Size detection method and system for dynamic screening of glass fiber particles based on machine vision

Abstract

This invention discloses a machine vision-based method and system for dynamic screening and size detection of glass fiber particles, specifically relating to the field of glass fiber particle size detection. The method includes real-time monitoring of particle adhesion and dynamic adjustment of the vibration dispersing mechanism, further integrating vibration signals and image features for analysis, and correcting motion blur and refractive index distortion in the image to accurately measure the physical size of the particles. The machine vision-based method and system for dynamic screening and size detection of glass fiber particles achieves coordinated control of vibration and imaging by dynamically adjusting vibration parameters through the adhesion index, effectively solving the problem of lack of coordination between the vibration dispersing mechanism and the imaging mechanism. Through distortion correction algorithms, it compensates for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index, reducing measurement distortion. Finally, through a reinforcement learning model, it minimizes size measurement errors while ensuring high dispersion.

Description

Technical Field

[0001] This invention relates to the field of glass fiber particle size detection technology, and more specifically, to a size detection method and system for dynamic screening of glass fiber particles based on machine vision. Background Technology

[0002] With the increasing demands for quality in the composite materials industry, the accuracy of online dynamic detection of glass fiber particle size is particularly important. Traditional technology uses an offline detection architecture that combines offline static sampling with optical microscopy. This involves pausing the production line, manually placing samples, and collecting static images to achieve size analysis. However, traditional methods suffer from technical bottlenecks such as secondary particle adhesion leading to insufficient effective monomer rate and low detection efficiency, and manual intervention causing a mismatch between the detection cycle and the production line speed. These methods cannot guarantee real-time quality monitoring for continuous production.

[0003] To address the inefficiency of offline inspection, existing technologies employ a synchronously triggered high-speed industrial camera linked with a conveyor belt encoder to capture motion sequence images of glass fiber particles during continuous production line operation. Motion compensation algorithms are introduced to eliminate displacement ambiguity, enabling continuous sampling of moving targets. By replacing manual sieving with an automated dispersing device and dynamic imaging with static acquisition, integrated inspection of the production line has been initially achieved.

[0004] However, in actual use, it still has some shortcomings, such as the lack of coordinated control between the vibration dispersing mechanism and the imaging mechanism, which leads to insufficient breaking of the adhering particles, directly causing the failure of individual imaging, resulting in size measurement distortion and increased sorting misjudgment rate. Especially in high-throughput production scenarios, it is difficult to balance the contradiction between particle dispersion and imaging clarity. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides a size detection method and system for dynamic screening of glass fiber particles based on machine vision, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A machine vision-based method for dynamic screening of glass fiber particles includes:

[0008] S1: Extract the first glass fiber particle feature in real time at the outlet of the target glass fiber particle flow vibration and dispersion, and output the adhesion index corresponding to the first glass fiber particle feature.

[0009] S2: The vibration dispersing mechanism is dynamically adjusted according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and the image of the first glass fiber particles is acquired. The timing of the first glass fiber particle image is synchronously triggered according to the vibration dispersing mechanism.

[0010] S3: Extract the features of the second glass fiber particles from the first glass fiber particle image by time-frequency analysis and convolutional neural network fusion;

[0011] S4: Analyze the features of the second glass fiber particles to obtain an image of the second glass fiber particles;

[0012] S5: Perform a distortion correction mechanism on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index.

[0013] S6: Based on the third glass fiber particle image, determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image;

[0014] S7: Dynamically update the vibration disintegration mechanism and distortion correction mechanism through reinforcement learning model.

[0015] Preferably, in step S1, the first glass fiber particle feature includes the area of ​​the glass fiber particle cluster convex hull. Minimum circumscribed rectangle area Number of contour dimples and the number of outline pixels .

[0016] Preferably, in step S1, based on the characteristics of the first glass fiber particles, an adhesion index is calculated and output. Specifically, it is expressed as:

[0017] ,

[0018] in, Represented as area feature weights, It is represented as the concave feature weight.

[0019] Preferably, in step S2, the vibration dispersing mechanism is implemented by an electromagnetic vibrator, specifically including:

[0020] When the adhesion index is greater than 0.6, the amplitude of the electromagnetic vibrator is increased to twice the preset amplitude reference value and the vibration frequency is reduced to 10% of the preset lower limit value.

