Fire retardant particle quality on-line monitoring and sorting system based on AI visual identification

Abstract

The invention relates to the technical field of flame retardant particle quality monitoring, in particular to a flame retardant particle quality online monitoring and sorting system based on AI visual identification. The system comprises a polarized light field imaging module, a refractive index gradient field construction module, a dynamic sorting decision module and an online self-optimization module. A Stokes parameter image and a multispectral transmittance distribution diagram of the particles are obtained through a polarized light field imaging module; the refractive index gradient field construction module is used for converting optical data into a refractive index gradient field and accurately identifying internal defects; the dynamic sorting decision module generates a sorting instruction according to the defect characteristics; and the online self-optimization module realizes continuous optimization of the system. The method solves the problem that a traditional method is difficult to accurately identify transparent defects, can effectively identify micron-sized pure bubbles and transparent cracks, realizes defect classification and cause analysis, can meet the real-time quality monitoring requirement of a flame retardant production line, and provides an intelligent solution for product quality control.

Description

Technical Field

[0001] The present invention relates to the technical field of flame retardant particle quality monitoring, and in particular to an online flame retardant particle quality monitoring and sorting system based on AI visual recognition. Background Art

[0002] Flame retardant particles are key additives for polymer materials, and their quality directly affects the fire resistance of the final product. With the development of intelligent manufacturing technology, online monitoring systems based on AI visual recognition have gradually replaced traditional manual spot checks. They use high-resolution industrial cameras to capture particle surface features (such as cracks, color differences, etc.) and combine them with deep learning algorithms to achieve real-time sorting. However, for translucent / transparent flame retardant particles (such as polycarbonate flame retardant masterbatch and optical-grade flame retardant microspheres), conventional surface detection technology cannot penetrate the interior of the material. The current mainstream solution uses X-ray microfocus tomography technology to reconstruct the internal structure of the particles in three dimensions and identify defects based on density differences.

[0003] While existing X-ray microfocus tomography technology can analyze internal structures, it still has significant shortcomings in detecting abnormal refractive index defects in transparent materials. Specifically, X-ray imaging relies on the physical principle of material density differences. However, defects such as pure bubbles and transparent impurities have minimal density differences from the base material, making their internal refractive index gradients difficult to characterize. Such defects lack significant contrast in X-ray images, resulting in a high rate of missed detection of tiny transparent bubbles and cracks. Furthermore, it is difficult to distinguish between true internal defects and surface optical artifacts, severely restricting the demand for high-precision transparent flame-retardant material quality control. Summary of the Invention

[0004] The present invention provides an online monitoring and sorting system for flame retardant particle quality based on AI visual recognition. This system decomposes mixed polarization signals into azimuthal and radial components to separate surface noise, then fuses characteristic band spectral data to construct an optical feature volume. Finally, a dual-channel network is used to generate a refractive index gradient field and analyze defect characteristics, thereby solving the problems raised in the above-mentioned background technology, namely:

[0005] Micron-level defects inside translucent flame retardant particles (such as pure bubbles and transparent cracks) are difficult to detect accurately.

[0006] To achieve the above objectives, the flame retardant particle quality online monitoring and sorting system includes a polarized light field imaging module, which is used to collect and pre-process the transmitted light image of the flame retardant particles to generate a Stokes parameter image and a multispectral transmittance distribution map. It also includes a refractive index gradient field construction module and a dynamic sorting decision module.

[0007] The refractive index gradient field construction module is used to perform defect mechanism analysis on the refractive index gradient field to generate a three-dimensional defect heat map, and the three-dimensional defect heat map includes:

[0008] A defect space coordinate layer for locating the physical position of the defect space;

[0009] A type code layer used to identify the physical category of the defect;

[0010] Probability of cause layer used to associate defects with production process parameters;

[0011] The steps of constructing the refractive index gradient field by the refractive index gradient field construction module include:

[0012] Perform polarization feature decoupling processing on the Stokes parameter image to separate the azimuthal polarization component and the radial polarization component;

[0013] Perform characteristic band extraction processing on the multi-spectral transmittance distribution map and screen the characteristic band spectral data;

[0014] The radial polarization component is fused with the characteristic band spectral data to generate an optical feature volume, and a dual-channel network architecture is used to construct a refractive index gradient field;

[0015] The dynamic sorting decision module sorts out flame retardant particles with internal defects based on the defect three-dimensional thermal map.

