Coal body structure identification method based on improved convolutional neural network

Abstract

The invention discloses a coal body structure identification method based on an improved convolutional neural network, and belongs to the technical field of geological exploration. The method comprises the following steps of: (1) screening out a logging curve sensitive to coal body structure characteristic response; (2) carrying out wavelet decomposition and noise reduction processing on the logging curve; (3) carrying out normalization processing on the logging curve; (4) reading feature vectors of samples formed by logging data according to a window, and forming training sample data by the feature vectors and coal body structure type labels obtained by coal core logging; (5) constructing a deep learning coal body structure recognition model based on the improved convolutional neural network; and (6) training the convolutional neural network model by using the training sample data, obtaining optimal network parameters through multiple times of network parameter iteration updating, returning the optimal network parameters to the prediction model, and obtaining an optimal model by properly adjusting the structure and hyper-parameters of the model. And inputting the new logging data into the prediction model to obtain a coal body structure identification result. By adopting the method, the coal body structure identification precision can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of geological exploration technology, and specifically to a coal body structure recognition method based on an improved convolutional neural network. By constructing an improved convolutional neural network deep learning coal body structure recognition model, and utilizing its powerful learning ability and nonlinear mapping ability, the implicit information in well logging data is deeply mined, and a mapping relationship between well logging and coal body structure is established, thereby realizing the prediction of well logging coal body structure. Background Art

[0002] Against the backdrop of the continuous growth of global energy demand and the increasing pressure on environmental protection, coalbed methane, as a clean energy source, has gradually become a hot topic in the energy field due to its large resources, wide distribution, high calorific value and low pollution. The development and utilization of coalbed methane has multiple benefits, such as ensuring safe production in coal mines, increasing the supply of clean energy and reducing greenhouse gas emissions. Coal body structure is an important factor affecting the permeability of coal seams, the enrichment and high yield of coalbed methane, and is also a key geological factor in the reformability of coalbed methane reservoirs. At present, coalbed methane development prefers coal seams with relatively complete coal body structure, high permeability, low development difficulty and good reformability, while coal seams with a high degree of coal body structure damage need to take avoidance measures under traditional technical conditions. Accurately identifying the coal body structure of coal seams has important practical significance for the exploration and development of coal and coalbed methane.

[0003] Well logging interpretation is a feasible and widely used technology for identifying coal body structure in the coalbed methane exploration process. It is efficient and accurate, breaking through the limitations of traditional ground exploration, and can conduct field exploration deep underground. It uses well logging to accurately divide the coal body structure, helping to grasp the vertical distribution characteristics of coal with different coal body structures, which can significantly improve the efficiency of coalbed methane exploration and development and reduce development costs. In theory, there are multiple well logging curves that are sensitive to the type of coal body structure, but in reality, the distribution of coal body structure is very complex, and a single well logging curve is also affected by a variety of geological and engineering factors and varies. There is overlap in the logging values ​​of well logging curves in different coal body structures, and traditional well logging coal body structure identification is prone to multiple solutions, which brings challenges to coal body structure identification. Summary of the invention

[0004] Technical problem: The purpose of the present invention is to overcome the problems of low efficiency, insufficient accuracy, complex feature engineering, etc. of traditional coal structure identification caused by the complexity of geological conditions and the limitations of logging curves, and to provide a coal structure identification method based on an improved convolutional neural network. By establishing a deep learning neural network model and utilizing its powerful learning ability and nonlinear mapping ability, the implicit information of logging data is deeply mined, and a mapping relationship between logging and coal structure is established, thereby realizing intelligent prediction of complex coal structure in logging.

[0005] Technical solution: A coal structure recognition method based on an improved convolutional neural network of the present invention comprises the following steps:

[0006] (a) Collect well logging curves of core drilling holes in the mining area and surrounding geological exploration, and select five well logging curves that are sensitive to coal body structure response as the five features of sample data;

[0007] (b) determining the location of the coal seam in the formation encountered by the borehole based on the drilling core log or the logging curve shape, and selecting the logging data of the coal seam section of the borehole;

[0008] (c) Eliminate the dimension and scale differences between different logging curves;

[0009] (d) reading multiple lines of logging data with the same logging sampling point labels and continuous sampling depths in the coal seam section to form a logging feature vector of a sample;

[0010] (e) construct a convolutional neural network coal structure recognition model, which consists of a convolutional layer, a pooling layer and two fully connected layers. The convolutional layer input is set as a single-channel image to process the logging feature vector of the sample;

[0011] (f) Use grid search to find the best hyperparameter combination;

[0012] (g) Use the training set to train the convolutional neural network model and input the new logging data into the prediction model to obtain the predicted coal structure label.

