A method and system for estimating the overall particle size distribution of industrial particles

Abstract

This invention discloses a method and system for estimating the overall particle size distribution (PSD) of industrial particles. By simultaneously acquiring surface and depth images of industrial particles, the surface images are segmented to obtain the particle size distribution. The number of particle stack layers is calculated based on the depth image. A layered stacking model characterizing the relationship between the surface particle size distribution and the overall particle size distribution is established based on the surface particle size distribution and the number of stack layers. Finally, the overall particle size distribution of the particle stack is obtained based on the surface particle size distribution and the layered stacking model. This solves the technical problem of existing technologies being unable to accurately detect the overall particle size distribution of a particle stack online. By modeling the relationship between the surface and overall PSD from the perspectives of random sampling, particle penetration, and spatial distribution of the particle stack, it achieves accurate online estimation of the current particle stack's overall particle size distribution with only surface and depth images as input, and the estimation results have high reliability and robustness.

Description

Technical Field

[0001] This invention mainly relates to the field of blast furnace smelting technology, specifically a method and system for estimating the overall particle size distribution of industrial particles. Background Technology

[0002] Particle size distribution (PSD) is a crucial parameter reflecting the porosity, spatial structure, and packing density of a particle pack, and it holds significant research value in many industrial fields. For example, in a blast furnace (a core piece of equipment in the ironmaking industry), the particle size distribution of its material is a key performance indicator of the charging system, and the accuracy and continuity of its measurement are essential for monitoring gas permeability and the stability of internal reactions. Similarly, in mineral processing, particle size distribution is also an important indicator for selecting useful minerals. Therefore, researching an accurate online estimation method for particle size distribution has broad application value in industrial fields. Existing methods are mainly divided into three categories: physical methods, instrumental methods, and segmentation methods. Most ironmaking plants use physical detection methods, utilizing the physical properties of material particles, such as gravity and density, to detect particle size through screening, sedimentation, and other methods. This method is easy to implement but has low accuracy and can only be used offline. Instrumental methods mainly utilize instruments and equipment based on principles such as laser scattering and diffraction for offline detection. While they can obtain relatively accurate detection results, their offline-only nature makes them unsuitable for the needs of blast furnace operations. Segmentation is a particle size detection method based on image segmentation. While it offers advantages such as speed and low cost, it can only obtain the particle size distribution on the surface of the particle pile. Segmentation cannot be directly applied to estimating the overall particle size distribution of the particle pile. Therefore, this invention proposes a layered packing model based on the spatial distribution and formation mechanism of particle packing. This model shows the relationship between the surface particle size distribution and the overall particle size distribution. Finally, the particle size distribution of the particle pile is estimated using the surface particle size distribution calculated by the segmentation model.

[0003] Patent application CN112528913A discloses an image-based sand and gravel particle size detection and analysis system, including a frame, a feeding device, an image acquisition device, a supplementary lighting device, an image processing device, a display, a slider, and a recycling bin. The image processing device performs segmentation processing on the sand and gravel image, comprising: first, using ResNet based on Feature Pyramid Network (FPN) as the feature extraction network; second, proposing an adaptive ROIAlign module to map each proposal box output by RPN to all feature levels of FPN, and using a multi-scale feature fusion method by taking the maximum value through a parameter layer; finally, adding an IoU prediction branch to improve the network's localization performance and achieve more accurate segmentation; and proposing an NMS algorithm based on localization confidence to prioritize retaining detection boxes with higher localization accuracy for the target detection branch and mask branch, further improving segmentation accuracy. This invention can be applied to real-time online analysis of sand and gravel particle size in factories.

[0004] This invention only uses the existing Mask-RCNN algorithm to complete the target detection and segmentation on the surface of the ore, without analyzing the particle size distribution inside the particle pile.

[0005] The invention patent with publication number CN107766794B discloses a training method for an image segmentation model based on Mask RCNN and a particle size detection method. The training method includes: acquiring detection images of stones and sample images; inputting the detection images into a CNN, where the CNN extracts features from the detection images to obtain feature maps; generating anchor boxes through RPN, and correcting the anchor boxes according to the output of the RPN regression branch, and selecting a predetermined number of anchor boxes to form suggestion windows; mapping the suggestion windows onto the feature map of the last layer of the CNN, and generating a fixed-size RoI feature map for each suggestion window through the RoI Align layer; classifying and regressing RoIs using fully connected layers, generating a mask for each RoI through FCN, and segmenting to obtain stone contours; calculating the offset between the stone contours and the true contours marked in the sample images, and updating the parameters of the Mask RCNN neural network according to the offset.

[0006] This invention only uses the existing Mask-RCNN algorithm to complete the target detection on the surface of the ore, but does not solve the problem of calculating the overall particle size distribution of the particle pile. Summary of the Invention

[0007] This invention provides a method and system for estimating the overall particle size distribution of industrial particles, which solves the technical problem that existing technologies cannot accurately detect the overall particle size distribution of particle piles online.

