Coal density prediction method and system based on improved convolutional neural network

Abstract

The invention discloses a coal density prediction method and system based on an improved convolutional neural network, and the method comprises the following steps: collecting a CT image of a coal sample of which the density is to be predicted, carrying out the computer image preprocessing, carrying out the robust gray value normalization processing of the image according to a gray histogram, and carrying out the prediction of the density of the coal sample; obtaining original input data for improving the convolutional neural network; marking the position of the coal in the image; training the improved convolutional neural network model by using the obtained training data set, and obtaining a weight file of the improved convolutional neural network model after training is completed; and the improved convolutional neural network model automatically marks the density of each coal sample in the image by using the obtained weight file. According to the method, the problem of density prediction error caused by coal dust attached to the surface of the coal sample and coal surface texture information is effectively avoided; the method is suitable for predicting the density of batch stacked coal particles on a coal dressing site; and the method has strong generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of coal digital image processing, and in particular to a coal density prediction method and system based on an improved convolutional neural network. Background Art

[0002] Density composition is one of the important indicators of coal gravity separation. The principle of heavy medium separation is the density difference between coal and gangue. At present, the density composition of coal is still measured by floating and sinking experiments, which consumes a lot of manpower, material resources and time. Although the coal preparation plant uses a simplified fast floating method to estimate the density composition of the input coal, it still takes one or two hours, which cannot meet the requirements of real-time production guidance.

[0003] The currently known coal density prediction methods based on convolutional neural networks use camera photography as the image collection method, which has the disadvantage of requiring pre-processing of coal samples, such as cleaning the coal particles attached to the surface of the coal samples and washing them clean, otherwise the prediction accuracy will decrease. These pre-processing will also make the prediction not timely. Summary of the invention

[0004] In order to solve the above-mentioned technical difficulties in coal density prediction, the present invention provides a coal density prediction method and system based on an improved convolutional neural network.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A coal density prediction method based on an improved convolutional neural network comprises the following steps:

[0007] S1: Collect CT images of coal samples whose density is to be predicted;

[0008] S2: Perform computer image preprocessing on the collected CT images of the coal samples, including obtaining grayscale histograms and grayscale cumulative histograms; perform robust grayscale value normalization on the images according to the grayscale histograms to obtain the original input data for improving the convolutional neural network;

[0009] S3: Annotate the position of coal in the image; annotation refers to recording the density and position information of each coal sample in the image to distinguish between coal samples and between coal samples and background images, and obtain a convolutional neural network model training data set;

[0010] S4: using the obtained training data set to train the improved convolutional neural network model, and obtaining a weight file of the improved convolutional neural network model after the training is completed;

[0011] S5: Improve the convolutional neural network model to automatically mark the density of each coal sample in the image using the obtained weight file.

[0012] Preferably, step S1 specifically includes: using computer tomography technology to obtain a grayscale image of the coal sample whose density is to be predicted.

[0013] Preferably, in step S2, the image is robustly normalized in grayscale value according to the grayscale histogram, and the formula for obtaining the original input data for improving the convolutional neural network is:

[0014] a=(2 b -1) / (g max -g min )

[0015] b=-ag min

[0016] f(g)=min(max([ag+b+0.5],0),2 b -1)

[0017] G = f(g)

[0018] Where g is the gray value of the original image, f(g) is the gray value change function of the image, G is the gray value of the transformed image, and g max is the cumulative grayscale histogram of the image p l Gray value at position, g min is the image grayscale cumulative histogram p m Gray value at position; p l , p m The value is manually set to a number between 0 and 1 and satisfies p l > m .

[0019] Preferably, in step S5, automatically marking the density of each coal sample in the image means using a rectangular frame to frame the boundary of each coal sample to distinguish different coal samples; and displaying the density of the coal block to be predicted above each rectangular frame.

[0020] Preferably, in step S4, the improved convolutional neural network model includes an input layer, a convolution layer, a pooling layer, an activation function, a fully connected layer, and a loss function;

[0021] The input layer is used to scale the image to a standard size so that the image size is consistent so that the subsequent learning process can proceed normally;

[0022] The convolution layer uses convolution operations to extract various information from the image during the calculation process;

[0023] The pooling layer is used to reduce the dimension of data and reduce the feature image to increase the calculation speed;

[0024] The activation function is used to increase the nonlinear component so that the improved convolutional neural network model can learn deeper information;

[0025] The fully connected layer is used to classify the learned results and obtain the final prediction results;

[0026] The loss function is used to guide the update of model parameters and promote generalization ability.

