ResNet-KAN-based proppant particle identification method

Abstract

The invention belongs to the technical field of oil production engineering, and provides a method for identifying proppant particles based on ResNet-KAN. The identification method comprises the following steps: acquiring logging rock debris samples at different well depths in a drilling and coring process; screening and photographing the single particles of the logging rock debris sample; extracting an image of the single particle, measuring the particle size of the single particle, and storing; making a training set and a verification set of the model; a ResNet-KAN residual error neural network model is constructed; training the model by using a training set, and verifying by using a verification set; and identifying the image of the single particle by using the trained model so as to identify the proppant particles. According to the method, the KAN network is used for replacing a linear full connection layer of a traditional ResNet network, so that further optimization of the ResNet network is achieved, and the method has the advantage that the propping agent particles in the coring well logging rock debris in the hydraulic fracturing test field can be automatically, rapidly and accurately recognized.

Description

Technical Field

[0001] The present disclosure relates to the technical field of oil production engineering, and in particular to a method for identifying proppant particles based on ResNet-KAN. Background Art

[0002] In order to better study the crack expansion law and influencing factors of crack morphology in shale oil reservoirs during hydraulic fracturing, evaluate the adaptability of horizontal well hydraulic fracturing technology in shale oil reservoirs, and further optimize fracturing process parameters, coring of high-angle wells in hydraulic fracturing test sites is the most direct research method. It can more intuitively understand the characteristics of the fracture network, analyze and evaluate the migration and laying effects of proppants, and play a great guiding role in the subsequent fracturing construction design. Therefore, it is very important to clarify the distribution of proppant content in different parts of the coring well. At present, the method of determining the proppant content in different parts of the coring well is to rely on manual naked eye identification of proppant particles in logging cuttings at different parts of the coring well under a microscope. The identification speed is slow, the efficiency is low, it takes a long time, and it consumes a lot of manpower. Summary of the invention

[0003] The present disclosure provides a method for identifying proppant particles based on ResNet-KAN to at least solve the above technical problems existing in the prior art.

[0004] According to a first aspect of the present disclosure, a method for identifying proppant particles based on ResNet-KAN is provided, comprising the following steps: S1: Obtain logging cuttings samples at different well depths during drilling and coring; S2: screening the logging rock cuttings sample and taking photos of single particles of the logging rock cuttings sample; S3: extracting the image of the single particle, measuring the particle size of the single particle, and saving it; S4: Make training set and validation set of the model; S5: Build the ResNet-KAN residual neural network model; S6: Using the training set to train the ResNet-KAN residual neural network model, and using the validation set to validate it; S7: using the ResNet-KAN residual neural network model trained in step S6 to identify the image of the single particle to identify the proppant particle.

[0005] In one embodiment, in step S1, during the coring process, logging cuttings sampling is performed with a sampling length of 2 to 3 m per section, the returned cuttings are sampled and the well depth position is recorded, and each sample is weighed and recorded after the cuttings are dried.

[0006] Specifically, 300-400 g of logging cuttings are taken from each well section.

[0007] More specifically, 300 g of logging cuttings were collected from each well section.

[0008] In one embodiment, step S2 first cleans the logging rock cuttings sample, and then sieves and dries the logging rock cuttings sample.

[0009] Specifically, step S2 is to clean the logging rock cuttings sample, which specifically includes the following steps: cleaning the logging rock cuttings sample with water, oil cleaning agent, and flushing spray gun in sequence.

[0010] Specifically, step S2 first places the logging rock cuttings sample in a reagent bottle, then adds water, shakes for 30 to 60 seconds, and then pours the liquid in the reagent bottle into a waste liquid bucket through a sieve; then adds water, and repeats the washing for 5 to 6 times to wash away the mud in the logging rock cuttings sample; then adds an oil cleaning agent, and rinses for 10 to 12 minutes to remove the oily substance attached to the logging rock cuttings sample; finally, the logging rock cuttings sample is cleaned with a flushing spray gun for 5 to 8 minutes to wash away the residual oil cleaning agent.

[0011] Specifically, pouring the liquid in the reagent bottle into the waste liquid bucket through the screen is to prevent logging rock cuttings samples from leaking into the waste liquid bucket when pouring the liquid.

[0012] In one embodiment, step S2 uses a combination of 20 mesh, 40 mesh, 70 mesh and 140 mesh sieves to sieve the logging cuttings sample to obtain sample particles of 20 / 40 mesh, 40 / 70 mesh and 70 / 140 mesh.