[0013] In step (a), the data of logging curves of coring boreholes for geological exploration in and around the mining area are collected, the vertical change characteristics of the coal body structure of the borehole coal seam are obtained through the coal core logging data, and the changes in the logging values ​​of coals with different coal body structures are statistically analyzed, and the logging curves with better regularity of logging value changes with increasing degree of coal body structure damage are selected as the characteristic curves for coal body structure logging identification; the logging curves include five logging curves: density logging curve, sonic time difference logging curve, caliper logging curve, resistivity logging curve and natural gamma logging curve.

[0014] In step (b), the logging data of the coal seam section of the borehole is selected, and each original logging curve is decomposed using wavelet decomposition to remove high-frequency noise in the logging data, retain the low-frequency approximate curve after secondary decomposition, and improve the data signal-to-noise ratio; that is, the depth range of the coal seam in the borehole is determined according to the core logging data or the logging curve morphology, and the logging data in the depth range of the coal seam is selected, and the original logging curve is decomposed using sym8 wavelet decomposition, each level of decomposition obtains a high-frequency signal curve and a low-frequency similarity curve, the original logging curve is subjected to secondary decomposition, the high-frequency curve is removed, and the similarity curve after the secondary decomposition is retained for coal body structure identification.

[0015] In the step, the dimensional differences and scale differences between different logging curves are eliminated: the logging data is preprocessed using data normalization to make the model easier to converge and improve the training efficiency. The data normalization subtracts the minimum value from the logging curve and divides it by the difference between the maximum and minimum values ​​to scale the data to the range of [0, 1], eliminating the dimensional differences between different logging curves.

[0016] In the step, the sampling depth interval of the logging sampling point is 0.05m, and each logging sampling point is composed of the logging values ​​of five logging curves to form five characteristics of the sampling point. Each sampling point generates a row of logging data, and the n sampling points in the coal seam section form n rows of logging data. The coal body structure information of each sampling point is obtained through the coal core logging data as the label of the sampling point; the sample reading process starts from the third row of logging data, and determines whether the label of the i-th logging sampling point is consistent with the first two sampling points and the last two sampling points, and whether the logging sampling depths of the five sampling points are continuous. If the conditions are met, then Read the logging data of the i-th logging sampling point and the two sampling points above and below the i-th sampling point to form a 5×5 logging feature vector of a sample, and the label of the i-th sampling point is used as the sample label of the sample; repeat the above operation to read the next sample with one logging sampling point interval, until the n-2-th row of logging data is read to obtain the last sample; read all the coal seam logging data in the geological exploration coring borehole to generate samples to obtain a sample set; encode the sample labels of the samples in the sample set, and finally randomly select 20% of the samples from the sample set as the test set, and the remaining samples as the training set.

[0017] The sample labels use the numbers "1", "2" and "3" to represent three different types of coal body structures. Before inputting the samples into the model, the categorical data is first converted into integers between 0 and n-1 using the LabelEncoder of the sklearn (scikit-learn) library, that is, the three types of coal body structure sample labels "1", "2" and "3" are converted into three category labels "0", "1" and "2"; secondly, the sample labels are converted into one-hot encoding (One-HotEncoding) using the LabelEncoder of the sklearn library; 20% of the samples are extracted from the sample set as a test set, and the sample segmentation tool of the sklearn library is used to set the test set ratio to 0.2. The algorithm automatically extracts 20% of the samples as a test set for evaluating the model in the model training process, and the remaining samples are used as a training set for updating the model network parameters in the model training process.

[0018] In step (e), the model convolution layer accepts a single-channel image with an input size of 5×5, the number of convolution kernels in the convolution layer is set to 64, the convolution kernel size is set to 2×2, the convolution kernel moving step is set to 1, and the convolution kernel of the convolution layer performs convolution calculation on the logging feature vector of each input sample to obtain 64 feature maps; the pooling layer is set to maximum pooling, the pooling kernel function size is set to 2×2, the moving step is set to 1, and each feature map is processed by pooling to obtain a pooled feature map. After the 64 feature maps are processed by the pooling layer, 64 pooled feature maps are output. Feature map; 64 pooled feature maps are flattened into one-dimensional feature vectors through the flattening layer and then input into the fully connected layer. The fully connected layer consists of two layers. The first layer contains 64 neurons and the second layer contains 32 neurons. Finally, the output is obtained through the Softmax function. Add optimization algorithm to optimize the convolutional neural network model. Add ReLU activation function and L2 regularization to the convolution layer, add ReLU activation function and Dropout to the fully connected layer, and the Dropout ratio is 0.5. Use Adam optimizer and the learning rate is set to 0.01.