[0021] When the adhesion index is less than 0.3, the amplitude will be reduced to 10% of the preset amplitude reference value and the vibration frequency will be increased to 10% of the preset frequency upper limit value.

[0022] Preferably, in step S3, the second glass fiber particle characteristics include real-time vibration velocity vector, main frequency band energy ratio, frequency centroid, frequency band entropy value, imaging exposure time, glass fiber surface defects, and adhesion type.

[0023] Preferably, S5, the motion blur correction in the distortion correction mechanism, specifically includes:

[0024] The vibration adaptive PSF reconstruction technology is adopted, based on the acquired real-time vibration velocity vector. and imaging exposure time Calculate at position Blur intensity at the location Specifically, it is expressed as:

[0025] ,

[0026] in, Represented as the length of the motion trajectory. Represented as a unit impulse function, It is represented as an index of the imaging exposure time.

[0027] Preferably, S5, the correction of optical distortion in the distortion correction mechanism, specifically includes:

[0028] The edge region and the central region are distinguished based on the incident position of the incident light on the surface of the glass fiber particles;

[0029] For rays with large incident angles in the edge region, aspherical refraction calculations are performed, and the ray deflection angle is determined by iteratively solving Snell's law.

[0030] For rays with small incident angles in the central region, the paraxial approximation is used to simplify the calculation;

[0031] The displacement of light rays due to differences in refractive index during propagation within glass fiber particles is geometrically corrected, specifically as follows:

[0032] ,

[0033] in, This indicates the position after distortion correction. This indicates the location of the second glass fiber particle in the image. Expressed as an empirical coefficient determined based on the refractive index of glass fiber. Indicated as in position The size of the distortion at that location, Represented as position The radial angle relative to the center of the particle.

[0034] Preferably, S7, the construction of the reinforcement learning model, specifically includes:

[0035] Vibration parameters, imaging parameters, and environmental indicators are integrated into a normalized vector as the state input;

[0036] The action space is defined as a series of executable action vectors, which include vibration adjustment, imaging adjustment, and distortion correction adjustment.

[0037] Design a reward function to evaluate the quality of actions performed in a given state, specifically expressed as follows:

[0038] ,

[0039] in, Represented as the value of the reward function, This is expressed as the deviation between the measured dimension and the target dimension. This represents the maximum possible dimensional measurement error. This represents the improvement in particle dispersion after vibration dispersion. Expressed as the energy cost of performing continuous actions, , , These represent the weighting coefficients for dimensional accuracy, dispersion improvement, and energy consumption, respectively.

[0040] To achieve the above objectives, the present invention provides the following technical solution: a size detection system for dynamic screening of glass fiber particles based on machine vision, comprising the following steps for implementing the aforementioned size detection method for dynamic screening of glass fiber particles based on machine vision:

[0041] Adhesion status monitoring module: used to extract the first glass fiber particle features in real time at the vibration dispersal outlet of the target glass fiber particle flow, and output the adhesion index corresponding to the first glass fiber particle features;

[0042] Collaborative imaging control module: used to dynamically adjust the vibration dispersing mechanism according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and to acquire images of the first glass fiber particles. The timing of the images of the first glass fiber particles is synchronously triggered according to the vibration dispersing mechanism.

[0043] Multimodal feature fusion module: used to extract the features of the second glass fiber particle from the first glass fiber particle image by fusing time-frequency analysis and convolutional neural network;

[0044] Monomerized particle screening module: used to analyze based on the characteristics of the second glass fiber particles to obtain images of the second glass fiber particles;

[0045] Dynamic distortion correction module: used to perform distortion correction mechanism on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by the difference in glass fiber refractive index.

[0046] Physical size calculation module: used to determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image based on the third glass fiber particle image;

[0047] Reinforcement learning optimization module: used to dynamically update the vibration disintegration mechanism and distortion correction mechanism through a reinforcement learning model.

[0048] Preferably, the size detection system for dynamic screening of glass fiber particles based on machine vision further includes:

[0049] Central processing unit (CPU) is used to output text instructions from various modules in the central control system.

[0050] The database is used to store all data text of the machine vision-based dynamic screening size detection system for glass fiber particles, and to collect information text output by each module in real time.

[0051] The visual information terminal is a device used to receive information output from a machine vision-based dynamic screening system for glass fiber particles.

