Coal ash prediction method and system based on deep learning

Abstract

The invention discloses a coal ash prediction method and system based on deep learning. The method comprises the following steps: collecting a CT image data set of a coal sample; preprocessing the CT image data set to enhance image features; s2, constructing an MSA-ViT model, inputting the image preprocessed in the step S2, extracting features through a 3D convolution layer, an SPP layer and two different encoders, and outputting a prediction result; and analyzing a prediction result, and introducing an error index for evaluation. The problems of low efficiency, insufficient accuracy, dependence and high cost in the aspect of coal ash detection in the prior art are effectively solved, and an efficient, accurate and convenient coal ash detection solution is provided for users.

Description

Technical Field

[0001] The present invention relates to the technical field of coal ash data processing, and in particular to a coal ash prediction method and system based on deep learning. Background Art

[0002] The ash content of coal is one of the important indicators for evaluating coal quality. Coal with high ash content will produce more waste residue during combustion, reduce combustion efficiency and increase environmental pollution. Therefore, accurate prediction of coal ash content is of great significance for coal procurement, use and environmental protection control.

[0003] Traditional coal ash detection methods mainly include laboratory analysis and on-site rapid detection. Laboratory analysis usually uses high-temperature ignition loss method or chemical analysis method, which are highly accurate but require a long experimental operation and are costly. On-site rapid detection methods such as X-ray ash analyzers are easy to operate, but are greatly affected by coal types and external conditions and have limited accuracy.

[0004] In recent years, with the development of computed tomography (CT) technology, it has become possible to predict the physical properties of coal through CT images. CT scanning can obtain the internal structure information of coal samples non-destructively, and then analyze the density, porosity and other parameters of coal. These parameters have a certain correlation with the ash content of coal, so the ash content of coal can be indirectly predicted by analyzing CT images.

[0005] However, most existing coal ash prediction methods based on CT images rely on manual feature extraction and traditional machine learning models, which have limitations in processing high-dimensional image data and capturing complex nonlinear relationships in the data. In addition, traditional machine learning models also need to be improved in terms of model generalization ability and prediction accuracy.

[0006] Therefore, it is necessary to develop a new method that can effectively utilize CT image data, capture the complex characteristics of the data through a deep learning model, and improve the accuracy and efficiency of coal ash prediction. Summary of the invention

[0007] In order to solve the above-mentioned technical difficulties in the field of coal ash detection, the present invention provides a coal ash prediction method and system based on deep learning, which can predict the ash content of coal samples by analyzing CT images.

[0008] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] A coal ash prediction method based on deep learning includes the following steps:

[0010] S1: Collect CT image dataset of coal samples;

[0011] S2: Preprocess the CT image dataset to enhance image features;

[0012] S3: Build the MSA-ViT model, input the image preprocessed in step S2, extract features through the 3D convolution layer, SPP layer, and two different encoders, and output the prediction results;

[0013] S4: Analyze the prediction results and introduce error indicators for evaluation.

[0014] Preferably, step S1 specifically includes: using an industrial CT scanner to obtain CT image data sets of coal samples of different coal types and origins according to set resolutions and parameter settings.

[0015] Preferably, step S2 specifically includes:

[0016] S21: De-noising the CT image to reduce the noise level in the image so as to better extract the characteristics of the coal sample;

[0017] S22: Image enhancement technology is used to improve the contrast and clarity of the image, making the texture features of the coal sample more obvious;

[0018] S23: Crop the original CT image into small blocks of the same size to eliminate irrelevant information at the edge of the image and reduce the amount of calculation;

[0019] S24: standardized cropped image;

[0020] S25: Use rotation, flipping, and scaling data augmentation methods on the preprocessed images to expand the data set;

[0021] S26: Segment the preprocessed image into several regions, and extract texture and shape features of each region respectively.

[0022] Preferably, step S3 specifically includes:

[0023] S31: Receive the preprocessed CT image data, select the dataset images as the training set, validation set and test set in a ratio of 6:3:1; random sampling is used in each process to ensure the distribution of the data and avoid data bias affecting the generalization ability of the model;

[0024] S32: The image is passed through multiple 3D convolutional layers to extract the initial feature map. The convolutional layer uses ReLU as the activation function for nonlinear transformation to enhance the expression ability of the model. The output of the convolutional layer is used as input to the SPP layer.

