Combustion state identification method based on improved InceptionNeXt

Abstract

The invention relates to the technical field of countercurrent combustors, in particular to a combustion state recognition method based on improved InceptionNeXt, which comprises the following steps: acquiring a combustion state image of a countercurrent combustor; constructing an improved InceptionNeXt model, wherein the improved InceptionNeXt model comprises a preprocessing module, a feature extraction module and a classification head module; the preprocessing is formed by 4 * 4 convolution and a normalization layer; the feature extraction module comprises a first triple attention mechanism module and a first improved MetaNeXt module which are connected in a three-level manner; the first triplet attention mechanism module, the first downsampling module and the first improved MetaNeXt module are cascaded in three stages, the second triplet attention mechanism module, the first downsampling module and the second improved MetaNeXt module are cascaded in three stages, the third triplet attention mechanism module and the second downsampling improved MetaNeXt module are cascaded in twenty-seven stages, and the fourth triplet attention mechanism module, the third downsampling module and the third improved MetaNeXt module are cascaded in three stages. According to the invention, the problem of low recognition precision in a combustion state recognition task in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of backflow burners, and in particular to a combustion state recognition method based on an improved InceptionNeXt. Background Art

[0002] The counterflow burner is a combustion device used for combustion research. Due to its simplicity, reliability and excellent performance, it is widely used to study combustion characteristics, pollutant generation, flame propagation speed and combustion stability.

[0003] Malpica et al. proposed a gradient-free combustion state recognition method, but due to the lack of prior knowledge, this method is only applicable to specific combustion conditions; Sitaraman et al. proposed a machine learning method for identifying the combustion state in the RCCI engine combustion device, but due to the lack of prior knowledge, the recognition accuracy is low; Bhattacharya et al. proposed a combustion state recognition method based on probabilistic finite state automata, but the required data volume is large and the computational efficiency is low; Compais et al. proposed a machine learning method for identifying the combustion state in a laboratory combustion device, which can still achieve high accuracy with a small amount of data, but it requires a lot of computing resources and is not conducive to real-time application; Bukkarapu et al. proposed a decision tree model for identifying the combustion state of the burner. This method achieved high recognition accuracy, but the required computational cost was high.

[0004] InceptionNeXt is widely used due to its high computational efficiency, high accuracy and strong feature extraction capabilities, and has achieved good experimental results. However, despite the good results, InceptionNeXt still has some shortcomings in combustion state recognition, such as difficulty in adapting to the dynamic changes of flames, easy loss of key information of flame images, and small receptive field. Summary of the invention

[0005] Aiming at the shortcomings of the existing methods, the present invention solves the problem that the existing methods have low recognition accuracy when performing combustion state recognition tasks.

[0006] The technical solution adopted by the present invention is: based on the improved InceptionNeXt combustion state recognition method, the following steps are included:

[0007] Step 1, collecting a combustion state image of a counter-flow burner;

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

[0009] Step 2, build an improved InceptionNeXt model, including: preprocessing module, feature extraction module and classification head module; the preprocessing consists of 4x4 convolution and normalization layer; the feature extraction module includes the first triple attention mechanism module, the first improved MetaNeXt module three-stage cascade, the second triple attention mechanism module, the first downsampling, the second improved MetaNeXt module three-stage cascade, the third triple attention mechanism module, the second downsampling improved MetaNeXt module twenty-seven-stage cascade, the fourth triple attention mechanism module, the third downsampling and the third improved MetaNeXt module three-stage cascade.

[0010] As a preferred implementation of the present invention, improving the MetaNeXt module includes: replacing the normalization layer and the multilayer perceptron in the MetaNeXt module with a SimC module.

[0011] As a preferred embodiment of the present invention, the formula of the SimC module is:

[0012]

[0013] In the formula, Used to determine pixel similarity, i is the pixel number, represents the variance of all neurons on a single channel, λ is the regularization term, and t i represents the i-th pixel of the input feature map on a single channel, represents the mean value of all pixels on a single channel, is the enhanced tensor, E is The sum of ⊙ is the Hadamard multiplication, C is the channel dimension, b and γ are used to change the ratio of the number of channels to the convolution kernel size, odd means an odd number, Conv1D k () is a one-dimensional convolution operation with kernel size k.

[0014] As a preferred embodiment of the present invention, the classification head module includes: a multi-scale hole parallel integration module, a global average pooling, a first fully connected layer, an activation layer, a feature expression module, a normalization layer, a random flexibility, and a second fully connected layer.

[0015] As a preferred embodiment of the present invention, the multi-scale hole parallel integration module includes: multi-scale hole convolution and parallel channel and spatial attention mechanism; wherein the formula of multi-scale hole convolution is:

[0016] F1(x)=ReLU(BN(Conv2D(x,1×1,dilation=1))) (11)

[0017] F2(x)=ReLU(BN(Conv2D(x,3×3,dilation=6))) (12)

[0018] F3(x)=ReLU(BN(Conv2D(x,3×3,dilation=12))) (13)

[0019] F4(x)=ReLU(BN(Conv2D(x,3×3,dilation=18))) (14)

[0020] F5(x)=ReLU(BN(Conv2D9GlobalAvgPool(x),1×1))) (15)

[0021] Fconcat(x)=Concat(F1(x),F2(x),F3(x),F4(x),F5(x)) (16)

[0022] Among them, ReLU() is a nonlinear activation function, BN() is batch normalization, Conv2D() is a two-dimensional convolution, dilation is the void rate, GlobalAvgPool() is the global average pooling, and Concat() refers to splicing 5 tensors in dimension.