[0013] Specifically, the sample particles sieved above correspond to the proppants of 20 / 40 mesh, 40 / 70 mesh and 70 / 140 mesh used in the fracturing site, respectively.

[0014] In one possible implementation manner, after the screening is completed, step S2 dries the screened logging cuttings sample at 70-75° C.

[0015] In one possible implementation manner, step S2 divides the logging rock cuttings sample into 9 regions on average, and samples are taken from each region to obtain the single particle.

[0016] Specifically, the mass of the single particle is 0.2-0.5 g.

[0017] Specifically, after sampling in each area, the obtained single particle is weighed and recorded.

[0018] In one embodiment, step S2 uses an electron microscope to take a picture of the single particle.

[0019] In one possible implementation mode, step S3 uses an automatic image segmentation program to pre-process the image of the single particle, identify the edge contour of each single particle, and segment and extract the image of the single particle, and save it in a single particle image database; and automatically measure the particle size of each single particle and save it in a particle size database.

[0020] Specifically, in step S3, Python and OpenCV libraries are used to compile an automatic image segmentation program, and the automatic image segmentation program performs the pre-processing on the input single particle image through grayscale processing, image closing operation, and Gaussian filtering.

[0021] Specifically, step S3 uses the Canny edge recognition algorithm in the automatic image segmentation program to identify the edge contour of each single particle, and makes a circumscribed rectangle of the edge contour to segment and extract the image of the single particle, and save it in the single particle image database; and automatically measures the particle size of each single particle and saves it in the particle size database.

[0022] In one embodiment, step S4 manually selects 500 to 700 proppant particle images and 500 to 700 formation rock cuttings particle images from the single particle image database of step S3, and then uses an image enhancement program to expand the data sets of the two images, and prepares the training set and the validation set in a ratio of 8:2 between the number of proppant particle image data sets and the number of formation rock cuttings particle image data sets.

[0023] Specifically, Python is used to compile the image enhancement program, and then the image enhancement program is used to perform enhancement operations of randomly scaling, rotating and flipping the proppant particle images and the formation rock cuttings particle images to expand the data sets of the two images. Finally, the size of all images is adjusted to 200×200, and the training set and the validation set are prepared in a ratio of 8:2 between the number of the proppant particle image data set and the number of the formation rock cuttings particle image data set.

[0024] Specifically, the ResNet-KAN residual neural network model constructed in step S5 can effectively avoid the "degeneration phenomenon" that occurs in the convolutional neural network as the number of network layers increases. It solves the problem of difficulty in training neural networks with excessive depth and too many neural units to a large extent through "quick connections".

[0025] In one embodiment, the ResNet residual neural network is composed of residual blocks, each of which is represented as: ; in, f is the activation function; xi and y i is the input and output of the i-th layer function; ω i is the weight matrix; Indicates the weight is ω i The residual mapping learned by the layer, function h is .

[0026] Specifically, the most prominent innovation of the ResNet-KAN residual neural network model is that the gradient propagation path of the neural network is branched, which effectively solves the "gradient explosion" and "gradient vanishing" problems that occur with the increase in the number of network layers of the ResNet model. The model connects the residual blocks and shortcuts, allowing some layers of the neural network to skip the connection with the next layer of neurons, connecting the layers in between, and weakening the strong connection between each layer. It successfully solves the "gradient vanishing" and other problems that occur with the increase in the number of ResNet layers. Therefore, as the number of neural network layers increases, its accuracy will not drop significantly. The basic unit of the neural network model is the residual unit. The residual unit constructed in the present disclosure is improved on the basis of the classic residual unit, and the number of convolutional layers and batch normalization layers is increased to better extract the features of each dimension of the image and improve the recognition accuracy.

[0027] In one embodiment, the ResNet-KAN residual neural network model includes: an input layer, a convolution layer, a BN layer, a maximum pooling layer, a residual part RS-1, a residual part RS-2, a residual part RS-3, a residual part RS-4, a global average pooling layer, a Flatten layer, and a KAN network.

[0028] Specifically, the input layer is used to input a three-channel single particle color image.

[0029] Specifically, the convolution layer is a 7×7×3 convolution layer. The convolution layer performs convolution on the three-channel single-particle color image to obtain a feature map output, thereby extracting features of each dimension of the image.

[0030] Specifically, the BN layer, i.e., batch normalization layer, can solve the problem of changes in the data distribution of the intermediate layer during the training process, so as to prevent the gradient from disappearing or exploding and speed up the training.

[0031] Specifically, the residual part RS-1 consists of three type A residual blocks.

