Copper-clad plate prepreg defect prediction method based on interface semantic features, server and storage medium

Abstract

The application discloses a copper-clad plate prepreg defect prediction method based on interface semantic features, a server and a storage medium. The method first acquires interface tension field distribution information of a surface of a prepreg blank to be detected in a copper-clad plate prepreg manufacturing process and performs micro-area physical disturbance decoupling, extracts local interface free energy fluctuation features and stress concentration core tendency features, and obtains a micro-area latent precursor semantic feature set. Secondly, implicit feature association processing of Rayleigh instability texture is performed, an energy accumulation tendency space-time evolution mode under time boundary constraints is constructed, and defect evolution semantic feature descriptions are obtained. A defect initiation risk map is generated through defect initiation risk mapping processing. The defect initiation risk map and the interface tension gradient space change rate are combined to determine a defect evolution probability distribution in a preset time period to generate a defect prediction instruction set. Through multi-level physical field decoupling and space-time evolution modeling, the accuracy and timeliness of defect initiation risk prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, a server, and a storage medium. Background Technology

[0002] Copper-clad laminate prepreg is a key intermediate material in printed circuit board manufacturing, and its quality directly affects the electrical performance and reliability of the final product. During the prepreg manufacturing process, impregnation, drying, and other processes cause complex interfacial physicochemical interactions between the reinforcing material and the resin system, which in turn create residual stress fields and interfacial energy distributions on and inside the blank. The abnormal evolution of these microscopic physical fields is the root cause of defects such as interlayer microcracks, porosity, and delamination.

[0003] To identify defect risks in the early stages of manufacturing, the industry primarily employs methods such as infrared thermal imaging, ultrasonic scanning, and laser speckle interferometry to screen billets for macroscopic defects. These methods rely on defects reaching a certain scale before detection and are insensitive to microscopic defect precursors in their latent or early stages. Other quality monitoring methods based on statistical process control indirectly infer defect likelihood by monitoring process parameter deviations, but they cannot establish a direct physical correlation between process parameters and the risk of defects at specific spatial locations. Therefore, existing technologies struggle to accurately extract latent precursor features from physical field distribution data before defects undergo substantial evolution, and they also cannot effectively predict the spatiotemporal evolution of defects in subsequent processes. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, as well as a server and storage medium, to solve the problems mentioned in the background art.

[0005] This invention provides a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, comprising: To obtain the interfacial tension field distribution information on the surface of the prepreg blank to be tested during the manufacturing process of copper clad laminate prepreg; The interface tension field distribution information is subjected to micro-region physical perturbation decoupling processing. The interface tension vector data of each micro-region unit in the interface tension field distribution information is decomposed into steady-state interface tension reference component and local perturbation residual component. Based on the local perturbation residual component, the local interface free energy fluctuation characteristics and stress concentration kernel tendency characteristics are extracted to obtain the micro-region latent precursor semantic feature set. The latent precursor semantic feature set of the micro-region is subjected to Rayleigh instability texture latent feature association processing to construct the spatiotemporal evolution mode of energy accumulation tendency under temporal boundary constraints, and obtain the semantic feature description of defect evolution. The semantic feature description of defect evolution includes the energy accumulation tendency intensity distribution map and the stress concentration nucleus germination probability distribution map. Based on the semantic feature description of defect evolution, a defect emergence risk mapping process is performed to generate a defect emergence risk cloud map. The defect emergence risk cloud map includes a distribution of defect emergence probability values ​​and a defect category tendency coding distribution associated with spatial coordinates. By combining the defect initiation risk cloud map and the spatial change rate of the interface tension gradient in the interface tension field distribution information, the defect evolution probability distribution within a preset time period is determined, and a defect prediction instruction set is generated based on the defect evolution probability distribution.

[0006] This invention provides a defect prediction server for copper-clad laminate prepregs, comprising: A processor; a storage device on which a computer program is stored; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the above-described method for predicting defects in copper-clad laminate prepregs based on interface semantic features.

[0007] The present invention provides a readable storage medium on which a program or instruction is stored, and when the program or instruction is executed by a processor, it implements the above-mentioned method for predicting defects in copper-clad laminate prepreg based on interface semantic features.

