Visual sorting method and system for pollutants in insulating material particles

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a visual sorting method and system for pollutants in insulating material particles, and the method comprises the steps: obtaining the particle characterization size, the number of adjacent particles and the distance between the adjacent particles; obtaining a target congestion degree; calculating a target deviation index for characterizing the moving environment of the target particle; obtaining a real-time movement distance according to the position change of the target particle before and after movement, and obtaining a total movement distance from a database; obtaining target dynamic noise; outputting and rejecting the final coordinates of the target; judging whether the elimination is successful or not and storing the structured data; and optimizing the motion prediction model. The technical problems that in a traditional sorting mode, particle motion trail prediction is not accurate, and complex environment changes cannot be adapted are solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a visual sorting method and system for contaminants in insulating material particles. Background Technology

[0002] In the production process of high-performance insulating materials, the removal of trace contaminants from raw materials is a core step in ensuring the quality and performance of the final product. To this end, the industry generally adopts color sorters or sorting systems based on machine vision. These systems use cameras to capture the falling material flow, identify contaminant particles with abnormal color or shape, and use prediction systems to calculate their trajectory. Finally, they drive high-pressure air nozzles to spray and remove the contaminants at precise times and locations.

[0003] However, traditional visual sorting systems are mostly passive prediction systems, assuming that particles undergo ideal free fall or parabolic motion under gravity. They struggle to detect abnormal states during the prediction process, such as trajectory deviations caused by collisions with other particles or uneven airflow disturbances. These unconsidered random disturbances, while having a small impact in a single event, accumulate into significant prediction errors in large-scale, high-density production flows, ultimately leading to inaccurate rejection and potential product quality issues. To address this deficiency, some existing technologies employ advanced prediction algorithms such as Kalman filtering to provide early warnings of anomalies. This involves comparing and correcting the real-time particle positions with a dynamic model based on fixed parameters.

[0004] However, this judgment method based on fixed parameter models has inherent limitations. In real industrial environments, the intensity of disturbances experienced by particles is not constant but is affected by various dynamic factors such as material flow density, electrostatic interactions between particles, and ambient humidity. Using single, fixed model parameters, such as a fixed process noise covariance, cannot adapt to these dynamic changes. In conditions where the material flow becomes congested and collisions occur frequently, the system becomes overly reliant on model predictions, resulting in insufficient correction to the actual observed positions and missed detections, i.e., failure to remove contaminants. Conversely, in conditions where the material flow is sparse and the flight trajectory is stable, overly conservative model parameters may lead to excessive oscillations in the predicted trajectory, resulting in false alarms, i.e., incorrectly removing good products. This lack of ability to perceive and adaptively adjust to the current real flight environment of particles makes it difficult for existing technologies to achieve a balance between high rejection accuracy and high material yield, severely restricting the reliability and practicality of intelligent visual sorting technology. Summary of the Invention

[0005] To address the technical problems of inaccurate prediction of particle movement trajectories and inability to adapt to complex environmental changes in traditional sorting methods, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a visual sorting method for contaminants in insulating material particles, comprising: The target particle is initially located using a neural network model. Based on the morphological characteristics and distribution of the target particle and its neighboring particles, the particle's size, the number of neighboring particles, and the distance between them are obtained. The congestion degree is obtained by summing the congestion contribution of each individual neighboring particle within a pre-defined neighborhood, considering their size and distance. A target deviation index, representing the moving environment of the target particle, is calculated based on the degree of deviation of the target congestion degree from historical conditions, depending on whether historical conditions are stable. The real-time moving distance is obtained based on the positional changes of the target particle before and after movement, and the total moving distance is retrieved from the database. The pre-acquired baseline process noise is adaptively adjusted by fusing the real-time movement progress, historical conditions, and the target deviation index to obtain the target dynamic noise. The motion prediction model is updated based on the target dynamic noise, and the final target coordinates are output and removed. The success of the removal is determined, and structured data is stored. Regression analysis is performed on multiple failed data points in the structured data to optimize the motion prediction model.

[0007] This invention effectively solves the problems of inaccurate prediction of particle trajectory and inability to adapt to complex environmental changes in traditional sorting methods. Traditional methods struggle to handle particle collisions and airflow interference, easily leading to sorting deviations, missed sorting, and missorting. This solution accurately identifies particle information, assesses environmental congestion, evaluates risks based on historical data, and flexibly adjusts relevant parameters to optimize sorting strategies in real time, ensuring precise execution of sorting actions. Simultaneously, by verifying sorting results and optimizing the model, it continuously improves sorting accuracy and stability, reducing missed and missorting, and resolving the difficulty of balancing sorting accuracy and material utilization in complex working conditions under traditional sorting methods, ensuring efficient and reliable sorting operations.