[0032] Specifically, the residual part RS-2 consists of one B-1 type residual block and three B-2 type residual blocks.

[0033] Specifically, the residual part RS-3 consists of one C-1 type residual block and five C-2 type residual blocks.

[0034] Specifically, the residual part RS-4 consists of one D-1 type residual block and three D-2 type residual blocks.

[0035] Specifically, the Flatten layer converts multi-dimensional input into one dimension.

[0036] Specifically, if the traditional ResNet network is used, the next layer should be a linear fully connected layer, but the present disclosure innovatively uses a KAN network with a hidden layer containing 64 neurons to replace the linear fully connected layer, thereby further optimizing the ResNet network. The final output neural unit of the network is two and uses a Softmax activation function, thereby achieving a binary classification task of whether a single particle image is a proppant particle.

[0037] Specifically, the design of the KAN network is inspired by the Kolmogorov-Arnold representation theorem, which profoundly reveals that multivariate continuous functions can be represented by a finite combination of single-variable continuous functions and binary addition operations. The core of the theorem is to accurately select the appropriate single-variable function Φ p (internal function) and Φ q (External function). In the process of constructing the KAN network, B-spline is used as the function basis. The B-spline shows consistent continuity in the domain and at the nodes, and its polynomial representation strictly follows the mathematical description of the Cox-de Boor recursive formula.

[0038]

[0039]

[0040] in, t i is a node in the node vector, k is the order of the B-spline, B i,k (x) It is i indivual k Order B-spline basis function, x is the independent variable.

[0041] The internal function of the KA theorem can be defined as a matrix calculation with parameters. Each element in the matrix is ​​actually a function or operator, where the KAN layer can be defined as: .

[0042] in, Φ prepresents the pth internal univariate function, there are n in An internal univariate function; Φ q represents the qth external univariate function, there are n out An external univariate function.

[0043] The KAN network applies a learnable activation function to the weights instead of using a fixed activation function on the nodes (neurons), processing and learning the complex relationships of the input data in a more flexible way that is closer to the KA representation theorem. Compared with the traditional linear fully connected network, the KAN network has the advantages of higher accuracy, better interpretability, faster neural scaling law, etc., and its performance improves faster as the model size increases. At the same time, it can achieve the accuracy of the original linear fully connected layer with fewer iterations, so the optimized ResNet network can also be called the ResNet-KAN residual neural network.

[0044] In one embodiment, step S6 uses an automatic learning rate update algorithm during the iterative training process to automatically select the optimal batch size batchsize in the range of 24 to 128, and to automatically overwrite and save better network parameters, weights, etc. during the training process, so that the optimal network parameters can be obtained after the iterative training is completed.

[0045] Specifically, thanks to the advantages of the KAN network, the number of iterations of the ResNet-KAN neural network training has been greatly reduced, and the accuracy of its validation set can reach 99.21% after 20 iterations. The accuracy of the traditional ResNet neural network is only 96.18%, and the improved ResNet-KAN neural network can improve the accuracy by 3.03 percentage points.

[0046] In one embodiment, step S7 uses the ResNet-KAN residual neural network model trained in step S6 to identify all the single particle images in the single particle image database to identify all the proppant particles, and combines with the particle size database to count the particle sizes of all the proppant particles.

[0047] Specifically, the above method is used to identify proppant particles, which can provide basic data for subsequent hydraulic fracturing test field analysis and evaluation of proppant migration and placement effects.

[0048] According to an implementable embodiment of the present disclosure, at least the following beneficial effects are achieved: The present invention uses a KAN network with a hidden layer containing 64 neurons to replace the linear fully connected layer of the traditional ResNet network, thereby further optimizing the ResNet network. Compared with the traditional linear fully connected network, the KAN network has the advantages of higher accuracy, better interpretability, faster neural scaling law, etc., and its performance improves faster as the model scale increases. At the same time, it can achieve the accuracy of the original linear fully connected layer with fewer iterations.

[0049] The ResNet-KAN residual neural network model disclosed in the present invention branches the gradient propagation path of the neural network, effectively solving the problems of "gradient explosion" and "gradient vanishing" that appear as the number of network layers of the ResNet model increases. The model connects the residual block and the shortcut, allowing some layers of the neural network to skip the connection of the next layer of neurons, connecting the layers, weakening the strong connection between each layer, and successfully solving the problems of "gradient vanishing" that appear as the number of ResNet layers increases. Therefore, as the number of neural network layers increases, its accuracy will not drop significantly. In addition, the residual unit constructed in the present invention is improved on the basis of the classic residual unit, and the number of convolutional layers and batch normalization layers is increased to better extract the features of each dimension of the image and improve the recognition accuracy.

