Transition material blasting particle size regulation method based on real-time identification of blasting pile image

Abstract

The application discloses a transition material blasting particle size regulation method based on real-time identification of blasting pile images, belongs to the technical field of blasting engineering, and particularly relates to an automatic regulation method of transition material blasting particle size. The method solves the problem of being difficult to stably and accurately obtain transition material meeting strict grading requirements, collects multi-view digital images of a blasting pile after each round of blasting; identifies stone slag particle outlines based on the images and generates grading data by counting; compares the grading data with a preset target grading range of the transition material, judges whether there is an out-of-tolerance index; when there is an out-of-tolerance index, adjustment values of hole pattern parameters and / or charging parameters for the next round of blasting are generated; then the next round of blasting is performed according to the adjusted parameters, and the foregoing steps are repeated until the grading meets design requirements. The application is mainly used for transition material mining and production in water conservancy and hydropower, mine and other engineering, realizes closed-loop intelligent control of blasting particle size based on real-time image feedback, and significantly improves the output quality and stability of the transition material.

Description

Technical Field

[0001] This invention belongs to the field of blasting engineering technology and relates to a method for controlling the particle size of transition material blasting based on real-time image recognition of blasted piles. Background Technology

[0002] In the construction of large-scale water conservancy and hydropower projects such as pumped storage power stations, transition material is a key material for dam body filling. Its particle size distribution must meet strict continuous gradation requirements, such as a maximum particle size not exceeding 300mm and fine particle content controlled within a specific range. Currently, transition material is typically obtained by controlled blasting extraction of weakly weathered to fresh rock. However, in actual production, the particle size distribution of the blasted rock fragments is affected by various complex factors such as geological conditions, explosive consumption, and borehole parameters, resulting in significant uncertainty.

[0003] Traditional methods for verifying the effectiveness of blasting primarily rely on manual on-site inspections after blasting, combined with indoor sieving tests. This method suffers from significant time lag; often, substandard gradation is only discovered after blasting is complete and samples are sent for testing. By then, substandard material has already been produced, making real-time remediation impossible, necessitating secondary processing or disposal, leading to increased costs and project delays. To address this issue, projects typically require repeated blasting tests to determine suitable construction parameters, a process that is both time-consuming and costly.

[0004] In recent years, although some have attempted to use digital cameras to capture images of blasted piles and perform two-dimensional image processing to estimate particle size, numerous difficulties have been encountered in practical engineering applications. For example, the high dust levels and poor lighting conditions at blasting sites result in poor image quality, affecting recognition accuracy. Furthermore, two-dimensional images cannot effectively handle the stacking and occlusion of stone particles, easily leading to the underestimation of large particles or the omission of small particles, distorting statistical results. Therefore, existing methods struggle to achieve accurate and rapid identification of blasted pile gradation, and consequently, cannot promptly guide the adjustment of blasting parameters for the next round based on real-time results. Consequently, the quality control and construction efficiency of transition material production still need improvement. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0006] This addresses the problem of the lack of a method in existing technologies that can provide real-time feedback on the particle size of the blasted stone and automatically adjust the parameters for the next round of blasting, thereby stably and accurately producing transition materials that meet the gradation requirements.

[0007] The goal is to solve the problem of high-precision identification of the outline of stone particles in complex and variable images of blast piles, especially to obtain their three-dimensional morphological information, in order to overcome the problem of large errors in two-dimensional image analysis.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for controlling the blast particle size of transition material based on real-time image recognition of blasted piles includes the following steps: 1) After each round of blasting is completed, multi-view digital images of the blast pile formed by blasting are collected. The collection of multi-view digital images of the blast pile includes: collecting multi-view image data covering the entire blast pile area by using a fixed camera array and / or a mobile UAV-borne camera deployed around the blast area; and transmitting the collected multi-view image data to a remote server or edge computing node in real time through a wireless communication network. 2) Based on the multi-view digital images, identify the outlines of the stone chips in the blast pile, calculate the equivalent particle size of each particle according to the identified outlines, and then statistically generate gradation data reflecting the particle size distribution of the stone chips in the current round of blasting. 3) Compare the generated gradation data with the preset transition material target gradation range to determine whether there are any gradation indicators that exceed the allowable deviation; 4) When it is determined that there is a gradation index that exceeds the allowable deviation, the hole mesh parameter adjustment value and / or charge parameter adjustment value are generated for the next round of blasting operation according to the preset parameter-gradation mapping relationship. 5) Carry out the next round of blasting operations according to the adjusted blasting parameters; Repeat steps 1) to 5) until the blasting slag gradation meets the design requirements.

[0009] Preferably, in the transition material blasting particle size control method based on real-time image recognition of blasted piles, step 2), identifying the contours of the stone chips, includes: The acquired multi-view images are input into a pre-trained deep learning image segmentation model, which outputs a pixel-level segmentation mask for the stone particles in each image. Cross-view matching and fusion of segmentation masks for the same particle in images from different viewpoints are performed to generate the three-dimensional contour information of each particle.

[0010] The deep learning image segmentation model uses the U-Net architecture as the backbone network, and introduces dilated convolution in the encoder part to expand the receptive field, while introducing skip connections in the decoder part to fuse multi-scale features.

[0011] The identification of the contours of stone particles in the blast pile is alternatively achieved using a morphology-based particle edge detection algorithm, including: Morphological gradient operations are performed on the preprocessed image to enhance edge information; The adaptive Canny edge detection operator is applied to extract particle edges, and broken edges are connected by morphological closing operations. Watershed transform is applied to the edge image to segment interconnected particles.

[0012] Preferably, in the transition material blasting particle size control method based on real-time image recognition of blasted piles, step 2) involves calculating the equivalent particle size of each particle, including: Based on the three-dimensional contour information of each particle, the minimum bounding geometry of the particle is fitted, and the characteristic size of the bounding geometry is used as the equivalent particle size. According to the preset particle size classification standard, the equivalent particle size of all particles is classified into the corresponding particle size range, and the number of particles or cumulative mass of each particle size range is counted to generate gradation data.

[0013] Preferably, in the transition material blasting particle size control method based on real-time image recognition of blasted piles, step 2) further includes an image preprocessing step before recognizing the outline of the stone chips: The acquired raw images are subjected to denoising, contrast enhancement, and geometric correction to improve the recognition accuracy of particle contours.

[0014] The noise reduction process employs an adaptive median filtering algorithm; the contrast enhancement process employs a contrast-limited adaptive histogram equalization algorithm; and the geometric correction process is achieved through perspective transformation based on camera calibration parameters.

[0015] Preferably, in the method for controlling the blasting particle size of transition material based on real-time image recognition of blasted piles, in step 3), the preset target gradation range of the transition material includes: The maximum particle size is no more than 300 mm, the content of particles smaller than 5 mm is 15% to 25%, the content of particles smaller than 0.075 mm is no more than 5%, and the uniformity coefficient Cu > 5, and the curvature coefficient Cc is between 1 and 3.

[0016] Preferably, in the transition material blasting particle size control method based on real-time image recognition of blasted piles, in step 4), the preset parameter-gradation mapping relationship is established in advance in the following way: Collect blasting parameters and corresponding blasting debris gradation data from each blasting test to construct a training dataset; The training dataset is trained using regression analysis or machine learning algorithms to establish a quantitative mapping relationship model between blasting parameters and gradation characteristics; The process of generating adjustment values ​​includes: inputting the deviation between the current round gradation data and the target gradation into the mapping relationship model, and calculating the corrected blasting parameters.

[0017] Preferably, in the method for controlling the particle size of transition material blasting based on real-time image recognition of blasted piles, the blasting parameters include at least one of explosive consumption per unit volume, hole spacing, row spacing, and packing length; the generated adjustment values ​​include: When the proportion of oversized stone chips exceeds the standard, an adjustment command is generated to reduce the mesh parameters or increase the explosive consumption. When the fine particle content is detected to be too low, an adjustment command is generated to increase the explosive consumption or decrease the mesh parameters. When an excessively high content of fine particles is detected, an adjustment command is generated to reduce the explosive consumption or increase the mesh parameters.

[0018] Preferably, in the transition material blasting particle size control method based on real-time image recognition of blasting piles, in step 1), the wireless communication network adopts 5G or Wi-Fi 6 communication protocol, and ensures the complete transmission of image data by means of breakpoint resume transmission.

