An image-based proppant transport condition prediction method and system

Abstract

The application provides a proppant migration state prediction method and system based on images, and relates to the technical field of oil and gas development.The method comprises the following steps: processing and predicting a real-time image sequence currently collected by using a prediction neural network model to obtain current placement form prediction data of the proppant in a complex fracture; obtaining a regulation instruction based on the current placement form prediction data, and sending the regulation instruction to a field fracturing construction control unit to close-loop control the proppant migration process.The application realizes accurate prediction and closed-loop control of the proppant migration state through downhole image collection, preprocessing, segmentation correction, multi-source data fusion modeling and real-time prediction and regulation.

Description

Technical Field

[0001] This invention relates to the field of oil and gas development technology, and in particular to an image-based method and system for predicting proppant migration status. Background Technology

[0002] The efficient development of unconventional oil and gas resources such as shale oil and shale gas relies on hydraulic fracturing technology. The proppant migration state within the fracture directly affects the final placement effect and fracture conductivity. Therefore, real-time and accurate prediction of proppant migration state is key to achieving intelligent control of fracturing operations.

[0003] With the development of downhole visualization imaging technology, some field sites have begun to attempt to collect real-time image data of proppant migration using downhole camera equipment. However, existing image-based prediction methods have the following shortcomings in data processing: random noise, salt-and-pepper noise, and uneven illumination introduced by the complex downhole environment are difficult to suppress effectively; single filtering or simple histogram equalization may lead to the loss of particle edge details while denoising; traditional threshold segmentation or machine learning methods are difficult to accurately distinguish particles, wall impurities, and background under low contrast conditions, and particle adhesion, over-segmentation, and under-segmentation are common; furthermore, the identification of preset spatial reference markers within the fracture is difficult, which may lead to inaccuracies in the number of extracted particles and the contour of the sand embankment. Key parameters have significant errors (field tests show that the sand embankment height error exceeds 15%); image distortion caused by lens distortion and installation angle is not corrected, and position and velocity parameters in the image coordinate system cannot be accurately converted to the physical space of the fracture, which may lead to deviations between dynamic characteristics and the actual physical process; the prediction relies solely on image features and does not integrate multi-dimensional parameters such as fracturing fluid viscosity and construction displacement, as well as CFD-DEM (Computational Fluid Dynamics and Discrete Element Method) numerical simulation data, which may make it difficult to fully reflect the coupling effects of multiple factors and result in poor model generalization ability; the complex processing flow causes a single-frame image prediction lag of more than 30 seconds, which cannot match the real-time acquisition frequency and may make it difficult to support closed-loop control. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an image-based method and system for predicting the migration state of proppant. Through downhole image acquisition, preprocessing, segmentation and correction, multi-source data fusion modeling and real-time prediction and control, the method and system can achieve accurate prediction and closed-loop control of the migration state of proppant.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, an image-based method for predicting proppant transport states is provided, the method comprising: Acquire real-time image sequences of proppant migration within downhole fracturing fractures; Each frame of the real-time image sequence is sequentially subjected to adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement to obtain the proppant image to be segmented. Image segmentation based on convolutional neural networks is performed on the proppant image to be segmented to identify the proppant particle region in the image and multiple spatial reference markers pre-set in the crack, resulting in a segmented image containing the reference markers; a spatial correction grid is constructed based on the reference markers in the segmented image, and the regional distortion features are calculated to obtain correction coefficients; dynamic migration feature parameters of proppant particles are extracted from the segmented image, and the dynamic migration feature parameters are geometrically corrected using the correction coefficients to obtain the feature parameters to be fused. The process involves obtaining the feature parameters to be fused, the multivariate parameters affecting proppant migration, and a simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; associating and fusing the feature parameters to be fused with the multivariate parameters, and combining them with the simulation dataset to construct a feature dataset; and using the feature dataset to train a pre-defined neural network to obtain a predictive neural network model. The current real-time image sequence is processed and predicted using a predictive neural network model to obtain the current proppant placement pattern prediction data in complex fractures; based on the current placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

[0006] Secondly, an image-based proppant transport state prediction system includes: The acquisition module is used to acquire real-time image sequences of proppant migration within downhole fracturing fractures; The preprocessing module is used to sequentially perform adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement on each frame of the real-time image sequence to obtain the proppant image to be segmented. The segmentation and correction module is used to perform image segmentation based on convolutional neural networks on the proppant image to be segmented, identify the proppant particle region in the image and multiple spatial reference markers preset in the crack, and obtain a segmented image containing the reference markers; construct a spatial correction grid based on the reference markers in the segmented image, calculate the regional distortion features, and obtain correction coefficients; extract the dynamic migration feature parameters of the proppant particles from the segmented image, and use the correction coefficients to perform geometric correction on the dynamic migration feature parameters to obtain the feature parameters to be fused. The fusion module is used to obtain the feature parameters to be fused, the multivariate parameters affecting proppant migration, and the simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; it associates and fuses the feature parameters to be fused with the multivariate parameters, and combines them with the simulation dataset to construct a feature dataset; it uses the feature dataset to train a preset neural network to obtain a predictive neural network model; The control module is used to process and predict the currently acquired real-time image sequence using a predictive neural network model to obtain the current proppant placement pattern prediction data in complex fractures; based on the current placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

[0007] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: High-temperature and high-pressure downhole visualization imaging equipment was deployed and images were acquired simultaneously during the sand-addition stage to obtain a continuous real-time image sequence of proppant migration containing timestamps and location information, providing a true and complete foundation of image data for subsequent data processing. Adaptive median filtering, homomorphic filtering, and adaptive histogram equalization were used to process the images sequentially, filtering out image noise, correcting uneven illumination, and stretching grayscale differences to obtain high-quality proppant images for segmentation, providing data support for image segmentation. Convolutional neural network image segmentation techniques were employed to identify proppant particles and spatial reference markers, obtaining segmented images containing the reference markers. A correction grid was constructed using the reference markers, and correction coefficients were calculated to correct geometric distortion of the images. Feature parameters were extracted and combined with geometric correction and polar coordinate transformation using the correction coefficients to obtain the feature parameters to be fused, providing a basis for multi-source data. The system integrates standardized and accurate image feature data; it merges the feature parameters to be merged with diverse construction engineering parameters to expand the feature dimensions and achieve organic integration of images and engineering parameters; it combines a simulation dataset constructed based on computational fluid dynamics and discrete element coupling methods to supplement working condition data and improve feature coverage; it uses the technique of training a pre-set neural network with the feature dataset to allow the network to learn the movement law and obtain a prediction model adapted to complex working conditions, realizing the collaborative utilization of multi-source data; it uses a prediction neural network model to process the current real-time image sequence, quickly outputting proppant placement pattern prediction data to achieve real-time determination of placement pattern; it generates control instructions based on the prediction data and sends them to the construction control unit to adapt construction parameters to the movement state, forming a closed loop throughout the process, realizing closed-loop control of proppant movement, and ensuring the proppant placement effect. Attached Figure Description

[0010] Figure 1This is a schematic flowchart of an image-based proppant migration state prediction method provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of an image-based proppant migration state prediction system provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention proposes an image-based proppant transport state prediction method, the method comprising the following steps: Step 100: Obtain a real-time image sequence of proppant migration within the downhole fracturing fracture; Step 200: For each frame of the real-time image sequence, perform adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement in sequence to obtain the proppant image to be segmented. Step 300: Perform image segmentation based on convolutional neural network on the proppant image to be segmented, identify the proppant particle region in the image and multiple spatial reference markers preset in the crack, and obtain a segmented image containing the reference markers; construct a spatial correction grid based on the reference markers in the segmented image, and calculate the regional distortion features to obtain correction coefficients; extract the dynamic migration feature parameters of the proppant particles from the segmented image, and use the correction coefficients to perform geometric correction on the dynamic migration feature parameters to obtain the feature parameters to be fused. Step 400: Obtain the feature parameters to be fused, and obtain the multivariate parameters that affect proppant migration, as well as the simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; associate and fuse the feature parameters to be fused with the multivariate parameters, and combine them with the simulation dataset to construct a feature dataset; use the feature dataset to train a preset neural network to obtain a predictive neural network model; Step 500: The current real-time image sequence is processed and predicted using a predictive neural network model to obtain the current proppant placement pattern prediction data in complex fractures; based on the current placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

[0014] In this embodiment of the invention, a technique for acquiring real-time image sequences of proppant migration within downhole fracturing fractures is used to obtain continuous dynamic proppant migration data, providing accurate and complete raw data for subsequent data processing and state prediction. Each frame of the real-time image sequence is sequentially subjected to adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement. This filters out image noise, corrects illumination deviations, and improves contrast, resulting in a clear and stable image to be segmented, providing high-quality data for subsequent segmentation. Convolutional neural network image segmentation technology is employed to identify proppant particle regions and spatial reference markers, obtaining segmented images. A spatial correction grid is constructed using the reference markers, and distortion features are calculated to obtain correction coefficients, achieving image geometric distortion correction. The process involves: extracting dynamic migration feature parameters and correcting them with correction coefficients to ensure that the feature parameters accurately reflect the actual migration state of the proppant; acquiring the feature parameters to be fused, multiple influencing parameters, and coupling methods to construct a simulation dataset to expand the data dimensions; using multi-source data association and fusion combined with the simulation dataset to construct a feature dataset to achieve multi-source data integration; training a pre-set neural network using the feature dataset to enable the model to fully learn the proppant migration law and improve its adaptability to complex scenarios; using a predictive neural network model to process real-time image sequences and make predictions to quickly obtain the current proppant placement morphology data; generating control instructions based on the predicted data and sending them to the construction control unit to achieve closed-loop control of proppant migration and ensure the proppant placement effect.

[0015] In a preferred embodiment of the present invention, step 100 includes: Step 101: Deploy high-temperature and high-pressure resistant downhole visualization imaging equipment in advance at key cluster locations in the fracturing well section to obtain the deployed imaging equipment. Specifically, this includes: conducting preliminary geological surveys and construction design for the fracturing well section to determine the key cluster locations for fracturing operations. These key cluster locations need to cover the main paths of proppant migration and key areas prone to accumulation and retention, ensuring that the imaging range can comprehensively capture the core process of proppant migration; selecting downhole visualization imaging equipment with high-temperature and high-pressure resistant specifications that are compatible with the actual downhole environment. This equipment needs to meet the long-term stable operation requirements under high-temperature and high-pressure conditions in the downhole, effectively resist downhole fluid erosion and impurity interference, and ensure the continuity and stability of image acquisition.

[0016] The selected imaging equipment was pre-deployed according to preset installation specifications. Specifically, the wellbore walls at key cluster locations were cleaned and polished to remove mud, rock debris, and protruding impurities, ensuring a smooth and flat installation surface. The imaging equipment was then securely connected to the wellbore wall using high-temperature, high-pressure resistant fasteners. During connection, the installation angle was adjusted so that the lens was directly facing the core area of ​​proppant migration within the fracture. After installation, a sealing test was performed, using downhole-specific sealants to seal the equipment interfaces and prevent downhole fluid from seeping into the equipment and causing damage. The equipment was then started for debugging and calibration, testing the image acquisition clarity, signal transmission stability, and angle appropriateness. Any non-compliant parts were adjusted to ensure secure installation, reasonable acquisition angles, and stable operation. This resulted in a deployed and operational imaging device, providing reliable hardware support for subsequent real-time image acquisition.

