Method and system for screening activated carbon raw material based on multi-modal image feature fusion

Abstract

The present application relates to the field of active carbon raw material screening, and provides an active carbon raw material screening method and system based on multi-modal image feature fusion, which comprises the following steps: collecting edge contour original images and surface texture original images, constructing an edge contour binary image set and a surface texture gray image set; extracting an edge chain code to generate an edge curvature sequence, extracting an effective cutting feature vector based on a coating damage critical threshold, and mapping to generate a physical wear risk index; mapping a texture region of interest to a three-dimensional gray topological surface, calculating a fractal dimension characteristic value and a moisture absorption texture proportion, and coupling to generate an environmental moisture absorption preemption index; calculating a discoloration pearl compatibility score through an environment-medium double constraint weighted algorithm, combining an optimal threshold, an elimination threshold and physicochemical detection data to perform a multi-path shunt decision, and generating a particle grading control instruction. Through simultaneous time division light field imaging and multi-modal feature coupling, the present application realizes intelligent screening and resource scheduling of active carbon particle micro-morphology.

Description

Technical Field

[0001] This invention relates to the field of machine screening of activated carbon raw materials, and in particular to a method and system for screening activated carbon raw materials based on multimodal image feature fusion. Background Technology

[0002] With the widespread application of activated carbon and color-changing catalyst blends in air purification, environmental monitoring, and other fields, compatibility assessment and precise sorting of components have become crucial for ensuring the quality of finished products. In the dynamic production environment of activated carbon particles, how to overcome the limitations of traditional physicochemical testing indicators in identifying microscopic morphological defects, and how to identify and remove potentially incompatible particles in real time, has become a critical technical challenge that urgently needs to be addressed in the process of moving from single-indicator testing to refined microscopic morphological quality control in activated carbon raw material screening.

[0003] Chinese patent CN116843692B, authorized by patent number CN116843692B, provides an artificial intelligence-based method for detecting the state of regenerated activated carbon. This method includes: acquiring several grayscale images of activated carbon; training a neural network on the grayscale images to obtain several activated carbon regions; obtaining anomaly reference values ​​based on the activated carbon regions; obtaining an initial anomaly degree based on the anomaly reference values; obtaining a reference pixel column based on the columnar direction of each activated carbon region; obtaining an initial mold degree based on the initial anomaly degree; obtaining a mold reference pixel column based on the reference pixel column; obtaining the mold degree based on the mold reference pixel column and the initial mold degree; and obtaining an enhancement coefficient based on the mold degree; thereby performing activated carbon state detection.

[0004] However, current technology still faces many challenges. On automated sorting production lines for formaldehyde-removing activated carbon bags, existing machine vision systems typically rely on macroscopic physicochemical indicators, such as iodine value, for raw material screening, making it difficult to identify microscopic morphological defects on the surface of activated carbon particles. When activated carbon particles have microscopic, sharp edges or dense, highly hygroscopic nanopores, these particles can cause mechanical damage to the indicator coating of the color-changing beads during mixing, packaging, and transportation, or absorb the crucial moisture required to maintain color change within a short period. If the screening system cannot remove these abnormal particles with potential bioincompatibility risks at the microscopic level, the color-changing beads in the finished activated carbon bags may experience false color change or pulverization failure without adsorbing formaldehyde, directly interfering with the judgment of product status and causing batch quality risks due to the failure of the product's indicator function. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a method for screening activated carbon raw materials based on multimodal image feature fusion, the specific technical solution of which is as follows:

[0006] The original edge contour image and the original surface texture image of the activated carbon particle flow to be screened are acquired in a single frame. The topology-preserving segmentation algorithm based on variational level set is performed on the original edge contour image to construct a set of binary edge contour images. The original surface texture image is introduced with a homomorphic filtering algorithm to construct a set of grayscale surface texture images.

[0007] The edge contour binary image set is traversed to extract the edge chain code. The edge chain code is then reconstructed by subpixel-level parameterization to generate an edge curvature sequence. Based on a preset coating damage critical threshold, an effective cutting feature vector is extracted from the edge curvature sequence. A physical wear risk index is generated through probability mapping.

[0008] The system iterates through the set of grayscale surface textures, maps the regions of interest in the textures to three-dimensional grayscale topological surfaces, and uses the differential box-counting dimension method to calculate the fractal dimension eigenvalues. It then performs a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, constructs a local frequency domain energy distribution field based on a preset sliding window step size, and calculates the proportion of moisture-absorbing textures through binarization segmentation. Finally, it uses exponential multiplicative coupling logic to perform nonlinear probability mapping on the fractal dimension eigenvalues ​​and the proportion of moisture-absorbing textures to generate an environmental moisture-absorbing preemption index.

[0009] The physical wear risk index and environmental moisture absorption preemption index are obtained. An environmental-medium dual-constraint weighted algorithm is introduced to calculate the compatibility score of the color-changing beads. Combined with the preset optimization threshold and elimination threshold and the physicochemical test data of the activated carbon raw material batch, a multi-path diversion decision is executed to generate particle classification control instructions pointing to different diversion states.

[0010] Furthermore, the method for constructing the surface texture grayscale image set includes:

[0011] Two light field slice data acquisitions were performed on the activated carbon particle flow to be screened at the same spatial location. During the first light field slice data acquisition, the backlight illumination module was used to acquire a single frame of the original edge contour image. During the second light field slice data acquisition, the side structured light illumination module was switched to acquire a single frame of the original surface texture image.

[0012] Based on the original image of a single frame edge contour, a topology-preserving segmentation algorithm based on variational level sets is constructed to generate a binary image of the single frame edge contour and to build a set of binary edge contour images.

[0013] A pixel-by-pixel logical AND operation is performed on the original single-frame surface texture image using the single-frame edge contour binary image as a spatial topological constraint. A homomorphic filtering algorithm based on frequency domain signal processing is introduced to generate a single-frame surface texture grayscale image and construct a set of surface texture grayscale images.

[0014] Furthermore, the step of switching to the lateral structured light illumination module to acquire a single-frame original image of surface texture during the second light field slice data acquisition includes:

[0015] Within a preset time interval after the first light field slice data acquisition is completed, the lateral structured light illumination module is triggered to control the incident light to irradiate the activated carbon particle flow at a preset incident grazing angle; the incident grazing angle is the geometric angle between the main optical axis of the lateral structured light illumination module and the transmission plane of the activated carbon particle flow, and its value is limited to a closed range of 10° to 15°.

[0016] Based on the inverse application principle of Lambert's cosine law, the light signal reflected by the micro-pore structure on the surface of activated carbon particles is received, and the surface depth undulation is converted into brightness gradient changes to generate a single-frame original surface texture image; the micro-pore structure is mapped as gray value pixel clusters in the single-frame original surface texture image.

[0017] Extract the lateral texture reflection brightness response value at each pixel coordinate of the original image of the surface texture of a single frame, and obtain the backlight residual brightness response value at the first light field slice data acquisition.

[0018] A texture confidence enhancement unit is introduced, which performs sinusoidal weighted modulation on the lateral texture reflection brightness response value based on the sine value of the incident grazing angle, and divides the sinusoidal weighted modulation result by the sum of the backlight residual brightness response value and the preset numerical stability constant to obtain a ratio result; the ratio result is multiplied by the preset photoelectric conversion gain constant to generate a feature enhancement contrast coefficient; the feature enhancement contrast coefficient is the moisture absorption porosity feature weight of the pixel in the original image of the single frame surface texture.

[0019] Furthermore, the method for generating the physical wear risk index includes:

[0020] Particle connected components are extracted from single-frame edge contour binary maps in the edge contour binary map set to mark the particle objects to be analyzed, and the edge chain code of the particle objects to be analyzed is extracted. The edge chain code is mapped to a sub-pixel level parameterized curve using a B-spline curve fitting algorithm. The sub-pixel level parameterized curve is solved based on the central difference method, and the local discrete curvature value is calculated to construct the edge curvature sequence.

[0021] The local discrete curvature values ​​in the edge curvature sequence are traversed, and effective cutting points are extracted by filtering using a preset coating damage critical threshold. A single-point mechanical damage force unit is introduced to physically weight the effective cutting points to calculate the cumulative attack potential energy, and this is combined with the total number of effective cutting points and the previous... Construct an effective cutting feature vector using the mean of maximum curvature;

[0022] A comprehensive risk score is generated by weighting and aggregating the effective cutting feature vectors using a preset risk weight vector, and then the comprehensive risk score is mapped to a physical wear risk index using a Sigmoid nonlinear activation function.

[0023] Furthermore, the calculation steps for the accumulated attack potential energy include:

[0024] Perform a local maximum search on the edge curvature sequence and extract local maximum points whose local discrete curvature values ​​are greater than the preset coating failure critical curvature threshold as effective cutting points;

[0025] For each selected effective cutting point, the difference between its local discrete curvature value and the critical curvature threshold for coating failure is calculated to generate the over-limit sharpness.

[0026] Using a preset contact mechanical index as the exponent, a power operation is performed on the excessive sharpness to obtain the theoretical destructive potential energy response value; the contact mechanical index is a preset constant greater than 1.

[0027] The theoretical failure potential energy response value is multiplied and corrected using a preset brittle fracture correction factor to generate a single-point failure potential energy. The single-point failure potential energy of all selected effective cutting points is then summed to output the cumulative attack potential energy.

[0028] Furthermore, the method for generating the environmental moisture absorption preemption index includes:

[0029] Traverse the single-frame surface texture grayscale images in the set of surface texture grayscale images, use the single-frame edge contour binary image as a spatial index mask, clip the texture region of interest in the single-frame surface texture grayscale image and map it into a three-dimensional grayscale topological surface, perform differential box-counting dimension operation based on multi-scale mesh division on the three-dimensional grayscale topological surface, and calculate the fractal dimension eigenvalue.

[0030] A spatial sliding window is used to traverse single-frame surface texture grayscale images in the surface texture grayscale image set. A two-dimensional discrete Fourier transform is performed to generate a local complex spectrum matrix. Based on the local complex spectrum matrix, the local frequency domain energy value is calculated. According to the preset sliding step size of the spatial sliding window, all local frequency domain energy values ​​on the single-frame surface texture grayscale image are traversed to construct a local frequency domain energy distribution field. A moisture absorption hotspot determination threshold is introduced to lock the moisture absorption hotspot region through binarization segmentation. The moisture absorption texture ratio of the moisture absorption hotspot region is calculated.

[0031] Based on the fractal dimension eigenvalues ​​and the proportion of moisture-absorbing texture, the physical potential energy value is calculated using an exponential multiplicative coupling term, and the physical potential energy value is converted into an environmental moisture-absorbing preemption index through a probability mapping activation layer.

[0032] Furthermore, the calculation steps for the local frequency domain energy value include:

[0033] Traverse each frequency domain coordinate in the local complex spectrum matrix, extract the real and imaginary part values ​​of the complex elements at each frequency domain coordinate, calculate the arithmetic square root of the sum of the squares of the real and imaginary part values, and generate the spectral amplitude of each frequency domain coordinate.

[0034] The preset texture cutoff frequency is retrieved as the lower limit parameter for frequency domain integration, and the Nyquist frequency of the imaging system is retrieved as the upper limit parameter for frequency domain integration to construct the frequency band of interest.

[0035] All spectral amplitudes falling within the frequency band of interest are selected, and the local frequency domain energy value is calculated by performing square integration on all selected spectral amplitudes.

[0036] Furthermore, the method for generating the particle classification control command includes:

[0037] Based on the physical wear risk index and the environmental moisture absorption preemption index, coating physical strength data and average relative humidity data are introduced as external working condition constraint parameters. Data fusion is performed using an environment-medium dual constraint weighting algorithm to generate a color-changing bead compatibility score.