[0019] A transition material blasting particle size control system based on real-time image recognition of blasted piles, used to perform any of the methods described above, comprising: The image acquisition module is used to acquire multi-view digital images of the blast pile after each round of blasting operations. The particle size analysis module is used to identify the outline of stone particles in the blast pile based on the multi-view digital images, calculate the equivalent particle size, and generate gradation data. The deviation determination module is used to compare the generated gradation data with the preset transition material target gradation range to determine whether there are gradation indicators that exceed the allowable deviation. The parameter adjustment module is used to generate hole mesh parameter adjustment values ​​and / or charge parameter adjustment values ​​for the next round of blasting operations based on the preset parameter-gradation mapping relationship when it is determined that there are gradation indicators that exceed the allowable deviation. The feedback control module is used to output the adjusted blasting parameters to the blasting operation execution end to guide the next round of blasting operations.

[0020] The image acquisition module includes a fixed camera array and / or a mobile UAV-borne camera deployed around the blasting area, as well as a wireless communication unit connected thereto. The wireless communication unit uses 5G or Wi-Fi 6 communication protocols to transmit the acquired image data to a remote server or edge computing node.

[0021] The particle size analysis module includes: The preprocessing unit is used to perform noise reduction, contrast enhancement, and geometric correction on the original image; The particle segmentation unit is used to identify the outline of stone particles using a deep learning image segmentation model or a morphological edge detection algorithm. The particle size calculation unit is used to calculate the equivalent particle size based on the particle profile and to statistically generate gradation data.

[0022] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0023] Compared with the prior art, the present invention has the following advantages: This invention constructs a complete closed-loop control process from blast image acquisition, particle size analysis, deviation determination to parameter adjustment, thereby realizing real-time monitoring and intelligent control of the blasting particle size of the transition material. This significantly improves the accuracy of blasting operations and the stability of transition material quality, and reduces subsequent processing costs.

[0024] This invention effectively improves the recognition accuracy of stone particle outlines in complex environments by introducing deep learning image segmentation models or multi-view matching technology, or by using morphological edge detection algorithms, and can generate three-dimensional outline information, providing a more accurate data basis for particle size calculation.

[0025] This invention transforms image information into quantifiable gradation data by fitting the smallest circumscribed geometry of the particles and performing statistical analysis according to particle size classification standards. This enables a scientific assessment of the particle size distribution of blasted rock slag and provides a reliable basis for deviation determination and parameter adjustment.

[0026] This invention effectively improves image quality and enhances particle edge features by performing preprocessing such as adaptive median filtering, limited contrast histogram equalization, and perspective transformation before particle recognition. This significantly improves the system's adaptability and recognition accuracy in harsh environments.

[0027] This invention provides a clear and quantitative basis for deviation judgment by setting a target gradation range for transition materials, including specific indicators such as maximum particle size, fine particle content, and non-uniformity coefficient, thus ensuring that the produced stone slag strictly meets the engineering design requirements.

[0028] This invention establishes a parameter-gradation mapping model by collecting historical data and using regression analysis or machine learning algorithms. This enables the scientific calculation of the adjustment amount of blasting parameters from gradation deviation, replacing the traditional experience-based trial and error method and significantly improving the efficiency and accuracy of control.

[0029] This invention provides specific parameter adjustment instructions for different gradation deviations (such as oversize and abnormal fine particle content), making the control process more refined and operable, and ensuring that the closed-loop control strategy can be effectively executed.

[0030] This invention significantly improves the stability and reliability of image data transmission by adopting 5G or Wi-Fi 6 communication protocols and combining them with a breakpoint resume mechanism, ensuring the complete and timely uploading of data required for real-time analysis.

[0031] This invention, through modular system design, clarifies the composition and connection relationships of functional units such as image acquisition, particle size analysis, deviation determination, parameter adjustment, and feedback control, forming a complete and engineerable automated control system.

[0032] This invention provides a storage medium containing relevant computer programs, enabling the control method to be easily loaded into various computing devices for execution, greatly enhancing the portability and application prospects of the technology.

[0033] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0034] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0035] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0036] According to one embodiment of the present invention, a method for controlling the blasting particle size of transition material based on real-time identification of blasting images includes the following steps: 1) After each round of blasting is completed, multi-view digital images of the blast pile formed by blasting are collected. The collection of multi-view digital images of the blast pile includes: collecting multi-view image data covering the entire blast pile area by using a fixed camera array and / or a mobile UAV-borne camera deployed around the blast area; and transmitting the collected multi-view image data to a remote server or edge computing node in real time through a wireless communication network. 2) Based on the multi-view digital images, identify the outlines of the stone chips in the blast pile, calculate the equivalent particle size of each particle according to the identified outlines, and then statistically generate gradation data reflecting the particle size distribution of the stone chips in the current round of blasting. 3) Compare the generated gradation data with the preset transition material target gradation range to determine whether there are any gradation indicators that exceed the allowable deviation; 4) When it is determined that there is a gradation index that exceeds the allowable deviation, the hole mesh parameter adjustment value and / or charge parameter adjustment value are generated for the next round of blasting operation according to the preset parameter-gradation mapping relationship. 5) Carry out the next round of blasting operations according to the adjusted blasting parameters; Repeat steps 1) to 5) until the blasting slag gradation meets the design requirements.

[0037] Taking the blasting operation of transition material mining in a water conservancy project as an example, the method of this invention is used for blasting particle size control. A fixed camera array is pre-installed around the blasting area to ensure coverage of the entire blast pile area. Simultaneously, a vertical take-off and landing (VTOL) drone equipped with a high-definition camera is deployed. After each round of blasting, the drone is immediately launched and flies along a preset route, simultaneously triggering the fixed camera to acquire high-resolution digital images of the blast pile from multiple angles. All image data is transmitted in real-time via the on-site 5G wireless communication network to edge computing nodes deployed near the construction site. During transmission, a breakpoint resume function is enabled to cope with temporary network fluctuations and ensure data integrity.

[0038] After receiving multi-view images, the edge computing node first preprocesses the images, including noise reduction and contrast enhancement. Then, it uses a deep learning image segmentation model to identify the contour of each stone chip particle in the image. Because of the use of multi-view images, feature matching can be used to fuse the contours of the same particle from different views, generating three-dimensional contour information for each particle. Based on this three-dimensional contour, the equivalent particle size of each particle is calculated, for example, by fitting its minimum circumscribed cuboid and using the longest side as the equivalent particle size. Subsequently, all particles are graded according to their size, and the mass proportion of each size range is statistically calculated, generating a full gradation curve reflecting the particle size distribution of the blasted stone chips in this round. Key indicators include the maximum particle size, the content of particles smaller than 5 mm, the content of particles smaller than 0.075 mm, the non-uniformity coefficient Cu, and the curvature coefficient Cc.

[0039] The generated gradation data is compared with the preset target gradation range for transition materials. The preset target gradation range is as follows: maximum particle size not greater than 300mm, content of particles smaller than 5mm 15% to 25%, content of particles smaller than 0.075mm not greater than 5%, uniformity coefficient Cu greater than 5, and curvature coefficient Cc between 1 and 3. If all the gradation data in this round falls within this range and the deviation is within the allowable error, the gradation is deemed qualified, and the next round of blasting maintains the original parameters. If there are out-of-tolerance indicators, such as detecting that the proportion of oversized stone chips exceeds the allowable value, the system automatically calculates the parameter adjustment values ​​required for the next round of blasting based on the parameter-gradation mapping relationship model established in advance through machine learning. This mapping model is trained by collecting data from previous blasting tests and can recommend new explosive consumption, hole spacing, row spacing, or packing length based on the deviation between the current gradation and the target.

[0040] Assuming the current analysis reveals an excessive proportion of oversized particles, the model will recommend reducing the hole spacing from 4.5 meters to 4.2 meters, while simultaneously increasing the explosive consumption per cubic meter from 0.45 kg to 0.48 kg. After reviewing this recommendation, the blasting engineer will design the hole mesh and charge parameters for the next blast and organize the operation. Following the completion of the blast, the image acquisition, analysis, comparison, and adjustment steps are repeated until the gradation consistently meets the design requirements across multiple blasting runs. Throughout this process, data is automatically recorded, and control decisions are executed rapidly, achieving closed-loop intelligent control of the blast particle size.

