A rotary kiln combustion image classification method based on color correction and convolutional network

Abstract

The present invention relates to a rotary kiln combustion image classification method based on color correction and convolutional network, which belongs to the field of kiln image analysis technology, including obtaining and processing flame image data sets; introducing color correction methods and convolutional networks to build a network model; using the obtained image data sets as input to finally generate a three-dimensional prediction; taking the combustion state with the highest prediction score as the combustion state in the rotary kiln at this time, and outputting the final result. The kiln combustion process is a highly complex nonlinear process, with unstable combustion, harsh environment and outdated shooting equipment. It aims to solve the problems of poor visual effects of kiln flame images, whitish image exposure, unclear colors, etc. The existing technology is often ineffective in classifying kiln images. The present invention can improve the quality of recognition and the accuracy of classification.

Description

Technical Field

[0001] The invention relates to a rotary kiln combustion image classification method based on color correction and convolutional network, belonging to the technical field of kiln image analysis. Background Art

[0002] The rotary kiln is a building material equipment used in industrial production processes, mainly used to calcine cement clinker, alumina and other industrial raw materials. The factory fire watcher controls the combustion state of the rotary kiln by adjusting the oxygen supply and fuel amount of the burner, and analyzes the temperature,

[0003] Parameters such as oxygen content are used to determine changes in the combustion state and to promptly adjust combustion control parameters. The stability of the combustion state within a rotary kiln has a significant impact on sintering quality, energy consumption, and pollutant emissions. With increasing environmental awareness, many researchers are working to optimize combustion control and improve fuel selection to reduce emissions and enhance the environmental performance of rotary kilns. Comprehensive consideration and analysis of multiple factors, and in-depth research on the combustion state of rotary kilns, can significantly contribute to improving product quality, reducing energy consumption, and improving environmental performance. Typically, workers install cameras within the kiln flame to capture flame images and observe the combustion state. However, due to the limited photography conditions within rotary kilns, flame image data is often difficult to obtain. Photography within rotary kilns is significantly affected by factors such as fuel type, camera type, and background color, resulting in defects such as color cast and image blur. Some flame images obtained show little difference in characteristics, making them visually difficult to distinguish. Therefore, accurately determining the combustion state of rotary kilns remains a major challenge. Rotary kiln combustion states are generally categorized into three types: normal combustion, under-combustion, and over-combustion. The earliest rotary kiln appeared in cement production. Since then, due to its good thermal conductivity and strong mixing ability, the rotary kiln has been widely used in industrial production fields such as power generation, metallurgy and cement, which has promoted the development of the industrial industry. However, due to the complexity of the rotary kiln structure and the nonlinearity of combustion, the combustion status in the kiln often relies on "manual monitoring". This method leads to waste of human resources and an increased risk of misjudgment. In the 1970s, CCD cameras began to be used in the industrial field, and research on visual flame monitoring began at home and abroad. In 1989, Tsinghua University first studied the relationship between small flame temperature distribution and image brightness, and obtained a polynomial regression model. It began to analyze the burning status with the burning zone flame image as the research object. Flame image processing has since attracted widespread attention from domestic scholars. Coal-fired flame monitoring is mostly used in industrial boilers, mainly by analyzing the combustion flame image data to achieve CO and NO X Identification of emissions and coal types.

[0004] With the continuous development of the computer science field, a combustion condition recognition method based on a generalized learning vector neural network (GNN) has been proposed based on the texture features of pulverized coal combustion flame images during the oxide pellet sintering process in a rotary kiln. Simultaneously, numerous methods for extracting flame image features have been proposed and studied. In particular, three luminescence features and four dynamic features were extracted from the flame region using a series of blurred flame images for temperature detection in rotary kilns. However, these methods require extensive preprocessing and manual parameter adjustment. In recent years, deep learning has become a hot topic and has been widely adopted and successfully applied in many fields. For example, a deep learning-based rotary kiln combustion state monitoring system, through an end-to-end network, eliminates the complex procedures of traditional feature extraction methods. Furthermore, a proposed convolutional recurrent neural network (CRNN) can effectively extract features from flame image sequences to predict the combustion state within the rotary kiln. Practice has proven that deep learning methods can more quickly and accurately detect the combustion state of rotary kiln flames. Existing techniques for classifying kiln images often perform poorly due to the harsh environments in which kiln images are captured. Compared to traditional networks, the classification accuracy is improved due to the use of deeper networks. The expressiveness is increased by using an attention mechanism to focus on important features and suppress unnecessary features. Summary of the Invention