[0016] The above-mentioned technical solution, through the innovative combination of multimodal optical feature fusion and physical constraint modeling, overcomes the dual limitations of traditional detection methods in terms of single information dimension and algorithmic interpretability. Relying solely on single polarization imaging, while it can resolve light wave phase changes, cannot eliminate the interference of surface scattering on internal defect signals, making it difficult to distinguish between tiny bubbles and surface scratches. Using spectral analysis alone, while it can reflect material absorption characteristics, it loses the vectorial characteristics of defect morphology, making it impossible to obtain key information such as crack propagation direction. This solution constructs a multi-dimensional fused optical feature volume by decoupling polarization features to remove surface noise and filtering characteristic wavelengths to enhance refractive index correlation, fundamentally solving the problem of extracting effective information from mixed signals. The dual-channel network design achieves a breakthrough at the algorithmic level: the optical feature channel retains the detail sensitivity of the measured data, while the material property channel embeds the inherent constraints of physical laws. The synergy between the two avoids the over-reliance on manual experience of traditional threshold segmentation methods while overcoming the failure risk of purely data-driven models for unknown defect types.

[0017] On this basis, the dual-channel network architecture includes an optical characteristic channel and a material characteristic channel. The optical characteristic channel processes the polarization phase angle change data, and the material characteristic channel generates physical constraints by combining the characteristic band transmittance ratio and the material intrinsic parameters.

[0018] In another technical solution, the defect mechanism analysis includes:

[0019] The refractive index gradient field construction module calculates the second-order derivative of each spatial point and locates the abnormal mutation area based on the vector matrix of the refractive index gradient field;

[0020] The refractive index gradient field construction module distinguishes radially divergent defects from linearly arranged defects based on the gradient vector distribution pattern of the abnormal mutation area;

[0021] The refractive index gradient field construction module performs back propagation calculation based on the direction of the gradient vector to determine the physical origin position of the defect.

[0022] This technical solution overcomes the structural defects of traditional defect detection methods in feature characterization and cause tracing through a hierarchical design of dual-channel collaborative analysis and gradient field vector analysis. If only the optical feature channel is used, although the polarization phase change caused by the defect can be captured, the process fluctuation will be misjudged as a defect due to the lack of material property constraints; if only the material characteristic channel is relied upon, although the consistency of physical laws can be guaranteed, the morphological characteristics of subtle defects will be lost, resulting in the missed detection of micron-level cracks. This solution uses a dual-path verification mechanism that extracts measured signals through the optical channel and injects physical priors into the material channel, which not only prevents the risk of overfitting driven by pure data, but also avoids the lack of sensitivity of pure theoretical models. In the defect analysis stage, second-order derivative calculations locate the basic defects in the abnormal area, gradient vector pattern analysis realizes the intelligent classification of bubbles and cracks, and backpropagation tracing reflects the process cause of the defect, so that the system can not only detect the existence of defects, but also understand the nature of the defects.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] Through multi-dimensional analysis of the refractive index gradient field, an integrated diagnostic capability has been achieved, from three-dimensional morphological reconstruction of flame retardant particle quality defects to inference of physical causes and then to process correlation analysis. Specifically, the construction of the gradient vector field not only reflects the geometric characteristics of the defects, but also reveals the potential correlation between defects and process conditions through the constraints of the material's intrinsic parameters. The collaborative mechanism of the dual-channel network breaks through the traditional algorithm's dilemma of choosing between detection sensitivity and result credibility, giving the system the ability to adaptively identify unknown defect types. This analytical paradigm, which deeply couples optical characteristics, material properties, and production processes, provides complete technical support for flame retardant production, from simple sorting to quality traceability, and realizes the transformation of quality control from "post-elimination" to "process prevention." BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the overall process structure of the present invention;

[0026] Figure 2 This is a schematic diagram of the process of the polarized light field imaging module of the present invention;

[0027] Figure 3 This is a schematic diagram of the refractive index gradient field construction module flow of the present invention.