[0008] To address the aforementioned technical problems, this invention proposes a method for estimating the overall particle size distribution of industrial particles, comprising:

[0009] Simultaneously acquire surface images and depth maps of industrial particles.

[0010] Image segmentation is performed on the particle surface image to obtain the particle size distribution on the particle surface.

[0011] The number of particle stacks is calculated based on the depth map.

[0012] Based on the particle size distribution on the particle surface and the number of particle layers, a layered packing model is established to characterize the relationship between the surface particle size distribution and the overall particle size distribution.

[0013] Based on the particle size distribution on the particle surface and the stratified packing model, the overall particle size distribution of the particle pile is obtained.

[0014] Furthermore, image segmentation is performed on the particle surface image to obtain the particle size distribution on the particle surface, including:

[0015] Create an instance segmentation dataset based on particle surface images.

[0016] An industrial particle segmentation model was obtained by fine-tuning the SAM pre-trained model vit-h using an instance segmentation dataset.

[0017] An industrial particle segmentation model is used to segment the particle surface image to obtain the particle size distribution on the particle surface.

[0018] Furthermore, calculating the number of particle stacks based on the depth map includes:

[0019] Based on the depth information of the pixels in the depth map, calculate the overall maximum and minimum values ​​of the depth information. The overall maximum and minimum values ​​are the average values ​​of the top 10% and bottom 10% of the data in terms of depth value size, respectively.

[0020] The number of particle layers is calculated based on the overall maximum and minimum values. The specific calculation formula is as follows:

[0021]

[0022] Where H is the number of layers in the particle stack, h max and h min These are the overall maximum and minimum values, d. m The average size of the surface particles. This indicates rounding down to the nearest integer.

[0023] Furthermore, establishing a layered packing model characterizing the relationship between surface particle size distribution and overall particle size distribution includes:

[0024] The particle stack is defined as the permeation and superposition of multiple layers of particles, and the overall particle size distribution of the particle stack is expressed as a function of the ideal particle size distribution and particle permeability, specifically:

[0025]

[0026]

[0027]

[0028]

[0029] Among them, D total This refers to the overall particle size distribution of the particle pile. Let represent the particle size distribution of the entire particle pack in the k-th particle size interval. This represents the actual particle size distribution of the i-th layer particles in the k-th particle size interval. and Let N represent the ideal particle size distribution of the i-th layer and the (i-1)-th layer in the k-th particle size interval. i,k P represents the number of particles in the i-th layer within the k-th particle size interval. i,k and P (i-1),k Let H and k be the particle permeability of the i-th and (i-1)-th layers in the k-th particle size range, respectively, where H is the number of layers in the particle stack, and k is a preset positive integer. This indicates the rounding up operation.

[0030] Based on the spatial distribution characteristics of the particle stack, the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution is obtained.

[0031] Based on the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution, the analytical relationship between the surface particle size distribution and the overall particle size distribution is obtained.

[0032] The surface particle permeability is derived from the state difference between random loose packing and random compacted packing.

[0033] Based on the analytical relationship between surface particle size distribution and overall particle size distribution, and the permeability of surface particles, a layered packing model is established.

[0034] Furthermore, based on the spatial distribution characteristics of the particle stack, the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution is obtained as follows:

[0035] Based on the spatial distribution characteristics of the particle stack, the volume relationship of each layer of the particle stack in the ideal particle size distribution is derived as follows:

[0036]

[0037] V1 id and V i id S1 and S2 are the ideal volumes of the first and i-th particle layers, respectively. iThese are the cross-sectional areas of the first and i-th particle layers, respectively.

[0038] Based on the volume relationship of each layer of the particle stack in the ideal particle size distribution, the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution is derived.

[0039]

[0040] in and These are the ideal particle size distributions for the first and i-th particle layers, respectively.

[0041] Furthermore, based on the particle size distribution relationship of each layer of the particle pack in the ideal particle size distribution, the analytical relationship between the surface particle size distribution and the overall particle size distribution is obtained, specifically:

[0042]

[0043] Where O represents a 1*k dimensional unit vector matrix, The actual particle size distribution of the first particle layer and P1 is the particle permeability of the first layer of particles, and P1 = [P 1,1 ,P 1,2 ,...,P 1,k ], f rec This represents a function that calculates the reciprocals of matrix elements.

[0044] Furthermore, the permeability of surface particles is derived from the state difference between random loose packing and random compacted packing, including:

[0045] Based on the state difference between random loose packing and random compacted packing, the average particle permeability is obtained, specifically:

[0046]

[0047] Among them, P m ρ represents the average particle permeability of the particle stack. c and ρ l V represents the packing density under random compact packing and random loose packing, respectively. p This represents the total volume of particles in a particle pile. This represents the void volume in the particle stack.

