A real-time detection method for soil particle gradation characteristics

Abstract

The present invention relates to the field of intelligent detection technology for building materials, and in particular to a real-time detection method for gradation characteristics of slag particles. The method comprises the following steps: S1. acquisition of original point cloud data; S2. clustering of particle point clouds; S3. determination of particle size after screening; S4. extraction of particle morphological characteristics; S5. correction of particle quality; S6. output of gradation characteristics. The present method has the characteristics of high precision, high efficiency, low cost, and wide applicability in particle gradation characteristic detection, and is particularly suitable for gradation characteristic detection of residual mud and slag particles. It can also be applied to real-time detection requirements involving granular materials in other fields, such as mineral particle screening, construction sand and gravel detection, agricultural granular fertilizer grading, etc. It can realize rapid and accurate detection of particle gradation characteristics in different industries, and meet the needs of various industrial production and engineering applications.

Description

Technical Field

[0001] The invention relates to the technical field of intelligent detection of building materials, and in particular to a real-time detection method for gradation characteristics of slag particles. Background Art

[0002] The gradation characteristics of particulate matter play a key role in the performance of various engineering materials. Reasonable particle gradation can significantly improve the mechanical properties, durability, impermeability and construction operability of the material, thereby improving the overall performance and economic benefits of the material. However, traditional particle detection methods (such as screening methods) have problems such as limited sampling range, low detection efficiency, and complex operation, which are difficult to meet the needs of modern engineering for rapid and accurate detection.

[0003] In recent years, with the development of sensor technology and computer vision technology, CT scanning, digital image processing and 3D point cloud technology have begun to be applied to particle grading feature detection. However, these technologies still have the following shortcomings in practical applications:

[0004] CT scanning: Although it can obtain relatively accurate three-dimensional structural information, it is expensive, slow to detect, and is not suitable for large-scale real-time detection.

[0005] Digital image processing: mainly based on the analysis of two-dimensional images, it cannot accurately reflect the true shape of three-dimensional particles, resulting in errors between the detection results and the actual results, especially when detecting particles with complex shapes.

[0006] 3D point cloud technology: Although 3D point cloud technology can provide 3D information of particles and is considered an effective means of particle detection, the existing 3D point cloud processing algorithms still have obvious deficiencies in segmentation efficiency, particle size definition accuracy, and particle mass calculation, making it difficult to achieve ideal accuracy and speed in real-time detection. The specific problems are as follows:

[0007] ① The segmentation algorithm is not efficient enough. Most existing point cloud segmentation algorithms rely on density clustering or geometric feature segmentation methods, but these methods often require a lot of computing resources when processing high-density or complex-shaped point cloud data, resulting in low segmentation efficiency. For example, the density clustering method may generate more redundant calculations in high-density areas, while the geometric feature law requires complex parameter debugging and high algorithm complexity. Therefore, it is difficult for existing segmentation algorithms to achieve fast and accurate segmentation effects when dealing with the real-time detection requirements of a large number of particles.

[0008] ② The particle size definition is not accurate. In terms of particle size definition, although three-dimensional data is used, existing methods generally rely on two-dimensional plane projection to calculate the particle size, and usually use the width of the best fit rectangle to approximate the particle screening size. This method ignores the spatial advantages of three-dimensional data and fails to fully reflect the true three-dimensional structure of the particles, resulting in a large deviation between the actual test results and the true particle size, which is a waste of three-dimensional data resources.

[0009] ③ Inaccurate particle mass calculation. In terms of particle mass calculation, existing methods often ignore the integrity of three-dimensional data, usually using plane area as an approximation of volume or estimating particle volume through partial surface reconstruction. Since this method fails to utilize the complete information of the three-dimensional point cloud, it often leads to significant errors in particle samples with complex morphology. Therefore, existing mass calculation methods have great limitations in particle volume and density estimation, and it is difficult to accurately output the mass distribution of particles.

[0010] In view of the above problems, the existing 3D point cloud processing technology is still difficult to achieve efficient and accurate real-time detection of particle gradation characteristics. Therefore, there is an urgent need for a 3D point cloud detection method that has higher accuracy and applicability in segmentation efficiency, particle size definition and mass calculation to meet the needs of real-time detection of particle gradation on engineering sites. Summary of the invention

[0011] The present invention provides a real-time detection method for particle gradation characteristics of slag, aiming to solve the technical defects existing in the application of existing sensing technology and computer vision technology in particle gradation characteristic detection.

[0012] The present invention provides a method for real-time detection of soil particle gradation characteristics, comprising the following steps:

[0013] S1. Original point cloud data acquisition: The 3D camera scans the soil particles, collects and obtains the original point cloud data of the soil particles;

[0014] S2. Particle point cloud clustering: The original point cloud data is identified and segmented into particles of different densities and shapes through the particle clustering segmentation algorithm;

[0015] S3. Determine the particle size after particle screening: Calculate the oriented bounding box of each particle through the oriented bounding box algorithm in the three-dimensional point cloud data, and determine the particle screening size through the minimum control cross section of the oriented bounding box;

[0016] S4. Particle morphology feature extraction: The overall outline of the particle is extracted through Alpha shape reconstruction, and the missing point cloud data at the bottom is filled by uniform sampling. A closed three-dimensional grid is generated using Poisson surface reconstruction. Based on the reconstructed watertight three-dimensional grid, a complete and stable particle volume is obtained using the spatial integration method. The morphological parameters of the particle are obtained based on the particle volume.