[0041] Traditional methods for controlling particle size during blasting typically rely on the experience of blasting engineers to initially design parameters. After blasting, manual sampling is performed on the surface of the blast pile using a grid method or by measuring a small number of large pieces with a ruler to estimate the proportion of large pieces and the content of fine particles. Then, parameters are roughly adjusted for the next round based on experience. The drawbacks of this method are: limited sampling points, poor representativeness, and inability to accurately obtain gradation information for the entire blast pile; low efficiency of manual measurement, long feedback cycle, often taking several days to complete one round of analysis, resulting in a lag in parameter adjustment; and adjustment decisions depend on personal experience, lacking quantitative basis, making it difficult to consistently produce transition materials that meet strict gradation requirements.

[0042] In comparison, this invention employs real-time acquisition of multi-view digital images, covering the entire blast pile area and avoiding sampling bias; it shortens the feedback cycle to several hours or even less through real-time wireless network transmission and edge computing processing; it automatically determines deviations based on a preset target gradation, and generates precise adjustment values ​​using a data-driven parameter-gradation mapping model, enabling the control process to move from empirical qualitative to scientific quantitative; repeated execution of steps forms a closed loop, gradually approaching and stabilizing within the target gradation range. This invention significantly improves the quality stability, consistency, and efficiency of transition material production, reduces labor costs and reliance on experience, and provides reliable technical support for high-quality construction of large-scale projects.

[0043] This embodiment constructs a closed-loop control system from data acquisition to parameter adjustment and feedback to the next round of operations. This method can acquire the particle size distribution information of the blasting debris after each round of blasting in real time and accurately, automatically diagnose deviations based on preset targets, and intelligently decide on parameter adjustments using a pre-established parameter-gradation quantitative relationship to guide subsequent blasting operations. This cycle continues until the quality is qualified. Its beneficial effects include: achieving online monitoring and automatic control of blasting particle size, greatly shortening the feedback adjustment cycle; improving the accuracy and representativeness of gradation analysis through multi-view image coverage and three-dimensional particle size calculation; using data models to drive parameter optimization, making control more scientific and precise, avoiding the uncertainty of human experience; and ultimately, continuously and stably producing blasting debris that meets the design requirements of the transition material, effectively ensuring the smooth operation of downstream processes and significantly improving project quality and economic benefits.

[0044] According to one embodiment of the present invention, a method for controlling the particle size of transition material blasting based on real-time recognition of blasting images, in step 2), the recognition of the contour of the stone particles includes: The acquired multi-view images are input into a pre-trained deep learning image segmentation model, which outputs a pixel-level segmentation mask for the stone particles in each image. Cross-view matching and fusion of segmentation masks for the same particle in images from different viewpoints generate 3D contour information for each particle. For segmentation masks in images from different viewpoints, their geometric center points and shape descriptors are extracted. A feature matching method based on epipolar geometric constraints is used to determine mask pairs belonging to the same particle. A multi-view stereo vision (MVS) algorithm is employed, based on camera calibration parameters and epipolar geometric constraints, to triangulate the matched particle masks from different viewpoints, generating dense point clouds, and thus constructing the 3D contour of the particle.

[0045] Preferably, cross-view matching is performed on the segmentation mask of the same particle in images from different viewpoints, using a matching method based on a combination of scale-invariant feature transform (SIFT) and epipolar geometric constraints, specifically including the following sub-steps: (1) Feature extraction: For each particle segmentation mask in each viewpoint image, extract the SIFT feature points within the mask area, and use the geometric center of the mask as the candidate matching center of the particle. For particles with small mask areas (e.g., less than 500 pixels), treat the entire mask as a feature region and calculate its shape descriptor (e.g., Hu moment).

[0046] (2) Initial matching: Using the FLANN (Fast Nearest Neighbor Search Library) matcher, SIFT feature points extracted from images from different perspectives are matched, and a threshold strategy of the ratio of nearest neighbor distance to second nearest neighbor distance of less than 0.7 is used to filter the initial matching point pairs.

[0047] (3) Epipolar geometric constraints: Based on the pre-calibrated camera intrinsic parameters and pose information, calculate the fundamental matrix between different viewpoints. For each candidate matching point pair, calculate its distance to the corresponding epipolar line, and remove mismatched point pairs with a distance greater than 3 pixels.

[0048] (4) Mask matching confirmation: For the matching point pairs after being filtered by epipolar geometry constraints, count the number of matching points belonging to the same particle mask region. If the number of valid matching points between two mask regions exceeds a preset threshold (e.g., 5), then the two masks are determined to belong to the same physical particle.

[0049] (5) 3D contour generation: For the segmentation mask of the same particle that is successfully matched in multiple views, the 3D coordinates of each matching point on the mask are calculated by combining the camera projection matrix of each view and the triangulation algorithm to generate a sparse point cloud. Then, the sparse point cloud is expanded into a dense point cloud by the multi-view stereo vision (PMVS) algorithm. Finally, the 3D mesh model of the particle is generated by the Poisson surface reconstruction algorithm as the 3D contour information of the particle.

[0050] The deep learning image segmentation model uses the U-Net architecture as the backbone network, and introduces dilated convolution in the encoder part to expand the receptive field, while introducing skip connections in the decoder part to fuse multi-scale features.

[0051] As a preferred embodiment, the specific configuration of the U-Net architecture is as follows: The encoder consists of four downsampling stages, each consisting of two 3×3 convolutional layers (each followed by a batch normalization layer and a ReLU activation function) and a 2×2 max pooling layer. As downsampling proceeds, the number of feature channels doubles in each stage, for example, from the initial 64 channels to 512 channels.

[0052] The decoder consists of four upsampling stages, each of which includes a 2×2 deconvolutional layer, a skip connection (concatenated along the channel dimension) to the corresponding layer feature map of the encoder, and two 3×3 convolutional layers (each convolutional layer is followed by a batch normalization layer and a ReLU activation function).

[0053] The dilated convolution is inserted into the last two downsampling stages of the encoder in the form of hybrid dilated convolution (HDC), for example, using convolution kernels with dilation rates of 1, 2, and 4 in the second 3×3 convolutional layer of each stage.

[0054] The output layer uses a 1×1 convolutional layer to map the feature map of the last layer of the decoder to a single-channel segmentation mask, and outputs the probability value of each pixel belonging to a particle through the Sigmoid activation function.

[0055] Preferably, the deep learning image segmentation model is trained and validated in the following manner: (1) Data collection: At multiple water conservancy, hydropower and mining blasting project sites, fixed cameras and UAV-borne cameras were used to collect multi-view images of blast piles under different lighting conditions, different dust concentrations and different rock types. A total of 8,000 original images were collected.

[0056] (2) Data annotation: Three trained rock mechanics or blasting engineering professionals used LabelMe image annotation software to perform pixel-level polygonal contour annotation on all visible stone particles in each image. During annotation, for particles that were stacked or occluded, only the visible contours were annotated, along with the occlusion relationships. Approximately 240,000 particles were annotated in total. After the annotation was completed, a fourth expert conducted a random inspection of the annotation results, with a sampling rate of 10%, and the annotation consistency requirement was to reach over 95%.

[0057] (3) Dataset partitioning: The labeled image dataset was randomly divided into a training set (4800 images), a validation set (1600 images), and a test set (1600 images) in a ratio of 6:2:2. During partitioning, it was ensured that images from different engineering scenarios and under different lighting conditions were evenly distributed in each subset.

[0058] (4) Model Training: The U-Net architecture was used as the backbone network. Dilated convolutions were introduced in the encoder (with dilation rates set to 1, 2, and 4 respectively), and skip connections were introduced in the decoder. A composite loss function combining binary cross-entropy loss and Dice loss was used. The optimizer was Adam, with an initial learning rate of 0.0001, a batch size of 8, and 200 training epochs. The model performance was evaluated on the validation set every 10 epochs, and the model with the highest Dice coefficient on the validation set was saved as the final model.

[0059] (5) Model testing: The final model is tested on the test set. The test indicators include the mean intersection-over-union ratio (mIoU) and pixel accuracy. When the mIoU on the test set reaches 0.85 or above, the model is considered to meet the usage requirements and can be used for the segmentation of stone particles in actual blast pile images.