[0023] As a preferred embodiment of the present invention, the feature expression module includes:

[0024] First, calculate the channel attention mechanism, the formula is:

[0025] T c =Sigmoid(σ(W avg ·F avg +b avg )+σ(W max ·F max +B max ))⊙ T (17)

[0026] Where, T c is the weighted feature map; Sigmoid() and σ are activation functions; W avg and W max is the weight of the fully connected layer; F avg is the global average pooling, F max is the global maximum pooling; b avg and b max is the bias term; ⊙ is the element-by-element multiplication operation;

[0027] Secondly, perform channel shuffle operation, the formula is:

[0028] T s=Shuffle(T c ) (18)

[0029] Where, T s is the shuffled feature map; Shuffle() shuffles the grouped channels; T c It is the feature map after channel attention weighting;

[0030] Finally, the shuffled feature map is processed and output using spatial attention. The formula for spatial attention is:

[0031] T s '=Sigmoid(W a ·[AvgPool(T s );MaxPool(T s )]+b a )⊙T s (19)

[0032] Where, T s ' is the feature map after spatial attention processing; Sigmoid() is the activation function; W a is the convolution kernel; AvgPool(T s ) is the input feature map T s The global average pooling operation in the channel dimension, MaxPool(T s ) is the input feature map T s Global maximum pooling operation in the channel dimension; b a is the bias term; ⊙ is the element-by-element multiplication operation; T s is the input feature map.

[0033] Step 3: Use the loss function to update the parameters of the improved InceptionNeXt model;

[0034] As a preferred implementation of the present invention, the loss function includes: a soft target cross entropy loss function.

[0035] As a preferred implementation manner of the present invention, the loss function further includes: a Focal Loss loss function.

[0036] As a preferred embodiment of the present invention, an improved InceptionNeXt combustion state recognition system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the improved InceptionNeXt combustion state recognition method.

[0037] Beneficial effects of the present invention:

[0038] 1. The improved InceptionNeXt counterflow burner combustion state recognition model based on feature extraction optimization is proposed in this paper. First, the Triplet Attention mechanism is introduced to apply attention on the image width, height and RGB dimensions to enhance the model's ability to capture the dynamic characteristics of flames. Secondly, the SimC feature focusing module is added after the Inception deep convolution to make up for the defect that the Inception deep convolution loses key flame information and enhance the model's ability to recognize key information. Then, a multi-scale hole parallel fusion module is embedded after the feature extraction part of the model to expand the receptive field of the model, capture a wider range of contextual information in the flame image, and work with Focal Loss to improve the recognition performance.

[0039] 2. In the classification part, the ShuffleCS module based on channel shuffling is added to improve the problem of the original MlpHead ignoring channel interaction through information interaction of different channels. The recognition ability of MlpHead for different states is enhanced through feature mixing, further improving the performance of the model.

[0040] 3. Introduce the Focal Loss loss function to replace SoftTargetCrossEntropy to increase the model's attention to difficult-to-recognize states;

[0041] 4. The proposed model is effective in the combustion state recognition task, achieving an overall accuracy of 85.71%, and its performance is higher than that of the baseline model and the comparison model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the improved InceptionNeXt structure diagram of the present invention;

[0043] Figure 2 It is a schematic diagram of the structure of the InceptionNeXt baseline model of the present invention;

[0044] Figure 3 It is a structural diagram of the Triplet Attention module of the present invention;

[0045] Figure 4 It is the structural diagram of the SimC module of the present invention;

[0046] Figure 5 It is the MDCPI module structure diagram of the present invention;

[0047] Figure 6 It is the module structure diagram of ShuffleCS of the present invention;

[0048] Figure 7 This is the working principle diagram of the counter-flow burner;

[0049] Figure 8 It is the flame visible light image of six types of combustion states;

[0050] Fig. 9 This is a comparison chart of the Loss curve and Acc curve in the training of the baseline model and the improved model;

[0051] Fig.10 Confusion matrix of the improved model and the comparison model on the test set. DETAILED DESCRIPTION

[0052] 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.

[0053] like Figure 1 As shown, the combustion state recognition method based on the improved InceptionNeXt includes the following steps:

[0054] Step 1, collecting a combustion state image of a counter-flow burner;

[0055] The combustion states include: oxygen-rich stable combustion, high oxygen combustion, nitrogen-diluted high oxygen combustion, low oxygen and low methane combustion, high nitrogen-diluted medium combustion and balanced combustion.

[0056] Step 2: Build an improved InceptionNeXt model and use the improved InceptionNeXt model to identify the combustion state image;

[0057] The improved InceptionNeXt model includes: a preprocessing module, a feature extraction module and a classification head module; the preprocessing consists of a 4x4 convolution and a normalization layer; the feature extraction module includes a first triple attention mechanism module, a first improved MetaNeXt module in three cascades, a second triple attention mechanism module, a first downsampling, a second improved MetaNeXt module in three cascades, a third triple attention mechanism module, a second downsampling improved MetaNeXt module in twenty-seven cascades, and a fourth triple attention mechanism module, a third downsampling and a third improved MetaNeXt module in three cascades.