[0008] Compared with existing technologies, the beneficial effects of this invention are as follows: By acquiring the interface tension field distribution information and performing micro-region physical perturbation decoupling processing, the original tension vector data is decomposed into steady-state reference components and local perturbation residual components. From these components, local interface free energy fluctuation characteristics and stress concentration kernel tendency characteristics are extracted to form a micro-region latent precursor semantic feature set, achieving physical separation and quantitative characterization of hidden defect precursors. Based on this, by performing Rayleigh instability texture latent precursor semantic feature set latent precursor processing, a spatiotemporal evolution model of energy accumulation tendency under temporal boundary constraints is constructed, and a defect evolution semantic feature description is obtained, coupling isolated micro-region features into a defect embryo expression with spatiotemporal evolution attributes. Furthermore, a defect germination risk cloud map is generated through defect germination risk mapping processing, and combined with the spatial change rate of interface tension gradient to determine the defect evolution probability distribution, ultimately generating a defect prediction instruction set. The entire process, from macroscopic field distribution to microscopic perturbation decoupling, and then to cross-micro-region correlation and spatiotemporal evolution modeling, forms a multi-level, multi-physical field fusion defect prediction link. This overcomes the shortcomings of existing methods in being insensitive to latent defects and lacking spatiotemporal evolution prediction capabilities, and improves the accuracy, timeliness, and interpretability of defect initiation risk prediction. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, provided in an embodiment of this application.

[0012] Figure 3 The temporal partitioning multiphysics field linkage distribution diagram provided in the embodiments of this application.

[0013] Figure 4 This is a schematic diagram of the basic structure of a copper-clad laminate prepreg defect prediction server provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0015] Please see Figure 1 , Figure 1 This is a flowchart of a method for predicting defects in copper-clad laminate prepregs based on interface semantic features, provided in an embodiment of this application. The method can be executed by a copper-clad laminate prepreg defect prediction server, or by a copper-clad laminate prepreg defect prediction server and a server together. The method includes steps 110-150.

[0016] This invention obtains the interfacial tension field distribution information on the surface of the prepreg blank to be tested during the manufacturing process of copper clad laminate prepreg, and then sequentially performs micro-area physical disturbance decoupling, Rayleigh instability texture latent feature association, defect initiation risk mapping, and multi-factor fusion defect evolution probability prediction, finally generating a defect prediction instruction set, thereby achieving early and accurate identification of the risk of micro-defect initiation inside the prepreg blank.

[0017] Step 110: Obtain the interfacial tension field distribution information on the surface of the prepreg blank to be tested during the manufacturing process of copper clad laminate prepreg.

[0018] In this embodiment of the invention, the prepreg blank to be tested is a sheet-like composite material that has undergone upstream processes such as impregnation and drying before entering the next pressing or testing process. The blank is composed of reinforcing material (such as glass fiber cloth) and resin system attached thereto. During the manufacturing process, due to the flow of resin, the evaporation of solvent and the difference in thermal expansion coefficient between reinforcing material and resin, a complex residual stress field and interfacial energy distribution will be formed on the surface and near the surface of the blank. The spatial distribution of the above physical quantities is closely related to the initiation of subsequent defects (such as interlayer microcracks, pores, delamination, etc.).

[0019] To obtain distribution information, a pre-set interface characteristic sensing unit can be used to perform non-contact scanning of the billet surface. The working mechanism of this sensing unit can be either an optical measurement mode based on digital image correlation or an acoustic measurement mode based on ultrasonic surface wave reflection. For example, in the optical measurement mode, a high-resolution texture image sequence of the billet surface under a set light source is first acquired using a high-speed industrial camera. Then, a subset matching algorithm is used to track the displacement field of each pixel subset between consecutive image frames. The in-plane strain field is then calculated from the displacement field using numerical differentiation. Based on the obtained strain field and combined with the constitutive relation of the billet material, the stress tensor at each point on the surface can be inverted. The principal direction and component values ​​of the stress tensor are used as the interface tension vector at the corresponding sampling point. The interface tension vectors of all sampling points together constitute the interface tension field distribution information. This information is structured into a two-dimensional dataset, where each element corresponds to a sampling point, and the element value includes the spatial coordinates of the point, the interface tension amplitude, and the interface tension direction angle. For example, if a rectangular coordinate system is constructed on the surface of the billet with one corner of the billet as the origin, the information of a sampling point Pij can be represented as Pij (x-coordinate offset i, y-coordinate offset j, interfacial tension amplitude Fij, direction angle θij).