[0060] The identification of the contours of stone particles in the blast pile is alternatively achieved using a morphology-based particle edge detection algorithm, including: Morphological gradient operations are performed on the preprocessed image to enhance edge information; The adaptive Canny edge detection operator is applied to extract particle edges, and broken edges are connected by morphological closing operations. Watershed transform is applied to the edge image to segment interconnected particles.

[0061] In one specific embodiment, a deep learning image segmentation model is used for particle segmentation to accurately identify the contours of blasted slag particles. First, a neural network based on the U-Net architecture is constructed as the backbone network. The encoder part of this network introduces dilated convolutions, which effectively expand the receptive field while maintaining the feature map size by setting different dilation rates. This allows the model to capture a wider range of contextual information, which is beneficial for identifying the overall contours of large slag particles. The decoder part introduces skip connections, fusing the multi-scale features extracted from each layer of the encoder with the corresponding layers of the decoder. This restores spatial resolution while preserving rich detail information, resulting in finer segmentation boundaries. This network is trained on a large dataset of labeled blasted slag images, resulting in a model that can accurately output pixel-level segmentation masks for each slag particle.

[0062] Multi-view images acquired after the blasting operation are input into the model to obtain a 2D segmentation mask for all visible particles in each image. Subsequently, a cross-view matching and fusion step is performed. Utilizing multi-view geometric principles and particle feature matching algorithms, segmentation masks belonging to the same particle in images from different views are associated. Combined with camera pose information, 3D contour information for each particle is generated using methods such as triangulation or voxel reconstruction. This process effectively solves the problem of incomplete contours caused by particle occlusion in single-view images and overcomes the influence of particle projection deformation on size measurement in 2D images.

[0063] As an alternative, morphology-based particle edge detection algorithms can be used in scenarios with limited computing resources or requiring rapid processing. First, morphological gradient operations are performed on the preprocessed image, i.e., subtracting the erosion map from the dilation map, to enhance the edge information of particles in the image. Then, the adaptive Canny edge detection operator is applied. This operator can automatically adjust the high and low thresholds according to local regions of the image, more robustly extracting continuous single-pixel wide edges. Since the extracted edges may be broken, morphological closing operations are used—dilation followed by erosion—to connect adjacent broken edges, closing the particle contours. Finally, a watershed transform is performed on the edge image, treating the image as a topographic map, with gray values ​​corresponding to altitude. Starting from a local minimum, water is flooded, and watersheds are constructed at the points where different regions meet during the flooding process, effectively separating mutually adhering particles and preventing multiple particles from being misclassified as a single entity.

[0064] This invention employs an improved U-Net deep learning model, introducing dilated convolution and multi-scale feature fusion, significantly enhancing the segmentation accuracy of particle contours in complex backgrounds. It can better distinguish particles from the background and mutually occluded particles. Cross-view matching and 3D contour generation techniques fundamentally overcome the limitations of 2D analysis, providing a more realistic 3D size basis for particle size calculation. Furthermore, the combination of morphological gradient, adaptive Canny, and watershed transform in the alternative approach offers stronger adaptability and better segmentation of adhered particles compared to traditional fixed threshold methods, providing a flexible, efficient, and low-cost solution for specific scenarios.

[0065] This embodiment can adapt to the complexity and variability of populated images, transforming the original image into quantifiable particle contour information. Through deep learning models, especially the improved U-Net architecture, the accuracy and robustness of particle segmentation are significantly improved, reducing missegmentation and missed segmentation. Dilated convolution expands the receptive field, making the recognition of large particles more complete. Skip connections fuse multi-scale features, making the boundaries of small particles and details clearer. Cross-view matching and 3D contour generation techniques overcome the inherent occlusion and projection distortion problems of 2D images, providing a data foundation closer to the real 3D morphology for subsequent particle size calculation. At the same time, morphology-based alternatives provide flexible choices for different application scenarios, ensuring that relatively reliable particle segmentation results can still be obtained when computational resources are limited, thereby ensuring the overall accuracy and adaptability of the entire control system.

[0066] According to one embodiment of the present invention, a method for controlling the particle size of transition material blasting based on real-time recognition of blasting images, in step 2), calculating the equivalent particle size of each particle includes: Based on the three-dimensional contour information of each particle, the minimum bounding geometry of the particle is fitted, and the characteristic size of the bounding geometry is used as the equivalent particle size; preferably, the equivalent particle size is defined as the length of the longest side of the minimum bounding cuboid that can completely enclose all point clouds of the particle. According to the preset particle size classification standard, the equivalent particle size of all particles is classified into the corresponding particle size range, and the number of particles or cumulative mass of each particle size range is counted to generate gradation data.

[0067] In one specific embodiment, after obtaining the three-dimensional contour information of each particle, it is necessary to calculate the equivalent particle size and perform gradation data statistics. For the three-dimensional point cloud or mesh model of each stone particle reconstructed through deep learning multi-view matching, an algorithm is used to fit its minimum bounding geometry. For example, for particles with relatively regular shapes, a minimum bounding cuboid can be fitted; for particles that are closer to a sphere, a minimum bounding sphere can be fitted; and for elongated particles, a minimum bounding ellipsoid can be fitted. The fitting process uses an iterative optimization algorithm to find the geometry that can completely contain all the point clouds of the particle and has the smallest volume. Then, the characteristic size of the minimum bounding geometry is used as the equivalent particle size. For example, for the minimum bounding cuboid, the maximum value of the length, width, and height, i.e., the longest side size, can be taken as the equivalent particle size. This reflects the maximum extension scale of the particle in space and has a good correlation with the sieve aperture size passed through in actual screening.

[0068] After calculating the equivalent particle size of all particles, they are categorized according to a preset particle size classification standard. This standard can be set based on engineering requirements, such as dividing the particle size into intervals of 5mm, 20mm, 40mm, 80mm, 150mm, and 300mm. For each particle size interval, the number of particles falling within that interval is counted, and the mass of each particle is estimated based on rock density and particle volume. This results in the percentage of each interval's mass relative to the total mass, ultimately generating a full gradation curve or gradation table reflecting the particle size distribution of the blasted rock debris. This method, based on three-dimensional information and calculating and counting particle by particle, is more comprehensive and accurate than traditional sieving sampling or two-dimensional image statistical methods.

[0069] This invention, based on the three-dimensional contour information of particles, fits the minimum bounding geometry and uses its characteristic dimensions as the equivalent particle size, making the particle size value closer to the true maximum size of the particles and more consistent with the physical meaning of sieved particle size. The gradation of all particles is calculated and statistically analyzed for each particle individually, avoiding the random errors of sampling statistics, and ensuring that the generated gradation data accurately reflects the particle size distribution characteristics of the entire swarm. This refined three-dimensional statistical method provides reliable data support for subsequent accurate comparison with the target gradation and precise parameter control.

[0070] This embodiment uses the minimum circumscribed geometry and characteristic dimensions as the equivalent particle size, making the particle size definition more scientific and reasonable, and improving its correlation with sieving test results, thus avoiding systematic errors caused by two-dimensional projection. The method of statistically analyzing and accumulating data for each particle fully utilizes all the information obtained from image recognition, resulting in more representative and accurate gradation data. The clear particle size grading standard enables the generated gradation curve to seamlessly connect with engineering design requirements, providing a directly comparable input for subsequent deviation judgment. This refined statistical method based on three-dimensional information significantly improves the accuracy and reliability of the entire control system in assessing the quality of blasted rockfill, laying the foundation for precise control.

[0071] Preferably, in step 1), acquiring multi-view digital images of the exploded pile further includes: Simultaneously acquire 3D point cloud data of the burst pile; The acquisition of the three-dimensional point cloud data of the blast pile is achieved through a line laser scanner or a time-of-flight (ToF) depth camera deployed on a fixed camera array or an unmanned aerial vehicle (UAV) camera.

[0072] Preferably, in step 2), after generating the three-dimensional contour information of each particle and before calculating the equivalent particle size, the method further includes: The generated incomplete particle 3D contour information is input into a pre-trained point cloud completion network, and the complete particle 3D contour information is output.

[0073] The point cloud completion network is based on a generative adversarial network (GAN) or PointNet++ architecture and is trained using a large number of complete rock samples and artificially simulated occlusion samples.

[0074] Preferably, in step 2), after generating the gradation data reflecting the particle size distribution of the blasting debris in the current round, a real-time calibration step is also included: Rapid on-site screening was conducted in at least one representative area of ​​the explosive pile to obtain real gradation sample data; The gradation data generated by the image recognition is compared with the real gradation sample data to establish and dynamically update the deviation correction model; The deviation correction model is used to correct the gradation data generated by subsequent image recognition.

