A method and system for on-line quality detection and dynamic grading and batching of recycled aggregates

Abstract

The present application relates to the technical field of recycled aggregate intelligent mixing, and particularly relates to a recycled aggregate online quality detection and dynamic grading and batching method and system. The method comprises the following steps: extracting an attached mortar-crack symbiotic defect density matrix and a standard grain shape spectrum vector of a single particle; combining the symbiotic defect density matrix and the standard grain shape spectrum vector, and obtaining a quality grading membership degree vector of a current aggregate group through graph attention coupling reasoning; performing constraint interval fuzzy robust optimization on the basis of a quality interval parameter of a stock aggregate to obtain an initial batching proportion vector, and performing spatiotemporal texture manifold recognition on a mixing image sequence to obtain a mixing homogeneity manifold coordinate vector; dynamically correcting the initial batching proportion vector by using the mixing homogeneity manifold coordinate vector to obtain a homogeneity-oriented adaptive batching vector; and fusing the quality grading membership degree vector and the homogeneity-oriented adaptive batching vector, and obtaining a final batching execution scheme after Bayesian shift compensation correction.

Description

Technical Field

[0001] This invention relates to the field of intelligent mixing technology for recycled aggregates, specifically to a method and system for online quality detection and dynamic grading and batching of recycled aggregates. Background Technology

[0002] The resource utilization of construction waste to prepare recycled aggregate is the core approach to solving the shortage of aggregate resources and pollution from solid waste stockpiling. However, recycled aggregate has complex original defects, such as cement mortar adhering to the surface, coexisting micro-cracks inside, discrete particle shape and angles, large fluctuations in the proportion of needle-like and flaky particles, and continuous nonlinear fluctuations in batch quality of aggregate, which directly leads to significant instability in concrete strength, homogeneity, and workability.

[0003] Traditional recycled aggregate quality testing mainly relies on offline laboratory screening, crushing value tests, and manual image sampling. The testing lags behind production by more than 30 minutes and cannot achieve online real-time feedback. Existing aggregate grading algorithms perform hard classification by taking the average value of single particle indicators, without considering the spatial adjacency relationship of aggregates conveyed by belt conveyors and the mutual coupling effect of quality between particles. They cannot adapt to the continuous fluctuation characteristics of recycled aggregate quality, and the grading results cannot provide a basis for continuous adjustment of fine batching. Summary of the Invention

[0004] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a method for online quality detection and dynamic grading and batching of recycled aggregates, comprising the following steps: Surface defect detection is performed on multispectral images to obtain the density matrix of mortar-crack coexistence defects for individual particles. Three-dimensional particle shape characterization is performed on the three-dimensional point cloud to obtain the standard particle shape spectrum vector for individual particles. Combining the coexistence defect density matrix and the standard particle shape spectrum vector, graph attention coupling inference is used to obtain the quality grading membership vector of the current aggregate group. Constrained interval fuzzy robust optimization is performed based on the quality interval parameters of the stored aggregates to obtain the initial batching ratio vector. Spatiotemporal texture manifold identification is performed on the mixing image sequence to obtain the homogeneity manifold coordinate vector of the mixture. The initial batching ratio vector is dynamically corrected using the homogeneity manifold coordinate vector of the mixture to obtain the homogeneity-oriented adaptive batching vector. The quality grading membership vector and the homogeneity-oriented adaptive batching vector are fused, and the final batching execution scheme is obtained after Bayesian drift compensation correction.

[0005] Optionally, the step of obtaining the density matrix of mortar-crack coexistence defects for a single particle by performing surface defect detection on the multispectral image includes the following steps: Feature bands corresponding to crack prominence, mortar differentiation, and matrix verification are selected from multispectral images. After illumination normalization correction, aggregate region masks are extracted. Adaptive multi-scale disk structural elements are generated based on the equivalent particle size of single aggregate particles. Morphological operations guided by channel-space dual attention are used to separate attached mortar and microcracks, and a mortar-crack co-existence defect density matrix is ​​constructed.

[0006] Optionally, the step of performing three-dimensional grain shape characterization on the three-dimensional point cloud to obtain the standardized grain shape spectral vector of a single particle includes the following steps: Background removal, noise reduction, downsampling, and individual particle segmentation are performed on the 3D point cloud to obtain the 3D contour point cloud of a single aggregate particle; the 3D contour point cloud is then subjected to pose alignment, minimum bounding sphere size normalization, and conformal mapping to obtain a unit spherical point set; the unit spherical point set is then subjected to spherical harmonic decomposition, and multi-scale particle shape features are extracted and weighted to generate a standard particle shape spectral vector.

[0007] Optionally, the step of combining the co-occurrence defect density matrix and the standard particle shape spectrum vector to obtain the quality classification membership vector of the current aggregate group through graph attention coupling reasoning includes the following steps: Based on the defect statistical characteristics and standard particle shape spectrum vector of a single particle, the quality feature vector of a single particle is obtained and used as the attribute of a graph node; an undirected weighted adjacency graph is constructed based on particle spatial distance and feature similarity, and neighborhood information is aggregated through two layers of multi-head graph attention; the global quality features of the population are obtained through particle size weighted pooling, and the quality grading membership vector is obtained using a fuzzy membership classifier.

[0008] Optionally, the construction of the undirected weighted adjacency graph based on particle spatial distance and feature similarity includes the following steps: The K=6 nearest neighbor rule is used to construct the edges, and the edge weights are obtained by fusing the spatial Gaussian kernel weights and the feature cosine similarity in a 6:4 ratio. The fuzzy membership classifier is preset with four quality levels: excellent, first-class, qualified, and unqualified, and the sum of the components of the membership vector after normalization is 1.

[0009] Optionally, the step of obtaining the initial batching ratio vector by performing constrained interval fuzzy robust optimization based on the quality interval parameters of the stored aggregate includes the following steps: Based on historical storage data and real-time quality grading membership, a storage quality-reservoir database is constructed, including the content of attached mortar, the content of needle-like and flaky particles, and the range of crushing index. Rigid performance requirements are set as hard constraints, and elastic performance requirements are set as fuzzy constraints with trapezoidal membership. The most unfavorable quality condition is selected by the cut-set method, and the initial batching ratio vector is obtained by solving the constrained particle swarm algorithm.

[0010] Optionally, the step of performing spatiotemporal texture manifold identification on the mixed image sequence to obtain the homogeneity manifold coordinate vector of the mixture includes the following steps: Multi-scale gray-level co-occurrence matrix features, local binary pattern features, and temporal dynamic features are extracted from the mixing image sequence and fused to generate high-dimensional spatiotemporal texture features. The high-dimensional spatiotemporal texture features are mapped to a three-dimensional manifold space using a local tangent space alignment algorithm to generate a homogeneous manifold coordinate vector of the mixture containing the current homogeneity state value, the change trend value, and the target deviation.

[0011] Optionally, the step of dynamically correcting the initial ingredient ratio vector using the homogeneity manifold coordinate vector of the mixture to obtain a homogeneous guided adaptive ingredient vector includes the following steps: A Gaussian process response surface model is trained based on historical ingredient ratios and homogeneity samples. When the overall homogeneity deviation is greater than 0.15 and the trend of change is ≤0, a correction is triggered. Within the safe neighborhood of the initial ratio, the optimal ratio correction is searched through Bayesian optimization, and the homogeneity-oriented adaptive ingredient vector is obtained after constraint verification.

[0012] Optionally, the process of fusing the quality grading membership vector and the homogeneous guided adaptation batching vector, and obtaining the final batching execution scheme after Bayesian drift compensation correction, includes the following steps: A fusion feature vector is obtained by combining the quality grading membership vector, the homogeneous-oriented adaptive batching vector, and environmental operating parameters. A multilayer perceptron with quality-batch cross-attention is used to obtain the preliminary batching decision result. Based on the deviation between the factory slump and the measured strength value, a Bayesian online update is triggered. After compensating for system drift, the final batching execution scheme is obtained through full constraint verification.

[0013] Secondly, to efficiently execute the online quality detection and dynamic grading and batching method for recycled aggregates provided by this invention, this invention also provides an online quality detection and dynamic grading and batching system for recycled aggregates, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory stores program instructions for the online quality detection and dynamic grading and batching method for recycled aggregates. The online quality detection and dynamic grading and batching system for recycled aggregates of this invention has a compact structure and stable performance, and can stably execute the online quality detection and dynamic grading and batching method for recycled aggregates provided by this invention, further enhancing the overall applicability and practical application capability of this invention.

[0014] Beneficial effects: This invention integrates multispectral three-band images and three-dimensional laser point cloud dual-modal data, quantifies mortar-crack co-existing defects through an adaptive multi-scale morphological attention residual network, constructs a graph topology structure for single aggregate particles, establishes weighted edges by integrating spatial adjacency and quality similarity, and achieves nonlinear coupling evaluation of inter-particle quality through a multi-head graph attention network, accurately characterizing the overall quality fluctuation of the aggregate group and providing a quantitative basis for refined and continuous adjustment of batching; it uses Matern combined kernel Gaussian process regression to establish a nonlinear probability mapping between proportion and mixing homogeneity, and combines Bayesian optimization to quickly search for the optimal correction amount, using real-time manifold homogeneity coordinates as closed-loop feedback to dynamically and adaptively adjust the aggregate blending ratio of each compartment, significantly improving the utilization rate of recycled aggregates and reducing the dispersion of concrete performance. Attached Figure Description

[0015] Figure 1 This is a flowchart of an online quality detection and dynamic grading and batching method for recycled aggregates provided in an embodiment of the present invention; Figure 2 This is a framework diagram of an online quality detection and dynamic grading and batching system for recycled aggregates provided in an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 This invention provides a method for online quality detection and dynamic grading and batching of recycled aggregates, comprising the following steps: S1. Surface defect detection is performed on the multispectral image to obtain the density matrix of the mortar-crack coexistence defect of a single particle, and the three-dimensional particle shape is characterized by three-dimensional point cloud to obtain the standard particle shape spectral vector of a single particle.