[0048] Based on the average particle permeability, the surface particle permeability is obtained through linear transformation, specifically:

[0049]

[0050] Among them, P 1,jThis represents the particle permeability of the first layer of particles in the j-th particle size range.

[0051] The present invention provides an industrial particle size distribution estimation system comprising:

[0052] The present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the industrial particle size distribution estimation method provided by the present invention.

[0053] The present invention proposes a method and system for estimating the overall particle size distribution of industrial particles.

[0054] This invention addresses the technical challenge of accurately detecting the overall particle size distribution (PSD) of a particle stack online. By modeling the relationship between the surface and the overall PSD from the perspectives of random sampling, particle penetration, and spatial distribution, it enables accurate online estimation of the overall PSD of the current particle stack with only surface and depth images as input. The estimation results exhibit high reliability and robustness, which is of significant importance for subsequent research related to interdisciplinary integration.

[0055] The beneficial effects of this invention specifically include:

[0056] (1) An algorithm framework is proposed that only requires input of surface image and depth image to estimate the overall particle size distribution of the current particle pile.

[0057] (2) Based on the formation characteristics and spatial distribution of particle piles, the analytical relationship between surface particle size distribution and overall particle size distribution was constructed using the principle of particle permeation.

[0058] (3) The concept of particle permeability was proposed, and the analytical expression of particle permeability was derived using bulk density. The formation process of particle piles was further quantified, which provided a key role in analyzing the particle size distribution of particle piles. Attached Figure Description

[0059] Figure 1 This is the overall algorithm framework for the industrial particle size distribution estimation method in Embodiment 2 of the present invention;

[0060] Figure 2 This is a schematic diagram of an industrial particle surface image and laser detection system according to Embodiment 2 of the present invention;

[0061] Figure 3 This is a schematic diagram of the idealized particle stack spatial distribution of Embodiment 2 of the present invention;

[0062] Figure 4 This is a graph showing the test results for estimating the overall particle size distribution of industrial particles in Embodiment 2 of the present invention. Figure 4 (a) Figure 4 (b) and Figure 4 (c) shows the particle surface image, segmentation results, and estimated overall particle size distribution, respectively;

[0063] Figure 5 This is a structural block diagram of the industrial particle size distribution estimation system according to an embodiment of the present invention.

[0064] Figure label:

[0065] 10. Memory; 20. Processor. Detailed Implementation

[0066] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0067] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0068] Example 1

[0069] Embodiment 1 of the present invention provides a method for estimating the overall particle size distribution of industrial particles, comprising:

[0070] Step S101: Simultaneously acquire particle surface images and depth maps of industrial particles.

[0071] Step S102: Perform image segmentation on the particle surface image to obtain the particle size distribution on the particle surface.

[0072] Step S103: Calculate the number of particle stack layers based on the depth map.

[0073] Step S104: Based on the particle size distribution on the particle surface and the number of particle layers, establish a layered packing model that characterizes the relationship between the surface particle size distribution and the overall particle size distribution.

[0074] Step S105: Based on the particle size distribution on the particle surface and the layered packing model, obtain the overall particle size distribution of the particle pile.

[0075] The method for estimating the overall particle size distribution of industrial particles provided in this invention simultaneously acquires particle surface images and depth maps. It then segments the particle surface images to obtain the particle size distribution on the particle surface, calculates the number of particle stack layers based on the depth map, and establishes a layered stacking model characterizing the relationship between the surface particle size distribution and the overall particle size distribution based on the particle size distribution and the number of particle stack layers. Finally, it obtains the overall particle size distribution of the particle stack based on the particle surface particle size distribution and the layered stacking model. This solves the technical problem of existing technologies being unable to accurately detect the overall particle size distribution of particle stacks online. By modeling the relationship between the surface and overall particle size distribution (PSD) from the perspectives of random sampling, particle penetration, and spatial distribution of the particle stack, it achieves accurate online estimation of the current overall particle size distribution of the particle stack with only surface and depth images as input. Furthermore, the estimation results exhibit high reliability and robustness, which is of significant importance for subsequent research related to fusion and interdisciplinary studies.

[0076] Example 2

[0077] refer to Figure 1 The method for estimating the overall particle size distribution of industrial particles proposed in this invention includes the following steps:

[0078] The overall algorithm steps of this invention are as follows:

[0079] (1) Acquire RGB data of industrial particle surfaces and create a dataset. Images of industrial particle surfaces on a conveyor belt are acquired using an industrial camera, and an RGB instance segmentation dataset is created using annotation tools such as LabelMe. Then, the SAM pre-trained model vit-h is fine-tuned using the industrial particle dataset to obtain the industrial particle segmentation model f. seg (I). The particle size distribution D1 of the surface can be calculated by segmenting the surface industrial particle image. ac =f D (f seg (I)).