[0075] To address the issue of incomplete 3D contour information caused by particle stacking and occlusion in burst images, which leads to a systematic underestimation of the minimum bounding geometry, the following multi-level technical solutions can be adopted to resolve or compensate for this problem: First, active stereo vision or structured light scanning technology is introduced during the image acquisition and 3D reconstruction stages. A line laser scanner or depth camera (such as a ToF camera) is mounted on a drone or fixed support to actively project light spots or stripes, acquiring high-density 3D point cloud data of the blast pile surface. Since the active light source can penetrate some surface shadows and illuminate recessed areas, it can significantly reduce texture loss caused by uneven natural lighting. Simultaneously, the laser point cloud can directly reflect the true geometry of the particle surface; even if particles are stacked, the geometric information of their exposed parts can be captured more accurately, providing more reliable initial data for subsequent contour completion.

[0076] Secondly, a deep learning-based point cloud completion and shape generation network is employed. For partial point clouds obtained from multi-view image reconstruction or laser scanning, a 3D generative adversarial network (3D-GAN) or a point cloud completion network (such as PointNet++ combined with an encoder-decoder structure) is trained. By learning the geometric features of a large number of complete particle samples, the network can infer the reasonable shape of the occluded portion from local observations. Training data can come from various rock samples scanned in the laboratory, particle models generated from blasting simulations, or the complete outlines of partially occluded particles annotated manually. This network takes the incomplete point cloud as input and outputs a complete 3D shape, then fits the minimum bounding geometry, making the calculated equivalent particle size closer to the true value.

[0077] Third, a statistical correction model based on physical crushing laws is introduced. The shape of blasted stone chips often follows a certain statistical distribution (such as fractal dimension and aspect ratio distribution). By screening and 3D scanning a large amount of blasted stone chips from the same quarry in the early stage, a priori library of particle shapes for different particle size ranges is established (e.g., large-diameter particles tend to be blocky, while small-diameter particles tend to be flaky or needle-like). When the system detects that a particle is severely obscured, it can match the most similar probability density function from the priori library based on the size, shape, and particle size range of its visible part, and perform statistical extrapolation on the invisible part, thereby correcting the calculated value of the equivalent particle size.

[0078] Fourth, establish a real-time calibration mechanism between image recognition results and on-site sieving tests. After each round of blasting, in addition to image acquisition, rapid sieving is simultaneously performed in multiple representative areas of the blast pile (e.g., using a portable vibrating screen or grid method) to obtain the true particle size distribution of a small number of samples. The true particle size of these samples is compared with the estimated particle size of particles in the same area by the image system to establish a deviation correction curve (e.g., system estimate = true value × correction coefficient + intercept). As the number of rounds increases, the accumulated calibration data can be used to update the regression model online, dynamically compensate for systematic underestimation caused by occlusion, and ensure the accuracy of closed-loop control.

[0079] Finally, the fitting strategy for the minimum bounding geometry is improved at the algorithm level. Instead of directly fitting based solely on the visible point cloud, the contour boundary of the occluded area is inferred first using texture information (such as shadows and edge direction) from multi-view images. Constraint optimization methods based on symmetry assumptions (most natural rocks have approximate symmetry) or shape priors are adopted to ensure that the fitted bounding geometry can fit the visible part while also satisfying the reasonable constraints of the overall shape, thereby reducing the underestimation.

[0080] According to one embodiment of the present invention, a method for controlling the particle size of transition material blasting based on real-time image recognition of blasted piles, in step 2), before recognizing the contours of stone particles, an image preprocessing step is further included: The acquired raw images are subjected to denoising, contrast enhancement, and geometric correction to improve the recognition accuracy of particle contours.

[0081] The noise reduction process employs an adaptive median filtering algorithm; the contrast enhancement process employs a contrast-limited adaptive histogram equalization algorithm; and the geometric correction process is achieved through perspective transformation based on camera calibration parameters.

[0082] In a specific embodiment of the present invention, before inputting the acquired multi-view digital images into the particle recognition model, an image preprocessing step is first performed to improve image quality. The first step is denoising, employing an adaptive median filtering algorithm. Unlike traditional median filtering which uses a fixed window, this algorithm dynamically adjusts the filter window size based on the grayscale characteristics of the pixels within the window. This effectively removes salt-and-pepper noise and Gaussian noise caused by dust, uneven lighting, or sensor noise, while better preserving the detailed information of particle edges and avoiding edge blurring. The second step is contrast enhancement, employing a contrast-limited adaptive histogram equalization algorithm. This algorithm divides the image into multiple small regions, performs histogram equalization in each region, and limits the contrast amplification by cropping the histogram amplitude. This effectively enhances the contrast between the stone particles and the background in the image, making the particle outlines clearer and more discernible, while avoiding excessive enhancement that could lead to noise amplification. The third step is geometric correction, achieved through perspective transformation based on camera calibration parameters. Before acquiring images with a fixed camera and an UAV-mounted camera, the camera is pre-calibrated to obtain its intrinsic parameters and distortion coefficients. For the acquired raw images, distortion correction is performed using calibration parameters. Then, based on the pose information at the time of shooting, perspective transformation is used to uniformly correct images from different viewpoints to the same coordinate system, eliminating geometric distortion caused by camera tilt and height changes, and ensuring that the geometric size of particles in the image is accurately proportional to their actual physical size.

[0083] The images processed through the above three steps have clear particle edges, accurate geometric dimensions, and suitable contrast, providing high-quality input data for subsequent particle contour recognition and particle size calculation.

[0084] In this invention, adaptive median filtering preserves edges while denoising, outperforming traditional fixed-window filtering; contrast-limited adaptive histogram equalization enhances local contrast while suppressing noise amplification, outperforming global histogram equalization; and perspective transformation based on calibration parameters accurately corrects geometric distortion, outperforming simple affine transformation or ignoring distortion. The combined application of these preprocessing steps significantly improves the system's adaptability to harsh environments.

[0085] In this embodiment, adaptive median filtering effectively removes noise interference while protecting the sharpness of particle edges and reducing false recognition. Contrast-limited adaptive histogram equalization enhances the contrast between particles and the background, making it easier to detect small particles and weak edges, while avoiding excessive amplification of noise. Perspective transformation based on camera calibration corrects geometric distortion, ensuring consistency between image measurement and actual size, and eliminating systematic errors caused by differences in shooting angle and position. These three preprocessing steps work together to construct a robust image front-end processing flow, significantly improving the robustness and accuracy of the entire control system in complex field environments, enabling subsequent deep learning models or traditional algorithms to achieve optimal performance.

[0086] According to one embodiment of the present invention, a method for controlling the blasting particle size of transition material based on real-time image recognition of blasted piles, in step 3), the preset target gradation range of the transition material includes: The maximum particle size is no more than 300 mm, the content of particles smaller than 5 mm is 15% to 25%, the content of particles smaller than 0.075 mm is no more than 5%, and the uniformity coefficient Cu > 5, and the curvature coefficient Cc is between 1 and 3.

[0087] In a specific embodiment of the present invention, in order to achieve precise control of the blasting particle size of the transition material, it is first necessary to define the target gradation range of the transition material as the basis for quality judgment. Based on the special requirements of water conservancy and hydropower projects for transition materials, the preset target gradation range is set as follows: the maximum particle size is no greater than 300 mm, the content of particles smaller than 5 mm is 15% to 25%, the content of particles smaller than 0.075 mm is no greater than 5%, and the coefficient of uniformity Cu is greater than 5, and the coefficient of curvature Cc is between 1 and 3. This range comprehensively reflects the particle size composition characteristics that the transition material should possess as a seepage-proof transition layer. It requires a certain amount of coarse particles to form a skeleton, a moderate content of fine particles to fill the pores, and good gradation continuity to meet the functions of reverse filtration and transition.

[0088] After each round of blasting, the image acquisition and analysis system generates the gradation data for the current blast pile. The system automatically compares this data with the preset target gradation range item by item. First, it checks whether the maximum particle size exceeds 300mm. Then, it calculates whether the mass percentage of particles smaller than 5mm falls within the range of 15% to 25%. Next, it calculates whether the content of particles smaller than 0.075mm exceeds the standard. Finally, it calculates whether the non-uniformity coefficient and curvature coefficient meet the requirements. Only when all indicators fall within the allowable deviation range is the blasting gradation deemed qualified. If any indicator exceeds the allowable deviation, the system triggers a parameter adjustment process to optimize the parameters for the next round of blasting for the specific out-of-tolerance item.

