A granulometry detection method and system for a granulator

Abstract

This invention discloses a particle size detection method and system for a granulator, relating to the field of particle detection technology. The method involves acquiring multi-angle image sequences and obtaining a particle size distribution list through contour repair and 3D reconstruction. Simultaneously, the spindle speed and feeding rate are monitored during particle size list generation. When the ratio of these two values ​​exceeds a stable threshold, the particle size data is marked with operating condition information. Subsequently, stable and unstable operating condition data are distinguished based on the markings, and different filtering algorithms are used to remove outliers. Finally, the average particle size, distribution width, and pass rate are calculated by merging the data. This method dynamically correlates the detection data with the real-time production status, effectively eliminating interference from operating condition fluctuations through classification filtering. This allows the detection report to more accurately reflect the true particle quality under stable production conditions, providing a reliable basis for process optimization.

Description

Technical Field

[0001] This invention belongs to the field of particle detection technology, specifically a particle size detection method and system for a granulator. Background Technology

[0002] Traditional particle size detection in granulation processes largely relies on image analysis techniques. This involves acquiring particle images, converting them to grayscale, extracting contours, and reconstructing them in 3D to calculate particle size distribution. Existing methods focus on optimizing image processing algorithms, such as using morphological operations to repair contour breaks or voids, and mapping multi-angle image sequences to reconstruct point cloud models in 3D space to estimate the equivalent particle diameter. These techniques enable non-contact online measurement and can generate preliminary particle size distribution lists.

[0003] Existing technical solutions have shortcomings. Conventional detection methods typically operate independently and are not linked in real-time to the actual production parameters of the granulator. Particle formation is directly affected by operating parameters such as spindle speed and feed rate; when these parameters fluctuate, particle size may become abnormal. However, existing methods process all detection data uniformly, failing to distinguish between data differences under stable and unstable production conditions. While image processing can restore contours, it cannot identify actual particle variations or measurement artifacts caused by operating condition fluctuations. Consequently, final statistical indicators such as average particle size or pass rate may deviate from actual production levels, misleading process adjustments.

[0004] This invention addresses how to dynamically correlate particle size detection data with real-time production conditions and achieve data classification processing. Existing technologies lack a synchronous monitoring mechanism for the ratio of spindle speed to feed rate, resulting in particle size data failing to carry operational information. Furthermore, uniform data filtering methods struggle to adapt to outlier characteristics under different production conditions, affecting the accuracy of the final report. Therefore, a method is needed to mark operational conditions in real-time during particle size list generation and to clean the data based on these markings. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes a method for particle size detection in a granulator, comprising:

[0007] Collect multi-angle image sequences of material particles at the discharge port of the granulator, and perform grayscale processing to form a particle contour dataset;

[0008] Based on morphological operations, contour integrity restoration is performed on the particle contour dataset to eliminate contour breaks and holes caused by particle adhesion or image noise, and obtain complete particle contour data.

[0009] The repaired particle contour data is mapped to a preset three-dimensional particle space coordinate system, and the three-dimensional point cloud model of each particle is reconstructed according to the spatial distribution of the contour points.

[0010] Calculate the equivalent volume of the 3D point cloud model of each particle, and convert it into the equivalent sphere diameter based on the equivalent volume to generate a preliminary particle size distribution list.

[0011] During the generation of the preliminary particle size distribution list, the current spindle speed and feeding rate of the granulator are monitored simultaneously.

[0012] When the real-time ratio of the spindle speed to the feeding rate exceeds the preset stable production threshold range, the data in the preliminary particle size distribution list is marked to generate particle size data with working condition markings.

[0013] Based on the particle size data with operating condition markings, stable operating condition data and unstable operating condition data are distinguished, and different data filtering algorithms are applied to the data under the two operating conditions to remove abnormal particle size values.

[0014] The filtered stable and unstable operating condition data are merged, and the final average particle size, particle size distribution width, and pass rate are calculated to output a particle size test report.

[0015] Furthermore, the process of forming the granular contour dataset includes:

[0016] Collect a multi-angle image sequence of material particles at the discharge port of the granulator, wherein the multi-angle image sequence contains particle outline information under different lighting conditions;

[0017] The multi-angle image sequence is converted to grayscale to obtain a set of granular grayscale images, where each image in the set of granular grayscale images corresponds to a sampling angle.

[0018] The set of grayscale particle images is processed by an adaptive threshold segmentation algorithm to extract the binary contours of particles in each image, forming a particle contour dataset.

[0019] The step of performing grayscale processing on the multi-angle image sequence to obtain a set of granular grayscale images includes:

[0020] Read each frame of the original color image from the multi-angle image sequence;

[0021] For each frame of the original color image, convert it from the RGB color space to the YCbCr color space;

[0022] Extract the luminance component Y from the YCbCr color space as the initial grayscale image of each frame of the original color image;

[0023] The initial grayscale image is locally contrast-enhanced to improve the distinction between particle edges and the background, resulting in an enhanced grayscale image.

[0024] Brightness normalization is performed on all enhanced grayscale images to eliminate the overall brightness deviation caused by differences in lighting from different angles, forming a standardized grainy grayscale image.

[0025] All standardized grain grayscale images that have undergone brightness normalization are sorted and integrated according to their corresponding acquisition timestamps and angle codes to form the grain grayscale image set.

[0026] Further, the step of processing the set of grayscale images of particles using an adaptive threshold segmentation algorithm to extract the binarized contours of particles in each image and form a particle contour dataset includes:

[0027] For each standardized grainy grayscale image in the set of grainy grayscale images, it is divided into multiple non-overlapping image sub-blocks;

[0028] Calculate the mean gray value and standard deviation of the pixels within each image sub-block;

[0029] The local segmentation threshold of the image sub-block is dynamically calculated based on the mean gray value, standard deviation of gray value, and preset weight coefficient of each image sub-block.

[0030] Using the calculated local segmentation threshold, the corresponding image sub-blocks are binarized, and pixels with gray values ​​higher than the threshold are classified as foreground, i.e., granular regions, while pixels with gray values ​​lower than the threshold are classified as background.

[0031] Traverse all image sub-blocks to complete the binarization of the entire image, resulting in a granular binarized image;

[0032] Edge detection operators are used to extract edge pixels of the particle region from the particle binarized image, and the edge pixels are connected in sequence to form the binarized contour of a single particle.