[0080] (2) By analyzing the formation characteristics of particle piles, particle piles are defined as the permeation and superposition of multiple layers of particles, and the overall particle size distribution is expressed as an idealized particle size distribution and a function of particle permeability.

[0081] (3) By analyzing the spatial distribution characteristics of the particle stack, the volume relationship of each layer of the idealized particle stack (without considering particle penetration) is derived.

[0082] (4) From the perspective of probability and statistics, the process of removing particles from the silo can be regarded as a random resampling process according to rules (each time a fixed volume V of particles is removed), and from this, the idealized relationship between the particle size distributions of each layer can be derived.

[0083] (5) Based on the results derived above, the analytical relationship between the surface particle size distribution and the overall particle size distribution can be obtained. However, there is still an unknown factor, which is the particle permeability of the surface layer.

[0084] (6) Inspired by packing density, the average particle permeability is derived from the state differences between random loose packing and random compacted packing based on the definition of particle permeability. Then, through linear transformation, the particle permeability of the surface layer is obtained. Finally, the overall particle size distribution is estimated by inputting a surface image.

[0085] The specific implementation plan is as follows:

[0086] Step 1: Obtain RGB and depth data and create an instance segmentation dataset, and calculate the surface particle size distribution.

[0087] 1) Data Acquisition. To acquire RGB images, RGB image information is collected using hardware devices such as industrial cameras. The industrial camera used is the German-made Basler acA2500-14gm Basler industrial area scan camera, and the data acquisition software is Pylon Viewer. Communication is via Ethernet. Depth data is acquired using laser scanning, and the acquisition system is as follows: Figure 2 As shown.

[0088] 2) Create an instance segmentation dataset. Using annotation tools such as LabelMe, manually annotate the collected and processed images, labeling each industrial particle, and finally create an instance segmentation dataset through data augmentation.

[0089] 3) Apply a segmentation model to segment the RGB image, extract the contour of each particle, and then calculate the area of ​​each contour as the size of the particle, statistically representing the surface particle size distribution.

[0090] Step 2: Layered packing model based on particle permeation.

[0091] The accumulation of particles is similar to the construction of a tall building; subsequent particles accumulate on top of those that have settled below, layer upon layer, until all particles are at rest. Therefore, a stationary particle pile composed of particles of varying sizes can be considered to have a certain spatial stratification. Let the number of layers in a particle pile be H. The outermost layer is denoted as the first layer, and subsequent layers follow the same pattern. Therefore, the overall particle size distribution of the particle pile can be defined as:

[0092]

[0093]

[0094]

[0095] Among them, D total This refers to the overall particle size distribution of the particle pile. Let represent the particle size distribution of the entire particle pack in the k-th particle size interval. N represents the actual particle size distribution of the i-th layer particles in the k-th particle size interval. i,k H represents the number of particles in the i-th layer within the k-th particle size interval, H is the number of layers in the particle stack, and k is a preset positive integer, typically 4, 8, or 10. This indicates the rounding up operation.

[0096] The particle size distribution of the outermost layer of a particle pile can be calculated using a surface detection method based on visual perception, but the particle size distribution of other layers cannot be directly calculated. During the accumulation process, smaller particles in the upper layer have a certain probability of gradually penetrating the gaps between particles in the lower layer during movement and collisions, thus accumulating in the lower layer. Smaller particles are more likely to penetrate into the lower layer through the gaps between particles. Furthermore, as particles sink, the resistance to penetration increases, making penetration increasingly difficult. Therefore, for ease of calculation, only the penetration of the current layer into the next layer is considered, while penetration into the two layers below or even deeper layers is ignored (the probability is too low to be considered). Based on this theoretical foundation, this embodiment first proposes a probability describing the penetration of particles in the current layer into particles in the lower layer, called the particle permeability P. The particle size distribution of each layer is defined as:

[0097]

[0098] in and Let P be the ideal particle size distribution of the i-th layer and the (i-1)-th layer in the k-th particle size interval (when particle penetration does not occur). i,k and P (i-1),k denoted as , respectively, the particle permeability of the i-th and (i-1)-th layers in the k-th particle size interval. * indicates element-wise multiplication, also known as the Hadamard product.

[0099] Step 3: Spatial distribution of the layered stacking model.