[0017] S5. Particle mass correction: The PMM model identifies the particle type through the point cloud projection image, combines the morphological parameters and density parameters, corrects the particle mass / volume, and generates a mass distribution curve;

[0018] S6. Grading characteristic output: Based on the mass / volume of each particle and the mass / volume of all particles to be detected, the particle grading characteristics are calculated.

[0019] As a further improvement of the present invention, in step S2, the specific steps of the particle clustering segmentation algorithm include:

[0020] Input the original point cloud data, calculate the average point spacing factor ASF of the point cloud data, and compare the calculated ASF value with the optimal ASF value k. If the ASF value of the input original point cloud data is less than k, the point cloud is uniformly downsampled so that the ASF value is greater than k, and then the point cloud is density clustered to divide the point cloud into single particle point clouds; if the ASF value of the input original point cloud data is greater than k, the point cloud is directly density clustered to obtain the segmented point cloud data.

[0021] As a further improvement of the present invention, in step S2, the calculation process of the average point spacing factor ASF is:

[0022] ,

[0023] in, Represents a point in the point cloud p With any other point q The distance between the points is contained in the point cloud N points.

[0024] As a further improvement of the present invention, in step S3, the particle screening size calculation process includes:

[0025] Screening size d i Calculation:

[0026] ,

[0027] in, W Indicates the width of the oriented bounding box, H Indicates the height of the oriented bounding box;

[0028] Volume of the oriented bounding box V OBB The calculation formula is:

[0029] ,

[0030] in, W Indicates the width of the oriented bounding box, H Represents the height of the oriented bounding box, L Indicates the length of the oriented bounding box;

[0031] Screen size D Defined as:

[0032] ,

[0033] in S is the side length of the sieve hole;

[0034] The screening judgment rule is: if the diagonal length of the minimum control section of the particle oriented bounding box is d i Less than or equal to the diagonal length of the sieve hole D , then the particle is considered to be able to pass through the sieve hole, that is, it satisfies:

[0035] .

[0036] As a further improvement of the present invention, in step S4, the overall contour of the particle is extracted by Alpha shape reconstruction, specifically comprising:

[0037] Get the original point cloud data of the particles and define a complexity parameter a , to determine whether the distance between two discrete points in the original point cloud data is less than a , if less than a Then two discrete points are connected to form an edge, and an external boundary close to the actual shape of the particle is generated through all discrete points to form the overall outline of the particle;

[0038] The missing point cloud data at the bottom is filled by uniform sampling, including:

[0039] On the particle arrangement contour after Alpha shape reconstruction, uniform point cloud sampling is performed on the mesh surface of the particle arrangement contour: surface information is extracted from the mesh reconstructed by Alpha shape, data points are generated at equal distances on the mesh surface, points inside the mesh or discontinuous parts of the boundary are ignored, and the external form of the object is focused on; finally, continuous and complete point cloud data is generated on the mesh surface, and the mesh reconstructed by Alpha shape is deleted to obtain a continuous and complete new particle point cloud data;

[0040] Generate a closed 3D mesh using Poisson surface reconstruction, including:

[0041] On the point cloud data obtained by uniform sampling, the potential surface shape is inferred through the normal vector field of the point cloud, a smooth closed surface is generated from the point cloud, the gaps between the sampling points are filled, and a stable three-dimensional grid is formed.

[0042] As a further improvement of the present invention, in step S4, obtaining the morphological parameters of the particles based on the particle volume specifically includes:

[0043] After obtaining the volume of particle reconstruction V R After that, the morphological parameters l Obtained by the following formula:

[0044] ,

[0045] in V OBB is the volume of the oriented bounding box.

[0046] As a further improvement of the present invention, in step S5, the PMM model identifies the particle type through the point cloud projection image, including:

[0047] The PMM model uses a CNN network to quickly identify particle types and automatically match the morphological parameters preset in the database. l and density parameters r , the specific process is:

[0048] Establish a point cloud projection image database, select the projection of the particle point cloud in the direction of the maximum bottom surface of the oriented bounding box; divide the image data set into a training set and a validation set, set the training step size, training rounds, and CNN variants, and then compare the model performance and accuracy after training multiple variants to obtain the optimal image recognition model; according to the morphological parameters of different types of particles pre-determined and stored in the database l and density parameters r , apply the optimal image recognition model to identify the particle type and call the morphological parameters of the corresponding type of particles l and density parameters r .

[0049] As a further improvement of the present invention, in step S5, the PMM model is combined with the morphological parameters l and density parameters r , the calculation process for correcting the particle mass / volume is:

[0050] ,

[0051] ,

[0052] in, M i represents the estimated mass of the particle, V i represents the estimated volume of the particle, V OBB Represents the volume of an oriented bounding box.