[0019] In the sample processing process, the 5×5 well logging feature vector is regarded as a single-channel image of size 5×5 input to the model convolution layer; the model convolution layer uses "0" to fill the well logging feature vector of each input sample, and uses 64 2×2 convolution kernels to perform convolution calculation on the well logging feature vector to extract local feature information of the well logging feature vector. Each well logging feature vector is convoluted to obtain 64 5×5 feature maps; the pooling layer uses maximum pooling, the pooling kernel function size is 2×2, the moving step is 1, the pooling kernel function scans each feature map to output a 4×4 pooling feature map, and the pooling layer outputs a total of 64 4×4 pooling feature maps; The 64 pooled feature maps are flattened by the flattening layer and aggregated into a one-dimensional feature vector with a total length of 64×4×4=1024; the fully connected layer contains 64 and 32 neurons respectively, and the mapping relationship between the one-dimensional feature vector and the sample label is established between the neurons through correlation calculation; the output uses the Softmax activation function to convert the fully connected output into the probability of belonging to each category label, and the category with the largest probability is taken as the output category. The output category label is in the form of "0", "1", and "2", and the LabelEncoder inverse conversion is used to convert the output category label back to the coal body structure label in the form of "1", "2", and "3".

[0020] In step (e), the convolution layer of the model adds a ReLU activation function and an L2 regularization with a value of 0.001; a Dropout layer is added after each fully connected layer, and the Dropout ratio is 0.5. The Adma optimizer is added to the model to improve the model training efficiency and final performance. The model learning rate is set to 0.01, and the grid search is used to find the best hyperparameter combination. First, the convolution kernel, neurons, and learning rate range are defined. The automatic parameter tuner randomly combines hyperparameters and trains the model, automatically evaluates the model, and finally returns the optimal hyperparameter combination with the highest accuracy and the smallest loss function after the training.

[0021] In step (f), the grid search for the best hyperparameter combination is used. Before obtaining the best hyperparameter, the value range of each hyperparameter that needs to be tuned is first defined. The range of the number of convolution kernels is defined as 32, 64, and 128, the range of the number of neurons is defined as 32, 64, and 128, and the range of the learning rate is defined as 0.1, 0.01, and 0.001. The automatic parameter tuner randomly combines the hyperparameters and trains the model, evaluates the model, and finally returns the optimal hyperparameter combination; the evaluation model records the model accuracy and loss function values ​​after training under each hyperparameter combination, and returns the hyperparameter combination with the highest accuracy and the smallest loss function after training, which is the optimal hyperparameter combination.

[0022] In step (g), the convolutional neural network model is trained using the training set, the number of model iterations is set to 1000, the neural network parameters are updated using the small batch stochastic gradient descent method and the back propagation algorithm, the small batch size is set to 128, and the model updates the neural network parameters each time iteratively so that the loss function is reduced, and the loss function adopts the classification cross entropy loss function; the prediction accuracy and loss function changes after each iteration model parameter update are plotted, and the global minimum or approximate value of the loss is finally found after multiple iterations, and the accuracy and loss function curves converge. At this time, the training accuracy is the highest and the loss function is the lowest, and a trained coal body structure prediction model is obtained;

[0023] In each iteration of the convolutional neural network model, 128 samples are randomly selected from the training set to calculate the loss function of the predicted results and the actual results. The loss function adopts the classification cross entropy loss, uses the gradient descent method to update the network parameters, and adopts the back propagation algorithm to update the parameters of each layer of the network, so that the loss function gradually decreases. The difference between the model prediction results and the actual labels gradually decreases as the loss function decreases, and the prediction accuracy gradually increases. When the accuracy and loss function curves converge, the loss function reaches the global minimum, and the model training is completed. At this time, the model prediction accuracy is the highest.

[0024] Beneficial effect: Due to the adoption of the above technical solution, the present invention performs wavelet decomposition, denoising and normalization on the logging data, reads the logging data by window to form the feature vector of the training sample, establishes an improved convolutional neural network model to identify the coal body structure, uses the training sample composed of the feature vector and the label to train the established improved convolutional neural network coal body structure recognition model, obtains the trained coal body structure prediction model, inputs the new sample without label into the prediction model to obtain the prediction label, realizes the intelligent recognition of the coal body structure, and overcomes the problems of low efficiency, insufficient accuracy and complex feature engineering of traditional coal body structure recognition caused by the complexity of geological conditions and the limitations of logging curves. First, a deep learning neural network model is established, and its powerful learning ability and nonlinear mapping ability are used to deeply mine the implicit information of the logging data, establish the mapping relationship between logging and coal body structure, and realize the intelligent prediction of complex coal body structure; secondly, the logging data reading method is optimized, and each sample reads 5 lines of logging data to form a feature vector, which ensures that each sample has enough features for the neural network model to learn, and can make full use of the logging data, improving the model learning effect and prediction accuracy. The main advantages compared with the prior art are:

[0025] ① It realizes intelligent identification of coal body structure, which is fast and efficient, and significantly improves the prediction accuracy of coal body structure;

[0026] ② No additional feature engineering, avoiding the subjectivity caused by artificial division;

[0027] ③Combining multiple logging curves for analysis, with higher accuracy and stronger generalization ability;

[0028] ④ When reading sample data, each sample uses multiple rows of logging data to form a feature vector, and the spatial structure of the logging data is retained, so that the neural network obtains sufficient information;

[0029] ⑤The model can be trained and optimized based on coal structure data under different geological backgrounds to meet the needs of coal structure identification in different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the coal body structure recognition method based on the improved convolutional neural network of the present invention.