[0008] Preferably, obtaining particle characterization size, number of adjacent particles, and distance between adjacent particles includes: Based on the visual features of material particles such as color, texture, and shape in real-time materials, contaminant particles are accurately identified and segmented from the background of normal insulating material particles and recorded as target particles. The particle size is obtained by calculating the number of pixels contained in the segmentation mask of the material particles. The pixel coordinates of the target particles are output, and all neighboring particles are identified in the preset neighborhood of the target particles using image processing algorithms to obtain the number of neighboring particles. The pixel distance between all neighboring particles and the target particles is calculated and converted according to the camera calibration parameters. The pixel distance between each neighboring particle and the target particle is recorded as the neighboring particle distance.

[0009] Preferably, the target congestion degree satisfies the following expression: ; In the formula, Indicates the target congestion level; The target particle's first The particle size of each adjacent particle; Indicates the target particle and the first The distance between adjacent particles of a neighboring particle; This represents the pre-obtained distance decay index; Indicates the number of adjacent particles of the target particle; Represents a very small positive number, ensuring that the denominator is not zero; This represents the normalization function.

[0010] This invention, by accurately calculating the degree of congestion around particles, can truly reflect the environmental state of the particles and clearly demonstrate the impact of particles of different sizes on the target particles at different distances. Instead of simply judging particle distribution, it precisely calculates the degree of environmental congestion, allowing the system to better understand the potential interference to the target particles. This provides strong support for subsequent adjustments to sorting strategies, reduces sorting deviations caused by inaccurate environmental judgments, and improves the sorting process's adaptability to complex environments.

[0011] Preferably, the target deviation index satisfies the following expression: ; In the formula, Indicates the deviation index from the target; Indicates the target congestion level; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; , Represents a very small positive number, ensuring that the denominator is not zero; It is the hyperbolic tangent function.

[0012] This invention combines historical data to determine changes in the current environment of particles, intuitively reflecting the potential trajectory deviation risks during particle movement. It helps the system quickly determine whether the current environment is stable, promptly identify crowded or sparse environmental states, and prepare in advance to avoid inaccurate sorting due to ignoring environmental changes. The system can adjust its operation mode according to the level of risk, improving the flexibility and reliability of the sorting process and reducing sorting errors caused by environmental fluctuations.

[0013] Preferably, the real-time moving distance is obtained based on the positional change of the target particle before and after its movement, and the total moving distance is obtained from the database, including: The pixel coordinates of the target particle acquired initially are taken as the initial target coordinates. During the movement of the target particle, the real-time position of the target particle is located according to the optical flow method to obtain the real-time position coordinates. The Euclidean distance between the initial target coordinates and the real-time position coordinates is calculated to obtain the real-time movement distance. The system's inherent distance is obtained from the database and recorded as the total movement distance.

[0014] Preferably, the target dynamic noise satisfies the following expression: ; In the formula, For target dynamic noise; The deviation index from the target; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; Reference process noise; Indicates the real-time distance traveled; Indicates the total distance traveled; Represents the absolute value function; It represents a very small positive number, and guarantees that the denominator is not 0.

[0015] This invention flexibly adjusts relevant parameters based on the actual movement of particles, historical environmental conditions, and potential trajectory deviation risks. This allows the parameter settings to be flexible and adaptable to different production environments and particle movement states, avoiding the deterioration of sorting performance in complex environments due to fixed parameters. It enhances the system's adaptability to different working conditions, ensuring a good sorting state under various circumstances, reducing sorting errors caused by parameter mismatch, and improving sorting efficiency and quality.

[0016] Preferably, outputting the final coordinates of the targets and removing them includes: The system updates the motion prediction model based on the target dynamic noise and calculates the final coordinates of the target particle when it will arrive at the removal execution module at a certain predicted time. The final coordinates of the target particle are sent to the removal execution module. At the same time, the air nozzle that precisely corresponds to the final coordinates of the target particle is activated instantly and sprays air to accurately separate the target particle from the main material flow.

[0017] Preferably, determining whether the removal was successful and storing the structured data includes: A verification camera is set downstream of the rejection execution module. This camera is used to capture images of the material flow that has just passed through the rejection area. The system analyzes the images captured by the verification camera to determine whether the target particles have been completely separated from the main material flow and whether no new contaminant particles appear in the image. If the target particles are completely separated and there are no new contaminants, the rejection is considered successful. If the target particles are still in the main material flow or new contaminants appear, the rejection is considered a failure. The system records the verification results of this rejection task and all the parameters in this rejection task process, including target congestion, target deviation index, target dynamic noise, and target final coordinates, as structured data.