[0017] Step 102: Set the image acquisition frequency and acquisition period of the imaging device to synchronize it with the proppant addition stage of the fracturing operation, and obtain the synchronized acquisition parameters. Specifically, this includes: acquiring the proppant addition stage parameters in the fracturing operation plan, determining the start time, end time, and proppant ratio change pattern of the proppant addition stage. This proppant addition stage is the key stage where proppant migration is most active and the migration state changes most significantly, and it is also the core data source for subsequent proppant migration state prediction; setting the image acquisition frequency and acquisition period of the imaging device according to the duration of the proppant addition stage and the frequency of proppant ratio change, wherein the acquisition frequency must match the dynamic change rate of proppant migration to ensure that the subtle dynamic processes of proppant particle suspension, sedimentation, and accumulation can be captured, and the acquisition period corresponds to the start and end time of the proppant addition stage to avoid acquiring invalid image data of irrelevant time periods; verifying the set acquisition frequency and acquisition period to ensure that it is completely synchronized with the proppant addition stage of the fracturing operation, and thus obtaining the synchronized acquisition parameters to provide parameter guidance for subsequent acquisition of proppant migration images.

[0018] Step 103: Based on the synchronized acquisition parameters, continuously acquire image data of proppant migration with fracturing fluid within the fracture using the imaging device in real time to obtain raw image data. Specifically, this includes: starting the deployed downhole visualization imaging device, adjusting the acquisition mode of the device according to the synchronized acquisition parameters to ensure that the device can operate stably at the set acquisition frequency; after the fracturing operation enters the sand-addition stage, continuously acquire continuous image data of proppant migration with fracturing fluid by aligning the imaging device with the fracture interior area in real time, monitoring the device's operating status in real time during the acquisition process, and promptly handling abnormal situations such as signal interruption and image blurring that occur during device operation to ensure that the acquired raw image data can clearly reflect the morphology, distribution, and migration trajectory of proppant particles; thereby obtaining raw image data that can truly reflect the actual downhole migration state of the proppant. This raw image data serves as the basis for subsequent data processing and state prediction, directly determining the reliability of subsequent prediction results.

[0019] Step 104 involves encoding and storing the raw image data in chronological order to form a real-time image sequence containing timestamps and location information. Specifically, this includes: classifying and organizing the acquired raw image data, removing blurry, distorted, or invalid image data acquired during the acquisition process to ensure the validity of subsequent data processing; uniformly encoding the filtered raw image data according to the chronological order of image acquisition, incorporating the location information of the image acquisition during the encoding process. This location information corresponds to the deployment cluster position of the imaging device in step 101. To further improve the accuracy of the location information, a three-dimensional spatial registration algorithm is incorporated here. This algorithm involves calibrating the two-dimensional image coordinates acquired by the imaging device with the actual three-dimensional physical spatial coordinates of the underground fracture, eliminating positional deviations caused by the imaging device's installation angle, lens distortion, and interference from the underground environment, achieving accurate mapping between the two-dimensional image and the three-dimensional physical space, and ensuring that the location information of each frame of the image truly corresponds to the specific spatial location of the underground fracture.

[0020] The specific implementation involves acquiring the downhole three-dimensional physical coordinates of each imaging device deployment cluster. These coordinates are obtained in advance through downhole logging equipment. Simultaneously, the installation angle, lens parameters, and distance from the fracture wall of each imaging device are recorded. These parameters are input into a three-dimensional spatial registration algorithm to calibrate the coordinates of the two-dimensional images acquired by the imaging devices, correcting coordinate offsets caused by lens distortion. Based on the device installation angle and distance from the fracture wall, the coordinates of each pixel in the two-dimensional image are mapped to the acquired downhole three-dimensional physical coordinate system, achieving a one-to-one correspondence between the two-dimensional image and the three-dimensional physical space, thereby completing the image position information... The system ensures accurate registration of information. After registration, the registered three-dimensional spatial location information is integrated into the image encoding process to ensure that each frame of the image corresponds to a clear downhole three-dimensional acquisition location. At the same time, the encoded image data is stored in a standardized manner. During the storage process, a corresponding timestamp is added to each frame of the image. The timestamp corresponds precisely to the real-time time of image acquisition, thereby forming a real-time image sequence containing timestamps, two-dimensional image information, and three-dimensional spatial location information. This image sequence can fully present the temporal change process and spatial distribution characteristics of proppant migration, which is convenient for subsequent temporal tracing, dynamic analysis, and feature extraction of proppant migration status.

[0021] In a preferred embodiment of the present invention, step 200 includes: Step 201 involves performing adaptive median filtering on each frame of the real-time image sequence to remove random noise and salt-and-pepper noise introduced by the complex downhole environment, resulting in a denoised image. Specifically, this includes: acquiring the real-time image sequence; extracting individual frames from the sequence frame by frame to ensure each frame completely covers the proppant migration area and the preset spatial reference markers; performing adaptive median filtering on each extracted frame; dynamically adjusting the filtering window size based on the noise density of different regions of the image during the filtering process; appropriately increasing the filtering window in noisy areas to effectively remove random noise and salt-and-pepper noise, while reducing the filtering window in proppant particle edge areas to avoid loss of particle edge details; monitoring the image processing effect in real-time during the filtering process; performing edge contour detection on the filtered image to confirm that the proppant particle edges are not blurred, broken, or otherwise abnormal, resulting in a thoroughly denoised image with complete particle edge details, providing a high-quality image foundation for subsequent illumination correction.

[0022] Step 202 involves performing homomorphic filtering on the denoised image to correct the uneven brightness distribution caused by downhole light source reflection and fracture wall shadows, resulting in a uniformly illuminated image. Specifically, this includes importing the denoised image into a preset image processing module. This preset image processing module is a dedicated image data processing module built specifically for the processing characteristics of downhole fracturing fracture images. It integrates functions such as image region recognition, homomorphic filtering, parameter adjustment, and effect detection. It can achieve targeted processing of images acquired in complex downhole environments, adapt to the processing needs of uneven brightness and easy loss of details in downhole images, provide a standardized and efficient processing carrier for image illumination correction, and ensure the orderly implementation and processing effect of the illumination correction process.

[0023] The preset process of this image processing module is as follows: Based on the pixel features and brightness distribution characteristics of downhole fracturing fracture images, a basic processing framework is built, dividing the module into functional areas for image import, region analysis, filtering, parameter adjustment, and effect detection, and determining the processing logic and data transmission path for each area. An image region recognition algorithm is integrated into the module to quickly locate and mark core areas with uneven brightness in the image. Simultaneously, a homomorphic filtering algorithm is embedded, preset rules for separating and processing illuminance and reflection components, and reserved interfaces for manual and automatic adjustment of filtering parameters. Considering the actual needs of downhole image illumination correction, the module is designed... The image brightness uniformity detection standard is based on the image grayscale value. The grayscale value fluctuation threshold between different regions of the image is preset to 20 to 30 grayscale levels, and the grayscale value standard deviation threshold within a single region is preset to 15 to 20 grayscale levels. An image detail preservation verification mechanism is added. After the overall construction of the module function is completed, the module is debugged and optimized through multiple sets of denoised images actually acquired downhole. The threshold range can be flexibly adjusted according to different downhole construction conditions and imaging equipment acquisition parameters to ensure that the module functions operate stably and the processing effect meets the requirements of subsequent image processing, thus completing the final construction and deployment of the preset image processing module.

[0024] The core areas of uneven brightness in the image were identified, mainly concentrated in areas covered by fracture wall shadows and areas with strong light source reflection. Homomorphic filtering was performed on the denoised image, separating the illuminance and reflection components. The illuminance component was suppressed, while the reflection component was enhanced to correct the brightness deviation caused by downhole light source reflection and fracture wall shadows. During the processing, the filtering parameters were gradually adjusted to ensure that the overall image brightness was uniform while preserving the inherent grayscale differences between proppant particles and fracturing fluid background, avoiding image detail distortion due to brightness correction. After correction, the image processing module detected the uniformity of image brightness, confirming that the brightness fluctuations in each region of the image were within a preset reasonable range, thus obtaining a uniformly illuminated image with clear details, providing stable image data for subsequent contrast enhancement.

[0025] Step 203 involves performing adaptive histogram equalization on the uniformly illuminated image to stretch the grayscale difference between the proppant particles and the fracturing fluid background, thereby enhancing image contrast and obtaining the proppant image to be segmented. Specifically, this includes: performing global grayscale statistics on the uniformly illuminated image, traversing all pixels of the image and recording the grayscale value of each pixel, statistically obtaining the grayscale distribution ranges of the proppant particle region and the fracturing fluid background region, and locating low-discrimination regions where the grayscale overlap between the particles and the background is high and the differences are not obvious. These regions are key and difficult areas where subsequent segmentation is prone to misjudgment and adhesion.

[0026] Adaptive histogram equalization is performed on uniformly illuminated images. Based on the image size and the distribution of low-resolution areas, the entire image is divided into multiple non-overlapping rectangular sub-regions of the same size. Gray-level histogram statistics are performed on each sub-region separately to calculate the distribution probability of pixel gray levels within the sub-region. Based on this distribution probability, the pixel gray levels within the sub-region are remapped to increase the gray-level range, thereby stretching the gray-level difference between proppant particles and fracturing fluid background within the sub-region.

[0027] During processing, the sub-region size and equalization parameters are adjusted in real time based on the particle display effect within each sub-region. In areas with insufficient grayscale difference, the sub-region size is reduced to enhance local contrast, while in areas with normal grayscale difference, the original parameters are maintained to avoid overly bright, dark, or detail loss caused by global equalization. This ensures that the proppant particle outlines are clear and the fracturing fluid background and particle boundaries are clearly distinguishable. After contrast enhancement, the entire image is subjected to contrast detection, and the grayscale difference between the proppant particles and the background is compared region by region to confirm that the grayscale difference meets the preset qualified standard, with no blurring, over-enhancement, or under-enhancement. Finally, a proppant image to be segmented is obtained with appropriate contrast, clear particle boundaries, and no artifacts, providing stable and reliable image data support for the subsequent convolutional neural network image segmentation stage.

[0028] In a preferred embodiment of the present invention, step 300 includes: Step 301: Input the proppant image to be segmented into a pre-trained convolutional neural network segmentation model. Extract features from the input image using an encoder and decoder structure to obtain multi-scale feature data. Specifically, this includes: acquiring the proppant image to be segmented, and determining the construction, pre-training, and implementation process of the pre-trained convolutional neural network segmentation model. The model construction process is as follows: Based on the proppant image segmentation requirements, determine the overall architecture of the model and build a two-layer structure including an encoder and a decoder. The encoder structure is configured with multiple convolutional layers and pooling layers. The convolutional layers are used for feature extraction, and the pooling layers are used for feature filtering and dimensionality reduction. The decoder structure is configured with multiple deconvolutional layers for feature data recovery and fusion. Pixel-level classification layers are also configured for subsequent pixel category determination. After completing the model architecture construction, initialize the initial parameters of the model, determine the convolutional kernel size, stride, and pooling method of each layer, and construct the initial convolutional neural network segmentation model.