[0020] Step 120: Perform micro-region physical perturbation decoupling processing on the interface tension field distribution information, decompose the interface tension vector data of each micro-region unit in the interface tension field distribution information into steady-state interface tension reference component and local perturbation residual component, extract local interface free energy fluctuation features and stress concentration kernel tendency features based on the local perturbation residual component, and obtain micro-region latent precursor semantic feature set.

[0021] After obtaining the interfacial tension field distribution information, it is necessary to separate the macroscopic stable distribution determined by the inherent properties of the material and ideal process conditions, as well as the disturbance information caused by microscopic defect precursors, impurities, or local inhomogeneities. This decoupling process is the core of this step.

[0022] Therefore, step 120 can be further implemented through steps 121 to 124.

[0023] Step 121: Perform spatial micro-region division processing on the interfacial tension field distribution information. Based on the surface texture direction of the prepreg blank and the preset micro-region scale, divide the surface of the prepreg blank into multiple adjacent micro-region units with overlapping boundaries, and assign a spatial position index to each micro-region unit to obtain the micro-region unit division result.

[0024] Specifically, the main texture direction is first analyzed from the surface image or interfacial tension amplitude distribution map of the blank. For example, the gradient direction histogram method is used to statistically analyze the distribution of the interfacial tension direction angle θij at each sampling point throughout the entire surface, thereby determining the two main texture directions, typically the warp and weft directions.

[0025] Then, a non-orthogonal or orthogonal grid coordinate system is established based on the texture direction, with the coordinate axes aligned with the main texture direction. Based on a preset micro-area scale, for example, setting the micro-area unit to contain m rows and n columns of sampling points in the longitudinal and latitudinal directions respectively, the blank surface can be divided into a regularly arranged, adjacent, and overlapping array of micro-area units. For a sampling point matrix of size M rows × N columns, after partitioning, (M / m) × (N / n) micro-area units are obtained. A unique spatial location index is assigned to each micro-area unit, for example, using a tuple (u, v) consisting of its row number and column number as the index, where u is an integer variable starting from 1 representing the row order, and v is an integer variable starting from 1 representing the column order, thus obtaining a structured micro-area unit partitioning result.

[0026] Step 122: Perform steady-state trend surface fitting on the interfacial tension vector data in each micro-unit, and construct the steady-state reference surface of interfacial tension in the micro-unit using a local weighted regression algorithm to obtain the micro-unit steady-state interfacial tension reference component. The steady-state reference surface of interfacial tension represents the ideal distribution of the interfacial tension field in the micro-unit under undisturbed conditions.

[0027] For each micro-cell (u, v) in the aforementioned array of micro-cells, a locally weighted regression algorithm is run on the point set consisting of the spatial coordinates of all sampling points within that micro-cell and the interfacial tension amplitude F. The algorithm is implemented as follows: For a point within the micro-cell to be fitted, a window centered on that point is selected. Within the window, weights are assigned based on the geometric distance of other points to the center point; the closer the distance, the greater the weight. The weighting function can be a cubic kernel function.

[0028] Then, a low-order polynomial, such as a quadratic polynomial, is fitted using the weighted least squares method. The fitted value of this polynomial at the center point is the steady-state interfacial tension amplitude at that point. By traversing all points within the micro-cell, a smooth surface representing the ideal distribution pattern under undisturbed conditions can be constructed, namely the steady-state reference surface of interfacial tension. Each point on this surface corresponds to a steady-state amplitude Fbase and a steady-state direction θbase. These data together constitute the micro-cell steady-state interfacial tension reference component.

[0029] Step 123: Based on the steady-state interface tension reference component of the micro-region, perform residual separation processing on the interface tension vector data of each sampling point in the micro-region unit. Perform difference calculation between the interface tension amplitude and interface tension direction angle of each sampling point and the steady-state amplitude and steady-state direction at the corresponding positions on the steady-state reference surface of the interface tension to obtain the local disturbance residual component of the micro-region. The local disturbance residual component of the micro-region includes the disturbance amplitude residual and the disturbance direction residual.