[0033] For each standardized particle grayscale image, the operations of segmenting it into multiple image sub-blocks, calculating local segmentation thresholds, performing binarization processing, and extracting and connecting the particle binarized contours are performed. The resulting binarized contours of all particles are then summarized to form the particle contour dataset.

[0034] Furthermore, the step of performing contour integrity restoration on the particle contour dataset based on morphological operations to eliminate contour breaks and holes caused by particle adhesion or image noise, and obtaining complete particle contour data, includes:

[0035] Read the binarized contour of a single particle from the particle contour dataset;

[0036] A morphological closing operation is performed on the binarized contour image. The morphological closing operation is a dilation followed by erosion operation, which is used to fill small holes inside the contour caused by noise and connect adjacent small breaks.

[0037] A morphological opening operation is performed on the image after the closing operation is performed. The morphological opening operation is erosion followed by dilation, which is used to smooth the contour boundaries and eliminate isolated noise points at the edges.

[0038] Edge extraction is performed again on the processed image to obtain a clearer, continuous and complete particle outline;

[0039] For cases where two or more particle profiles are connected into a single overall profile due to particle adhesion, calculate the geometric features of the overall profile, including the ratio of profile perimeter to area and the number of indentations.

[0040] When the geometric features meet the preset adhesion determination conditions, a segmentation operation is performed at the concave point of the contour to separate the adhered particle contours into independent individual particle contours.

[0041] All the independent and complete individual particle contour data that have been repaired and segmented are re-integrated as the complete particle contour data.

[0042] Furthermore, the step of mapping the repaired particle contour data to a preset three-dimensional particle space coordinate system and reconstructing the three-dimensional point cloud model of each particle based on the spatial distribution of the contour points includes:

[0043] Establish a three-dimensional particle space coordinate system with the center of the granulator outlet as the origin;

[0044] Obtain the position and attitude parameters of the camera relative to the three-dimensional particle space coordinate system corresponding to each acquisition angle;

[0045] For each contour point in the complete particle contour data, based on its acquisition angle, a spatial ray in the three-dimensional particle space coordinate system is calculated by back projection using the position and attitude parameters.

[0046] For the same particle, the contour points extracted from images at different angles are calculated, and the intersection point or nearest neighbor point of the corresponding spatial rays is calculated. The three-dimensional coordinates of the intersection point or nearest neighbor point are used as a three-dimensional point cloud data point on the particle surface.

[0047] The back projection and intersection calculation process is repeated by traversing the contour points of a particle in all acquired angle images to obtain multiple three-dimensional point cloud data points that constitute the surface of the particle.

[0048] Using a surface reconstruction algorithm, multiple three-dimensional point cloud data points belonging to the same particle are triangularly meshed to construct a three-dimensional closed surface model of the particle, i.e., the three-dimensional point cloud model.

[0049] Further, the step of calculating the equivalent volume of the 3D point cloud model of each particle, and converting the equivalent volume into the equivalent sphere diameter to generate a preliminary particle size distribution list includes:

[0050] For each reconstructed 3D point cloud model of a particle, calculate the total surface area of ​​all triangular faces in its triangular mesh model.

[0051] Using the voxelization method or the Simpson integral method, the spatial volume occupied by the particles is calculated based on the three-dimensional point cloud model, and used as the equivalent volume;

[0052] Assuming the particle is an ideal sphere, the equivalent sphere diameter can be calculated by back-calculating the equivalent volume using the formula for the volume of a sphere.

[0053] Arrange the calculated equivalent sphere diameters of all particles in ascending or descending order according to their diameter;

[0054] Count the number of particles in different diameter ranges and calculate the percentage of particles in each range to the total number of particles.

[0055] The sorted list of equivalent sphere diameters, each diameter range, and the corresponding number and percentage of particles are integrated to generate the preliminary particle size distribution list containing details of particle size distribution.

[0056] Furthermore, the step of simultaneously monitoring the current spindle speed and feeding rate of the granulator during the generation of the preliminary particle size distribution list includes:

[0057] The encoder installed on the main shaft of the granulator collects the rotation pulse signal of the main shaft in real time.

[0058] The spindle rotation pulse signal is counted and timed to calculate the number of pulses per unit time, and the real-time value of the spindle speed is obtained by combining the encoder resolution.

[0059] The mass or volume of material passing through per unit time is collected in real time by a weighing sensor or flow meter installed on the feeding device.

[0060] The collected mass or volume data is processed by time differentiation to calculate the real-time value of the feeding rate;

[0061] Establish a time synchronization mechanism to ensure that the real-time value of the spindle speed and the real-time value of the feeding rate have the same time reference;

[0062] The real-time values ​​of the spindle speed and the feeding rate are recorded and correlated with the generation time of the preliminary particle size distribution list.

[0063] Further, the step of marking the data in the preliminary particle size distribution list and generating particle size data with operating condition markings when the real-time ratio of the spindle speed to the feeding rate exceeds a preset stable production threshold range includes:

[0064] Read the preset stable production threshold range, which defines the upper and lower limits of the allowable ratio of spindle speed to feed rate;

[0065] The ratio of the real-time spindle speed to the real-time feed rate is calculated to obtain the real-time operating condition ratio.

[0066] Compare the real-time operating condition ratio with the upper and lower allowable limits;

[0067] If the real-time operating condition ratio is greater than the upper limit or less than the lower limit, it is determined that the current operating condition is unstable.

[0068] The moment when an unstable operating condition is determined is obtained, and the preliminary particle size distribution list data generated within the time and a preset delay window thereafter is retrieved.

[0069] Add an unstable operating condition marker to the preliminary particle size distribution list data. The unstable operating condition marker includes the operating condition type and timestamp.

[0070] For data where the real-time operating condition ratio is between the upper and lower limits of the allowable value, a stable operating condition marker is added;

[0071] All particle size data with added operating condition tags are aggregated to form the particle size data with operating condition tags.

[0072] Further, the step of distinguishing between stable and unstable operating condition data based on the particle size data with operating condition labels, and applying different data filtering algorithms to the data under the two operating conditions to remove abnormal particle size values, includes:

[0073] The particle size data with operating condition labels is analyzed, and the data is divided into stable operating condition datasets and unstable operating condition datasets according to the operating condition type in the labels.