[0029] The pre-training process of the initial convolutional neural network segmentation model involves collecting a large dataset of labeled images under different downhole conditions and proppant migration states. This dataset includes labeled proppant particles, reference markers, and background regions. The dataset is divided into a training subset and a validation subset according to a preset ratio. The initial convolutional neural network segmentation model is trained using the training subset. The labeled images in the training subset are input into the model, and the encoder extracts image features layer by layer. After being restored and fused by the decoder, the classification layer outputs the pixel category prediction results. The deviation between the prediction results and the labeled information is calculated, and the network weight parameters are updated. The training is iterated until the model performance stabilizes. The model is then validated using the validation subset, and the model parameters are adjusted to optimize the model's segmentation accuracy. Finally, the pre-trained convolutional neural network segmentation model is obtained.

[0030] The implementation process of this convolutional neural network segmentation model is as follows: A pre-trained model is loaded into the image processing module using a model deployment tool, ensuring the model can stably receive input images and efficiently perform operations such as feature extraction and pixel classification. The proppant image to be segmented is normalized according to the input requirements of the pre-trained convolutional neural network segmentation model, ensuring that the size of the input image matches the model's input size. The normalized proppant image is then input into the pre-trained convolutional neural network segmentation model, which has a built-in encoder and decoder two-layer structure. The encoder structure includes multiple convolutional and pooling layers, extracting features from the input image through layer-by-layer convolution operations. This extracts shallow texture features, retains key features and reduces the amount of feature data through pooling layers, and further extracts deep semantic features through deep convolution operations. The decoder structure correspondingly includes multiple deconvolutional layers, performing layer-by-layer deconvolution operations on the feature data extracted by the encoder to achieve gradual recovery and fusion of feature data. Finally, shallow and deep features are integrated to obtain multi-scale feature data covering different scales, providing reliable feature support for subsequent pixel-level classification and target recognition.

[0031] Step 302: Based on the multi-scale feature data, after upsampling and pixel-level classification by the decoder, the category probability data of each pixel corresponding to proppant particles, reference markers, and background is obtained. Specifically, this includes: inputting the multi-scale feature data into the deconvolution layer of the decoder, performing step-by-step upsampling on the feature map through deconvolution operations, and restoring the size of the feature map at each upsampling until the size of the feature map is completely consistent with the size of the input proppant image to be segmented; after the size restoration of the feature map is completed, the upsampled feature map is input into the pixel-level classification layer of the convolutional neural network segmentation model. The classification layer analyzes each pixel in the image one by one, and calculates the category probability data of each pixel belonging to the three types of targets—proppant particles, reference markers, and background—in combination with the texture and semantic information contained in the multi-scale feature data, ensuring that each pixel corresponds to three sets of category probability data, and that each set of probability data can accurately represent the possibility that the pixel belongs to the corresponding category, providing a basis for subsequent pixel category determination.

[0032] Step 303: Classify each pixel according to the category probability data to obtain pixel-level classification labels. This specifically includes: sequentially traversing each pixel in the proppant image to be segmented, reading the category probability values ​​of the three target classes (proppant particles, reference marker, and background) corresponding to that pixel, and obtaining the complete probability results of that pixel under the three target classes; comparing the three sets of probability values ​​of that pixel item by item, sequentially comparing the probability of proppant particles with the probability of reference marker, the probability of proppant particles with the probability of background, and the probability of reference marker with the probability of background, and determining the largest probability item among the three sets of values ​​through step-by-step comparison; and assigning the largest probability item to the... The target type is determined as the final category of the pixel, completing the category determination for a single pixel. Following the same determination logic and calculation method, all pixels in the image are traversed and their categories are determined sequentially, without missing any pixel. After all pixels are determined, they are arranged in order according to their original positions in the image, and each pixel is labeled with its corresponding category identifier, ultimately forming a pixel-level classification label covering the entire image. This classification label can clearly distinguish the target category corresponding to each pixel, accurately dividing the pixel regions corresponding to proppant particles, reference markers, and background, providing an accurate basis for subsequent image segmentation.

[0033] Step 304: Based on the pixel-level classification labels, obtain a segmented image containing the proppant particle region and the position information of multiple spatial reference labels. Specifically, this includes: classifying and filtering the pixels of the entire image according to the pixel-level classification labels, filtering out all pixels labeled as proppant particles, integrating these pixels to form a complete proppant particle region; simultaneously filtering out all pixels labeled as reference labels, determining the pixel set corresponding to each reference label, and introducing a minimum bounding rectangle to further determine the position information of the reference labels and avoid positional deviations caused by pixel dispersion. The minimum bounding rectangle is the rectangle that can completely enclose the target pixel set and has the smallest area. This rectangle can quickly and accurately define the boundary range of the target region, providing a positional reference for subsequent spatial coordinate mapping.

[0034] The specific implementation method is as follows: For each pixel set corresponding to a reference marker, the x-coordinate and y-coordinate of all pixels in the set are extracted one by one. The maximum and minimum values ​​of the x-coordinate and y-coordinate are selected. The minimum x-coordinate is used as the left boundary of the rectangle, the maximum x-coordinate as the right boundary of the rectangle, the minimum y-coordinate as the lower boundary of the rectangle, and the maximum y-coordinate as the upper boundary of the rectangle. In this way, the minimum bounding rectangle of each reference marker pixel set is constructed. The boundary range and center position of each reference marker are determined by the coordinates of the four vertices of the minimum bounding rectangle, and thus the position information of each reference marker is determined. All pixels that are marked as background are removed to eliminate the interference of the background area. The integrated proppant particle area is associated and integrated with the minimum bounding rectangle of all reference markers and the corresponding position information. Finally, a segmented image containing the proppant particle area and the position information of multiple spatial reference markers is obtained, which provides accurate and reliable image data for subsequent spatial coordinate mapping and distortion correction.

[0035] Step 305: Extract the pixel coordinates of each spatial reference marker from the segmented image. Based on the preset actual physical coordinates of the reference markers, establish a mapping relationship between the pixel coordinate system and the fracture physical space coordinate system. Specifically, this includes: locating the pixel region corresponding to each spatial reference marker in the segmented image; selecting the center pixel of the minimum bounding rectangle of each reference marker, and extracting the x and y coordinates of this center pixel as the pixel coordinates of each spatial reference marker; simultaneously retrieving the preset actual physical coordinates of each spatial reference marker. The preset actual physical coordinates of the reference markers refer to the true three-dimensional coordinates of each spatial reference marker within the downhole fracture physical space. These coordinates characterize the actual position of the reference marker within the fracture, thus establishing a mapping relationship between the pixel coordinate system and the fracture physical space coordinate system. The mapping of the physical space coordinate system provides a true reference. The preset process is as follows: After the imaging equipment is deployed and before the formal acquisition of proppant migration images, the key cluster positions of the fractured well section are measured by the downhole logging equipment. The logging equipment must meet the requirements of downhole high temperature and high pressure conditions and be able to acquire three-dimensional coordinate data in the physical space of the fracture. During the measurement, the actual position of each preset spatial reference mark is located one by one, and the three-dimensional coordinates of each reference mark in the physical space of the fracture are recorded, namely the horizontal coordinate, vertical coordinate and depth coordinate. After the measurement is completed, the three-dimensional coordinate data of all reference marks are checked and verified, and data with excessive measurement deviation and distortion are eliminated. Missing data is supplemented and improved. The three-dimensional coordinates of each qualified reference mark are stored, and the preset of the actual physical coordinates of the reference mark is completed.

[0036] The actual physical coordinates are the three-dimensional coordinates of each reference mark in the physical space of the crack. The pixel coordinates of each spatial reference mark are associated with the corresponding preset actual physical coordinates. By analyzing multiple sets of corresponding coordinates, the mapping relationship between the pixel coordinate system and the crack physical space coordinate system is established, the coordinate correspondence rules between the two coordinate systems are determined, and the preliminary association between the image coordinates and the real physical coordinates is realized, laying the foundation for subsequent image distortion correction and coordinate transformation.

[0037] Step 306: Based on the mapping relationship, a regular correction grid covering the entire field of view is constructed within the fracture physical space, and the regular correction grid is back-projected onto the image pixel plane to obtain the corresponding distortion correction grid. Specifically, this includes: based on the mapping relationship between the pixel coordinate system and the fracture physical space coordinate system, the actual physical range of the imaging field of view is determined through imaging device parameters; a spatial range completely corresponding to the imaging field of view is delineated within the fracture physical space to ensure that the spatial range can completely cover all areas that the downhole imaging device can capture, including the core area of ​​proppant migration and the location of all spatial reference markers; a regular correction grid is constructed within the delineated spatial range. The regular correction grid adopts a uniformly distributed rectangular grid structure, with grid nodes arranged at equal intervals to uniformly cover the entire field of view. The size and density of the grid are set according to the actual size of the imaging field of view and the distribution density of the reference markers. The grid density is appropriately increased in areas with densely distributed reference markers and appropriately decreased in areas with sparsely distributed reference markers to ensure that the grid can accurately cover the area surrounding each reference marker while avoiding excessively high grid density that would increase the computational load.

[0038] After constructing the regular correction mesh, each mesh node in the regular correction mesh is back-projected from the crack physical space coordinate system to the image pixel plane according to the mapping relationship between the pixel coordinate system and the crack physical space coordinate system. The corresponding horizontal and vertical coordinates of each mesh node in the pixel coordinate system are calculated through the mapping relationship to determine the specific position of each mesh node in the pixel plane. Based on the corresponding coordinates of all mesh nodes, the nodes are connected in the image pixel plane according to the arrangement order of the mesh nodes to form the distortion correction mesh corresponding to the regular correction mesh. This distortion correction mesh can reflect the distortion distribution in different areas of the image, especially the local distortion caused by lens distortion and installation angle, providing a clear and reliable mesh basis for the subsequent calculation of local distortion coefficients.

[0039] Step 307: Divide the distortion correction mesh into multiple sub-regions. Within each sub-region, calculate the local distortion coefficient based on the mapping deviation between the pixel coordinates of the reference marker and the actual physical coordinates. Specifically, this includes: dividing the distortion correction mesh into sub-regions according to a division rule. The division uses a uniform segmentation method, dividing the entire distortion correction mesh into multiple independent sub-regions that continuously cover the entire mesh area. The shape of each sub-region is consistent, all being rectangles, and the size of each sub-region is uniform to ensure the consistency of subsequent distortion coefficient calculations. During the division process, it is necessary to ensure that each sub-region contains at least one spatial reference marker to avoid situations where distortion cannot be quantified due to the absence of a reference marker in a sub-region. If there is a situation where a sub-region lacks a reference marker, the sub-region division range needs to be adjusted until each sub-region contains at least one reference marker.