[0050] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which: In the drawings, the same or corresponding reference numerals represent the same or corresponding parts.

[0052] Figure 1 A schematic diagram of a single particle image automatically segmented and extracted in Example 1 of the present disclosure is shown; Figure 2 A basic structural diagram of a residual unit in Embodiment 1 of the present disclosure is shown; Figure 3 A schematic diagram showing the structure of a type A residual block in Embodiment 1 of the present disclosure is shown; Figure 4 A schematic diagram showing the structure of a B-1 type residual block in Embodiment 1 of the present disclosure is shown; Figure 5 A schematic diagram showing the structure of a B-2 type residual block in Embodiment 1 of the present disclosure is shown; Figure 6 A schematic diagram of the structure of a C-1 type residual block in Embodiment 1 of the present disclosure is shown; Figure 7 A schematic diagram of the structure of a C-2 type residual block in Embodiment 1 of the present disclosure is shown; Figure 8 A schematic diagram of the structure of a D-1 type residual block in Embodiment 1 of the present disclosure is shown; Fig. 9 A schematic diagram of the structure of a D-2 type residual block in Embodiment 1 of the present disclosure is shown; Fig.10 A schematic diagram of the structure of the ResNet-KAN residual neural network model in Embodiment 1 of the present disclosure is shown; Fig.11 The figure shows the accuracy test results of the validation set of the ResNet-KAN neural network and the traditional ResNet neural network in the iterative training process in Example 1 of the present disclosure. DETAILED DESCRIPTION

[0053] In order to make the purpose, features, and advantages of the present disclosure more obvious and easy to understand, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0054] Example 1 This embodiment provides a method for identifying proppant particles based on ResNet-KAN, which is as follows: (1) During the core well drilling process, logging cuttings sampling is carried out with a sampling length of 2 m per section. The returned drilling cuttings are sampled and the well depth position is recorded. About 300 g of cuttings are taken from each section. After the cuttings are completely dried, each cutting sample is weighed and recorded.

[0055] (2) Pour all the collected rock cuttings into a polyethylene reagent bottle with a capacity of 1000mL, then inject an appropriate amount of clean water into the bottle, cover the bottle cap and shake vigorously for 30 seconds, then pour the upper liquid in the bottle into the waste liquid bucket through a 140-mesh stainless steel screen (the screen is to prevent rock cuttings samples from leaking into the waste liquid bucket when pouring waste liquid); then add clean water to the reagent bottle and repeat this process 5 to 6 times until most of the soil in the rock cuttings is washed away. Then add oil cleaning agent to the reagent bottle and rinse repeatedly for 10 minutes to remove the oily substances (crude oil, oily mud, etc.) attached to the rock cuttings. Finally, use a flushing spray gun to repeatedly rinse the rock cuttings for 5 minutes to wash away the residual oil cleaning agent. The cleaned rock cuttings samples were screened, and the 20 mesh, 40 mesh, 70 mesh and 140 mesh screens were combined from top to bottom, and the cleaned rock cuttings samples were screened using the screen combination to screen out 20 / 40 mesh particles, 40 / 70 mesh particles and 70 / 140 mesh particles, which correspond to the 20 / 40 mesh proppant, 40 / 70 mesh proppant and 70 / 140 mesh proppant used at the fracturing site. After the screening is completed, the screened rock cuttings samples of different sizes are placed in different glass culture dishes, and then placed in a drying box, and the temperature is set at 70°C for drying until the sample particles are completely dry.

[0056] (3) Use a nine-square grid to divide the dried rock cuttings sample in the culture dish into 9 equal areas. Use a microsampling spoon to take samples from each area. The mass of the final single particle (i.e., the observation sample) is controlled to be around 0.2 g. It is weighed on an analytical balance with an accuracy of 0.0001 g and recorded.

[0057] (4) The obtained single particles are evenly distributed on a transparent glass plate with a size of 350 mm × 350 mm, the glass plate is placed on an LED light-emitting panel, and then the brightness of the LED light-emitting panel is adjusted to an appropriate level, and then photographed using a high-resolution industrial electron microscope.