[0074] For the stable operating condition dataset, a filtering algorithm based on statistical distribution is used to calculate the mean and standard deviation of all particle size values ​​in the stable operating condition dataset;

[0075] A multiple of the standard deviation is set as the stable operating condition filtering threshold. Data points whose particle size deviates from the mean by more than the stable operating condition filtering threshold are removed to complete the filtering of the stable operating condition dataset.

[0076] For the aforementioned unstable operating condition dataset, a median filtering algorithm based on a sliding window is adopted, with a fixed-size data sliding window set.

[0077] Slide the data sliding window in the unstable operating condition dataset, and take the particle size data in the window each time;

[0078] The particle size data within the window are sorted, and the median value is taken as the filtered value of the data at the center position of the sliding window. All data are processed in sequence to complete the filtering of the unstable working condition dataset.

[0079] The filtered stable operating condition dataset and the filtered unstable operating condition dataset are then merged again.

[0080] Furthermore, the present invention also includes a particle size detection system for a granulator, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the particle size detection method for a granulator as described above.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] During the generation of the preliminary particle size distribution list, the current spindle speed and feeding rate of the granulator are monitored simultaneously, and the real-time ratio is calculated. When this ratio exceeds the preset stable production threshold range, the system automatically marks the corresponding particle size data in the preliminary list, generating particle size data with operating condition markings. This technology associates each particle size data point with the instantaneous state during production, directly mapping fluctuations in equipment operating parameters to the detection results. The data marking process is completed automatically in real time, eliminating the need for manual comparison afterward, and establishing an immediate correspondence between particle size anomalies and operating condition anomalies. Through marking, subsequent processing can clearly identify which data originated from stable production conditions and which from unstable states, providing a fundamental data basis for accurate analysis.

[0083] Based on particle size data with operating condition labels, stable and unstable operating condition data are clearly distinguished. One filtering algorithm is applied to the stable operating condition data, and a different filtering algorithm is applied to the unstable operating condition data to remove outlier particle size values ​​from each type of data. Outliers in stable operating conditions may mainly originate from image noise or slight adhesion, while outliers in unstable operating conditions may contain truly formed outliers or serious measurement errors. Differentiated filtering is designed for the characteristics of different data sources, avoiding the insufficient adaptability of a single algorithm to complex situations. The filtered data are then combined to calculate the final average particle size, particle size distribution width, and pass rate indicators. The statistical results better reflect the true particle quality under stable production conditions, reducing the interference of operating condition fluctuations on the overall evaluation. Attached Figure Description

[0084] Figure 1 This is a flowchart illustrating the particle size detection method for the granulator described in this invention.

[0085] Figure 2 A flowchart illustrating the generation of a particle contour dataset using an adaptive threshold segmentation algorithm;

[0086] Figure 3 Scatter plot for structural element size correlation analysis during the particle contour repair stage of a granulator;

[0087] Figure 4 This is a 3D point cloud reconstruction result of granulator particles;

[0088] Figure 5 This is a composite data graph for monitoring the operating conditions of the granulator. Detailed Implementation

[0089] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] See Figure 1Multiple industrial cameras are deployed at the granulator outlet to simultaneously capture images of falling material particles from different angles, forming a multi-angle image sequence. This sequence is then converted into a particle contour dataset after grayscale processing. Next, morphological operations are applied to the particle contour dataset to repair breaks and fill voids, resulting in clear and complete particle contour data. Subsequently, these two-dimensional contour data are mapped to a preset three-dimensional particle space coordinate system according to camera calibration parameters, and a three-dimensional point cloud model corresponding to each particle is reconstructed using multi-view geometry principles. Based on each three-dimensional point cloud model, its equivalent volume is calculated and further converted into an equivalent sphere diameter, thereby generating a preliminary particle size distribution list containing all particle size information. Simultaneously, the system monitors and records the spindle speed and feeding rate during granulator operation. When the real-time ratio of spindle speed to feeding rate deviates from a preset stable production threshold range, the system adds a specific marker to the data in the preliminary particle size distribution list generated for that period, thus forming particle size data with operating condition markers. Next, based on the labels, the data is divided into two categories: stable operating condition data and unstable operating condition data. Different data filtering algorithms are then applied to each category to remove outlier particle size values. Finally, the two categories of filtered data are merged, and the final average particle size, particle size distribution width, and product qualification rate are calculated. A formatted particle size detection report is then output.

[0091] In one embodiment of the present invention, a multi-angle image sequence of material particles at the discharge port of a granulator is acquired. This multi-angle image sequence includes particle contour information under different lighting conditions. For example, three industrial cameras can be evenly arranged around the discharge port, with each camera installed at a 120-degree interval and equipped with auxiliary light sources at different angles. This allows for the synchronous acquisition of an image sequence containing information about the sides and bottom of the particles. The multi-angle image sequence is then processed into grayscale images to obtain a set of grayscale particle images. Each image in the set corresponds to a different acquisition angle. The grayscale processing aims to remove color information while retaining crucial brightness information to simplify subsequent calculations.

[0092] Each frame of the original color image is read from the multi-angle image sequence. For each frame, the original color image is converted from the RGB color space to the YCbCr color space, and the luminance component Y in the YCbCr color space is extracted as the initial grayscale image of each frame. In specific implementations, the conversion from RGB color space to YCbCr color space follows the International Telecommunication Union (ITU) standard, and the calculation of the luminance component Y is based on the weighted sum of the red, green, and blue channels. Local contrast enhancement is performed on the initial grayscale image to enhance the distinction between grain edges and the background, resulting in an enhanced grayscale image. One enhancement method is to calculate the grayscale value distribution in the neighborhood of each pixel in the image and apply a contrast stretching function. Luminance normalization is performed on all enhanced grayscale images to eliminate the overall brightness deviation caused by differences in illumination from different angles, forming a standardized grainy grayscale image. Luminance normalization adjusts the mean of the grayscale histogram of all images to a preset reference value.