[0005] In response to the above-mentioned problems of the prior art, the present invention provides a rotary kiln combustion image classification method based on color correction and convolutional network, which is more suitable for factory rotary kiln combustion images.

[0006] The present invention is achieved through the following technical solution, which specifically includes the following steps:

[0007] Step 1: Use a CCD camera (charge-coupled device camera, suitable for industrial cameras) to capture factory flame combustion videos and process them to obtain a rotary kiln flame image dataset;

[0008] Step 2: Introduce the two-dimensional gamma color correction method to process the brightness of the rotary kiln image;

[0009] Step 3: Import the corrected image into the convolutional classification network model based on the efficient neural network architecture EfficientNetV2 to obtain the processed image and feature map;

[0010] Step 4: Classify and predict the feature map, classify the combustion state through the softmax classifier, and output the final combustion state prediction classification result.

[0011] The flame image dataset obtained by the processing described in step 1 includes:

[0012] S1.1: Extract the flame burning video captured by the CCD camera into a continuous image sequence frame by frame and decompose it into an RGB image dataset of size 704×576;

[0013] S1.2, divide the image data set into three states: under-combustion, normal combustion, and over-combustion;

[0014] S1.3, divide the image dataset in S1.2 into a training dataset and a test dataset.

[0015] The two-dimensional gamma color correction method in step 2 includes converting the image from RGB format, i.e., an image format represented by three channels of red (R), green (G), and blue (B), into a YUV format for representing a color image, wherein the image is divided into three components: brightness (Y), chromaticity (blue difference) (U), and chromaticity (red difference) (V). The furnace image is captured by a camera and then decoded, so the original R, G, B image is converted into brightness (Y) and color saturation (U, V), and then gamma correction is performed on the brightness (Y).

[0016] The convolutional network classification model in step 3 primarily consists of the mobile inverted bottleneck convolution (MBConv) module and the coordinate attention and mobile inverted bottleneck convolution (CoFe-MBConv) module from the EfficientNetV2 convolutional network. The CoFe-MBConv module replaces the original shallow Fused-MBConv module in EfficientNetV2 to create the CoFe-EfficientNetV2 model. This new network not only considers important channel features but also spatial features and collects critical positional information.

[0017] The mobile flip bottleneck convolution CoFe-MBConv module with the added attention mechanism includes embedding the coordinate attention mechanism into the original Fused-MBConv module of EfficientNetV2, and naming the new module CoFe-MBConv.

[0018] The softmax classifier in step 4 is a commonly used classifier, which is usually used for multi-class classification problems. It implements the classification task based on the activation function of the softmax function.

[0019] In a softmax classifier, the features of the input sample undergo a linear transformation and are then normalized using the softmax function. The softmax function converts the scores for each category into probabilities, such that the sum of the probabilities for all categories is 1. Ultimately, the model selects the category with the highest probability as the prediction.

[0020] Take the feature map in step 2 as the output result and pass it through the fully connected layer;

[0021] After step 3, a three-dimensional prediction is generated, corresponding to the three combustion states of under-combustion, normal combustion, and over-combustion. After the feature layer is fully connected, it passes through the softmax classifier. The softmax function converts the input into a probability with a categorical distribution; finally, the predicted probabilities of the three combustion categories are output. The combustion state with the highest predicted probability is the combustion state in the rotary kiln at that time.