[0052] The technical effects and advantages of this invention are as follows:

[0053] 1. This invention dynamically adjusts vibration parameters through adhesion index via S1 and S2, and triggers imaging based on vibration phase and energy spectral density, thereby achieving coordinated control of vibration and imaging. This effectively solves the problem of lack of coordination between vibration dispersion mechanism and imaging mechanism, improves particle dispersion and imaging clarity, and reduces the failure of single-unit imaging.

[0054] 2. This invention uses the S5 distortion correction algorithm to compensate for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index, which significantly improves image quality, reduces measurement distortion, and enhances the accuracy of dimensional measurement.

[0055] 3. This invention uses the reinforcement learning model of S7 to dynamically optimize the best combination, comprehensively considering size accuracy, dispersion improvement and energy consumption cost. While ensuring high dispersion, it maximizes the improvement of imaging quality, minimizes size measurement error, and further optimizes system performance. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the steps of a machine vision-based dynamic screening method for glass fiber particle size detection according to an embodiment of this application.

[0057] Figure 2 This is a block diagram of a size detection system for dynamic screening of glass fiber particles based on machine vision, provided in an embodiment of this application.

[0058] Figure 3 This is a module collaboration diagram of a machine vision-based dynamic screening size detection system for glass fiber particles provided in an embodiment of this application. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0060] As attached Figure 1 The machine vision-based dynamic screening size detection method for glass fiber particles, as shown, monitors the particle adhesion state in real time and dynamically adjusts the vibration dispersing mechanism. It further integrates vibration signals and image features for analysis, and corrects motion blur and refractive index distortion in the image to accurately measure the physical size of the particles. Specifically, it includes the following steps:

[0061] S1: Extract the first glass fiber particle feature in real time at the outlet of the target glass fiber particle flow vibration and dispersion, and output the adhesion index corresponding to the first glass fiber particle feature.

[0062] S2: The vibration dispersing mechanism is dynamically adjusted according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and the image of the first glass fiber particles is acquired. The timing of the first glass fiber particle image is synchronously triggered according to the vibration dispersing mechanism.

[0063] S3: Extract the features of the second glass fiber particles from the first glass fiber particle image by time-frequency analysis and convolutional neural network fusion;

[0064] S4: Analyze the features of the second glass fiber particles to obtain an image of the second glass fiber particles;

[0065] S5: Perform a distortion correction mechanism on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index.

[0066] S6: Based on the third glass fiber particle image, determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image;

[0067] S7: Dynamically update the vibration disintegration mechanism and distortion correction mechanism through reinforcement learning model.

[0068] Specifically, in S1, when the target glass fiber particle flow leaves the outlet of the vibration dispersing device, the adhesion state of each glass fiber particle cluster in the target glass fiber particle flow is monitored in real time, so as to integrate the data information of the adhesion state of each glass fiber particle cluster as the first glass fiber particle feature.

[0069] In this embodiment, a high-speed linear array camera is deployed 15cm directly above the outlet of the vibration dispersing device. The high-speed linear array camera must meet the following requirements: it must match the projection velocity of the glass fiber particles ≥5m / s to ensure complete capture of the particle trajectory during high-speed movement; it must be adapted to the clear imaging requirements of 20μm-level glass fiber particles, so that a single target particle covers at least 15 pixels in the image; and it must isolate vibration interference below 50Hz to ensure that imaging stability can still be maintained in a vibration environment with acceleration up to 2g.

[0070] It should be noted that the steps for performing multi-scale filtering on the images acquired by the high-speed linear array camera are as follows: The images corresponding to each glass fiber particle cluster in the target glass fiber particle stream are input into filtering paths using multiple scale structural elements. In this embodiment, a 3x3 disk structural element is used for scale 1 filtering, a 5x5 cross structural element for scale 2 filtering, and a 7x7 rectangular structural element for scale 3 filtering. Corresponding morphological operations are performed on the filtering results at multiple scales. In this embodiment, an erosion operation is performed on the scale 1 results to eliminate noise points smaller than 50μm; a dilation operation is performed on the scale 2 results to bridge the tiny gaps on the particle contours; an opening operation is performed on the scale 3 results to effectively separate weakly bonded particle clusters; and feature fusion is performed on the results after multi-scale filtering and morphological operations to generate particle contour features corresponding to each glass fiber particle cluster, i.e., the first glass fiber particle features.