[0089] For example, after a blast, analysis showed that the maximum particle size was 320mm, slightly exceeding the standard, while the content of particles smaller than 5mm was only 12%, below the lower limit, although the coefficient of uniformity and the coefficient of curvature were within the required range. The system determined that there were two deviations: excessively large particle size and insufficient fine particle content. The parameters for the next round of blasting will be comprehensively adjusted based on the mapping relationship model in order to improve both problems simultaneously.

[0090] This embodiment ensures that the blasted material contains no excessively large stones by limiting the maximum particle size to no more than 300mm, thus avoiding difficulties in subsequent crushing or affecting the filling quality. It also ensures sufficient fine material fills the pores of coarse particles by limiting the content of particles smaller than 5mm to 15% to 25%, forming a dense structure. Furthermore, it prevents excessively high particle content from affecting permeability stability by limiting the content of particles smaller than 0.075mm to no more than 5%. Finally, it ensures good gradation continuity by limiting the coefficient of uniformity to greater than 5 and the coefficient of curvature to between 1 and 3, meeting the requirements for reverse filtration and transition functions. These specific indicators transform the quality control of the transition material from a vague concept into an executable technical standard, providing precise targets for the entire closed-loop control system and significantly improving the targeting and effectiveness of the control.

[0091] According to one embodiment of the present invention, a method for controlling the blasting particle size of transition material based on real-time identification of blasting images, in step 4), the preset parameter-gradation mapping relationship is established in advance in the following manner: Collect blasting parameters and corresponding blasting debris gradation data from each blasting test to construct a training dataset; The training dataset is trained using regression analysis or machine learning algorithms to establish a quantitative mapping relationship model between blasting parameters and gradation characteristics. The mapping relationship model is an inverse model, whose input is the deviation values ​​of n key indicators between the current gradation and the target gradation (including: maximum particle size deviation, content deviation of less than 5 mm, and content deviation of less than 0.075 mm), and the output is the adjustment values ​​of m adjustable parameters for the next round of blasting (including: explosive consumption adjustment, hole spacing adjustment, and row spacing adjustment).

[0092] The process of generating adjustment values ​​includes: inputting the deviation between the current round gradation data and the target gradation into the mapping relationship model, and calculating the corrected blasting parameters.

[0093] Preferably, the parameter-gradation mapping relationship model is constructed using a backpropagation neural network (BP neural network), and the node settings of its input and output layers are as follows: The input layer contains 4 nodes, which correspond to the deviation values ​​of 4 key gradation indicators between the current round of blasted rock slag gradation and the preset target gradation. The 4 key gradation indicators include: maximum particle size deviation, particle size content deviation of particles smaller than 5 mm, particle size content deviation of particles smaller than 0.075 mm, and non-uniformity coefficient Cu deviation. The output layer contains 3 nodes, which correspond to the adjustment amount of 3 adjustable blasting parameters in the next round of blasting operation. The 3 adjustable blasting parameters include: explosive consumption adjustment amount, hole spacing adjustment amount, and row spacing adjustment amount. The number of hidden layer nodes was determined to be 10 through trial and error. The network was trained using the Levenberg-Marquardt algorithm, with the training objective being to minimize the root mean square error between the predicted adjustment and the actual required adjustment. After training, the gradation deviation value of the current round was input into the network, and the adjustment amount of the blasting parameters for the next round could be directly calculated.

[0094] In one specific embodiment, to automatically generate blasting parameter adjustment values ​​based on gradation deviation, a parameter-gradation mapping relationship needs to be established in advance. First, a series of blasting tests are conducted, recording detailed blasting parameters for each test, including explosive consumption, hole spacing, row spacing, packing length, and initiation method. After blasting, the corresponding stone slag gradation data, including maximum particle size, passing rate of each sieve size, non-uniformity coefficient, and curvature coefficient, is obtained using the image analysis system of this invention. These data are collected to construct a training dataset containing hundreds or even more samples, each sample consisting of input parameters and output gradation features.

[0095] Then, regression analysis or machine learning algorithms are used to train the dataset. If regression analysis is used, a multiple linear regression or nonlinear regression model can be established to fit the functional relationship between each blasting parameter and the gradation characteristic index. If machine learning is used, algorithms such as random forest, support vector machine, or neural network can be selected to obtain a black-box model that can predict the gradation output based on the input parameters, or conversely, an inverse model that can recommend input parameters based on the target gradation. After training, a pre-defined parameter-gradation mapping relationship model is obtained, which can quantitatively describe the influence of changes in blasting parameters on the gradation result.

[0096] In actual control operations, when the gradation data of a certain round of blasting deviates from the target gradation, the system inputs the deviation value into the mapping model. Based on built-in functional relationships or learned patterns, the model calculates the amount of blasting parameter adjustment required to bring the gradation of the next round closer to the target range. For example, if there are too many particles larger than 300mm in the current gradation, the model might calculate that the hole spacing needs to be reduced by 0.3m and the explosive consumption per unit area increased by 0.03kg / m. 3 The proposed adjustments are then output for use in the next round of blasting design.

[0097] This embodiment constructs a training dataset by collecting data from previous blasting tests, enabling the model to be built on actual engineering data, thus possessing specificity and reliability. Employing regression analysis or machine learning algorithms, it can uncover complex nonlinear relationships between multiple variables, outperforming traditional empirical formulas. The trained mapping model can use the deviation between the current gradation and the target as input to directly calculate the required parameter adjustments, achieving an automated closed loop from deviation diagnosis to parameter correction. This data-driven modeling method can be continuously optimized and updated with data accumulation, enabling the control system to possess self-learning and continuous improvement capabilities. Over long-term operation, its accuracy and adaptability will increase, significantly improving the quality stability of transition material production.

[0098] According to one embodiment of the present invention, a method for controlling the particle size of transition material blasting based on real-time image recognition of blasted piles is provided, wherein the blasting parameters include at least one of explosive consumption per unit volume, hole spacing, row spacing, and packing length; the method for generating adjustment values ​​includes: When the proportion of oversized stone chips exceeds the standard, an adjustment command is generated to reduce the mesh parameters or increase the explosive consumption. When the fine particle content is detected to be too low, an adjustment command is generated to increase the explosive consumption or decrease the mesh parameters. When an excessively high content of fine particles is detected, an adjustment command is generated to reduce the explosive consumption or increase the mesh parameters.

[0099] In one specific embodiment, when the deviation determination module identifies a specific gradation index as out of tolerance, the parameter adjustment module generates a targeted adjustment instruction based on a preset mapping relationship model. The blasting parameters involved include at least one of explosive consumption, hole spacing, row spacing, and filling length. The adjustment instruction follows clear logical rules depending on the type of deviation.

[0100] After a blasting operation, image analysis and gradation statistics showed that the proportion of oversized rock fragments (greater than 300mm) in the blast pile exceeded the upper limit of the target range. The system determined that the proportion of oversized rock fragments exceeded the limit, meaning the blasting energy was insufficient to fully break up large rocks, or the borehole layout was too sparse, resulting in uneven breaking. In this case, the system generated adjustment instructions to either reduce the borehole mesh parameters or increase the explosive consumption per unit volume. Reducing the borehole mesh parameters means decreasing the hole spacing and row spacing, making the borehole layout denser and distributing the explosive energy more evenly in the rock mass, thus enhancing the rock-breaking effect. Increasing the explosive consumption per unit volume directly increases the explosive energy obtained per unit volume of rock; both can effectively reduce the yield of large rocks. The mapping model calculates specific adjustment amounts based on the current degree of deviation, for example, suggesting reducing the hole spacing from 4.5m to 4.2m, the row spacing from 3.5m to 3.3m, and simultaneously reducing the explosive consumption per unit volume from 0.45kg / m³. 3 Increased to 0.48 kg / m 3 .