[0019] In this embodiment, the original images in the 500-1000nm band acquired by a linear array multispectral camera deployed directly above the belt conveyor are used as input. Based on the spectral response characteristics of the recycled aggregate components, the band selection and precise segmentation of the target area are completed.

[0020] Specifically, based on the optical properties of the core components of recycled aggregate, three characteristic bands—550nm green light, 660nm red light, and 850nm near-infrared—were selected to form a complementary feature group. The 550nm green light band provides the highest contrast between light and dark areas of surface micro-depressions and micro-cracks, clearly highlighting crack boundaries. The 660nm red light band shows the largest difference in reflectivity between the virgin aggregate matrix and the attached cement mortar, making it the core band for distinguishing mortar areas. The 850nm near-infrared band can penetrate the thin slurry layer on the aggregate surface, reflecting the true texture of the matrix and used to remove false defects caused by surface dust and water stains.

[0021] Then, illumination normalization correction is performed. Using a standard polytetrafluoroethylene white board fixedly installed within the camera's field of view as the reflectivity benchmark, the illumination gain coefficient of each band is calculated. The original image is then subjected to pixel-by-pixel reflectivity normalization to offset the grayscale deviation caused by ambient light fluctuations and camera dark current drift during belt operation, and all pixel values ​​are uniformly mapped to the relative reflectivity range of 0 to 1.

[0022] Next, an aggregate region mask is generated, and a ratio calculation is performed on the red and near-infrared bands to enhance the grayscale difference between the aggregate and the belt background. The Otsu adaptive thresholding method is used to complete the initial binary segmentation, and then a morphological opening operation is performed through a disk structuring element with a radius of 2 pixels to remove noise spots caused by surface dust and small debris. The 8-neighbor contour tracking algorithm is used to extract the contours of all connected regions, and connected regions with an area smaller than the pixel area corresponding to the equivalent particle size of 5mm are screened out to obtain a pixel-level mask for each single aggregate particle. In this embodiment, only the effective pixels within the mask are retained for subsequent defect calculation, avoiding invalid areas such as belt background and frame shadow.

[0023] Furthermore, using a single aggregate mask as input, morphological calculation scales are dynamically matched based on the actual physical dimensions of the aggregate.

[0024] For each aggregate mask, the total area of ​​its connected pixels is calculated, and then converted into the actual physical area by combining the physical resolution of the pixels calibrated by the camera (e.g., 0.1 mm / pixel). The equivalent particle size of a single particle is then calculated according to the formula for the equivalent circle diameter, that is, the particle size value is equal to twice the square root of (area divided by pi). This is used as the benchmark parameter for scale matching.

[0025] Then, a disk-shaped structural element is used to match the isotropic characteristics of aggregate surface defects. Based on three fixed proportional coefficients of 2%, 5%, and 10% of the equivalent particle size, three sets of structural elements of fine, medium, and coarse scales are generated, corresponding to the characteristic scales of three typical defects: microcracks, fine cracks, and mortar adhesion patches, respectively. The embodiment also sets scale boundary constraints: the minimum structural element radius is not less than 1 pixel, and the maximum does not exceed 15% of the equivalent particle size, to avoid feature distortion caused by scale overflow.

[0026] A grading mapping table was established for the particle size range of commonly used recycled aggregates in engineering. Fine aggregates with an equivalent particle size of 5-20 mm correspond to a structural element physical scale of 0.1-2 mm; medium aggregates with an equivalent particle size of 20-40 mm correspond to 0.4-4 mm; and coarse aggregates with an equivalent particle size of 40-80 mm correspond to 1-8 mm. Based on the automatic placement of a single particle into the corresponding range, three sets of disk structural element parameters for that particle were obtained. This ensures that large aggregates will not miss macroscopic cracks due to excessively small structural elements, and that small aggregates will not misjudge normal surface textures as defects due to excessively large structural elements.

[0027] Then, using the single-particle multispectral feature map and the corresponding adaptive multiscale structural element as input, morphological operations guided by channel-space dual attention are performed to accurately separate two types of defects: attached mortar and microcracks, including: For the effective aggregate region images of the three characteristic bands, top-cap and bottom-cap transformations were performed using three sets of structuring elements. The top-cap transformation extracted a candidate set of bright regions with brightness higher than the surrounding matrix, corresponding to light-colored attached cement mortar; the bottom-cap transformation extracted a candidate set of dark regions with brightness lower than the surrounding matrix, corresponding to concave microcracks. The candidate regions of the same type in the three bands were extracted and fused at the pixel level to initially obtain mortar candidate images and crack candidate images, covering all defect responses at multiple scales.

[0028] Then, a dual-channel attention branch was constructed, with channel weights set for the mortar and crack targets respectively. The mortar branch assigned the initial maximum weight of 0.6 to the red light channel, with the near-infrared and green light channels used as auxiliary verifications; the crack branch assigned the initial maximum weight of 0.6 to the green light channel, with the red light and near-infrared channels used as auxiliary verifications.

[0029] Next, an edge distance weight map is generated based on the aggregate mask, with pixels closer to the aggregate edge having higher weights. The edge region is a high-incidence area for mortar accumulation and microcrack initiation during the crushing process of recycled aggregate. The basic weight of the central region is retained to avoid missing internal defects of the aggregate. The generated spatial attention mask is multiplied pixel by pixel with the channel-weighted candidate map to further highlight high-probability defect areas and suppress the response of non-defect areas.

[0030] In this embodiment, the original aggregate image is also subjected to 5×5 Gaussian blur to obtain a matrix texture reference image that retains only the macroscopic texture of the matrix and filters out small defects. The attention-enhanced candidate image is compared with the reference image to perform residual operation, which cancels the pseudo-defect response caused by the roughness and uneven color of the aggregate surface itself. Then, the residual image is binarized by local adaptive thresholding to obtain a pixel-level precision binary segmentation image of the attached mortar and a binary segmentation image of the microcrack.

[0031] Based on this, using the binary map of single-particle attached mortar and the binary map of microcracks as input, a coexisting defect density matrix with spatial distribution information and defect coupling characteristics is constructed through equal physical size gridded statistics and coexisting defect weighting.

[0032] Specifically, the smallest bounding rectangle of the aggregate mask is used as the boundary, and grid cells are divided according to a fixed physical size (e.g., 0.8mm × 0.8mm). The number of grid cells is dynamically adjusted according to the aggregate particle size, but is ultimately uniformly normalized to a standard dimension of 16 × 16. Incomplete grid cells at the edges are checked for the effective pixel ratio; grid cells with less than 50% effective pixels are not included in the statistics, ensuring that the physical scale of all statistical cells is consistent and eliminating statistical bias caused by particle size differences.

[0033] Then, for each effective grid cell, the following core indicators are calculated: First, the mortar area ratio, which is the ratio of the number of mortar pixels in the grid to the number of effective pixels in the grid, reflecting the degree of mortar adhesion in the area; Second, the relative length of the crack, which is obtained by first extracting the skeleton of the crack binary image to obtain the center line of the crack of a single pixel, counting the number of crack skeleton pixels in the grid, and then using the ratio of the number of pixels of the grid diagonal as the relative length, reflecting the degree of crack development in the area.

[0034] By traversing all grid cells, when the mortar area ratio in the same grid is ≥5% and the relative length of the crack is ≥3%, it is determined to be a mortar-crack coexisting grid. The defect density value of this type of grid is multiplied by a coexisting strengthening coefficient of 1.5 based on the weighted sum of the two indicators. In engineering, the mortar in the coexisting area will aggravate crack propagation and the crack will accelerate mortar spalling. The deterioration effect on the mechanical properties of aggregate is much greater than that of a single defect. Therefore, its contribution to quality is amplified by weighting.

[0035] The defect density values ​​of all grids are normalized and mapped to the 0~1 range. The final output is a mortar-crack coexisting defect density matrix with a fixed dimension of 16×16. The matrix not only preserves the spatial distribution of defects on the aggregate surface, but also characterizes the degree of coexistence and coupling of the two types of defects.

[0036] In another embodiment, the original 3D point cloud output by a line laser 3D scanner deployed directly above the belt conveyor is used as input. Background removal, noise filtering, and individual particle clustering are then used to obtain a set of 3D contour points for each individual aggregate particle, corresponding one-to-one with the multispectral particles. The original point cloud contains four-dimensional information including spatial coordinates (x, y, z) and laser reflection intensity. The coordinate system has the scanner's optical center as the origin, the belt running direction as the x-axis, the lateral direction as the y-axis, and the height direction as the z-axis.

[0037] First, the Random Sampling Consensus (RANSAC) algorithm is used to fit the belt reference plane. The distance threshold between points in the plane is set to 0.5 mm, and the number of iterations is 100. The z-axis reference plane where the belt is located is obtained by fitting. All points with z coordinates 0.2 mm below the reference plane and all points in the plane are marked as belt background and removed. Only the foreground aggregate point set with z coordinates 0.2 mm above the reference plane is retained to eliminate the interference of the conveyor background on the aggregate profile.