[0058] (5) Use Python and OpenCV library to compile an automatic image segmentation program. After the input single particle image is preprocessed by grayscale processing, image closing operation, Gaussian filtering and other preprocessing operations, the program uses the Canny edge recognition algorithm in the automatic image segmentation program to identify the edge contour of each single particle, and then creates a circumscribed rectangle of the contour to segment and extract the single particle image. This program can automatically identify and segment the single particle images taken previously, and extract the single particle images (such as Figure 1 ) is saved in the single particle image database; and the particle size of each single particle can be automatically measured and saved in the particle size database.

[0059] (6) Producing training and validation sets: About 500 proppant particle images and 500 formation rock cutting particle images were manually selected from the single particle image database in the previous step. Then, an image enhancement program was compiled using Python. The program randomly scaled, rotated, and flipped these images (proppant particle images and formation rock cutting particle images) to expand the dataset. Finally, all images were resized to 200 × 200. The training and validation sets were then prepared in a ratio of 8:2 between the number of proppant particle image datasets and the number of formation rock cutting particle image datasets.

[0060] (7) Construct the ResNet-KAN residual neural network model, which can effectively avoid the "degeneration phenomenon" that occurs in convolutional neural networks as the number of network layers increases. It solves the problem of difficulty in training neural networks with too large depth and too many neural units to a large extent through "quick connection". The ResNet residual neural network is composed of residual blocks, and each residual block can be expressed as:

[0061] in, f is the activation function; x i and y i is the input and output of the i-th layer function; ω i is the weight matrix; Indicates the weight is ω i The residual mapping learned by the layer, function h is .

[0062] The most prominent innovation of the ResNet-KAN residual neural network is that it branches the gradient propagation path of the neural network, effectively solving the "gradient explosion" and "gradient vanishing" problems that occur with the increase in the number of network layers of the ResNet model. The model connects the residual block and the shortcut, allowing some layers of the neural network to skip the connection with the next layer of neurons, connecting the layers in between, and weakening the strong connection between each layer. It successfully solves the problems such as "gradient vanishing" that occur with the increase in the number of ResNet layers. Therefore, as the number of neural network layers increases, its accuracy will not drop significantly. The basic unit of the neural network model is the residual unit. The residual unit constructed in this embodiment is improved on the basis of the classic residual unit, and the number of convolutional layers and batch normalization layers is increased to better extract the features of each dimension of the image and improve the recognition accuracy. The structure of the residual unit is as follows Figure 2As shown in the figure (where weight layer represents the convolutional layer, relu represents the activation function, and identity represents the addition of the transformed input X to the output of the convolutional layer).

[0063] The structure of the constructed ResNet-KAN residual neural network model (see Fig.10 )for: The first layer is the input layer, and the input data is a three-channel single-particle color image of size 200×200×3; The second layer is the convolution layer. The input is a three-channel single-particle color image of size 200×200×3. The convolution kernel size is 7×7×3, the number of convolution kernels is 64, the padding is 3, and the stride is 2. After convolution, the feature map output of size 100×100×64 is obtained, thereby extracting the features of each dimension of the image. The third layer is the BN layer, i.e. the batch normalization layer, which can solve the problem of changes in the data distribution of the intermediate layer during the training process, so as to prevent the gradient from disappearing or exploding and speed up the training; The fourth layer is the maximum pooling layer with an input size of 100×100×64, a pooling kernel size of 3×3, a padding of 1, and a stride of 2. After the maximum pooling operation, the feature map output of the pooling layer is 50×50×64.

[0064] After that, the feature map data enters the first residual part RS-1 of the entire neural network model, which consists of three A-type residual blocks with exactly the same structure. The structure of the A-type residual block is as follows (see Figure 3 ): The first layer is the convolution layer. The input feature map size is 50×50×64, the convolution kernel size is 3×3×64, the number of convolution kernels is 64, the padding is 1, and the stride is 1. After convolution, the feature map output size is 50×50×64. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 50×50×64. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 50×50×64. The third layer is the convolution layer. The input feature map size is 50×50×64, the convolution kernel size is 3×3×64, the number of convolution kernels is 64, the padding is 1, and the stride is 1. After convolution, the feature map output size is 50×50×64. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 50×50×64. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 50×50×64. The fifth layer is the convolution layer. The input data size is 50×50×64, the convolution kernel size is 3×3×64, the number of convolution kernels is 64, the padding is 1, and the stride is 1. After convolution, the feature map output is 50×50×64. The sixth layer is the batch normalization layer. The input size of this layer is 50×50×64. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 50×50×64. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer to obtain the final output feature map and activate it through the Relu activation function. The output feature map size is 50×50×64.