[0101] After another round of blasting, gradation analysis showed that the content of fine particles smaller than 5mm was only 10%, lower than the target lower limit of 15%. The system determined that there was a problem with the excessively low content of fine particles, indicating that the blasting process was too coarse, the crushing effect was insufficient, and it failed to produce enough fine material to fill the pores of coarse particles. According to the logic of the amended claim 7, the explosive consumption per unit volume should be increased or the mesh parameters should be decreased to enhance the crushing effect. The system generated an adjustment command, suggesting that the explosive consumption per unit volume be reduced from 0.45kg / m³. 3 Increased to 0.50 kg / m 3 Meanwhile, the hole spacing was reduced from 4.5m to 4.3m to provide more energy for the fine fragmentation of the rock.

[0102] After another round of blasting, gradation analysis showed that the content of fine particles smaller than 5mm reached 30%, exceeding the target upper limit of 25%. The system determined that there was an excessively high content of fine particles, which meant that the blasting was excessively crushing, wasting explosive energy and potentially causing the permeability of the transition material to fail to meet requirements. In this case, the explosive consumption per unit volume should be reduced or the mesh parameters should be increased to weaken the crushing intensity. The system generated an adjustment command, suggesting reducing the explosive consumption per unit volume from 0.45kg / m³. 3 Reduced to 0.42 kg / m 3 Furthermore, the hole spacing was increased from 4.5m to 4.8m, and the row spacing was increased from 3.5m to 3.7m, making the blasting action more gentle and reducing the generation of fine particles.

[0103] In each adjustment, the system's output instructions are fed back to the blasting operation execution end to guide the next round of blasting design. Through multiple rounds of closed-loop control, the blasting mix gradually approaches and stabilizes within the target range.

[0104] This embodiment clearly defines that energy input should be increased when the ultra-large particle size exceeds the standard, energy input should also be increased when the fine particle content is too low, and energy input should be reduced when the fine particle content is too high. This logical rule is simple and clear, easy to understand and implement. By correctly associating the adjustment direction with the deviation type, it avoids the logical errors commonly found in traditional experience-based adjustments, improving the success rate and efficiency of control. This basic logic remains applicable even under different engineering conditions, exhibiting broad universality. When used in conjunction with a mapping relationship model, the model only needs to learn the specific amplitude based on the correct direction, reducing modeling difficulty and improving the model's stability and reliability. The entire closed-loop control system thus possesses inherent logical consistency, ensuring that every parameter adjustment is justified, gradually approaching the optimal blasting effect, and ultimately achieving stable and controlled transition material quality.

[0105] According to one embodiment of the present invention, a method for controlling the particle size of transition material blasting based on real-time identification of blasting images is provided. In step 1), the wireless communication network adopts the 5G or Wi-Fi 6 communication protocol and ensures the complete transmission of image data by means of breakpoint resume transmission.

[0106] In one specific embodiment, to achieve real-time and stable transmission of multi-view image data of the blast site, the wireless communication network adopts fifth-generation mobile communication technology, namely the 5G communication protocol or the Wi-Fi 6 communication protocol. The fixed camera arrays deployed around the blast area and the mobile UAV-borne cameras all integrate communication modules supporting 5G or Wi-Fi 6, enabling rapid uploading of the acquired high-definition multi-view image data.

[0107] During data transmission, the breakpoint resume function is enabled. When a drone or fixed camera encounters a temporary network signal interruption or instability during data transmission, such as in areas with weak 5G signal coverage at a mountain blasting site, the communication module automatically stores the untransmitted image data in a local cache. Once the network connection is restored, the system automatically resumes uploading the remaining data from the point of interruption without restarting the transmission of the entire file. This ensures that all acquired image data ultimately reaches the remote server or edge computing node intact, avoiding data loss or duplicate transmissions caused by network fluctuations.

[0108] For example, during a blasting operation, the drone passed through a signal blind spot during its flight, and some images were not transmitted in real time. These images were automatically saved in the drone's built-in storage. When the drone flew out of the blind spot and reconnected to the network, the system immediately detected the breakpoint and automatically resumed transmission. All images arrived completely on the server within minutes, ensuring the timely execution of subsequent particle size analysis.

[0109] This invention employs 5G or Wi-Fi 6 communication protocols. These two next-generation wireless communication technologies feature high bandwidth, low latency, and a large number of connections, enabling them to meet the demands of simultaneous uploading of multiple high-definition video streams. They are particularly suitable for scenarios like those with high-volume network congestion where rapid transmission of large amounts of image data is required. The addition of a resume function further enhances transmission reliability, ensuring data integrity and transmission efficiency even in unstable network environments.

[0110] This embodiment utilizes the high bandwidth of 5G or Wi-Fi 6 to support the simultaneous and rapid uploading of multiple high-definition images, shortening the waiting time from acquisition to analysis and significantly reducing the feedback cycle. The breakpoint resume mechanism effectively addresses the problem of unstable networks in the field, avoiding data loss or retransmission due to signal interruption, and improving the system's robustness and data integrity in harsh environments. Real-time and stable transmission lays the foundation for rapid response in closed-loop control, enabling the control system to truly achieve timely analysis and adjustment after each round of blasting, rather than being delayed for many days, significantly improving the timeliness and effectiveness of control.

[0111] According to one embodiment of the present invention, a transition material blasting particle size control system based on real-time image recognition of blasted piles is used to perform any of the methods described above, comprising: The image acquisition module is used to acquire multi-view digital images of the blast pile after each round of blasting operations. The particle size analysis module is used to identify the outline of stone particles in the blast pile based on the multi-view digital images, calculate the equivalent particle size, and generate gradation data. The deviation determination module is used to compare the generated gradation data with the preset transition material target gradation range to determine whether there are gradation indicators that exceed the allowable deviation. The parameter adjustment module is used to generate hole mesh parameter adjustment values ​​and / or charge parameter adjustment values ​​for the next round of blasting operations based on the preset parameter-gradation mapping relationship when it is determined that there are gradation indicators that exceed the allowable deviation. The feedback control module is used to output the adjusted blasting parameters to the blasting operation execution end to guide the next round of blasting operations.

[0112] The image acquisition module includes a fixed camera array and / or a mobile UAV-borne camera deployed around the blasting area, as well as a wireless communication unit connected thereto. The wireless communication unit uses 5G or Wi-Fi 6 communication protocols to transmit the acquired image data to a remote server or edge computing node.

[0113] The particle size analysis module includes: The preprocessing unit is used to perform noise reduction, contrast enhancement, and geometric correction on the original image; The particle segmentation unit is used to identify the outline of stone particles using a deep learning image segmentation model or a morphological edge detection algorithm. The particle size calculation unit is used to calculate the equivalent particle size based on the particle profile and to statistically generate gradation data.

[0114] In one specific embodiment, to implement the method described in any of the above, a transition material blasting particle size control system based on real-time image recognition of blasted piles was constructed. This system consists of an image acquisition module, a particle size analysis module, a deviation determination module, a parameter adjustment module, and a feedback control module.

[0115] The image acquisition module includes a fixed camera array deployed around the blast area and mobile drone-borne cameras. These cameras are equipped with wireless communication units, using 5G or Wi-Fi 6 communication protocols to transmit multi-view image data in real time to a remote server or edge computing node. The fixed camera array can cover a fixed area of ​​the blast pile, while the drone can flexibly adjust its flight path to acquire images from different angles as needed. The combination of these two systems achieves multi-view coverage of the entire blast pile area.

[0116] The particle size analysis module further includes a preprocessing unit, a particle segmentation unit, and a particle size calculation unit. The preprocessing unit performs adaptive median filtering for noise reduction, contrast-limiting adaptive histogram equalization to enhance contrast, and perspective transformation geometric correction based on camera calibration on the received raw image. The particle segmentation unit uses a pre-trained deep learning image segmentation model, such as an improved U-Net architecture, to identify the contours of stone particles and can perform cross-view matching and fusion to generate 3D contour information. Alternatively, morphological edge detection algorithms can be used for segmentation. The particle size calculation unit fits the minimum bounding geometry of the particle contour, uses the feature size as the equivalent particle size, and generates gradation data according to a preset grading standard. The deviation judgment module compares the generated gradation data item by item with the preset transition material target gradation range. The preset range includes a maximum particle size of no more than 300 mm, a particle content of 15% to 25% smaller than 5 mm, a particle content of no more than 5% smaller than 0.075 mm, a non-uniformity coefficient greater than 5, and a curvature coefficient between 1 and 3, determining whether any indicators exceed the allowable deviation.