[0038] To address the scattered outlier noise points generated by aggregate surface dust and laser diffuse reflection, a statistical outlier removal algorithm is adopted: for each foreground point, the average Euclidean distance to its 15 nearest neighbors is calculated, the global mean μ and standard deviation σ of the average distance of all points are calculated, and outliers with an average distance greater than μ+2σ are removed. This parameter configuration can remove more than 95% of surface noise while preserving the details of aggregate edges and corners, and avoid pseudo-contours interfering with the subsequent mapping accuracy.

[0039] To balance contour accuracy and computational efficiency, voxel grid downsampling was used to sparsify the foreground point set. The voxel side length was set to 0.3 mm, and the centroid point within each voxel was retained as the sampling output. After downsampling, the point cloud density was reduced to 1 / 9 of the original, but the macroscopic contour and mesoscale angular features of the aggregate were still completely preserved, meeting the accuracy requirements of particle shape characterization.

[0040] For aggregate groups with a single layer of paving and a small number of adjacent aggregates on the conveyor belt, Euclidean clustering algorithm is used to achieve single-particle separation: the cluster neighborhood distance threshold is set to 2mm (greater than the surface point spacing and less than the minimum physical gap between aggregates), the minimum number of cluster points threshold is 200 (the minimum number of points corresponding to an equivalent aggregate size of 5mm), and the maximum number of cluster points threshold is 30000 (the maximum number of points corresponding to an equivalent aggregate size of 80mm). After clustering, each point cloud cluster corresponds to an independent aggregate, and the planar centroid coordinates (x,y) of each aggregate are obtained.

[0041] Furthermore, using the three-dimensional contour point cloud cluster of a single aggregate particle as input, aggregate point clouds of different particle sizes and placement orientations are mapped onto a unit spherical space. While preserving local shape and angular relationships, size and orientation interference are eliminated to obtain a standardized spherical mapping point set, including: The centroid is obtained by calculating the arithmetic mean of the coordinates of all points in the point cloud cluster. The point cloud is then translated so that the centroid coincides with the origin of the coordinate system. Principal component analysis (PCA) is then used to solve for the three principal axes of inertia of the point cloud. The point cloud is then rotated and transformed so that the major axis is aligned along the x-axis, the median axis along the y-axis, and the minor axis along the z-axis.

[0042] Then, the Welzl minimum bounding sphere algorithm is used to solve for the minimum bounding sphere of the point cloud after attitude alignment, and the coordinates of the sphere center and the radius R of the bounding sphere are obtained. The three-dimensional coordinates of all points in the point cloud are uniformly divided by R, so that the aggregate point cloud is constrained within a unit sphere with a radius of 1.

[0043] Then, Poisson surface reconstruction is performed on the normalized aggregate point cloud to generate a closed triangular mesh model. Then, a conformal spherical parameterization algorithm based on harmonic mapping is used to map all vertices of the triangular mesh onto the unit sphere.

[0044] During the mapping process, the positive endpoint of the major axis is fixed as the equatorial reference point and the centroid projection is the pole. Conformal mapping is achieved by minimizing the angular distortion of the mesh triangle, ensuring that the sharpness of the aggregate surface edges and the curvature of the surface maintain geometric similarity after mapping, and avoiding the shape distortion of the edge region caused by conventional radial projection.

[0045] To address the issue of uneven density of spherical points after mapping (dense in angular areas and sparse in flat areas), spherical Fibonacci sampling is used to generate 1024 uniformly distributed spherical sampling points. The radial height deviation of the original mapping points is assigned to the sampling points through neighbor interpolation, resulting in a fixed number of uniformly distributed unit spherical mapping points. The radial height deviation value carried by each sampling point corresponds to the amount of unevenness of the original aggregate surface relative to the standard sphere.

[0046] Next, using a uniform set of sampling points on a unit sphere and its radial height deviation as input, the spatial domain shape features are converted into frequency domain spectral features through spherical harmonic decomposition. Grain shape attributes are then assigned to different scales according to their order, generating a multi-scale grain shape spectral feature sequence, including: Construct a spherical scalar function, taking the radial height deviation at each sampling point on a unit sphere as a scalar value, and then construct a height deviation function defined across the entire unit sphere. ,in It is the polar angle (0~π). The azimuth angle is 0~2π. This function fully describes the shape undulation of the aggregate surface relative to a standard sphere and is the input object for spherical harmonic decomposition.

[0047] Orthogonal basis pairs of spherical harmonic functions Perform a series expansion, and set the decomposition order range to... to The spherical harmonic coefficients of each order are calculated using spherical numerical integration. ,in For harmonic orders, Order (-l≤m≤l). It is negatively correlated with shape scale: the lower the order, the larger the corresponding shape scale; the higher the order, the smaller the corresponding shape scale.

[0048] Then, a clear correspondence is established between the order interval and the engineering particle shape attribute: Order: Zero-order DC component, corresponding to the global mean of height deviation, does not carry shape difference information, and is removed during subsequent feature construction; The lower order corresponds to the large-scale overall deformation of the aggregate. The coefficient amplitude reflects the sphericity, flatness and slenderness ratio. The anisotropy of the l=2 order coefficient directly corresponds to the degree of needle-like and flaky texture. The intermediate level (intermediate level) corresponds to the surface undulation of aggregate at the medium scale. The coefficient energy reflects the number of edges and corners, the sharpness of the edges and corners and the distribution of broken surfaces. It is the core characterization range of the engineering angularity index. Rank (higher order): corresponds to the surface texture of aggregate at the microscale. The coefficient energy reflects the surface roughness and tiny pits and protrusions, characterizing the micromorphological features of the fractured surface.

[0049] For each order Calculate all of this order The sum of squares of the coefficients yields the shape energy of that order. right to The energy sequence of each order is normalized using the L2 norm to obtain the relative energy percentage sequence of each order.

[0050] Finally, using the multi-scale particle shape spectrum feature sequence as input, and combining it with engineering application weights for weighted adjustment, a standardized particle shape spectrum vector is constructed.

[0051] Specifically, based on the influence weight of recycled aggregate concrete performance, differentiated weighting coefficients are first set for four order intervals: the needle-like and flaky interval (… The weight of 0.35 has the greatest impact on the workability and strength of concrete; the sphericity range ( A weight of 0.3 determines the aggregate porosity; angularity range ( The weight is 0.25, affecting interfacial bonding and aggregate interlocking force; the micro-roughness range ( A weight of 0.1 has the weakest impact on macroscopic performance. The weighting coefficient is obtained through experimental calibration and can be dynamically adjusted according to the target concrete strength grade.

[0052] Then, the normalized energy value of each order is multiplied by the weight coefficient of its interval to strengthen the feature dimension corresponding to the key engineering indicators and weaken the interference of minor micro-textures, so that the feature space is more in line with the needs of engineering quality evaluation.

[0053] Next, the weighted fine-grained spectral features are aggregated by interval to generate macro-grain shape factors (sphericity factor, needle-like factor, angularity factor, roughness factor); then the macro-factors are concatenated with the fine-grained spectral features to form a 36-dimensional standardized grain shape spectral vector.

[0054] In the embodiments, Z-score standardization can also be performed on the standardized particle shape spectrum vector based on the characteristic mean and standard deviation of the historical aggregate sample library, so as to unify the values ​​of each dimension to a distribution space with zero mean and unit variance.

[0055] S2. Combining the co-existing defect density matrix and the standard particle shape spectrum vector, the quality grading membership vector of the current aggregate group is obtained through graph attention coupling reasoning.

[0056] In this embodiment, the global physical coordinate system of the belt conveyor is first used as a reference. The aggregate centroid pixel coordinates extracted from the multispectral image are calibrated and converted into physical coordinates. Hungarian matching is then performed with the aggregate centroid coordinates obtained by three-dimensional point cloud clustering. The matching distance threshold is set to 3mm, and a unique ID index is generated for each aggregate.

[0057] Then, for the co-occurrence defect density matrix of each particle, the following statistical features are extracted: ① Global scale features: global average mortar proportion, global average relative length of cracks, and global mean co-occurrence defect density, which characterize the overall defect level of the particle; ② Distribution morphology features: standard deviation, skewness, and kurtosis of defect density, which characterize the dispersion and aggregation tendency of defects on the particle surface; ③ Local extreme value features: maximum mortar proportion of the unit, maximum crack length of the unit, peak density of co-occurrence units, proportion of co-occurrence units in the total effective mesh, proportion of high defect density units (mesh proportion with density > 0.7), and mean defect density of the edge region, which reflect the quality impact of the worst local defects and the high incidence of defects in the edge areas.

[0058] Then, Z-score standardization is performed on the defect statistical features and the standard particle shape spectral vector respectively to eliminate the weight imbalance caused by the difference in dimensions. Then, the two are spliced ​​together in the order of defect features first and particle shape features last to form a single particle multidimensional quality feature vector.

[0059] Furthermore, an undirected weighted adjacency graph is constructed using the initial multi-dimensional quality feature vectors of all particles and the centroid plane coordinates of each particle in the global coordinate system of the belt. Specifically: Graph node definition: Each valid independent aggregate in the current detection field is defined as a graph node. The total number of nodes N is equal to the number of valid aggregate particles in the field. The attribute vector of the node is the corresponding initial multi-dimensional quality feature vector.

[0060] K-Nearest Neighbor Edge Generation: Physical edges between nodes are constructed using the K-nearest neighbor rule, with K=6 fixed. This means that each node establishes undirected edges only with its six nearest neighbors in terms of spatial Euclidean distance; no edges are established between particles outside the K-nearest neighbor range. Furthermore, a maximum adjacency distance threshold of 50mm is set; neighboring nodes exceeding this distance are also removed from the connection, eliminating invalid associations due to excessive spatial distance.