[0038] The compatibility scores of the color-changing beads are compared numerically based on preset preferred and eliminated thresholds. Combined with the physicochemical test data of the activated carbon raw material batch, a multi-path diversion decision is executed to generate particle grading control instructions. If the compatibility score of the color-changing beads is not lower than the preferred threshold, it is mapped to the first diversion instruction state value; if the compatibility score of the color-changing beads is lower than the eliminated threshold and the physicochemical test data meets the standard, it is mapped to the second diversion instruction state value; otherwise, it is mapped to the third diversion instruction state value.

[0039] The physicochemical testing data of the activated carbon raw material batches are read through an industrial IoT interface or a laboratory information management system data port, which reads batch attribute data associated with the activated carbon raw material batches, including the batch average iodine value and standard adsorption value.

[0040] Furthermore, the logical steps of the environment-medium dual-constraint weighted algorithm include:

[0041] The physical strength data and average relative humidity data of the coating are read as external working condition constraint parameters. Based on the preset negative correlation mapping, the physical strength data and average relative humidity data of the coating are mapped to the coating mechanical strength weight coefficient and the environmental humidity sensitivity weight coefficient, respectively.

[0042] The physical wear risk index is weighted by the coating mechanical strength weighting coefficient, and the environmental moisture absorption preemption index is weighted by the environmental humidity sensitivity weighting coefficient. The weighted multiplication results of the two are summed to generate a comprehensive risk weighted sum.

[0043] The sum of the coating mechanical strength weighting coefficient and the environmental humidity sensitivity weighting coefficient is used as the normalized denominator. The weighted sum of the comprehensive risks is divided by the normalized denominator to obtain the linear risk mean.

[0044] Perform a power operation on the linear risk mean with a preset risk sensitivity index as the exponent, and use the value 1 to subtract the result of the power operation to generate a color-changing bead compatibility score.

[0045] The activated carbon raw material screening system based on multimodal image feature fusion is used to implement the activated carbon raw material screening method based on multimodal image feature fusion mentioned above. It includes an image acquisition module, a physical wear assessment module, a moisture absorption risk assessment module, and a compatible diversion decision module.

[0046] The image acquisition module is used to acquire single-frame edge contour original images and single-frame surface texture original images of activated carbon particle flow. It performs a topology-preserving segmentation algorithm based on variational level sets on the single-frame edge contour original images to construct a set of edge contour binary images, and introduces a homomorphic filtering algorithm on the single-frame surface texture original images to construct a set of surface texture grayscale images.

[0047] The physical wear assessment module is used to traverse the set of binary edge contour maps to extract edge chain codes, perform sub-pixel-level parameterized reconstruction of the edge chain codes to generate an edge curvature sequence, extract effective cutting feature vectors from the edge curvature sequence based on a preset coating damage critical threshold, and generate a physical wear risk index through probability mapping.

[0048] The moisture absorption risk assessment module is used to: traverse the set of surface texture grayscale images, map the regions of interest of the textures into three-dimensional grayscale topological surfaces, calculate the fractal dimension eigenvalues ​​using the differential box-counting method, perform a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, construct a local frequency domain energy distribution field based on a preset sliding window step size, calculate the proportion of moisture-absorbing textures through binarization segmentation, and perform nonlinear probability mapping on the fractal dimension eigenvalues ​​and the proportion of moisture-absorbing textures using exponential multiplicative coupling logic to generate an environmental moisture absorption preemption index.

[0049] The compatibility diversion decision module is used to obtain the physical wear risk index and the environmental moisture absorption preemption index, introduce an environment-medium dual constraint weighted algorithm to calculate the compatibility score of the color-changing beads, and combine the preset preferred threshold and elimination threshold with the physicochemical test data of the activated carbon raw material batch to perform multi-path diversion decision and generate particle classification control instructions pointing to different diversion states.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention constructs a nanosecond-level synchronized time-division multiplexed optical field imaging environment and combines variational level set topology preservation and frequency domain homomorphic filtering algorithms to achieve orthogonal decoupling of the geometric contour information and surface texture topology information of activated carbon particles. This solves the problem of bioincompatibility between components caused by the inability of traditional physicochemical detection indicators to identify microscopic knife-edge-shaped edges and deep hygroscopic pores.

[0052] This invention transforms the microscopic geometric sharpness into the cumulative attack potential energy of the coating by performing sub-pixel-level parameterized reconstruction and contact mechanical equivalent weighting of the discrete edges of activated carbon particles. This avoids the physical wear failure of the color-changing bead coating caused by ignoring the microscopic knife-edge-like physical characteristics in traditional physicochemical screening.

[0053] This invention quantifies the instantaneous moisture plundering potential of activated carbon particles on the microscopic contact surface by nonlinearly coupling the fractal dimension feature, which characterizes the overall pore complexity, and the frequency domain energy distribution, which characterizes the local moisture absorption hotspots. This avoids the pseudo-discoloration failure of catalyst beads caused by excessive moisture absorption due to the microporous structure.

[0054] This invention introduces a dual constraint mechanism of environment and medium to dynamically weight the microscopic morphological characteristics of activated carbon particles, and combines macroscopic physicochemical indicators to perform multi-level diversion decisions that decouple physical properties. This enables particles with physical aggression but excellent chemical adsorption performance to be directed to single-component product lines, solving the problem of resource waste caused by high-quality raw materials being misjudged as waste due to microscopic morphological defects in the traditional binary screening mode. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the principle of the activated carbon raw material screening method based on multimodal image feature fusion of the present invention.

[0057] Figure 2 This is a schematic diagram of the optical path for the first optical field slice data acquisition of the present invention;

[0058] Figure 3 This is a schematic diagram of the optical path for the second optical field slicing data acquisition in this invention;

[0059] Figure 4 This is a functional block diagram of the activated carbon raw material screening system based on multimodal image feature fusion of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1:

[0062] Please see Figure 1 As shown, this embodiment provides a method for screening activated carbon raw materials based on multimodal image feature fusion, including:

[0063] Step S1000: Collect the activated carbon particle flow to be screened. single-frame edge contour original image and the original image of surface texture in a single frame For the original image of the edge contour of a single frame Perform a topology-preserving segmentation algorithm based on variational level sets to construct a set of binary edge contour maps. For a single frame of the original surface texture image Homomorphic filtering algorithm is introduced to construct a set of surface texture grayscale images. .

[0064] Specifically, this step aims to transform the dynamically transporting flow of activated carbon particles to be screened in the physical world. Through photoelectric conversion and algorithm decoupling, the data is transformed into digital samples that can be processed by computer vision, and a set of binary edge contour maps that can characterize the physical sharpness of particles is extracted. To investigate potential physical wear and tear, and simultaneously extract a set of surface texture gradient maps that characterize the hygroscopic properties of particle specific surface area. To identify potential false color changes, we can achieve an enhanced mapping from physical entities to data features, providing a data foundation for subsequent steps.

[0065] Further, step S1000 includes:

[0066] Step S1100: Filter the activated carbon particles to be screened at the same spatial location. Two light field slicing data acquisitions were performed. During the first light field slicing data acquisition, a backlight illumination module was used to acquire a single frame of raw edge contour image. During the second light field slice data acquisition, the system switched to the lateral structured light illumination module to acquire a single frame of raw surface texture image. .

[0067] Specifically, this step aims to transform the discretely distributed, dynamically flowing activated carbon particles to be screened on a conveyor belt in the physical world... As the data acquisition object, the differences in the interaction between light field and matter are utilized to map the microscopic knife-edge-like physical features of the potential, easily scratched color-changing bead layer of particles within the current field of view into a single-frame original image of the edge contour. The binarized gradient boundary signal in the image is used to map the potential, invisible, and pseudo-discoloration-inducing hygroscopic pore features into a single-frame original surface texture image. The surface texture gradient signal in the data source achieves orthogonal decoupling of geometric morphology information and surface topology information, providing real-time frame-level data input for subsequent steps.

[0068] In the specific implementation process, this step constructs a time-division multiplexing imaging environment based on nanosecond-level precise synchronization. Through the logical coordination of the image acquisition terminal and the dual-channel illumination system, the flow of activated carbon particles to be screened at the same spatial location is observed. Perform two physically independent light field slice data acquisitions.

[0069] The first light field slice data acquisition, i.e., transmitted light field truncation acquisition, involves controlling the backlight module, i.e., the telecentric backlight, to emit a parallel beam of light through the particle stream. Utilizing the opaque physical property of the activated carbon particles, the particles themselves form a light energy blockage in the optical path, thus projecting a dark area with near-zero light intensity onto the photosensitive array of the image sensor. The data acquired at this time is defined as the single-frame edge contour raw image. The physical significance of this process lies in using backlighting to eliminate interference from surface textures, forcing the micron-level jagged edges, sharp angles, and irregular protrusions of the particle edges—that is, potential sources of physical wear—to appear as a clear binary bimodal distribution in the image histogram, thereby obtaining geometric contour data that maximizes the gradient.

[0070] Further, please refer to Figure 2 As shown, Figure 2 This is a schematic diagram of the optical path for the first optical field slice data acquisition of the present invention.

[0071] The second light field slicing data acquisition, namely the reflected light field gradient acquisition, occurs within a tiny time interval after the first light field slicing data acquisition. This tiny time interval is set to be less than 50 microseconds to suppress pixel displacement caused by conveyor belt movement. The system immediately cuts off the backlight illumination module and triggers the side structured light illumination module. The light is controlled to be incident at a preset grazing angle. Irradiation of a stream of activated carbon particles on a conveyor belt, wherein the incident grazing angle It is the main optical axis of the lateral structured light illumination module and the flow of activated carbon particles to be screened. The geometric angle formed between transmission planes is physically constrained. to Within the closed interval. Based on the inverse application of Lambert's Cosine Law, when light rays are at the stated incident grazing angle... When the light is scanned across the particle surface, the microscopic deep pores, cracks, and irregular burrs on the particle surface, which have strong instantaneous hygroscopicity, block the light, thus forming a projection area on the backlit side that is proportional to its physical depth. This transforms the slight surface depth undulations into brightness gradient changes on the image plane. The data acquired at this time is defined as a single-frame raw image of the surface texture. The original image of the surface texture in this single frame. In this process, the microscopic pore structure that is not visible under normal diffuse reflection light is transformed into a cluster of gray-scale pixels with high frequency distribution, i.e., hygroscopic pore features. This light and shadow conversion mechanism intuitively maps the chemical potential of activated carbon particles to absorb water bound by catalyst beads, such as CATALYST color-changing beads, into surface texture gradient signals in digital images.

[0072] Further, please refer to Figure 3 As shown, Figure 3 This is a schematic diagram of the optical path for the second optical field slice data acquisition in this invention.

[0073] To quantitatively evaluate the original image of a single frame surface texture To enhance the characterization ability of each pixel for hygroscopic pore features and suppress the interference of dark current in the image sensor and ambient stray light on the micro-texture signal, this step further introduces a texture confidence enhancement unit. This unit establishes a digital mapping relationship between optical projection geometry and the hygroscopic pore features of activated carbon micropores, calculating pixel coordinates through a nonlinear modulation algorithm based on prior knowledge of the physical light field. Microscopic features enhance contrast coefficient This transforms the image from a simple grayscale distribution matrix into a weighted matrix representing the hygroscopic porosity characteristics that easily induce pseudo-discoloration. Among these, This represents a pair of pixel coordinate indices on the photosensitive target surface of the image sensor. The range of values ​​is determined by the resolution of the industrial camera and is used to determine the microscopic physical feature points on the surface of activated carbon particles. The x-coordinate of the pixel coordinate. The vertical coordinate represents the pixel coordinate.