[0040] After the region is divided, for each sub-region, the positions of all spatial reference markers within that sub-region are located, and the pixel coordinates corresponding to each reference marker are extracted. These pixel coordinates are the coordinates of the center pixel of the smallest bounding rectangle of the reference marker determined in step 305. At the same time, the actual physical coordinates corresponding to each reference marker preset in step 305 are retrieved. The deviation values ​​of the pixel coordinates of each reference marker and the corresponding actual physical coordinates in the horizontal and vertical directions are calculated one by one, that is, the difference between the pixel horizontal coordinate and the physical horizontal coordinate, and the difference between the pixel vertical coordinate and the physical vertical coordinate are calculated respectively. The deviation values ​​of all reference markers in the horizontal and vertical directions in each sub-region are statistically analyzed, and abnormal data with abnormally large deviation values ​​are removed to avoid abnormal data affecting the accuracy of the distortion coefficients. The average value of all remaining deviation values ​​is taken as the local distortion coefficient of the sub-region. This completes the calculation of the local distortion coefficient of each sub-region, ensuring that the local distortion coefficient can characterize the degree of image distortion of the corresponding sub-region, and providing a reliable basis for the subsequent integration of the overall correction coefficients.

[0041] Step 308: Combine the local distortion coefficients of all sub-regions into correction coefficients. This includes: after calculating the local distortion coefficients of all sub-regions, numbering all sub-regions in an orderly manner, with the numbering order consistent with the arrangement order of the sub-regions in the distortion correction grid, i.e., numbering each sub-region from left to right and from top to bottom; organizing the local distortion coefficients of all sub-regions in an orderly manner according to this numbering order, associating each sub-region's number, coordinate range in the distortion correction grid, and local distortion coefficient with that sub-region to determine the spatial range corresponding to each local distortion coefficient; integrating all associated local distortion coefficients to form a complete coefficient set, which is the correction coefficient covering the entire imaging area. This correction coefficient contains the distortion data of each sub-region, which can comprehensively and accurately reflect the distortion situation of different areas of the image, especially the local distortion differences caused by lens distortion and mounting angle, providing a unified and reliable basis for the subsequent geometric correction of dynamic movement feature parameters, and ensuring that subsequent coordinate transformation can specifically correct the distortion deviations of different areas.

[0042] Step 309 involves identifying and labeling each proppant particle from the segmented image to obtain particle identification label data containing the centroid coordinates of the particles. Specifically, this includes: identifying each proppant particle region in the segmented image based on the segmented image; distinguishing between independent proppant particles and adhered particles by analyzing the pixel grayscale features, contour features, and pixel distribution density of the proppant particle regions; for adhered particles, separating them using a region growing method by analyzing the grayscale differences and contour gaps between particles to ensure that each proppant particle can be identified individually, avoiding errors in subsequent parameter extraction due to particle adhesion; after identification, each proppant particle is individually labeled, and each particle is assigned a unique label. The labels are numbered sequentially to facilitate matching of particles in adjacent frames.

[0043] By calculating the geometric center of the pixel region corresponding to each proppant particle, the centroid position of each particle is determined. During the calculation, the x and y coordinates of all pixels in the pixel region of each proppant particle are extracted. The average of all x coordinates is taken as the x coordinate of the centroid, and the average of all y coordinates is taken as the y coordinate of the centroid. The x and y coordinates of the centroid position are then extracted to form the centroid coordinates of each proppant particle. The labels of all proppant particles are associated and integrated with the corresponding centroid coordinates, and the data is arranged into an ordered set according to the label order. Finally, particle identification label data containing particle centroid coordinates is obtained, which provides target point information for the subsequent calculation of dynamic movement feature parameters.

[0044] Step 310: Based on the particle identification marker data, extract the centroid coordinates of the same proppant particles between adjacent frames, calculate the displacement vector between the two frames and decompose it into horizontal and vertical velocity components. Simultaneously, combine the image acquisition time interval to obtain the instantaneous transport velocity and its velocity components. Based on the particle identification marker data, count the number of particles in each region and calculate the proppant concentration distribution. Specifically, this includes: based on the particle identification marker data, extracting the particle identification marker data corresponding to two adjacent segmented images from the real-time image sequence to determine the order and time nodes of the two images; by comparing the markers of proppant particles in the two images, matching the same proppant particles between adjacent frames, i.e., finding particles with the same markers in the two images, ensuring... To avoid matching errors, the matched particles are the same proppant particle. The centroid coordinates of the same particle are extracted in two adjacent frames. The x-coordinate and y-coordinate of the centroid of the particle in the previous frame are recorded in the next frame. The coordinate differences of the centroid coordinates of the same particle in the horizontal and vertical directions are calculated between the two frames. That is, the horizontal coordinate difference is obtained by subtracting the horizontal coordinate of the centroid of the previous frame from the horizontal coordinate of the centroid of the next frame, and the vertical coordinate difference is obtained by subtracting the vertical coordinate of the centroid of the previous frame from the horizontal coordinate of the centroid of the next frame. The displacement vector of the particle between the two frames is determined based on the coordinate differences in the horizontal and vertical directions. The displacement vector is decomposed into horizontal and vertical displacement components.

[0045] Simultaneously, the image acquisition time interval set in step 102 is retrieved. This time interval is the acquisition time difference between two adjacent frames. The horizontal displacement component is divided by the acquisition time interval to calculate the velocity component of the particle in the horizontal direction. The vertical displacement component is divided by the acquisition time interval to calculate the velocity component of the particle in the vertical direction. The instantaneous migration velocity of the particle is obtained by combining the velocity components in the two directions. At the same time, according to the particle identification marker data, the segmented image is divided into multiple statistical regions of the same size according to the region division rules. The number of proppant particles in each statistical region is counted. Combined with the actual area of ​​each statistical region, the number density of proppant particles in each region is calculated by dividing the number of particles in each region by the area of ​​the region. This yields the proppant concentration distribution data, comprehensively characterizing the spatial distribution of proppant within the crack.

[0046] Step 311: Identify the sand embankment outline from the segmented image, measure the vertical distance between the highest point of the sand embankment and the bottom of the fracture to obtain the sand embankment height, and track the change of the leading edge position of the sand embankment over time to calculate the leading edge advancement distance. Specifically, this includes: identifying the pixel region corresponding to the sand embankment from the segmented image based on the pixel grayscale features and outline features of the sand embankment region. The pixel grayscale of the sand embankment region is significantly different from the fracturing fluid background and fracture wall. By combining grayscale thresholding with outline extraction, the complete outline of the sand embankment is delineated to ensure that the outline is unbroken and complete; determine the boundary pixel coordinates of the sand embankment outline, and record the sand embankment outline one by one. The x and y coordinates of each boundary pixel on the embankment outline are calculated. The pixel corresponding to the highest point in the sand embankment outline is located. By comparing the y coordinates of all boundary pixels of the sand embankment outline, the pixel with the largest y coordinate is found. This pixel is the highest point of the sand embankment, and its y coordinate is extracted. At the same time, the pixel corresponding to the bottom of the crack is located. The boundary line of the bottom of the crack is found according to the contour features of the crack wall. The y coordinates of the pixels on the bottom boundary line of the crack are extracted. The average value of the y coordinates of the boundary line pixels is taken as the y coordinate of the bottom of the crack. The difference between the y coordinate of the highest point of the sand embankment and the y coordinate of the bottom of the crack is calculated. This difference is the height of the sand embankment.

[0047] In the segmented images of consecutive frames in a real-time image sequence, the boundary pixel coordinates corresponding to the leading edge of the sand embankment are continuously tracked. The horizontal and vertical coordinates of the boundary pixels of the leading edge of the sand embankment are recorded in each frame to determine the position information of the leading edge of the sand embankment at different times. The coordinate difference of the leading edge of the sand embankment in the horizontal direction at adjacent times is calculated, that is, the horizontal coordinate of the leading edge of the sand embankment at the next time is subtracted from the horizontal coordinate of the leading edge of the sand embankment at the previous time. Combined with the time interval between adjacent times, the horizontal coordinate difference is divided by the time interval to calculate the advancing distance of the leading edge of the sand embankment. This comprehensively characterizes the accumulation morphology and migration trend of the proppant and reduces the extraction error of sand embankment related parameters.

[0048] Step 312 involves using the centroid coordinates of the proppant particles, instantaneous migration velocity and its velocity components, proppant concentration distribution, sand embankment height, and leading edge advance distance as dynamic migration feature parameters. Specifically, after steps 310 and 311 complete the calculation of the relevant parameters, the calculated parameters are categorized and organized. The parameters obtained in step 310 are divided into three categories: proppant particle centroid coordinates, instantaneous migration velocity and its horizontal and vertical velocity components, and proppant concentration distribution. The parameters obtained in step 311 are divided into two categories: sand embankment height and sand embankment leading edge advance distance. All parameters are standardized in format, and the numerical units and representation methods are unified to ensure consistency in the representation methods of various parameters and avoid deviations in subsequent data fusion due to inconsistent formats. These parameters are collectively used as dynamic migration feature parameters that comprehensively reflect the proppant migration state. They are organized in an orderly manner according to parameter type to form a complete parameter set, providing a complete feature parameter foundation for subsequent parameter correction and multi-source data fusion, ensuring that subsequent model training can acquire proppant migration feature information.

[0049] Step 313: Based on the correction coefficients, construct the coordinate transformation relationship from the image coordinate system to the crack physical space coordinate system. Specifically, this includes: extracting the local distortion coefficients of each sub-region based on the correction coefficients; determining the correspondence between the local distortion coefficients of each sub-region and the coordinate range of that sub-region within the crack physical space, ensuring that the distortion coefficients of each sub-region accurately match their corresponding spatial range; combining the mapping relationship between the pixel coordinate system and the crack physical space coordinate system, integrating the distortion correction rules corresponding to the correction coefficients, and determining the deviation correction method corresponding to different distortion coefficients; and constructing the coordinate transformation relationship from the image coordinate system to the crack physical space coordinate system. The system determines the transformation rules between the horizontal and vertical coordinates in the image coordinate system and the three-dimensional coordinates in the physical space coordinate system of the crack. The transformation rules need to be adjusted in conjunction with the installation angle of the imaging equipment and the lens parameters. At the same time, the correction ratio of coordinate transformation in different regions is determined by combining the local distortion coefficient of each sub-region. The correction ratio is appropriately increased in regions with larger distortion coefficients and appropriately decreased in regions with smaller distortion coefficients. This ensures that the coordinate transformation relationship can effectively correct the deviation caused by image distortion, provide a reliable transformation basis for the spatial correction of dynamic movement characteristic parameters, and ensure that the transformed parameters can truly reflect the state of the proppant in the physical space of the crack.