[0061] Bilateral weight fusion calculation: The weight of each edge is obtained by weighting spatial adjacency weight and quality feature similarity weight in a 6:4 ratio. Spatial adjacency weight is calculated using a Gaussian kernel function; the closer the particle spacing, the higher the weight. When the spacing is 0, the weight is 1, and the weight approaches 0 when the maximum adjacency distance is reached. Quality feature similarity weight is calculated using cosine similarity to determine the similarity between the quality feature vectors of two nodes. The closer the features, the higher the weight; when they are completely identical, the weight is 1. The final edge weight ranges from 0 to 1. A higher weight indicates that the two aggregates are spatially closer and have more similar quality characteristics.

[0062] Adjacency graph structured output: Based on the node set, edge set and corresponding weights, an undirected weighted adjacency graph of the current aggregate group is constructed and stored in the form of a sparse adjacency matrix. This graph simultaneously encodes the spatial distribution structure and quality similarity relationship of the aggregate group.

[0063] Based on this, using an undirected weighted adjacency graph of N nodes and the initial quality feature vector of the nodes as input, a multi-head graph attention mechanism is used to achieve nonlinear coupling and aggregation of neighborhood quality information, thereby obtaining a single-particle comprehensive quality feature that integrates the group association effect.

[0064] Specifically, the initial features of all nodes are first mapped to a 64-dimensional hidden feature space through a shared linear transformation matrix. Then, four independent attention branches are used to split the hidden features into channels, each corresponding to a 16-dimensional feature subspace. Attention aggregation is performed in parallel to capture the particle quality correlation patterns under different feature subspaces.

[0065] For each attention head, the attention coefficient between each node and its neighboring nodes is calculated. The calculation process incorporates a two-dimensional coupling mechanism: ① Intra-feature coupling: The defect feature subset and grain shape feature subset of the node itself are cross-productted to generate a quality coupling term, which represents the superimposed degradation effect of high defects and poor grain shape, avoiding the underestimation problem of independent evaluation of the two types of features; ② Inter-node coupling: The coupling term of neighboring nodes is weighted by combining the spatial-feature fusion weight of the adjacent edges. The final attention coefficient is jointly determined by the node's own quality coupling strength, the feature similarity between nodes, and the spatial adjacency weight. After softmax normalization, the output is ensured that the sum of the attention coefficients of all neighboring nodes of each node is 1.

[0066] Then, a two-layer graph attention layer is stacked to achieve multi-order neighborhood information aggregation. The first graph attention layer aggregates the quality information of first-order nearest neighbor particles, outputting 64-dimensional node features. The second layer, based on the output of the first layer, aggregates the quality information of second-order neighbors, while introducing residual connections to add the initial mapped features to the aggregated features of the second layer element-wise, avoiding feature degradation in deep networks. After two layers of propagation, the features of each node not only include its own quality attributes but also incorporate the quality influence of surrounding neighboring particles, reflecting the local quality aggregation effect of the aggregate group.

[0067] The output features of the four attention heads are concatenated to obtain a final 64-dimensional single-particle comprehensive quality feature vector. Using the 64-dimensional comprehensive quality feature vectors of all nodes as input, global pooling aggregation and fuzzy membership inference are performed to obtain the soft-grading membership vector of the aggregate group, including: A particle size-weighted global average pooling is performed on the comprehensive quality feature vector of all nodes. That is, the feature vector of each node is weighted by the square of the equivalent particle size of a single particle (corresponding to the quality weight), and then the arithmetic mean is taken. The N 64-dimensional vectors are aggregated into a 64-dimensional aggregate group global quality feature vector. After weighting, the proportion of large-particle aggregates in the group evaluation is higher, which is consistent with the influence of aggregate quality proportion on concrete performance in actual engineering.

[0068] A 64-dimensional global quality feature vector is input into a pre-trained Gaussian fuzzy membership classifier. The classifier has four preset quality levels: excellent, first-class, qualified, and unqualified. Each level corresponds to a set of pre-labeled feature center vectors and Gaussian width parameters. These parameters are obtained by training historical labeled samples (including laboratory-measured crushing values, mortar content, particle shape index, etc.) through fuzzy C-means clustering. The membership degree of the global feature vector to each of the four levels is calculated. A higher membership degree indicates a higher degree of conformity of the batch of aggregates to the corresponding level. In this embodiment, the membership degree values ​​of the four levels are also normalized so that the sum of the four values ​​is 1. The values ​​are then arranged in the order of excellent, first-class, qualified, and unqualified to form the aggregate group quality grading membership vector.

[0069] S3. Based on the quality range parameters of the stored aggregate, constrained interval fuzzy robust optimization is performed to obtain the initial batching ratio vector. Spatiotemporal texture manifold identification is performed on the mixing image sequence to obtain the homogeneity manifold coordinate vector of the mixture.

[0070] In one embodiment, a camera installed on the side observation window of the forced mixer acquires continuous grayscale images at a frame rate of 25fps. The field of view covers one-third of the main material tumbling area inside the mixing cylinder. Using two fixing bolt holes on the inner wall of the mixing cylinder as feature anchor points, affine transformation is used to complete image registration frame by frame to offset pixel offset caused by camera loosening and vibration. A fixed rectangular detection area is delineated in the registered image, avoiding the fixed motion trajectory of the mixing shaft and mixing blades, and only retaining the central effective area where the material is most fully tumbling.

[0071] In this embodiment, for the working conditions of dust filling the mixing tank and uneven lighting, homomorphic filtering is used to compress the low-frequency lighting component of the image and enhance the high-frequency texture component to eliminate the interference of material pile shadows and light spots; then, 5×5 Gaussian filtering is used to suppress the high-frequency noise caused by surface dust particles, retain the macroscopic texture boundary of mortar and aggregate, and avoid dust pseudo-texture interference with subsequent feature extraction.

[0072] Next, a Gaussian mixture model was used to model the fixed background of the cylinder wall. During the training phase, 200 frames of images of the empty cylinder and the unloaded stirring were collected to update the background model. Three Gaussian components were set and the learning rate was 0.005. During online detection, background subtraction was performed frame by frame. Regions with a pixel-level foreground probability greater than 0.7 were identified as effective regions of the mixture, and a material mask was generated. Then, fixed background pixels such as the cylinder wall and residual clumps were removed, and the real material surface was retained to participate in texture calculation.

[0073] Then, using the synchronous trigger signal of the mixer's main shaft encoder as the time reference, for each complete cycle pulse received (corresponding to a 2-second cycle at a forced mixer speed of 30 rpm, totaling 50 frames), 50 consecutive frames of the detection area image within that cycle are captured, forming a set of spatiotemporal image block sequences. The sequence is updated using a sliding window mechanism with a sliding step size of 10 frames to ensure the temporal resolution and real-time performance of state identification. Full-cycle capture ensures that the features include the complete material tumbling cycle, avoiding state deviations caused by half-cycle sampling.

[0074] Furthermore, using a single set of 50 spatiotemporal image block sequences as input, multi-scale static textures are extracted from the spatial dimension and dynamic difference features are extracted from the temporal dimension. After fusion, a high-dimensional spatiotemporal texture feature set is formed to characterize the mortar distribution, aggregate exposure, and mixing dynamic characteristics.

[0075] Specifically, multi-scale gray-level co-occurrence matrix (GLCM) spatial features are first extracted. For each frame of the sequence, three step sizes (1 pixel, 3 pixels, and 5 pixels) are used, corresponding to the micro, meso, and macro scales, respectively. The GLCM is calculated along four directions (0°, 45°, 90°, and 135°). For each matrix, four statistical measures are extracted: energy, contrast, correlation, and inverse moment. Energy reflects the uniformity of texture, and contrast reflects the roughness of surface. A total of 48-dimensional GLCM features are output across the three scales and four directions. High energy and low contrast correspond to sufficient mortar coverage and a smooth surface, while low energy and high contrast correspond to more exposed aggregate and greater surface unevenness.

[0076] Then, the uniform local binary pattern texture distribution was extracted. For each frame, uniform LBP operators with radii of 1, 2, and 3 were applied, corresponding to three-scale local texture patterns. LBP histograms were calculated at each scale, with each histogram containing 59 uniform pattern dimensions, resulting in a total of 177 LBP features across the three scales. The concentration of the histogram distribution corresponds to the uniformity of the material components: a more concentrated distribution indicates a more uniform mixing of mortar and aggregates; a more dispersed distribution indicates the presence of local aggregate agglomeration or mortar enrichment.

[0077] Next, the temporal dynamic difference features are calculated. The cosine distance between the feature vectors of adjacent frames is calculated frame by frame along the time dimension for the GLCM feature sequence of 50 frames of images, resulting in a 49-dimensional inter-frame texture difference sequence. The mean, variance, peak value and periodic fluctuation coefficient are extracted from this sequence. At the same time, the temporal intersection-union ratio mean of the LBP histogram is calculated, resulting in a total of 12-dimensional temporal dynamic features. The smaller the temporal variance and the lower the fluctuation coefficient, the more stable the mixing process and the more uniform the material state. Conversely, it indicates the existence of dynamic non-uniform phenomena such as agglomeration and segregation.

[0078] The 48-dimensional GLCM features and 177-dimensional LBP features of a single frame are concatenated with the 12-dimensional temporal features of the entire sequence to form a 237-dimensional single-sample high-dimensional spatiotemporal texture feature vector. Z-score standardization is then performed on all feature dimensions. The standardization parameters are obtained from the historical sample database to eliminate the weight interference of different feature magnitudes on the subsequent manifold embedding weights.