[0074] The feature enhances the contrast coefficient The specific calculation logic is as follows: Extract the surface texture brightness data of the current frame from the second light field slice data acquisition, that is, the lateral texture reflection brightness response value. and the lateral texture reflection brightness response value With incident glancing angle The sine value, i.e., the sine geometric weighted term. The process involves performing multiplication, utilizing the monotonically increasing characteristic of the sine function at small angles to nonlinearly modulate the original reflected light field. This amplifies the signal amplitude of deep-pore structures sensitive to the incident grazing angle, simulating the capillary trapping effect of micropores on the activated carbon surface on the moisture in the catalyst beads. Subsequently, the amplification value of the nonlinearly modulated signal is divided by the backlight residual brightness response value from the first light field slice data acquisition. The division process, summing the numerical stability constants, utilizes the object occlusion characteristics during backlit imaging to construct a pixel-level signal-to-noise ratio enhancement mechanism. Specifically, it minimizes the denominator in the particle region to exponentially increase the result value, while maximizing the denominator in the background region to suppress the result value. Finally, the result of the division operation is multiplied by the photoelectric conversion gain constant for dimensional normalization, yielding the feature enhancement contrast coefficient. The lateral texture reflection brightness response value is mentioned above. It is the image sensor at pixel coordinates The received raw grayscale data containing the surface texture and topological information of activated carbon is used to characterize the incident grazing angle. Under light irradiation, the scattering and absorption characteristics of photons in the micro-regions of the particle surface are directly related to the physical roughness of the particle surface and the geometric cross-sectional area of ​​the pore openings. Represents the digital sine function operation; the backlight residual brightness response value It is the image sensor at pixel coordinates The received signal, which includes the superimposed signal of system ambient light noise and backlight transmission residue, approaches zero in the ideal shading area; the numerical stability constant is a minimal correction term to prevent the denominator from becoming zero and to suppress sensor dark current noise, used to ensure the backlight residual brightness response value. In the ideal occlusion state with an absolute zero value, the denominator is not zero; the photoelectric conversion gain constant is a system calibration factor that correlates the quantum efficiency (QE) of the image sensor and the analog amplification gain, and its value is a positive real number preset according to the imaging hardware specifications.

[0075] For example, suppose a batch of activated carbon raw materials to be screened contains incompatible particles with high iodine values ​​but defective microstructures. When these incompatible particles flow through this step, during the first light field slicing data acquisition, a single-frame original image of the edge contour is generated. The incompatible particle's edge is clearly delineated by three sharp corners with a radius of curvature of less than 5 micrometers, exhibiting microscopic knife-edge-like physical characteristics, directly indicating a risk of physical wear. During the second light field slice data acquisition, although the particle appeared entirely black, its surface texture was clearly visible in the generated single-frame original image. In the image, a dense cluster of grayscale pixels appears in the central region of the surface. Using a nonlinear modulation algorithm based on prior knowledge of the physical light field, the feature enhancement contrast coefficient of the pixels in this region is calculated. The abnormally high value is attributed to the lateral texture reflection brightness response value. at the angle of incidence The shadow effect is strong, and the backlight residual brightness response value is high. The extremely low denominator due to occlusion indicates the presence of a deep capillary structure in this region, which readily absorbs moisture from the color-changing beads upon mixing. Based on this, this step simultaneously identifies the microscopic, knife-edge-like physical characteristics and hygroscopic porosity of the particle within milliseconds through a single pass-through inspection, thereby issuing a rejection instruction to prevent potential batch-level false color changes during subsequent finished product storage.

[0076] Step S1200, based on the original image of the single-frame edge contour A topology-preserving segmentation algorithm based on variational level sets is constructed to generate single-frame edge contour binary maps. And construct a set of binary edge contour maps. .

[0077] Specifically, this step aims to process the single-frame edge contour raw image acquired in step S1100. As input to the algorithm, the data is transformed into feature analysis data with physical feature fidelity through a topology-preserving segmentation algorithm based on variational level sets, from the original edge contour image of a single frame. A single-frame binary image of edge contours, preserving microscopic knife-edge-like physical features, is segmented from the middle. And the single-frame edge contour binary image Import edge contour binary image set .

[0078] In the specific implementation process, for the original image of the edge contour of a single frame Considering that the Gaussian smoothing operation introduced by conventional edge detection operators during discretization inevitably erodes the micron-level serrated features of activated carbon particle edges, potentially causing physical scratches on catalyst beads, such as CATALYST color-changing beads, and the loss of the microscopic knife-edge-like physical features of the coating in digital mapping, this step constructs a topology-preserving segmentation algorithm based on variational level sets. The core of this algorithm is to transform the image segmentation problem into a problem of minimizing the edge total energy functional of the level set function. By introducing a geometric topological constraint term specifically for geometric singularities, i.e., sharp corners, in addition to traditional length and area constraints, the algorithm enables the dynamic segmentation contour to overcome the geometric contraction effect caused by minimizing the contour perimeter when approaching the particle's physical edge, forcing the segmentation boundary to reside on geometric singularities with sharp protrusions. Here, the level set function is the original image of the edge contour in a single frame. domain The real-valued function of the spatial location information of all pixels represents the iterative shrinkage process of the dynamically segmented contour. The set of points where the value of this function is zero constitutes the zero-level set contour for tracking the contour of activated carbon particles.

[0079] The total edge energy functional is a dynamic competition mechanism constrained by the image gray-level gradient distribution and contour geometric features. The gradient descent iterative algorithm is used to iteratively solve this total edge energy functional until it converges to a minimum value. The specific iterative solution calculation logic is as follows: For a single frame of the original edge contour image... Domain Each pixel within the array undergoes integration. First, the Dirac distribution function is used to focus the computation on the current dynamically segmented contour, i.e., the zero-level set contour. Then, the computation is decomposed into two parallel weighted branches: the first branch calculates the magnitude of the image's grayscale gradient. The edge stopping function is used, and weighted by a preset length term weighting coefficient. The purpose of this branch is to drive the dynamically segmented contour to converge to the region with a large image gray-level gradient in order to fit the edge, and to utilize the image gray-level gradient magnitude. The minimization property of the first branch makes the segmentation contour tend to be smooth; the second branch introduces physical-geometric cross-modal constraints, that is, calculating the absolute value of the local curvature of the current pixel on the dynamic segmentation contour. The difference between the value and the preset microscopic acute angle curvature threshold is used for binarization logic judgment through the Heaviside step function. The value is determined only when the absolute value of the local curvature is... When the microscopic acute angle curvature threshold is exceeded, i.e., when a potential geometric singularity capable of piercing the color-changing bead coating is detected (i.e., a microscopic blade-like physical feature), the Heaviside step function outputs a logical truth value, thereby activating the preset sharpness preservation weight coefficients in the calculation. Through the above calculation, the finally converged zero-level set contour, while eliminating background noise, completely preserves the geometric singularities of the activated carbon particles, achieving a mapping from physical damage risk to digital topological features.

[0080] Wherein, the domain It is a single-frame edge contour original image The two-dimensional spatial region in which it is located covers the entire set of pixel coordinates on the photosensitive target surface of the image sensor, and is used to define the computational boundary of the functional integration operation of the total edge energy; the Dirac distribution function is a generalized function with compact support properties, whose value tends to infinity when the independent variable is zero and is zero at other locations, and is used to strictly mathematically constrain the computational range of the cost value within the local pixel neighborhood of the zero-level set contour, ensuring that the computational power is concentrated only in the particle physical edge region; the image grayscale gradient magnitude It is the original image reflecting the edge contours of a single frame. The physical quantity representing the spatial rate of change of pixel grayscale values ​​is calculated by convolution using the Sobel or Prewitt operators. It is used to quantify the contrast difference in optical transmittance between activated carbon particles and the background area of ​​the conveyor belt. The edge stopping function is a boundary indicator constructed using image grayscale discontinuities. Its value ranges from 0 to 1 in a closed interval and is related to the magnitude of the image grayscale gradient. The length term weight coefficient is a constant that adjusts the smoothness of the edges. Its value is a positive real number preset based on the image signal-to-noise ratio, used to balance the competition between the constraint of minimizing the perimeter of the segmented contour (i.e., the global smoothness trend) and the edge fit. The differential geometric quantity characterizing the geometrical sharpness of particle edges is obtained by calculating the second derivative of the level set function and is used to identify protruding structures on the surface of activated carbon particles. The value of the micro-acute angle curvature threshold is physically calibrated based on the material hardness characteristics of the catalyst beads and serves as a quantitative benchmark for determining whether activated carbon particles have a risk of physical wear failure. The Heaviside step function is a binary logic switch function with an output value of 0 or 1, used to implement logic gating judgment for micro-knife-shaped physical features. The sharpness retention weight coefficient is a control parameter preset based on the physical tolerance of the catalyst bead coating. Its value is set to a positive real number that is significantly greater than the length term weight coefficient, used to force the algorithm to preferentially retain acute angle features that can cause physical damage during the energy minimization process.

[0081] When the above iterative calculation terminates, i.e., when the total edge energy functional reaches its minimum value, the system locks the current level set function, extracts the connected region enclosed by its zero level set contour, sets the pixel values ​​within the connected region to 1, and sets the external pixel values ​​to 0, thereby generating a single-frame edge contour binary map. As the activated carbon particles to be screened flow along the conveyor belt... The image frames from each acquisition cycle are continuously transmitted, and the above segmentation operation is performed on each frame. This generates multiple frames of edge contour binary images in a time sequence. The data is sequentially imported into the data storage queue, thereby constructing a set of structured binary edge contour maps. This provides a data foundation that preserves physical sharpness information for subsequent batch feature analysis.

[0082] Step S1300: Using a single-frame edge contour binary image As a spatial topological constraint on the original image of surface texture in a single frame A pixel-by-pixel logical AND operation is performed, and a homomorphic filtering algorithm based on frequency domain signal processing is introduced to generate a single-frame surface texture grayscale image. And construct a set of surface texture grayscale images. .

[0083] Specifically, this step aims to process the single-frame raw surface texture image acquired in step S1100. As input to the algorithm, the single-frame edge contour binary image segmented in step S1200 is used. As a spatial topological constraint, the original image of surface texture in a single frame Perform region of interest extraction, remove background noise, and analyze the incident light component. and surface reflection component Frequency domain signal processing generates a single-frame surface texture grayscale image that highlights the hygroscopic pore characteristics. and the single-frame surface texture grayscale image Import surface texture grayscale image set .

[0084] Wherein, the incident light component The luminous flux distribution formed by lateral structured light on the three-dimensional curved surface of activated carbon particles due to the gradually changing incident angle is characterized in the frequency domain as a low-frequency spatially varying signal with concentrated energy. The presence of this component compresses the dynamic range of pixels in the particle edge region, resulting in a visual masking phenomenon where pixel grayscale value attenuation obscures the hygroscopic pore features; the surface reflection component... It is a signal of sudden change in local reflectivity caused by the microporous geometry of the activated carbon particle surface with strong instantaneous hygroscopicity. In the frequency domain, it is a high-frequency spatial change signal with energy dispersion. It is the core feature data characterizing whether the particles will induce hygroscopic pore characteristics.

[0085] In the specific implementation, a single-frame edge contour binary image that retains the microscopic blade-like physical features is used. As a binary spatial mask, i.e., a spatial topological constraint, it is used in conjunction with the original image of the surface texture of a single frame. Perform a pixel-by-pixel logical AND operation. This operation utilizes a single-frame edge contour binary image. The logical truth value inside the zero-level set contour is selected from the original image of the surface texture in a single frame. The corresponding pixel data of the activated carbon particle surface is used, and the background area is forcibly masked by using logical false values ​​outside the contour, so as to obtain a masked image containing only particle body information.