[0093] All standardized grainy grayscale images, after brightness normalization, are sorted and integrated according to their corresponding acquisition timestamps and angle codes to form the grainy grayscale image set. In some embodiments, the acquisition timestamps are generated by a unified system clock, and the angle codes are associated with the physical installation location identifier of the industrial camera. The integration process establishes an image data stack indexed by time and angle. The grainy grayscale image set is processed by an adaptive threshold segmentation algorithm to extract the binary contours of the grains in each image, forming a grain contour dataset. The adaptive threshold segmentation algorithm dynamically determines the segmentation threshold based on the local characteristics of different regions of the image. For each standardized grainy grayscale image in the grainy grayscale image set, it is segmented into multiple non-overlapping image sub-blocks. For example, a 1024x1024 pixel image is divided into 256 64x64 pixel image sub-blocks. The grayscale mean and grayscale standard deviation of the pixels in each image sub-block are calculated. The grayscale mean reflects the overall brightness level of the region, and the grayscale standard deviation reflects the contrast strength of the region.

[0094] Based on the mean gray value, standard deviation of gray value, and preset weight coefficient of each image sub-block, the local segmentation threshold of the image sub-block is dynamically calculated. One calculation method can be expressed as follows:

[0095]

[0096] in: Represents the local segmentation threshold. Represents the average gray level of a sub-region of the image. The standard deviation of gray levels representing a sub-block of an image. This represents a preset weight coefficient. Using the calculated local segmentation threshold, the corresponding image sub-blocks are binarized. Pixels with grayscale values ​​higher than the threshold are classified as foreground (i.e., particle regions), and pixels with grayscale values ​​lower than the threshold are classified as background. All image sub-blocks are traversed to complete the binarization of the entire image, resulting in a particle binarized image. Since each image sub-block uses an independently calculated segmentation threshold, this process can adapt to uneven lighting or varying background complexity in the image. An edge detection operator is used to extract edge pixels from the particle binarized image, and these edge pixels are connected sequentially to form the binarized contour of a single particle. For each standardized particle grayscale image, the operations of segmenting into multiple image sub-blocks, calculating local segmentation thresholds, performing binarization, and extracting and connecting particle binarized contours are performed. All obtained particle binarized contours are summarized to form the particle contour dataset, which stores the geometric information of each contour in the form of a coordinate point sequence.

[0097] See Figure 2 In one embodiment of the present invention, for each standardized grainy grayscale image in the grainy grayscale image set, it is divided into multiple non-overlapping image sub-blocks. The size of the image sub-blocks is set based on the resolution of the standardized grainy grayscale image and the expected physical size of the grains. For example, a standardized grainy grayscale image of 2048 pixels by 2048 pixels is divided into 1024 image sub-blocks, each containing 64 pixels by 64 pixels. The grayscale mean and grayscale standard deviation of the pixels in each image sub-block are calculated. The grayscale mean is obtained by summing the grayscale values ​​of all pixels in the image sub-block and dividing by the total number of pixels. The grayscale standard deviation is obtained by taking the square root of the average of the sum of the squares of the differences between the grayscale values ​​of each pixel and the grayscale mean. The local segmentation threshold of an image sub-block is dynamically calculated based on the mean gray level, standard deviation gray level, and preset weight coefficients. The dynamic calculation process is achieved by adding the products of the mean gray level and the standard deviation gray level multiplied by the weight coefficient. For example, for an image sub-block, if the mean gray level is 150, the standard deviation gray level is 20, and the weight coefficient is 0.3, then the local segmentation threshold is 156. The calculated local segmentation threshold is used to binarize the corresponding image sub-block. Pixels with gray levels higher than the local segmentation threshold are classified as foreground (i.e., grain regions), and pixels with gray levels lower than the local segmentation threshold are classified as background. The binarization of the entire standardized grain gray-level image is completed by traversing all image sub-blocks, resulting in a grain binarized image. In the grain binarized image, white pixel regions represent grains, while black pixel regions represent the background.

[0098] In some embodiments, the image sub-blocks are divided using a fixed grid segmentation method. The number of rows and columns of the grid is determined based on the divisibility between the total image size and the sub-block size. It is understood that the fixed grid segmentation method is easy to implement and computationally efficient. Optionally, the image sub-blocks can also be divided using an adaptive-size region growing method, where the seed point selection is based on the image grayscale gradient. Edge detection operators are used to extract edge pixels from the particle binarized image. Edge detection operators, such as the Canny operator, locate edge pixels by calculating the image grayscale gradient and applying dual thresholds. The edge pixels are then connected sequentially to form the binarized contour of a single particle. The connection process uses an edge tracking algorithm to search for adjacent edge points along the gradient direction from the initial edge point until a closed loop is formed. For each standardized particle grayscale image, the operations of segmenting into multiple image sub-blocks, calculating local segmentation thresholds, performing binarization processing, and extracting and connecting the particle binarized contours are performed. The resulting binarized contours of all particles are summarized to form a particle contour dataset. The particle contour dataset stores the pixel coordinate sequence of each contour in an array structure. In some embodiments, the particle contour dataset also includes the minimum bounding rectangle information of each contour for fast spatial lookup. It is understood that the minimum bounding rectangle information of the contour helps to quickly locate the position of the particle in the image in subsequent processing. Optionally, contour area filtering is applied after extracting the binarized contours to remove contours with an area smaller than a preset threshold to eliminate the influence of image noise.

[0099] In one embodiment of the present invention, the specific process of performing contour integrity repair on a particle contour dataset based on morphological operations to obtain complete particle contour data involves reading the binary contour of a single particle in the particle contour dataset. The binary contour exists in the form of a binary image where white pixels represent the contour line and black pixels represent the background. Morphological closing operations are performed on the binary contour image. Morphological closing operations are a composite operation of dilation followed by erosion. The dilation operation uses a structuring element of a preset shape and size, such as a square structuring element with a side length of 3 pixels. It traverses each pixel of the image and aligns the origin of the structuring element with the current pixel. If there are white pixels within the coverage area of ​​the structuring element, the current pixel is output as white. The erosion operation uses the same structuring element to traverse the image. The current pixel is output as white only when all pixels within the coverage area of ​​the structuring element are white. The morphological closing operation is used to fill small holes inside the contour caused by noise and connect adjacent small fractures. After the closing operation is performed, a morphological opening operation is performed on the image. The morphological opening operation is a composite operation of erosion and dilation. It also uses a structuring element, such as a circular structuring element with a radius of 2 pixels. The erosion operation is performed first to remove isolated noise points at the edges, and then the dilation operation is performed to smooth the contour boundaries. Edge extraction is performed again on the processed image to obtain a clearer, continuous and complete grain contour. Edge extraction can use the Sobel operator to calculate the image gradient and track the contour points.