[0071] In one possible implementation, the first glass fiber particle feature includes the area of ​​the glass fiber particle cluster convex hull. Minimum circumscribed rectangle area Number of contour dimples and the number of outline pixels Based on the characteristics of the first glass fiber particles, the adhesion index is calculated and output. Specifically, it is expressed as:

[0072] ,

[0073] in, Represented as area feature weights, The area feature weight and the concave feature weight are represented as concave point feature weights; in this embodiment, the area feature weight and the concave point feature weight are set to 0.7 and 0.3, respectively.

[0074] It should be noted that the adhesion index is a continuous value between 0 and 1, where 0 represents complete monomerization.

[0075] Specifically, in S2, the adjustable parameters in the vibration dispersing mechanism are dynamically adjusted according to the adhesion index output in real time, and the first glass fiber particle image is acquired simultaneously.

[0076] In one possible implementation, the adjustable parameters of the vibration dispersing mechanism include vibration frequency and amplitude, and are implemented by an electromagnetic vibrator. The implementation steps include: when the adhesion index > 0.6, indicating that there is significant adhesion of the glass fiber particle clusters and the breaking effect needs to be strengthened, the amplitude of the electromagnetic vibrator is increased to twice the preset amplitude reference value to enhance the impact force on the particle clusters, and the vibration frequency is reduced to 10% of the preset lower frequency limit to prolong the single vibration time and enable the glass fiber particles to obtain greater kinetic energy separation; when the adhesion index < 0.3, indicating that the glass fiber particle clusters have basically reached the individualization state, the amplitude is reduced to 10% of the preset amplitude reference value to avoid secondary splashing of the individualized particles due to excessive vibration, and the vibration frequency is increased to 10% of the preset upper frequency limit.

[0077] In this embodiment, the preset amplitude reference value is 1mm, the preset frequency lower limit is 20Hz, and the preset frequency upper limit is 200Hz.

[0078] Furthermore, the vibration energy spectral density is calculated based on the accelerometer to determine the motion state of the target glass fiber particle flow. The specific calculation method includes performing a fast Fourier transform on the acceleration signal, identifying the dominant frequency, and calculating the sum of squares of energy within a bandwidth of ±5Hz of the dominant frequency, i.e., the vibration energy spectral density. The vibration energy spectral density is positively correlated with the breaking effect of the target glass fiber particle flow. The timing of the acquisition of the first glass fiber particle image is synchronously triggered according to the adjustable parameters of the vibration breaking mechanism. A mapping relationship between the vibration phase and the imaging timing is established based on identifying the optimal imaging phase, which is the vibration trough phase. At the vibration trough, the glass fiber particle cluster is ejected to its maximum height and is in a relatively suspended state.

[0079] In this embodiment, when the vibration energy spectral density is greater than 3000 (m / s²)² / Hz, it is determined to be a strong fragmentation state, indicating that the target glass fiber particle flow is moving violently, triggering the high-speed linear array camera to perform imaging at a high frame rate to capture the instantaneous state; when the vibration energy spectral density is less than 800 (m / s²)² / Hz, it is determined to be a stable flow dynamic, indicating that the target glass fiber particle flow is moving relatively smoothly, triggering the high-speed linear array camera to perform imaging at a lower frame rate to reduce the amount of data and maintain processing efficiency;

[0080] It should be noted that the method for identifying the vibration trough phase includes: detecting the zero-crossing point through the accelerometer signal and determining whether the signal is at the rising edge or falling edge, thereby marking the current vibration phase as a peak or trough.

[0081] Furthermore, based on the vibration phase and vibration energy spectral density, a trigger condition decision tree is used to determine the imaging timing of the first glass fiber particle image. The specific judgment logic is as follows: when the vibration phase is at a trough and the vibration energy spectral density is less than a preset energy spectral density threshold, imaging is triggered immediately; otherwise, the exposure is delayed until the next vibration cycle, waiting for the trigger condition to be met. In this embodiment, the preset energy spectral density threshold is set to 1500 (m / s²)² / Hz.