[0086] Subsequently, to address the non-uniform illumination phenomenon in the masked image, where the grayscale response of pixels in the central region of the activated carbon particles is high while that in the edge region is weak, due to the three-dimensional curved surface distribution of the activated carbon particle surface, this step introduces a homomorphic filtering algorithm. This homomorphic filtering algorithm is constructed based on the multiplicative coupling characteristic of the optical imaging process, i.e., the activated carbon surface in pixel coordinates... Optical response at the location Physically, it is not a single signal, but rather a combination of incident light components. and surface reflection component It is formed by coupling through multiplication.

[0087] In order to decouple the aforementioned multiplicative relationship and directionally enhance the surface reflection component that characterizes the hygroscopic porosity... This step performs the following frequency domain signal processing: First, a natural logarithmic transformation is performed on the masked image to transform the multiplicative coupling into an additive coupling. Second, the image data after additive coupling is converted from pixel coordinates in the spatial domain using a Fast Fourier Transform (FFT). Transform to frequency domain coordinates This generates a complex spectrum matrix, i.e., spectrum data. In this spectrum data, the frequency domain coordinates... Instead of geometric positions, these represent sinusoidal frequency indices in the horizontal and vertical directions, respectively, used to characterize the amplitude and phase information of the corresponding frequency components in the image. The third step involves introducing the response value of a pre-defined homomorphic filter transfer function to perform point-to-point multiplicative weighted modulation on the spectral data. In this multiplicative weighted modulation, the response value of the homomorphic filter transfer function is used to attenuate the spectral coefficients in the low-frequency index region to suppress the uneven lighting background caused by the three-dimensional curvature of the particle surface; simultaneously, the spectral coefficients in the high-frequency index region are amplified to enhance the reflectivity abrupt signal caused by micropores. The fourth step involves performing an inverse Fast Fourier Transform (IFFT) and exponential transform on the multiplicative weighted modulated spectral data to restore it to the spatial domain, generating a single-frame surface texture grayscale image that enhances the details of the micropore texture. In this image, the deep pore texture that was originally hidden in the edge shadows is reconstructed into high-contrast pixel clusters, thus visually reflecting the true instantaneous moisture absorption potential of the particle surface.

[0088] The specific calculation logic for the response value of the homomorphic filter transfer function is as follows: using frequency domain coordinates... The square of the ratio of the Euclidean distance to the cutoff frequency constant is used, and a preset sharpening constant is introduced to scale this squared ratio to construct a dimensionless frequency metric. Subsequently, an inverse Gaussian attenuation operation is performed, using the aforementioned frequency metric as the negative exponential variable of the natural exponential function to obtain a Gaussian attenuation factor. This Gaussian attenuation factor is then subtracted from the value 1 to generate a filter operator that approaches 0 in the low-frequency region and 1 in the high-frequency region. Finally, linear stretching and bias modulation are performed on this filter operator. The difference between the high-frequency gain coefficient and the low-frequency gain coefficient is used as a dynamic range scaling factor to stretch the amplitude of the filter operator, and the low-frequency gain coefficient is superimposed as the base bias. Through these operations, the homomorphic filter transfer function response value forces a gain value less than 1 in the low-frequency range to compress the dynamic range of illumination, while forcing a gain value greater than 1 in the high-frequency range to amplify the details of the pore texture. This achieves a digital characterization of the hygroscopic pore characteristics of the activated carbon surface while eliminating non-uniform illumination background.

[0089] Wherein, the Euclidean distance is a frequency domain coordinate. To the center of the spectrum, i.e., the horizontal sinusoidal frequency index. and the frequency index of the sinusoidal wave in the vertical direction The Euclidean distance at which all points are zero is a quantitative indicator measuring the rate of spatial change of the current frequency component; the cutoff frequency constant is used to strictly distinguish the components representing incident light illumination. The low-frequency region and the surface reflection component The frequency domain boundary threshold of the high-frequency region is determined by physical calibration based on the spatial frequency corresponding to the average pore size of the micropores on the surface of activated carbon particles that have strong instantaneous hygroscopicity; the sharpening constant is used to adjust the frequency range of the incident light component. to surface reflection component The frequency domain transition smoothness; the high-frequency gain coefficient is for the surface reflection component that can characterize the hygroscopic pore features. The set amplification factor, a positive real number greater than 1.0, is used to directionally amplify the amplitude of high-frequency signals far from the center of the spectrum, in the final single-frame surface texture grayscale image. The enhanced texture contrast of the deep-pore structure of the catalyst beads may instantly dry out the moisture; the low-frequency gain coefficient is for the incident light component caused by the three-dimensional curved surface of the activated carbon particles. The compression ratio is set to a positive real number less than 1.0 to attenuate the low-frequency signal energy near the center of the spectrum, thereby compressing the dynamic range of illumination and eliminating texture occlusion caused by the brightness attenuation at the edge of the particles.

[0090] As the activated carbon particles to be screened flow on the conveyor belt The continuous transmission of images, performing the above operations on each image frame of each acquisition cycle, generates multiple frames of surface texture grayscale images sequentially. The data is sequentially imported into the data storage queue, thereby constructing a structured collection of surface texture grayscale images. .

[0091] Step S2000: Traverse the set of binary edge contour images. The edge chain code is extracted and then reconstructed using sub-pixel-level parameterization to generate an edge curvature sequence. Based on a preset coating damage critical threshold, from the edge curvature sequence Extracting effective cutting feature vectors Physical wear risk index is generated through probability mapping. .

[0092] Specifically, this step aims to set up the binary edge contour map output in step S1200. As data input, this set of binary images of edge contours... Each single-frame edge contour binary image Perform connected component extraction and differential geometry analysis to quantify the sharpness of the activated carbon particle surface and output a physical wear risk index for each individual particle. Based on this physical wear risk index Identify and label aggressive particle indices.

[0093] Further, step S2000 includes:

[0094] Step S2100: Process the set of binary edge contour maps. Binary image of edge contour in a single frame Particle connected components are extracted to label the particle objects to be analyzed, and the edge chain codes of the particle objects are extracted. The edge chain codes are mapped to subpixel-level parameterized curves using a B-spline curve fitting algorithm. The subpixel-level parameterized curves are solved based on the central difference method to calculate the local discrete curvature values. To construct edge curvature sequences .

[0095] Specifically, this step aims to extract the set of binary edge contour maps output from step S1200, which retains the geometric singularities of the activated carbon particles. As a data source, the set of binary edge contour maps is extracted using a granular connected component extraction algorithm. Binary image of single-frame edge contour The data is deconstructed into independent particle objects to be analyzed. Then, a B-spline curve fitting algorithm is used to reconstruct the discrete pixel-level edges into continuous sub-pixel-level parametric curves. Finally, an edge curvature sequence that can quantify the microscopic knife-edge-like physical characteristics is calculated. This provides differential geometric data support for determining whether particles have the physical aggression to scratch the color-changing bead coating.

[0096] In the specific implementation process, this step includes cascaded processing logic from image frames to grain features, as follows:

[0097] The first step involves separating the particle objects to be analyzed based on the particle connected component extraction. This processing step starts from the set of binary edge contour maps. Retrieve the current single-frame edge contour binary image Due to the binary image of the edge contour in this single frame. The image contains multiple discretely distributed activated carbon particles that exist simultaneously on a conveyor belt. A particle connectivity extraction algorithm, such as 8-connected component labeling (8-CCL), is executed to identify all independent closed white regions in the single frame image. Each closed white region is labeled as an independent particle object to be analyzed, thus achieving single-object separation of multi-target images.

[0098] The second step involves edge chain code extraction and sub-pixel-level parameterized reconstruction. For each particle object to be analyzed, a boundary tracking algorithm, such as Moore's neighborhood tracking, is executed. This involves traversing the edge pixels of the particle object in a counter-clockwise direction to extract the closed edge chain code composed of discrete coordinate pairs. These discrete coordinate pairs refer to the row and column address indices of the discrete photodiode array on the photosensitive target surface of the image sensor, and their values ​​are physically restricted to integers by the data format.

[0099] However, because the physical edge of activated carbon particles is a continuous optical geometric trajectory at the microscale, and image sensors perform spatial sampling through a fixed-interval photosensitive array, this continuous optical geometric trajectory is forcibly aligned to the nearest integer grid point. This discretized spatial sampling process introduces spatial sampling aliasing artifacts into the digital image, causing what were originally smooth or sharp physical edges to be distorted into stepped discrete paths composed of alternating horizontal and vertical line segments. This makes it impossible to distinguish between the digital right-angle features caused by the pixel grid and the microscopic knife-edge-like physical sharp corner features of the activated carbon surface that can physically scratch the color-changing bead coating, leading to misjudgments of physical wear risk.

[0100] To eliminate the aforementioned spatial sampling aliasing artifacts and restore the true physical edge morphology, this processing step introduces a cubic B-spline curve fitting algorithm. This algorithm maps the discrete closed edge chain code coordinates to fitting control vertices in the normalized parameter domain, constructing a sub-pixel-level parameterized curve with second-order differentiability. The normalized parameter domain refers to the range of values ​​for the independent variable of the B-spline curve, used to achieve the mathematical transformation from a discrete pixel index domain to a continuous digital function. The fitting control vertices are a set of weighted geometric guide points calculated based on the closed edge chain code, used to determine the geometric shape and local curvature of the sub-pixel-level parameterized curve. Second-order differentiability means that the first derivative (tangent vector) and the second derivative (curvature vector) of the sub-pixel-level parameterized curve at any parameter point remain numerically continuous without abrupt changes.

[0101] The third step is the calculation of the discrete curvature field based on the central difference method. Based on the aforementioned sub-pixel-level parameterized curve, this step performs equally spaced parameter sampling to generate a series of sub-pixel-level geometric sampling points located on the curve trajectory. The local discrete curvature value at each sub-pixel-level geometric sampling point is then calculated using the central difference method. The subpixel-level geometric sampling points are actual physical coordinate points analytically calculated along the path of the subpixel-level parameterized curve based on the sampling step size. Their set constitutes the continuous geometric boundary of the activated carbon particles after digital reconstruction. The sampling index represents the subpixel geometric sampling point, which is used to identify the arrangement order of the subpixel geometric sampling points on the continuous subpixel level parameterized curves. In the algorithm operation, each microscopic physical feature point is uniquely digitally addressed. Indicates the first Local discrete curvature values ​​at sub-pixel geometric sampling points.

[0102] The local discrete curvature The calculation logic is as follows: Define a local sliding window covering the preceding and following neighborhoods. Perform a weighted central difference operation using the coordinate data of the sub-pixel geometric sampling points within this local sliding window to calculate the first derivative vector (tangent vector) and the second derivative vector (acceleration vector) of the current sub-pixel geometric sampling point. Then, use the cross product magnitude of the first and second derivative vectors as the numerator and the cube of the first derivative vector's magnitude as the denominator, performing a division operation to obtain the local discrete curvature value of the sub-pixel geometric sampling point. .

[0103] Finally, by traversing all sub-pixel geometric sampling points on the sub-pixel level parameterized curve, an edge curvature sequence describing the distribution of the sharpness of the particle surface is generated. .in, This represents the total number of subpixel geometric sampling points, and its value depends on the physical perimeter of the particle and the preset subpixel sampling density.

[0104] Step S2200: Traverse the edge curvature sequence Local discrete curvature values ​​in Effective cutting points are extracted by screening using a preset coating damage threshold. A single-point mechanical damage force unit is introduced to physically weight the effective cutting points to calculate the cumulative attack potential energy. And combined with the total number of effective cutting points and before Mean of maximum curvature Constructing effective cutting feature vectors .