[0050] Step 314: Based on the coordinate transformation relationship, perform a region-by-region coordinate transformation on the dynamic transport feature parameters, converting the parameter values ​​in the image coordinate system to actual values ​​in the crack physical space coordinate system, to obtain the physical space dynamic feature parameters. Specifically, this includes: based on the coordinate transformation relationship, dividing the dynamic transport feature parameters into sub-regions according to the distortion correction grid; associating the dynamic transport feature parameters corresponding to each sub-region with the local distortion coefficient of that sub-region to determine the distortion correction standard corresponding to each parameter; and performing a region-by-region coordinate transformation on the dynamic transport feature parameters within each sub-region according to the coordinate transformation relationship and the local distortion coefficient, converting the centroid coordinates of the proppant particles, and converting the horizontal and vertical coordinates of the centroid in the image coordinate system to the crack physical space coordinate system. The system employs three-dimensional coordinates in a spatial coordinate system. Instantaneous migration velocity and its components are transformed, and the velocity values ​​in the image coordinate system are converted to actual velocity values ​​within the physical space of the fracture, based on the proportional relationship of the coordinate transformation. Prop concentration distribution, sandbank height, and leading-edge advance distance are sequentially transformed, converting the parameter values ​​in the image coordinate system to actual values ​​in the physical space coordinate system of the fracture. After completing the transformation of all sub-region parameters, the transformed parameters are integrated, organized according to parameter type, and abnormal data generated during the transformation process are removed to ensure the completeness and accuracy of parameter information. This yields dynamic characteristic parameters of the physical space, effectively eliminating parameter deviations caused by lens distortion and installation angle, and ensuring that the parameters truly reflect the migration state of the proppant within the physical space of the fracture.

[0051] Step 315: Perform polar coordinate transformation on the physical space dynamic feature parameters, converting the proppant position, velocity components, and distribution parameters in rectangular coordinate form to polar coordinate form to obtain feature parameters to be fused containing polar coordinate information. Specifically, this includes: classifying and organizing the physical space dynamic feature parameters, filtering out parameters represented in rectangular coordinate form, including: the centroid position of proppant particles, the horizontal and vertical components of instantaneous migration velocity, and coordinate parameters related to proppant concentration distribution, excluding parameters that do not require polar coordinate transformation; performing polar coordinate transformation on these rectangular coordinate parameters, and based on the correspondence between rectangular and polar coordinates, converting each… The rectangular coordinates of each parameter are converted to their corresponding polar coordinates. During the conversion, the representation of each parameter in the polar coordinate system is determined, that is, the spatial position and orientation of the parameter are represented by the polar radius and polar angle. After the conversion, all parameters are integrated, and the numerical units and formats of the parameters are unified to ensure that the parameter format is consistent and the information is complete, and data with excessive deviations during the conversion process are eliminated. Finally, the feature parameters to be fused containing polar coordinate information are obtained. These parameters can better adapt to the needs of subsequent multi-source data fusion and predictive neural network model training, provide the model with comprehensive and standardized feature input, and ensure that the model can make full use of the spatial feature information of proppant transport.

[0052] In a preferred embodiment of the present invention, step 400 includes: Step 401: Obtain the feature parameters to be fused, and acquire real-time monitoring data of fracturing fluid viscosity, proppant particle size, flow rate, sand ratio, and fracture width during the fracturing operation. This data serves as multi-source parameter data. Specifically, this includes: retrieving the feature parameters to be fused containing polar coordinate information; initially organizing these parameters to ensure consistent format and complete information, avoiding missing or inconsistent parameters; simultaneously, collecting various key engineering parameters during the fracturing operation using downhole real-time monitoring equipment, including real-time monitoring data of fracturing fluid viscosity, proppant particle size, flow rate, sand ratio, and fracture width. During data collection, the validity of the data is verified in real-time, eliminating abnormal fluctuations and distorted data, and supplementing missing data. These verified and supplemented real-time monitoring data are then uniformly categorized as multi-source parameter data, achieving the initial acquisition and organization of the feature parameters to be fused and multi-source engineering parameters, laying the foundation for subsequent multi-source data fusion.

[0053] Step 402: Obtain a pre-constructed proppant migration simulation dataset based on the computational fluid dynamics and discrete element method (CFD-DEM). The simulation dataset contains data on the proppant migration trajectory, sedimentation distribution, and concentration field within the fracture under different combinations of construction parameters. Specifically, the pre-constructed proppant migration simulation dataset is a numerical simulation dataset of proppant migration covering different combinations of fracturing construction parameters, built based on the computational fluid dynamics and discrete element method (CFD-DEM). Its core includes data on the proppant migration trajectory, sedimentation distribution, and concentration field within the fracture under different operating conditions. It can reflect the proppant migration law under the influence of different construction parameters and can effectively make up for the shortcomings of the actual downhole monitoring data being limited by the construction scenario and the limited coverage. It provides comprehensive simulation data support for the subsequent multi-source data fusion to construct a feature dataset.

[0054] The pre-construction process of this simulation dataset involves: identifying key parameters affecting proppant migration during fracturing operations; determining multi-dimensional combinations of construction parameters covering different fracturing fluid viscosities, proppant particle sizes, fracturing flow rates, sand ratios, and fracture widths to ensure that these parameter combinations cover various operating conditions that may occur during on-site fracturing operations; employing a computational fluid dynamics and discrete element method coupled approach to numerically simulate the entire proppant migration process within the fracture corresponding to each set of construction parameter combinations; recreating the physical environment of the downhole fracture and the actual fracturing operation process during the simulation, accurately simulating the entire process of proppant suspension, movement, sedimentation, and accumulation within the fracture along with the fracturing fluid; after the simulation is completed, data is extracted from the simulation results corresponding to each set of parameter combinations, focusing on recording proppant migration trajectories, sedimentation distributions, and concentration field data; all extracted simulation data undergoes comprehensive screening and organization, eliminating invalid data with excessive simulation deviations or inconsistent with actual fracturing operation scenarios, and standardizing the data format to ultimately form a comprehensive and accurate proppant migration simulation dataset covering multiple operating conditions, which is then stored.

[0055] The pre-built proppant migration simulation dataset is directly retrieved from the storage medium. The proppant migration trajectory, sedimentation distribution, and concentration field data under different combinations of construction parameters are specifically retrieved from the dataset. At the same time, the range of parameters of the retrieved simulation data is checked to ensure that it can fully cover the parameter combinations that may occur during the current fracturing operation. This provides simulation data support that fits the actual construction for the construction of subsequent feature datasets, effectively making up for the limited coverage of actual downhole monitoring data scenarios.

[0056] Step 403 involves vectorizing and splicing the feature parameters to be fused and the multivariate parameter data to obtain a fused feature vector. Specifically, this includes: vectorizing the two types of data based on the feature parameters to be fused and the multivariate parameter data. Various feature indicators in the feature parameters to be fused, including proppant position, velocity components, and distribution parameters in polar coordinates, are converted into vector forms one by one. Simultaneously, parameters in the multivariate parameter data, such as fracturing fluid viscosity, proppant particle size, construction displacement, sand ratio, and fracture width, are also converted into corresponding vector forms, ensuring that both types of data are presented in a unified vector format. The preset fusion rules are standardized rules pre-defined to achieve orderly and effective splicing of the feature parameter vectors to be fused and the multivariate parameter data vectors. The core rules include vector dimension matching rules, vector splicing order rules, and data format uniformity verification rules. The formulation of these rules can solve the problem that multi-source data cannot be effectively fused due to differences in format and dimension, leading to predictions relying only on a single image feature. This ensures the organic integration of image features and engineering parameters, improving the correlation and completeness of feature data.

[0057] The pre-set process of this fusion rule is as follows: Combining the actual working conditions of fracturing operations with the needs of proppant migration characteristic analysis, the importance ranking of the feature parameters to be fused and the multivariate parameter data in subsequent model training is determined, and the basic order of vector splicing is determined. The feature parameter vector reflecting the real-time movement of the proppant is used as the core feature segment, and the multivariate engineering parameter vector affecting proppant migration is used as the auxiliary feature segment. The splicing connection position of the two vector segments is defined. Based on the vector dimension characteristics of the two types of data after vectorization, dimension matching rules are formulated. If the dimensions of the two types of vectors are inconsistent, the lower-dimensional vector is standardized and expanded without changing the original characteristics of the data through feature dimension completion, ensuring that the dimensions of the two types of vectors are completely matched before splicing. Data format uniformity verification rules are formulated, determining the uniform standards for the numerical units, data precision, and data types of vector data. Format verification of the two types of vectors is required before splicing, and data that does not meet the standards is standardized and converted.

[0058] After vector transformation, according to the preset fusion rules, the vectors corresponding to the feature parameters to be fused and the vectors corresponding to the multivariate parameter data are double-checked in terms of format and dimension. Vectors that do not meet the rule requirements are standardized. The two types of vectors are then ordered and concatenated according to feature importance. During the concatenation process, the defined connection positions are followed to ensure that the dimensions of the vectors match and the order is consistent, avoiding vector misalignment and concatenation chaos. After concatenation, a fused feature vector with a unified structure and complete information is obtained, realizing the organic integration of image features and engineering parameters and improving the correlation and completeness of feature data.

[0059] Step 404: Extract simulated feature data matching the current construction stage from the simulated dataset, and align the simulated feature data with the fused feature vector to construct a feature dataset containing image features, engineering parameters, and simulated data. Specifically, this includes: determining the current construction stage of the fracturing operation and identifying key parameters such as fracturing fluid viscosity, proppant particle size, and construction displacement corresponding to the current construction stage; based on these key parameters, selecting simulated feature data matching the parameter combinations of the current construction stage from the simulated dataset to ensure that the selected simulated feature data fits the current actual construction scenario and improves the relevance of the simulated data; after selection, standardizing the format of the simulated feature data to be consistent with the format of the fused feature vector obtained in step 403, aligning the simulated feature data with the fused feature vector, and associating them one-to-one according to parameter type and data dimension to ensure that each set of fused feature vectors corresponds to the matched simulated feature data; integrating the associated fused feature vectors with the simulated feature data to construct a feature dataset containing image features, engineering parameters, and simulated data, realizing the collaborative utilization of multi-source data and improving the coverage of feature data.

[0060] Step 405: Obtain the actual proppant placement morphology data corresponding to each sample in the feature dataset as the label value. Use the fused feature vector in the feature dataset as the input feature, which, together with the label value, constitutes the training sample set and the validation sample set. Specifically, this includes: collecting the actual proppant placement morphology data corresponding to each sample in the feature dataset using downhole testing equipment. During the collection process, ensure that the measured data corresponds one-to-one with the construction parameters and time nodes of the feature dataset samples to avoid data misalignment. Use the collected actual proppant placement morphology data as the label value for subsequent training and validation of the neural network model. Use the fused feature vector in the feature dataset as the input feature, and associate each input feature with the corresponding label value to form complete sample data.

[0061] The preset division ratio is a pre-defined rule for dividing the complete sample data into training and validation sample sets. It determines the proportion of the training and validation sample sets in the total sample data. This preset ratio can solve the problem that the model training has poor generalization ability and low prediction accuracy of proppant migration state in complex cracks due to unreasonable sample division. It ensures that the training sample set is sufficient to support the neural network to learn the proppant migration law, while the validation sample set can effectively verify the prediction performance of the model.