[0025] S33: The SPP layer is applied after the convolution layer to divide the feature map into different regions, and each region is pooled at different scales, and the maximum pooling operation is used to retain the main information of the image features;

[0026] S34: The output of the SPP layer is used as input to two ViT encoders, one of which introduces the MSA mechanism to enhance the model's attention to key features. The MSA mechanism calculates the similarity between Query, Key, and Value and uses the Softmax function to normalize the similarity to determine the attention weight. The attention weight is used to generate a new feature representation by weighted summation, thereby enhancing the model's ability to understand the image.

[0027] S35: Fusion of the outputs of the two encoders ensures that the model can combine global and local features when making predictions;

[0028] S36: The encoder output passes through a fully connected layer to convert the result into the final coal ash estimate.

[0029] Preferably, step S4 specifically includes:

[0030] S41: Introduce error indicators MAE, RMSE, and R2 for error analysis and result visualization;

[0031] S42: Setting different prediction scenario models and comparing the prediction performance of the models under different prediction scenarios.

[0032] Preferably, in step S32, an LBP method is used to extract texture features; the LBP method generates texture features by comparing the value of each pixel with the values ​​of its surrounding pixels:

[0033] LBP(x,y,P,R)=sgn(I(x,y)-I(x+i,y+j)) (1)

[0034] Where x, y are the positions of the center pixel, P is the number of pixels, R is the radius, I(x, y) is the value of the center pixel, and the values ​​of the surrounding pixels are represented by I(x, y)-I(x+i, y+i); in the feature extraction process, a center pixel (x, y) is selected, and then a radius R and a number of pixels P are selected; then, the value of the center pixel I(x, y) is compared with the values ​​of the surrounding P pixels I(x, y)-I(x+i, y+i), where i and j are integers from -1 to 1; by comparison, a sign function value is obtained, which represents the relative size relationship between the center pixel and its surrounding pixels; these sign function values ​​are combined to obtain an LBP feature vector, which represents the texture characteristics of a specific area in the image;

[0035] The SC method is used to extract shape features, and its expression is:

[0036]

[0037] The distance between the central pixel and all other pixels in the image is calculated and these distance values ​​are weighted summed, where the weight is Indicates the similarity between pixels; in this way, a shape context feature vector is obtained, which can represent the shape characteristics of a specific area in the image.

[0038] Preferably, the detailed parameters of the 3D convolutional neural network and SPP layer of step S33 are: the input size is 512×512×512; the hidden layer contains two convolutional layers, the first convolutional layer contains 32 3×3×3 convolutional kernels, and the second convolutional layer contains 64 3×3×3 convolutional kernels, with a step size of 1; the SPP layer contains three pooling windows of 1×1×1, 2×2×2, and 3×3×3; the training batch is 16, the number of neural network training times is 1000, and the learning rate is 0.001.

[0039] Preferably, in step S34, a CT image I is provided, wherein the value of each pixel i is represented as I i ; Divide the image into multiple image blocks B m , where m is the index of the image block; each image block B m Contains N pixels; for each image block B m Calculate the similarity between pixels:

[0040]

[0041] Among them, B m and B n are two different image patches, W Q is the query weight matrix, d k is the dimension of the key; calculate the image block B m The attention representation is:

[0042]

[0043] Where A(m) is the image block B m The attention representation is represented by M, and M is the total number of image blocks. By combining the self-attention mechanism with the encoder, the model’s ability to capture spatial relationships in CT images is enhanced, thereby improving the model’s ability to understand CT images.

[0044] The present invention also provides a coal ash prediction system based on deep learning, comprising:

[0045] An image acquisition module, used to collect CT image data sets of coal samples;

[0046] An image preprocessing module, used for preprocessing the CT image data set;

[0047] Model training module, used to build and train the MSA-ViT model;

[0048] The prediction module is used to input the CT image data to be predicted, perform prediction through the trained model, and output the predicted value of coal ash.

[0049] Preferably, it also includes a user interface for receiving CT image data input by a user and displaying prediction results.