[0065] After being processed by the three A-type residual blocks, the feature map data enters the second residual part RS-2 of the entire neural network model, which consists of a B-1-type residual block and three B-2-type residual blocks. The structure of the B-1-type residual block is as follows (see Figure 4 ): The first layer is the convolution layer. The input feature map size is 50×50×64, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, the stride is 2, and after convolution, the feature map output size is 25×25×128. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 25×25×128. The third layer is the convolution layer. The input feature map size is 50×50×64, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, the stride is 2, and after convolution, the feature map output size is 25×25×128. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 25×25×128. The fifth layer is the convolution layer. The input data size is 25×25×128, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, and the stride is 1. After convolution, the feature map output is 25×25×128. The sixth layer is the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 25×25×128. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer, but the width, height and number of channels of the two are different, one is 50×50×64 and the other is 25×25×128. Therefore, 128 convolution kernels are used, the convolution kernel size is 1×1×128, filled with 0, and the stride is 2, and the output feature map is 25×25×128. At this time, the two sets of feature maps can be added and convolved to obtain a 25×25×128 feature map, and then the output feature map is activated by the Relu activation function. After activation, the output feature map size is 25×25×128.

[0066] The structure of the B-2 residual block is as follows (see Figure 5 ): The first layer is the convolution layer. The input feature map size is 25×25×128, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, and the stride is 1. After convolution, the feature map output size is 25×25×128. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 25×25×128. The third layer is the convolution layer. The input feature map size is 25×25×128, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, and the stride is 1. After convolution, the feature map output size is 25×25×128. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 25×25×128. The fifth layer is the convolution layer. The input data size is 25×25×128, the convolution kernel size is 3×3×128, the number of convolution kernels is 128, the padding is 1, and the stride is 1. After convolution, the feature map output is 25×25×128. The sixth layer is the batch normalization layer. The input size of this layer is 25×25×128. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 25×25×128. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer to obtain the final output feature map and activate it through the Relu activation function. The output feature map size is 25×25×128.

[0067] After being processed by the above-mentioned B-1 residual block and three B-2 residual blocks, the feature map data enters the third residual part RS-3 of the entire neural network model, which consists of one C-1 residual block and five C-2 residual blocks. The structure of the C-1 residual block is as follows (see Figure 6 ): The first layer is the convolution layer. The input feature map size is 25×25×128, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, the stride is 2, and after convolution, the feature map output size is 13×13×256. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 13×13×256. The third layer is the convolution layer. The input feature map size is 25×25×128, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, the stride is 2, and after convolution, the feature map output size is 13×13×256. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 13×13×256. The fifth layer is the convolution layer. The input data size is 13×13×256, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, and the stride is 1. After convolution, the feature map output is 13×13×256. The sixth layer is the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 13×13×256. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer, but the width, height and number of channels of the two are different, one is 25×25×128 and the other is 13×13×256. Therefore, 256 convolution kernels are used, the convolution kernel size is 1×1×256, the padding is 0, the stride is 2, and the convolution outputs a feature map of 13×13×256. At this time, the two sets of feature maps can be added to obtain a feature map of 13×13×256, and then the output feature map is activated by the Relu activation function. After activation, the output feature map size is 13×13×256.

[0068] The structure of the C-2 residual block is as follows (see Figure 7 ): The first layer is the convolution layer. The input feature map size is 13×13×256, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, and the stride is 1. After convolution, the feature map output size is 13×13×256. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 13×13×256. The third layer is the convolution layer. The input feature map size is 13×13×256, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, and the stride is 1. After convolution, the feature map output size is 13×13×256. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 13×13×256. The fifth layer is the convolution layer. The input data size is 13×13×256, the convolution kernel size is 3×3×256, the number of convolution kernels is 256, the padding is 1, and the stride is 1. After convolution, the feature map output is 13×13×256. The sixth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 13×13×256. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 13×13×256. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer to obtain the final output feature map and activate it through the Relu activation function. The output feature map size is 13×13×256.

[0069] After being processed by the above-mentioned C-1 residual block and five C-2 residual blocks, the feature map data enters the fourth residual part RS-4 of the entire network model, which consists of one D-1 residual block and three D-2 residual blocks. The structure of the D-1 residual block is as follows (see Figure 8 ): The first layer is the convolution layer. The input feature map size is 13×13×256, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, the stride is 2, and after convolution, the feature map output size is 7×7×512. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 7×7×512. The third layer is the convolution layer. The input feature map size is 13×13×256, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, the stride is 2, and after convolution, the feature map output size is 7×7×512. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 7×7×512. The fifth layer is the convolution layer. The input data size is 7×7×512, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, and the stride is 1. After convolution, the feature map output is 7×7×512. The sixth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 7×7×512. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer, but the width, height and number of channels of the two are different, one is 13×13×256 and the other is 7×7×512. Therefore, 512 convolution kernels are used, the convolution kernel size is 1×1×512, the padding is 0, the stride is 2, and the convolution outputs a feature map of 7×7×512. At this time, the two sets of feature maps can be added to obtain a feature map of 7×7×512, and then the output feature map is activated by the Relu activation function. After activation, the output feature map size is 7×7×512.