[0062] The process of setting the partition ratio involves combining the model structure and training requirements of the proppant migration prediction neural network, as well as the total number and distribution characteristics of the feature dataset. It comprehensively considers the required scale of the training dataset and the required effectiveness of the validation dataset. Scale requires a sufficient number of samples in the training dataset to allow the model to fully learn the intrinsic relationship between proppant migration and placement patterns under different construction scenarios and parameter combinations. Effectiveness requires the number of samples in the validation dataset to reflect the model's prediction bias under different working conditions. The process involves analyzing the number of samples for different fracturing construction parameter combinations and different proppant migration states to ensure that the preset partition ratio evenly covers various construction scenarios and parameter combinations in both the training and validation datasets, avoiding excessive concentration or absence of samples for a particular working condition in a single dataset. Based on the above analysis, the partition ratio range suitable for training this model is determined, thus completing the preset process.

[0063] According to the preset division ratio, all sample data are divided into training sample set and validation sample set. During the division process, samples are extracted according to the ratio. At the same time, the distribution of the two types of sample sets is verified to ensure that the distribution of the two types of sample sets is uniform and covers samples with different construction scenarios and different parameter combinations. This provides standardized and reliable sample support for the initial training and validation of the neural network, and ensures that the training process can be carried out in an orderly and efficient manner.

[0064] Step 406: Construct an initial neural network. Train the initial neural network using the training sample set. Input the fused feature vectors from the training sample set into the initial neural network for forward propagation calculation to obtain the layup morphology prediction data. Specifically, this includes: Constructing an initial neural network. To address the problems of existing prediction models relying solely on single image features, having poor generalization ability, and experiencing prediction lag due to complex processing procedures, thus failing to meet the real-time control requirements of fracturing construction, and considering the actual engineering needs of proppant migration prediction, the initial neural network is constructed from three core dimensions: network structure selection, layer and node design, and functional definition of each layer. The specific construction process is as follows: Network structure selection. Select a deep neural network structure that is suitable for multi-source fusion feature extraction and nonlinear law fitting to ensure that the network can uncover the inherent correlation between image features, engineering parameters, and simulation data in the fused feature vectors, while also considering network inference efficiency and avoiding subsequent prediction lag due to overly complex network structure, thus meeting the needs of real-time prediction in fracturing construction.

[0065] The number of network layers and nodes in each layer are determined. Based on the dimension of the fused feature vector and the feature complexity of proppant placement morphology prediction, the number of input, hidden, and output layers is rationally designed. The hidden layers employ a multi-level structure to achieve deep feature mining and transformation. The number of nodes in each layer is matched to the feature processing requirements of each layer. The number of nodes in the input layer is consistent with the dimension of the fused feature vector to ensure complete reception of all information from the fused feature vector. The number of nodes in the hidden layer is gradually adjusted according to the depth requirements of feature mining, achieving a progressive transformation from shallow features to deep semantic features. The number of nodes in the output layer matches the dimension of the proppant placement morphology prediction data to ensure the output of prediction results related to placement morphology. The core functions of the input, hidden, and output layers are determined. The input layer is used to fully receive the fused feature vector, achieving efficient input and transmission of multi-source fused feature data. The hidden layer is used to perform layer-by-layer feature mining, transformation, and integration of the input fused feature vector, uncovering the coupling relationships between different features and extracting the core rules relating proppant migration and placement morphology. The output layer is used to perform final mapping and calculation on the feature data processed by the hidden layer, outputting the proppant placement morphology prediction data.

[0066] The initial neural network is constructed using the above method, ensuring that it can fully learn the intrinsic relationships between multi-source features while maintaining inference efficiency, thus meeting the real-time prediction requirements of downhole fracturing operations. After the initial neural network is constructed, it is trained using a training sample set. The fused feature vectors from the training sample set are input into the initial neural network one by one, passing through the input layer to the hidden layer. The hidden layer performs feature mining, transformation, and integration layer by layer on the input fused feature vectors, and then passes them to the output layer. The output layer performs forward propagation calculations to obtain the proppant placement pattern prediction data corresponding to each input fused feature vector, completing the initial training of the initial neural network and allowing the network to initially learn the intrinsic relationship between fused features and proppant placement pattern.

[0067] Step 407 involves calculating the loss function between the predicted proppant morphology data and the corresponding label values, and updating the network weight parameters. Specifically, this includes: comparing the predicted proppant morphology data for each training sample with the corresponding label value, calculating the deviation based on the comparison results, and then calculating the loss function based on this deviation. The loss function is a core function that quantifies the degree of deviation between the predicted proppant morphology data and the actual proppant proppant morphology data. Its role is to reflect the fit between the initial neural network's prediction results and the actual working conditions, compensating for the lack of effective quantitative evaluation standards in model training and the difficulty in accurately correcting prediction deviations. The value generated by this function is used to intuitively reflect the magnitude of the prediction deviation; a larger value indicates a higher degree of deviation between the predicted data and the actual data, and a worse prediction effect of the neural network; a smaller value indicates a higher degree of fit between the predicted data and the actual data, and a better prediction effect of the neural network. Based on the value generated by the loss function, the current training fit effect of the initial neural network can be determined, providing a quantitative basis for subsequent adjustments and updates to the network weight parameters, making parameter updates more targeted, and avoiding problems such as low model training efficiency and poor prediction accuracy caused by blind adjustments.

[0068] After the loss function is calculated, the network weight parameters of the initial neural network are adjusted and updated according to the calculation results of the loss function using a preset parameter update rule. The preset parameter update rule is a standardized rule pre-defined to achieve accurate and orderly adjustment of the neural network weight parameters. The core of the rule includes the matching rule between the deviation magnitude and the update amplitude, the adjustment priority rule of each layer weight parameter, and the step size constraint rule of parameter update. The preset rule can solve the problems of blind adjustment of model parameters and low efficiency of fitting effect optimization, so that the weight parameter update is highly adapted to the prediction deviation and efficiently corrects the prediction deviation of the initial neural network.

[0069] The preset process for this parameter update rule is as follows: Combining the hierarchical structure of the proppant transport prediction neural network and the feature processing functions of each layer's nodes, the adjustment priority of the weight parameters for each layer (input, hidden, and output) is determined. Priority is given to adjusting parameters at layers that have a greater impact on the prediction of the layup morphology, making parameter updates more targeted. Based on the range of deviation reflected by the loss function, the adjustment intensity of weight parameters corresponding to different deviation levels is divided, determining the strengthening adjustment level for large deviations, the regular adjustment level for moderate deviations, and the fine adjustment level for small deviations, achieving a precise match between the deviation magnitude and the update amplitude. Considering the training convergence requirements of the neural network, step size constraints for parameter updates are set, determining the maximum and minimum step size for a single update to avoid problems such as model training oscillations due to excessively large step sizes and low training efficiency due to excessively small step sizes. Simultaneously, the update method and verification standard for parameters at each layer are unified, completing the overall preset of the parameter update rule.

[0070] During the adjustment process, the preset parameter update rules are followed. The update magnitude is matched according to the magnitude of the deviation of the loss function. When the deviation is large, the update magnitude is increased appropriately, and when the deviation is small, the update magnitude is decreased. At the same time, the weight parameters of each layer are adjusted according to the preset priority, taking into account the step size constraint requirements. By updating the weight parameters, the prediction deviation of the initial neural network is corrected, and the fitting effect of the network on the feature data is optimized.

[0071] Step 408 involves iteratively inputting the fused feature vector from the training sample set into the initial neural network, performing forward propagation to obtain the layup pattern prediction data, calculating the loss function between the layup pattern prediction data and the corresponding label values, and updating the network weight parameters until the loss function converges, resulting in a trained prediction neural network model. Specifically, this includes iteratively executing the relevant calculation processes of steps 406 and 407, inputting the fused feature vector from the training sample set again into the neural network after the weight parameter update, performing forward propagation to obtain new layup pattern prediction data, and calculating the new layup pattern prediction data and the corresponding label values. The loss function is calculated and compared with the previous calculation to determine if the loss function shows a convergence trend. Based on the current loss function calculation result, the network weight parameters are updated again to continuously optimize the network performance. The above process of forward propagation calculation, loss function calculation and weight parameter update is repeated until the loss function calculation result tends to be stable and no longer shows a significant decrease or fluctuation, that is, the loss function converges. At this time, the iteration is stopped, and the trained predictive neural network model is obtained. This ensures that the model can stably output the proppant placement morphology prediction data that meets the requirements, providing reliable model support for subsequent proppant migration state prediction.

[0072] In a preferred embodiment of the present invention, step 500 includes: Step 501: Acquire a real-time image sequence of proppant migration within the downhole fracturing fracture during the current fracturing operation, as the current image sequence. This specifically includes: activating the downhole visualization imaging equipment deployed in step 101; acquiring image data of proppant migration with fracturing fluid within the downhole fracturing fracture during the current fracturing operation according to the synchronous acquisition parameters set in step 102; monitoring the equipment's operating status in real time during acquisition to ensure that the equipment's acquisition frequency is synchronized with the proppant addition stage, avoiding image acquisition interruptions or distortion; receiving and initially processing the continuously acquired image data in real time, removing blurry or invalid images during acquisition, sorting the valid images according to chronological order, and adding corresponding timestamps and acquisition location information to each frame to form the current image sequence under the current fracturing operation scenario. This provides the latest, accurate, and complete original image data for subsequent real-time data processing, ensuring that data processing can keep up with the real-time pace of on-site construction.

[0073] Step 502: For each frame of the current image sequence, perform adaptive median filtering for denoising, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement in sequence to obtain the current proppant image to be segmented. Specifically, this includes: extracting single-frame images from the current image sequence frame by frame, and sequentially performing adaptive median filtering for denoising, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement on each frame to ensure that each frame meets the quality requirements of subsequent processing; during the adaptive median filtering for denoising, dynamically adjust the filter window size according to the noise density in different regions of the current image, appropriately increasing the window size in areas with dense noise to thoroughly filter out random noise and spurs introduced by the complex downhole environment. Salt noise is eliminated by narrowing the window in the edge region of proppant particles to preserve particle edge details. After denoising, homomorphic filtering is performed on the denoised image to separate the illuminance and reflectance components. The illuminance component is suppressed, and the reflectance component is enhanced to correct the uneven brightness distribution caused by downhole light source reflection and fracture wall shadows. After illumination correction, adaptive histogram equalization is performed on the uniformly illuminated image to divide the image into multiple non-overlapping sub-regions. The histogram of each sub-region is equalized to stretch the grayscale difference between the proppant particles and the fracturing fluid background, thereby enhancing image contrast. Finally, a clear, stable, and detailed image of the proppant to be segmented is obtained, providing a high-quality image foundation for subsequent image segmentation and feature extraction.