[0018] This invention analyzes sorting failure data to continuously optimize the system's predictive capabilities, identifies key factors leading to sorting failures, and adjusts system settings accordingly to prevent similar errors in subsequent operations. It enhances the system's self-improvement capabilities, moving away from fixed settings and enabling continuous improvement based on actual operating conditions. This adapts to different production needs and environmental changes, maintaining a consistently efficient and accurate sorting state, reducing the sorting failure rate, and improving overall sorting performance.

[0019] Preferably, optimizing the motion prediction model includes: The system periodically collects and analyzes structured data, including failed data with failure labels. It then performs regression analysis on the trajectory data of multiple failed data with failure labels to update the target dynamic noise in the motion prediction model and optimize the motion prediction model.

[0020] Secondly, the present invention provides a visual sorting system for contaminants in insulating material particles, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned visual sorting method for contaminants in insulating material particles is implemented.

[0021] By adopting the above technical solution, a computer program is generated for the visual sorting method of pollutants in insulating material particles, and stored in a memory for loading and execution by a processor. Terminal devices are then manufactured based on the memory and processor for convenient use.

[0022] The beneficial effects of this invention are as follows: From an overall production perspective, this solution improves the automation and intelligence level of insulating material particle sorting, reduces the uncertainty and errors caused by manual intervention, removes impurities through precise sorting, ensures the quality of insulating material products, provides high-quality raw materials for subsequent production stages, helps improve the performance and reliability of the final product, and enhances the product's competitiveness in the market. At the same time, the continuously optimized sorting system can adapt to different production conditions, improve production efficiency, and reduce material waste and production costs caused by sorting errors. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a visual sorting method for contaminants in insulating material particles according to the present invention. Detailed Implementation

[0024] This invention discloses a visual sorting method for contaminants in insulating material particles, referring to... Figure 1 This includes steps S1-S4: S1: Based on the neural network model, the target particle is initially located. According to the morphological characteristics and distribution of the target particle and its neighboring particles, the particle size, the number of neighboring particles, and the distance between neighboring particles are obtained. The crowding degree of the target particle is obtained by accumulating the crowding contribution of a single neighboring particle within the preset neighborhood of the target particle, which is obtained by considering the size and distance of the neighboring particles.

[0025] It should be noted that existing technologies use Kalman filters with fixed parameters, much like a physicist predicting the landing point of a particle in a vacuum. The trajectory is a perfect parabola, making prediction very simple. However, in real production material flows, the target particle does not fall alone but is deeply embedded in a particle swarm consisting of thousands of accompanying particles. The random collisions and compressions of these accompanying particles constantly interfere with the target particle's ideal trajectory. The core idea of ​​this invention is to enable the prediction system to perceive the intensity of collisions and compressions on the target particle in real time, i.e., the degree of crowding, and dynamically adjust the prediction model accordingly, thereby accurately locking onto the initially determined, moved target particle amidst the chaos.

[0026] Specifically, the target particle is initially located based on a neural network model. Based on the morphological characteristics and distribution of the target particle and its neighboring particles, the particle's size, the number of neighboring particles, and the distance between neighboring particles are obtained, including: This step first identifies the target particles and accompanying particles. The system captures real-time material images from the camera, and the real-time material image stream is fed into a pre-trained deep convolutional neural network model. Based on visual features such as color, texture, and shape, the deep convolutional neural network model accurately identifies and segments contaminant particles from the background of normal insulating material particles, which are recorded as target particles. The particle size is obtained by calculating the number of pixels contained in the segmentation mask of the material particles. The pixel coordinates of the target particles are output, and the image processing algorithm is used to identify all neighboring particles in the preset neighborhood of the target particles to obtain the number of neighboring particles. The pixel distance between all neighboring particles and the target particles is calculated and converted according to the camera calibration parameters. The pixel distance between each neighboring particle and the target particle is recorded as the neighboring particle distance.