[0050] The present invention uses deep learning technology to achieve rapid prediction of coal ash content and significantly improve detection efficiency. At the same time, the accuracy of prediction is improved by optimizing the model training and verification process. In addition, the present invention reduces the dependence of detection on professionals by developing a user-friendly interface and realizing seamless integration of system modules, thereby improving the flexibility and scalability of detection. Finally, by reducing dependence on professional equipment and personnel, the cost of detection is reduced. In general, the present invention effectively solves the problems of low efficiency, lack of accuracy, dependence and high cost in the existing technology in coal ash detection, and provides users with an efficient, accurate and convenient coal ash detection solution.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention aims to improve the efficiency and accuracy of coal ash detection. Through deep learning technology, fast and accurate prediction of coal ash is achieved, greatly improving detection efficiency and accuracy. The present invention adopts comprehensive data preprocessing and feature extraction technology to fully explore and utilize useful information in CT image data, and provide high-quality data support for model training. The present invention adopts advanced model training strategies, including adaptive learning rate adjustment strategy, appropriate loss function, data enhancement, regularization, hyperparameter tuning and other technologies to improve the performance and generalization ability of the model. During the model training process, the present invention uses a validation set and a test set to evaluate and test the model to ensure the generalization ability and prediction performance of the model, thereby ensuring the reliability and accuracy of the model in practical applications. At the same time, the present invention has also developed a user-friendly interface, allowing users to upload CT images, view the prediction results of coal density, and obtain the predicted value of coal ash based on these results, so that users can operate and use it conveniently. In addition, the present invention also realizes the seamless integration of the user interface with other modules of the system, provides a one-stop coal ash detection service, and provides users with a comprehensive and convenient solution. In general, the present invention, through its innovative technical solution, effectively solves the shortcomings of the existing technology in coal ash detection, improves detection efficiency and accuracy, and provides users with an efficient, accurate and convenient coal ash detection solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 This is a flow chart of the coal ash prediction method based on deep learning of the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of the coal ash prediction system based on deep learning in the present invention. DETAILED DESCRIPTION

[0056] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0057] like Figure 1 As shown, a coal ash prediction method based on deep learning includes the following steps:

[0058] S1: Collect CT image datasets of coal samples: Use an industrial CT scanner to obtain CT image datasets of coal samples of different coal types and origins according to the set resolution and parameter settings.

[0059] S2: Preprocess the CT image dataset to enhance image features;

[0060] Specifically include:

[0061] S21: De-noising the CT image by filtering to reduce the noise level in the image so as to better extract the characteristics of the coal sample;

[0062] S22: Image enhancement technology is used to improve the contrast and clarity of the image, making the texture features of the coal sample more obvious;

[0063] S23: Crop the original CT image into small blocks of the same size to eliminate irrelevant information at the edge of the image and reduce the amount of calculation;

[0064] S24: standardized cropped image;

[0065] S25: Use rotation, flipping, and scaling data augmentation methods on the preprocessed images to expand the data set;

[0066] S26: Divide the preprocessed image into several regions, and extract the texture and shape features of each region respectively, which are used to verify the preliminary feature map extracted by the 3D convolution layer in step S33.

[0067] S3: Build the MSA-ViT model, input the image preprocessed in step S2, extract features through the 3D convolution layer, SPP layer, and two different encoders, and output the prediction results;

[0068] Specifically include:

[0069] S31: Receive the preprocessed CT image data, select the dataset images as the training set, validation set and test set in a ratio of 6:3:1; random sampling is used in each process to ensure the distribution of the data and avoid data bias affecting the generalization ability of the model;

[0070] S32: The image is passed through multiple 3D convolutional layers to extract the initial feature map. The convolutional layer uses ReLU as the activation function for nonlinear transformation to enhance the expression ability of the model. The output of the convolutional layer is used as input to the SPP layer:

[0071] The Local Binary Patterns (LBP) method is used to extract texture features. The LBP method generates texture features by comparing the value of each pixel with the values ​​of its surrounding pixels:

[0072] LBP(x,y,P,R)=sgn(I(x,y)-I(x+i,y+j)) (1)

[0073] Where x, y are the positions of the center pixel, P is the number of pixels, R is the radius, I(x, y) is the value of the center pixel, and the values ​​of the surrounding pixels are represented by I(x, y)-I(x+i, y+i); in the feature extraction process, a center pixel (x, y) is selected, and then a radius R and a number of pixels P are selected; then, the value of the center pixel I(x, y) is compared with the values ​​of the surrounding P pixels I(x, y)-I(x+i, y+i), where i and j are integers from -1 to 1; by comparison, a sign function value is obtained, which represents the relative size relationship between the center pixel and its surrounding pixels; these sign function values ​​are combined to obtain an LBP feature vector, which represents the texture characteristics of a specific area in the image;

[0074] The shape context (SC) method is used to extract shape features, and its expression is:

[0075]

[0076] The distance between the central pixel and all other pixels in the image is calculated and these distance values ​​are weighted summed, where the weight is Indicates the similarity between pixels; in this way, a shape context feature vector is obtained, which can represent the shape characteristics of a specific area in the image.