[0074] Step 503: For the current proppant image to be segmented, perform image segmentation, reference marker recognition, spatial correction mesh construction, regional distortion feature calculation, dynamic transport feature parameter extraction, and geometric correction to obtain the current feature parameters to be fused. Specifically, based on the current proppant image to be segmented, perform operations sequentially according to the standardized processing flow from steps 301 to 315 to ensure that the processing logic and operation standards of the current image are completely consistent with the previous feature extraction flow, ensuring the uniformity and accuracy of the extracted parameters from the source, and providing a precise feature foundation for subsequent real-time prediction. The specific implementation process is as follows: The current proppant image to be segmented is normalized according to the input requirements of the pre-trained convolutional neural network segmentation model, and then input into the pre-trained convolutional neural network segmentation model. Through the encoder of the model, multiple layers of convolution and pooling operations are used to extract the shallow texture and deep semantic features of the image layer by layer to obtain multi-scale feature data. After deconvolution and upsampling by the decoder, the feature map is restored to the same size as the input image. Through the pixel-level classification layer, the probability of each pixel belonging to the category of proppant particles, reference markers and background is calculated. The pixel-level category is determined based on the probability value to obtain the pixel-level classification label. The pixel regions of proppant particles and reference markers are selected according to the classification labels. The minimum bounding rectangle is constructed for each reference marker to determine its position information. Finally, a segmented image containing the position information of proppant particle regions and multiple spatial reference markers is obtained.

[0075] The center pixel coordinates of the minimum bounding rectangle of each reference marker are extracted from the segmented image. Combined with the pre-defined actual physical coordinates of the reference marker, a mapping relationship between the pixel coordinate system and the crack physical space coordinate system is established. Based on this mapping relationship, the actual physical range of the imaging field of view is determined. A regular correction grid covering the entire field of view is constructed within the crack physical space. The grid nodes are back-projected onto the image pixel plane to obtain the distortion correction grid. The distortion correction grid is uniformly divided into multiple sub-regions, ensuring that each sub-region contains at least one reference marker. The horizontal and vertical deviation values ​​of the reference marker pixel coordinates and the actual physical coordinates in each sub-region are calculated. After removing abnormal deviations, the average value is taken to obtain the local distortion coefficient of each sub-region. The local distortion coefficients of all sub-regions are integrated in the order of arrangement to form the correction coefficients covering the entire imaging area.

[0076] Propionate particle regions are identified one by one from the segmented images. Independent particles and adherent particles are distinguished by grayscale and contour features. Adhesive particles are separated using region growing. Each particle is assigned a unique label and its centroid coordinates are calculated to obtain particle identification label data. Based on this data, the same propionate particle in adjacent frames is matched, the centroid coordinates are extracted, the displacement vector is calculated and decomposed into horizontal and vertical components, and the instantaneous migration velocity and velocity components of the particles are obtained by combining the image acquisition time interval. At the same time, the number of particles is counted by region and the propionate concentration distribution is calculated by combining the actual area of ​​the region. Sand embankment pixel regions are identified from the segmented images based on grayscale and contour features, and the complete sand embankment contour is delineated. The pixel coordinates of the highest point of the sand embankment and the bottom of the crack are located, and the vertical difference between the two is calculated to obtain the sand embankment height. The pixel coordinates of the leading edge boundary of the sand embankment are tracked in the segmented images of consecutive frames, and the horizontal position difference between adjacent time points is calculated to obtain the leading edge advancement distance of the sand embankment. Finally, the centroid coordinates of propionate particles, instantaneous migration velocity and components, propionate concentration distribution, sand embankment height, and leading edge advancement distance of the sand embankment are integrated to form dynamic migration feature parameters.

[0077] The distortion correction ratio for each sub-region is determined based on the correction coefficients. A coordinate transformation relationship from the image coordinate system to the crack physical space coordinate system is constructed by combining the mapping relationship between the pixel coordinate system and the crack physical space coordinate system. The dynamic transport feature parameters are then transformed region by region according to this transformation relationship, converting all parameter values ​​in the image coordinate system to actual physical values ​​in the crack physical space coordinate system, thus obtaining the physical space dynamic feature parameters. These physical space dynamic feature parameters are then classified and filtered, extracting the proppant position, velocity components, and distribution parameters represented in rectangular coordinates. Polar coordinate transformation is then performed, converting the rectangular coordinate values ​​to polar coordinate values ​​represented by polar radius and polar angle. Finally, all parameters are formatted and outlier data is removed, ultimately yielding the current feature parameters to be fused, containing polar coordinate information. This provides standardized and reliable feature input for subsequent model forward inference.

[0078] Step 504: Input the current feature parameters to be fused into the predictive neural network model. Through forward inference calculation, obtain the predicted data of the current proppant placement morphology in the complex crack. Specifically, this includes: performing format verification on the current feature parameters to be fused to ensure that their format and dimensions are consistent with the input requirements of the predictive neural network model obtained in step 408, avoiding problems such as parameter dimension mismatch and format confusion; after verification, input the current feature parameters to be fused one by one into the trained predictive neural network model. The data is passed from the input layer to the hidden layer. The hidden layer performs deep feature mining, transformation, and integration on the input current feature parameters to be fused, making full use of the intrinsic relationship between proppant migration and placement morphology learned during model training. Forward inference calculation is performed through the output layer to quickly output the predicted data of the current proppant placement morphology in the complex crack. This data corresponds to the current feature parameters to be fused. The entire inference calculation process is efficient and fast, effectively shortening the prediction time of a single frame image.

[0079] Step 505: Based on the current proppant placement pattern prediction data, obtain control instructions for adjusting on-site fracturing construction parameters according to preset control rules. Specifically, this includes: analyzing the prediction data based on the current proppant placement pattern prediction data to extract key information such as the current proppant concentration distribution, sand embankment height, and leading edge advance distance, and determining the specific state of the current proppant placement pattern; retrieving preset control rules, which are pre-established construction parameter adjustment criteria based on fracturing construction process requirements and ideal proppant placement standards. The core of these rules is to match corresponding fracturing construction parameter adjustment schemes for different proppant placement patterns and different migration anomalies, determining the adjustment range, adjustment magnitude, and adjustment direction of key parameters such as fracturing fluid viscosity, construction flow rate, and sand ratio. The preset rules can solve the problem of lacking real-time and precise control basis in fracturing construction, and the inability to effectively close the loop control of the proppant migration process, ensuring that the construction parameter adjustments are highly adapted to the actual proppant migration state, and guaranteeing the proppant placement effect.

[0080] The pre-set process for this control rule involves: reviewing the core technological requirements of fracturing operations, determining the ideal standard for proppant placement, and defining reasonable ranges for key indicators such as concentration distribution, sand embankment height, and leading edge advance distance; combining field construction experience and numerical simulation results to analyze the causes of deviations from the ideal standard in different placement patterns, as well as the impact of adjusting parameters such as fracturing fluid viscosity, construction flow rate, and sand ratio on proppant migration; and setting corresponding construction parameter adjustment schemes for different placement pattern scenarios, including normal scenarios with parameters within reasonable ranges, uneven concentration distribution, excessively high or low sand embankments, and excessively fast or slow leading edge advance, determining the adjustment range, magnitude, and direction of key parameters in each scheme, and defining the limit thresholds for parameter adjustment in conjunction with construction safety regulations, thus completing the formulation and storage of the pre-set control rule.

[0081] Based on the key information of the current proppant configuration obtained from the analysis, subsequent operations are carried out according to the generation logic of the control instructions. The generation logic of the control instructions is as follows: the extracted key proppant indicators are compared with the ideal standard range defined in the preset control rules to determine whether the current proppant configuration is in a normal or abnormal scenario; if it is an abnormal scenario, the corresponding construction parameter adjustment scheme in the matching rules is used, and combined with the actual parameters of the current fracturing construction, the specific parameter adjustment values ​​adapted to the current working conditions are calculated within the adjustment range and amplitude defined in the scheme; the specific adjustment values ​​are integrated with the corresponding adjustment parameters and execution requirements to form a standardized and executable control instruction; through this generation logic, combined with the actual parameters of the current construction, the control instructions used to adjust the on-site fracturing construction parameters are calculated to ensure that the control instructions can adapt to the current proppant migration state and provide a basis for on-site construction parameter adjustment.

[0082] Step 506 involves sending the control command to the on-site fracturing construction control unit to adjust fracturing construction parameters in real time and achieve closed-loop control of the proppant migration process. Specifically, this includes: verifying the format and validity of the control command to ensure that the parameter adjustment range and direction comply with the safety regulations and process requirements of on-site fracturing construction, thus avoiding any unauthorized adjustments; after successful verification, sending the control command to the on-site fracturing construction control unit in real time via a pre-set signal transmission module to ensure stable and delay-free command transmission; the pre-set signal transmission module is a dedicated data transmission module pre-built to adapt to the complex high-temperature and high-pressure downhole construction environment and meet the real-time control requirements of fracturing construction. Its core consists of a downhole data transmission unit, a surface signal receiving unit, and a data interaction protocol, enabling high-speed, stable, and delay-free transmission of control commands from the data processing end to the on-site construction control unit.

[0083] The preset process for this signal transmission module is as follows: Considering the high temperature, high pressure, and strong interference environment of downhole fracturing operations, high-temperature resistant, high-pressure resistant, and electromagnetic interference resistant transmission hardware is selected. A hardware transmission link, including a downhole data transmitter and a surface signal receiver, is built. Simultaneously, suitable signal amplification and anti-interference components are matched to ensure stable operation of the hardware link in the complex downhole environment. Based on the real-time control requirements of fracturing operations, a dedicated data interaction protocol is developed, determining the transmission format, encoding method, verification rules, and transmission rate of control commands to ensure no data loss or corruption during transmission, while also ensuring the transmission rate matches the real-time requirements of the control operations. The hardware transmission link and the dedicated data interaction protocol are integrated and debugged to complete the overall construction of the signal transmission module. Finally, the module is tested and connected with the data processing terminal and the on-site fracturing control unit to verify the stability, real-time performance, and accuracy of command transmission, completing the final deployment of the preset signal transmission module.

[0084] After receiving the control command, the on-site fracturing construction control unit adjusts the fracturing construction parameters in real time according to the command requirements, including key parameters such as fracturing fluid viscosity, construction flow rate, and sand ratio, so that the construction parameters can be adapted to the current migration state of the proppant in a timely manner. At the same time, it collects real-time image sequences of the proppant migration after adjustment, repeats the processing flow from steps 501 to 506, and forms a closed loop of real-time acquisition, real-time processing, real-time prediction, and real-time control, so as to realize closed-loop control of the proppant migration process, ensure the proppant placement effect, and promote intelligent control of fracturing construction.