[0100] Directly calculating the particle size distribution of each layer is extremely difficult. This embodiment discovers that if each particle is considered a uniform sphere, there is a quantitative relationship between particle size and volume. Therefore, this embodiment examines a particle pile on a plane from a spatial geometry perspective. Thus, the geometry of the particle pile can be roughly modeled as a superposition of multiple cylindrical layers. Taking a three-layer particle pile as an example, without considering particle permeation, the relationship between the idealized volumes of each layer can be calculated. Based on the volume formula of a cylinder and trigonometric function formulas, we can obtain:

[0101]

[0102] Where F1, F2, and F3 represent respectively Figure 3 The sector area of ​​each layer F1 = S1 + S2 + S3 + Z1, F2 = S2 + S3 + Z2, F3 = S3 + Z3. Z1, Z2, and Z3 represent... Figure 3 Area of ​​the isosceles triangles below each layer S represents the ideal volume of the i-th particle layer. i Let L represent the cross-sectional area of ​​the i-th particle layer, and L represent the length of the particle stack on the conveyor belt. In the current particle stack, L is the same for all layers. Figure 3 As shown. Figure 3 In this context, 'l' represents the height of a single particle layer. Therefore, in this embodiment, the idealized volume relationship of each layer is expressed as:

[0103]

[0104] S i It can be done directly using known parameters ( S i+1 Calculations were performed using (etc.). Although the relationship between the idealized volumes of each layer is known, it is still difficult to correlate it with the particle size distribution in this embodiment.

[0105] Step 4: Random resampling before forming the particle pile.

[0106] If we temporarily disregard the infiltration of particles during the accumulation process, the process of removing particles from the silo is essentially a random process.

[0107] Therefore, in order to explore the particle size distribution of each layer of the theory, this embodiment considers modeling the particle accumulation process from a statistical and probabilistic perspective. At each moment, several particles of different sizes fall from the hopper onto the conveyor belt due to gravity. The process of particles flowing out of the hopper can be regarded as a random resampling process (without considering the motion changes of the particle accumulation process). Thus, the hopper can be regarded as a box filled with balls of different sizes, and the particles falling from the hopper at a fixed flow rate Q can be regarded as randomly sampling several balls from the box with replacement according to certain rules (the sampled volume at each moment is a fixed value V).

[0108]

[0109] Where ρ represents the density of the particle, T i N represents the ideal time to remove all particles from the i-th layer. i v represents the total number of particles in the i-th layer. j Let d represent the volume of the j-th particle.j Let n represent the diameter of the j-th particle, k represent the number of particle size intervals (for example, if k = 8, it means there are 8 particle size levels, which is a fixed value, and each layer has 8 particle size levels), and n represent the diameter of the j-th particle. i,q This represents the number of particles in the i-th layer within the k-th particle size interval. Clearly, as the volume of the particle pile increases, the number of particles within each particle size interval also increases, therefore:

[0110] n i,1 ,n i,2 ,…,n i,k ∝V i id (8)

[0111] Where ∝ represents a positive correlation. As the particle size distribution increases, the number of particles in each size range also increases. And according to the definition of particle size distribution, we can obtain:

[0112]

[0113] in and These are the ideal particle size distributions for the first and i-th particle layers, respectively.

[0114] According to equation (4), we can obtain:

[0115]

[0116]

[0117] Further derivation yields:

[0118]

[0119] Where f rec This function represents the reciprocal of each element of a matrix. Represents a 1*k dimensional matrix. The actual particle size distribution of the first particle layer and P1 is the particle permeability of the first layer of particles, and P1 = [P 1,1 ,P 1,2 ,...,P 1,k ] represents the surface particle permeability. * indicates element-wise multiplication, also known as the Hadamard product.

[0120] Substituting the specific value of i into equation (4), we get:

[0121]

[0122] Then, by superimposing the above equations, we can obtain:

[0123]

[0124] Therefore, the overall particle size distribution can be written as:

[0125]

[0126] At this point, only the surface particle permeability P1 is unknown in the formula. It can be calculated through segmentation algorithms and statistics.

[0127] Step 5: Use bulk density to characterize particle permeability.

[0128] The calculation of particle permeability is relevant to the problem of particle bulk density. Bulk density refers to the ratio of particle volume to the total volume of the particle packing.

[0129]

[0130] Where ρ represents the bulk density, V represents the actual volume of the particle pile, and V p This represents the total volume of all particles. This represents the void volume in the particle stack.

[0131] The mathematical problem of particle packing density can be traced back to Kepler's conjecture proposed in the early 17th century: for an ideal sphere of equal diameter, no other packing method produces a higher packing density than face-centered cubic or hexagonal compact packing. Through a large number of experimental studies, many scholars have found that the packing density of a single sphere of equal diameter is independent of the particle size, but there are three repeatable packing patterns: (1) when a large number of spheres are gently poured into a container, the spheres will form a random loose packing with a packing density ρ≤0.60; (2) after the container is vibrated, shaken, patted or tamped, the spheres will form a random dense packing with a packing density ρ≤0.64; (3) if the spheres are carefully arranged one by one according to face-centered cubic or hexagonal compact packing, an ordered dense packing is formed with a packing density ρ≤0.74.