[0053] As a further improvement of the present invention, step S6 specifically includes:

[0054] In the grading calculation, X i It indicates the percentage or proportion of particle size distribution during the screening process, based on the mass / volume of each particle ( M i / V i ) and the mass / volume of all particles to be detected ( M / V ) and the particle size distribution characteristics are calculated by the following formula:

[0055] (1)

[0056] (2)

[0057] Therefore, formula (1) expresses that the sieve size is d i When the particle size is smaller than the sieve size, the percentage of the total mass of the particles; if the particles to be tested are of the same type, that is, the density of the particles r , morphological characteristics l If they are consistent, the gradation characteristics can be directly obtained by formula (2); using the oriented bounding box of the particle V OBB The volume is taken as the equivalent volume of the particle.

[0058] The beneficial effects of the present invention are as follows: the present method has the characteristics of high precision, high efficiency, low cost, and wide applicability in particle grading feature detection, and is particularly suitable for grading feature detection of residual mud and slag particles. It can also be applied to real-time detection needs involving particulate matter in other fields, such as mineral particle screening, construction sand and gravel detection, agricultural granular fertilizer grading, etc. It can realize rapid and accurate detection of particle grading characteristics in different industries and meet the needs of various industrial production and engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flowchart of the real-time detection method of the particle gradation characteristics of slag in the present invention;

[0060] Figure 2 It is a schematic diagram of the overall structure of the real-time detection method of the particle gradation characteristics of slag in the present invention;

[0061] Figure 3 It is a schematic diagram of obtaining particle point cloud data in the present invention;

[0062] Figure 4It is a technical detail diagram of the CNN image recognition model in the present invention;

[0063] Figure 5 It is a schematic diagram of an application scenario of the real-time detection method of slag particle gradation characteristics of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0065] like Figure 1 and Figure 2 As shown, a real-time detection method for particle gradation characteristics of slag of the present invention is specifically a real-time particle gradation characteristic detection algorithm based on three-dimensional point cloud data, which realizes automatic segmentation of particles, characteristic screening size extraction and mass calculation, and gradation characteristic output.

[0066] The specific real-time detection method of the particle gradation characteristics of the slag soil comprises the following steps:

[0067] S1. Original point cloud data acquisition: The 3D camera scans the slag particles, collects and acquires the original point cloud data of the slag particles.

[0068] S2. Particle point cloud clustering: The original point cloud data is identified and segmented into particles of different densities and shapes through the particle clustering segmentation algorithm.

[0069] Particle Clustering Segmentation Algorithm (PPC): This algorithm is an optimization of the density-based clustering (DBSCAN) method, which realizes the fast and effective segmentation of particles in 3D point cloud data. The advantage and innovation of the PPC clustering algorithm is that it optimizes the segmentation efficiency and accuracy by dynamically adjusting the average point spacing factor (ASF) of the particle point cloud data, so that particles of different densities and shapes can be accurately identified and quickly segmented.

[0070] Specifically, if Figure 2 As shown in (a) in the figure, the specific steps of the PPC clustering algorithm are as follows: input the original point cloud data, the program first calculates the average point spacing factor (ASF) of the point cloud, and then compares the calculated ASF with the optimal ASF value (k for short). If the ASF value is less than k, the point cloud is uniformly downsampled so that the ASF is greater than k, and then the point cloud is subjected to DBSCAN clustering calculation to segment the point cloud into individual particle point clouds. In particular, if the ASF value of the input original point cloud data is greater than k, the algorithm directly performs DBSCAN clustering operation on it to obtain the segmented point cloud data. k is obtained through experiments on the influence of ASF on the particle size distribution curve, and k=0.15 is preferably selected. This k value can take into account both the clustering operation speed and the accuracy of the point cloud segmentation results. The PPC clustering algorithm can greatly improve the efficiency of point cloud clustering segmentation.

[0071] The calculation process of the average point spacing factor ASF is:

[0072]

[0073] in, Represents a point in the point cloud p With any other point q The distance between the points is contained in the point cloud N points. Different 3D cameras have different ASFs in their output point clouds due to differences in built-in hardware and software parameter settings. The ASF value can be freely adjusted by simply upsampling or downsampling the point cloud data. A large number of experiments have shown that there is an optimal ASF value that can balance clustering segmentation speed and segmentation accuracy. The optimal ASF value is represented by the letter k.

[0074] S3. Determination of particle size after particle screening: The oriented bounding box algorithm in the three-dimensional point cloud data is used to calculate the oriented bounding box of each particle, and the particle screening size is determined by the minimum control cross section of the oriented bounding box.

[0075] like Figure 2 As shown in (b) of FIG. 1 , the particle screening size determination method (PSSD) of the present invention uses an oriented bounding box (OBB) algorithm in three-dimensional point cloud data to accurately calculate the oriented bounding box of each particle. Specifically, the PSSD method uses the minimum cross section (i.e., the minimum control cross section) of the oriented bounding box to determine the screening size of the particle, thereby accurately reflecting the actual three-dimensional shape of the particle.