[0105] Specifically, this step aims to transform the edge curvature sequence, which characterizes the micro-geometry of the particles, calculated in step S2100. As data input, from a massive amount of local discrete curvature values The method uses threshold logic to filter out truly effective cutting points with physical destructive power, and introduces the concept of mechanical equivalence from contact mechanics. It then performs nonlinear weighted aggregation on the selected effective cutting points to construct an effective cutting feature vector describing the overall wear and damage capability of the activated carbon particles on the catalyst bead coating. .

[0106] In the specific implementation process, this step traverses the edge curvature sequence. The goal is to find local maxima in the sequence. Not all protrusions will scratch the color-changing beads; only local discrete curvature values ​​will. Only protrusions exceeding certain physical limits pose a threat. Therefore, a critical curvature threshold for coating failure is defined, based on the physical hardness and yield strength of the indicator coating on the color-changing beads surface. All protrusions satisfying the locally discrete curvature values ​​are then selected. Local maxima exceeding the critical curvature threshold for coating failure are defined as effective cutting points. These effective cutting points physically correspond to microscopic, knife-like physical features on the particle surface that are sharp enough to generate enough pressure at the moment of contact to pierce the coating.

[0107] For each selected effective cutting point, it is no longer treated as an equivalent counting unit, but rather a single-point mechanical destructive force unit is introduced for physical equivalent weighting. The construction of the single-point mechanical destructive force unit is based on the following physical law: on the one hand, the sharper the microscopic protrusions on the surface of activated carbon particles, i.e., the higher the local discrete curvature value... The larger the diameter, the smaller the contact area when it comes into contact with the catalyst beads, resulting in an exponential increase in local pressure and a stronger destructive force on the coating. On the other hand, activated carbon itself is a brittle porous material. If the geometric structure of the micro protrusions is too thin and sharp, that is, the aspect ratio is too large, it is very easy to undergo brittle fracture during mechanical extrusion, thereby weakening its ability to continuously cut the color-changing beads.

[0108] Therefore, this step introduces the contact mechanics index and the brittle fracture correction factor to calculate the cumulative attack potential energy of a single activated carbon particle. The accumulated attack potential This is a quantitative method for calculating the comprehensive physical potential of an indicator coating to undergo irreversible damage or pulverization during random physical contact between a single activated carbon particle and a catalyst bead. The specific calculation logic is as follows: First, calculate the difference between the local discrete curvature value and the critical curvature threshold for coating damage at each selected effective cutting point, i.e., the over-limit sharpness. Then, perform a power operation based on the contact mechanics exponent to simulate the nonlinear exponential amplification effect of contact pressure increasing with the local discrete curvature value, obtaining the theoretical damage potential energy at that point. Next, multiply the theoretical damage potential energy using a brittle fracture correction factor to obtain the single-point damage potential energy, eliminating false attack forces caused by the brittle fracture of the activated carbon particle itself. Finally, sum the single-point damage potential energies of all selected effective cutting points to obtain the cumulative attack potential energy characterizing the overall wear risk of the particle. This calculation process transforms discrete geometric curvature signals into continuous physical energy signals, achieving cross-dimensional feature mapping from microscopic morphological observation to macroscopic wear prediction. The contact mechanical index is a preset constant greater than 1, corresponding to the stress distribution law in Hertzian contact theory, used to characterize the nonlinear growth trend of physical destructive force with increasing curvature; the brittle fracture correction factor is a factor with a value range of... The dimensionless coefficient is used to correct the influence of the mechanical properties of activated carbon materials on the wear effect.

[0109] Finally, an effective cutting feature vector was constructed to describe the overall physical wear risk of this single activated carbon particle. The specific construction logic is as follows: First, count the total number of valid cutting points that have been selected. First, it serves as a frequency characteristic describing the wear frequency; second, it utilizes the calculated cumulative attack potential energy. The energy characteristic is used to describe the overall wear intensity; then, the local discrete curvature values ​​of all selected effective cutting points are sorted in descending order, and the top-ranked values ​​are extracted. The local discrete curvature values ​​are calculated and their arithmetic mean is obtained to generate the previous value. Mean of maximum curvature Finally, the three independent physical characteristics mentioned above are combined to obtain the total number of effective cutting points. Accumulated attack potential and the former Mean of maximum curvature Construct an effective cutting feature vector .

[0110] Step S2300: Use a preset risk weight vector to evaluate the effective cutting feature vector. A weighted aggregation operation is performed to generate a comprehensive risk score, and the comprehensive risk score is mapped to a physical wear risk index using a Sigmoid nonlinear activation function. .

[0111] Specifically, this step aims to transform the effective cutting feature vector constructed in step S2200 into... As input, a probability mapping unit that has converged in advance based on historical failure samples is used to nonlinearly map the high-dimensional feature space into a one-dimensional normalized scalar, namely the physical wear risk index. This physical wear and tear risk index The aim is to quantify the probability that current particles will act as abrasives in future mixed-cargo logistics scenarios, causing catalyst bead coating to detach, thereby providing a basis for rejection in subsequent steps.

[0112] In the specific implementation process, this step calls the preset risk weight vector and effective cutting feature vector. The inner product operation is performed, and a bias term is added to generate a comprehensive risk score whose value range covers the set of real numbers. This process achieves a weighted average of wear failure contributions across different physical characteristic dimensions, that is, assigning a specific number of effective cutting points based on historical data statistical patterns. Accumulated attack potential and the former Mean of maximum curvature Different levels of wear and tear contribute to failure. For example, accumulated attack potential is learned through training. Its contribution to wear failure is higher than the total number of effective cutting points alone. This results in a larger numerical coefficient being assigned to it in the risk weight vector. The risk weight vector is a set of feature weight parameters obtained by a supervised learning algorithm during the offline phase through training on a large number of aggressive and harmless particle samples that cause discoloration beads to pulverize. It is used to characterize the effective cutting feature vector. The wear fault contribution weights of each feature component; the bias term is the intercept parameter used to define the basic sensitivity threshold of the classifier.

[0113] To transform the aforementioned comprehensive risk score into a control signal with clear probabilistic meaning, a Sigmoid nonlinear activation operation is introduced. The specific logic is as follows: First, the comprehensive risk score is inverted. Second, using the natural constant as the base, an exponential operation is performed on the inverted comprehensive risk score. This process utilizes the monotonically increasing characteristic of the exponential function to map the linear comprehensive risk score into a nonlinear response value. Third, the nonlinear response value of the exponential operation is added to the value 1 to construct a normalized denominator. Fourth, the reciprocal operation is performed on the normalized denominator, i.e., divided by 1. Through the above series of mathematical transformations, the originally discrete and dimensionlessly variable geometric and mechanical characteristics are forcibly compressed and mapped into a standardized probability exponent between 0 and 1, namely the physical wear risk index. This enables the identification and removal of aggressive activated carbon particles that can easily act as abrasives and damage the color-changing bead coating, based on a unified risk threshold.

[0114] Step S3000: Traverse the set of surface texture grayscale images The region of interest in the texture is mapped to a three-dimensional grayscale topological surface, and the fractal dimension eigenvalues ​​are calculated using the box-counting method. It performs a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, constructs a local frequency domain energy distribution field based on a preset sliding window step size, and calculates the proportion of hygroscopic texture through binarization segmentation. Using exponential multiplicative coupled logic to analyze fractal dimension eigenvalues and the proportion of moisture-absorbing texture Perform a nonlinear probability mapping to generate an environmental moisture absorption preemption index. .

[0115] Specifically, this step aims to collect the surface texture grayscale images output in step S1300. As a feature extraction source, the fractal dimension eigenvalues ​​characterizing the global porosity complexity are calculated using the differential box-counting dimension method. The frequency domain energy sensing algorithm was used to calculate the proportion of moisture absorption texture representing local strong moisture absorption areas. Finally, the feature parameters of the above two dimensions are mapped to the environmental moisture absorption and preemption index, which quantifies the instantaneous water-grabbing ability of particles, through a nonlinear weighting unit. This provides a quantitative basis for eliminating potential particles that cause false color changes in color-changing beads.

[0116] Further, step S3000 includes:

[0117] Step S3100: Traverse the set of surface texture grayscale images single-frame surface texture grayscale image Using a single-frame edge contour binary image As a spatial index mask, in the single-frame surface texture grayscale image The region of interest of the texture is clipped and mapped to a 3D grayscale topological surface. A differential box-counting dimension operation based on multi-scale meshing is then performed on the 3D grayscale topological surface to calculate the fractal dimension eigenvalues. .

[0118] Specifically, this step aims to collect the surface texture grayscale images output in step S1300. As a data source, the set of binary edge contour maps output in step S1200 is used. Binary image of single-frame edge contour grayscale image of surface texture in a single frame The region of interest (ROI) for the texture of activated carbon particles was located and segmented. Subsequently, the two-dimensional grayscale distribution within this ROI was mapped to a three-dimensional grayscale topological surface, and its fractal dimension eigenvalue was calculated using the box-counting method. This technology transforms the invisible hygroscopic pore features on the surface of activated carbon into quantifiable texture topological complexity using computer vision, enabling the identification of potentially incompatible particles with extremely complex surface microporous structures and strong capillary siphon effects.

[0119] In the specific implementation process, this step involves using a set of surface texture grayscale images. Retrieve the current single-frame surface texture grayscale image Due to the grayscale image of the surface texture in this single frame. It is field-of-view data containing the background and multiple particles, and it uses binary edge contour maps of a single frame acquired at the same timestamp. As a spatial index mask, using a single-frame edge contour binary image. The location information of connected components in the grayscale image of surface texture in a single frame. The region of interest for each individual activated carbon particle texture is cut out to ensure that the subsequent fractal dimension calculation is only for the particle body surface, eliminating the dilution effect of the conveyor belt background on the texture complexity.

[0120] For each region of interest in the texture of the cut-out activated carbon particles, this step converts the pixel coordinates of the two-dimensional image plane. Mapped to three-dimensional Euclidean space shaft and axis, the pixel coordinate The grayscale values ​​are mapped to a three-dimensional Euclidean space. The axis is used to construct a three-dimensional grayscale topological surface that characterizes the microscopic undulations of the particle surface.

[0121] Subsequently, the differential box-counting dimension method was used to perform multi-scale meshing analysis on the three-dimensional grayscale topological surface to generate the number of local boxes. The specific logic of the multi-scale meshing analysis is as follows: First, all pixels within each grid are traversed, the maximum and minimum grayscale values ​​are extracted, and the difference between them is calculated to obtain the local grayscale amplitude difference. This local grayscale amplitude difference is used to characterize the physical depth span of the micropore structure in the local region. Second, using the box height normalized unit as a scale, a division operation is performed on the local grayscale amplitude difference to eliminate the influence of image resolution and grayscale level on the dimensionality of the measurement result, achieving a unification of spatial scale and grayscale scale. Finally, the result of the division operation is rounded up to convert the continuous depth value into a discrete number of local boxes. This calculation process transforms the pore depth information of the activated carbon surface into discretized spatial filling volume information.

[0122] The box height normalization unit is a calibration constant that aligns the numerical scale of the grayscale dimension with the numerical scale of the spatial dimension. Its specific calculation logic is as follows: The ratio of the total number of grayscale levels in the image to the spatial resolution scale of the texture's region of interest is calculated. This ratio defines the average grayscale dynamic range carried by a unit pixel width, i.e., establishing a normalization ratio coefficient between vertical grayscale variation and horizontal spatial distance. Subsequently, the normalization ratio coefficient is multiplied by the current grid scale, thereby linearly mapping the box's spatial dimensions on the plane to its grayscale dimensions in height. The total number of grayscale levels in the image is the number of grayscale levels on the three-dimensional grayscale topological surface. The maximum theoretical height in the axial direction is used to define the numerical limit of micro-texture undulations. Its value usually depends on the image bit depth, for example, 256 levels for an 8-bit image; the spatial resolution scale of the texture region of interest is the total width of pixels in the horizontal or vertical direction of the activated carbon particle texture region of interest; the grid scale is the basic spatial sampling interval set when dividing the texture region of interest into grids, used to detect the detail self-similarity of the activated carbon surface texture at different magnifications.