[0132] Particle permeability describes the probability that particles will fill the underlying voids during the transition from a loose to a compacted packing state. Therefore, the following holds true:

[0133]

[0134] Where P m ρ represents the average particle permeability of the particle pack. c ρ represents the packing density under random compact packing. l This represents the packing density under random, loose packing. V represents the volume of the voids between particles.p This represents the total volume of the particles. Before permeation, the gaps between particles are relatively large; after permeation, the gaps between particles become smaller. The degree to which these gaps become smaller can be expressed as particle permeability.

[0135]

[0136] therefore, It represents the permeability of the i-th particle layer to particles in the j-th particle size range. In actual calculations, this embodiment defines T. c It is actually an H*k matrix. If we construct it as a linear function of i and j, that is... Where a, b, and c are all parameters to be determined. In addition... We can obtain:

[0137]

[0138] Let represent the packing density of the random compact packing in the j-th grain size interval of the i-th grain layer. Then, let It can be solved Therefore, we can conclude that:

[0139]

[0140] P 1,j This represents the particle permeability of the first layer (surface layer) to particles in the j-th particle size range, and P1 = [P 1,1 ,P 1,2 ,...,P 1,k According to P] 1,j The value of P1 can be calculated naturally, and finally D can be obtained. total The estimation results. Figure 4 The image shows the test results obtained using the method described in this embodiment.

[0141] This embodiment first analyzes the relationship between surface PSD and overall PSD from a mathematical modeling perspective, and then uses the surface PSD obtained through a segmentation algorithm to predict the overall PSD of the particle pile. To model the mathematical relationship between surface PSD and overall PSD, this embodiment starts from the essence of particle pile formation to explore clues. From the perspective of a single particle, the formation of a particle pile can be seen as the result of countless random samplings. From the perspective of inter-particle correlation, the formation of a particle pile can be seen as the process of small particles penetrating into voids. From an overall perspective, the formation of particle piles exhibits certain patterns in spatial distribution. Inspired by the above, this embodiment models the relationship between surface and overall PSD from the perspectives of random sampling of particle piles, particle penetration, and spatial distribution.

[0142] The method for estimating the overall particle size distribution of industrial particles provided in this invention simultaneously acquires particle surface images and depth maps. It then segments the particle surface images to obtain the particle size distribution on the particle surface, calculates the number of particle stack layers based on the depth map, and establishes a layered stacking model characterizing the relationship between the surface particle size distribution and the overall particle size distribution based on the particle size distribution and the number of particle stack layers. Finally, it obtains the overall particle size distribution of the particle stack based on the particle surface particle size distribution and the layered stacking model. This solves the technical problem of existing technologies being unable to accurately detect the overall particle size distribution of particle stacks online. By modeling the relationship between the surface and overall particle size distribution (PSD) from the perspectives of random sampling, particle penetration, and spatial distribution of the particle stack, it achieves accurate online estimation of the current overall particle size distribution of the particle stack with only surface and depth images as input. Furthermore, the estimation results exhibit high reliability and robustness, which is of significant importance for subsequent research related to fusion and interdisciplinary studies.

[0143] Specifically, the beneficial effects of this embodiment include:

[0144] 1. An algorithm framework is proposed to estimate the overall particle size distribution of a particle pile from surface images. This method can not only obtain the particle size distribution on the surface of the particle pile, but also estimate the overall particle size distribution of the particle pile.

[0145] 2. The concept of particle permeability was proposed and derived through bulk density calculation. The introduction of particle permeability further analyzes the formation process of particle piles and helps to quantitatively characterize the analytical relationship between surface particle size distribution and overall particle size distribution.

[0146] 3. A layered packing model is proposed from the perspectives of particle penetration, spatial distribution, and random resampling to quantitatively characterize the analytical relationship between surface particle size distribution and overall particle size distribution.

[0147] Example 3

[0148] This example uses a 2650m³ iron smelter in a certain ironworks. 3 The method for estimating the overall particle size distribution of industrial particles in this embodiment was verified using a large blast furnace. The specific steps are as follows:

[0149] 1. Dataset Creation. To acquire RGB images of blast furnace sinter and laser detection data, this invention uses industrial cameras and laser scanners to collect RGB image information and laser depth information on the blast furnace charging conveyor belt of a steel plant. The laser detection equipment used is the SICK LMS4111R-13000 laser line scanner, and the data acquisition software is SOPAS Engineering Tool. The industrial camera used is the German Basler acA2500-14gmBasler industrial area array camera, and the data acquisition software is Pylon Viewer. Both communicate via Ethernet. A total of 8 sets of data were collected within a concentrated time period using the industrial camera and laser scanner (a new batch of industrial particles was used after each set of data was collected). Each set of data includes 1000-2000 images (the acquisition time varies, with an acquisition frequency of 60ms, similar to the laser scanning frequency; the acquisition frequency of the laser scanning data fluctuates between 55ms and 65ms due to hardware limitations). The image size is 2592*2048. The laser scan data contains 1000-2000 lines of laser messages, including hardware information, ranging data, and configuration data. The ranging data provides the depth information we need. A joint calibration algorithm is then used to obtain the registration dataset.