[0030] After obtaining the steady-state reference surface for interfacial tension of the micro-unit (u, v), for each sampling point Pij within the unit, the originally measured interfacial tension vector is (Fij, θij), while the reference vector information for the same location indexed from the steady reference surface is (Fbase_ij, θbase_ij). Residual separation is performed by executing two independent difference operations: first, the amplitude residual ΔFij is calculated, which is the original amplitude Fij minus the reference amplitude Fbase_ij; then, the direction residual Δθij is calculated, which is the angular difference between the original direction angle θij and the reference direction angle θbase_ij. This difference needs to be processed through a wraparound method to ensure that the result falls within ±180 degrees.

[0031] After processing, the micro-cell forms a perturbation residual dataset. Each record in the dataset contains the perturbation amplitude residual ΔF and the perturbation direction residual Δθ. This dataset constitutes the micro-cell local perturbation residual component of the micro-cell.

[0032] Step 124: Based on the residual components of the local disturbance in the micro-region, perform interface free energy fluctuation quantization and stress concentration kernel tendency analysis on the micro-region unit to obtain local interface free energy fluctuation characteristics and stress concentration kernel tendency characteristics. Combine the local interface free energy fluctuation characteristics and the stress concentration kernel tendency characteristics to generate a set of semantic features for latent precursors in the micro-region.

[0033] The above step 124 performs deep feature mining on the local perturbation residual components of each micro-region unit, specifically implemented by steps 1241 to 1244.

[0034] Step 1241: Based on the disturbance amplitude residual in the local disturbance residual component of the micro-region, perform interfacial free energy fluctuation quantification processing on each micro-region unit, correlate and map the square of the disturbance amplitude residual with the surface tension coefficient of the semi-cured sheet blank to generate local interfacial free energy fluctuation characteristics. The local interfacial free energy fluctuation characteristics characterize the degree of fluctuation of interfacial free energy in space and the tendency of energy accumulation in the micro-region unit.

[0035] For a micro-cell (u, v), the residual perturbation amplitude ΔF of all sampling points within it is extracted. The interfacial free energy fluctuation is a physical quantity positively correlated with the square of the perturbation amplitude. The process involves calculating the square of the residual perturbation amplitude for each sampling point within the micro-cell. This squared value is then multiplied by a surface tension coefficient gamma, representing an intrinsic property of the prepreg material. The surface tension coefficient gamma can be obtained in advance through independent material testing experiments, such as by contact angle measurement.

[0036] Thus, the sampling point yields an instantaneous interfacial free energy fluctuation metric, Eud_ij, constituting an energy fluctuation scalar field. To generate a feature vector characterizing the overall fluctuation features of this micro-region, a statistical transformation is performed on this energy fluctuation scalar field. Specifically, the mean, standard deviation, skewness, kurtosis, and full width at half maximum (FWHM) of the scalar field can be calculated, and these statistics are concatenated in a preset order to form a multidimensional feature vector. This vector represents the local interfacial free energy fluctuation characteristics of the micro-region unit, characterizing the intensity of the interfacial free energy fluctuations within the micro-region space and the presence of locally anomalously clustered energy clusters.

[0037] Step 1242: Based on the disturbance direction residual in the local disturbance residual component of the micro-region, stress concentration kernel tendency analysis is performed on each micro-region unit. The spatial divergence of the disturbance direction residual is used to determine the disturbance direction convergence region within the micro-region unit. Stress concentration kernel tendency features are generated based on the area connectivity and directional convergence density of the disturbance direction convergence region. The stress concentration kernel tendency features characterize the probability and approximate spatial range of stress concentration kernel initiation within the micro-region unit.

[0038] Similarly, for a micro-cell (u, v), the perturbation direction residual Δθ of all sampling points within it is extracted. First, the discrete perturbation direction residual field within this micro-cell is processed into a continuous vector field. Then, spatial divergence is calculated on this vector field. In the divergence field, regions with negative divergence values ​​and large absolute values ​​represent regions where the vector field arrows converge, i.e., the perturbation direction convergence region. By setting a preset divergence threshold, for example, taking the mean of the divergence values ​​of all sampling points minus 1 standard deviation as the lower limit of the threshold, the perturbation direction convergence region is selected.