[0027] Preferably, the activation function used is Leaky Relu, and the formula of Leaky Relu is:

[0028] f(x)=max(αx,x)

[0029] The value of α is 0.05, and x is the input variable;

[0030] Preferably, the loss function uses a rectangular box loss, and the calculation of the rectangular box loss includes calculating four losses, as follows:

[0031] S41: Calculate angle loss α

[0032]

[0033] Where h is the height difference between the center points of the real box and the predicted box, and d is the distance between the center points of the real box and the predicted box;

[0034] S42: Calculate distance loss β

[0035]

[0036]

[0037]

[0038] Where D and H are the width and height of the minimum bounding rectangle of the real box and the predicted box;

[0039] S43: Calculate shape loss γ

[0040] γ=(1-e A ) e +(1-e B ) e

[0041]

[0042]

[0043] Among them, A and B are intermediate calculation variables; a1 and b1 are the length and width of the real box, and a2 and b2 are the length and width of the predicted box;

[0044] S44: Calculate the intersection-over-union loss δ

[0045]

[0046] Where s1 is the overlapping area of ​​the real box and the predicted box, and s2 is the sum of the areas of the real box and the predicted box;

[0047] In summary, the calculation formula for the rectangular box loss is:

[0048]

[0049] The present invention also provides a coal density prediction system based on an improved convolutional neural network, comprising:

[0050] Image acquisition module: used to acquire the grayscale image of the coal sample CT whose density is to be predicted;

[0051] Image preprocessing module: used for computer image preprocessing of the collected CT images of coal samples;

[0052] Control module: used to adjust the input image size and control the training strategy;

[0053] Improved convolutional neural network module: used to mark the processing results obtained by the image preprocessing module, and mark the location information and density information of each coal sample;

[0054] Detection module: It is used to automatically predict the density of the coal sample to be predicted and automatically label it using the weight file obtained through training of the improved convolutional neural network module.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. The present invention can be applied to different coal samples in different regions. The training samples can be prepared quickly, the model training time is short, and the prediction effect is good.

[0057] 2. Since the present invention uses CT images as raw data, it can effectively avoid problems such as incorrect density prediction caused by coal powder attached to the surface of the coal sample and coal surface texture information.

[0058] 3. Since the model prediction speed of the present invention is only related to the image size, it can be applied to the density prediction of batch stacked coal particles at the coal preparation site.

[0059] 4. The present invention can improve the coal sample CT image preprocessing and improved convolutional neural network module according to the characteristics of coal samples in different regions to improve the prediction effect, and has strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] 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.

[0061] Figure 1 It is a flow chart of the present invention.

[0062] Figure 2 This is a diagram showing the predicted effect of an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0064] like Figure 1 As shown, a coal density prediction method based on an improved convolutional neural network includes the following steps:

[0065] S1: Collecting a CT image of a coal sample whose density is to be predicted; specifically, obtaining a grayscale image of the coal sample whose density is to be predicted by using computer tomography technology.

[0066] S2: Perform computer image preprocessing on the collected CT images of the coal samples, including obtaining the grayscale histogram and the grayscale cumulative histogram; perform robust grayscale value normalization on the image according to the grayscale histogram to obtain the original input data for improving the convolutional neural network; perform robust grayscale value normalization on the image according to the grayscale histogram to obtain the formula for the original input data for improving the convolutional neural network:

[0067] a=(2 b -1) / (g max -g min )

[0068] b=-ag min

[0069] f(g)=min(max([ag+b+0.5],0),2 b -1)

[0070] G = f(g)

[0071] Where g is the gray value of the original image, f(g) is the gray value change function of the image, G is the gray value of the transformed image, and g max is the cumulative grayscale histogram of the image p l Gray value at position, g min is the image grayscale cumulative histogram p m Gray value at position; pl , p m The value is manually set to a number between 0 and 1 and satisfies p l > m In this embodiment, p l =0.8, p m =0.2.

[0072] S3: marking the position of the coal in the image; marking refers to recording the density and position information of each coal sample in the image to distinguish between coal samples and between coal samples and background images; in this embodiment, three density sets of coal samples are marked, namely 1.35-1, 45, 1.45-1.55, 1.8+ (unit: g / cm 3 ).