[0058] MetaNeXt module triple cascade refers to the sequential connection of three MetaNeXt modules.

[0059] like Figure 2The large-core depthwise convolution has been widely studied and adopted due to its large receptive field. However, due to its high memory access, the efficiency of the model is low. If the memory access can be reduced, the purpose of improving the detection speed can be achieved. Therefore, InceptionNeXt proposes an Inception depthwise convolution, which significantly improves the efficiency of the model while maintaining the performance of large-core depthwise convolution.

[0060] Assume that the input is X, and the operation of the traditional large kernel depth convolution is shown in formula (1):

[0061]

[0062] In the formula, X' is the result after the large kernel depth convolution operation, represents a depthwise convolution with a kernel size of k×k, and X represents the input and output channels.

[0063] The operation of Inception deep convolution is shown in formula (2):

[0064]

[0065] Where, X hW ,X W ,X h ,X id are the four convolution branches after decomposition, k s ,k b Represents the size of the decomposed convolution kernel, and Concat() represents connecting the output of each branch.

[0066] From the comparison between equation (1) and equation (2), we can see that the traditional large-kernel deep convolution requires k×k operations each time, while the Inception deep convolution decomposes the traditional large-kernel deep convolution into four parallel convolution branches: small square convolution (kernel size is k) s ×k s ), two strip convolutions (kernel size is 1×k b and k b×1) and identity mapping, the amount of calculation of each parallel branch is much smaller than k×k, so the amount of calculation is significantly reduced; among them, the small square convolution is used to retain the small-scale feature capture of the traditional convolution operation, the two orthogonal strip convolutions are used to expand the receptive field and reduce the computational cost, and the identity mapping is used to keep some channels unchanged; this design solves the problems of high memory access cost and decreased computing speed of traditional large-core deep convolution, and by decomposing the large-core deep convolution into multiple small-core convolutions and one-dimensional strip convolutions, the advantage of the large receptive field of the large-core deep convolution is maintained, and the computational complexity is reduced; InceptionNeXt is used as the benchmark model, which provides a good foundation for the subsequent recognition of combustion status.

[0067] The InceptionNeXt image recognition model consists of three parts: image preprocessing (Input), feature extraction (Stages) and classification head (MlpHead). The Input part performs preprocessing on the flame image of the counterflow burner, such as resizing, convolution operation and batch normalization. The Stages part uses the MetaNeXtBlock module to extract the feature information of the flame image and output the feature map. The MlpHead part performs a global average pooling operation on the feature map, and then performs feature transformation through two fully connected layers, and finally outputs the recognition result of size Num_classes.

[0068] For the task of recognizing the combustion state of the reverse flow burner, the triple attention mechanism is introduced to improve the accuracy of the model in recognizing the combustion state. The purpose of introducing triplet attention in the stages part of the model is to enhance the model's ability to capture the dynamic features of the flame image. The module structure of triplet attention is as follows: Figure 3 shown.

[0069] The InceptionNeXt baseline model has difficulty capturing the dynamic features of flames. To solve this problem, Triplet Attention is introduced in the Stages part of the model to enhance the model's ability to capture the dynamic features of flame images. Triplet Attention applies attention in the three dimensions of height, width, and RGB to enhance its ability to model the dynamic features of flames.

[0070] Triplet Attention is based on multi-dimensional attention to the input features, dividing the feature map into three dimensions, namely height, width and channel, and then applying attention to each dimension through a three-branch structure. C×H×W , firstly, the three branches are operated, and then the average is taken to get the final weight. The calculation formula is shown in formulas (3) to (6):

[0071]

[0072] M * =Mσ(ω3*(Z-pool(M))) (5)

[0073]

[0074] In the formula, H - It means rotating 90° counterclockwise along the H axis, W - represents a 90° counterclockwise rotation along the W axis, σ represents the activation function, ω1 and ω2 represent the convolution kernel, * represents the convolution operation, H + It means rotating 90° clockwise along the H axis, W + It means rotating 90° clockwise along the W axis, and Z-pool is the feature-preserving representation.

[0075] Based on the three-branch design of Triplet Attention, it can pay attention to different spatial dimensions (height, width and channel) respectively. At different times and burning states, the shape and color of the flame will change. In terms of shape, the flame features change in the vertical and horizontal directions. Triplet Attention can extract attention information from the height and width of the flame image respectively, and dynamically adjust the model's sensitivity to these changing features, so as to better capture the changes in flame shape. In terms of color, the change in flame color under different burning states is essentially a change in the RGB component, which will be reflected through the channel dimension of the image. Triplet Attention extracts attention information from the channel dimension and can adapt to the color changes of the flame more flexibly. In short, the model's ability to capture the dynamic features of flame images is enhanced from these two aspects.

[0076] The Inception deep convolution used in the InceptionNeXt baseline model is to decompose the large kernel deep convolution, and the key features in the image will be ignored in the decomposition process; in order to improve the problem that the multi-branch (Inception) deep convolution is easy to ignore the key information, thereby further improving the recognition accuracy of the model, the present invention designs a SimC feature focusing module based on the energy function, and embeds it into the multi-branch deep convolution, that is, replaces the normalization layer and the multi-layer perceptron in the MetaNeXt module with the SimC focusing module; the basic principle of the SimC feature focusing module is to enhance the pixels and channels containing important information, and its structure is as follows Figure 4 shown.