[0079] Based on this, using 237-dimensional high-dimensional spatiotemporal texture features as input, the high-dimensional features are mapped to a three-dimensional manifold space with explicit physical semantics through local tangent space alignment, obtaining standardized three-dimensional manifold coordinates, including: Labeled samples with different mix proportions, mixing times, and aggregate qualities were pre-collected. Each sample group corresponded to a set of 237-dimensional texture features and three manually calibrated state scores (mortar encapsulation, aggregate dispersion uniformity, and mixing dynamic stability), with a score range of 0 to 10. Local Tangent Space Alignment (LTSA) was selected as the manifold embedding algorithm. By fitting the local tangent space of each sample and then globally aligning it, the continuous nonlinear mixing state manifold was restored while maintaining the local geometric structure.

[0080] Then, the nearest neighbor number k=12 is set to ensure that the neighborhood of each sample covers the local continuous state change range, and the 237-dimensional features are mapped to the three-dimensional low-dimensional manifold space through LTSA.

[0081] After mapping, semantic dimension alignment is performed using labeled samples: the first dimension coordinate is linearly correlated with the mortar encapsulation score, and the higher the coordinate value, the more fully the mortar encapsulates the aggregate; the second dimension coordinate is linearly correlated with the aggregate dispersion uniformity score, and the higher the coordinate value, the more uniformly the aggregate is distributed in space and the less agglomeration occurs; the third dimension coordinate is linearly correlated with the mixing dynamic stability score, and the higher the coordinate value, the more stable the material agitation and the less significant segregation fluctuation occurs.

[0082] Furthermore, 10 parallel samples with standard proportions and optimal mixing time were selected as the reference sample set, and their three-dimensional coordinate mean was calculated as the reference origin of the manifold space. Before the start of each batch of production, the steady-state sample of the first 30 seconds of the batch was used to perform Procrustes analysis, and the manifold coordinates of the current batch were rigidly aligned to the reference coordinate system to eliminate global coordinate drift caused by fogging of the observation window, light aging, and slight dirt on the camera, so as to ensure that the coordinates of different time periods and different batches are comparable.

[0083] In another embodiment, during online detection, new samples are not retrained on the manifold. Instead, nearest neighbor interpolation mapping is performed based on the local tangent space of the existing trained manifold to project the new high-dimensional features onto the existing three-dimensional manifold coordinate system. The single sample mapping takes less than 10ms, which meets the requirements for online real-time performance.

[0084] Furthermore, based on the single-sample three-dimensional manifold coordinates, a three-dimensional homogeneous manifold coordinate vector containing the current state, trend of change, and target deviation is generated, including: Gaussian weighted temporal smoothing is performed on the continuously output 3D manifold coordinates using a sliding window of 5 frames. The weight of the center frame of the window is 0.4, which decreases frame by frame towards both sides to suppress coordinate jumps caused by material splashing and local reflections in a single frame.

[0085] The current homogeneity state value is calculated by weighting the smoothed three-dimensional coordinates according to engineering weights: mortar encapsulation weight 0.4, aggregate dispersion uniformity weight 0.4, and mixing dynamic stability weight 0.2. The weighted result is then normalized to the min-max range of 0-1 to obtain the current homogeneity state value. 1 corresponds to the theoretical optimal homogeneity state, and 0 corresponds to a completely discrete unmixed state. This value serves as the first component of the vector, representing the overall homogeneity level of the current mixing.

[0086] The trend component is calculated by performing a univariate linear regression on the homogeneity state value sequence of the past 10 frames (approximately 2 seconds). The regression slope is used as the trend value, with the unit being the change per second. A positive slope indicates that homogeneity is improving and the mixing is still developing towards uniformity; a negative slope indicates that homogeneity is decreasing and segregation or bleeding may occur; a slope approaching 0 indicates that the mixing has entered a steady state.

[0087] The target baseline deviation is calculated by taking the preset target homogeneity baseline coordinates as a reference, calculating the Mahalanobis distance between the current three-dimensional coordinates and the baseline point, and considering the variance differences and correlations of the three dimensions. The distance value is normalized to the range of 0 to 1 to obtain the target deviation degree. The larger the value, the farther the current state is from the target homogeneity state, and the greater the correction required.

[0088] The final generated homogeneity manifold coordinate vector of the mixture is arranged in the order of [current homogeneity state value, trend value, and target deviation].

[0089] S4. The initial ingredient ratio vector is dynamically corrected using the homogeneity manifold coordinate vector of the mixture to obtain a homogeneous guided adaptive ingredient vector.

[0090] In a preferred embodiment, based on the historical offline detection sample library of each aggregate bin, the real-time aggregate group quality grading membership vector at the discharge port, and the storage volume data collected by the level gauge, a storage quality-storage volume database containing uncertainties is constructed, including: First, a quantitative mapping relationship between front-end visual inspection features and conventional engineering quality parameters is established through laboratory sampling and comparison experiments: the global mean of the coexisting defect density matrix is ​​linearly regressed with the laboratory measured percentage of attached mortar mass to obtain the conversion coefficient from defect density to mortar content; the needle-like and flaky factors and sphericity factors in the standard particle shape spectrum vector are multivariately regressed with the needle-like and flaky particle content and standard particle shape index measured by sieving method; after fusing defects and particle shape features, a prediction model is established with the crushing value test results.

[0091] In this embodiment, for each independent aggregate bin, offline test samples and online visual inspection historical data from the past three months were retrieved. The mean and standard deviation of three core quality parameters—attached mortar content, flaky and needle-like particle content, and crushing index—were calculated. The upper and lower limits corresponding to the 95% confidence level were taken as the initial quality parameter range for that bin, with the upper limit corresponding to the most unfavorable quality condition and the lower limit corresponding to the optimal quality condition. At the same time, based on the sieving test results, the sieve residue of each bin was also represented as an interval number, covering the gradation fluctuation range of the aggregate within the bin.

[0092] During online production, for the current batch of aggregate, the quality grade membership degree is used as a weighting factor to calculate the expected quality parameters of the aggregate at that grade. Using this expected value as the center of the interval, the upper and lower limits of the interval are shifted synchronously based on historical fluctuations to achieve real-time updates to the quality interval. When the membership degree of non-conforming aggregate increases, the upper limit of the interval shifts upwards, amplifying the boundary of the most unfavorable operating conditions; when the membership degree of superior aggregate increases, the lower limit of the interval shifts downwards, appropriately reducing the degree of conservatism. The corrected interval retains the historical statistical fluctuation patterns while incorporating the real-time quality characteristics of the current batch.

[0093] Furthermore, a structured database is established based on the aggregate bin ID. Each record includes: particle size range, upper and lower limit ranges of three core quality parameters, sieve residue range for each sieve aperture, current available reserves, and reserve warning threshold.

[0094] Then, using concrete design technical requirements, storage quality-reservoir database, and pre-calibrated quality-performance influence coefficient matrix as inputs, rigid requirements and elastic performance requirements are transformed into hard constraints and fuzzy constraints respectively, while embedding the performance degradation effect of quality fluctuations to form a complete constraint system.

[0095] Specifically, the requirements that must be met unconditionally are first set as 0-1 type hard constraints, and violations are judged as infeasible solutions. These include: 1. Storage capacity constraint: the mass taken from a single storage bin must not exceed the current available storage capacity; 2. Total quantity constraint: the sum of the extraction ratios of all aggregate bins must be equal to 100%; 3. Minimum replacement rate constraint for recycled aggregate: the proportion of the total mass of recycled aggregate to the total mass of aggregate must not be lower than the minimum value required by the design; 4. Unit cost constraint: the unit cost of aggregate components must not exceed the set cost upper limit; 5. Gradation range constraint: the upper and lower limits of the cumulative sieve residue after mixing in proportion in each bin must all fall within the gradation range required by the specification, and the most unfavorable gradation combination must not exceed the boundary of the gradation zone.

[0096] Then, a fuzzy constraint membership function is constructed, and the performance requirement that allows a certain fluctuation range is set as a continuous fuzzy constraint. The degree of satisfaction is described by a trapezoidal membership function, and the membership value range is 0 to 1.

[0097] For strength indicators: the membership degree is 0 when below the design strength value, increases linearly from the design strength to the design strength plus surplus strength, and is 1 when above the surplus strength. For slump indicators: the membership degree is 0 when below the lower construction limit or above the upper construction limit, and is 1 within the optimal slump range, with linear changes in the upper and lower transition sections. For paste-aggregate ratio indicators: with the benchmark paste-aggregate ratio as the center, the membership degree is 1 within a reasonable fluctuation range, and decreases linearly beyond that range. The width of all transition sections is determined according to the engineering acceptance specifications and construction tolerance rate.

[0098] Next, the quality fluctuation influence coefficient is embedded. First, the quality-performance influence coefficient matrix is ​​calibrated through mix proportion tests to characterize the quantitative impact of the deterioration of a single quality parameter on concrete performance: for example, for every 1 percentage point increase in the content of adhering mortar, the strength reduction coefficient increases by a fixed amount, and the water demand per cubic meter increases by a fixed value; for every 1 percentage point increase in the content of needle-like and flaky particles, the slump decreases by a fixed value. During constraint verification, all quality parameters are substituted into the influence coefficient at the upper limit of the interval (the most unfavorable value) to calculate the predicted values ​​of strength, water demand, and slump under the most unfavorable working condition. These predicted values ​​are then used in the fuzzy constraint membership calculation to ensure that the solution is feasible throughout the entire quality fluctuation range.

[0099] Furthermore, all hard and fuzzy constraints are uniformly incorporated into the optimization constraint set, and a global minimum acceptable membership degree threshold is set, with a default value of 0.7. That is, the solution is considered valid and feasible only when the membership degree of all fuzzy constraints is not lower than 0.7. The threshold can be adjusted according to the importance level of the project. The threshold can be appropriately increased for important projects to improve robustness, while it can be appropriately decreased for ordinary projects to improve economy.

