flame combustion state recognition method of lightweight INCN according to claim 1 is characterized in that: The feature extraction network includes: a first downsampling, a first star operation, a second star operation, a third star operation, a second downsampling, a first normalization layer, a fourth star operation, a fifth star operation, a sixth star operation, a third downsampling, a second normalization layer, a seventh star operation, an eighth star operation and a ninth star operation cascaded.

3. The flame combustion state identification method of lightweight INCN according to claim 1 is characterized in that: The classification head includes: a global average pooling layer, a linear transformation and activation layer, a tensor reshaping 4D layer, a rectangular self-calibration module, a flattening and normalization layer, and a cascade of fully connected layers.

4. The flame combustion state identification method of lightweight INCN according to claim 3 is characterized in that: The rectangular self-calibration module includes: the input feature passes through the rectangular self-calibration attention area layer, the batch normalization layer and the multi-layer perceptron, and then is multiplied element-by-element with the input feature rectangle.

5. The flame combustion state identification method of lightweight INCN according to claim 1 is characterized in that: The image preprocessing module includes: 4x4 convolution layer and normalization layer.

6. The flame combustion state recognition method of lightweight INCN according to claim 1 is characterized in that: The combustion states include: high oxygen combustion, nitrogen diluted high oxygen combustion, oxygen-rich stable combustion, high nitrogen diluted medium combustion, balanced combustion and low oxygen and low methane combustion.

7. The flame combustion state identification method of lightweight INCN according to claim 1 is characterized in that: Perform noise enhancement on flame images.

8. Lightweight INCN flame combustion state recognition system, characterized by: include: a memory for storing instructions executable by a processor; A processor is used to execute instructions to implement the flame combustion state recognition method of the lightweight INCN according to any one of claims 1 to 7.

9. A computer readable medium storing a computer program code, characterized in that: When the computer program code is executed by a processor, the method for identifying the flame combustion state of a lightweight INCN as claimed in any one of claims 1 to 7 is implemented.

Images