[0031] Figure 2 This is the improved convolutional neural network model diagram of the present invention.

[0032] Figure 3 It is a diagram of data processing and sample reading of the present invention.

[0033] Figure 4 This is a data reading flow chart and a model training process chart of the present invention.

[0034] In the figure: 1-well logging curve; 2-coal seam; 3-normalization; 4-wavelet decomposition; 5-original well logging curve; 6-similar curve; 7-well logging sampling depth; 8-feature; 9-well logging feature vector; 10-label; 11-sample label; 12-sample; 13-sample set; 14-training set; 15-test set; 16-small batch; 17-convolution layer; 18-pooling layer; 19-fully connected layer; 20-convolution kernel; 21-feature map; 22-pooling kernel function; 23-pooling feature map; 24-flattening layer; 25-neuron; 26-output.

[0035] Figure 2 Middle: f 1 : The first row of features of the sample, corresponding to the first row of well logging data in the sample; GR, CAL, DEN, AC, R: well logging curves; K 1 : The first convolution kernel; Padding: Padding around the feature map; Convolution: Convolution operation; ReLU: Activation function; Kernel: Convolution kernel; 2×2: indicates the vector size; Stride: Convolution kernel moving step; Feature Map, F 1 : feature map; 64@(5×5): 64 feature maps of size 5×5; pool size: pooling kernel function size; PooledMap, P 1 : Pooling feature map; n 1 : Neuron; Fully connected: fully connected; P, F, T: coal body structure type label.

[0036] Figure 3 In: Depth: depth of logging sampling point; feature: feature of a single sampling point; label: label of a single row of logging data; Feature, F 1 : sample features; Label: sample label; Sample: sample.

[0037] Figure 4 In: i: the i-th row of logging data; n: the number of logging data rows; Train Samples: training sample data set; Training Set: training set; Test set: test set; CNN Model: convolutional neural network model; Loss: loss function; Validation Loss: validation loss; Mini-bach SDG: mini-batch gradient descent; One Epoch: single iteration; Preset Epoch: preset number of iterations. DETAILED DESCRIPTION

[0038] An embodiment of the present invention is further described below in conjunction with the accompanying drawings:

[0039] like Figure 1 As shown, the coal body structure recognition method based on the improved convolutional neural network of the present invention performs wavelet decomposition, denoising and normalization on the logging data, reads the feature vectors of the training samples composed of the logging data according to the window, establishes an improved convolutional neural network model to recognize the coal body structure, uses the training samples composed of the feature vectors and labels to train the established improved convolutional neural network coal body structure recognition model, obtains the trained coal body structure prediction model, inputs the new samples without labels into the prediction model to obtain the prediction labels, and realizes the intelligent recognition of the coal body structure. The specific steps are as follows:

[0040] (a) When the coal reservoir is destroyed by tectonic action, the porosity, density, radioactivity and other properties of the coal reservoir change with the increase of the degree of coal body structure destruction. Well logging can capture these formation properties and reflect them as changes in amplitude; collect well logging curves 1 data and coal core logging data of geological exploration coring boreholes in the mining area and surrounding areas, and obtain the vertical change characteristics of the coal body structure of the borehole coal seam 2 through the coal core logging data; count the changes in well logging values ​​of coal with different coal body structures, and screen out the well logging curves 1 that are sensitive to the coal body structure response. The well logging curves 1 with better regularity of well logging value changes with the increase of coal body structure destruction include density logging curve DEN, acoustic time difference logging curve AC, caliper logging CAL, resistivity logging R, and natural gamma logging GR. Therefore, the above five well logging curves 1 are selected as the five characteristic 8 curves of sample data.

[0041] (b) According to the drilling core logging data or the morphological characteristics of the logging curve 1, the position of the coal seam 2 in the stratum encountered by the borehole is determined, the logging data of the drilling coal seam 2 section is selected, and the original logging curve 5 is subjected to secondary decomposition using the sym8 wavelet decomposition 4. Each level of decomposition can obtain a high-frequency signal curve and a low-frequency similarity curve 6; the high-frequency curve is removed, and the similarity curve 6 after the secondary decomposition is retained to achieve the curve filtering effect, remove the high-frequency noise in the logging signal that is not very helpful for identifying the coal body structure type, and improve the data signal-to-noise ratio;

[0042] (c) Use normalization 3 to preprocess the logging data. Each logging curve 1 is scaled to the range of [0, 1] by subtracting the minimum value and dividing it by the difference between the maximum and minimum values, as shown in Formula 1. This eliminates the dimensional differences between different logging curves, making the model easier to converge and improving training efficiency.

[0043]

[0044] Where: x is the original data, x norm is the normalized data, x max is the maximum value of the data x min is the minimum value of the data.