[0077] The specific calculation formula of the SimC focusing module is shown in equations (7) to (10), where equation (7) is the energy function.

[0078]

[0079] In the formula, Used to determine pixel similarity, i is the pixel number, represents the variance of all neurons on a single channel, λ is the regularization term, and t i represents the i-th pixel of the input feature map on a single channel, represents the mean value of all pixels on a single channel, is the enhanced tensor, E is The sum of ⊙ is the Hadamard multiplication, C is the channel dimension, b and γ are used to change the ratio of the number of channels to the convolution kernel size, odd means an odd number, Conv1D k () is a one-dimensional convolution operation with kernel size k.

[0080] For the input feature tensor, the SimC module finds pixels with high similarity based on the energy function of local image similarity and enhances them in subsequent operations, thereby highlighting the key features that are important for accurate identification; after obtaining the feature tensor containing enhanced pixels, in order to maintain the lightweight advantage of Inception deep convolution, the importance of each channel relative to other channels is learned through lightweight and fast one-dimensional convolution operations, thereby increasing the channel weights related to key information, and finally achieving the goal of focusing on important information in the feature map; the SimC module adaptively weights the input feature map and dynamically adjusts the channel weights, effectively improving the model's sensitivity to key image information, thereby improving the problem of Inception deep convolution ignoring key information.

[0081] In order to expand the receptive field of the model and make it pay more attention to the contextual information of the flame image, the present invention designs a multi-scale hole parallel integration module (MDCPI) based on hole convolution. MDCPI enhances the model's perception of multi-scale contextual information by utilizing multiple hole rates and integrating channel and spatial attention mechanisms. The structure of the MDCPI module is as follows: Figure 5 shown.

[0082] The workflow of the MDCPI module is divided into two steps. The first step is multi-scale dilated convolution, which uses five parallel convolution branches with different dilation rates to extract features of different scales, then concatenates them in the channel dimension, and finally outputs a comprehensive feature map. The expressions of multi-scale dilated convolution are shown in Equations (11) to (16).

[0083] F1(x)=ReLU(BN(Conv2D(x,1×1,dilation=1))) (11)

[0084] F2(x)=ReLU(BN(Conv2D(x,3×3,dilation=6))) (12)

[0085] F3(x)=ReLU(BN(Conv2D(x,3×3,dilation=12))) (13)

[0086] F4(x)=ReLU(BN(Conv2D(x,3×3,dilation=18))) (14)

[0087] F5(x)=ReLU(BN(Conv2D(GlobalAvgPool(x),1×1))) (15)

[0088] F concat (x)=Concat(F1(x),F2(x),F3(x),F4(x),F5(x)) (16)

[0089] In equations (12) to (16), ReLU() is a nonlinear activation function, BN() is batch normalization, Conv2D() is a two-dimensional convolution, dilation is the dilation rate, GlobalAvgPool() is the global average pooling, and Concat() refers to concatenating 5 tensors in a specific dimension.

[0090] However, in the experiment, it was found that the feature map output by the dilated convolution was missing some pixels; this is because the dilated convolution has a grid effect. Due to the setting of the dilated rate, the convolution kernel will skip several pixels for sampling, causing the sampled point distribution to present a grid shape rather than a continuous pixel area, which in turn leads to the loss of local information; therefore, in order to solve the problem in the first step, the parallel channel spatial attention mechanism is used in the second step; the feature map output by the first step is calibrated by the parallel channel and spatial attention mechanism, and then the feature maps output by the two parallel branches are fused through element-wise addition operations, and the final output feature map is generated through integration and enhancement.

[0091] The MDCPI module significantly expands the receptive field of the model through multi-scale dilated convolution, so that it can capture a wider range of contextual information in the flame image; because of the grid effect of dilated convolution, the feature map processed by multi-scale dilated convolution is calibrated using parallel channel and spatial attention mechanisms; channel attention alleviates the problem of local information lost due to dilated convolution by enhancing global contextual information, making the features more comprehensive in the channel dimension; spatial attention corrects the discontinuity of features in the spatial dimension, making the feature map more consistent in space and offsetting the local non-smoothness caused by the grid effect; finally, the feature outputs of the two parallel branches are integrated to ensure that the model expands the receptive field and can effectively capture the multi-scale contextual information of the flame feature map.

[0092] The present invention designs a feature expression module (ShuffleCS) based on channel shuffling, which enhances the recognition ability of different states by feature mixing. The module structure is as follows: Figure 6 shown.

[0093] In the original InceptionNeXt model, the classification head only relies on global average pooling (GAP) to extract global features. It isolates all operations and makes it impossible to exchange information between groups, thereby reducing the expressiveness of the input features. In order to enhance the cross-channel communication of features and further improve the performance of the model, a feature expression module ShuffleCS based on channel shuffling is designed. The specific calculation formula of the ShuffleCS module is shown in equations (17) to (19):

[0094] The first step is to calculate the channel attention mechanism. For the input feature map T,

[0095] T c =Sigmoid(σ(W avg ·F avg +b avg )+σ(W max ·F max +b max ))⊙T#(17)

[0096] Where, T c is the weighted feature map; Sigmoid() and σ are activation functions; W avg and W max is the weight of the fully connected layer; F avg is the global average pooling, F max is the global maximum pooling; b avg and b max is the bias term; ⊙ is the element-wise multiplication operation.