[0100] Based on this, using the selected optimization objective, the integrated constraint system, and the warehouse quality-storage database as input, the cut-set method is employed to handle fuzzy uncertainties. Optimization is sought under the most unfavorable quality conditions to obtain feasible solutions for the initial ingredient ratios that satisfy all constraints, including: Set optimization objectives and define decision variables. Decision variables are defined as the proportion of total aggregate mass taken from each aggregate bin, with the dimension consistent with the number of aggregate bins, and values ​​ranging from 0 to 1. The optimization objective supports switching between two engineering modes: Mode 1 aims to minimize the cost per cubic meter of aggregate, suitable for cost control scenarios; Mode 2 aims to maximize the comprehensive utilization rate of recycled aggregate, suitable for environmental indicator control scenarios. Alternatively, the two objectives can be merged into a single objective through weighted coefficients, with the weights set according to engineering requirements.

[0101] Then, a fuzzy cut-set method is used to perform a deterministic transformation of the constraints, selecting the parameter range with a membership degree not lower than the global threshold as the effective fluctuation range. Within the effective range, the most unfavorable combination of quality parameters is selected, i.e., all parameters that negatively affect performance are taken at the upper bound of the range, corresponding to the working condition with the lowest intensity, worst workability, and highest water demand. This most unfavorable combination is used as the constraint verification benchmark to ensure that the solved ratio meets the performance requirements throughout the entire quality fluctuation range, avoiding the problem of the mean meeting the standard but the inferior value exceeding the standard that occurs in conventional mean optimization.

[0102] Furthermore, a particle swarm optimization algorithm with embedded constraint penalty mechanism is used for solving the problem. The population size is set to 60, the maximum number of iterations is set to 400, the inertia weight decreases linearly with the number of iterations, and the cognitive factor and social factor are set to 1.5 and 1.7, respectively. The fitness function is composed of the objective function value and the constraint penalty term: individuals that violate hard constraints are subject to high penalties and are directly excluded from the feasible region; individuals with fuzzy constraint membership degrees below the threshold are subject to gradient penalties according to the degree of violation. During the iteration process, the optimal solution of the individual and the global optimal solution are updated generation by generation, and the optimal individual in the feasible region is retained.

[0103] After the iteration terminates, the proportion of each bin corresponding to the global optimal feasible solution is extracted, and normalization verification is performed to ensure that the sum of all proportions is exactly equal to 1. The specific mass consumption of each aggregate per cubic meter of concrete is then calculated to form the basic data of the initial batching scheme.

[0104] Then, based on the initial batching ratio, quality parameter range and concrete performance calculation model obtained from the solution, the initial batching ratio vector is obtained by verifying the performance under multiple working conditions, gradation compliance and storage feasibility.

[0105] Specifically, the upper, average, and lower limits of the quality parameter range are taken respectively, corresponding to the most unfavorable, normal, and optimal working conditions. They are then substituted into the empirical formula for concrete performance for calculation: the strength is calculated using the Balromy formula with embedded quality reduction coefficient, and the workability is calculated using the water demand method to predict the slump value. The strength range and slump range under the three working conditions are obtained, which intuitively presents the performance fluctuation range caused by quality fluctuation.

[0106] Based on the performance calculation results, two core criteria are determined: first, the strength guarantee rate, which requires that the probability of the predicted strength value being higher than the design strength is not less than 95%, corresponding to the strength still meeting the design requirements under the most unfavorable working conditions; second, the workability qualification rate, which requires that the probability of the slump falling within the allowable range for construction is not less than 90%.

[0107] It can also verify the synthetic gradation curve, confirm that the cumulative sieve residue of each sieve aperture falls within the standard gradation zone, and that the predicted porosity value meets the design requirements.

[0108] If the verification fails, adjustments are made step by step according to priority: first, the threshold of the fuzzy constraint is slightly reduced to expand the feasible region, and the optimization solution is re-executed; if there is still no feasible solution after the threshold is reduced to 0.5, the natural aggregate supplementation mechanism is triggered, and the upper limit of the amount of natural aggregate bins is appropriately relaxed to reduce the overall quality fluctuation range by adding natural aggregates; if it still cannot be satisfied, feedback is given to improve the quality level of recycled aggregate entering the bins, until all verification items pass. After the verification passes, the final initial batching ratio vector is obtained, and each component is the percentage of the corresponding bin aggregate in the total aggregate mass.

[0109] In another embodiment, a Gaussian process response surface model is initialized. Valid samples from the steady-state mixing stage are selected from historical batches, eliminating abnormal batches due to equipment failure, insufficient material quantity, or manual intervention. The initial training set sample size is no less than 120 groups. On the input side, a logarithmic ratio transformation of the component proportions of each bin is performed to adapt the proportion data to the Euclidean space assumption of the Gaussian process. On the output side, Z-score standardization is performed on the three-dimensional manifold coordinates. The standardization parameters are obtained statistically from historical samples, eliminating differences in dimensions and amplitudes across different dimensions.

[0110] A hybrid kernel function combining Matern5 / 2 kernel and linear kernel is used. The Matern5 / 2 kernel is used to capture the strong nonlinear relationship between the proportion of ingredients and homogeneity. Its differentiable order is adapted to the moderate smoothness of industrial processes and is more suitable for the non-smooth response caused by the quality fluctuation of recycled aggregates than the radial basis kernel. The linear kernel is used to capture the linear trend term caused by the overall proportion deviation, which improves the prediction stability of large proportion adjustment range.

[0111] The kernel function contains three types of hyperparameters: feature length scale, signal variance, and noise variance. The initial values ​​are solved on the training set through maximum likelihood estimation, which correspond to the sensitivity of the proportion of each aggregate bin to homogeneity, the overall fluctuation amplitude of the response, and the observation noise level, respectively.

[0112] The Gaussian process regression model is initialized and trained offline based on the preprocessed training set. The model outputs the predicted mean and predicted variance of the three-dimensional homogeneous coordinates. The variance value represents the confidence level of the prediction result under this ratio. The model has the ability to generalize with small samples. Even when the sample size is less than 60 groups in the early stage of production line commissioning, it can still output a statistically significant prediction interval, avoiding the situation where no model is available during the cold start stage.

[0113] In this embodiment, the initial model is an offline baseline model. Subsequently, every 10 valid production batches that have been finally verified are added to the sample set with the actual ingredient ratio and the corresponding steady-state homogeneity coordinates. Incremental hyperparameter fine-tuning is performed, and the update adopts a sparse Gaussian process approximation algorithm. The update time per round is less than 500ms, which does not block the online real-time correction link and realizes the continuous accuracy evolution of the model as production data accumulates.

[0114] Furthermore, based on parallel test samples under standard mix proportions and optimal mixing processes, the mean vector and covariance matrix of the three-dimensional manifold coordinates were statistically analyzed. A three-dimensional ellipsoidal target qualified domain was constructed with the mean vector as the target center point and ±0.5 times the standard deviation as the boundary. These three dimensions correspond to mortar encapsulation, aggregate dispersion uniformity, and mixing dynamic stability, respectively. The baseline domain range can be dynamically scaled according to concrete strength grade and construction slump requirements. For important projects, it can be tightened to ±0.3 times the standard deviation to improve homogeneity requirements.

[0115] Then, for the current 3D manifold coordinates after Gaussian smoothing, the difference between the current coordinates and the target center point is calculated dimension by dimension. Negative values ​​represent values ​​below the target level, and positive values ​​represent values ​​above the target level.

[0116] The deviation type is attributed based on the deviation ratio: if the first dimension, mortar coating deviation, is negative and contributes more than 60% of the total deviation, it is determined to be a mortar insufficient coating type deviation, and the correction direction is to increase the proportion of fine aggregate or finely adjust the paste content; if the second dimension, aggregate dispersion uniformity deviation, is negative and contributes more than 60% of the total deviation, it is determined to be an aggregate agglomeration type deviation, and the correction direction is to optimize the particle size distribution and improve the particle packing gradation; if the third dimension, dynamic stability deviation, is significant and the first two dimensions are normal, it is determined to be a mixing process type deviation, and the mixing speed and time should be adjusted first, with the mix proportion only used for auxiliary correction.

[0117] Then, the Mahalanobis distance between the current coordinates and the target center point is calculated. By integrating the correlation and variance differences of the three-dimensional coordinates, a normalized overall target deviation of 0 to 1 is obtained. Compared with Euclidean distance, Mahalanobis distance can avoid the large variance dimension dominating the deviation evaluation, and is more in line with the physical characteristics of homogeneous multi-dimensional coupling.

[0118] Furthermore, a dual triggering judgment is made by combining the homogeneity change trend component: when the overall deviation is greater than 0.15 and the change trend component is ≤0 (homogeneity does not show a self-convergence trend), the ingredient correction process is officially triggered; if the deviation exceeds the standard but the trend component is positive and the convergence rate is fast, the proportioning intervention is temporarily not executed, and the mixing self-steady state is waited for to be reached to avoid excessive adjustment causing oscillation.