[0028] The meaning of each number in the figure is:

[0029] 100. Polarized light field imaging module; 200. Refractive index gradient field construction module; 300. Dynamic sorting decision module; 400. Online self-optimization module. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Currently, it is difficult to accurately detect micron-level defects (such as pure bubbles and transparent cracks) inside translucent flame retardant particles. The present invention provides an online monitoring and sorting system for flame retardant particle quality based on AI visual recognition. Figure 1 As shown, it includes a polarized light field imaging module 100, a refractive index gradient field construction module 200, a dynamic sorting decision module 300 and an online self-optimization module 400. Through the collaborative work of multiple modules, high-precision identification and automatic sorting of internal defects of translucent flame retardant particles are achieved.

[0032] The polarized light field imaging module 100 of the present invention is located above the production line conveyor belt and is primarily used to obtain high-quality optical characteristic data of translucent flame retardant particles. Through its specially designed imaging system, the polarized light field imaging module 100 is able to penetrate the particle surface and capture internal optical information that is inaccessible to traditional vision systems.

[0033] like Figure 2 As shown, the core of the polarization light field imaging module 100 consists of three components: an adjustable polarization light source system, a multispectral polarization camera array, and a real-time image processing unit. The adjustable polarization light source system uses a circular array of LEDs that can generate incident light with different polarization directions (0°, 45°, 90°, and 135°). Through precise control, the polarization state can be rapidly switched. The multispectral polarization camera array consists of four highly sensitive industrial cameras, each equipped with a specific wavelength filter and polarizer, which can simultaneously capture images of transmitted light in different polarization states.

[0034] During operation, flame retardant particles on a conveyor belt first pass through an area illuminated by a polarized light source. The system sequentially emits incident light with four different polarization directions, while a multispectral camera array simultaneously captures corresponding transmission images. This multi-angle, multispectral imaging method maximizes the extraction of internal optical characteristics of the particles. The captured raw images are immediately transmitted to a real-time image processing unit for pre-processing.

[0035] The image processing unit performs three key steps on the raw data: first, background noise is removed to eliminate ambient light interference; second, image registration is performed to ensure strict alignment of images in different polarization states; and finally, a specific algorithm is used to synthesize a Stokes parameter image containing complete polarization information. A multispectral transmittance distribution map is then calculated based on the transmitted light intensity at different wavelengths. These steps ensure that subsequent modules receive accurate and reliable input data.

[0036] After processing, the module outputs two sets of key data: a Stokes parameter image and a multispectral transmittance distribution map. The Stokes parameter image fully records the optical response characteristics of the particle under various polarization states and contains rich internal structural information. The multispectral transmittance distribution map reflects the particle's absorption characteristics for light of different wavelengths. The image processing unit calculates the ratio of transmitted light intensity at each wavelength by analyzing the ratio of transmitted light intensity at each wavelength.

[0037] Particularly noteworthy is the module's unique ring-shaped light source design and multi-camera synchronous acquisition solution, ensuring clear images of internal structures even at high production line speeds (particle movement speeds of up to 3m / s). Furthermore, the module's built-in self-calibration function automatically compensates for imaging errors caused by ambient temperature fluctuations or equipment vibration, ensuring long-term stable operation.

[0038] However, these optical signatures alone cannot directly determine internal defects in particles. While Stokes parameter images record rich polarization information, the surface scattering effects they contain can interfere with accurate identification of internal defects. Multispectral transmittance profiles, while reflecting the optical properties of the material, lack a precise description of refractive index variations. To accurately detect internal defects, these optical signatures must be further converted into physical fields that can characterize the material's internal structure.

[0039] To this end, we introduced a refractive index gradient field construction module 200. The polarized light field imaging module 100 transmits the processed Stokes parameter image and multispectral transmittance distribution map to the refractive index gradient field construction module 200. The Stokes parameter image provides the optical response characteristics of the particles under different polarization states, and the multispectral transmittance distribution map contains information on the light absorption characteristics of the material. These two sets of data together constitute the basic input for constructing the refractive index gradient field.

[0040] The refractive index gradient field construction module 200 is the core processing unit of the entire inspection process. Its primary function is to convert the optical characteristic data acquired by the polarized light field imaging module 100 into a physical field that accurately characterizes the material's internal structure. Through innovative algorithm design, this module achieves precise identification and location of defects within translucent flame retardant particles.