[0076] The particle size calculation formula of the PSSD method is as follows:

[0077] Screening size d i Calculation:

[0078] ,

[0079] in, W Indicates the width of the oriented bounding box, H Indicates the height of the oriented bounding box.

[0080] In addition, the volume of the oriented bounding box V OBB The calculation formula is:

[0081] ,

[0082] in, W Indicates the width of the oriented bounding box, H Represents the height of the oriented bounding box,L Indicates the length of the oriented bounding box.

[0083] In order to accurately determine whether a particle passes through a specific sieve hole, the sieve hole size D Defined as:

[0084] ,

[0085] in S is the side length of the sieve hole. The sieve is composed of small square sieve holes of the same size, so the two S The length of the side of the sieve hole is the same as the diagonal length of the sieve hole. D Screening size d i (which is also the diagonal line of the minimum cross section of the oriented bounding box) is used to determine whether the particles can pass through the sieve holes, thereby obtaining the grading result that is closest to the actual screening.

[0086] According to the above formula, the screening judgment rule is: if the diagonal length of the minimum control section of the particle oriented bounding box is d i Less than or equal to the diagonal length of the sieve hole D , then the particle is considered to be able to pass through the sieve hole, that is, it satisfies:

[0087] .

[0088] At this time, the program will determine that the actual screening particle size of the particle is less than S gear, instead of D . formula This is an intermediate process of judgment. If the formula is satisfied, the particle size can be classified as smaller than D Corresponding to the side length of the sieve hole S This gear means that the particles can pass through the sieve hole.

[0089] Through this method, PSSD is more accurate than traditional methods in determining particle screening size because it fully considers the three-dimensional structure of the particles rather than just based on two-dimensional plane projection. Traditional methods (image method and existing three-dimensional point cloud technology) usually use the minimum width of the two-dimensional image to approximate the particle screening size. Such simplification will lead to errors, especially when the particle shape is complex or irregular. The PSSD method constructs a realistic particle size representation through the spatial data of the three-dimensional point cloud, thereby significantly improving the accuracy of particle classification and size distribution measurement.

[0090] In order to achieve efficient and accurate particle mass distribution detection in mixed samples, this paper proposes a particle morphology feature extraction and mass correction model (PCE-PMM). l value) and density parameter ( r The model uses the APR algorithm (Alpha-Poisson Reconstruction) to perform three-dimensional morphological reconstruction to support the accurate extraction of morphological parameters.

[0091] S4. Particle morphological feature extraction: The overall outline of the particle is extracted through Alpha shape reconstruction, and the missing point cloud data at the bottom is filled by uniform sampling. A closed three-dimensional grid is generated using Poisson surface reconstruction. Based on the reconstructed watertight three-dimensional grid, the complete and stable particle volume is obtained using the spatial integration method. The morphological parameters of the particle are obtained based on the particle volume.

[0092] Particle morphology feature extraction (PCE method): The core of the PCE method lies in the morphology parameters l The morphology of different particles (such as shape and density) will affect the volume and mass calculation. The existing methods introduce errors because they fail to fully consider the morphological differences. The PCE method uses the APR algorithm to reconstruct the three-dimensional structure of particles to obtain reliable morphological parameters. The specific steps are as follows:

[0093] APR algorithm reconstruction process: The APR algorithm combines Alpha shape reconstruction (ASR), uniform sampling and Poisson surface reconstruction (PSR) to achieve efficient reconstruction of the three-dimensional volume of particles. First, the overall outline of the particle is extracted through Alpha shape reconstruction, and the missing point cloud data at the bottom is filled by uniform sampling. Poisson surface reconstruction is used to generate a closed three-dimensional grid. Finally, based on the watertight three-dimensional grid reconstructed by the APR algorithm, the complete and stable particle volume can be obtained using the spatial integration method. V R The unique advantage of the APR algorithm is that it can obtain a stable three-dimensional mesh even when the bottom data of the point cloud is missing.

[0094] A watertight mesh is a closed geometry with no gaps or holes. For a 3D model, if its surface (mesh) has holes or disconnected parts, the computer cannot determine the inside and outside of the model, and therefore cannot accurately calculate the volume, surface area or other spatial integrals.

[0095] Since volume calculation usually relies on spatial integration methods, if there are holes in the mesh model (for example, the connection between triangles is incomplete), the integral formula will not converge due to the lack of necessary boundary conditions. This may lead to calculation errors or program crashes. The APR algorithm is responsible for reconstructing a watertight mesh. Only when the watertight mesh is reconstructed can the computer use the spatial integration method to calculate the volume. If the reconstructed mesh has holes, the computer integration process will not converge at this time, and the program will report an error.

[0096] like Figure 2 As shown in (c) in the figure, the APR algorithm has three main steps:

[0097] S41. Alpha Shape Reconstruction (ASR): Extract the overall outline of the particle through Alpha shape reconstruction, including:

[0098] Get the original point cloud data of the particles and define a complexity parameter a , to determine whether the distance between two discrete points in the original point cloud data is less than a , if less than a Then two discrete points are connected to form an edge, and an external boundary close to the actual shape of the particle is generated through all discrete points to form the overall outline of the particle.