[0097] After being processed by the channel attention mechanism, the channel shuffle operation is performed, and the calculation formula is as follows:

[0098] T s =Shuffle(T c ) (18)

[0099] Where, T s is the shuffled feature map; Shuffle() shuffles the grouped channels; T c It is the feature map after channel attention weighting.

[0100] After the channel shuffling operation, in order to make full use of the spatial information in the feature map, the shuffled feature map is processed and output using spatial attention. The calculation formula of spatial attention is as follows:

[0101] T s '=Sigmoid(W a ·[AvgPool(T s );MaxPool(T s )]+b a )⊙T s (19)

[0102] Where, T s ' is the feature map after spatial attention processing; Sigmoid() is the activation function; W a is the convolution kernel; AvgPool(T s ) is the input feature map T s The global average pooling operation in the channel dimension, MaxPool(T s ) is the input feature map T s Global maximum pooling operation in the channel dimension; b a is the bias term; ⊙ is the element-by-element multiplication operation; T s is the input feature map.

[0103] Step 3: Use the loss function to update the parameters of the improved InceptionNeXt model;

[0104] The loss function in the InceptionNeXt baseline model is SoftTargetCrossEntropy (soft target cross entropy loss function). Its basic principle is to calculate the difference between the predicted probability distribution and the target probability distribution, and update the model parameters through this difference. Its formula is shown in formula (20):

[0105]

[0106] Where L is the loss value, C is the number of categories, and y i is the target probability distribution, that is, the probability that the sample belongs to the i-th category, pi is the probability distribution predicted by the model, that is, the probability that the model predicts that the sample belongs to the i-th category.

[0107] In the combustion state recognition task, the recognition difficulty of the six states varies; however, SoftTargetCrossEntropy cannot effectively emphasize the attention to the difficult-to-recognize state; in order to solve the problem that the model does not pay enough attention to the difficult-to-recognize samples, the present invention adopts the Focal Loss loss function to replace SoftTargetCrossEntropy; FocalLoss dynamically adjusts the weight of the loss function to make the model pay more attention to the difficult-to-recognize state, and its formula is shown in formula (21):

[0108]

[0109] Among them, p t is the probability of the logit output after the softmax function conversion, C is the number of categories, α t It is a parameter for adjusting the weight of positive and negative samples, and γ is a parameter for adjusting the weight of difficult and easy samples.

[0110] Experimental results and analysis:

[0111] In order to verify the effectiveness of the improved InceptionNeXt model for the combustion state recognition task, experiments were conducted on the flame visible light image dataset of the counterflow burner. Based on the 64-bit Ubuntu operating system and NVIDIA GeForce RTX4060 Ti graphics card, the PyTorch deep learning framework and Python compiled language were used for the experiment. The experimental environment configuration is shown in Table 1:

[0112] Table 1 Experimental environment configuration

[0113] Table 1 Experimental environment configuration

[0114]

[0115]

[0116] The dataset used is from Kang et al., which contains visible light images of flames in six combustion states of a counterflow burner collected by a FASTEC TS-5 high-speed CMOS camera. The camera is sensitive to light in the wavelength range of 350 to 950 nm, covering the visible light band, and has a high sampling rate of 100 frames per second. After data acquisition, the video data is processed by frame extraction to screen out 2,640 visible light images of six types of combustion states suitable for this study. In order to expand the dataset, a data enhancement method is used, and finally an enhanced dataset covering six combustion states with a total of 5,400 images is obtained, with an image size of 640×480. The principle of the counterflow burner is as follows: Figure 7 shown.

[0117] The names of the six types of combustion states in the data set are as follows: 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); this data set basically covers various combustion conditions used in combustion experiments. Selecting this data set for experiments is conducive to verifying the effectiveness of the combustion state recognition model of the present invention; the flame visible light images of the six types of combustion states are shown in Figure 2. Figure 8 As shown, in order to simplify the subsequent discussion, the following section will use the corresponding English abbreviations to refer to each combustion state.

[0118] Evaluation indicators

[0119] To further verify the effectiveness and accuracy of the model, the effect of the model in identifying combustion status is evaluated by precision (Precision, P), recall (Recall, R), F1 score (F1-score, F1), overall accuracy (Accuracy, Acc), TOP-3 accuracy (TOP-3Accuracy, TOP-3), logarithmic loss (Log Loss) and parameter quantity (Params).

[0120] Among them, precision refers to the proportion of images that actually belong to a certain combustion state when the model predicts that the image is in that state. A high precision indicates that the model has a low false recognition rate. Recall refers to the proportion of images that are correctly recognized by the model among images that actually belong to a certain combustion state. A high recall indicates that the model can better recognize all images that actually belong to a certain combustion state with fewer missed reports. F1-score refers to the harmonic mean of precision and recall, which comprehensively measures the performance of the model in the recognition of six types of combustion states. Overall accuracy refers to the proportion of images that are correctly recognized by the model to all flame images. TOP-3 accuracy refers to the proportion of the correct category in the top three most likely categories predicted by the model. Log loss refers to the difference between the state predicted by the model and the actual state. A high log loss indicates that the confidence of the model prediction is low. The specific formulas of the above evaluation indicators are shown in equations (22) to (26).