[0022] The beneficial effects of the present invention are as follows: color correction is selected in combination with the improved convolutional network EfficientNetV2 as the basic network. The improved network combines the attention mechanism to update the shallow basic module, enabling the network model to obtain richer feature information in the image, better distinguish between flame images that look similar, and further improve network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the present invention.

[0024] Figure 2 Schematic diagram of the overall network structure of the method of the present invention.

[0025] Figure 3 This is a comparison diagram of the flame image of the present invention and the original flame image.

[0026] Figure 4 Schematic diagram of the two-dimensional gamma color correction network structure.

[0027] Figure 5 This is the structural diagram of MBConv and CoFeMBConv.

[0028] Figure 6 This is the prediction process diagram of the softmax classifier. DETAILED DESCRIPTION

[0029] In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention; in addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The present invention will be described in more detail below with reference to the accompanying drawings. In each of the drawings, the same elements are represented by similar figure marks. For the sake of clarity, the various parts in the drawings are not drawn to scale.

[0030] The following is combined with Figures 1 to 6 To describe the present rotary kiln combustion image classification method based on color correction and convolutional network, as Figure 1 As shown, the present invention provides a rotary kiln combustion image classification method based on color correction and convolutional network to classify rotary kiln flame combustion pictures. The specific steps are as follows:

[0031] Step 1: Use a CDD camera (charge-coupled device camera, suitable for industrial cameras) to capture factory flame combustion videos and process them to obtain a rotary kiln flame image dataset;

[0032] Step 2: Introduce the two-dimensional gamma color correction method to process the brightness of the rotary kiln image;

[0033] Step 3: Import the corrected image into the convolutional classification network model based on the efficient neural network architecture EfficientNetV2 to obtain the processed image and feature map;

[0034] Step 4: Classify and predict the feature map, classify the combustion state through the softmax classifier, and output the final combustion state prediction classification result.

[0035] From steps 1 to 4 above, it's clear that in the image classification process of the present invention, the classification model is based on a combination of color correction and a convolutional network. Using a softmax classifier, it ultimately outputs three burn category prediction scores. The one with the highest classification accuracy is ultimately used as the final result. By incorporating an attention mechanism into the EfficientNetV2 convolutional network module, the overall network framework's feature extraction capabilities are enhanced.

[0036] The flame image dataset obtained by the processing described in step 1 includes:

[0037] S1.1: Extract the flame combustion video recorded and captured by the CCD camera into a continuous image sequence frame by frame. The size of each image is 704×576, and the image size is adjusted to 128×128×3;

[0038] S1.2, experienced kiln workers mark and divide the image data set into three states: under-burning, normal burning, and over-burning; Figure 6 Shown are three typical examples of combustion conditions;

[0039] S1.3. Divide the image dataset in S1.2 into a training dataset and a test dataset. Extract multiple spatiotemporally continuous images from the video frame by frame as the training set, totaling 10,000 images (including 1,927 overburned samples, 5,275 normal combustion samples, and 2,798 underburned samples). Similarly, extract 5,000 spatiotemporally continuous images from the video frame by frame as the test set (1,134 samples for overburned states, 2,352 samples for normal combustion states, and 1,514 samples for underburned states).

[0040] like Figure 2 As shown in the figure, the two-dimensional gamma color correction method is first used to process the input image, and the processed image is then input into the improved EfficientNetV2 network to obtain the feature map, and finally classification is performed.

[0041] The two-dimensional gamma color correction method in step 2 includes converting the image from RGB format to YUV format, that is, converting the image format represented by three channels of red (R), green (G), and blue (B) into a color coding format for representing color images, wherein the image is divided into three components: brightness (Y), chromaticity blue difference (U), and chromaticity red difference (V). The furnace image is captured by a camera and then decoded, so the original R, G, B image is converted into brightness (Y) and color saturation (U, V), and then gamma correction is performed on the brightness (Y).