[0150] 2. The SAM pre-trained model vit-h was fine-tuned and inferred using this dataset. Adam was used as the optimization algorithm during fine-tuning and inference. The learning rate was set to 10^-4, with a cosine annealing learning rate scheduling strategy to adjust the learning rate, where the minimum value was 10^-6. The weight decay was set to 10^-8. Fine-tuning yielded a segmentation model that better represented industrial particle images, capable of extracting the contours of all industrial particles in the surface image. Then, the size of each industrial particle was calculated using a single-particle calculation method, and finally, the surface particle size distribution of the industrial particle pile was statistically analyzed.

[0151] 3. Calculate the number of strata H in the particle stack based on the depth information. The specific process is as follows:

[0152] (1) First, a depth map (laser scan map) of the surface image is obtained through data acquisition. The pixels of this depth map represent the depth information of each point (acquired almost simultaneously with the RGB image captured by the camera, and can be obtained during the data acquisition stage).

[0153] (2) Calculate the overall maximum and minimum values ​​of the depth information. Taking a 512*512 depth map as an example, calculate the mean of the top 10% of data with larger depth values ​​as the overall maximum value. Calculate the mean of the bottom 10% of data with smaller depth values ​​as the overall minimum value.

[0154]

[0155] Where h t,j h represents the j-th depth value among the top 10% of data with the largest depth values. b,j The depth value represents the j-th depth value in the last 10% of the data with smaller depth values. N represents the total number of pixels in the depth map. For example, in a 512*512 depth map, N = 512*512.

[0156] (3) Calculate the number of stratifications H in the particle stack:

[0157]

[0158] Where d m The average size of the surface particles can be calculated from the surface particle size distribution. This indicates rounding down to the nearest integer.

[0159] In this embodiment, the number of particle gradations k is determined manually (the number of particle gradations is determined before calculation; the larger the value, the finer the particle size distribution). In this case, k is set to 8. Then, the surface particle size distribution result is input into the layered packing model proposed in this embodiment, and the overall particle size distribution of the particle pack can be calculated.

[0160] 4. Experimental Results and Analysis. Experiments were conducted on the same experimental platform using the same dataset, based on ADPSD, CPSD, LUNet, RLPNet, MASU R-CNN, and MP. ADPSD and CPSD are traditional segmentation-based PSD detection methods, LUNet and RLPNet are semantic segmentation-based PSD detection methods, and MASU R-CNN and MP are instance segmentation-based PSD detection methods. Table 1 shows that the proposed method achieves the best results on KL, WD, and mDis compared to the latest PSD detection methods. Using SAM, a mainstream large-scale visual model, as a baseline, our method improves mDis by 2.49% compared to SAM, but only by 0.81% on mIoU. This is mainly because we do not directly replace the overall PSD with the surface PSD, but instead use a mathematical model to calculate a more reasonable result based on the surface PSD. Furthermore, according to Table 1, the higher the accuracy of the segmentation model, the higher the accuracy of the predicted PSD. This also underscores the necessity of fine-tuning the SAM pre-trained model vit_h. We fine-tuned the SAM pre-trained model vit_h using an industrial particle dataset to improve its segmentation accuracy on industrial particle samples, thereby improving the overall PSD calculation accuracy in our method. The time consumption of our method mainly includes the inference of the segmentation model, the calculation of particle size, and the calculation of the overall PSD. The parameters of our segmentation model are the same as those of SAM, approximately equal to the number of parameters in vit_h. A 2176*808 image achieved 10.8GB of VRAM on a 2080ti graphics card during testing. The time and space complexity of the particle size and overall PSD calculation processes are both O(n), consuming relatively little time. As shown in Table 1, the method in this embodiment is only 0.01 FPS lower than SAM.

[0161] This invention takes industrial particle piles on conveyor belts as the research object and proposes a method for estimating the overall particle size distribution of industrial particles based on surface images and stacking mechanisms. Compared with existing methods, the method proposed in this invention has higher estimation accuracy, and because the estimation is based on the characteristics formed inside the particle pile, the estimation results have higher reliability and robustness, which is of great significance for subsequent research related to integration and interdisciplinary fields. Finally, the accuracy of the estimation results is evaluated using the KL and WD indices, as shown in Table 1.

[0162] Table 1

[0163]

[0164] Reference Figure 5The industrial particle size distribution estimation system proposed in this embodiment includes a memory 10, a processor 20, and a computer program stored on the memory 10 and run on the processor 20. When the processor 20 executes the computer program, it implements the steps of the industrial particle size distribution estimation method proposed in this embodiment.