[0099] Alpha shape reconstruction is used to extract the outer contour of particles from the original particle point cloud data (single particle point cloud data obtained by PPC clustering, which is characterized by the missing bottom point cloud). When processing particle point clouds, Alpha shape is a geometric model based on point cloud data, which can be defined by a parameter a to control the complexity of the shape. a When is larger, the reconstructed shape will be simpler; when a When the point cloud is smaller, the shape will be more complex and detailed. In this step, the Alpha shape reconstruction method is used to extract the overall outline of the particle, which is the basis for subsequent reconstruction. This method can handle the discreteness of the point cloud and generate an outer boundary close to the actual shape of the particle.

[0100] S42. Uniform sampling: fill the missing point cloud data at the bottom through uniform sampling, including:

[0101] On the particle outline after Alpha shape reconstruction, the reconstructed particle outline is actually a (non-watertight) mesh with small holes. Uniform point cloud sampling is performed on the surface of this mesh. Specifically, surface information is extracted from the mesh reconstructed by the Alpha shape, and data points are generated at equal distances on the mesh surface. The points inside the mesh or the discontinuous parts of the boundary are ignored, and the focus is on the external form of the object. Finally, continuous and complete point cloud data (point cloud is a general term for a large number of data points) is generated on the surface of the mesh. Then, the mesh reconstructed by the Alpha shape is deleted, and a new particle point cloud data with higher quality (reflected in the bottom point cloud that could not be scanned originally is completed, and the density of the sampled point cloud is more uniform) and continuous and complete can be obtained.

[0102] The key advantages of uniform point cloud sampling include:

[0103] (1) Completing the bottom point cloud: For some areas that are difficult to scan, uniform sampling can generate point clouds in these areas, making the entire data set more complete;

[0104] (2) Improve the uniformity of point cloud density: Through uniform sampling, the uneven distribution of point clouds caused by inappropriate scanning angles or positions can be eliminated, making the point cloud more uniform over the entire object surface;

[0105] (3) Improving continuity: Uniform point cloud sampling can ensure the continuity of point cloud data, which is crucial for subsequent mesh reconstruction, morphological analysis, and other tasks.

[0106] S43. Poisson Surface Reconstruction (PSR): Generate a closed 3D mesh using Poisson Surface Reconstruction, including:

[0107] On the point cloud data obtained by uniform sampling, the potential surface shape is inferred through the normal vector field of the point cloud, a smooth closed surface is generated from the point cloud, the gaps between the sampling points are filled, and a stable three-dimensional grid is formed.

[0108] Poisson surface reconstruction is a point cloud-based surface reconstruction method that infers the underlying surface shape through the normal field of the point cloud. The goal of this method is to generate a smooth closed surface from the point cloud, filling the gaps between the sampling points to form a stable three-dimensional mesh. In this step, Poisson surface reconstruction uses the point cloud data obtained by uniform sampling to generate a closed three-dimensional mesh to ensure that the surface of the particles is continuous and complete. Even if the original point cloud data is missing, Poisson reconstruction can still generate a stable three-dimensional mesh.

[0109] Finally, after Alpha shape reconstruction, uniform sampling and Poisson surface reconstruction (collectively known as the APR algorithm), a complete and watertight three-dimensional grid is finally obtained. Then, the three-dimensional volume of the particle can be quickly and stably (reflected in the program without error) calculated through the spatial integration method. V R The grid represents the true three-dimensional shape of the particle and can effectively make up for the missing parts that may exist in the point cloud data, especially when the data at the bottom of the particle is missing.

[0110] The present invention is the first to combine the three algorithms of Alpha Shape Reconstruction (ASR), Uniform Sampling and Poisson Surface Reconstruction (PSR). Figure 3 As shown, the 3D camera is scanning from directly above, and the point cloud data at the bottom of the particle is missing. The 3D camera can only scan and output the point cloud on the top and sides of the particle. The APR algorithm is a surface reconstruction algorithm proposed by the present invention for particles with missing bottom point clouds. Its unique advantage is that even if there is data missing at the bottom or other parts of the point cloud, the algorithm can still generate a stable and accurate three-dimensional grid. This enables the APR algorithm to still ensure the accuracy and reliability of the reconstruction results when processing incomplete or limited point cloud data.

[0111] Morphological parameters l Calculation: After obtaining the particle reconstruction volume V R After that, the morphological parameters l It can be obtained by the following formula:

[0112] ,

[0113] in V OBB is the volume of the oriented bounding box. l ,The PCE method effectively compensates for the volume error caused by the difference in particle morphology and ensures the accuracy of mass estimation.

[0114] S5. Particle mass correction: Figure 2 As shown in (d), the PMM model identifies the particle type through the point cloud projection image, combines the morphological parameters and density parameters, corrects the particle mass / volume, and generates a mass distribution curve.

[0115] Particle mass correction (PMM model):

[0116] Through the classification of point cloud projection images, the PMM model uses a simple CNN network to quickly identify particle types and automatically match the morphological parameters preset in the database. l and density parameters r , ensuring the accuracy and efficiency of quality estimation.