[0042] The principle of the present invention is that the advantage of converting a picture into the YUV format is that it separates the brightness and color information of the image. Since the human eye is more sensitive to brightness, the brightness information can be compressed more efficiently, while the chromaticity information can be compressed less, thereby achieving a better compression rate. YUV divides the brightness (Y) and chromaticity (U, V) of the image into independent components, which allows different processing strategies to be adopted for different components when processing the image. The brightness component contains the grayscale information of the image, while the chromaticity component contains color information. By separating the brightness and chromaticity, the image can be better processed and optimized. The YUV format is more robust to some image processing operations than RGB. For example, when adjusting the brightness or contrast of an image, only the value of the Y channel needs to be adjusted, while adjusting the U and V channels will not affect the brightness information of the image, which can more accurately control the performance of the image. Figure 3 The following image shows a comparison of image data before and after processing. The original kiln flame image exhibits poor visual quality, with a whitish overexposure and unclear colors. Color correction preprocessing corrects these uneven colors, significantly improving the visual quality of the image in terms of edge segmentation, background segmentation, and color processing. In YUV space, the image better preserves color sharpness and captures a wealth of features and details when segmenting the flame and its background, resulting in clear edges. The color-corrected image also features a more defined shape and color in the flame's core (the whitish and yellowish center). This eliminates irrelevant artifacts in subsequent network classification, improving classification accuracy.

[0043] like Figure 4 The figure shows the network structure diagram of two-dimensional gamma color correction. First, the kiln flame image is input, the RGB format rotary kiln flame image is converted to YUV format, and then the brightness (Y) is gamma corrected separately. Finally, the YUV format is converted to RGB format for output.

[0044] The convolutional network classification model in step 3 primarily consists of the mobile inverted bottleneck convolution (MBConv) module and the coordinate attention and mobile inverted bottleneck convolution (CoFe-MBConv) module from the EfficientNetV2 convolutional network. The CoFe-MBConv module replaces the original shallow Fused-MBConv module in EfficientNetV2 to create the CoFe-EfficientNetV2 model. This new network not only considers important channel features but also spatial features and collects critical positional information.

[0045] The mobile flip bottleneck convolution CoFe-MBConv module with the added attention mechanism includes embedding the coordinate attention mechanism into the original Fused-MBConv module of EfficientNetV2, and naming the new module CoFe-MBConv. Figure 5 As shown in the figure, the first input part passes through a 3×3 convolution, four times per channel. It then passes through the CoFe attention mechanism, followed by a 1×1 convolution. Finally, it is fused with the input of the previous layer.

[0046] The softmax classifier in step 4 is a commonly used classifier, often used for multi-class classification problems. It implements the classification task based on the softmax activation function. In the softmax classifier, the features of the input sample undergo a linear transformation and are then normalized using the softmax function. The softmax function converts the scores for each class into probabilities, such that the sum of all class probabilities is 1. Ultimately, the model selects the class with the highest probability as the prediction.

[0047] Take the feature map in step 3 as the output result and pass it through the fully connected layer;

[0048] The fully connected layer is used to learn the high-level features of the convolutional layer output, that is, the possible nonlinear functional relationships. After a series of training, the model is able to distinguish the main features and some low-level features in the image and classify them using the softmax classifier.

[0049] like Figure 3The following are examples of original flame images of three typical combustion states: normal combustion, overburning, and underburning, from left to right. The flame image of the overburning state has a clear flame core, while the underburning state has no flame core.

[0050] Step 4 generates a three-dimensional prediction corresponding to the three combustion states of underburn, normal combustion, and overburn. After fully connecting the feature layer, the prediction passes through a softmax classifier, which converts the input into probabilities with a categorical distribution. The final output is the predicted probabilities for the three combustion categories. The combustion state with the highest predicted probability is the combustion state in the rotary kiln at that time.