[0039] Next, connectivity analysis is performed on the above regions, marking spatially adjacent clusters of convergent pixels as the same connected region. For each connected region, its area is calculated, and the convergence density of the perturbation direction vectors within it is statistically analyzed. This convergence density can be defined as the negative divergence integral of the unit vector of the directional residuals of all sampling points within the connected region divided by the area of ​​the connected region. Finally, by combining the area and convergence density of each connected region, a set of characteristic parameters is formed to describe whether there is a high-probability stress concentration core and its range within the micro-region. This set of parameters constitutes the stress concentration core tendency feature of the micro-region.

[0040] Step 1243: Semantic feature encapsulation processing is performed on the local interface free energy fluctuation features and stress concentration core tendency features. The local interface free energy fluctuation features and stress concentration core tendency features are associated and combined with the spatial location index of the corresponding micro-region unit to obtain the latent precursor semantic feature record of the micro-region unit.

[0041] For a micro-unit (u, v), after obtaining its local interface free energy fluctuation characteristics (defined as a feature vector V_N containing several statistics) and stress concentration kernel tendency characteristics (defined as a feature vector V_L containing several connected component description parameters), an encapsulation operation is performed. Specifically, one data record is created, containing three main fields. The first field is the spatial location index, whose value is the binary identifier (u, v) of the micro-unit. The second field is the interface free energy fluctuation characteristic field, whose value is V_N. The third field is the stress concentration kernel initiation tendency characteristic field, whose value is V_L. This record, which strongly correlates physical characteristics with spatial location, is the latent precursor semantic feature record of the micro-unit.

[0042] Step 1244: Traverse all micro-region units and aggregate the latent precursor semantic feature records of each micro-region unit to generate a micro-region latent precursor semantic feature set. Each record in the micro-region latent precursor semantic feature set includes a micro-region unit spatial location index field, an interface free energy fluctuation feature field, and a stress concentration nucleus germination tendency feature field.

[0043] For all micro-regions obtained in step 121, repeat steps 1241 to 1243. For each latent precursor semantic feature record obtained, add it to a general list or database table. After traversing all micro-regions, the set containing all records constitutes the micro-region latent precursor semantic feature set. This feature set is a structured data entity, where each row corresponds to one micro-region, and the columns strictly correspond to the spatial location index field, the interface free energy fluctuation feature field, and the stress concentration nucleus germination tendency feature field.

[0044] Step 130: Perform Rayleigh instability texture latent feature association processing on the semantic feature set of latent precursors of the micro-region to construct a spatiotemporal evolution mode of energy accumulation tendency under temporal boundary constraints, and obtain a semantic feature description of defect evolution. The semantic feature description of defect evolution includes an energy accumulation tendency intensity distribution map and a stress concentration nucleus germination probability distribution map.

[0045] After obtaining the latent precursor semantics of each isolated micro-region, the next step is to analyze the physical interaction between adjacent micro-regions on a spatial scale. This interaction follows the Rayleigh instability principle in materials science, that is, the perturbation growth and mass transport driven by interface energy tend to couple with each other under certain conditions to form a larger-scale defect precursor. The above step 130 can be further realized through steps 131 to 134.

[0046] Step 131: Perform spatial adjacency relationship construction processing on the semantic feature set of the micro-region latent precursors. Based on the spatial location index of each micro-region unit, determine the set of neighboring micro-region units that are spatially adjacent to the micro-region unit, and establish the micro-region spatial adjacency topology to obtain the micro-region adjacency relationship graph.

[0047] Based on the spatial location index (u, v) of each record in the micro-region latent precursor semantic feature set, the spatial topology of the entire billet surface can be established. For a micro-region cell with index (u, v), its neighbor cell set is defined according to its position in the grid array. If the 4-neighborhood rule is adopted, its neighbor set includes micro-region cells with indices (u-1, v), (u+1, v), (u, v-1), and (u, v+1), while the boundary cells naturally have fewer neighbors. Traversing all micro-region cells, each micro-region cell is treated as a graph node. If two micro-region cells are neighbors, an undirected edge is created between them. Thus, a graph data structure representing global spatial adjacency relationships is constructed, namely the micro-region adjacency graph. The nodes of this graph store the latent precursor semantic features of the micro-region cell, and the edges represent spatial adjacency relationships.