[0117] CNN (Convolutional Neural Network) is a deep learning model that is widely used in image recognition, video analysis, natural language processing and other fields, and is particularly good at processing data with a grid structure (such as images). The present invention proposes to use the projection image of the particle point cloud to identify the particle type, thereby avoiding the high computational complexity of the particle point cloud (three-dimensional data) recognition and processing method. Specifically, Figure 4 As shown, the training steps are as follows:

[0118] First, it is necessary to establish a point cloud projection image database (no less than 100 images of each type), such as pebbles, machine-made sand, gravel, etc. To ensure the effectiveness of the projection image, it is necessary to select the projection image of the particle point cloud in the direction of the maximum bottom surface of the oriented bounding box. It should be noted that the type of sand and gravel is based on the actual application scenario and can also be replaced by other granular materials, such as expanded clay, rice, fruit and other items from other industries.

[0119] Next, the image dataset is divided into two parts, a training set (83.3%) and a validation set (16.7%). The training step size Lr is set to 0.001 and the training round Epoch is set to 60. ResNet-18 / 34 / 50 / 101 / 154 and VGG16 / 19 are all variants of CNN. After training multiple variants, the model performance and accuracy are compared to obtain the optimal image recognition model.

[0120] Finally, the optimal image recognition model is used to identify the particle type, and different types of results are obtained. The morphological parameters of different types of particles l and density parameters r It is pre-determined and stored in the database, so once the CNN model identifies the particle type, it can quickly call the correct morphological parameters l and density parameters r .

[0121] In obtaining the morphological parameters l and particle type, the PMM model uses the density parameter r Further corrections are made to the particle mass. Density parameters r It is obtained through actual measurement and stored in the database as known parameters of the PMM model, which is convenient for PMM model call. The PMM model quickly identifies the particle type through point cloud projection images and combines morphological parameters l and density parameters r , generating an accurate mass distribution curve. The specific mass estimation / volume formula is as follows:

[0122] ,

[0123] ,

[0124] in, M i represents the estimated mass of the particle, V i represents the estimated volume of the particle, V OBB Represents the volume of the oriented bounding box. The PMM model is applicable to granular materials of different densities and shapes, and can quickly generate mass distribution data, especially for complex mixed particle samples.

[0125] S6. Grading characteristic output: Based on the mass / volume of each particle and the mass / volume of all particles to be detected, the particle grading characteristics are calculated.

[0126] In the grading calculation, X i It indicates the percentage or proportion of particle gradation during the screening process. Specifically, X i It is the percentage of particles smaller than a certain sieve size in the total mass or volume. The sieve size is obtained by the particle sieve size determination method (PSSD). Finally, the mass / volume of each particle obtained by the program ( M i / V i ) and the mass / volume of all particles to be detected ( M / V ), the particle grading characteristics can be calculated by the following formula:

[0127] (1)

[0128] (2)

[0129] Therefore, formula (1) expresses that the sieve size is d i When the particle size is smaller than the sieve size, the total mass of the particles is the percentage of the total mass. In particular, if the particles to be tested are of the same type, that is, the density characteristics of the particles are r , morphological characteristics lIf the values ​​of the particle size distribution of the particle are consistent, the grading characteristics can be directly obtained by using formula (2). Formula (1) can be used to calculate the particle size distribution of different types of particles and the same type of particles. In the special case of the same type of particles, the process can be simplified and the calculation can be directly performed by using formula (2). In order to further reduce the computational complexity of the program and improve the processing speed, this algorithm does not need to calculate the exact particle volume, but uses the particle oriented bounding box OBB volume as the equivalent volume of the particle. Compared with the method of reconstructing the particle point cloud grid, obtaining the watertight body and then calculating the particle volume, this algorithm has a great speed advantage.

[0130] With the support of the APR algorithm, the PCE-PMM model has the following significant advantages in particle volume and mass calculation:

[0131] High precision: morphological parameters l The introduction of effectively corrects the morphological error and makes the quality estimation more accurate.

[0132] High efficiency: Combined with two-dimensional projection operations, the calculation speed is greatly improved to meet real-time detection needs.

[0133] Wide applicability: The model is applicable to various types of particles with complex shapes and is suitable for multi-field applications in engineering, mineral screening, construction sand and gravel, and agricultural particles.

[0134] In summary, the PCE-PMM model uses the APR algorithm to achieve efficient morphological feature extraction and mass correction of particles, which is particularly suitable for generating high-precision mass distribution curves in real-time detection of complex particle samples.

[0135] The real-time detection method of soil particle gradation characteristics of the present invention innovatively combines the particle clustering segmentation algorithm (PPC), the particle screening size determination method (PSSD), and the morphological parameter extraction and quality correction model (PCE-PMM), specifically:

[0136] (1) Particle Clustering Segmentation Algorithm (PPC): This algorithm is an innovative optimization of the traditional density clustering method (DBSCAN) and is specifically designed to achieve fast and accurate segmentation of particles in three-dimensional point cloud data. The PPC algorithm innovatively introduces a dynamically adjusted average point spacing factor (ASF), which greatly optimizes segmentation efficiency and accuracy. Through ASF adjustment, the PPC algorithm can accurately identify and segment particles of different densities and shapes, adapting to the real-time detection needs of complex particle samples.