[0045] (d) The sampling depth 7 of the original logging data is 0.05m interval, that is, a logging sampling is performed every 0.05m interval and a row of logging data is formed. Each row of logging data contains five features 8 composed of five logging curves 1 logging values; n sampling points in the coal seam 2 section form n rows of logging data. Assume that i is the i-th row of logging data, and use f i represents the logging value of the i-th row, f i =[GR i ,DEN i ,AC i ,R i ,CAL i ], representing the five features of the i-th row in the well logging data 8; the coal structure information of each sampling point is obtained through the coal core logging data, label i Indicates the coal body structure label 10 corresponding to the i-th row of logging data, and divides the coal body structure into three categories, with labels 10 of "1", "2", and "3" respectively. When reading sample 12, start reading from the third row of logging data to determine whether the label 10 of a certain logging sampling point is consistent with the first two sampling points and the last two sampling points, and whether the sampling depths 7 of the five sampling points are continuous; when reading the i-th row of logging data in the actual algorithm design, the depth Depth of the i+2th row of data is calculated. i+2 Subtract the depth of row i-2 i-2 Whether it is equal to 0.4m indirectly determines whether the sampling depth 7 of the five lines of data is continuous. If the label 10 of the i-th line is the same as the label 10 of the two lines before and after and the sampling depth 7 of the five lines of data is continuous, then read the i-th well logging sampling point and the two sampling points above and below the i-th sampling point, a total of five lines of well logging data, to form the j-th 5×5 well logging feature vector 9F j =[f i-2 ,f i-1 ,f i ,f i+1 ,f i+2 ] T , read the label 10label of the i-th row of logging data i As the sample label 11Label of the jth sample 12 j , well logging feature vector 9F j and sample label 11Label j Composed of j-th sample 12Sample jThat is, five rows of logging values ​​with the same label 10 and continuous sampling depth 7 are read each time to form a 5×5 sample logging feature vector 9, which is combined with the coal body structure type sample label 11 obtained from the coal core logging data to form a sample 12. The above operation is repeated every other row of logging data to determine whether the labels 10 of the five rows of logging data are the same and the sampling depth 7 is continuous. If the condition is met, the second sample 12 is read. If not, the next row of logging data is read and judged until the n-2th row of logging data is read to obtain the last sample 12. The logging data of the second section of the coal seam is read to form a sample set 13. After obtaining the sample set 13, firstly, use the LabelEncoder of the sklearn (scikit-learn) library to convert the categorical data into integers between 0 and n-1, that is, convert the coal body structure labels 10 "1", "2", and "3" into "0", "1", and "2" respectively; secondly, use the LabelEncoder of the sklearn library to convert the label 10 into a one-hot encoding (One-Hot Encoding), converting each categorical value into a binary column, in which only one column is 1 (indicating the category), and the rest are 0. Finally, use the sample segmentation tool to segment the sample set 13, set the test set ratio to 0.2, and the algorithm randomly selects 20% of the samples 12 from the sample set 13 as the test set 15, and the remaining samples 12 as the training set 14.

[0046] (e) Construct a convolutional neural network coal structure recognition model with a single-channel input. Set the convolution layer 17 of the coal structure recognition model to accept a single-channel image with an input size of 5×5. When the sample 12 is input into the model, the model can treat the 5×5 well logging feature vector 9 of the sample 12 as a single-channel image with a size of 5×5 for processing. Enable the padding operation of the convolution layer 17. The convolution layer 17 first uses "0" to fill the feature vector 9 around each sample 12. Set the number of convolution kernels 20 to 64, the size to 2×2, and the moving step to 1. Each convolution kernel 20 of the convolution layer 17 scans the well logging feature vector 9 from top to bottom and from left to right, respectively, and extracts the features of the vector 9 through cross-correlation operation. The convolution layer 17 finally outputs 64 feature maps 21 with a size of 5×5. Add the ReLU activation function to the convolution layer 17 to enhance the nonlinear fitting ability of the model. Add L2 regularization with a value set to 0.001. The pooling layer 18 is set to maximum pooling, the size of the pooling kernel function 22 is 2×2, the moving step is 1, and the pooling kernel function 22 performs a scan and dimensionality reduction process on each feature map 21 to obtain 64 4×4 pooling feature maps 23. The 64 pooling feature maps 23 are flattened by the flattening layer 24 and aggregated into a one-dimensional feature vector with a total length of 64×4×4=1024 and passed to the fully connected layer 19. The fully connected layer 19 establishes a correlation between the one-dimensional feature vector and the sample label 11. It contains two layers. The first layer contains 64 neurons 25, which are connected to the one-dimensional feature vector in front and the 32 neurons 25 in the second layer in the back. The 32 neurons 25 in the second layer are then connected to the sample label 11. The fully connected layer 19 adds a ReLU activation function, and a Dropout layer is added after each fully connected layer 19, and the Dropout ratio is 0.5. Output 26 uses the Softmax activation function to convert the original output into a function of probability distribution, that is, the probability of belonging to a certain category. The category with the largest probability is taken as the output 26 category. The output 26 category label is in the form of "0", "1", and "2". The LabelEncoder inverse conversion is used to convert the output 26 category label back to the coal structure label 11 in the form of "1", "2", and "3". In order to improve the model training efficiency and final performance, the Adam optimizer is added to the model. The Adam optimizer adaptively adjusts the learning rate for each parameter by calculating the exponentially weighted gradient average and the exponentially weighted gradient square average.