[0027] It should be noted that the neural network model mentioned in this application is merely a conventional preliminary technique for obtaining basic image features, such as initial target particle localization, obtaining particle segmentation masks and pixel sizes, and is not the core inventive point that this application aims to improve. The core of this invention lies in proposing a complete set of subsequent trajectory prediction and optimization mechanisms based on physical congestion calculation, deviation index calculation, and dynamic noise adaptive adjustment, for harsh working conditions such as complex airflow disturbances and particle collisions in practical applications after the target particles are identified and located by the neural network. The system completes the identification, segmentation, and localization of target particles through this neural network model, providing source data support for subsequent environmental perception, trajectory prediction, and rejection execution throughout the entire process. In this invention, the system actually applies the basic feature parameters output by this model to the entire sorting method, thereby leading to the neural network model. Therefore, the original internal logic of the specification is that by processing the material images acquired in real time, the identification, segmentation, and spatial localization of pollutant target particles are completed, and the basic features and coordinate data of the particles are output, providing a stable data source for the calculation of core parameters of the entire sorting process. This neural network model and the subsequent core algorithms are not isolated concepts, but rather related to the foundational data and subsequent core innovations, basic feature extraction, and the implementation of the entire technology process. The neural network model of this invention adopts a single-input, single-output target particle instance segmentation and localization functional structure, which can be divided into three functional layers: an input layer, a core computation processing layer, and an output layer. The input layer is a real-time material image input stream, serving as the basic image data source for the model, corresponding to real-time frame images of the falling insulating material particles captured by industrial cameras on the production line. The core computation processing layer employs a three-step serial processing architecture of morphological feature recognition, segmentation mask calculation, and pixel coordinate mapping to accurately identify, segment, and locate contaminant target particles in the background of normal insulating particles. The output layer outputs the particle characterization size and initial positioning coordinates, directly connecting to the subsequent target congestion calculation stage, providing core basic data support for the material flow environment risk perception and trajectory prediction optimization of the entire solution. This neural network model structure discloses its functional architecture, data flow, and execution logic. Furthermore, this model and the subsequent core algorithms form a complete technical closed loop, ensuring the feasibility and industrial application value of the entire technical solution.

[0028] It should be noted that traditional sorting systems, due to their use of fixed parameter models, cannot accurately adapt to the complex collisions and interferences between particles in high-density material flows. This results in a low correlation between congestion assessment and rejection failure risk. The target congestion expression addresses this deficiency by determining the optimal distance attenuation index through historical data statistical regression and constructing a weighted nonlinear calculation method based on the size and distance of adjacent particles. This enables accurate calculation of the collision risk between particles, allowing the congestion index to truly reflect the comprehensive interference intensity of particles of different sizes at different distances.

[0029] Preferably, the target congestion degree is obtained by accumulating the congestion contribution of individual adjacent particles within a preset neighborhood of the target particle, based on the size and distance of adjacent particles. This includes: By conducting statistical regression analysis on a large amount of historical pollutant particle-related data and removal structure-related data, we sought the coefficient value that showed the strongest correlation between crowding and removal failure rate as the distance decay index. It should be noted that traditional models only consider the distance between particles, ignoring the influence of particle size; while the target crowding degree expression incorporates two key factors: particle size and the distance between adjacent particles. In actual production, collisions of large-sized particles have a more significant impact on the trajectory of the target particle. By incorporating the size factor, the calculated crowding degree can more realistically reflect the interaction between particles. In addition, by introducing a distance decay exponent to achieve weighted nonlinear calculation, when the exponent is greater than 1, the influence of nearby particles is amplified, which is more in line with the scenario where physical collisions are the primary cause.

[0030] It should be noted that the target congestion expression in this invention is constructed based on the laws of particle flow dynamics and the general mathematical principles of industrial visual measurement. Firstly, the expression follows the natural law of particle collision interference: the larger the size of adjacent particles and the closer they are to the target particle, the higher the risk of collision interference with the target particle's trajectory, and the greater their contribution to congestion. The expression accurately reflects this objective physical law through a mathematical structure where particle size is positively correlated with congestion contribution and particle distance is negatively correlated with congestion contribution. Secondly, particle size is represented by area, the distance between adjacent particles by length, and the contribution of a single particle by length. These dimensions are then summed and normalized using a function. After processing, the final output target congestion degree is a dimensionless standardized index, which conforms to the rules of dimensional operation; thirdly, the introduction of the distance decay index can be adapted to material flow scenarios with different densities through historical data regression, which conforms to the engineering practice rules of industrial particle sorting.

[0031] The target congestion level satisfies the following expression: ; In the formula, Indicates the target congestion level; The target particle's first The particle size of each adjacent particle; Indicates the target particle and the first The distance between adjacent particles of a neighboring particle; This represents the pre-obtained distance decay index; Indicates the number of adjacent particles of the target particle; Represents a very small positive number, ensuring that the denominator is not zero; This represents the normalization function.

[0032] In the formula, This represents the weighted nonlinear contribution of a single adjacent particle to the target crowding degree. Unlike traditional models, this contribution considers both the particle's size and the distance to adjacent particles. Including particle size in the calculation ensures that larger particles contribute more, while incorporating the distance between adjacent particles adjusts the influence of distance. When the value is greater than 1, the impact of nearby particles is amplified dramatically, making it more suitable for simulating scenarios primarily based on physical collisions; when... When the value approaches 1, the system degenerates into a conventional model that is inversely proportional to the distance between adjacent particles; Indicates all The weighted nonlinear contributions of each adjacent particle are accumulated to calculate the local, multi-factor collision risk into a macroscopic, measurable, and more physically complete crowding index, namely the target crowding.