[0077] S33: The SPP layer is applied after the convolution layer to divide the feature map into different regions, and each region is pooled at different scales, and the maximum pooling operation is used to retain the main information of the image features; the detailed parameters of the 3D convolutional neural network and the SPP layer are: the input size is 512×512×512; the hidden layer contains two convolutional layers, the first convolutional layer contains 32 3×3×3 convolutional kernels, and the second convolutional layer contains 64 3×3×3 convolutional kernels with a step size of 1; the SPP layer contains three pooling windows of 1×1×1, 2×2×2, and 3×3×3; the training batch is 16, the number of neural network training times is 1000, and the learning rate is 0.001.

[0078] S34: The output of the SPP layer is used as input to two ViT encoders, one of which introduces the MSA mechanism to enhance the model's attention to key features. The MSA mechanism calculates the similarity between Query, Key, and Value and uses the Softmax function to normalize the similarity to determine the attention weight. The attention weight is used to generate a new feature representation by weighted summation, thereby enhancing the model's ability to understand the image. Suppose there is a CT image I, where the value of each pixel i is represented by I i ; Divide the image into multiple image blocks B m , where m is the index of the image block; each image block B m Contains N pixels; for each image block B m Calculate the similarity between pixels:

[0079]

[0080] Among them, B m and B n are two different image patches, W Q is the query weight matrix, d k is the dimension of the key; calculate the image block B m The attention representation is:

[0081]

[0082] Where A(m) is the image block B m The attention representation is represented by M, and M is the total number of image blocks. By combining the self-attention mechanism with the encoder, the model’s ability to capture spatial relationships in CT images is enhanced, thereby improving the model’s ability to understand CT images.

[0083] In step S34, two encoders are used to extract multi-scale features of the image respectively; the first encoder is a common ViT encoder, which includes a self-attention layer and a feedforward network layer. The multi-head attention layer contains 8 attention heads, which are used to calculate the similarity between the Query, Key and Value vectors, and normalize the similarity using the softmax function to determine the attention weight; the attention weight is used for weighted summation to generate a new feature representation; the feedforward network layer contains two linear layers, each with 2048 neurons; the second encoder applies the self-attention mechanism to the ViT encoder to improve the model's ability to focus on key features; the output layer is used to predict coal ash content, and contains a fully connected layer with 1 output node; finally, the outputs of the two encoders are fused to generate the final prediction feature representation for predicting coal ash content.

[0084] S35: Fusion of the outputs of the two encoders ensures that the model can combine global and local features when making predictions;

[0085] S36: The encoder output passes through a fully connected layer to convert the result into the final coal ash estimate.

[0086] S4: Analyze the prediction results and introduce error indicators for evaluation; specifically include:

[0087] S41: Introduce error indicators MAE, RMSE, and R2 for error analysis and result visualization;

[0088] S42: Setting different prediction scenario models and comparing the prediction performance of the models under different prediction scenarios.

[0089] The present invention also provides a coal ash prediction system based on deep learning, comprising:

[0090] An image acquisition module, used to collect CT image data sets of coal samples;

[0091] An image preprocessing module, used for preprocessing the CT image data set;

[0092] Model training module, used to build and train the MSA-ViT model;

[0093] The prediction module is used to input the CT image data to be predicted, make predictions through the trained model, and output the predicted value of coal ash;

[0094] The invention also comprises a user interface for receiving CT image data input by a user and displaying the prediction result.

[0095] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A coal ash prediction method based on deep learning, characterized in that: The steps include: S1: Collect CT image dataset of coal samples; S2: Preprocess the CT image dataset to enhance image features; S3: Build the MSA-ViT model, input the image preprocessed in step S2, extract features through the 3D convolution layer, SPP layer, and two different encoders, and output the prediction results; S4: Analyze the prediction results and introduce error indicators for evaluation.

2. The coal ash prediction method based on deep learning according to claim 1 is characterized in that: Step S1 specifically includes: using an industrial CT scanner to obtain CT image data sets of coal samples of different coal types and origins according to set resolutions and parameter settings.

3. The coal ash prediction method based on deep learning according to claim 1 is characterized in that: Step S2 specifically includes: S21: De-noising the CT image to reduce the noise level in the image so as to better extract the characteristics of the coal sample; S22: Image enhancement technology is used to improve the contrast and clarity of the image, making the texture features of the coal sample more obvious; S23: Crop the original CT image into small blocks of the same size to eliminate irrelevant information at the edge of the image and reduce the amount of calculation; S24: standardized cropped image; S25: Use rotation, flipping, and scaling data augmentation methods on the preprocessed images to expand the data set; S26: Segment the preprocessed image into several regions, and extract texture and shape features of each region respectively.