[0100] For cases where two or more particle contours merge into a single contour due to particle adhesion, the geometric features of the overall contour are calculated. These features include the contour perimeter-to-area ratio and the number of indentations. The contour perimeter is obtained by summing the distances between adjacent pixels on the contour, the contour area is obtained by counting the number of pixels inside the contour, and the number of indentations is determined by calculating the difference between the contour's convex hull and the original contour and finding the point of maximum local curvature. In some embodiments, the adhesion determination criteria are set to a contour perimeter squared-to-area ratio greater than a preset threshold and the number of indentations greater than or equal to two. It can be understood that adhered particle contours typically have a higher perimeter-to-area ratio and exhibit obvious indentation features. When the geometric features meet the preset adhesion determination criteria, a segmentation operation is performed at the indentations of the contour to separate the adhered particle contours into independent individual particle contours. In practice, segmentation is achieved by finding pairs of contour depressions and connecting them with a dividing line. For example, for two adhered spherical particles, a narrow region is formed at the point of adhesion, and the contour points on both sides of this region are the depressions. The straight line connecting these two depressions can serve as the dividing line. Optionally, the dividing line can also be the optimal path obtained based on contour local curvature analysis. All the independent and complete individual particle contour data after repair and segmentation are re-integrated as complete particle contour data. The complete particle contour data is stored in the form of a list of coordinate points, with each list representing the closed boundary of an independent particle. In some embodiments, the size of the structuring element in the morphological operation is related to the image resolution and the minimum expected particle size. The formula is:

[0101]

[0102] in: Represents the side length or radius of the structural element. This represents the minimum expected physical diameter of the particle known through process knowledge. Represents the spatial resolution of an image, symbol Representing the rounding up operation, this association ensures that morphological operations can specifically process image features that match the particle size, while avoiding over-smoothing or eliminating effective contour details. Optionally, multiple sets of structural element size parameters can be preset for particle products of different size ranges.

[0103] See Figure 3This is a scatter plot showing the correlation between structural element dimensions and particle contour repair in a granulator. As the minimum expected physical diameter of the particles increases, the image spatial resolution also increases, showing a positive correlation. When the particle diameter is smaller and the image resolution is lower, the structural element size is closer to 1.30; conversely, when the particle diameter is larger and the image resolution is higher, the structural element size is closer to 1.00. This result perfectly aligns with the process logic of particle contour repair in granulators: smaller particles require finer structural elements to preserve contour details, while larger particles can be handled with coarser structural elements. This plot visually verifies the effectiveness of the structural element size formula, providing a precise basis for selecting structural element parameters for combinations of different particle sizes and image resolutions. It ensures that morphological operations can effectively repair contour voids and breaks without over-smoothing or losing effective contour details, making it a key process parameter optimization tool in granulator particle size detection.

[0104] In one embodiment of the present invention, a three-dimensional particle space coordinate system is established with the center of the granulator outlet as the origin. The Z-axis of the three-dimensional particle space coordinate system is parallel to the direction of gravity, and the X-axis and Y-axis are perpendicular to each other in the horizontal plane. The position parameters and attitude parameters of the camera relative to the three-dimensional particle space coordinate system corresponding to each acquisition angle are obtained. The position parameters are obtained by calibrating the correspondence between the world coordinates and image coordinates at the outlet using a three-dimensional calibration plate. The attitude parameters are expressed in the form of rotation matrix and translation vector. For each contour point in the complete particle contour data, the contour point is stored in the form of pixel coordinates (u,v). According to its acquisition angle, using the position parameters and attitude parameters, a spatial ray of the contour point in the three-dimensional particle space coordinate system is calculated by back projection. The back projection calculation involves the camera intrinsic parameter matrix, rotation matrix, and translation vector, transforming the pixel coordinates to the normalized camera coordinate system and then to the three-dimensional particle space coordinate system. For the same particle extracted from images at different angles, the intersection or nearest neighbor of the corresponding spatial rays is calculated. In practice, when two spatial rays should theoretically intersect at a single point, due to noise and calibration errors, the least squares method is used to find the coordinates of the 3D point that minimizes the sum of the squared distances of all corresponding rays. The 3D coordinates of the intersection or nearest neighbor are then used as a 3D point cloud data point on the particle surface. The back-projection and intersection calculation process is repeated for the contour points of a particle in all acquired angle images, resulting in multiple 3D point cloud data points constituting the particle surface. A particle surface typically consists of hundreds to thousands of 3D point cloud data points.

[0105] Using surface reconstruction algorithms, multiple 3D point cloud data points belonging to the same particle are triangularly meshed to construct a 3D closed surface model of the particle, i.e., a 3D point cloud model. Surface reconstruction algorithms, such as Poisson surface reconstruction, convert the point cloud into a triangular mesh model with normal vectors by solving implicit functions. For each reconstructed 3D point cloud model of a particle, the total surface area of ​​all triangular facets in its triangular mesh model is calculated. The total surface area is obtained by summing the areas of each triangular facet, which is calculated based on the coordinates of its three vertices using the modulus of the cross product. Using voxelization or Simpson's integral method, the spatial volume occupied by the particle is calculated based on the 3D point cloud model as the equivalent volume. In specific implementations, the voxelization method divides the bounding box containing the particle model into a fine cubic grid, and the equivalent volume is obtained by counting the number of voxels belonging to the particle and multiplying them by the volume of a single voxel. Assuming the particle is an ideal sphere, the equivalent sphere diameter is calculated by inversely from the equivalent volume according to the formula for the volume of a sphere. The formula for calculating the equivalent sphere diameter is:

[0106]

[0107] in: Represents the equivalent sphere diameter. Represents the equivalent volume of particles calculated from a 3D point cloud model, denoted by [symbol]. Representing pi. The equivalent sphere diameters of all calculated particles are sorted in ascending or descending order to generate an ordered diameter list. The number of particles within different diameter ranges is counted, and the percentage of particles in each range relative to the total number of particles is calculated. The division of diameter ranges can be set according to process requirements, for example, with intervals of 0.1 mm. The sorted equivalent sphere diameter list, each diameter range, and its corresponding particle count and percentage information are integrated to generate a preliminary particle size distribution list containing detailed particle size distribution information. See Table 1 for an example snippet of a preliminary particle size distribution list.