[0082] Specifically, in S3, the second glass fiber particle features are extracted from the acquired first glass fiber particle image using multimodal feature fusion technology. The multimodal features fuse the time-frequency analysis features of the vibration signal and the convolutional neural network features of the optical image. The second glass fiber particle features include real-time vibration velocity vector, main frequency band energy ratio, frequency centroid, frequency band entropy value, imaging exposure time, glass fiber surface defects, and adhesion types.

[0083] Furthermore, the extraction of vibration signal features through time-frequency analysis includes: capturing vibration signals of 0.5–10 kHz in real time by installing a triaxial piezoelectric accelerometer on the main shaft of the vibrating disk; configuring an accelerometer with a sampling rate of 20 kHz and installing it at 5 cm from the center of mass of the vibrating disk to directly sense the interaction force between the particles and the disk body; decomposing the signal in the time-frequency domain through wavelet packet transform; and using a convolutional neural network to receive images from the first glass fiber particles. The convolutional neural network follows a feature evolution mechanism: the shallow layer of the convolutional neural network extracts edges and corners; the middle layer extracts scratches and defects on the glass fiber surface; and the deep layer distinguishes various adhesion types, including but not limited to point adhesion, line adhesion, and surface adhesion.

[0084] Furthermore, the multi-fusion extraction of the second glass fiber particle features adopts a gated attention fusion network architecture, including: calculating the weights of the time-frequency analysis features of the vibration signal and the convolutional neural network features of the optical image through an attention gating mechanism to improve weighted fusion and generate a weighted feature vector; and outputting the second glass fiber particle features through a fully connected layer.

[0085] Specifically, in S4, analysis is performed based on the features of the second glass fiber particles to generate an optimized image of the glass fiber particles.

[0086] Furthermore, the acquisition of the second glass fiber particle image includes: analyzing feature saliency, retaining features including but not limited to edge sharpness, shape regularity, etc.; identifying the correlation between features through correlation analysis, including but not limited to the correlation between frequency band entropy value and dispersion; diagnosing the adhesion state based on multiple feature combination rules; and analyzing the particle surface texture using local binary mode.

[0087] In this embodiment, significant features include edge gradient variance in the image domain and the dominant frequency band energy ratio in the vibration domain; correlated features include edge sharpness and texture uniformity; and dispersive features include frequency band entropy and contour convexity. When the edge gradient variance is less than 15 and the frequency band entropy is greater than 4, the diagnostic result is "surface adhesion," indicating that the particles may have large-area adhesion. If the shape regularity is greater than 0.7 and the dominant frequency band energy ratio is less than 0.3, the diagnostic result is "monolithic particles," indicating that the particles have been successfully dispersed and fluidized stably.

[0088] Furthermore, the acquisition of the second glass fiber particle image also includes: adopting a dual-path generation architecture, selecting an image generation path based on the results of feature analysis. In this embodiment, for the target of "single particles", the super-resolution reconstruction path is used; for the target of "adhesive particles", the physical simulation auxiliary path is used. The super-resolution reconstruction path uses an ESRGAN network to improve the detail clarity of single particles and provide higher quality image data for subsequent accurate size measurement. The physical simulation auxiliary path combines a discrete element model and uses a DEM model to simulate the breakage and separation process under specific vibration, generating a virtual separation image based on the simulation results.

[0089] Specifically, in S5, dynamic distortion correction is performed on the second glass fiber particle image to generate a third glass fiber particle image.

[0090] In one possible implementation, the distortion correction mechanism includes measures for motion blur and optical distortion, including: employing a vibration-adaptive PSF reconstruction technique based on the acquired real-time vibration velocity vector. and imaging exposure time Calculate at position Blur intensity at the location Specifically, it is expressed as:

[0091] ,

[0092] in, Represented as the length of the motion trajectory. Represented as a unit impulse function, Represented as an index of the imaging exposure time; by reconstructing the time... The motion is analyzed by using a blind deconvolution algorithm to recover the second glass fiber particle image. Through multiple iterations, the potential image is estimated, and the backpropagation error is corrected using the known PSF.