[0121]

[0122] In the formula, TP (True Positive) represents samples that belong to a certain combustion state and are correctly identified as this class, FP (False Positive) represents samples that do not belong to a certain combustion state but are mistakenly identified as this class, TN (True Negative) represents samples that do not belong to a certain combustion state and are identified as other classes, FN (False Negative) represents samples that belong to a certain combustion state but are mistakenly identified as other classes, N is the number of samples, C is the number of classes, and y is the number of samples. i,C represents the true label of the i-th sample in category C, p i,C is the probability that the model predicts category C.

[0123] Since the dataset contains six categories of flame images, it is necessary to calculate the macro average and weighted average of the precision, recall, and F1 score of each category. The macro average is a simple average of various indicators without considering the number of category samples, reflecting the balanced performance of the model in each category. The weighted average is weighted based on the number of category samples, which is closer to the performance of the model in real scenarios.

[0124] Training process:

[0125] When training the model on the dataset, the learning rate (model_lr) is set to 0.0001, the batch size (batchsize) is set to 16, the number of iterations (epochs) is set to 150 rounds, and the gradient clipping parameter is set to 5.0; in terms of data preprocessing and enhancement, random rotation of 10 degrees, Gaussian blur, adjustment of color saturation and brightness, and Mixup methods are used, followed by Resize and normalization operations; during training, the AdamW optimizer combined with L2 regularization is selected, and the cosine annealing learning rate scheduler is used to adjust the learning rate; the EMA (Exponential Moving Average) method is used to smooth the time series data, and the decay coefficient of the EMA method (model_ema_decay) is set to 0.9998 to make the model more continuous and stable. The Loss curve and Acc curve of the InceptionNeXt baseline model and the improved model are shown in Figure 2. Fig. 9 shown.

[0126] Depend on Fig. 9 From the Loss curve, we can see that the improved model converges after 80 epochs, which means that the improved InceptionNeXt model has successfully converged on the data set, and its parameters can adapt to the characteristics of the flame image and effectively complete the recognition task. At the same time, the convergence loss value of the improved model is stable at around 0.22, which is much lower than the original model, indicating that the improved model has better fitting ability and can more accurately capture the characteristics of the flame image and reduce the error. In addition, the improved model converges faster, showing its improvement in optimization efficiency. Fig. 9 From the Acc curve, we can see that the overall accuracy of the improved model is higher than that of the original model, which indicates that the feature extraction ability of the improved model has been improved.

[0127] Comparative experiment:

[0128] In order to verify the superiority of the improved InceptionNeXt model in the combustion state recognition task, other classic models were selected for comparative experiments on the same dataset, including the classic image classification and recognition models RepVgg, SeaFormer, SGFormer and VanillaNet; the improved InceptionNeXt model and other models were evaluated by the macro-average and weighted average of the precision, recall and F1 scores of each category, as well as the overall accuracy, TOP-3 accuracy, logarithmic loss and parameter quantity. All models used the same parameters for training and testing on the dataset, and other models used the original loss function. The comparative experimental results are shown in Table 2.

[0129] From the overall results of the comparative experiments, the improved model is basically superior to the comparative model in various indicators; the improved model performs outstandingly in terms of macro-average and weighted average F1 scores. Compared with the models RepVgg, SeaFormer, SGFormer and VanillaNet, the macro-average F1 scores are increased by 9.99%, 4.72%, 11.05% and 9.19% respectively, reflecting the more balanced classification performance of the improved model in each category. The weighted average F1 scores are increased by 9.93%, 4.69%, 11.00% and 9.09% respectively, reflecting that the improved model still achieves high recognition performance under the premise of considering the number of samples in each category, which verifies the effectiveness of expanding the receptive field in the model. In terms of overall classification accuracy, the improved model performs particularly well, reaching 85.71%, which is 4.76% higher than the second-best VanillaNet's 80.95%, which shows that the improved model has stronger accuracy for the combustion state recognition task and verifies the effectiveness of channel shuffling in improving performance. In terms of Top-3 accuracy, the improved model reached 95.95%, which is 2.14% higher than the second-best SeaFormer's 93.81%. This shows that even if the Top-1 recognition result of the improved model is incorrect, the correct combustion state can still be included in the Top-3 prediction, which enhances the reliability of the model and verifies the effectiveness of focusing on important information of the image. In terms of Log Loss, the improved model scored 0.4901. Compared with the models RepVgg, SeaFormer, SGFormer and VanillaNet, LogLoss was reduced by 0.1991, 0.0773, 0.2535 and 0.3506 respectively, indicating that the improved model has higher confidence in combustion state recognition. Although the improved model has a large number of parameters (256.59MB), it has shown higher accuracy and stability in multiple performance indicators. In comparison, the SeaFormer parameter size is only 6.18MB, but its precision and recall are lower than those of the improved model.