[0051] like Figure 6 The figure shows the prediction process of the Softmax classifier. If the probability of an input belonging to a particular class is greater than the probability of belonging to other classes, the value corresponding to that class approaches 1, while the values ​​of other classes approach 0. Softmax converts the logit (the numerical output of the last linear layer of a multi-class classification neural network) into probabilities by taking the exponential of each output. Each number is then normalized by the sum of these exponentials, so that the sum of all probabilities equals 1. Cross-entropy loss is typically used as the loss function for this type of multi-class classification problem. Softmax is often added to the last layer of an image classification network. A softmax classifier is used to predict the combustion state, ultimately outputting prediction scores for three combustion categories. For example, the trained model might infer that an image of an overburned flame has an 80% probability of representing overburn, but a 10% probability of representing normal combustion (because normal and overburned flames have similar flame cores), making the probability of representing other numbers even lower. Therefore, the value corresponding to the maximum probability, 80%, is taken to indicate that the combustion state in the kiln at that moment is overburned.

[0052] This invention improves the feature extraction performance of the EfficientNetV2 model by combining a color correction method with an EfficientNetV2 network with an added attention mechanism. The kiln combustion process is a highly complex, nonlinear process characterized by unstable combustion, harsh environments, and outdated camera equipment. This invention addresses issues such as poor visual quality, whitish image exposure, and unclear colors in kiln flame images. The rotary kiln combustion image classification method based on color correction and a convolutional network is particularly suitable for classifying factory rotary kiln combustion images.

[0053] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

Claims

1. A rotary kiln combustion image classification method based on color correction and convolutional network, characterized in that: The following steps are involved: Step 1: Use a CCD camera to capture factory flame combustion videos and process them to obtain a rotary kiln flame image dataset; Step 2: Introduce the two-dimensional gamma color correction method to process the brightness of the rotary kiln image; The two-dimensional gamma color correction method involves converting an image from RGB format to YUV format. That is, converting an image format represented by three channels (red, green, and blue) into a color-coded format for color images, where the image is divided into three components: luminance (Y), chrominance (blue difference, U), and chrominance (red difference, V). The furnace image is captured by a camera and then decoded. Therefore, the original R, G, and B image is converted to a luminance (Y) and saturation (U, V) format, and then gamma correction is performed on the luminance (Y). Step 3: Import the corrected image into the convolutional classification network model based on the efficient neural network architecture EfficientNetV2 to obtain the processed image and feature map; The convolutional network classification model mainly includes the moving flip bottleneck convolution module in the EfficientNetV2 convolutional network, namely MBConv; and the moving flip bottleneck convolution module with attention mechanism, namely CoFe-MBConv, including embedding the coordinate attention mechanism into the Fused-MBConv module of EfficientNetV2; Step 4: Classify and predict the feature map, classify the combustion state through the softmax classifier, and output the final combustion state prediction classification result.

2. The rotary kiln combustion image classification method based on color correction and convolutional network according to claim 1 is characterized in that The flame image dataset obtained by the processing described in step 1 includes: S1.1: Extract the flame burning video captured by the CCD camera into a continuous image sequence frame by frame and decompose it into an RGB image dataset of size 704×576; S1.2, divide the image data set into three states: under-combustion, normal combustion, and over-combustion; S1.3, divide the image dataset in S1.2 into a training dataset and a test dataset.

3. The rotary kiln combustion image classification method based on color correction and convolutional network according to claim 1 is characterized in that: The softmax classifier in step 4 is an activation function based on the softmax function to implement the classification task; In the softmax classifier, the features of the input sample are linearly transformed and then normalized by the softmax function; The softmax function converts the score of each category into a probability value so that the sum of the probabilities of all categories is 1; Ultimately, the model selects the class with the highest probability value as the prediction result.

4. The rotary kiln combustion image classification method based on color correction and convolutional network according to claim 1 is characterized in that: The classification prediction in step 4 includes: The feature map in step 3 is used as the output result and passed through the fully connected layer to generate a three-dimensional prediction corresponding to the three combustion states of under-combustion, normal combustion and over-combustion. After the feature layer is fully connected and passed through the softmax classifier, the softmax function converts the input into a probability with a classification distribution; finally, the predicted probabilities of the three combustion categories are output, and the combustion state with the highest predicted probability is the combustion state in the rotary kiln at that time.

Images