[0123] Finally, a series of different mesh scales are constructed, and for each specific mesh scale, the surface texture grayscale image of a single frame is traversed. For all local regions, i.e., the regions of interest for activated carbon particle texture, the total number of boxes required to cover the 3D grayscale topological surface at this grid scale is calculated, which is the sum of the number of local boxes for each region of interest for activated carbon particle texture. Then, a double logarithmic coordinate system is constructed, with the logarithm of the relative scale index as the x-axis and the logarithm of the total number of boxes as the y-axis. The calculation results at different scales are plotted as a scatter set. Next, the least squares method is used to linearly fit the scatter set to obtain the slope of the fitted line. This slope value is the fractal dimension eigenvalue. This method is used to reveal the growth rate of activated carbon surface texture complexity with varying observation scale. The relative scale index is the ratio of the spatial resolution scale of the texture region of interest to the grid scale, used to eliminate computational errors caused by differences in the physical resolution of the images.

[0124] From a physicochemical perspective, this fractal dimension eigenvalue The larger the value, the more the microscopic pore details are revealed as the observation scale shrinks. This means that there are a large number of nanoscale capillary structures on the particle surface. These capillary structures are the physical cause of the instantaneous loss of moisture from the environment of the color-changing beads.

[0125] Step S3200: Traverse the set of surface texture grayscale images using a spatial sliding window. single-frame surface texture grayscale image A two-dimensional discrete Fourier transform is performed to generate a local complex spectrum matrix. Based on the local complex spectrum matrix, the local frequency domain energy value is calculated. According to the preset sliding step size of the spatial sliding window, the surface texture grayscale image of a single frame is traversed. All local frequency domain energy values ​​are used to construct a local frequency domain energy distribution field, and a moisture absorption hotspot determination threshold is introduced. The moisture absorption hotspot region is locked through binarization segmentation, and the moisture absorption texture ratio of the moisture absorption hotspot region is calculated. .

[0126] Specifically, this step aims to collect the surface texture grayscale images output in step S1300. As a data source, the spatial domain image data indexed by pixel coordinates is transformed into a local complex spectrum matrix indexed by frequency domain coordinates through sliding window frequency domain analysis. This allows for the location of densely populated micropore regions on the particle surface where the distribution density of nanoscale micropores exceeds the preset texture medium frequency, and outputs the moisture absorption texture ratio. To compensate for the fractal dimension eigenvalues ​​in step S3100 The inability to characterize local spatial distribution features is a limitation. The densely packed micropore regions physically constitute hygroscopic hotspots with hygroscopic porosity, and are the dominant factor leading to localized water loss from the color-changing beads in contact with them. The sustained water-grabbing capacity of the particles is assessed by quantifying the area proportion of these hygroscopic hotspots.

[0127] In the specific implementation process, this step involves using a set of surface texture grayscale images. Retrieve single-frame surface texture grayscale image A spatial sliding window with a preset size is set. This spatial sliding window slides across a single frame of surface texture grayscale image with a preset sliding step size. The image is discretized into a series of local sub-image regions by scanning row by row and column by column on a two-dimensional pixel plane. For each local sub-image region covered by a spatial sliding window, this step performs a two-dimensional discrete Fourier transform (2D-DFT) or a fast Fourier transform (FFT) to transform the pixel grayscale data within that region from pixel coordinates... The spatial domain indexed is transformed to the frequency domain coordinates. For the frequency domain indexed, a corresponding local complex spectrum matrix is ​​generated. Each element in this local complex spectrum matrix is ​​a complex number, consisting of a real part and an imaginary part, used to resolve the spectral components of the local texture.

[0128] For each local complex spectrum matrix corresponding to a local sub-image region, its local frequency domain energy value is calculated. The specific calculation logic is as follows: First, spectral amplitude analysis. Perform complex modulo operation on each frequency domain coordinate in the local complex spectrum matrix, that is, calculate the real and imaginary parts of the complex element at that frequency domain coordinate, and calculate the arithmetic square root of the sum of their squares to obtain the spectral amplitude corresponding to that frequency domain coordinate. This spectral amplitude characterizes the oscillation intensity of the texture signal at a specific spatial frequency in the current local sub-image region. Second, high-frequency energy integration. Set the texture cutoff frequency as the frequency domain boundary line distinguishing between macroscopic surface undulations and microscopic hygroscopic pores. On the frequency domain plane, select the frequency band of interest where the spectral amplitude is greater than the texture cutoff frequency and less than the Nyquist frequency. Perform square integration on all spectral amplitudes falling within this frequency band of interest to calculate the energy value in the local sub-image region in pixel coordinates. The local frequency domain energy value is the anchor point. This local frequency domain energy value quantifies the intensity of texture oscillations caused by nanoscale pores within a local micro-region. The texture cutoff frequency is a lower limit parameter for frequency domain integration with a clear physical direction. Its value is determined based on the effective micropore diameter of activated carbon, which generates strong capillary action, typically 2nm-50nm, and its mapping relationship with the optical magnification of the imaging system. The Nyquist frequency is an upper limit parameter for frequency domain integration, determined by the physical resolution of the image sensor, and is used to characterize the theoretical limit of the most subtle texture changes that the imaging system can capture.

[0129] Based on the preset sliding step size, traverse the grayscale images of surface textures in a single frame. pixel coordinates of the center of all sliding windows in space Each calculated local frequency domain energy value is mapped back to its corresponding spatial location, constructing a grayscale image corresponding to the surface texture of a single frame. The local frequency domain energy distribution field is strictly aligned with the pixel coordinate system. In this local frequency domain energy distribution field, each numerical point no longer represents grayscale brightness, but rather the micropore distribution density at the corresponding physical geometric location, that is, the potential for plundering moisture from the catalyst beads at that location.

[0130] Based on the local frequency domain energy distribution field, a hygroscopic hotspot determination threshold is introduced to perform binarization segmentation. When a pixel coordinate... When the local frequency domain energy value is greater than the moisture absorption hotspot detection threshold, the pixel location is determined to belong to a highly predatory moisture absorption hotspot region; otherwise, it is determined to be a normal surface region. Finally, the total number of pixels marked as moisture absorption hotspot regions is counted and divided by the total projected area of ​​the activated carbon particles to calculate the moisture absorption texture ratio. The total projected area of ​​the activated carbon particles refers to the area within a single frame of the binary image showing the edge contour. The total number of pixels contained in the connected domain of this particle.

[0131] Step S3300, based on fractal dimension eigenvalues and the proportion of moisture-absorbing texture The physical potential energy value is calculated using an exponential multiplicative coupling term, and then converted into an environmental moisture absorption preemption index through a probability mapping activation layer. .

[0132] Specifically, this step aims to extract the fractal dimension eigenvalues ​​representing the global topological complexity from step S3100. The proportion of hygroscopic texture representing the coverage of local hotspots extracted in step S3200. As input to multidimensional image features, a nonlinear weighted coupling unit conforming to physicochemical laws is constructed to map the geometric multidimensional image features to the physical dimension of the environmental moisture absorption preemption index. This allows for a comprehensive assessment of the overall risk of activated carbon particles instantly drying out the moisture in catalyst beads at the microscopic contact surface, thus compensating for the limitations of relying on a single feature in predicting complex physicochemical reactions.

[0133] In the specific implementation process, this step no longer involves fractal dimension eigenvalues. and the proportion of moisture-absorbing texture Instead of a simple linear superposition, this step constructs a nonlinear weighted coupling logic based on the physical mechanism of moisture preemption effect: only when the particle surface simultaneously possesses an extremely complex microporous structure (i.e., a high moisture absorption rate) and a wide distribution area (i.e., a high moisture absorption capacity) will a moisture-preventing flow sufficient to cause instantaneous water loss from the color-changing beads be formed. Therefore, this step constructs an exponential multiplicative coupling term, i.e., calculates the fractal dimension eigenvalue. The power of the fractal sensitivity index and the proportion of hygroscopic texture The product of these terms yields the physical potential energy value. This exponential multiplicative coupling term simulates the physical process by which complex microporous structures exhibit nonlinear explosive growth, limited by the effective contact area, as the complexity of the microporous structure increases. The fractal sensitivity index is a preset constant greater than 1, used to evaluate the fractal dimension eigenvalue. Exponential weighting is applied to reflect the physical characteristic that the finer the micropores, the more non-linearly the capillary suction increases.

[0134] To transform the wide-range physical potential energy value calculated by the aforementioned exponential multiplicative coupling term into a standardized control signal usable in industrial settings, the system introduces the Sigmoid function as a probabilistic mapping activation layer. The specific operational logic is as follows: First, the physical potential energy value is multiplied by the moisture absorption conversion gain and a bias term is added to complete the linear transformation; second, the result after adding the bias term is inverted and an exponential operation with the natural constant as the base is performed; third, the result of the exponential operation is added to the value 1; finally, the reciprocal of the sum is taken. This mapping process forcibly compresses the input feature, i.e., the physical potential energy value, into a numerical range of (0, 1), generating the final environmental moisture absorption preemption index. This achieves the conversion from physical potential to failure probability. The moisture absorption conversion gain is a scaling factor used to map the product of geometric eigenvalues, i.e., the physical potential value, to the risk probability space. Its value is obtained through backpropagation training or experimental calibration on known pseudo-discoloration failure samples. The bias term is used to set the basic tolerance or sensitivity threshold for the risk of microenvironment humidity preemption.

[0135] Step S4000: Obtain the physical wear and tear risk index. and the index of moisture absorption in the environment An environment-medium dual-constraint weighted algorithm is introduced to calculate the compatibility score of color-changing beads. Combined with preset optimization and rejection thresholds and physicochemical testing data of activated carbon raw material batches, the system performs multi-path diversion decisions and generates particle classification control instructions pointing to different diversion states. .

[0136] Specifically, this step aims to utilize the physical wear risk index, which characterizes the physical sharpness of particles, extracted in step S2300. The environmental moisture preemption index, which characterizes the instantaneous moisture absorption capacity of particles, extracted in step S3300. As multimodal input data, it is mapped to a unified color-changing bead compatibility score through a weighted fusion algorithm. And based on the compatibility rating of the color-changing beads The comparison results with the preset classification threshold generate granular hierarchical control commands to drive the actuator's actions. This diverts activated carbon granules to the color-changing formaldehyde-removing charcoal bag production line, the ordinary deodorizing charcoal bag production line, or the waste treatment channel.

[0137] Further, step S4000 includes:

[0138] Step S4100, based on the physical wear risk index and the index of moisture absorption in the environment Coating physical strength data and average relative humidity data are introduced as external operating condition constraint parameters. A dual-constraint weighted algorithm based on environment and medium is used for data fusion to generate a color-changing bead compatibility score. .

[0139] Specifically, this step aims to address the lack of generalization in using a single fixed threshold screening method, which cannot adapt to the characteristics of different batches of color-changing beads and different warehousing and logistics environments. It will utilize the physical wear risk index, representing the geometric dimensions, output from step S2300. The environmental moisture preemption index representing the texture dimension output by step S3300 As multimodal heterogeneous feature input, data fusion is performed using an environment-medium dual-constraint weighted algorithm. By dynamically adjusting risk weights, a uniform quantitative score for the coexistence stability of activated carbon particles and catalyst beads is generated to assess the compatibility of the color-changing beads. This provides numerical data for subsequent hierarchical decision-making.