[0165] The specific working process and working principle of the industrial particle overall size distribution estimation system in this embodiment can be referred to the working process and working principle of the industrial particle overall size distribution estimation method in this embodiment.

[0166] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the overall particle size distribution of industrial particles, characterized in that, The method includes: Simultaneously acquire surface images and depth maps of industrial particles; Image segmentation is performed on the particle surface image to obtain the particle size distribution on the particle surface; Calculate the number of particle layers based on the depth map; Based on the particle size distribution on the particle surface and the number of particle layers, a layered packing model is established to characterize the relationship between the surface particle size distribution and the overall particle size distribution. This layered packing model includes: The particle stack is defined as the permeation and superposition of multiple layers of particles, and the overall particle size distribution of the particle stack is expressed as a function of the ideal particle size distribution and particle permeability, specifically: ， ， ， ， in, This refers to the overall particle size distribution of the particle pile. For the entire particle stack in the first Particle size distribution within each particle size range, For the first Layer particles in the first Actual particle size distribution in each particle size range and The first Layer particles and the first Layer particles in the first Ideal particle size distribution within each particle size range For the first Layer particles in the first The number of particles within each particle size range and The first Layer and first Layer particles in the first Particle permeability in each particle size range The number of layers in the particle stack. It is a preset positive integer. This indicates the rounding up operation; Based on the spatial distribution characteristics of the particle stack, the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution is obtained. This includes obtaining the following: Based on the spatial distribution characteristics of the particle stack, the volume relationship of each layer of the particle stack in the ideal particle size distribution is derived as follows: ， in and The first Layer and first The ideal volume of the granular layer, and The first Layer and first Cross-sectional area of ​​the granular layer; Based on the volume relationship of each layer of the particle stack in the ideal particle size distribution, the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution is derived. ， in and The first Layer and first The ideal particle size distribution of the layered particles, and ; Based on the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution, the analytical relationship between the surface particle size distribution and the overall particle size distribution is obtained. Specifically, the analytical relationship between the surface particle size distribution and the overall particle size distribution is obtained based on the particle size distribution relationship of each layer of the particle stack in the ideal particle size distribution. ， in, Indicates 1 A 1D unit vector matrix, For the first The actual particle size distribution of the layered particles and , Let be the particle permeability of the first layer of particles, and , This represents a function that calculates the reciprocals of matrix elements. The permeability of surface particles was derived from the state difference between random loose packing and random compacted packing. Based on the analytical relationship between surface particle size distribution and overall particle size distribution and the permeability of surface particles, a layered packing model is established. Based on the particle size distribution on the particle surface and the stratified packing model, the overall particle size distribution of the particle pile is obtained.

2. The method for estimating the overall particle size distribution of industrial particles according to claim 1, characterized in that, Image segmentation of the particle surface image is performed to obtain the particle size distribution on the particle surface, including: Create an instance segmentation dataset based on particle surface images; An industrial particle segmentation model was obtained by fine-tuning the SAM pre-trained model vit-h using an instance segmentation dataset. An industrial particle segmentation model is used to segment the particle surface image to obtain the particle size distribution on the particle surface.

3. The method for estimating the overall particle size distribution of industrial particles according to claim 2, characterized in that, Calculating the number of particle stacks based on the depth map includes: Based on the depth information of the pixels in the depth map, calculate the overall maximum and minimum values ​​of the depth information. The overall maximum and minimum values ​​are the average values ​​of the top 10% and bottom 10% of the data in terms of depth value size, respectively. The number of particle layers is calculated based on the overall maximum and minimum values. The specific calculation formula is as follows: ， in, The number of layers in the particle stack. and These are the overall maximum and the overall minimum values, respectively. The average size of the surface particles. This indicates rounding down to the nearest integer.

4. The method for estimating the overall particle size distribution of industrial particles according to claim 3, characterized in that, The permeability of surface particles can be derived from the state differences between random loose packing and random compacted packing, including: Based on the state difference between random loose packing and random compacted packing, the average particle permeability is obtained, specifically: , Among them, This represents the average particle permeability of the particle stack. and These represent the packing densities under random compact packing and random loose packing, respectively. This represents the total volume of particles in a particle pile. This represents the void volume within the particle stack. Based on the average particle permeability, the surface particle permeability is obtained through linear transformation, specifically: ， in, This indicates that the first layer of particles is in the [missing information]. Particle penetration rate in each particle size range.

5. An industrial particle size distribution estimation system, the system comprising: The memory (10), the processor (20), and the computer program stored in the memory (10) and executable on the processor (20) are characterized in that the processor (20) implements the steps of the method according to any one of claims 1 to 4 when executing the computer program.

Images