4. The coal ash prediction method based on deep learning according to claim 1 is characterized in that: Step S3 specifically includes: S31: Receive the preprocessed CT image data, select the dataset images as the training set, validation set and test set in a ratio of 6:3:1; random sampling is used in each process to ensure the distribution of the data and avoid data bias affecting the generalization ability of the model; S32: The image is passed through multiple 3D convolutional layers to extract the initial feature map. The convolutional layer uses ReLU as the activation function for nonlinear transformation to enhance the expression ability of the model. The output of the convolutional layer is used as input to the SPP layer. S33: The SPP layer is applied after the convolution layer to divide the feature map into different regions, and each region is pooled at different scales, and the maximum pooling operation is used to retain the main information of the image features; S34: The output of the SPP layer is used as input to two ViT encoders, one of which introduces the MSA mechanism to enhance the model's attention to key features. The MSA mechanism calculates the similarity between Query, Key, and Value and uses the Softmax function to normalize the similarity to determine the attention weight. The attention weight is used to generate a new feature representation by weighted summation, thereby enhancing the model's ability to understand the image. S35: Fusion of the outputs of the two encoders ensures that the model can combine global and local features when making predictions; S36: The encoder output passes through a fully connected layer to convert the result into the final coal ash estimate.

5. The coal ash prediction method based on deep learning according to claim 1 is characterized in that: Step S4 specifically includes: S41: Introduce error indicators MAE, RMSE, and R2 for error analysis and result visualization; S42: Setting different prediction scenario models and comparing the prediction performance of the models under different prediction scenarios.

6. The coal ash prediction method based on deep learning according to claim 4 is characterized in that: In step S32, the LBP method is used to extract texture features; The LBP method generates texture features by comparing the value of each pixel with the values ​​of its surrounding pixels: LBP(x,y,P,R)=sgn(I(x,y)-I(x+i,y+j)) (1) Where x, y are the positions of the center pixel, P is the number of pixels, R is the radius, I(x, y) is the value of the center pixel, and the values ​​of the surrounding pixels are expressed as I(x, y)-I(x+i, y+i); In the feature extraction process, a central pixel (x, y) is selected, and then a radius R and a number of pixels P are selected; then, the value of the central pixel I(x, y) is compared with the values ​​of the surrounding P pixels I(x, y)-I(x+i, y+i), where i and j are integers from -1 to 1; through comparison, a sign function value is obtained, which represents the relative size relationship between the central pixel and its surrounding pixels; these sign function values ​​are combined to obtain an LBP feature vector, which represents the texture characteristics of a specific area in the image; The SC method is used to extract shape features, and its expression is: The distance between the central pixel and all other pixels in the image is calculated and these distance values ​​are weighted summed, where the weight is Indicates the similarity between pixels; in this way, a shape context feature vector is obtained, which can represent the shape characteristics of a specific area in the image.

7. The coal ash prediction method based on deep learning according to claim 4 is characterized in that: The detailed parameters of the 3D convolutional neural network and SPP layer of step S33 are as follows: the input size is 512×512×512; the hidden layer contains two convolutional layers, the first convolutional layer contains 32 3×3×3 convolutional kernels, and the second convolutional layer contains 64 3×3×3 convolutional kernels, with a step size of 1; the SPP layer contains three pooling windows of 1×1×1, 2×2×2, and 3×3×3; the training batch is 16, the number of neural network training times is 1000, and the learning rate is 0.

001.

8. The coal ash prediction method based on deep learning according to claim 4 is characterized in that: In step S34, a CT image I is assumed, in which the value of each pixel i is represented by I i ; Divide the image into multiple image blocks B m , where m is the index of the image block; each image block B m Contains N pixels; for each image block B m Calculate the similarity between pixels: Among them, B m and B n are two different image patches, W Q is the query weight matrix, d k is the dimension of the key; calculate the image block B m The attention representation is: Where A(m) is the image block B m The attention representation is represented by M, where M is the total number of image blocks. By combining the self-attention mechanism with the encoder, the model’s ability to capture spatial relationships in CT images is enhanced, thereby improving the model’s ability to understand CT images.

9. A coal ash prediction system based on deep learning, characterized in that: include: An image acquisition module, used to collect CT image data sets of coal samples; An image preprocessing module, used for preprocessing the CT image data set; Model training module, used to build and train the MSA-ViT model; The prediction module is used to input the CT image data to be predicted, perform prediction through the trained model, and output the predicted value of coal ash.

10. The coal ash prediction system based on deep learning according to claim 9, characterized in that: The invention also comprises a user interface for receiving CT image data input by a user and displaying the prediction result.

Images