[0041] like Figure 3 As shown, the refractive index gradient field construction module 200 first receives two sets of key input data from the polarization light field imaging module 100: the Stokes parameter image and the multi-spectral transmittance distribution map, and performs initial processing on them. In the initial processing stage, the module first performs polarization feature decoupling processing on the Stokes parameter image. The mixed optical signal is accurately separated into azimuthal polarization component and radial polarization component through the vector decomposition algorithm. Among them, the azimuthal polarization component mainly reflects the scattering noise characteristics of the particle surface, while the radial polarization component focuses on the changes in the refractive effect caused by internal defects. At the same time, the multi-spectral transmittance distribution map is processed through characteristic band extraction to screen out the characteristic band spectral data that is most relevant to the refractive index change of the material.

[0042] Next, the refractive index gradient field construction module 200 deeply fuses the processed radial polarization component with the characteristic band spectral data to generate an optical feature body. This feature body accurately records the key optical properties of each position inside the particle, including core parameters such as polarization phase angle change data and specific wavelength transmittance ratio. Through this fusion process, the system effectively distinguishes between surface scattering effects and internal defect characteristics, ensuring that subsequent analysis is only targeted at real internal defect signals, significantly improving the accuracy of detection. Among them, the polarization phase angle change data characterizes the birefringence effect caused by the defect, and the characteristic band transmittance ratio is directly related to the local refractive index change of the material. These two types of parameters together constitute the basic feature set for physical modeling.

[0043] In view of the two key parameters contained in the optical feature body, namely the polarization phase angle change data and the characteristic band transmittance ratio, the module adopts an innovative dual-channel network architecture to realize the reconstruction of the refractive index gradient field, among which: the optical feature channel (lower branch) specializes in processing the polarization phase angle change data, and adaptively extracts the gradient characteristics of particles of different particle sizes through the deformable convolution layer. The size of the receptive field dynamically adjusted by this channel directly depends on the spatial resolution recorded by the optical feature body, ensuring that the feature extraction accuracy strictly matches the input data characteristics. The material characteristic channel (upper branch) simultaneously utilizes the characteristic band transmittance ratio and the pre-stored material intrinsic parameters (dielectric constant, magnetic permeability), and converts these physical quantities into physical constraints through a special encoder. This processing enables the known correlation between the characteristic band transmittance ratio and the local refractive index to be directly embedded in the network structure.

[0044] When the two channels perform feature stitching at the fusion layer, they prioritize preserving the measured gradient features extracted by the optical feature channel while simultaneously verifying and correcting them using the physical constraints generated by the material property channel. This dual verification mechanism ensures that the final output refractive index gradient field is both faithful to the measured data of the optical feature volume and strictly conforms to the physical laws of electromagnetic wave propagation.

[0045] The generated refractive index gradient field accurately characterizes the refractive index variation characteristics at each location within the particle in the form of a three-dimensional vector matrix. Each spatial point corresponds to a three-dimensional vector, whose components represent the refractive index gradient of that point in three orthogonal directions of the spatial rectangular coordinate system. The vector modulus quantitatively characterizes the severity of the local refractive index change, while the vector direction reveals the energy diffusion trend in the defect area. This vector field-based expression method not only breaks through the limitation of traditional grayscale images that can only reflect a single dimension, but also captures the three-dimensional expansion characteristics of defects through the spatial correlation of vectors, providing richer physical information support for subsequent analysis.

[0046] The refractive index gradient field construction module 200 performs in-depth defect mechanism analysis on the constructed refractive index gradient field. First, the second-order derivative of the gradient field is calculated based on the vector matrix of the refractive index gradient field. By setting a dynamic threshold related to the material's elastic modulus, the region of abnormal mutation is precisely located. Next, based on the gradient vector distribution pattern in the region of abnormal mutation, radially divergent defects are distinguished from linearly arranged defects. Bubble defects typically exhibit a radially divergent vector field distribution, while crack defects exhibit a linearly ordered vector arrangement. Finally, a backpropagation calculation is performed along the gradient vector direction to trace the defect's origin and, in combination with a database of production process parameters, infer the possible cause. Through this analysis, the system generates a three-dimensional defect heat map containing three key information layers: a defect spatial coordinate layer precisely locates the abnormal region using three-dimensional coordinates; a type code layer identifies the defect category based on the gradient field's characteristic patterns; and a cause probability layer, combined with a process knowledge base, assesses the impact of various production parameters. This structured data is directly transmitted to the dynamic sorting decision module 300, providing comprehensive data support for automated sorting. The entire process is accelerated by industrial-grade GPUs to ensure the real-time performance of the production line is met.