[0108] Table 1: Preliminary Particle Size Distribution List

[0109]

[0110] In some embodiments, the Simpson integral method is used to calculate the equivalent volume. It is understood that the Simpson integral method estimates the volume by parameterizing the particle surface and performing numerical integration over the parameter domain. Optionally, during triangulation, the closure and manifold of the mesh are ensured to meet the requirements of volume calculation. It is understood that the conversion of the equivalent sphere diameter provides a single characteristic size equivalent to the particle volume, facilitating statistical analysis and process benchmarking. In some embodiments, in addition to the equivalent sphere diameter, the preliminary particle size distribution list may also include other geometric features calculated based on the 3D point cloud model, such as the maximum Feret diameter or surface area-to-volume ratio. Optionally, the list integration process also includes associating each particle entry with its unique identifier and the storage path of the 3D point cloud model.

[0111] See Figure 4 This is a 3D point cloud reconstruction result of granulator particles, showing the 3D spatial distribution of five particles with different sizes. The point cloud is distributed in a spherical shape, with the point cloud of each particle roughly clustered around its own center, consistent with the 3D morphology of real particles. The particles maintain independent spatial positions without obvious adhesion, indicating that the initial contour repair and adhesion segmentation processes were effective. This is a key output of the 3D reconstruction stage in the granulator particle size detection process. The point cloud data can be directly used to calculate core indicators such as the equivalent volume and equivalent diameter of the particles. By comparing the point cloud density and spatial distribution of different particles, the accuracy of the 3D reconstruction algorithm can be verified, providing a reliable data foundation for subsequent particle size distribution statistics. By observing whether the shape of the point cloud is regular, and whether there are any deformities or damages, the discharge status of the granulator can be directly reflected, such as whether abnormal temperature or pressure has caused particle morphology defects.

[0112] In one embodiment of the invention, an encoder installed on the main shaft of the granulator collects the rotation pulse signal of the main shaft in real time. The encoder generates a fixed number of pulses per revolution, such as 1024 pulses. The main shaft rotation pulse signal is counted and timed to calculate the number of pulses per unit time. Combined with the encoder resolution, the real-time value of the main shaft speed is obtained. The unit time is set to one second, and the encoder resolution is 1024 pulses per revolution. If 20480 pulses are counted in one second, the real-time value of the main shaft speed is 20 revolutions per second. A weighing sensor or flow meter installed on the feeding device collects the mass or volume of material passing through per unit time in real time. For example, the weighing sensor collects the cumulative mass value of the material at 0.1-second intervals. The collected mass or volume data is processed by time differentiation to calculate the real-time value of the feeding rate. Time differentiation is achieved by calculating the mass difference between adjacent sampling times and dividing by the time interval. If the mass increases by 5 kg in 0.1 seconds, the real-time value of the feeding rate is 50 kg per second. A time synchronization mechanism is established to ensure that the real-time values ​​of the spindle speed and the feeding rate have the same time reference. This mechanism is achieved by sending synchronization trigger signals from the system master clock to the encoder counter and the weighing sensor data acquisition card. The real-time values ​​of the spindle speed and feeding rate are recorded in real time and correlated with the generation time of the preliminary particle size distribution list. The timestamp accuracy is at the millisecond level, ensuring that each set of particle size data corresponds to a set of real-time spindle speed and feeding rate.

[0113] The system reads a preset stable production threshold range, which defines the upper and lower limits of the allowable ratio of spindle speed to feed rate. For example, the upper limit is set to 1.1 and the lower limit to 0.9. The system calculates the real-time ratio of the spindle speed to the feed rate to obtain the real-time operating condition ratio. If the real-time spindle speed is 20 revolutions per second and the real-time feed rate is 20 kilograms per second, the real-time operating condition ratio is 1.0. The system compares the real-time operating condition ratio with the allowable upper and lower limits. If the real-time operating condition ratio is greater than the allowable upper limit or less than the allowable lower limit, the system determines that it is currently in an unstable operating condition. For example, if the real-time operating condition ratio reaches 1.15, which is greater than the allowable upper limit of 1.1, it is determined to be in an unstable operating condition. The system obtains the moment when the system is determined to be in an unstable operating condition and retrieves the preliminary particle size distribution list data generated within a preset delay window at or after that moment. The preset delay window is used to compensate for the transmission time of the material from the granulation chamber to the image acquisition position, for example, it is set to 2 seconds. An unstable operating condition marker is added to the preliminary particle size distribution list data. This marker includes the operating condition type and a timestamp. The operating condition type is indicated by a flag "0", signifying an unstable operating condition, and the timestamp records the precise moment the unstable operating condition was determined. For data where the real-time operating condition ratio falls between the upper and lower allowable limits, a stable operating condition marker is added, indicated by a flag "1". All particle size data with added operating condition markers are then aggregated to form the particle size data with operating condition markers. This data is a list containing the equivalent sphere diameter, the operating condition marker, and the timestamp.

[0114] The particle size data with operating condition labels is analyzed. Based on the operating condition type in the labels, the data is divided into stable operating condition datasets and unstable operating condition datasets. The stable operating condition dataset contains all particle size values ​​with a flag of "1", while the unstable operating condition dataset contains all particle size values ​​with a flag of "0". For the stable operating condition dataset, a statistical distribution-based filtering algorithm is used. The mean and standard deviation of all particle size values ​​in the stable operating condition dataset are calculated. A multiple of the standard deviation is set as the stable operating condition filtering threshold. Data points whose particle size values ​​deviate from the mean by more than the stable operating condition filtering threshold are removed, thus completing the filtering of the stable operating condition dataset. The stable operating condition filtering threshold is defined as follows. The calculation can be expressed as:

[0115]

[0116] in: Represents the filtering threshold under stable operating conditions. This represents the arithmetic mean of all particle size values ​​in the stable operating condition dataset. This represents the standard deviation of all particle size values ​​in the steady-state dataset. Represents a preset multiple coefficient. For the non - stable working condition data set, a median filtering algorithm based on a sliding window is adopted. A data sliding window with a fixed size is set. For example, the window size is 5 data points. Slide the data sliding window in the non - stable working condition data set. Each time, take the particle size data within the window, sort the particle size data within the window, and take the median as the filtered value of the data at the center position of the data sliding window. Process all the data in turn to complete the filtering of the non - stable working condition data set. In some embodiments, the moving step of the sliding window is 1. After each slide, the oldest data point within the window is removed, and the latest data point is added. It can be understood that median filtering can effectively suppress the pulse - type outliers that may appear under non - stable working conditions. Re - merge the filtered stable working condition data set and the filtered non - stable working condition data set. The merging operation rearranges the data points of the two data sets in the original time order of the data.