[0119] Based on this, using the initial ingredient ratio vector, deviation discrimination results, and the initialized Gaussian process response surface model as input, the optimal ratio correction is searched through Bayesian optimization within the safe neighborhood of the initial ratio, achieving rapid convergence of homogeneity towards the target domain. Specifically, the incremental adjustment of the proportion of each aggregate bin is used as the decision variable, and a constrained optimization domain is constructed based on the initial mix design. The default upper limit for the proportion adjustment of a single bin is ±10% of the initial value, and the sum of the proportions of all aggregate bins always satisfies the closed constraint that the sum is 1. The neighborhood boundary ensures that the adjusted mix design does not deviate from the safe envelope of the robust mix design in step five, avoiding dynamic adjustments that could cause hard constraints such as strength and cost to exceed the limits. If the deviation is greater than 0.3 and is judged as a serious deviation, it can be temporarily relaxed to ±15%, but a mandatory constraint check is triggered simultaneously. The sand ratio and water consumption of the mortar components are auxiliary decision variables, with an adjustment range limited to ±2%, and are only activated when aggregate adjustments fail to meet the standards.

[0120] For each candidate ratio within the optimization domain, after component log ratio transformation, it is input into the Gaussian process response surface model to calculate the predicted mean and predicted variance of the corresponding three-dimensional homogeneity coordinates. Based on the weighting rule, the three-dimensional coordinates are fused into a comprehensive homogeneity score, and the 95% confidence interval of the score is output to provide an uncertainty reference for optimization.

[0121] Then, the Bayesian optimization iteration is driven. The acquisition function simultaneously weighs the improvement of the predicted mean against the exploratory value of the predicted variance, balancing development and exploration in small-sample optimization to avoid getting trapped in local optima. The optimization objective is to minimize the deviation between the overall homogeneity score and the target center, while incorporating a correction magnitude regularization term to prioritize feasible solutions with smaller correction amounts. The maximum number of iterations is set to 20, or it is terminated early when the overall score enters the target domain and the improvement is less than 0.01 for three consecutive steps. The total time for a single round of correction is controlled within 100ms.

[0122] The optimization process follows the gradation principle of prioritizing aggregate gradation and making minor adjustments to the slurry as a fallback: priority is given to adjusting the blending ratio of aggregate bins of different quality grades and particle sizes, and homogeneity is improved by increasing the proportion of high-quality aggregates and optimizing particle packing density; only when the deviation after adjusting the aggregate ratio still does not enter the target range is the fine-tuning ratio of sand ratio and unit water consumption used to help improve the mortar's coating and fluidity.

[0123] In the embodiment, a closure check can also be performed on the optimized output of each aggregate bin ratio to correct the slight deviation of the sum from 1 caused by numerical iteration according to the weight of each bin ratio, so as to ensure that the sum of the ratios of all aggregate bins is exactly equal to 100%.

[0124] Specifically, hard constraint checks are performed sequentially. If any check fails, the current optimal solution is rejected, the search domain is reduced by 1 / 2, and the solution is iterated again. The maximum number of retries is 3. If the retries fail, the solution is rolled back to the initial ratio. ① Warehouse storage quantity verification: The planned amount to be used in this batch in a single warehouse shall not exceed the current available storage quantity; ② Gradation verification: Calculate the cumulative sieve residue of each sieve hole according to the corrected ratio, and confirm that the full gradation curve falls within the gradation zone required by the specification; ③ Performance robustness verification: Based on the most unfavorable value of the current aggregate quality range, the predicted strength and slump still meet the minimum membership requirements of the fuzzy constraints; ④ Cost verification: The total cost per cubic meter of aggregate does not exceed the preset cost limit.

[0125] After the verification is passed, a homogeneous-oriented adaptive batching vector is generated, with each component representing the percentage of aggregate in the corresponding bin to the total aggregate mass. Simultaneously, three auxiliary parameters are output: the predicted homogeneity score, the predicted confidence level, and the total correction magnitude relative to the initial batching ratio, providing a reference for uncertainty and adjustment magnitude for subsequent fusion decisions.

[0126] S5. The quality grading membership vector and the homogeneous guiding adaptation batching vector are fused together, and the final batching execution scheme is obtained after Bayesian drift compensation correction.

[0127] First, the aggregate group quality grading membership vector, homogeneous-oriented adaptive batching vector, and environmental and process parameters collected by on-site sensors are divided into three feature branches, corresponding to the three core influencing sources of batching decisions: Quality characteristic branches: Based on the membership vector, they are arranged in the order of excellent, first-class, qualified, and unqualified, representing the quality distribution and grade uncertainty of the current batch of aggregates, and reflecting the basic influence of the intrinsic properties of aggregates on concrete performance. Ingredient feature branch: Based on homogeneity-oriented adaptive ingredient vector, it reflects the homogeneity adaptation status of the current initial ratio; Auxiliary operating condition branch: includes ambient temperature, ambient relative humidity, theoretical mixing time and percentage of mixing time. Temperature and humidity are collected minute by minute by the on-site weather station of the mixing plant to correct the impact of water evaporation and hydration rate on workability. The mixing time parameter characterizes the homogeneity development stage and assists in decision-making on when to make corrections.

[0128] To address the closed-loop characteristic of component data where the sum of the characteristic branch components of the ingredients is 1, the simplex proportional space is mapped to Euclidean space through central logarithmic ratio transformation. This avoids weight distortion caused by proportional constraints during neural network fitting and ensures that the characteristic contributions of each batch conform to the linear superposition assumption.

[0129] Then, Z-score standardization is performed on the three branches respectively. The standardization parameters are obtained by global statistics from the offline training set to eliminate the pull of different magnitude features on the network weights. The first and last branches are spliced ​​together in the order of quality branch → batching branch → working condition branch to form a multi-source fusion feature vector.

[0130] Next, using the multi-source fusion feature vector as input, a multilayer perceptron with a quality-ingredient cross-attention structure is used to complete the nonlinear mapping, and the preliminary ingredient execution ratio and corresponding performance prediction value are obtained.

[0131] Specifically, the multilayer perceptron network adopts a structured architecture of input splitting, cross-attention enhancement, and dual-branch output. The input layer first splits the fused features into three independent sub-vectors along the dimension: quality, ingredients, and working conditions. Each sub-vector is mapped to a unified 32-dimensional feature subspace through a shared fully connected layer, thus completing dimensional alignment and preliminary feature upsizing.

[0132] Then, using the quality subspace features as the query vector and the batching subspace features as both the key and value vectors, a single-head cross-attention operation is performed. During the operation, the higher the membership degree of non-conforming and conforming aggregates in the quality vector, the higher the attention weight of the corresponding low-quality aggregate bin dimension. The model automatically amplifies the adjustment weight of the batching ratio for this type of bin. After the attention weights are normalized by Softmax, they are used to perform weighted aggregation on the batching value vector, outputting quality-aware batching enhancement features. It can be understood that this structure breaks the feature weight dispersion problem of general fully connected layers, forcing the batching decision to always take aggregate quality fluctuations as the core basis, ensuring that the engineering logic of more aggressive adjustments for worse quality is implemented.

[0133] The quality-enhancing mix design features are then concatenated with the working condition subspace features and fed into two cascaded fully connected hidden layers. Each layer contains 64 neurons, with ReLU as the activation function. Dropout layers with an interlayer inactivation rate of 0.2 are used to suppress overfitting. The two hidden layers perform high-order nonlinear mapping of the features, fitting the coupling relationship between quality, mix proportion, working condition, and concrete performance to obtain global decision features.

[0134] The output layer is split into two independent parallel branches: ① Proportioning output branch: outputs an M-dimensional vector, with the end using a Softmax activation function to ensure that the sum of all components is 1. Its physical meaning is consistent with the homogeneous guiding adaptation batching vector, representing the mass proportion of each aggregate bin in the total aggregate; ② Performance prediction branch: outputs a 2-dimensional vector, with the end using a linear activation function, corresponding to the 28-day compressive strength prediction value and slump prediction value, respectively.

[0135] In this embodiment, the model is trained offline using historical production samples from the past 12 months. The samples cover different quality grades, mix proportions, and seasonal operating conditions, with a total sample size of no less than 800 groups. The optimizer used is Adam, with an initial learning rate of 0.001. The loss function is a weighted sum of the mix proportion cross-entropy loss and the performance prediction mean square error loss, with a mix proportion loss weight of 0.6 and a performance loss weight of 0.4. After training, the validation set strength prediction coefficient of determination is no less than 0.88, and the collapse prediction coefficient of determination is no less than 0.85. Only after meeting the engineering accuracy requirements can the model be deployed online.

[0136] To address slow, time-varying system biases such as front-end detection drift, gradual changes in aggregate source, and equipment wear, a Bayesian incremental learning mechanism is adopted to update only the output layer weights. This achieves online adaptive correction of system biases while preserving the model's core feature extraction capabilities, ensuring long-term operational accuracy.

[0137] In this embodiment, the compensation object is limited to the output layer weights of the MLP, while the hidden layer feature mapping parameters remain fixed to avoid the overall model deviation caused by changes in the underlying logic. Update triggering employs a dual-threshold determination: after each batch of production is completed, two types of feedback are received: the measured slump value at the factory and the measured strength value after 28 days. When the relative deviation between the model's predicted value and the measured value exceeds 15% and 10% respectively, it is determined that there is a systematic drift, triggering a weight update; deviations within the threshold are considered normal random errors, and no update is performed to avoid frequent model oscillations caused by sample noise.

[0138] The weight matrix and bias term of the output layer are then modeled as a multivariate normal distribution. The weight values ​​obtained during offline training are used as the mean of the prior distribution, and the parameter covariance during training is used as the prior uncertainty. This setting assumes that the parameters follow a Gaussian distribution around the true optimal value, and the prior distribution represents the perceived state of the parameters before the update, providing a probabilistic basis for incremental updates.

[0139] After adding new valid feedback samples, the fusion features of the current batch are used as input, and the measured ratio and performance are used as labels. The posterior distribution of the output layer weights is derived based on the Bayesian formula: first, the likelihood function is calculated from the labels, and then the posterior mean and covariance matrix are analytically solved by combining the prior distribution.