[0047] (f) To quickly find the optimal value of each hyperparameter, grid search is used to find the best hyperparameter combination. First, the possible value range of each hyperparameter that needs to be tuned is defined. The range of the number of convolution kernels 20 is defined as 32, 64, and 128, the range of the number of neurons 25 is defined as 32, 64, and 128, and the range of the learning rate is defined as 0.1, 0.01, and 0.001. The automatic parameter tuner randomly combines hyperparameters and trains the model, and records the accuracy and loss function of the model prediction results after each model network parameter update to evaluate the model, draws the accuracy and loss function change curves, and the training ends when the accuracy and loss function curves converge, and records the accuracy and loss function of the model prediction at this time. After the grid search is completed, the hyperparameter combination with the highest accuracy and the lowest loss function after the training is compared, which is the optimal hyperparameter combination.

[0048] (g) The convolutional neural network model is trained using the training set 14, and the back-propagation algorithm is used to update the parameters of each layer of the network so that the loss function gradually decreases. The core of the back-propagation algorithm is to calculate the gradient of each layer, that is, the partial derivative of the loss function with respect to the parameters of each layer. Through the chain rule, the gradient of the loss function can be back-propagated from the output layer to the input layer layer by layer, and the parameters of each layer are updated according to the gradient. Using the mini-batch 16 stochastic gradient descent method, in each iteration, the model first randomly samples a mini-batch 16β consisting of training samples of size |β|; then, the derivative or gradient of the average loss of the mini-batch 16 with respect to the model parameters is calculated, and finally, the gradient is multiplied by a predetermined positive number η and subtracted from the current parameter value. The mathematical formula of this update process is expressed as follows:

[0049]

[0050] Among them l (i) is the loss of the ith sample, w and b represent the weight and bias in the network connection respectively, both are vectors, and the model will randomly initialize them before training; |β| represents the number of samples in each small batch of 16, which is set to 128 in the model; η represents the learning rate, which is set to 0.01 in the model. Both are manually specified during the model construction process and can be adjusted but not updated during the training process. They are called hyperparameters. Other hyperparameters used in the training model are shown in Table 1. After the number of training iterations reaches the set 1000 times, the difference between the model prediction result and the actual value gradually decreases, and finally the global minimum or approximate value of the loss is found. The final hidden parameter estimate of the model is obtained After entering new feature data, the model can be used Calculate the predicted results. Represent the weights and biases of the connections used in the network.

[0051] The loss function uses the classification cross entropy loss function to evaluate the difference between the probability distribution predicted by the model and the probability distribution of the true label. Minimizing the loss function means the best fit, and the corresponding model parameters are the optimal parameters.

[0052]

[0053] Where C is the total number of categories, y i is the value of the i-th element in the one-hot encoding of the true label (this value is 1 for the correct category and 0 for other categories), is the probability of the i-th category in the probability distribution predicted by the model.

[0054] When evaluating the model, the accuracy and loss function of the model's prediction results for the training set and test set are recorded after each update of the model network parameters, and the training accuracy and verification accuracy, training loss function and verification loss function are returned. The accuracy and loss function change curves are plotted. The training ends when the accuracy and loss function curves converge. The structure and hyperparameters of the model can be appropriately adjusted according to the sample data to obtain the best model for the data set and achieve better prediction results. Finally, the new logging data is input into the trained model under the optimal hyperparameters to obtain the coal structure prediction results.

[0055] Table 1 Model hyperparameter description and value table

[0056]

Claims

1. A coal structure recognition method based on an improved convolutional neural network, characterized in that The steps include: (a) Collect well logging curves (1) of core drilling holes in and around the mining area for geological exploration, and select five well logging curves (1) that are sensitive to coal body structure response as five features (8) of sample data; (b) determining the position of the coal seam (2) in the formation encountered by the borehole according to the drilling core log or the shape of the logging curve (1), and selecting the logging data of the coal seam (2) section of the borehole; (c) Eliminate the dimension and scale differences between different well logging curves (1); (d) reading a plurality of lines of logging data having the same logging sampling point label (10) in the coal seam section and continuous sampling depths (7) to form a logging feature vector (9) of a sample (12); (e) constructing a convolutional neural network coal structure recognition model, which consists of a convolutional layer (17), a pooling layer (18) and two fully connected layers (19). The input of the convolutional layer (17) is set to a single-channel image to process the well logging feature vector (9) of the sample; (f) Use grid search to find the best hyperparameter combination; (g) Use the training set (14) to train the convolutional neural network model and input the new logging data into the prediction model to obtain the predicted coal structure label.