[0070] The structure of the D-2 residual block is as follows (see Fig. 9 ): The first layer is the convolution layer. The input feature map size is 7×7×512, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, and the stride is 1. After convolution, the feature map output size is 7×7×512. The second layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, it is activated by the Relu activation function. The output feature map size remains unchanged, which is still 7×7×512. The third layer is the convolution layer. The input feature map size is 7×7×512, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, and the stride is 1. After convolution, the feature map output size is 7×7×512. The fourth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, it is activated by the Relu activation function. The size of the output feature map remains unchanged, which is still 7×7×512. The fifth layer is the convolution layer. The input data size is 7×7×512, the convolution kernel size is 3×3×512, the number of convolution kernels is 512, the padding is 1, and the stride is 1. After convolution, the feature map output is 7×7×512. The sixth layer is the BN layer, i.e. the batch normalization layer. The input size of this layer is 7×7×512. After the input feature map is batch normalized, the output feature map size remains unchanged, which is 7×7×512. The seventh layer is the addition activation layer. In this layer, the residual block adds the initial input feature map and the feature map output by the sixth layer to obtain the final output feature map and activate it through the Relu activation function. The output feature map size is 7×7×512.

[0071] After being processed by the above-mentioned D-1 residual block and three D-2 residual blocks, the feature map data enters the global average pooling layer, and the input is 7×7×512. After pooling, a feature map output of size 1×1×512 is obtained, and then the feature map enters the Flatten flattening layer, which converts the multi-dimensional input into one dimension, that is, flattens the 1×1×512 feature map input into a 1×512 output. If the traditional ResNet network is followed, the next should be a linear fully connected layer, but this embodiment uses a KAN network with a hidden layer containing 64 neurons to replace the linear fully connected layer, thereby achieving further optimization of the ResNet network. The final output neural unit of the network is two and uses the Softmax activation function, thereby achieving the binary classification task of whether the single particle image is a proppant particle.

[0072] The design of the KAN network is inspired by the Kolmogorov-Arnold representation theorem, which profoundly reveals that multivariate continuous functions can be represented by a finite combination of single-variable continuous functions and binary addition operations. The core lies in the accurate selection of appropriate single-variable functions Φp (internal function) and Φq (external function). In the construction of the KAN network, B-spline is used as the function basis. The B-spline shows consistent continuity in the domain and at the nodes, and its polynomial representation strictly follows the mathematical description of the Cox-de Boor recursive formula.

[0073]

[0074]

[0075] in, t i is a node in the node vector, k is the order of the B-spline, B i,k (x) It is i indivual k Order B-spline basis function, x is the independent variable.

[0076] The internal function of the KA theorem can be defined as a matrix calculation with parameters. Each element in the matrix is ​​actually a function or operator, where the KAN layer can be defined as:

[0077] in, Φ p represents the pth internal univariate function, there are n in An internal univariate function; Φ q represents the qth external univariate function, there are n out An external univariate function.

[0078] The KAN network applies a learnable activation function to the weights instead of using a fixed activation function on the nodes (neurons), processing and learning the complex relationships of the input data in a more flexible way that is closer to the KA representation theorem. Compared with the traditional linear fully connected network, the KAN network has the advantages of higher accuracy, better interpretability, faster neural scaling law, etc., and its performance improves faster as the model size increases. At the same time, it can achieve the accuracy of the original linear fully connected layer with fewer iterations. Therefore, the optimized ResNet network is also called the ResNet-KAN network. Its overall network structure is as follows: Fig.10 shown.