[0033] For example, suppose the system identifies n=2 neighboring particles within the neighborhood of the target particle, with the following parameters: Neighboring Particle 1: Distance between neighboring particles Particle size characterization The dimension is area; adjacent particle 2: distance between adjacent particles Particle size characterization The dimension is area; the system's preset distance attenuation index. The original congestion level After normalization , The result is rounded to five decimal places. Therefore, although the distance between adjacent particles of particle 1 is only half that of adjacent particle 2, its size is twice that of adjacent particle 2, and... Under the nonlinear effect, its final contribution to the crowding degree is more than 5 times that of the adjacent particle 2.

[0034] S2: Based on whether historical working conditions are stable, and considering the adjustment effect of the deviation of the target congestion degree from historical conditions, calculate the target deviation index to characterize the moving environment of the target particles.

[0035] It should be noted that the target congestion degree obtained in step S1 is a physical quantity. In order to make it easier for the system to apply this physical quantity to actual sorting, this invention introduces the concept of target deviation index. It is a normalized evaluation index, which aims to transform the target congestion degree into a state perception tool that can intuitively reflect the risk of target particle trajectory deviation. This allows the system to determine, based on its historical experience, whether the target congestion degree is in a stable sparse flow state or a volatile dense flow state during its movement from the initial positioning to the final sorting position.

[0036] Specifically, based on whether historical operating conditions are stable, and considering the moderating effect of the deviation of the target congestion degree from historical conditions, a target deviation index is calculated to characterize the movement environment of target particles, including: It should be noted that traditional methods lack effective utilization of historical operating conditions, often relying on a single, fixed judgment standard, which cannot adapt to changes in particle trajectories under different production conditions. In contrast, the target deviation index expression fully utilizes historical data, normalizing the current target congestion level through historical average congestion and historical congestion standard deviation. Facing different production conditions such as changes in material flow density, alterations in inter-particle electrostatic interactions, and fluctuations in environmental humidity, it accurately reflects the degree of deviation of the target particle movement environment from historical conditions, transforming it into an intuitive target deviation index. A positive index indicates that the target particle movement environment is more congested than historically, with a high risk of trajectory deviation; a negative index indicates that the target particle movement environment is sparser than historically, with a low risk of trajectory deviation.

[0037] It should be noted that the construction of the target deviation exponential expression in this invention is based on the objective laws of statistical process control and the general mathematical principles of industrial anomaly detection. Firstly, the expression follows the characteristics of anomaly identification: the greater the deviation of the current material flow congestion from historical normal conditions, the higher the risk of the target particle trajectory deviating from the ideal model; the more stable the historical operating conditions and the smaller the standard deviation of historical congestion, the more sensitive the system is to abnormal deviations in congestion. The standardized deviation term constructed from historical mean and standard deviation precisely conforms to this objective statistical law. Secondly, all input terms in the expression are dimensionless congestion-related indicators, which are processed by the hyperbolic tangent function. Mapped to the (-1,1) interval, the final output target deviation index is a dimensionless standardized index, which conforms to the rules of dimensional calculation; thirdly, The introduction of the function achieves smooth normalization of the degree of anomaly, avoids model oscillation caused by extreme deviations, and conforms to the engineering practice rules of industrial real-time control systems.

[0038] The target deviation index satisfies the following expression: ; In the formula, Indicates the deviation index from the target; Indicates the target congestion level; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; , Represents a very small positive number, ensuring that the denominator is not zero; It is the hyperbolic tangent function.

[0039] In the formula, This indicates the percentage deviation of the target congestion level from the standard deviation of historical congestion levels. As a sensitivity regulator, when historical operating conditions are stable... Small, It will increase, thus amplifying small congestion deviations and making the system more alert to anomalies; The function compresses this amplified deviation value into a range of -1 to +1, forming a standardized target deviation index. This indicates that the target particle's moving environment is more crowded and risky than in the past. This indicates that the target particle's moving environment is sparser than in the past, and the risk is low.

[0040] For example, the original congestion level calculated in the previous step can be used. Assuming historical average congestion Historical Crowding Standard Deviation This indicates that historical operating conditions are relatively stable, and the target deviation index is... , Rounded to two decimal places, this value close to +1 clearly indicates that the target particle's moving environment is much more crowded than usual, and the risk of trajectory deviation is very high.