[0048] Step 132: In the micro-region adjacency graph, Rayleigh instability condition determination processing is performed on each pair of adjacent micro-region units. Based on the difference in the local interface free energy fluctuation characteristics and the directional consistency of the stress concentration core tendency characteristics of adjacent micro-region units, adjacent micro-region unit pairs that satisfy the Rayleigh instability perturbation growth criterion are identified, and Rayleigh unstable micro-region associated edge sets are obtained.

[0049] In the micro-region adjacency graph, each edge and the pair of micro-region nodes connected by that edge are traversed. Two criteria are extracted from the characteristics of this pair of nodes. The first criterion is the difference in the local interface free energy fluctuation characteristics. The absolute value of the difference between the components representing the total fluctuation energy in the V_N eigenvectors of each pair of nodes is calculated. If this difference exceeds a preset threshold, it indicates that there is a significant chemical potential gradient or energy difference between the two micro-regions, constituting the driving force for perturbation propagation. The second criterion is the directional consistency of the stress concentration core tendency characteristics. This involves analyzing whether there are connected stress concentration regions pointing towards each other in the V_L eigenvectors of each pair of nodes, or whether the angle between the principal stress concentration trend directions of the two is less than a certain acute angle threshold, such as 45 degrees. If both the energy difference condition and the directional consistency condition are satisfied, then this pair of adjacent micro-region units is determined to satisfy the Rayleigh instability perturbation growth criterion. Edges between all pairs of micro-region nodes that satisfy this criterion are assigned a set label, and the set of these labeled edges is the Rayleigh unstable micro-region associated edge set.

[0050] Step 133: Perform connected component extraction processing on the associated edge set of Rayleigh unstable micro-regions, merge the associated edges with shared micro-region units, and generate multiple non-overlapping Rayleigh unstable perturbation connected domains. Each Rayleigh unstable perturbation connected domain contains a set of spatially continuous micro-region units with uniform perturbation growth conditions.

[0051] The subgraph formed by the Rayleigh unstable micro-region associative edge set obtained in step 132 and the nodes it connects is then subjected to a connected component extraction algorithm. The algorithm's logic is to find all maximal connected subgraphs within this subgraph. Within a maximal connected subgraph, any two nodes can be connected to each other through several marked edges, while there are no such connecting paths between nodes in different maximal connected subgraphs. The set of all micro-region units contained in each maximal connected subgraph constitutes a Rayleigh unstable perturbation connected domain.

[0052] In this way, the originally discrete micro-unit groups on the billet surface, which satisfy the Rayleigh unstable coupling condition, are identified and merged into several non-overlapping large spatially continuous regions. Each of these connected regions represents a "defect embryo" region that may independently evolve defects.

[0053] Step 134: For each Rayleigh instability perturbation connected domain, generate an energy accumulation tendency intensity distribution map and a stress concentration kernel germination probability distribution map, and determine the semantic feature description of defect evolution through temporal partitioning and spatiotemporal correlation modeling.

[0054] Step 134 above refines and models the internal features and spatiotemporal evolution of each Rayleigh unstable perturbation connected domain, specifically implemented by steps 1341 to 1345.

[0055] Step 1341: Perform energy accumulation tendency accumulation processing on the local interface free energy fluctuation characteristics in each Rayleigh unstable perturbation connected domain. Then, perform spatial weighted accumulation of the interface free energy fluctuation characteristic values ​​of all micro-units in the Rayleigh unstable perturbation connected domain, and use the spatial gradient direction of the interface free energy fluctuation characteristic values ​​as the weighting direction to generate the energy accumulation tendency intensity distribution map of the Rayleigh unstable perturbation connected domain.