[0073] S4: The improved convolutional neural network model is trained using the obtained training set, and the weight file of the improved convolutional neural network model is obtained after the training is completed; the improved convolutional neural network model includes an input layer, a convolution layer, a pooling layer, an activation function, a fully connected layer, and a loss function; the improved convolutional neural network model is improved by modifying the loss function calculation method of the model during the training process, so that the model can extract deeper information during the training process, thereby improving recognition accuracy and training speed;

[0074] The input layer is used to scale the image to a standard size so that the image size is consistent so that the subsequent learning process can proceed normally;

[0075] The convolution layer uses convolution operations to extract various information from the image during the calculation process; to speed up model training, all convolution kernels are set to a size of 3×3.

[0076] The pooling layer is used to reduce the dimension of data and reduce the feature image to increase the calculation speed;

[0077] The activation function is used to increase the nonlinear component so that the improved convolutional neural network model can learn deeper information. The activation function used is Leaky Relu, which has the following advantages: it solves the problem of neuron death when the input value of the ReLU activation function is negative; Leaky ReLU can converge quickly due to its linear and non-saturated properties; it has low computational complexity and does not require exponential operations. The formula for Leaky Relu is:

[0078] f(x)=max(αx,x)

[0079] The value of α is 0.05, and x is the input variable;

[0080] The fully connected layer is used to classify the learned results and obtain the final prediction results;

[0081] The loss function is used to guide the update of model parameters and promote generalization ability. The loss function uses rectangular box loss. The calculation of rectangular box loss includes calculating four losses, as follows:

[0082] S41: Calculate angle loss α

[0083]

[0084] Where h is the height difference between the center points of the real box and the predicted box, and d is the distance between the center points of the real box and the predicted box;

[0085] S42: Calculate distance loss β

[0086]

[0087]

[0088]

[0089] Where D and H are the width and height of the minimum bounding rectangle of the real box and the predicted box;

[0090] S43: Calculate shape loss γ

[0091] γ=(1-e A ) e +(1-e B ) e

[0092]

[0093]

[0094] Among them, A and B are intermediate calculation variables; a1 and b1 are the length and width of the real box, and a2 and b2 are the length and width of the predicted box;

[0095] S44: Calculate the intersection-over-union loss δ

[0096]

[0097] Where s1 is the overlapping area of ​​the real box and the predicted box, and s2 is the sum of the areas of the real box and the predicted box;

[0098] In summary, the calculation formula for the rectangular box loss is:

[0099]

[0100] S5: Improve the convolutional neural network model to automatically mark the density of each coal sample in the image using the obtained weight file; automatically mark the density of each coal sample in the image, which means to use a rectangular frame to frame the boundary of each coal sample and distinguish different coal samples; at the same time, the density of the coal block to be predicted is displayed above each rectangular frame, such as Figure 2 shown.

[0101] The present invention also provides a coal density prediction system based on an improved convolutional neural network, comprising:

[0102] Image acquisition module: used to acquire grayscale images of coal sample CT whose density is to be predicted; grayscale images refer to images represented by grayscale, which are divided into several grayscale levels in the range of 0 to 255, where 0 represents black and 255 represents white.

[0103] Image preprocessing module: used to perform computer imaging preprocessing on the CT images of the collected coal samples; this makes the difference between coal samples of different densities greater, making it easier to improve the convolutional neural network module to extract more information.

[0104] Control module: used to adjust the input image size and control the training strategy; the hyperparameter settings are as follows: training rounds: 400 rounds, batch sample size: 4, image size: 640 pixels, learning rate: 0.01.

[0105] Improved convolutional neural network module: used to annotate the processing results obtained by the image preprocessing module, annotate the location information and density information of each coal sample; construct training set, validation set, and test set in a ratio of 3:1:1. Use the training set and validation set to train the improved convolutional neural network model so that it can learn the difference between coal samples of different densities. Use the test set to test the generalization ability of the model.

[0106] Detection module: It is used to automatically predict the density of the coal sample to be predicted and automatically label it using the weight file obtained through training of the improved convolutional neural network module.