[0041] S3: Obtain the real-time moving distance based on the positional changes of the target particle before and after its movement, and retrieve the total moving distance from the database; by integrating the real-time movement progress of the target particle, historical working conditions, and target deviation index, adaptively adjust the pre-acquired baseline process noise to obtain the target dynamic noise.

[0042] It should be noted that traditional sorting systems often ignore the dynamic changes in the impact of particle movement progress on disturbances in trajectory prediction, making it difficult to adapt to the environmental differences of particles from their initial position to their rejection position. This invention calculates the real-time movement distance by capturing the positional changes of particles before and after movement in real time, and completes normalization processing by combining it with the system's inherent total movement distance, thus accurately depicting the real-time movement progress of particles.

[0043] Specifically, the real-time moving distance is obtained based on the positional changes of the target particle before and after its movement, and the total moving distance is retrieved from the database, including: The pixel coordinates of the target particle acquired initially are taken as the initial target coordinates. During the movement of the target particle, the real-time position of the target particle is located according to the optical flow method to obtain the real-time position coordinates. The Euclidean distance between the initial target coordinates and the real-time position coordinates is calculated to obtain the real-time movement distance. The system's inherent distance is obtained from the database and recorded as the total movement distance.

[0044] It should be noted that the target dynamic noise is the core of the adaptive prediction in this invention. Its physical meaning in the scene is to dynamically adjust the weight of the process noise covariance in the motion prediction model. When the target deviation index is high, the noise weight is increased to make the model more dependent on the observed values; conversely, the weight is decreased to make the model more dependent on the model prediction. The baseline process noise represents the initial balanced state under ideal conditions, i.e., when the particle falls in a near-vacuum environment. The target dynamic noise in this invention satisfies the following condition: when the target particle's moving environment is crowded and the target deviation index is high, the target dynamic noise increases accordingly. This is equivalent to adding weight to the real-time observation position, making the system more confident in the observed actual position, thus enabling a rapid response to trajectory changes caused by collisions between the target particle and adjacent particles. Conversely, when the target particle's moving environment is sparse and the target deviation index is low, the target dynamic noise decreases, which is equivalent to adding weight to the model prediction trajectory, making the system more inclined to trust the smooth physical model.

[0045] It should be noted that traditional models use fixed process noise parameters, which cannot adapt to the environmental differences throughout the particle's movement, easily leading to an imbalance in the trust weights between the predicted trajectory and the actual observed position. In contrast, the target dynamic noise expression achieves adaptive adjustment of process noise through multi-dimensional dynamic correction. Based on the baseline process noise, the target dynamic noise expression integrates trajectory deviation risk assessment results, historical operating condition characteristics, and the real-time particle movement progress to construct a dynamic correction mechanism. Specifically, trajectory deviation risk directly regulates the basic correction direction of the noise parameters, historical operating condition characteristics determine the system's sensitivity to risk changes, and particle movement progress dynamically weakens the impact of the initial environmental assessment on the overall judgment through an attenuation term. This allows the noise parameters to be optimized in real time as the particle moves, better reflecting the scenario in actual production where particle trajectories are subject to dynamic interference from multiple factors.

[0046] Preferably, the target dynamic noise is obtained by adaptively adjusting the pre-acquired baseline process noise by fusing the real-time motion progress of the target particle, historical working conditions, and target deviation index, including: The baseline process noise is obtained by statistically analyzing the deviation between the historically least crowded sample trajectory and the ideal physical model.

[0047] The target dynamic noise satisfies the following expression: ; In the formula, For target dynamic noise; The deviation index from the target; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; Reference process noise; Indicates the real-time distance traveled; Indicates the total distance traveled; Represents the absolute value function; It represents a very small positive number, and guarantees that the denominator is not 0.

[0048] In the formula, As a basic adjustment amount, it directly reflects the level of risk. If the noise is amplified, the noise in the reference process is amplified; otherwise, it is reduced. This indicates that the real-time movement distance of the target particle has been normalized. Similar to the previous step, this is a sensitivity regulator. The more stable the historical operating conditions, the larger this value, and the higher the system's sensitivity to environmental anomalies. Part of it constitutes a dynamic decay term, which makes The influence gradually weakens as the target particle moves, which is consistent with the physical law that the influence of the initial environmental assessment will decrease over time, thus avoiding the excessive influence of the initial local state on the overall judgment. This represents a target dynamic noise calculation model based on baseline process noise plus multi-factor dynamic correction.

[0049] S4: Update the motion prediction model based on the target's dynamic noise, output the target's final coordinates and remove it; determine whether the removal was successful and store the structured data; perform regression analysis on multiple failed data in the structured data to optimize the motion prediction model.