[0093] In one possible implementation, the distortion correction mechanism includes measures for motion blur and optical distortion, and further includes: distinguishing between edge regions and central regions based on the incident position of the incident light rays on the surface of the glass fiber particles; performing aspherical refraction calculations on large incident angle light rays in the edge regions, and determining the light deflection angle by iteratively solving Snell's law; simplifying the calculations by using paraxial approximation for small incident angle light rays in the central regions; and performing geometric correction on the displacement of the light rays due to refractive index differences during propagation within the glass fiber particles, specifically expressed as:

[0094] ,

[0095] in, This indicates the position after distortion correction. This indicates the location of the second glass fiber particle in the image. Expressed as an empirical coefficient determined based on the refractive index of glass fiber. Indicated as in position The size of the distortion at that location, Represented as position The radial angle relative to the center of the particle.

[0096] Specifically, in S6, high-precision physical size measurement and verification are performed based on the third glass fiber particle image.

[0097] Furthermore, a judgment strategy is executed based on the adhesion index to distinguish and obtain images of individual glass fiber particles, including: when the adhesion index is less than 0.3, the watershed algorithm is used for segmentation; when the adhesion index is between 0.3 and 0.6, the concave point cutting method is used to cut by identifying concave points on the particle cluster contour as potential connection points; when the adhesion index is greater than 0.6, a semantic segmentation network based on U-Net is used to identify and segment highly adhered particle regions.

[0098] Furthermore, based on the identified and segmented images of individual glass fiber particles, physical size measurements are performed. The measurement method includes: calculating the equivalent diameter of the spherical particles. Specifically, it is expressed as:

[0099] ,

[0100] in, Represented as projected area, It is represented as the glass fiber shape factor, and the glass fiber shape factor is set to 0.82 for glass fiber particles with an aspect ratio greater than or equal to 20.

[0101] Specifically, in S7, the goal of the reinforcement learning model is to dynamically optimize the best combination of vibration parameters and imaging parameters in order to maximize imaging quality while ensuring high dispersion, thereby minimizing the final size measurement error.

[0102] In one possible implementation, the construction of the reinforcement learning model includes: integrating vibration parameters, imaging parameters, and environmental indicators into a normalized vector as the state input; defining the action space as a continuous vector of executable actions, which includes vibration modulation, imaging modulation, and distortion correction modulation; and designing a reward function to evaluate the quality of actions performed in the state, specifically expressed as:

[0103] ,

[0104] in, Represented as the value of the reward function, This is expressed as the deviation between the measured dimension and the target dimension. This represents the maximum possible dimensional measurement error. This represents the improvement in particle dispersion after vibration dispersion. Expressed as the energy cost of performing continuous actions, , , These represent the weighting coefficients for dimensional accuracy, dispersion improvement, and energy consumption, respectively.

[0105] In this embodiment, the weighting coefficients for dimensional accuracy, dispersion improvement, and energy consumption are 0.6, 0.3, and 0.1, respectively.

[0106] It should be noted that the vibration disintegration mechanism and distortion correction mechanism are updated using a rolling optimization process. Every unit of time, the latest 100,000 sets of data are sampled from the database and the corresponding error is calculated. If the average error is greater than the preset threshold, it indicates that the vibration disintegration mechanism and distortion correction mechanism are not performing well on the latest data and need to be updated.

[0107] As attached Figure 2 The machine vision-based dynamic screening size detection system for glass fiber particles shown includes: an adhesion state monitoring module, a collaborative imaging control module, a multimodal feature fusion module, a single-unit particle screening module, a dynamic distortion correction module, a physical size calculation module, and a reinforcement learning optimization module.

[0108] The adhesion state monitoring module is used to extract the first glass fiber particle features in real time at the vibration and dispersal outlet of the target glass fiber particle flow, and output the adhesion index corresponding to the first glass fiber particle features.

[0109] The collaborative imaging control module is used to dynamically adjust the vibration dispersing mechanism according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and to acquire images of the first glass fiber particles. The timing of the images of the first glass fiber particles is synchronously triggered according to the vibration dispersing mechanism.

[0110] The multimodal feature fusion module is used to extract the features of the second glass fiber particle from the first glass fiber particle image by combining time-frequency analysis and convolutional neural network.

[0111] The monomerized particle screening module is used to analyze the characteristics of the second glass fiber particles to obtain images of the second glass fiber particles.

[0112] The dynamic distortion correction module is used to perform distortion correction on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index.

[0113] The physical size calculation module is used to determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image based on the third glass fiber particle image.

[0114] The reinforcement learning optimization module is used to dynamically update the vibration disintegration mechanism and distortion correction mechanism through a reinforcement learning model.