[0079] (8) The ResNet-KAN network is trained using the training set prepared in step (6), and is validated using the validation set. During the iterative training process, the CPU used is Intel i5-14600KF, the GPU is NVIDIA GeForce RTX4070Ti 12G, and the learning rate automatic update algorithm is used. The program can automatically select the optimal batch size in the range of 24 to 128. At the same time, during the training process, whenever a better result is obtained in the iteration, that is, when the accuracy of the validation set is higher, the better network parameters, weights, etc. will be saved, so that the optimal network parameters can be saved after the iterative training is completed. Thanks to the advantages of the KAN network, the number of iterative training of the ResNet-KAN neural network is greatly reduced. After 20 iterations, the accuracy of the validation set can reach 99.21%; while the accuracy of the traditional ResNet neural network is only 96.18%. The improved ResNet-KAN neural network can improve the accuracy by 3.03 percentage points (see Fig.11 ).

[0080] (9) The trained ResNet-KAN residual neural network model is used to identify all single-particle images in the single-particle image database, identify all the proppant particles therein, and combine with the particle size database to count the particle sizes of all proppant particles. This can provide basic data for the subsequent analysis and evaluation of proppant migration and laying effects in hydraulic fracturing test sites.

[0081] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of this disclosure can be achieved, and this document is not limited here.

[0082] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present disclosure, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0083] The above is only a specific embodiment of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present disclosure, which should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for identifying proppant particles based on ResNet-KAN, characterized in that: The following steps are involved: S1: Obtain logging cuttings samples at different well depths during drilling and coring; S2: screening the logging rock cuttings sample and taking photos of single particles of the logging rock cuttings sample; S3: extracting the image of the single particle, measuring the particle size of the single particle, and saving it; S4: Make training set and validation set of the model; S5: Build the ResNet-KAN residual neural network model; S6: Using the training set to train the ResNet-KAN residual neural network model, and using the validation set to validate it; S7: using the ResNet-KAN residual neural network model trained in step S6 to identify the image of the single particle to identify the proppant particle.

2. The identification method according to claim 1, characterized in that: Step S2: firstly clean the logging rock cuttings sample, and then screen and dry the logging rock cuttings sample.

3. The identification method according to claim 1, characterized in that: Step S2: sieve the logging rock cuttings sample with a combination of 20 mesh, 40 mesh, 70 mesh and 140 mesh sieves to obtain sample particles of 20 / 40 mesh, 40 / 70 mesh and 70 / 140 mesh.

4. The identification method according to claim 1, characterized in that: Step S2: the logging rock cuttings sample is evenly divided into 9 areas, and samples are taken from each area to obtain the single particle; the mass of the single particle is 0.2-0.5 g.

5. The identification method according to claim 1, characterized in that: In step S3, the image of the single particle is preprocessed by an automatic image segmentation program, the edge contour of each single particle is identified, and the image of the single particle is segmented and extracted, and saved in a single particle image database; and the particle size of each single particle is automatically measured and saved in a particle size database; The automatic image segmentation program performs the pre-processing on the input single particle image by grayscale processing, image closing operation and Gaussian filtering.

6. The identification method according to claim 5, characterized in that: Step S4: manually selecting 500 to 700 proppant particle images and 500 to 700 formation rock cutting particle images from the single particle image database of step S3, and then using an image enhancement program to expand the data sets of the two images, and preparing the training set and the validation set in a ratio of 8:2 between the number of proppant particle image data sets and the number of formation rock cutting particle image data sets; The image enhancement program is used to perform enhancement operations of random scaling, rotation and flipping on the proppant particle image and the formation rock cutting particle image to expand the data sets of the two images.

7. The identification method according to claim 1, characterized in that: The ResNet-KAN residual neural network model includes: an input layer, a convolution layer, a BN layer, a maximum pooling layer, a residual part RS-1, a residual part RS-2, a residual part RS-3, a residual part RS-4, a global average pooling layer, a Flatten flattening layer, and a KAN network.

8. The identification method according to claim 7, characterized in that: In the ResNet-KAN residual neural network model, the residual part RS-1 consists of three class A residual blocks; the residual part RS-2 consists of one class B-1 residual block and three class B-2 residual blocks; the residual part RS-3 consists of one class C-1 residual block and five class C-2 residual blocks; the residual part RS-4 consists of one class D-1 residual block and three class D-2 residual blocks.

9. The identification method according to claim 1, characterized in that: In step S6, during the iterative training process, an automatic learning rate update algorithm is used to automatically select the optimal batch size in the range of 24 to 128, and to automatically overwrite and save better network parameters and weights during the training process, so that the optimal network parameters can be obtained after the iterative training is completed.

10. The identification method according to claim 5, characterized in that: Step S7 uses the ResNet-KAN residual neural network model trained in step S6 to identify all the single particle images in the single particle image database to identify all the proppant particles, and combines the particle size database to count the particle sizes of all the proppant particles.

Images