[0050] It should be noted that motion prediction models with fixed parameters cannot cope with the sudden changes in trajectory caused by particle collisions in high-density material flows, and are prone to rejection position deviations. To address this issue, this invention incorporates target dynamic noise into the motion prediction model update process. By adjusting the model's trust weights for the predicted trajectory and the actual observed position in real time through target dynamic noise, the model can accurately calculate the time and final coordinates of particles reaching the rejection module, ensuring that the air nozzle is activated at the precise moment, and achieving accurate separation of contaminant particles.

[0051] It should be noted that the target dynamic noise calculated in the preliminary steps of the specification is itself a core parameter prepared for the motion prediction model. In the previous steps, the system completed the parameter calculation preparation; in this invention, the system applies the calculated parameters to actually generate the motion prediction model. Therefore, the inherent logic of the original description in the specification is that by processing the data obtained from regression analysis, the target dynamic noise in the motion prediction model is updated, thereby optimizing the motion prediction model as a whole. The two are not isolated or contradictory concepts, but rather represent the relationship between local parameters and the overall model, and between the calculation of preliminary parameters and the application of subsequent models. The motion prediction model of this invention adopts a closed-loop adaptive functional structure with dual inputs and a single output, which can be divided into three functional layers: an input layer, a core operation and processing layer, and an output layer. The input layer adopts a dual-path parallel input architecture. One path is the real-time position coordinates obtained through optical flow, which serves as the basic observation input for the model; the other path is the target dynamic noise calculated by fusing multi-dimensional working condition features, which serves as the core input for adaptive adjustment of the model, realizing real-time perception of material flow environmental risks. The core processing layer employs a three-step serial processing architecture: state prediction update, dynamic weighting of noise covariance, and trajectory trend extrapolation. This architecture enables adaptive adjustment of the prediction model and accurate calculation of the target particle trajectory. The output layer outputs the final coordinates of the target, directly connecting to the rejection execution module to complete the sorting action. This motion prediction model structure discloses its functional architecture and execution logic, breaking through the technical limitations of traditional fixed-parameter prediction models, forming a complete technical closed loop, and ensuring the feasibility and industrial application value of the solution.

[0052] Specifically, the motion prediction model is updated based on the target's dynamic noise, and the final coordinates of the target are output and then discarded, including: The system updates the motion prediction model based on the target dynamic noise and calculates the final coordinates of the target particle when it will arrive at the removal execution module at a certain predicted time. The final coordinates of the target particle are sent to the removal execution module. At the same time, the air nozzle that precisely corresponds to the final coordinates of the target particle is activated instantly and sprays air to accurately separate the target particle from the main material flow.

[0053] It should be noted that this step aims to visually confirm the direct consequences of the sorting operation and to correlate the sorting results with key parameters in the prediction process, providing high-quality input data for subsequent system self-optimization. This invention closely connects the sorting results in the physical world with the prediction model in the digital world, which is a key link in realizing system intelligence.

[0054] Preferably, determining whether the removal was successful and storing the structured data includes: A verification camera is set downstream of the rejection execution module. This camera is used to capture images of the material flow that has just passed through the rejection area. The system analyzes the images captured by the verification camera to determine whether the target particles have been completely separated from the main material flow and whether no new contaminant particles appear in the image. If the target particles are completely separated and there are no new contaminants, the rejection is considered successful. If the target particles are still in the main material flow or new contaminants appear, the rejection is considered a failure. The system records the verification results of this rejection task and all the parameters in this rejection task process, including target congestion, target deviation index, target dynamic noise, and target final coordinates, as structured data.

[0055] It should be noted that this step utilizes a large amount of structured data with success or failure labels to optimize and adjust the core model of the entire sorting system, so as to continuously improve the accuracy and robustness of sorting. This self-evolution capability based on long-term data statistics and machine learning ideas enables the sorting system of this invention to adapt to the ever-changing production environment and upgrade from a fixed parameter model to an adaptive model with continuous optimization capabilities.

[0056] Preferably, regression analysis is performed on multiple failed data points in the structured data to optimize the motion prediction model, including: The system periodically collects and analyzes structured data, including failed data with failure labels. It then performs regression analysis on the trajectory data of multiple failed data with failure labels to update the target dynamic noise in the motion prediction model and optimize the motion prediction model.

[0057] This invention also discloses a visual sorting system for contaminants in insulating material particles, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a visual sorting method for contaminants in insulating material particles according to the present invention.