[0117] The parameter adjustment module has a built-in parameter-gradation mapping model trained with historical data. When the deviation judgment module detects an out-of-tolerance index, this module inputs the gradation deviation into the model and calculates and generates the hole mesh parameter adjustment value and / or charge parameter adjustment value for the next round of blasting operations. The adjustment rules provide specific adjustment directions for situations such as excessive ultra-large particle size or abnormal fine particle content.

[0118] The feedback control module outputs the adjusted blasting parameters to the blasting operation execution terminal to guide the next round of blasting design and construction, completing a closed-loop control cycle. The entire system can repeat the above process until the blasted rock aggregate gradation stably meets the design requirements.

[0119] This embodiment ensures comprehensive data acquisition and real-time transmission through the multi-view camera configuration and high-speed wireless communication of the image acquisition module. The preprocessing, segmentation, and calculation units built into the particle size analysis module modularize complex image processing tasks, facilitating maintenance and upgrades. The separate design of the deviation judgment module and parameter adjustment module clarifies the responsibilities of quality judgment and decision optimization, allowing the model to be optimized independently. The feedback control module realizes closed-loop output of control commands, enabling the system to truly possess automatic control capabilities. The entire system adopts a modular design, with each module capable of working collaboratively or being upgraded and replaced independently. This facilitates customized configuration according to different engineering needs, improving the system's flexibility and scalability, and laying the foundation for large-scale application.

[0120] According to one embodiment of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0121] In one specific embodiment of the present invention, a computer-readable storage medium is provided to implement any of the methods described above. This storage medium may be in various forms such as a USB flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk. A computer program is stored on the storage medium, which includes instruction code capable of executing the methods.

[0122] When the storage medium is connected to a computer, server, edge computing node, or embedded device, the processor reads and executes the computer program contained therein. This enables a series of operations, including automatic reception of images acquired from a burst pile, image preprocessing, particle contour recognition, equivalent particle size calculation, gradation statistics, comparison with the target gradation, deviation determination, generation of parameter adjustment values ​​based on a preset mapping relationship, and output of adjustment commands. Operators only need to insert the storage medium into the corresponding computing device to enable that device to implement the method of this invention, without the need for redeveloping software.

[0123] For example, in a newly constructed engineering project, the construction unit only needs to purchase a storage medium containing the computer program of this invention, insert it into the edge computing server on site, and connect it to hardware devices such as cameras to quickly deploy a complete intelligent control system for the blasting particle size of transition material, which greatly shortens the system development and debugging cycle.

[0124] This embodiment utilizes a standardized storage medium, enabling the technical solution of this invention to be copied, distributed, and installed like ordinary software, greatly improving the dissemination and ease of use of the technology. Users do not need professional software development capabilities; they only need to insert the storage medium and run the program to enable existing computing devices to possess all the functions described in this invention, reducing the technical threshold and labor costs for technology application. The computer program is independent of specific hardware and can run on various general-purpose computing platforms, enhancing system compatibility and flexibility. The storage medium can be mass-produced, facilitating the rapid promotion and application of the technical solution of this invention in multiple engineering projects, and promoting technological progress and standardization throughout the industry.

[0125] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A method for controlling the particle size of transition material explosion based on real-time image recognition of blasted piles, characterized in that, Includes the following steps: 1) After each round of blasting is completed, multi-view digital images of the blast pile formed by blasting are collected. The collection of multi-view digital images of the blast pile includes: collecting multi-view image data covering the entire blast pile area by using a fixed camera array and / or a mobile UAV-borne camera deployed around the blast area; and transmitting the collected multi-view image data to a remote server or edge computing node in real time through a wireless communication network. 2) Based on the multi-view digital images, identify the outlines of the stone chips in the blast pile, calculate the equivalent particle size of each particle according to the identified outlines, and then statistically generate gradation data reflecting the particle size distribution of the stone chips in the current round of blasting. 3) Compare the generated gradation data with the preset transition material target gradation range to determine whether there are any gradation indicators that exceed the allowable deviation; 4) When it is determined that there is a gradation index that exceeds the allowable deviation, the hole mesh parameter adjustment value and / or charge parameter adjustment value are generated for the next round of blasting operation according to the preset parameter-gradation mapping relationship. 5) Carry out the next round of blasting operations according to the adjusted blasting parameters; Repeat steps 1) to 5) until the blasting slag gradation meets the design requirements.

2. The method for controlling the particle size of transition material explosion based on real-time image recognition of the explosion pile as described in claim 1, characterized in that, In step 2), the outline of the stone chips is identified, including: The acquired multi-view images are input into a pre-trained deep learning image segmentation model, which outputs a pixel-level segmentation mask for the stone particles in each image. Cross-view matching and fusion of segmentation masks for the same particle in images from different viewpoints are performed to generate the three-dimensional contour information of each particle.

3. The method for controlling the blasting particle size of transition material based on real-time image recognition of blasting piles as described in claim 2, characterized in that, In step 2), the equivalent particle size of each particle is calculated, including: Based on the three-dimensional contour information of each particle, the minimum bounding geometry of the particle is fitted, and the characteristic size of the bounding geometry is used as the equivalent particle size. According to the preset particle size classification standard, the equivalent particle size of all particles is classified into the corresponding particle size range, and the number of particles or cumulative mass of each particle size range is counted to generate gradation data.

4. The method for controlling the blasting particle size of transition material based on real-time image recognition of blasting piles as described in claim 2, characterized in that, Step 2) includes an image preprocessing step before identifying the contours of the stone chips: The acquired raw images are subjected to denoising, contrast enhancement, and geometric correction.

5. The method for controlling the particle size of transition material explosion based on real-time image recognition of the explosion pile as described in claim 1, characterized in that, In step 3), the preset target gradation range for the transition material includes: The maximum particle size is no more than 300 mm, the content of particles smaller than 5 mm is 15% to 25%, the content of particles smaller than 0.075 mm is no more than 5%, and the uniformity coefficient Cu > 5, and the curvature coefficient Cc is between 1 and 3.

6. The method for controlling the blasting particle size of transition material based on real-time image recognition of blasting piles as described in claim 1, characterized in that, In step 4), the preset parameter-gradation mapping relationship is established in advance in the following way: Collect blasting parameters and corresponding blasting debris gradation data from each blasting test to construct a training dataset; The training dataset is trained using regression analysis or machine learning algorithms to establish a quantitative mapping relationship model between blasting parameters and gradation characteristics; The process of generating adjustment values ​​includes: inputting the deviation between the current round gradation data and the target gradation into the mapping relationship model, and calculating the corrected blasting parameters.

7. The method for controlling the blasting particle size of transition material based on real-time image recognition of blasting piles as described in claim 6, characterized in that, The blasting parameters include at least one of explosive consumption per unit volume, hole spacing, row spacing, and packing length; the generated adjustment values ​​include: When the proportion of oversized stone chips exceeds the standard, an adjustment command is generated to reduce the mesh parameters or increase the explosive consumption. When the fine particle content is detected to be too low, an adjustment command is generated to increase the explosive consumption or decrease the mesh parameters. When an excessively high content of fine particles is detected, an adjustment command is generated to reduce the explosive consumption or increase the mesh parameters.

8. The method for controlling the particle size of transition material blasting based on real-time image recognition of blasting piles as described in claim 1, characterized in that, In step 1), the wireless communication network adopts the 5G or Wi-Fi 6 communication protocol and ensures the complete transmission of image data by resuming interrupted transmission.

9. A transition material blasting particle size control system based on real-time image recognition of blasted piles, used to execute the method according to any one of claims 1 to 8, characterized in that, include: The image acquisition module is used to acquire multi-view digital images of the blast pile after each round of blasting operations. The particle size analysis module is used to identify the outline of stone particles in the blast pile based on the multi-view digital images, calculate the equivalent particle size, and generate gradation data. The deviation determination module is used to compare the generated gradation data with the preset transition material target gradation range to determine whether there are gradation indicators that exceed the allowable deviation. The parameter adjustment module is used to generate hole mesh parameter adjustment values ​​and / or charge parameter adjustment values ​​for the next round of blasting operations based on the preset parameter-gradation mapping relationship when it is determined that there are gradation indicators that exceed the allowable deviation. The feedback control module is used to output the adjusted blasting parameters to the blasting operation execution end to guide the next round of blasting operations.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method as claimed in any one of claims 1 to 8.