[0140] In the specific implementation process, in order to determine the physical wear risk index and the index of moisture absorption in the environment In the final evaluation system, this step introduces an environment-medium dual-constraint weighted algorithm to assess the relative contribution to damage. It reads the physical strength data of the catalyst beads used in the current production batch, such as the wear resistance coefficient, and the average relative humidity data of the storage and logistics environment, as external operating condition constraint parameters. Based on these two external operating condition constraint parameters, the coating mechanical strength weight coefficient and the environmental humidity sensitivity weight coefficient are dynamically calculated and adjusted. Then, a weighted fusion calculation is performed to map the two independent risk indicators to a unified evaluation scalar, namely the color-changing bead compatibility score. The coating mechanical strength weighting coefficient is a weighting factor set based on the coating physical strength data, and it is negatively correlated with the physical wear resistance of the indicator coating on the color-changing beads surface. When the externally read coating physical strength data indicates that the coating of the current batch of color-changing beads is brittle or thin, the value of this coefficient is automatically increased, adding a physical wear risk index to the overall score calculation. The penalty weighting applies a more stringent geometric screening standard. The environmental humidity sensitivity weighting coefficient is a weighting factor set based on average relative humidity data and is negatively correlated with the real-time relative humidity of the warehousing and logistics environment. When externally read average relative humidity data indicates a dry logistics environment, the color-changing beads are in a critical state of water loss and are extremely sensitive to external moisture absorption. The coefficient is automatically increased to combat the high environmental moisture absorption preemption index. The particles prevent false discoloration.

[0141] The compatibility rating of the color-changing beads It is a dimensionless evaluation index with a value range within the closed interval [0, 1], used to quantitatively characterize the coexistence stability of activated carbon particles and color-changing beads in a mixed system. Its specific calculation logic is as follows: First, using the coating mechanical strength weighting coefficient and the environmental humidity sensitivity weighting coefficient determined by external working conditions, respectively, the input physical wear risk index is... and the moisture absorption competition index of the microenvironment First, a weighted multiplication operation is performed, and the two weighted results are added together to obtain a comprehensive risk weighted sum. Second, this comprehensive risk weighted sum is divided by the sum of the coating mechanical strength weighting coefficient and the environmental humidity sensitivity weighting coefficient, and a linear normalization operation is performed to obtain a linear risk mean between 0 and 1. This linear risk mean reflects the average hazard level of particles under the current specific working conditions. Third, a power operation based on a preset risk sensitivity index is performed on this linear risk mean, utilizing the non-linear characteristics of the power function to amplify the signal strength of high-risk values. Finally, the result of the power operation is subtracted from the value of 1, reversing the risk dimension to a compatibility dimension, generating the final color-changing bead compatibility score. The risk sensitivity is typically set to a positive real number less than 1 to adjust the convexity of the scoring function.

[0142] Step S4200: Scoring the compatibility of the color-changing beads based on preset preferred thresholds and elimination thresholds. Numerical comparisons are performed, and multi-path diversion decisions are executed in conjunction with the physicochemical testing data of activated carbon raw material batches to generate particle classification control instructions. If the color-changing beads have a compatibility rating If the value is not lower than the preferred threshold, it is mapped to the first shunt instruction status value. If the color-changing beads have a compatibility score If the value is below the elimination threshold and the physicochemical test data meets the standard, it is mapped to the second diversion command status value. Otherwise, it is mapped to the third branch instruction status value. .

[0143] Specifically, this step aims to improve the compatibility score of the color-changing beads output from step S4100. Using the physicochemical testing data of this batch of activated carbon raw materials as decision input, this approach breaks through the limitations of traditional machine vision screening's binary classification of "qualified" and "unqualified." It executes multi-path diversion decisions, identifying and separating specific particle groups that are deemed unsuitable for coexistence with catalyst beads due to their overly sharp microstructure or excessively rapid hygroscopic absorption, but which still exhibit excellent chemical adsorption performance. This generates differentiated particle grading control instructions. This allows them to be directed to different product production lines, realizing a shift from simple defect elimination to refined resource scheduling based on morphological characteristics.

[0144] The physicochemical testing data of the batch of activated carbon raw materials are read through the Industrial Internet of Things (IIoT) interface or the data port of the Laboratory Information Management System (LIMS) to obtain batch attribute data associated with the current activated carbon particle flow, including the batch average iodine value. and standard adsorption value The average iodine value of the batches. This refers to the macroscopic chemical adsorption capacity index of the batch of activated carbon raw materials currently flowing on the conveyor belt, usually expressed in milligrams per gram (mg / g). It is obtained by sampling and testing the raw materials using a standard chemical titration method during the raw material warehousing or feeding stage, and this value is then entered into the system database as a global attribute of the batch. The standard adsorption value... This indicates the minimum threshold for the adsorption performance of raw materials for ordinary deodorizing charcoal bags that do not contain color-changing beads. It is usually set as an industrial standard value, such as 1000 mg / g. The value is determined by: being pre-set according to the quality control standards of the final product and stored in the system's configuration parameters.

[0145] In the specific implementation process, this step will score the compatibility of the color-changing beads. The values ​​are compared with preset preferred and rejection thresholds, and combined with the batch average iodine value. and standard adsorption value Perform multi-path diversion decisions and generate granular hierarchical control instructions. The preferred threshold is a first-level judgment standard constant preset in the multi-path diversion decision, used to define the minimum quality boundary of the special raw materials for color-changing formaldehyde removal charcoal bags. Its value is usually set as a positive real number close to 1, and its value is based on the statistical confidence level required for the finished color-changing beads to maintain zero failures within a specific shelf life. The elimination threshold is a second-level judgment standard constant preset in the multi-path diversion decision, used to define the critical point at which activated carbon particles pose a substantial hazard to the color-changing beads. Its value is a positive real number significantly lower than the preferred threshold, and its value is based on the risk score lower limit corresponding to the physical yield limit and critical water loss humidity of the color-changing bead coating.

[0146] The specific multi-path routing decision is as follows:

[0147] The first diversion stage targets color-changing formaldehyde-removing charcoal pack products. This path determines the compatibility score of the color-changing beads. Is it greater than or equal to a preset preferred threshold? If this inequality condition holds, it indicates that the activated carbon particles simultaneously satisfy the dual constraints of geometric passivation and gradual moisture absorption in terms of microscopic morphological characteristics, i.e., its physical wear risk index... and the moisture absorption competition index of the microenvironment All are within a safe range that is harmless to the color-changing beads. Physically, this means that the particle surface lacks sharp microscopic blades that can pierce the coating, and does not possess dense nanopores that can trigger a strong instantaneous capillary siphon effect. Therefore, it will not cause physical damage or moisture loss to the catalyst beads during microscopic physical contact.

[0148] Based on this determination, the first branch instruction status value is generated for this path. Particle grading control instruction for "diverting to the color-changing formaldehyde removal activated carbon bag production line" This particle grading control command By directing these particles into high-end product channels, we can ensure from the source that the carrier and functional beads can coexist stably for a long time in a mixed system containing color-changing beads, thus preventing false color changes and coating wear failures caused by microscopic mutual harm between components.

[0149] Second diversion state: For ordinary deodorizing charcoal bags. This path determines the compatibility score of color-changing beads. Whether it is less than the preset elimination threshold, and at the same time determine the average iodine value of the batch. Is it greater than or equal to the preset standard adsorption value? If both of the above inequalities hold simultaneously, it indicates that although the activated carbon particles have defects in terms of biocompatibility—namely, their physical aggression towards the color-changing beads due to the presence of microscopic razor-like structures or strong hygroscopic hot spots—this microscopic aggression often stems physically from their extremely well-developed pore structure and huge specific surface area. This characteristic, considered a risk factor in two-component blends containing color-changing beads, transforms into excellent formaldehyde adsorption performance in ordinary carbon packs of single-component products without color-changing beads.

[0150] Based on this determination, this path generates a second branching instruction status value. Particle grading control instruction for "diverting to ordinary deodorizing charcoal bag production line" This particle grading control command By guiding the particles into the secondary product channel, the source of failure in high-end products is eliminated, and the cost of treating high-adsorption raw materials as waste is avoided, thus achieving refined material scheduling based on micromorphology.

[0151] Third diversion state: for invalid waste. This path performs logical complement operation, if the activated carbon granules' color-changing bead compatibility score... If either of the criteria for the first and second diversion states cannot be met, the third diversion state is triggered.

[0152] The logical complement operation covers two physical scenarios: first, the severe defect scenario. (Color-changing bead compatibility rating) Less than the elimination threshold, and the average iodine value of the batch is It is also less than the standard adsorption value. At this point, the particle exhibits physical aggression towards the color-changing beads in terms of microscopic morphology, and simultaneously fails to meet the basic adsorption standards for air purification in terms of chemical properties, thus constituting a substandard product that must be discarded. Secondly, the intermediate ambiguous state scenario. Color-changing bead compatibility rating. It falls between the elimination threshold and the selection threshold. At this point, although the particle does not exhibit extreme physical aggression, its biocompatibility is insufficient to support its entry into the high-end color-changing product line, and it lacks clear high adsorption performance characteristics to support its entry into the diversion and reuse product line. It belongs to mediocre particles that lack specific application value.

[0153] Based on this determination, this path generates a third branch instruction status value. Particle grading control command for "rejection" This particle grading control command The particles are diverted to a waste collection bin or a crushing and reprocessing process, thereby physically removing them completely from the finished product stream.

[0154] Example 2:

[0155] This embodiment, based on Embodiment 1, provides an activated carbon raw material screening system based on multimodal image feature fusion, such as... Figure 4 As shown, the system includes an image acquisition module, a physical wear assessment module, a moisture absorption risk assessment module, and a compatible diversion decision module;

[0156] The image acquisition module is used to acquire the particle flow of activated carbon to be screened. single-frame edge contour original image and the original image of surface texture in a single frame For the original image of the edge contour of a single frame Perform a topology-preserving segmentation algorithm based on variational level sets to construct a set of binary edge contour maps. For a single frame of the original surface texture image Homomorphic filtering algorithm is introduced to construct a set of surface texture grayscale images. .

[0157] The physical wear assessment module is used to traverse the set of binary edge contour images. The edge chain code is extracted and then reconstructed using sub-pixel-level parameterization to generate an edge curvature sequence. Based on a preset coating damage critical threshold, from the edge curvature sequence Extracting effective cutting feature vectors Physical wear risk index is generated through probability mapping. .

[0158] The moisture absorption risk assessment module is used to traverse the set of surface texture grayscale images. The region of interest in the texture is mapped to a three-dimensional grayscale topological surface, and the fractal dimension eigenvalues ​​are calculated using the box-counting method. It performs a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, constructs a local frequency domain energy distribution field based on a preset sliding window step size, and calculates the proportion of hygroscopic texture through binarization segmentation. Using exponential multiplicative coupled logic to analyze fractal dimension eigenvalues and the proportion of moisture-absorbing texture Perform a nonlinear probability mapping to generate an environmental moisture absorption preemption index. .

[0159] The compatible flow splitting decision module is used to obtain the physical wear risk index. and the index of moisture absorption in the environment An environment-medium dual-constraint weighted algorithm is introduced to calculate the compatibility score of color-changing beads. Combined with preset optimization and rejection thresholds and physicochemical testing data of activated carbon raw material batches, the system performs multi-path diversion decisions and generates particle classification control instructions pointing to different diversion states. .