[0047] Since the three-dimensional defect thermal map output by the refractive index gradient field construction module 200 only contains the physical characteristic information of the defects and cannot directly guide the sorting equipment to perform operations, the present invention designs a dynamic sorting decision module 300, which receives the three-dimensional defect thermal map from the upstream, including three types of structured information: defect spatial coordinate layer, type code layer and cause probability layer.

[0048] The dynamic sorting decision module 300 first normalizes the coordinate system of the defect space coordinate layer, converting the three-dimensional thermal map coordinates into a two-dimensional mapping corresponding to the physical coordinates of the production line conveyor belt. This process dynamically adjusts the coordinate offset based on the current conveyor belt speed to ensure positioning accuracy is not affected by material movement. Simultaneously, the type code layer data is parsed by a classifier and converted into command codes recognizable by the sorting equipment. Bubble defects are assigned a gentle spray mode, while crack defects trigger a strong rejection command.

[0049] The core sorting process of the dynamic sorting decision module 300 utilizes a multi-dimensional decision-making mechanism. The system first calculates the precise sorting trigger time based on the defect's spatial coordinate layer data and real-time conveyor speed. Simultaneously, it automatically matches the preset sorting intensity level based on the defect category identified in the type code layer. For defects marked as highly process-relevant in the cause probability layer, the system prioritizes sorting resources, ensuring that critical quality issues are addressed promptly.

[0050] The resulting sorting instruction set uses a structured data format and contains three core fields: a time control field records the trigger timestamp and duration with millisecond-level accuracy; an execution parameter field specifies the valve injection force level and target defect type code; and a quality control field carries the inspection confidence score and process association identifier. These instructions are transmitted to the sorting actuator in real time via industrial Ethernet, with a transmission cycle strictly controlled within 50ms to ensure timing accuracy.

[0051] During system operation, a real-time monitoring mechanism for sorting effectiveness is established, continuously tracking key indicators such as command response latency, classification rejection success rate, and false rejection rate. If the system detects that the same defect type has not been effectively rejected three times in a row, the system automatically initiates an intensity escalation protocol: first, the system increases the sorting intensity level for that defect type within the safety threshold, simultaneously marks the relevant case as a sample for optimization, and sends a priority processing request to the online self-optimization module 400.

[0052] All sorting operations generate complete, time-stamped records, which are synchronized in real time to the online self-optimization module 400 via a standardized interface. Each record consists of four components: the original sorting instructions, the executed equipment status parameters, the actual sorting verification results, and the system's automatic adjustment log. This data, along with the corresponding original defect 3D heat map and a snapshot of the production line's process parameters, constitutes a closed-loop optimization data package, providing a comprehensive basis for the system's continuous self-optimization.

[0053] The online self-optimization module 400, serving as the system's intelligent optimization core, receives the sorting instruction set and corresponding 3D defect thermal map data transmitted by the dynamic sorting decision module 300 in real time. The sorting instruction set fully records the execution parameters of each sorting operation, including the trigger timestamp accurate to the millisecond, the air valve injection intensity level adjustable from 1 to 8 levels, and the defect type code. The received 3D defect thermal map contains the raw inspection data corresponding to the sorting operation, namely the complete spatial coordinate layer, type code layer, and cause probability layer generated by the refractive index gradient field construction module 200. This data is transmitted in encrypted form via industrial Ethernet to ensure data integrity and timeliness.

[0054] Based on the received data, the online self-optimization module 400 performs intelligent optimization analysis. First, it establishes a sorting effect evaluation model to analyze the actual rejection performance of different defect types under different sorting parameters. Then, it combines the characteristic layer data from the defect 3D heat map to identify the optimization space for the existing sorting rules. Finally, it generates two sets of optimization solutions: a dual-channel network parameter update solution for the refractive index gradient field construction module 200, focusing on optimizing defect detection accuracy; and a sorting rule adjustment recommendation for the dynamic sorting decision module 300, improving the matching between intensity level and defect type. All optimization instructions are fed back to the corresponding modules through standard interfaces, forming a closed-loop optimization system that continuously improves itself.