[0058] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0059] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for visually sorting contaminants in insulating material particles, characterized in that, include: The target particle is initially located based on a neural network model. The particle size, number of adjacent particles, and distance between adjacent particles are obtained based on the morphological characteristics and distribution of the target particle and its neighboring particles. The crowding degree of the target particle is obtained by accumulating the crowding contribution of each individual neighboring particle within the preset neighborhood of the target particle, based on the size and distance of the adjacent particles. Based on whether historical working conditions are stable, and considering the moderating effect of the deviation of the target congestion degree from historical conditions, a target deviation index is calculated to characterize the moving environment of target particles. The real-time moving distance is obtained based on the positional changes of the target particle before and after its movement, and the total moving distance is obtained from the database. The target dynamic noise is obtained by adaptively adjusting the pre-acquired baseline process noise by fusing the real-time movement progress of the target particle, historical working conditions, and target deviation index. The motion prediction model is updated based on the target's dynamic noise, the final coordinates of the target are output and then removed; the success of the removal is determined and the structured data is stored. Regression analysis was performed on multiple failed data points in the structured data to optimize the motion prediction model.

2. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The process of obtaining particle characterization size, number of adjacent particles, and distance between adjacent particles includes: Based on the visual features of material particles such as color, texture, and shape in real-time materials, contaminant particles are accurately identified and segmented from the background of normal insulating material particles and recorded as target particles. The particle size is obtained by calculating the number of pixels contained in the segmentation mask of the material particles. The pixel coordinates of the target particles are output, and all neighboring particles are identified in the preset neighborhood of the target particles using image processing algorithms to obtain the number of neighboring particles. The pixel distance between all neighboring particles and the target particles is calculated and converted according to the camera calibration parameters. The pixel distance between each neighboring particle and the target particle is recorded as the neighboring particle distance.

3. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The target congestion level satisfies the following expression: ； In the formula, Indicates the target congestion level; The target particle's first The particle size of each adjacent particle; Indicates the target particle and the first The distance between adjacent particles of a neighboring particle; This represents the pre-obtained distance decay index; Indicates the number of adjacent particles of the target particle; Represents a very small positive number, ensuring that the denominator is not zero; This represents the normalization function.

4. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The target deviation index satisfies the following expression: ； In the formula, Indicates the deviation index from the target; Indicates the target congestion level; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; , Represents a very small positive number, ensuring that the denominator is not zero; It is the hyperbolic tangent function.

5. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The step of obtaining the real-time moving distance based on the positional changes of the target particle before and after its movement, and retrieving the total moving distance from the database, includes: The pixel coordinates of the target particle acquired initially are taken as the initial target coordinates. During the movement of the target particle, the real-time position of the target particle is located according to the optical flow method to obtain the real-time position coordinates. The Euclidean distance between the initial target coordinates and the real-time position coordinates is calculated to obtain the real-time movement distance. The system's inherent distance is obtained from the database and recorded as the total movement distance.

6. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The target dynamic noise satisfies the following expression: ； In the formula, For target dynamic noise; The deviation index from the target; Indicates the historical average congestion level; Indicates the standard deviation of historical crowding; Reference process noise; Indicates the real-time distance traveled; Indicates the total distance traveled; Represents the absolute value function; It represents a very small positive number, and guarantees that the denominator is not 0.

7. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The final coordinates of the output target are then discarded, including: The system updates the motion prediction model based on the target dynamic noise and calculates the final coordinates of the target particle when it will arrive at the removal execution module at a certain predicted time. The final coordinates of the target particle are sent to the removal execution module. At the same time, the air nozzle that precisely corresponds to the final coordinates of the target particle is activated instantly and sprays air to accurately separate the target particle from the main material flow.

8. The visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The process of determining whether the removal was successful and storing structured data includes: A verification camera is set downstream of the rejection execution module. This camera is used to capture images of the material flow that has just passed through the rejection area. The system analyzes the images captured by the verification camera to determine whether the target particles have been completely separated from the main material flow and whether no new contaminant particles appear in the image. If the target particles are completely separated and there are no new contaminants, the rejection is considered successful. If the target particles are still in the main material flow or new contaminants appear, the rejection is considered a failure. The system records the verification results of this rejection task and all the parameters in this rejection task process, including target congestion, target deviation index, target dynamic noise, and target final coordinates, as structured data.

9. A visual sorting method for contaminants in insulating material particles according to claim 1, characterized in that, The optimized motion prediction model includes: The system periodically collects and analyzes structured data, including failed data with failure labels. It then performs regression analysis on the trajectory data of multiple failed data with failure labels to update the target dynamic noise in the motion prediction model and optimize the motion prediction model.

10. A visual sorting system for contaminants in insulating material particles, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a visual sorting method for contaminants in insulating material particles according to any one of claims 1-9.

Images