[0115] It should be noted that the machine vision-based dynamic screening size detection system for glass fiber particles also includes a central processing unit, a database, and a visual information terminal; as shown in the attached document. Figure 3 The module collaboration diagram shown includes a central processing unit (CPU) for outputting text instructions from each module in the central control system; a database for storing all data text of the machine vision-based dynamic screening size detection system for glass fiber particles, and for collecting information text output from each module in real time; and a visual information terminal for receiving information output from the machine vision-based dynamic screening size detection system for glass fiber particles.

[0116] The central processing unit (CPU) is the core computing and control unit of the entire machine vision-based dynamic screening size detection system for glass fiber particles. It includes one or more processing cores connected to key components within the system, such as the status monitoring module, collaborative imaging control module, multimodal feature fusion module, individual particle screening module, dynamic distortion correction module, physical size calculation module, and reinforcement learning optimization module, via a PCIe 4.0 bus and a gigabit industrial Ethernet interface. By running or executing instructions, programs, code sets, or instruction sets stored in a database, and by calling the data stored therein, it performs various functions of the dynamic screening system for glass fiber particles, including real-time collaborative control, multimodal feature fusion, and reinforcement learning decision-making, ensuring that the size measurement error remains stable within ±0.8 μm in high-throughput scenarios exceeding 200 particles / second.

[0117] The database is used to store a large amount of data related to glass fiber particle size detection, including mathematical model data based on physical optics principles, mathematical model data based on material dynamics principles, historical operation datasets including hundreds of thousands of labeled particle images, vibration parameter-size error mapping tables, and reinforcement learning strategy iterative version library. When the central processing unit executes various functions, it will frequently call these data from the system operation database to perform operations such as vibration parameter optimization, optical distortion compensation, and sorting path planning, thereby realizing closed-loop control of the entire glass fiber particle dispersion, imaging, measurement, and sorting.

[0118] The visual information terminal connects to external devices such as displays and cameras via standard wired or wireless interfaces, providing users with interactive functions including: displaying vibration energy spectral density curves and size distribution heatmaps, modifying reinforcement learning reward weights via touchscreen, and initiating AI-assisted maintenance based on device status images captured by cameras.

[0119] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0120] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A size detection method for dynamic screening of glass fiber particles based on machine vision, characterized in that, include: S1: Extract the first glass fiber particle feature in real time at the outlet of the target glass fiber particle flow vibration dispersion, and output the adhesion index corresponding to the first glass fiber particle feature; the first glass fiber particle feature includes the area of ​​the glass fiber particle cluster convex hull. Minimum circumscribed rectangle area Number of contour dimples and the number of outline pixels Based on the characteristics of the first glass fiber particles, the adhesion index is calculated and output. Specifically, it is expressed as: in, Represented as area feature weights, Represented as concave feature weights; S2: The vibration dispersing mechanism is dynamically adjusted according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and the image of the first glass fiber particles is acquired. The timing of the first glass fiber particle image is synchronously triggered according to the vibration dispersing mechanism. S3: For the first glass fiber particle image, extract the second glass fiber particle features by fusing time-frequency analysis and convolutional neural network; the second glass fiber particle features include real-time vibration velocity vector, main frequency band energy ratio, frequency centroid, frequency band entropy value, imaging exposure time, glass fiber surface defects and adhesion type. S4: Analyze the features of the second glass fiber particles to obtain an image of the second glass fiber particles, specifically including: The process involves analyzing the saliency of features; identifying the associations between features through correlation analysis; diagnosing the adhesion state based on multiple feature combination rules; selecting the image generation path based on the results of feature analysis; simulating the breakup and separation process of the adhered particles under specific vibration, and generating a virtual separation image based on the simulation results. S5: Perform a distortion correction mechanism on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by differences in glass fiber refractive index. S6: Based on the third glass fiber particle image, determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image; S7: Dynamically update the vibration disintegration mechanism and distortion correction mechanism through reinforcement learning model.

2. The size detection method for dynamic screening of glass fiber particles based on machine vision according to claim 1, characterized in that: In step S2, the vibration dispersal mechanism is implemented by an electromagnetic vibrator, specifically including: When the adhesion index is greater than 0.6, the amplitude of the electromagnetic vibrator is increased to twice the preset amplitude reference value and the vibration frequency is reduced to 10% of the preset lower limit value. When the adhesion index is less than 0.3, the amplitude will be reduced to 10% of the preset amplitude reference value and the vibration frequency will be increased to 10% of the preset frequency upper limit value.