[0137] (2) Particle Screening Size Determination Method (PSSD): The Oriented Bounding Box (OBB) algorithm is used to accurately calculate the minimum bounding box of the particle, and its minimum control cross section is used to determine the particle screening size. The innovation of the PSSD method is that it uses the complete three-dimensional information of the three-dimensional point cloud data to reflect the true shape of the particles, significantly improving the accuracy of the screening size, making it highly consistent with the actual screening particle size, and overcoming the traditional method's reliance on two-dimensional plane features.

[0138] (3) Morphological Parameter Extraction and Quality Correction Model (PCE-PMM): The PCE-PMM model extracts morphological parameters by reconstructing a small-scale sample point cloud ( l The PCE method uses the APR algorithm (Alpha-Poisson Reconstruction) to effectively reconstruct the three-dimensional shape of particles, and even in the absence of bottom point clouds, it can still obtain a stable three-dimensional grid and accurate volume estimation. The PMM model combines point cloud projection images with the CNN network to quickly identify different types of particles, and applies a quality correction model to ensure that the test results of the mixed sample gradation characteristics are highly consistent with the actual situation.

[0139] like Figure 5 As shown, the schematic diagram of the application scenario of the present invention shows the process of uniform transportation of slag particles on a conveyor belt. The 3D camera synchronously scans the point cloud data of the slag particles in real time, and transmits the collected point cloud data to the algorithm for rapid processing. Through the real-time detection algorithm proposed in the present invention, the point cloud data is efficiently calculated and the grading characteristic results of the particles are output. The result can accurately reflect the particle size distribution and other key characteristics of the slag particles, provide a scientific basis for the subsequent selection of slag resource utilization paths and material optimization solutions, and then optimize the slag recovery and reuse process and improve resource utilization efficiency.

[0140] In summary, the present invention has the following significant advantages:

[0141] High precision: The present invention uses the oriented bounding box (OBB) algorithm in the three-dimensional point cloud data to accurately calculate the minimum cross-section of the particles to determine the screening size, thereby achieving high-precision particle screening. Test results show that the OBB-based particle screening size determination method (PSSD) has an accuracy of 98.28% to 99.16% on different particle samples, which is significantly better than the traditional two-dimensional image method, whose accuracy is usually less than 75%. Compared with the time-consuming traditional screening method, the average absolute error of the grading curve obtained by the present invention is reduced by more than 1.2%, greatly improving the accuracy of particle classification and size distribution measurement.

[0142] High efficiency: In the application scenario of real-time detection, the detection efficiency of the present invention far exceeds that of traditional screening methods. Traditional screening methods usually take at least 10 minutes to complete the detection of a batch of particles, while the present invention uses the PCE-PMM model and APR algorithm to process the same sample in only about 2.59 to 3 seconds, with an efficiency improvement of more than 99.6%. Among them, the PPC clustering algorithm shortens the clustering time to 1.87 seconds through the optimized DBSCAN segmentation method, making it suitable for rapid detection of high-density, large-volume particles.

[0143] Low cost: This method avoids the reliance on expensive equipment such as CT scanners, and does not require high-precision laser scanning equipment. It only requires ordinary 3D cameras (such as high-definition industrial cameras) and mid-range computing devices (such as Intel i7 processors and NVIDIA RTX 3060 graphics cards) to complete the inspection task, which significantly reduces the equipment investment cost and is suitable for engineering sites and inspection scenarios with limited resources. In addition, this method has low hardware requirements during operation and can be flexibly deployed on a variety of common devices.

[0144] Wide applicability: This invention is not only applicable to the detection of residual mud and slag particles in the construction industry, but also to multiple fields such as mineral particle screening, soil composition analysis, and chemical particle classification. Through the adaptation of the morphological parameter extraction and mass correction model (PCE-PMM) to the particle type and density, this method shows high applicability in dealing with scenes with different shapes, densities, and complex mixed particles. Test results show that the accuracy and efficiency of particle detection in a variety of application fields have been significantly improved, making it a universal solution for particle detection in multiple fields.

[0145] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A real-time detection method for soil particle gradation characteristics, characterized in that: The following steps are involved: S1. Original point cloud data acquisition: The 3D camera scans the soil particles, collects and obtains the original point cloud data of the soil particles; S2. Particle point cloud clustering: The original point cloud data is identified and segmented into particles of different densities and shapes through the particle clustering segmentation algorithm; S3. Determine the particle size after particle screening: Calculate the oriented bounding box of each particle through the oriented bounding box algorithm in the three-dimensional point cloud data, and determine the particle screening size through the minimum control cross section of the oriented bounding box; S4. Particle morphology feature extraction: The overall outline of the particle is extracted through Alpha shape reconstruction, and the missing point cloud data at the bottom is filled by uniform sampling. A closed three-dimensional grid is generated using Poisson surface reconstruction. Based on the reconstructed watertight three-dimensional grid, a complete and stable particle volume is obtained using the spatial integration method. The morphological parameters of the particle are obtained based on the particle volume. S5. Particle mass correction: The PMM model identifies the particle type through the point cloud projection image, combines the morphological parameters and density parameters, corrects the particle mass or volume, and generates a mass distribution curve; S6. Grading characteristic output: based on the mass or volume of each particle and the mass or volume of all particles to be detected, the particle grading characteristics are calculated.

2. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S2, the specific steps of the particle clustering segmentation algorithm include: Input the original point cloud data, calculate the average point spacing factor ASF of the point cloud data, and compare the calculated ASF value with the optimal ASF value k. If the ASF value of the input original point cloud data is less than k, the point cloud is uniformly downsampled so that the ASF value is greater than k, and then the point cloud is density clustered to divide the point cloud into single particle point clouds; if the ASF value of the input original point cloud data is greater than k, the point cloud is directly density clustered to obtain the segmented point cloud data.

3. The real-time detection method for soil particle gradation characteristics according to claim 2 is characterized in that: In step S2, the calculation process of the average point spacing factor ASF is: in, Represents a point in the point cloud p With any other point q The distance between the points is contained in the point cloud N points.

4. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S3, the particle sieving size calculation process includes: Screening size d i Calculation: ， Volume of the oriented bounding box V OBB The calculation formula is: ， in, W Indicates the width of the oriented bounding box, H Represents the height of the oriented bounding box, L Indicates the length of the oriented bounding box; Screen size D Defined as: ， in S is the side length of the sieve hole; The screening judgment rule is: if the diagonal length of the minimum control section of the particle oriented bounding box is d i Less than or equal to the mesh size D , then the particle is considered to be able to pass through the sieve hole, that is, it satisfies: 。 5. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S4, the overall contour of the particle is extracted by Alpha shape reconstruction, which specifically includes: Get the original point cloud data of the particles and define a complexity parameter α , to determine whether the distance between two discrete points in the original point cloud data is less than α , if less than α Then two discrete points are connected to form an edge, and an external boundary close to the actual shape of the particle is generated through all discrete points to form the overall outline of the particle; The missing point cloud data at the bottom is filled by uniform sampling, including: On the particle arrangement contour after Alpha shape reconstruction, uniform point cloud sampling is performed on the mesh surface of the particle arrangement contour: surface information is extracted from the mesh reconstructed by Alpha shape, data points are generated at equal distances on the mesh surface, points inside the mesh or discontinuous parts of the boundary are ignored, and the external form of the object is focused on; finally, continuous and complete point cloud data is generated on the mesh surface, and the mesh reconstructed by Alpha shape is deleted to obtain a continuous and complete new particle point cloud data; Generate a closed 3D mesh using Poisson surface reconstruction, including: On the point cloud data obtained by uniform sampling, the potential surface shape is inferred through the normal vector field of the point cloud, a smooth closed surface is generated from the point cloud, the gaps between the sampling points are filled, and a stable three-dimensional grid is formed.

6. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S4, the morphological parameters of the particles are obtained based on the particle volume, specifically including: After obtaining the volume of particle reconstruction V R After that, the morphological parameters λ Obtained by the following formula: ， in V OBB is the volume of the oriented bounding box.

7. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S5, the PMM model identifies the particle type through the point cloud projection image, including: The PMM model uses a CNN network to quickly identify particle types and automatically match the morphological parameters preset in the database. λ and density parameters ρ , the specific process is: Establish a point cloud projection image database, select the projection of the particle point cloud in the direction of the maximum bottom surface of the oriented bounding box; divide the image data set into a training set and a validation set, set the training step size, training rounds, and CNN variants, and then compare the model performance and accuracy after training multiple variants to obtain the optimal image recognition model; according to the morphological parameters of different types of particles pre-determined and stored in the database λ and density parameters ρ , apply the optimal image recognition model to identify the particle type and call the morphological parameters of the corresponding type of particles λ and density parameters ρ .

8. The real-time detection method for soil particle gradation characteristics according to claim 1 is characterized in that: In step S5, the PMM model is combined with the morphological parameters λ and density parameters ρ , the calculation process for correcting the particle mass / volume is: ， ， in, M i represents the estimated mass of the particle, V i represents the estimated volume of the particle, V OBB Represents the volume of an oriented bounding box.

9. The real-time detection method for soil particle gradation characteristics according to claim 8 is characterized in that: The step S6 specifically includes: In the grading calculation, X i1 or X i2 Indicates the percentage or proportion of particle size distribution during the screening process, based on the mass of each particle obtained. M i or volume V i And the mass of all particles to be tested M or volume V The particle size distribution characteristics are calculated by the following formula: (1) or (2) Therefore, formula (1) expresses that the sieve size is d i When the grading characteristic is X i1 ; If the particles to be detected are of the same type, that is, the density of the particles ρ , morphological characteristics λ If they are consistent, the grading characteristics can be directly obtained by formula (2): X i2 ; Using the particle's oriented bounding box V OBB The volume is taken as the equivalent volume of the particle.

Images