[0055] In summary, the present invention combines polarization optical detection with material physics modeling to generate a three-dimensional refractive index gradient field. This gradient field accurately quantifies the refractive index changes at each point inside the particle. It can not only penetrate the surface of the material to detect pure bubbles and transparent cracks, but also automatically distinguish the defect type based on the gradient distribution characteristics. This effectively solves the industry problem that traditional methods have difficulty in accurately identifying transparent defects, and realizes all-round quality monitoring and sorting of flame retardant particles from the surface to the inside.

[0056] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An online monitoring and sorting system for flame retardant particle quality based on AI visual recognition, comprising a polarized light field imaging module (100), wherein the polarized light field imaging module (100) is used to collect transmitted light images of flame retardant particles and pre-process the images to generate Stokes parameter images and multi-spectral transmittance distribution maps, characterized in that: It also includes a refractive index gradient field construction module (200) and a dynamic sorting decision module (300); The refractive index gradient field construction module (200) is used to perform defect mechanism analysis on the refractive index gradient field to generate a three-dimensional defect thermal map, wherein the three-dimensional defect thermal map includes: A defect space coordinate layer for locating the physical position of the defect space; A type code layer used to identify the physical category of the defect; Probability of cause layer used to associate defects with production process parameters; The refractive index gradient field construction module (200) constructs the refractive index gradient field in the following steps: Perform polarization feature decoupling processing on the Stokes parameter image to separate the azimuthal polarization component and the radial polarization component; Perform characteristic band extraction processing on the multi-spectral transmittance distribution map and screen the characteristic band spectral data; The radial polarization component is fused with the characteristic band spectral data to generate an optical feature volume, and a dual-channel network architecture is used to construct a refractive index gradient field; The dynamic sorting decision module (300) sorts out flame retardant particles with internal defects based on the defect three-dimensional thermal map.

2. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The polarized light field imaging module (100) comprises an adjustable polarized light source system, wherein the adjustable polarized light source system adopts an LED array arranged in a ring and is used for switching polarization states.

3. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 2 is characterized by: The polarized light field imaging module (100) further comprises a multispectral polarization camera array and a real-time image processing unit. The multispectral polarization camera array synchronously collects transmitted light images in different polarization states. The real-time image processing unit performs background noise elimination and image registration on the transmitted light images to synthesize the Stokes parameter image and the multispectral transmittance distribution map.

4. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The optical feature body records the optical characteristics of each position inside the particle, including polarization phase angle change data and specific wavelength transmittance ratio.

5. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The dual-channel network architecture includes an optical characteristic channel and a material characteristic channel. The optical characteristic channel processes polarization phase angle change data, and the material characteristic channel generates physical constraints by combining characteristic band transmittance ratios and material intrinsic parameters.

6. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The defect mechanism analysis includes: The refractive index gradient field construction module (200) calculates the second-order derivative of each spatial point and locates the abnormal mutation area based on the vector matrix of the refractive index gradient field; The refractive index gradient field construction module (200) distinguishes radially divergent defects from linearly arranged defects based on the gradient vector distribution pattern of the abnormal mutation area; The refractive index gradient field construction module (200) performs back propagation calculation based on the direction of the gradient vector to determine the physical origin position of the defect.

7. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The dynamic sorting decision module (300) receives a three-dimensional thermal map of defects and calculates a sorting triggering time point based on the defect spatial coordinate layer and in combination with the conveyor belt speed.

8. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 1 is characterized by: The dynamic sorting decision module (300) determines the sorting intensity level based on the type code layer, and generates a sorting instruction set including a time control field, an execution parameter field and a quality control field, which is transmitted to the sorting execution mechanism and the online self-optimization module (400).

9. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 8 is characterized by: The online self-optimization module (400) receives a sorting instruction set and its corresponding three-dimensional defect heat map, analyzes the correlation between the sorting effect and the defect characteristics, and generates a dual-channel network parameter update plan and a sorting rule adjustment suggestion.

10. The flame retardant particle quality online monitoring and sorting system based on AI visual recognition according to claim 9 is characterized in that: The online self-optimization module (400) transmits the network parameter update plan to the refractive index gradient field construction module (200), and transmits the sorting rule adjustment suggestion to the dynamic sorting decision module (300).

Images