3. The size detection method for dynamic screening of glass fiber particles based on machine vision according to claim 1, characterized in that: S5, the motion blur correction in the distortion correction mechanism, specifically includes: The vibration adaptive PSF reconstruction technology is adopted, based on the acquired real-time vibration velocity vector. and imaging exposure time Calculate at position Blur intensity at the location Specifically, it is expressed as: in, Represented as the length of the trajectory. Represented as a unit impulse function, This is represented as an index of the imaging exposure time.

4. The size detection method for dynamic screening of glass fiber particles based on machine vision according to claim 1, characterized in that: S5, the correction of optical distortion in the distortion correction mechanism, specifically includes: The edge region and the central region are distinguished based on the incident position of the incident light on the surface of the glass fiber particles; For rays with large incident angles in the edge region, aspherical refraction calculations are performed, and the ray deflection angle is determined by iteratively solving Snell's law. For rays with small incident angles in the central region, the paraxial approximation is used to simplify the calculation; The displacement of light rays due to differences in refractive index during propagation within glass fiber particles is geometrically corrected, specifically as follows: in, This indicates the position after distortion correction. This indicates the location of the second glass fiber particle in the image. Expressed as an empirical coefficient determined based on the refractive index of glass fiber. Indicated as in position The size of the distortion at that location, Represented as position The radial angle relative to the center of the particle.

5. The size detection method for dynamic screening of glass fiber particles based on machine vision according to claim 1, characterized in that: S7, the construction of the reinforcement learning model, specifically includes: Vibration parameters, imaging parameters, and environmental indicators are integrated into a normalized vector as the state input; The action space is defined as a series of executable action vectors, which include vibration adjustment, imaging adjustment, and distortion correction adjustment. Design a reward function to evaluate the quality of actions performed in a given state, specifically expressed as follows: in, Represented as the value of the reward function, This is expressed as the deviation between the measured dimension and the target dimension. This represents the maximum possible dimensional measurement error. This represents the improvement in particle dispersion after vibration dispersion. Expressed as the energy cost of performing continuous actions, , , These represent the weighting coefficients for dimensional accuracy, dispersion improvement, and energy consumption, respectively.

6. A size detection system for dynamic screening of glass fiber particles based on machine vision, used to implement the size detection method for dynamic screening of glass fiber particles based on machine vision as described in any one of claims 1-5, characterized in that, include: Adhesion status monitoring module: used to extract the first glass fiber particle features in real time at the vibration dispersal outlet of the target glass fiber particle flow, and output the adhesion index corresponding to the first glass fiber particle features; Collaborative imaging control module: used to dynamically adjust the vibration dispersing mechanism according to the adhesion index corresponding to the characteristics of the first glass fiber particles, and to acquire images of the first glass fiber particles. The timing of the images of the first glass fiber particles is synchronously triggered according to the vibration dispersing mechanism. Multimodal feature fusion module: used to extract the features of the second glass fiber particle from the first glass fiber particle image by fusing time-frequency analysis and convolutional neural network; Monomerized particle screening module: used to analyze based on the characteristics of the second glass fiber particles to obtain images of the second glass fiber particles; Dynamic distortion correction module: used to perform distortion correction mechanism on the second glass fiber particle image to generate a third glass fiber particle image. The distortion correction mechanism includes compensating for motion blur caused by vibration and optical distortion caused by the difference in glass fiber refractive index. Physical size calculation module: used to determine the physical size of the glass fiber particles corresponding to the third glass fiber particle image based on the third glass fiber particle image; Reinforcement learning optimization module: used to dynamically update the vibration disintegration mechanism and distortion correction mechanism through a reinforcement learning model.

7. The size detection system for dynamic screening of glass fiber particles based on machine vision according to claim 6, characterized in that, Also includes: Central processing unit (CPU) is used to output text instructions from various modules in the central control system. The database is used to store all data text of the machine vision-based dynamic screening size detection system for glass fiber particles, and to collect information text output by each module in real time. The visual information terminal is a device used to receive information output from a machine vision-based dynamic screening system for glass fiber particles.

Images