[0117] Refer to Figure 5 , this is a composite data graph of the working condition monitoring of a granulator, showing the real - time联动变化(这里原中文表述有误,推测是“联动变化”,英文可译为“linkage change”) of the main shaft speed and the feeding rate within 60 seconds, combined with the business background of the particle size detection of the granulator. It intuitively presents the production stability of the granulator, and can quickly locate the occurrence period of non - stable working conditions. It provides a working condition basis for the reliability of the particle size detection data, and avoids misjudging the abnormal particle size values under non - stable working conditions as normal production data. The synchronous fluctuation characteristics of the curve verify the rationality of the linkage control design of the main shaft speed and the feeding rate of the granulator. The synchronous fluctuation characteristics of the curve verify the linkage control logic of the main shaft speed and the feeding rate. If the fluctuation periods of the two do not match, the abnormal equipment control can be detected in time. By accumulating such curves in the long term, the optimal matching interval of the main shaft speed and the feeding rate can be analyzed, providing a quantitative basis for the iterative optimization of the granulation process parameters.

[0118] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for particle size detection in a granulator, characterized in that, include: Collect multi-angle image sequences of material particles at the discharge port of the granulator, and perform grayscale processing to form a particle contour dataset; Based on morphological operations, contour integrity restoration is performed on the particle contour dataset to eliminate contour breaks and holes caused by particle adhesion or image noise, and obtain complete particle contour data. The repaired particle contour data is mapped to a preset three-dimensional particle space coordinate system, and the three-dimensional point cloud model of each particle is reconstructed according to the spatial distribution of the contour points. Calculate the equivalent volume of the 3D point cloud model of each particle, and convert it into the equivalent sphere diameter based on the equivalent volume to generate a preliminary particle size distribution list. During the generation of the preliminary particle size distribution list, the current spindle speed and feeding rate of the granulator are monitored simultaneously. When the real-time ratio of the spindle speed to the feeding rate exceeds the preset stable production threshold range, the data in the preliminary particle size distribution list is marked to generate particle size data with working condition markings. Based on the particle size data with operating condition markings, stable operating condition data and unstable operating condition data are distinguished, and different data filtering algorithms are applied to the data under the two operating conditions to remove abnormal particle size values. The filtered stable and unstable operating condition data are merged, and the final average particle size, particle size distribution width, and pass rate are calculated to output a particle size test report.

2. The particle size detection method for a granulator according to claim 1, characterized in that, The formation of the particle contour dataset includes: Collect a multi-angle image sequence of material particles at the discharge port of the granulator, wherein the multi-angle image sequence contains particle outline information under different lighting conditions; The multi-angle image sequence is converted to grayscale to obtain a set of granular grayscale images, where each image in the set of granular grayscale images corresponds to a sampling angle. The set of grayscale particle images is processed by an adaptive threshold segmentation algorithm to extract the binary contours of particles in each image, forming a particle contour dataset. The step of performing grayscale processing on the multi-angle image sequence to obtain a set of granular grayscale images includes: Read each frame of the original color image from the multi-angle image sequence; For each frame of the original color image, convert it from the RGB color space to the YCbCr color space; Extract the luminance component Y from the YCbCr color space as the initial grayscale image of each frame of the original color image; The initial grayscale image is locally contrast-enhanced to improve the distinction between particle edges and the background, resulting in an enhanced grayscale image. Brightness normalization is performed on all enhanced grayscale images to eliminate the overall brightness deviation caused by differences in lighting from different angles, forming a standardized grainy grayscale image. All standardized grain grayscale images that have undergone brightness normalization are sorted and integrated according to their corresponding acquisition timestamps and angle codes to form the grain grayscale image set.

3. The particle size detection method for a granulator according to claim 2, characterized in that, The step of processing the set of grayscale images of particles using an adaptive threshold segmentation algorithm to extract the binarized contours of particles in each image and form a particle contour dataset includes: For each standardized grainy grayscale image in the set of grainy grayscale images, it is divided into multiple non-overlapping image sub-blocks; Calculate the mean gray value and standard deviation of the pixels within each image sub-block; The local segmentation threshold of the image sub-block is dynamically calculated based on the mean gray value, standard deviation of gray value, and preset weight coefficient of each image sub-block. Using the calculated local segmentation threshold, the corresponding image sub-blocks are binarized, and pixels with gray values ​​higher than the threshold are classified as foreground, i.e., granular regions, while pixels with gray values ​​lower than the threshold are classified as background. Traverse all image sub-blocks to complete the binarization of the entire image, resulting in a granular binarized image; Edge detection operators are used to extract edge pixels of the particle region from the particle binarized image, and the edge pixels are connected in sequence to form the binarized contour of a single particle. For each standardized particle grayscale image, the operations of segmenting it into multiple image sub-blocks, calculating local segmentation thresholds, performing binarization processing, and extracting and connecting the particle binarized contours are performed. The resulting binarized contours of all particles are then summarized to form the particle contour dataset.