[0130] Table 3 Comparative experimental results

[0131]

[0132] In the comparative experiment, the confusion matrix of the improved model and the comparative model on the test set is as follows: Fig.10As shown in the figure; from the overall observation of the number of accurate recognitions of each model in each state (OESC, HOC, ND-HOC, LO-LMC, HNMC, BC), it can be found that the total number of correct recognitions of the improved model, that is, the sum of the diagonal elements of all state categories, is higher than that of all the comparison models, reflecting that it has a relatively stable improvement in all state categories; the number of misidentifications of the improved model is more dispersed and not concentrated in certain specific states, indicating that the performance of the model in various combustion states is relatively balanced, rather than relying on the performance of a certain category, which verifies Focal Loss is effective in the model; from the perspective of specific state categories, the improved model has achieved significant advantages in the recognition of easily confused categories (OESC and LO-LMC); the number of OESC state samples is 69, and the number of misidentified samples of the improved model is 12, 6, 14 and 10 lower than that of other models (RepVgg, SeaFormer, SGFormer and VanillaNet, in the same order below); the number of LO-LMC state samples is 68, and the number of misidentified samples of the improved model is 12, 5, 15 and 16 lower than that of other models, indicating that in the easily confused state categories, the improved model is more accurate in the recognition of subtle differences, verifying the importance of focusing on the important features of local details of flame images for performance improvement.

[0133] Ablation experiment:

[0134] In order to verify the effectiveness of the proposed improved method for the InceptionNeXt model in combustion state recognition, InceptionNeXt is used as the baseline model. Under the premise of unchanged data set, the five improved strategies of Triplet Attention, SimC, MDCPI, ShuffleCS and Focal Loss are analyzed by ablation experiment to evaluate the contribution of each module to the performance. Groups in the table indicates the serial number of the ablation experiment, "√" indicates that the improved module is used, and "×" indicates that the improved module is not used. The results of the ablation experiment are shown in Table 2.

[0135] Table 2 Ablation experiment results

[0136]

[0137] The performance of the baseline model InceptionNeXt in various evaluation indicators is used as the analysis basis for the ablation experiment; first, the F1 score of the weighted average of Triplet Attention is slightly improved, the overall accuracy remains at 83.33%, and the number of model parameters remains unchanged, verifying the effectiveness of introducing this mechanism to improve model performance; secondly, after adding SimC, Log Loss drops to 0.5495, indicating that the stability of the model has improved; then, MDCPI and Focal Loss are added, the macro-average and weighted-average F1 scores reach 84.72% and 84.52%, respectively, which are increased by 1.48% and 1.3%, respectively, the overall accuracy rises to 84.52%, and Log Loss is significantly reduced to 0.5117; This shows that MDCPI and Focal Loss are similar. The combination of Loss has a significant effect on improving the recall rate and overall recognition performance, and is more effective in reducing the confidence of recognition errors, verifying the contribution of expanding the receptive field to improving model performance. Finally, after adding the ShuffleCS module, the macro-average and weighted average F1 scores reached 85.79% and 85.82%, respectively, an increase of 1.07% and 1.30%, respectively. The overall accuracy rate increased to 85.71%, and the TOP-3 accuracy rate also increased to 95.95%. Log Loss further decreased to 0.4901, and the prediction stability and recognition performance of the model were improved. Although the number of parameters increased slightly to 256.59MB, the performance improvement significantly verified the effectiveness of inter-channel information fusion for improving model performance.

[0138] In the experiment, it was found that the linkage synergy of the MDCPI module or Focal Loss, the introduction of the MDCPI module or Focal Loss alone would reduce the recognition accuracy of the model, but the combination of the two would improve the performance. MDCPI enhances the global and local representation of flame features, but at the same time introduces redundant information, resulting in increased noise and changes in feature distribution, affecting the model effect; Focal Loss gives too high a weight to difficult-to-recognize samples in this dataset, weakening the learning of easy-to-recognize samples; therefore, when the two are combined, Focal Loss can help the model better utilize the features provided by MDCPI, focus on error-prone categories, and thus improve recognition performance.

[0139] As a combustion experimental device widely used in the field of combustion research, the accurate identification of the internal combustion state of the counterflow burner is of great significance. Aiming at the problems existing in traditional methods and existing deep learning methods, the present invention adopts InceptionNeXt as the baseline model and improves the InceptionNeXt counterflow burner combustion state recognition model. Firstly, the Triplet Attention mechanism is introduced to apply attention on the image width, height and RGB dimensions to enhance the model's adaptability to the dynamic changes of the flame. Secondly, the SimC feature focusing module is added after the Inception deep convolution of the model to make up for the defect that the Inception deep convolution loses the key information of the flame. Then, the multi-scale hole parallel fusion (MDCPI) module is embedded after the feature extraction part of the model to expand the receptive field of the model, capture a wider range of contextual information in the flame image, and work with Focal Loss to improve the recognition performance. Finally, the ShuffleCS module based on channel shuffling is added to the classification part of the model to improve the problem that the original MlpHead ignores channel interaction, and the recognition ability of MlpHead for different states is improved by feature mixing, further improving the performance of the model.