[0140] Through continuous incremental updates, the output layer weights can adaptively adjust to changes in overall aggregate quality, overall shifts in the visual inspection system, and slow declines in mixing equipment efficiency, thus offsetting systemic slow drift across the entire chain. Simultaneously, upper and lower limits are set for weight updates; the update magnitude of a single parameter must not exceed ±20% of the initial value. Exceeding these limits triggers an alarm, prompting manual verification of any abnormal changes in the front-end inspection equipment or raw material source, preventing excessive shifts in the model due to erroneous feedback.

[0141] Finally, using the drift-compensated and corrected ingredient ratio vector and the corresponding performance prediction values ​​as input, and after verification, a final executable solution is generated, including: Rigid checks are performed sequentially, and the output can only proceed if all items pass: ① Warehouse inventory verification: Connect to the warehouse quality-storage database to verify that the planned amount of goods to be used in a single warehouse in this batch does not exceed the current available storage capacity. If it is insufficient, the allocation ratio of other warehouses will be scaled up proportionally to the upper limit of storage capacity. ② Gradation compliance verification: Calculate the cumulative sieve residue of each sieve according to the final ratio, and confirm that the synthetic gradation curve falls completely within the standard gradation zone, and the predicted value of needle and flaky content meets the upper limit requirement of the standard. ③Performance guarantee rate verification: Combine the most unfavorable value of the current aggregate quality range, cross-verify that the predicted strength and slump values ​​meet the minimum membership degree requirement of the fuzzy constraint (default is not less than 0.7). ④ Cost and environmental protection verification: The comprehensive cost per cubic meter of aggregate does not exceed the preset upper limit, and the total replacement rate of recycled aggregate is not lower than the minimum design requirement; ⑤ Correction range verification: The deviation of the single-compartment ratio between the final mix ratio and the initial robust mix ratio shall not exceed ±15% to avoid dynamic adjustments deviating from the safe operating condition boundary.

[0142] If any check fails, the adjustment range is reduced according to the constraint boundary and the reasoning is re-inferred. If all checks still fail after 3 consecutive reasoning attempts, the homogeneous-oriented adaptation of the batching vector is used as the final execution plan. At the same time, an exception log is generated and pushed to the operation and maintenance end to record the deviation items and trigger reasons, prompting manual verification of model parameters and raw material status to ensure that production continuity is not affected by model anomalies.

[0143] After verification, the following are generated: ① Batching execution instructions: including the unit concrete mass usage, sand ratio adjustment value, and unit water usage adjustment value for each aggregate bin. These instructions are sent to the batching plant PLC control system in a standardized Modbus instruction format and simultaneously written into the production batch file; ② Final quality grading results: based on the membership vector and combined with the quality weight correction of the batching decision for this batch, the final grade judgment (taking the grade corresponding to the maximum membership degree) and membership vector are output and stored in the aggregate quality database for inventory management and quality traceability.

[0144] It should be noted that the specific implementation methods described above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art, according to actual needs.

[0145] Please see Figure 2 In this embodiment, to efficiently execute the online quality detection and dynamic grading and batching method for recycled aggregates provided by this invention, the present invention also provides an online quality detection and dynamic grading and batching system for recycled aggregates, comprising: an input device 1, an output device 2, a processor 3, and a memory 4, wherein the input device 1, output device 2, processor 3, and memory 4 are interconnected, and the memory 4 stores program instructions for executing the steps of the online quality detection and dynamic grading and batching method for recycled aggregates. The online quality detection and dynamic grading and batching system for recycled aggregates of this invention has a compact structure and stable performance, and can stably execute the online quality detection and dynamic grading and batching method for recycled aggregates of this invention, further improving the overall applicability and practical application capability of this invention.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A method for online quality detection and dynamic grading and batching of recycled aggregate, characterized in that, Includes the following steps: Surface defect detection of multispectral images yields the density matrix of mortar-crack coexistence defects in a single particle, and three-dimensional particle shape characterization of the three-dimensional point cloud yields the standard particle shape spectral vector of a single particle. By combining the co-occurrence defect density matrix and the standard particle shape spectrum vector, the quality classification membership vector of the current aggregate group is obtained through graph attention coupling reasoning. Based on the quality range parameters of the stored aggregate, constrained interval fuzzy robust optimization is performed to obtain the initial batching ratio vector, and spatiotemporal texture manifold identification is performed on the mixing image sequence to obtain the homogeneity manifold coordinate vector of the mixture. The initial ingredient ratio vector is dynamically corrected using the homogeneity manifold coordinate vector of the mixture to obtain a homogeneous guided adaptive ingredient vector. The final batching execution scheme is obtained by fusing the quality grading membership vector with the homogeneous guiding adaptation batching vector and then correcting it using Bayesian drift compensation.

2. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The method of obtaining the density matrix of mortar-crack coexistence defects by detecting surface defects in multispectral images includes the following steps: Feature bands corresponding to crack prominence, mortar differentiation, and matrix verification are selected from multispectral images, and aggregate area masks are extracted after illumination normalization correction. Adaptive multi-scale disk structural elements are generated based on the equivalent particle size of single-particle aggregate. Morphological operations guided by channel-space dual attention are used to separate attached mortar and microcracks, and a defect density matrix of attached mortar-crack coexistence is constructed.

3. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The process of obtaining a standardized particle shape spectral vector for a single particle by performing three-dimensional particle shape characterization on a three-dimensional point cloud includes the following steps: Background removal, noise reduction, downsampling, and individual particle segmentation are performed on the 3D point cloud to obtain the 3D contour point cloud of a single aggregate particle; The three-dimensional contour point cloud is subjected to attitude alignment, minimum bounding sphere size normalization and conformal mapping to a spherical surface to obtain a unit spherical point set. The unit spherical point set is subjected to spherical harmonic decomposition, and multi-scale grain shape features are extracted and then weighted to generate a standard grain shape spectral vector.

4. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The process of combining the co-occurring defect density matrix and the standardized particle shape spectrum vector to obtain the quality classification membership vector of the current aggregate group through graph attention coupling reasoning includes the following steps: Based on the defect statistical characteristics of a single particle and the standard particle shape spectrum vector, the quality feature vector of a single particle is obtained and used as the attribute of a graph node. An undirected weighted adjacency graph is constructed based on particle spatial distance and feature similarity, and neighborhood information is aggregated through two layers of multi-head graph attention; The global quality features of the population are obtained by particle size weighted pooling, and the quality grading membership vector is obtained by using a fuzzy membership classifier.

5. The method for online quality detection and dynamic grading of recycled aggregate according to claim 4, characterized in that, The construction of an undirected weighted adjacency graph based on particle spatial distance and feature similarity includes the following steps: The K=6 nearest neighbor rule is used to construct the edges, and the edge weights are obtained by fusing the spatial Gaussian kernel weights and the feature cosine similarity in a 6:4 ratio. The fuzzy membership classifier is preset with four quality levels: excellent, first-class, qualified, and unqualified, and the sum of the components of the membership vector after normalization is 1.

6. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The initial batching ratio vector is obtained by constrained interval fuzzy robust optimization based on the quality interval parameters of stored aggregates, including the following steps: Based on historical storage testing data and real-time quality grading membership, a storage quality-storage database is constructed, which includes the content of attached mortar, the content of needle-like and flaky particles, and the crushing index range. The rigid performance requirements are set as hard constraints, and the elastic performance requirements are set as fuzzy constraints with trapezoidal membership degrees. The most unfavorable quality condition is selected by the cut set method, and the initial batching ratio vector is obtained by the constrained particle swarm algorithm.

7. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The step of performing spatiotemporal texture manifold identification on the mixed image sequence to obtain the homogeneity manifold coordinate vector of the mixture includes the following steps: Multi-scale gray-level co-occurrence matrix features, local binary pattern features, and temporal dynamic features are extracted from the mixed image sequence and fused to generate high-dimensional spatiotemporal texture features; The high-dimensional spatiotemporal texture features are mapped to a three-dimensional manifold space using a local tangent space alignment algorithm, generating a homogeneous manifold coordinate vector of the mixture that includes the current homogeneity state value, the change trend value, and the target deviation.

8. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The step of dynamically correcting the initial ingredient ratio vector using the homogeneity manifold coordinate vector of the mixture to obtain a homogeneous guided adaptive ingredient vector includes the following steps: A Gaussian process response surface model is trained based on historical ingredient ratios and homogeneity samples. When the overall homogeneity deviation is greater than 0.15 and the trend of change is ≤0, a correction is triggered. Within the safe neighborhood of the initial proportion, the optimal proportion correction is searched through Bayesian optimization, and the homogeneous guided adaptive batching vector is obtained after constraint verification.

9. The method for online quality detection and dynamic grading of recycled aggregate according to claim 1, characterized in that, The final batching execution plan is obtained by fusing the quality grading membership vector and the homogeneous guided adaptation batching vector, and then correcting for Bayesian drift. The plan includes the following steps: By combining the quality grading membership vector, the homogeneous-oriented adaptive batching vector, and environmental condition parameters, a fusion feature vector is obtained. A multilayer perceptron with quality-batch cross-attention is then used to obtain preliminary batching decision results. The Bayesian online update is triggered based on the deviation between the factory slump and the measured strength. After the system drifts, the final batching execution plan is obtained through full constraint verification.

10. An online quality inspection and dynamic grading and batching system for recycled aggregates, characterized in that, The online quality inspection and dynamic grading and batching system for recycled aggregates includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions, which are used to execute the online quality inspection and dynamic grading and batching method for recycled aggregates according to any one of claims 1-9.

Images