2. The method for identifying coal structure based on an improved convolutional neural network according to claim 1, characterized in that: In step (a), the data of logging curves (1) of coring boreholes for geological exploration in and around the mining area are collected, the vertical change characteristics of the coal body structure of the borehole coal seam (2) are obtained through the coal core logging data, and the changes in the logging values ​​of coals with different coal body structures are statistically analyzed, and the logging curve (1) with better regularity of logging value changes as the degree of coal body structure damage increases is selected as the characteristic (8) curve for coal body structure logging identification; the logging curve (1) includes five logging curves: density logging curve (DEN), acoustic time difference logging curve (AC), caliper logging curve (CAL), resistivity logging (R) and natural gamma logging (GR).

3. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (b), the well logging data of the coal seam (2) section of the borehole is selected and each original well logging curve (5) is decomposed using wavelet decomposition (4) to remove high-frequency noise in the well logging data, retain the low-frequency approximate curve (6) after secondary decomposition, and improve the data signal-to-noise ratio; that is, the depth range of the coal seam (2) in the borehole is determined based on the core logging data or the shape of the well logging curve (1), and the well logging data in the depth range of the coal seam (2) is selected, and the original well logging curve (5) is decomposed using sym8 wavelet decomposition (4), and each level of decomposition obtains a high-frequency signal curve and a low-frequency similarity curve (6), and the original well logging curve (5) is subjected to secondary decomposition to remove the high-frequency curve, and the similarity curve (6) after the secondary decomposition is retained for coal body structure identification.

4. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (c), the dimension differences and scale differences between different logging curves (1) are eliminated by preprocessing the logging data using data normalization (3), so that the model is more easily converged and the training efficiency is improved. The data normalization (3) subtracts the minimum value from the logging curve (1) and divides it by the difference between the maximum value and the minimum value to scale the data to the range of [0, 1], thereby eliminating the dimension differences between different logging curves (1).

5. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (d), the sampling depth (7) of the logging sampling point is spaced at 0.05 m, and each logging sampling point is composed of five characteristics (8) of the sampling point formed by the logging values ​​of five logging curves (1). Each sampling point generates a line of logging data, and the n sampling points of the coal seam (2) section form n lines of logging data. The coal body structure information of each sampling point is obtained through the coal core logging data as the label (10) of the sampling point; the sample (12) reading process starts from the third line of logging data, and it is determined whether the label (10) of the i-th logging sampling point is consistent with the first two sampling points and the last two sampling points, and whether the logging sampling depths (7) of the five sampling points are continuous. If the conditions are met, the i-th logging sampling point and the upper and lower i-th sampling points are read. The logging data of the i-th sampling points constitute a 5×5 logging feature vector (9) of a sample (12), and the label (10) of the i-th sampling point is used as the sample label (11) of the sample (12); the above operation is repeated with one logging sampling point interval to read the next sample (12) until the n-2-th row of logging data is read to obtain the last sample (12); the logging data of all coal seams (2) in the geological exploration core drilling hole are read to generate samples (12) to obtain a sample set (13); the sample labels (11) of the samples (12) in the sample set (13) are encoded, and finally 20% of the samples (12) are randomly selected from the sample set (13) as a test set (15), and the remaining samples (12) are used as a training set (14).

6. The method for identifying coal structure based on improved convolutional neural network according to claim 5 is characterized in that: The sample labels (11) use the numbers "1", "2", and "3" to represent three different types of coal body structures. Before the samples (12) are input into the model, the categorical data are first converted into integers between 0 and n-1 using the LabelEncoder of the sklearn (scikit-learn) library, that is, the three types of coal body structure sample labels (11) of "1", "2", and "3" are converted into three category labels of "0", "1", and "2". Secondly, the LabelEncoder of the sklearn library is used to convert the sample labels (11) into one-hot encoding (One-HotEncoding). 20% of the samples are extracted from the sample set (13) as a test set (15), and the sample segmentation tool of the sklearn library is used to set the test set ratio to 0.

2. The algorithm automatically extracts 20% of the samples as (12) as a test set (15) for evaluating the model in the model training process, and the remaining samples (12) are used as a training set (14) for updating the model network parameters in the model training process.

7. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (e), the model convolution layer (17) receives a single-channel image with a size of 5×5 as input, the number of convolution kernels (20) of the convolution layer (17) is set to 64, the size of the convolution kernel (20) is set to 2×2, the moving step of the convolution kernel (20) is set to 1, and the convolution kernel (20) of the convolution layer (17) performs convolution calculation on the well logging feature vector (9) of each input sample (12) to obtain 64 feature maps (21); the pooling layer (18) is set to maximum pooling, the size of the pooling kernel function (22) is set to 2×2, the moving step is set to 1, and each feature map (21) is pooled to obtain a pooled feature map (23). The 64 feature maps (21) are processed by the pooling layer ( 18) After processing, 64 pooled feature maps (23) are output; the 64 pooled feature maps (23) are flattened by the flattening layer (24) and converted into one-dimensional feature vectors, and then input into the fully connected layer (19). The fully connected layer (19) includes two layers, the first layer contains 64 neurons (25), and the second layer contains 32 neurons (25). Finally, the output (26) is obtained through the Softmax function; an optimization algorithm is added to optimize the convolutional neural network model, and a ReLU activation function and L2 regularization are added to the convolutional layer. A ReLU activation function and Dropout are added to the fully connected layer (19), and the Dropout ratio is 0.

5. The Adam optimizer is used, and the learning rate is set to 0.01; During the processing of the sample (12), the 5×5 well logging feature vector (9) is regarded as a 5×5 single-channel image input by the model convolution layer (17); the model convolution layer (17) uses "0" to fill the well logging feature vector (9) of each input sample, and uses 64 2×2 convolution kernels (20) to perform convolution calculation on the well logging feature vector (9) to extract local feature information of the well logging feature vector (9), and each well logging feature vector (9) is convoluted to obtain 64 5×5 feature maps (21); the pooling layer (18) uses maximum pooling, the pooling kernel function (22) has a size of 2×2, and a moving step size of 1, and the pooling kernel function (22) scans each feature map (21) to output a 4×4 pooling feature map (23), and the pooling layer (18) outputs a total of 64 4× 4 pooled feature map (23); 64 pooled feature maps (23) are flattened by a flattening layer (24) and aggregated into a one-dimensional feature vector with a total length of 64×4×4=1024; the fully connected layer (19) comprises 64 and 32 neurons (25) respectively, and a mapping relationship between the one-dimensional feature vector and the sample label (11) is established between the neurons through correlation calculation; the output (26) uses a Softmax activation function to convert the fully connected output into the probability of belonging to each category label, and the category with the largest probability is taken as the output (26) category, and the output (26) category label is in the form of "0", "1", and "2", and the output (26) category label is converted back to the coal body structure label (11) in the form of "1", "2", and "3" using LabelEncoder inverse conversion.

8. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (e), the model convolution layer (17) adds a ReLU activation function and adds L2 regularization with a value of 0.001; a Dropout layer is added after each fully connected layer (19), and the Dropout ratio is 0.

5. The Adma optimizer is added to the model to improve the model training efficiency and final performance. The model learning rate is set to 0.01, and a grid search is used to find the best hyperparameter combination. First, the convolution kernel (20), neurons (25), and the range of learning rates are defined. The automatic parameter tuner randomly combines hyperparameters and trains the model, automatically evaluates the model, and finally returns the optimal hyperparameter combination with the highest accuracy and the smallest loss function after the training is completed.

9. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (f), the grid search is used to search for the best hyperparameter combination. Before obtaining the best hyperparameter, the value range of each hyperparameter that needs to be tuned is first defined. The number range of convolution kernels (20) is defined as 32, 64, and 128, the number range of neurons (25) is defined as 32, 64, and 128, and the learning rate range is defined as 0.1, 0.01, and 0.

001. The automatic parameter tuner randomly combines hyperparameters and trains and evaluates the model, and finally returns the optimal hyperparameter combination; the evaluation model records the model accuracy and loss function values ​​after training under each hyperparameter combination, and returns the hyperparameter combination with the highest accuracy and the smallest loss function after training, which is the optimal hyperparameter combination.

10. The method for identifying coal structure based on improved convolutional neural network according to claim 1, characterized in that: In step (g), the convolutional neural network model is trained using the training set (14), the number of model iterations is set to 1000, the neural network parameters are updated using a small batch stochastic gradient descent method and a back propagation algorithm, the small batch (16) size is set to 128, the model updates the neural network parameters each iteration so that the loss function decreases, and the loss function adopts a classification cross entropy loss function; the prediction accuracy and loss function changes after each iteration of the model parameter update are plotted, and after multiple iterations, the global minimum value or approximate value of the loss is finally found, and the accuracy and loss function curves converge. At this time, the training accuracy is the highest and the loss function is the lowest, and a trained coal body structure prediction model is obtained; In each iteration of the convolutional neural network model, 128 samples are randomly selected from the training set to calculate the loss function of the predicted results and the actual results. The loss function adopts the classification cross entropy loss, uses the gradient descent method to update the network parameters, and adopts the back propagation algorithm to update the parameters of each layer of the network, so that the loss function gradually decreases. The difference between the model prediction result and the actual label (11) gradually decreases as the loss function decreases, and the prediction accuracy gradually increases. When the accuracy and loss function curves converge, the loss function reaches the global minimum and the model training is completed. At this time, the model prediction accuracy is the highest.

Images