4. The particle size detection method for a granulator according to claim 1, characterized in that, The step of performing contour integrity restoration on the particle contour dataset based on morphological operations to eliminate contour breaks and holes caused by particle adhesion or image noise, and obtaining complete particle contour data, includes: Read the binarized contour of a single particle from the particle contour dataset; A morphological closing operation is performed on the binarized contour image. The morphological closing operation is a dilation followed by erosion operation, which is used to fill small holes inside the contour caused by noise and connect adjacent small breaks. A morphological opening operation is performed on the image after the closing operation is performed. The morphological opening operation is erosion followed by dilation, which is used to smooth the contour boundaries and eliminate isolated noise points at the edges. Edge extraction is performed again on the processed image to obtain a clearer, continuous and complete particle outline; For cases where two or more particle profiles are connected into a single overall profile due to particle adhesion, calculate the geometric features of the overall profile, including the ratio of profile perimeter to area and the number of indentations. When the geometric features meet the preset adhesion determination conditions, a segmentation operation is performed at the concave point of the contour to separate the adhered particle contours into independent individual particle contours. All the independent and complete individual particle contour data that have been repaired and segmented are re-integrated as the complete particle contour data.

5. The particle size detection method for a granulator according to claim 1, characterized in that, The step of mapping the repaired particle contour data to a preset three-dimensional particle space coordinate system and reconstructing the three-dimensional point cloud model of each particle based on the spatial distribution of the contour points includes: Establish a three-dimensional particle space coordinate system with the center of the granulator outlet as the origin; Obtain the position and attitude parameters of the camera relative to the three-dimensional particle space coordinate system corresponding to each acquisition angle; For each contour point in the complete particle contour data, based on its acquisition angle, a spatial ray in the three-dimensional particle space coordinate system is calculated by back projection using the position and attitude parameters. For the same particle, the contour points extracted from images at different angles are calculated, and the intersection point or nearest neighbor point of the corresponding spatial rays is calculated. The three-dimensional coordinates of the intersection point or nearest neighbor point are used as a three-dimensional point cloud data point on the particle surface. The back projection and intersection calculation process is repeated by traversing the contour points of a particle in all acquired angle images to obtain multiple three-dimensional point cloud data points that constitute the surface of the particle. Using a surface reconstruction algorithm, multiple three-dimensional point cloud data points belonging to the same particle are triangularly meshed to construct a three-dimensional closed surface model of the particle, i.e., the three-dimensional point cloud model.

6. The particle size detection method for a granulator according to claim 1, characterized in that, The steps of calculating the equivalent volume of the 3D point cloud model of each particle, converting the equivalent volume into the equivalent sphere diameter, and generating a preliminary particle size distribution list include: For each reconstructed 3D point cloud model of a particle, calculate the total surface area of ​​all triangular faces in its triangular mesh model. Using the voxelization method or the Simpson integral method, the spatial volume occupied by the particles is calculated based on the three-dimensional point cloud model, and used as the equivalent volume; Assuming the particle is an ideal sphere, the equivalent sphere diameter can be calculated by back-calculating the equivalent volume using the formula for the volume of a sphere. Arrange the calculated equivalent sphere diameters of all particles in ascending or descending order according to their diameter; Count the number of particles in different diameter ranges and calculate the percentage of particles in each range to the total number of particles. The sorted list of equivalent sphere diameters, each diameter range, and the corresponding number and percentage of particles are integrated to generate the preliminary particle size distribution list containing details of particle size distribution.

7. The particle size detection method for a granulator according to claim 1, characterized in that, The step of simultaneously monitoring the current spindle speed and feeding rate of the granulator during the generation of the preliminary particle size distribution list includes: The encoder installed on the main shaft of the granulator collects the rotation pulse signal of the main shaft in real time. The spindle rotation pulse signal is counted and timed to calculate the number of pulses per unit time, and the real-time value of the spindle speed is obtained by combining the encoder resolution. The mass or volume of material passing through per unit time is collected in real time by a weighing sensor or flow meter installed on the feeding device. The collected mass or volume data is processed by time differentiation to calculate the real-time value of the feeding rate; Establish a time synchronization mechanism to ensure that the real-time value of the spindle speed and the real-time value of the feeding rate have the same time reference; The real-time values ​​of the spindle speed and the feeding rate are recorded and correlated with the generation time of the preliminary particle size distribution list.

8. The particle size detection method for a granulator according to claim 1, characterized in that, The step of marking the data in the preliminary particle size distribution list and generating particle size data with operating condition markings when the real-time ratio of the spindle speed to the feeding rate exceeds the preset stable production threshold range includes: Read the preset stable production threshold range, which defines the upper and lower limits of the allowable ratio of spindle speed to feed rate; The ratio of the real-time spindle speed to the real-time feed rate is calculated to obtain the real-time operating condition ratio. Compare the real-time operating condition ratio with the upper and lower allowable limits; If the real-time operating condition ratio is greater than the upper limit or less than the lower limit, it is determined that the current operating condition is unstable. The moment when an unstable operating condition is determined is obtained, and the preliminary particle size distribution list data generated within the time and a preset delay window thereafter is retrieved. Add an unstable operating condition marker to the preliminary particle size distribution list data. The unstable operating condition marker includes the operating condition type and timestamp. For data where the real-time operating condition ratio is between the upper and lower limits of the allowable value, a stable operating condition marker is added; All particle size data with added operating condition tags are aggregated to form the particle size data with operating condition tags.

9. The particle size detection method for a granulator according to claim 1, characterized in that, The step of distinguishing between stable and unstable operating condition data based on the particle size data with operating condition markings, and applying different data filtering algorithms to remove abnormal particle size values ​​for the two types of data respectively, includes: The particle size data with operating condition labels is analyzed, and the data is divided into stable operating condition datasets and unstable operating condition datasets according to the operating condition type in the labels. For the stable operating condition dataset, a filtering algorithm based on statistical distribution is used to calculate the mean and standard deviation of all particle size values ​​in the stable operating condition dataset; A multiple of the standard deviation is set as the stable operating condition filtering threshold. Data points whose particle size deviates from the mean by more than the stable operating condition filtering threshold are removed to complete the filtering of the stable operating condition dataset. For the aforementioned unstable operating condition dataset, a median filtering algorithm based on a sliding window is adopted, with a fixed-size data sliding window set. Slide the data sliding window in the unstable operating condition dataset, and take the particle size data in the window each time; The particle size data within the window are sorted, and the median value is taken as the filtered value of the data at the center position of the sliding window. All data are processed in sequence to complete the filtering of the unstable working condition dataset. The filtered stable operating condition dataset and the filtered unstable operating condition dataset are then merged again.

10. A particle size detection system for a granulator, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the particle size detection method for a granulator as described in any one of claims 1 to 9.

Images