[0085] like Figure 2 As shown, embodiments of the present invention also provide an image-based proppant transport state prediction system, comprising: The acquisition module is used to acquire real-time image sequences of proppant migration within downhole fracturing fractures; The preprocessing module is used to sequentially perform adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement on each frame of the real-time image sequence to obtain the proppant image to be segmented. The segmentation and correction module is used to perform image segmentation based on convolutional neural networks on the proppant image to be segmented, identify the proppant particle region in the image and multiple spatial reference markers preset in the crack, and obtain a segmented image containing the reference markers; construct a spatial correction grid based on the reference markers in the segmented image, calculate the regional distortion features, and obtain correction coefficients; extract the dynamic migration feature parameters of the proppant particles from the segmented image, and use the correction coefficients to perform geometric correction on the dynamic migration feature parameters to obtain the feature parameters to be fused. The fusion module is used to obtain the feature parameters to be fused, the multivariate parameters affecting proppant migration, and the simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; it associates and fuses the feature parameters to be fused with the multivariate parameters, and combines them with the simulation dataset to construct a feature dataset; it uses the feature dataset to train a preset neural network to obtain a predictive neural network model; The control module is used to process and predict the currently acquired real-time image sequence using a predictive neural network model to obtain the current proppant placement pattern prediction data in complex fractures; based on the current placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

[0086] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0087] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting proppant transport state based on images, characterized in that, The method includes: Acquire real-time image sequences of proppant migration within downhole fracturing fractures; Each frame of the real-time image sequence is sequentially subjected to adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement to obtain the proppant image to be segmented. Image segmentation based on convolutional neural networks is performed on the proppant image to be segmented to identify the proppant particle region in the image and multiple spatial reference markers pre-set in the crack, resulting in a segmented image containing the reference markers; a spatial correction grid is constructed based on the reference markers in the segmented image, and the regional distortion features are calculated to obtain correction coefficients; dynamic migration feature parameters of proppant particles are extracted from the segmented image, and the dynamic migration feature parameters are geometrically corrected using the correction coefficients to obtain the feature parameters to be fused. The process involves obtaining the feature parameters to be fused, the multivariate parameters affecting proppant migration, and a simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; associating and fusing the feature parameters to be fused with the multivariate parameters, and combining them with the simulation dataset to construct a feature dataset; and using the feature dataset to train a pre-defined neural network to obtain a predictive neural network model. The current real-time image sequence is processed and predicted using a predictive neural network model to obtain the current proppant placement pattern prediction data in complex fractures; based on the current placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

2. The image-based proppant transport state prediction method according to claim 1, characterized in that, Acquire real-time image sequences of proppant migration within downhole fracturing fractures, including: High-temperature and high-pressure resistant downhole visualization imaging equipment is pre-deployed at key cluster locations in the fracturing well section to obtain the deployed imaging equipment; Set the image acquisition frequency and acquisition time period of the imaging device to synchronize it with the sand addition stage of the fracturing operation, and obtain the synchronized acquisition parameters. Based on the synchronized acquisition parameters, the imaging device acquires continuous image data of the proppant migration with fracturing fluid within the fracture in real time to obtain the raw image data. The original image data is encoded and stored in chronological order to form a real-time image sequence containing timestamps and location information.

3. The image-based proppant transport state prediction method according to claim 2, characterized in that, Each frame of the real-time image sequence is sequentially subjected to adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement to obtain the proppant image to be segmented, including: Adaptive median filtering is applied to each frame of the real-time image sequence to remove random noise and salt-and-pepper noise introduced by the complex downhole environment, resulting in a denoised image. Homomorphic filtering is applied to the denoised image to correct the uneven brightness distribution caused by downhole light source reflection and fracture wall shadows, resulting in a uniformly illuminated image. Adaptive histogram equalization is performed on the uniformly illuminated image to stretch the grayscale difference between the proppant particles and the fracturing fluid background, thereby enhancing the image contrast and obtaining the proppant image to be segmented.

4. The image-based proppant transport state prediction method according to claim 3, characterized in that, Image segmentation based on a convolutional neural network is performed on the proppant image to be segmented to identify the proppant particle region in the image and multiple spatial reference markers pre-defined within the crack, resulting in a segmented image containing the reference markers, including: The proppant image to be segmented is input into a pre-trained convolutional neural network segmentation model. The input image is then processed by an encoder and decoder structure to extract features and obtain multi-scale feature data. Based on the multi-scale feature data, after upsampling by the decoder and pixel-level classification, the category probability data of each pixel corresponding to the proppant particles, the reference marker and the background are obtained; Each pixel is classified based on the category probability data to obtain a pixel-level classification label. Based on the pixel-level classification markers, a segmented image containing the proppant particle region and the location information of multiple spatial reference markers is obtained.

5. The image-based proppant transport state prediction method according to claim 4, characterized in that, A spatial correction grid is constructed based on the reference markers in the segmented image, and the regional distortion features are calculated to obtain the correction coefficients, including: The pixel coordinates of each spatial reference marker are extracted from the segmented image, and a mapping relationship between the pixel coordinate system and the crack physical spatial coordinate system is established based on the preset actual physical coordinates of the reference markers. Based on the mapping relationship, a regular correction grid covering the entire field of view is constructed in the physical space of the crack, and the regular correction grid is back-projected onto the image pixel plane to obtain the corresponding distortion correction grid; The distortion correction mesh is divided into multiple sub-regions. Within each sub-region, the local distortion coefficient is calculated based on the mapping deviation between the pixel coordinates of the reference marker and the actual physical coordinates. The local distortion coefficients of all sub-regions are combined into a correction coefficient.

6. The image-based proppant transport state prediction method according to claim 5, characterized in that, The dynamic transport feature parameters of proppant particles are extracted from the segmented image, and geometric correction is performed on the dynamic transport feature parameters using correction coefficients to obtain the feature parameters to be fused, including: Each proppant particle is identified and labeled from the segmented image to obtain particle identification label data containing the centroid coordinates of the particles; Based on the particle identification marker data, the centroid coordinates of the same proppant particle between adjacent frames are extracted, the displacement vector between the two frames is calculated and decomposed into velocity components in the horizontal and vertical directions, and the instantaneous transport velocity and its velocity components are obtained by combining the image acquisition time interval; the number of particles in each region is counted based on the particle identification marker data, and the proppant concentration distribution is calculated. The outline of the sand embankment is identified from the segmented image, the vertical distance between the highest point of the sand embankment and the bottom of the crack is measured to obtain the height of the sand embankment, and the change of the position of the leading edge of the sand embankment over time is tracked to calculate the advancing distance of the leading edge. The centroid coordinates, instantaneous migration velocity and its velocity components, proppant concentration distribution, sandbank height, and leading edge advance distance of the proppant particles are used as dynamic migration characteristic parameters. Based on the correction coefficients, a coordinate transformation relationship is constructed from the image coordinate system to the crack physical space coordinate system; Based on the coordinate transformation relationship, the dynamic movement feature parameters are transformed region by region to convert the parameter values ​​in the image coordinate system into actual values ​​in the crack physical space coordinate system, thereby obtaining the physical space dynamic feature parameters. The physical space dynamic feature parameters are transformed into polar coordinates, converting the proppant position, velocity components and distribution parameters in rectangular coordinates into polar coordinates to obtain the feature parameters to be fused containing polar coordinate information.

7. The image-based proppant transport state prediction method according to claim 6, characterized in that, The feature parameters to be fused are obtained, as well as the multivariate parameters affecting proppant migration, and the simulation dataset is constructed based on the coupling method of computational fluid dynamics and discrete element method. The feature parameters to be fused are associated and fused with multivariate parameters, and combined with a simulated dataset to construct a feature dataset, including: The characteristic parameters to be fused are obtained, and real-time monitoring data of fracturing fluid viscosity, proppant particle size, construction flow rate, sand ratio and fracture width during the fracturing operation are obtained as multivariate parameter data. Obtain a pre-constructed proppant migration simulation dataset based on the computational fluid dynamics and discrete element method. The simulation dataset contains proppant migration trajectory, settlement distribution and concentration field data in the crack under different combinations of construction parameters. The feature parameters to be fused are vectorized and concatenated with the multivariate parameter data to obtain a fused feature vector. The simulation feature data that matches the current construction stage is extracted from the simulation dataset, and the simulation feature data is associated and aligned with the fused feature vector to construct a feature dataset containing image features, engineering parameters and simulation data.

8. The image-based proppant transport state prediction method according to claim 7, characterized in that, A pre-defined neural network is trained using a feature dataset to obtain a predictive neural network model, including: The actual data of proppant placement morphology corresponding to each sample in the feature dataset is obtained as the label value. The fused feature vector in the feature dataset is used as the input feature, and together with the label value, they constitute the training sample set and the validation sample set. An initial neural network is constructed, and the initial neural network is trained using the training sample set. The fused feature vectors in the training sample set are input into the initial neural network for forward propagation calculation to obtain the layup pattern prediction data. Calculate the loss function between the predicted layup pattern data and the corresponding label values, and update the network weight parameters; The iterative execution involves inputting the fused feature vectors from the training sample set into the initial neural network, performing forward propagation to obtain the layup pattern prediction data, calculating the loss function between the layup pattern prediction data and the corresponding label values, and updating the network weight parameters until the loss function converges, thus obtaining the trained prediction neural network model.

9. The image-based proppant transport state prediction method according to claim 8, characterized in that, By using a predictive neural network model to process and predict the currently acquired real-time image sequence, the current proppant placement morphology prediction data in complex cracks is obtained. Based on the current proppant placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop, including: Acquire a real-time image sequence of proppant migration within the downhole fracturing fracture during the current fracturing operation, and use it as the current image sequence; For each frame of the current image sequence, adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement are performed sequentially to obtain the current proppant image to be segmented. For the current proppant image to be segmented, image segmentation, reference marker recognition, spatial correction mesh construction, regional distortion feature calculation, dynamic transport feature parameter extraction and geometric correction are performed to obtain the current feature parameters to be fused. The current feature parameters to be fused are input into the prediction neural network model, and the proppant current placement pattern prediction data in complex cracks is obtained through forward inference calculation of the model. Based on the current deployment pattern prediction data, control instructions for adjusting on-site fracturing construction parameters are obtained according to preset control rules; The control commands are sent to the on-site fracturing construction control unit to adjust the fracturing construction parameters in real time, thereby achieving closed-loop control of the proppant migration process.

10. An image-based proppant transport state prediction system, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire real-time image sequences of proppant migration within downhole fracturing fractures; The preprocessing module is used to sequentially perform adaptive median filtering for noise reduction, homomorphic filtering for illumination correction, and adaptive histogram equalization for contrast enhancement on each frame of the real-time image sequence to obtain the proppant image to be segmented. The segmentation and correction module is used to perform image segmentation based on convolutional neural networks on the proppant image to be segmented, identify the proppant particle region in the image and multiple spatial reference markers preset in the crack, and obtain a segmented image containing the reference markers; construct a spatial correction grid based on the reference markers in the segmented image, calculate the regional distortion features, and obtain correction coefficients; extract the dynamic migration feature parameters of the proppant particles from the segmented image, and use the correction coefficients to perform geometric correction on the dynamic migration feature parameters to obtain the feature parameters to be fused. The fusion module is used to obtain the feature parameters to be fused, as well as the multivariate parameters affecting proppant migration, and the simulation dataset constructed based on the coupling method of computational fluid dynamics and discrete element method; it associates and fuses the feature parameters to be fused with the multivariate parameters, and combines them with the simulation dataset to construct a feature dataset; A pre-defined neural network is trained using a feature dataset to obtain a predictive neural network model; The control module is used to process and predict the currently acquired real-time image sequence using a predictive neural network model to obtain the predicted data of the current placement morphology of the proppant in complex cracks. Based on the current proppant placement pattern prediction data, control instructions are obtained and sent to the on-site fracturing construction control unit to control the proppant migration process in a closed loop.

Images