[0160] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0161] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for screening activated carbon raw materials based on multimodal image feature fusion, characterized in that, include: The original edge contour image and the original surface texture image of the activated carbon particle flow to be screened are acquired in a single frame. The topology-preserving segmentation algorithm based on variational level set is performed on the original edge contour image to construct a set of binary edge contour images. The original surface texture image is introduced with a homomorphic filtering algorithm to construct a set of grayscale surface texture images. The edge contour binary image set is traversed to extract the edge chain code. The edge chain code is then reconstructed by subpixel-level parameterization to generate an edge curvature sequence. Based on a preset coating damage critical threshold, an effective cutting feature vector is extracted from the edge curvature sequence. A physical wear risk index is generated through probability mapping. Traverse the set of grayscale images of surface textures, map the regions of interest of the textures into three-dimensional grayscale topological surfaces, and use the differential box dimension method to calculate the fractal dimension eigenvalues. It performs a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, constructs a local frequency domain energy distribution field based on the preset sliding step size of the spatial sliding window, calculates the proportion of hygroscopic texture through binarization segmentation, and performs nonlinear probability mapping on the fractal dimension eigenvalues ​​and the proportion of hygroscopic texture using exponential multiplicative coupling logic to generate the environmental hygroscopic preemption index. The physical wear risk index and environmental moisture absorption preemption index are obtained. An environmental-medium dual-constraint weighted algorithm is introduced to calculate the compatibility score of the color-changing beads. Combined with the preset optimization threshold and elimination threshold and the physicochemical test data of the activated carbon raw material batch, a multi-path diversion decision is executed to generate particle classification control instructions pointing to different diversion states.

2. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 1, characterized in that, The method for constructing the surface texture grayscale image set includes: Two light field slice data acquisitions were performed on the activated carbon particle flow to be screened at the same spatial location. During the first light field slice data acquisition, the backlight illumination module was used to acquire a single frame of the original edge contour image. During the second light field slice data acquisition, the side structured light illumination module was switched to acquire a single frame of the original surface texture image. Based on the original image of a single frame edge contour, a topology-preserving segmentation algorithm based on variational level sets is constructed to generate a binary image of the single frame edge contour and to build a set of binary edge contour images. A pixel-by-pixel logical AND operation is performed on the original single-frame surface texture image using the single-frame edge contour binary image as a spatial topological constraint. A homomorphic filtering algorithm based on frequency domain signal processing is introduced to generate a single-frame surface texture grayscale image and construct a set of surface texture grayscale images.

3. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 2, characterized in that, The step of switching to the lateral structured light illumination module to acquire a single frame of original surface texture image during the second light field slice data acquisition includes: Within a preset time interval after the first light field slice data acquisition is completed, the lateral structured light illumination module is triggered to control the incident light to irradiate the activated carbon particle flow at a preset incident grazing angle; the incident grazing angle is the geometric angle between the main optical axis of the lateral structured light illumination module and the transmission plane of the activated carbon particle flow, and its value is limited to a closed range of 10° to 15°. Based on the inverse application principle of Lambert's cosine law, the light signal reflected by the micro-pore structure on the surface of activated carbon particles is received, and the surface depth undulation is converted into brightness gradient changes to generate a single-frame original surface texture image; the micro-pore structure is mapped as gray value pixel clusters in the single-frame original surface texture image. Extract the lateral texture reflection brightness response value at each pixel coordinate of the original image of the surface texture of a single frame, and obtain the backlight residual brightness response value at the first light field slice data acquisition. A texture confidence enhancement unit is introduced, which performs sinusoidal weighted modulation on the lateral texture reflection brightness response value based on the sine value of the incident grazing angle, and divides the sinusoidal weighted modulation result by the sum of the backlight residual brightness response value and the preset numerical stability constant to obtain a ratio result; the ratio result is multiplied by the preset photoelectric conversion gain constant to generate a feature enhancement contrast coefficient; the feature enhancement contrast coefficient is the moisture absorption porosity feature weight of the pixel in the original image of the single frame surface texture.

4. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 1, characterized in that, The method for generating the physical wear and tear risk index includes: Particle connected components are extracted from single-frame edge contour binary maps in the edge contour binary map set to mark the particle objects to be analyzed, and the edge chain code of the particle objects to be analyzed is extracted. The edge chain code is mapped to a sub-pixel level parameterized curve using a B-spline curve fitting algorithm. The sub-pixel level parameterized curve is solved based on the central difference method, and the local discrete curvature value is calculated to construct the edge curvature sequence. The local discrete curvature values ​​in the edge curvature sequence are traversed, and effective cutting points are extracted by filtering using a preset coating damage critical threshold. A single-point mechanical damage force unit is introduced to physically weight the effective cutting points to calculate the cumulative attack potential energy, and this is combined with the total number of effective cutting points and the previous... Construct an effective cutting feature vector using the mean of maximum curvature; A comprehensive risk score is generated by weighting and aggregating the effective cutting feature vectors using a preset risk weight vector, and then the comprehensive risk score is mapped to a physical wear risk index using a Sigmoid nonlinear activation function.

5. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 4, characterized in that, The calculation steps for the accumulated attack potential energy include: Perform a local maximum search on the edge curvature sequence and extract local maximum points whose local discrete curvature values ​​are greater than the preset coating failure critical curvature threshold as effective cutting points; For each selected effective cutting point, the difference between its local discrete curvature value and the critical curvature threshold for coating failure is calculated to generate the over-limit sharpness. Using a preset contact mechanical index as the exponent, a power operation is performed on the excessive sharpness to obtain the theoretical destructive potential energy response value; the contact mechanical index is a preset constant greater than 1. The theoretical failure potential energy response value is multiplied and corrected using a preset brittle fracture correction factor to generate a single-point failure potential energy. The single-point failure potential energy of all selected effective cutting points is then summed to output the cumulative attack potential energy.

6. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 1, characterized in that, The method for generating the environmental moisture absorption preemption index includes: Traverse the single-frame surface texture grayscale images in the set of surface texture grayscale images, use the single-frame edge contour binary image as a spatial index mask, clip the texture region of interest in the single-frame surface texture grayscale image and map it into a three-dimensional grayscale topological surface, perform differential box-counting dimension operation based on multi-scale mesh division on the three-dimensional grayscale topological surface, and calculate the fractal dimension eigenvalue. A spatial sliding window is used to traverse single-frame surface texture grayscale images in the surface texture grayscale image set. A two-dimensional discrete Fourier transform is performed to generate a local complex spectrum matrix. Based on the local complex spectrum matrix, the local frequency domain energy value is calculated. According to the preset sliding step size of the spatial sliding window, all local frequency domain energy values ​​on the single-frame surface texture grayscale image are traversed to construct a local frequency domain energy distribution field. A moisture absorption hotspot determination threshold is introduced to lock the moisture absorption hotspot region through binarization segmentation. The moisture absorption texture ratio of the moisture absorption hotspot region is calculated. Based on the fractal dimension eigenvalues ​​and the proportion of moisture-absorbing texture, the physical potential energy value is calculated using an exponential multiplicative coupling term, and the physical potential energy value is converted into an environmental moisture-absorbing preemption index through a probability mapping activation layer.

7. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 6, characterized in that, The steps for calculating the local frequency domain energy value include: Traverse each frequency domain coordinate in the local complex spectrum matrix, extract the real and imaginary part values ​​of the complex elements at each frequency domain coordinate, calculate the arithmetic square root of the sum of the squares of the real and imaginary part values, and generate the spectral amplitude of each frequency domain coordinate. The preset texture cutoff frequency is retrieved as the lower limit parameter for frequency domain integration, and the Nyquist frequency of the imaging system is retrieved as the upper limit parameter for frequency domain integration to construct the frequency band of interest. All spectral amplitudes falling within the frequency band of interest are selected, and the local frequency domain energy value is calculated by performing square integration on all selected spectral amplitudes.

8. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 1, characterized in that, The method for generating the particle classification control command includes: Based on the physical wear risk index and the environmental moisture absorption preemption index, coating physical strength data and average relative humidity data are introduced as external working condition constraint parameters. Data fusion is performed using an environment-medium dual constraint weighting algorithm to generate a color-changing bead compatibility score. The compatibility scores of the color-changing beads are compared numerically based on preset preferred and eliminated thresholds. Combined with the physicochemical test data of the activated carbon raw material batch, a multi-path diversion decision is executed to generate particle grading control instructions. If the compatibility score of the color-changing beads is not lower than the preferred threshold, it is mapped to the first diversion instruction state value; if the compatibility score of the color-changing beads is lower than the eliminated threshold and the physicochemical test data meets the standard, it is mapped to the second diversion instruction state value; otherwise, it is mapped to the third diversion instruction state value. The physicochemical testing data of the activated carbon raw material batches are read through an industrial IoT interface or a laboratory information management system data port, which reads batch attribute data associated with the activated carbon raw material batches, including the batch average iodine value and standard adsorption value.

9. The activated carbon raw material screening method based on multimodal image feature fusion according to claim 8, characterized in that, The logical steps of the environment-medium dual-constraint weighted algorithm include: The physical strength data and average relative humidity data of the coating are read as external working condition constraint parameters. Based on the preset negative correlation mapping, the physical strength data and average relative humidity data of the coating are mapped to the coating mechanical strength weight coefficient and the environmental humidity sensitivity weight coefficient, respectively. The physical wear risk index is weighted by the coating mechanical strength weighting coefficient, and the environmental moisture absorption preemption index is weighted by the environmental humidity sensitivity weighting coefficient. The weighted multiplication results of the two are summed to generate a comprehensive risk weighted sum. The sum of the coating mechanical strength weighting coefficient and the environmental humidity sensitivity weighting coefficient is used as the normalized denominator. The weighted sum of the comprehensive risks is divided by the normalized denominator to obtain the linear risk mean. Perform a power operation on the linear risk mean with a preset risk sensitivity index as the exponent, and use the value 1 to subtract the result of the power operation to generate a color-changing bead compatibility score.

10. A system for screening activated carbon raw materials based on multimodal image feature fusion, used to implement the method for screening activated carbon raw materials based on multimodal image feature fusion as described in any one of claims 1-9, characterized in that, The system includes an image acquisition module, a physical wear assessment module, a moisture absorption risk assessment module, and a compatible diversion decision module. The image acquisition module is used to acquire single-frame edge contour original images and single-frame surface texture original images of activated carbon particle flow. It performs a topology-preserving segmentation algorithm based on variational level sets on the single-frame edge contour original images to construct a set of edge contour binary images, and introduces a homomorphic filtering algorithm on the single-frame surface texture original images to construct a set of surface texture grayscale images. The physical wear assessment module is used to traverse the set of binary edge contour maps to extract edge chain codes, perform sub-pixel-level parameterized reconstruction of the edge chain codes to generate an edge curvature sequence, extract effective cutting feature vectors from the edge curvature sequence based on a preset coating damage critical threshold, and generate a physical wear risk index through probability mapping. The moisture absorption risk assessment module is used to traverse the set of surface texture grayscale images, map the texture region of interest into a three-dimensional grayscale topological surface, and use the differential box dimension method to calculate the fractal dimension eigenvalue. It performs a two-dimensional discrete Fourier transform to generate a local complex spectrum matrix, constructs a local frequency domain energy distribution field based on the preset sliding step size of the spatial sliding window, calculates the proportion of hygroscopic texture through binarization segmentation, and performs nonlinear probability mapping on the fractal dimension eigenvalues ​​and the proportion of hygroscopic texture using exponential multiplicative coupling logic to generate the environmental hygroscopic preemption index. The compatibility diversion decision module is used to obtain the physical wear risk index and the environmental moisture absorption preemption index, introduce an environment-medium dual constraint weighted algorithm to calculate the compatibility score of the color-changing beads, and combine the preset preferred threshold and elimination threshold with the physicochemical test data of the activated carbon raw material batch to perform multi-path diversion decision and generate particle classification control instructions pointing to different diversion states.

Images