[0140] Experimental results show that the method of the present invention is effective in the combustion state recognition task, achieving an overall accuracy of 85.71%, and its performance is higher than that of the baseline model and the comparison model; however, the improved model has a large number of parameters, and the improved method of the present invention is only considered from the perspective of optimizing feature extraction, which has limitations; so the next step is to find a balance between overall accuracy and lightweight to improve recognition performance, and explore methods to optimize model performance from other perspectives.

[0141] 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. Based on the improved InceptionNeXt combustion state recognition method, it is characterized by: The following steps are involved: Step 1, collecting a combustion state image of a counter-flow burner; Step 2: Build an improved InceptionNeXt model, including: preprocessing module, feature extraction module and classification head module; The preprocessing consists of 4x4 convolution and normalization layers; the feature extraction module includes a first triple attention mechanism module, a first improved MetaNeXt module in three cascades, a second triple attention mechanism module, a first downsampling, a second improved MetaNeXt module in three cascades, a third triple attention mechanism module, a second downsampling improved MetaNeXt module in twenty-seven cascades, and a fourth triple attention mechanism module, a third downsampling and a third improved MetaNeXt module in three cascades.

2. The combustion state recognition method based on improved InceptionNeXt according to claim 1 is characterized in that: Improvements to the MetaNeXt module include replacing the normalization layer and multilayer perceptron in the MetaNeXt module with the SimC module.

3. The combustion state recognition method based on improved InceptionNeXt according to claim 2 is characterized in that: The formula for the SimC module is: In the formula, Used to determine pixel similarity, i is the pixel number, represents the variance of all neurons on a single channel, λ is the regularization term, and t i represents the i-th pixel of the input feature map on a single channel, represents the mean value of all pixels on a single channel, is the enhanced tensor, E is The sum of ⊙ is the Hadamard multiplication, C is the channel dimension, b and γ are used to change the ratio of the number of channels to the convolution kernel size, odd means an odd number, Conv1D k () is a one-dimensional convolution operation with kernel size k.

4. The combustion state recognition method based on improved InceptionNeXt according to claim 1 is characterized in that: The classification head module includes: multi-scale void parallel integration module, global average pooling, first fully connected layer, activation layer, feature expression module, normalization layer, random flexibility, and second fully connected layer.

5. The combustion state recognition method based on improved InceptionNeXt according to claim 4 is characterized in that: The multi-scale hole parallel integration module includes: multi-scale hole convolution and parallel channel and spatial attention mechanism; the formula of multi-scale hole convolution is: F1(x)=ReLU(BN(Conv2D(x, 1×1, dilation=1))) (11) F2(x)=ReLU(BN(Conv2D(x, 3×3, dilation=6))) (12) F3(x)=ReLU(BN(Conv2D(x, 3×3, dilation=12))) (13) F4(x)=ReLU(BN(Conv2D(x, 3×3, dilation=18))) (14) F5(x)=ReLU(BN(conv2D(GlobalAvgPool(x), 1×1))) (15) F concat (x)=Concat(F1(x),F2(x),F3(x),F4(x),F5(x)) (16) Among them, ReLU() is a nonlinear activation function, BN() is batch normalization, Conv2D() is a two-dimensional convolution, dilation is the void rate, GlobalAvgPool() is the global average pooling, and Concat() refers to splicing 5 tensors in dimension.

6. The combustion state recognition method based on improved InceptionNeXt according to claim 4 is characterized in that: The formula of the feature expression module is: First, calculate the channel attention mechanism, the formula is: T c =Sigmoid(σ(W avg ·F avg +b avg )+σ(W max ·F max +b max ))⊙T (17) Where, T c is the weighted feature map; Sigmoid() and σ are activation functions; W avg and W max is the weight of the fully connected layer; F avg is the global average pooling, F max is the global maximum pooling; b avg and b max is the bias term; ⊙ is the element-by-element multiplication operation; Secondly, perform channel shuffle operation, the formula is: T s =Shuffle(T c ) (18) Where, T s is the shuffled feature map; Shuffle() shuffles the grouped channels; T c It is the feature map after channel attention weighting; Finally, the shuffled feature map is processed and output using spatial attention. The formula for spatial attention is: T s =Sigmoid(W a ·[AvgPool(T s );MaxPool(T s )]+b a )⊙T s (19) Where, T s ′ is the feature map after spatial attention processing; Sigmoid() is the activation function; W a is the convolution kernel; AvgPool(T s ) is the input feature map T s The global average pooling operation in the channel dimension, MaxPool(T s ) is the input feature map T s Global maximum pooling operation in the channel dimension; b a is the bias term; ⊙ is the element-by-element multiplication operation; T s is the input feature map.

7. The combustion state recognition method based on improved InceptionNeXt according to claim 1 is characterized in that: The loss functions include: soft target cross entropy loss function.

8. The combustion state recognition method based on improved InceptionNeXt according to claim 1 is characterized in that: The loss function also includes: Focal Loss loss function.

9. The combustion state recognition method based on improved InceptionNeXt according to claim 1 is characterized in that: The combustion states include: oxygen-rich stable combustion, high oxygen combustion, nitrogen-diluted high oxygen combustion, low oxygen and low methane combustion, high nitrogen-diluted medium combustion and balanced combustion.

10. Improved InceptionNeXt combustion state recognition system, characterized by: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the improved InceptionNeXt combustion state recognition method as described in any one of claims 1 to 9.

Images