[0107] 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 density prediction method based on improved convolutional neural network, characterized in that: The steps include: S1: Collect CT images of coal samples whose density is to be predicted; S2: Perform computer image preprocessing on the collected CT images of the coal samples, including obtaining grayscale histograms and grayscale cumulative histograms; perform robust grayscale value normalization on the images according to the grayscale histograms to obtain the original input data for improving the convolutional neural network; S3: Annotate the position of coal in the image; annotation refers to recording the density and position information of each coal sample in the image to distinguish between coal samples and between coal samples and background images, and obtain a convolutional neural network model training data set; S4: using the obtained training data set to train the improved convolutional neural network model, and obtaining a weight file of the improved convolutional neural network model after the training is completed; S5: Improve the convolutional neural network model to automatically mark the density of each coal sample in the image using the obtained weight file.

2. The coal density prediction method based on improved convolutional neural network according to claim 1 is characterized in that: Step S1 specifically includes: using computer tomography technology to obtain a grayscale image of the coal sample whose density is to be predicted.

3. The coal density prediction method based on improved convolutional neural network according to claim 1 is characterized in that: In step S2, the image is robustly normalized according to the grayscale histogram, and the formula for the original input data for improving the convolutional neural network is obtained as follows: a=(2 b -1) / (g max -g min ) b=-ag min f(g)=min(max([ag+b+0.5],0),2 b -1) G = f(g) Where g is the gray value of the original image, f(g) is the gray value change function of the image, G is the gray value of the transformed image, and g max is the cumulative grayscale histogram of the image p l Gray value at position, g min is the image grayscale cumulative histogram p m Gray value at position; p l , P m The value is manually set to a number between 0 and 1 and satisfies p l >p m .

4. The coal density prediction method based on improved convolutional neural network according to claim 1 is characterized in that: In step S5, the density of each coal sample is automatically marked in the image, which means that the boundary of each coal sample is framed by a rectangular frame to distinguish different coal samples; at the same time, the density of the coal block to be predicted is displayed above each rectangular frame.

5. The coal density prediction method based on improved convolutional neural network according to claim 1 is characterized in that: In step S4, the improved convolutional neural network model includes an input layer, a convolution layer, a pooling layer, an activation function, a fully connected layer, and a loss function; The input layer is used to scale the image to a standard size so that the image size is consistent so that the subsequent learning process can proceed normally; The convolution layer uses convolution operations to extract various information from the image during the calculation process; The pooling layer is used to reduce the dimension of data and reduce the feature image to increase the calculation speed; The activation function is used to increase the nonlinear component so that the improved convolutional neural network model can learn deeper information; The fully connected layer is used to classify the learned results and obtain the final prediction results; The loss function is used to guide the update of model parameters and promote generalization ability.

6. The coal density prediction method based on improved convolutional neural network according to claim 5 is characterized in that: The activation function used is Leaky Relu, and the formula of Leaky Relu is: f(x)=max(αx,x) The value of α is 0.05 and x is the input variable.

7. The coal density prediction method based on improved convolutional neural network according to claim 5 is characterized in that: The loss function uses a rectangular box loss. The calculation of the rectangular box loss includes calculating four losses, as follows: S41: Calculate angle loss α Where h is the height difference between the center points of the real box and the predicted box, and d is the distance between the center points of the real box and the predicted box; S42: Calculate distance loss β Where D and H are the width and height of the minimum bounding rectangle of the real box and the predicted box; S43: Calculate shape loss γ γ=(1-e A ) e +(1-e B ) e Among them, A and B are intermediate calculation variables; a1 and b1 are the length and width of the real box, and a2 and b2 are the length and width of the predicted box; S44: Calculate the intersection-over-union loss δ Where s1 is the overlapping area of ​​the real box and the predicted box, and s2 is the sum of the areas of the real box and the predicted box; In summary, the calculation formula for the rectangular box loss is:

8. A coal density prediction system based on improved convolutional neural network, characterized in that: include: Image acquisition module: used to acquire the grayscale image of the coal sample CT whose density is to be predicted; Image preprocessing module: used for computer image preprocessing of the collected CT images of coal samples; Control module: used to adjust the input image size and control the training strategy; Improved convolutional neural network module: used to mark the processing results obtained by the image preprocessing module, and mark the location information and density information of each coal sample; Detection module: It is used to automatically predict the density of the coal sample to be predicted and automatically label it using the weight file obtained through training of the improved convolutional neural network module.

Images