[0056] For each given Rayleigh-unstable perturbation connected domain, the local interface free energy fluctuation characteristics of all its micro-units are extracted. First, the spatial gradient field of the energy fluctuation characteristic values ​​within the connected domain is calculated. This gradient field provides an amplitude and a direction at each micro-unit, with the direction pointing towards the region of fastest energy increase. Then, isotropic spatial weighted accumulation is performed. Specifically, starting from the geometric center of the connected domain, a radial scan is performed outwards. For a given position on the scan line, the accumulated energy accumulation tendency intensity value is obtained by weighting and summing the interface free energy fluctuation characteristic values ​​of all micro-units within a certain step before and after that position, using the cosine of the angle between their gradient directions and the scan line direction as weights. The closer the gradient direction is to the scan line direction, the greater its contribution.

[0057] In detail, the energy accumulation tendency intensity value is a core scalar value characterizing the degree of energy accumulation, the magnitude of interfacial driving force, and the tendency of material transport at various spatial locations within a Rayleigh instability perturbation connected domain. Its physical essence is a spatial quantitative representation of the chemical potential gradient generated by the difference in interfacial free energy between adjacent micro-units. The calculation method is implemented using the technical logic of Rayleigh instability texture association in this embodiment of the invention: The geometric center of each Rayleigh instability perturbation connected domain is determined, and a polar coordinate system is established with this center as the origin. A radial scan is performed at 1° intervals across the entire angular range from 0° to 360°, with a preset radial scan step size of [missing value]. The edge length of the micro-unit is 0.5 times that of the micro-unit. For any position on the scan line, the local interface free energy fluctuation feature values ​​of all micro-units within one step before and after that position are extracted (i.e., the mean component in the six-dimensional statistical feature vector obtained by mapping the square of the residual of the disturbance amplitude and the surface tension coefficient). The cosine of the angle between the spatial gradient direction of the interface free energy fluctuation feature value of each micro-unit and the current scan line direction is used as the weight to weight all feature values. The weighted summation result is the energy accumulation tendency intensity value at that scan position. After the full-angle and full-radial scan is completed, a complete energy accumulation tendency intensity distribution map is generated.

[0058] After calculating all locations, a distribution map of the energy accumulation tendency intensity, with values ​​continuously distributed in space, is obtained. In the map, high-value regions indicate locations where the driving force of disturbance energy is strong, and matter and energy tend to accumulate there.

[0059] Step 1342: Perform probability density estimation on the stress concentration kernel tendency characteristics in each Rayleigh unstable perturbation connected domain. Based on the nonparametric kernel density estimation, construct a spatial continuous probability density field for the stress concentration kernel emergence tendency characteristics of each micro-unit in the Rayleigh unstable perturbation connected domain, and generate a probability distribution map of stress concentration kernel emergence in the Rayleigh unstable perturbation connected domain.

[0060] For a connected domain with the same Rayleigh instability perturbation, the stress concentration kernel tendency eigenvalue of each micro-unit within it is extracted. This value can be a scalar representing the convergence density of the most significant stress concentration kernel within that micro-unit. The discrete set of numerical points with spatial coordinates is then used as input to a nonparametric kernel density estimation algorithm. This algorithm places a smooth kernel function, such as a Gaussian kernel, at each data point location in the spatial domain and then spatially superimposes and normalizes all kernel functions. The bandwidth parameter for kernel density estimation can be adaptively determined based on the average nearest neighbor distance of the data points to avoid over-smoothing or under-smoothing.

[0061] This algorithm transforms the originally discrete stress concentration tendency at the micro-unit level into a spatially continuous probability density function that is differentiable everywhere in the entire connected domain. The value of this function at any spatial coordinate represents the probability density of that point as the location of stress concentration nucleus initiation. Visualizing or matrixing the probability density values ​​of each point in the entire connected domain generates a probability distribution map of stress concentration nucleus initiation in that domain.

[0062] Step 1343: Perform temporal boundary segmentation on the energy accumulation tendency intensity distribution map and stress concentration kernel germination probability distribution map of each Rayleigh unstable perturbation connected domain. Based on the time process stage of the prepreg manufacturing process, introduce temporal boundary constraint parameters corresponding to the process stage of the prepreg blank, and divide the energy accumulation tendency intensity distribution map and stress concentration kernel germination probability distribution map into temporal partitions corresponding to different process stages to obtain temporal partition energy accumulation tendency intensity distribution map and temporal partition stress concentration kernel germination probability distribution map.

[0063] This step introduces a temporal dimension to the manufacturing process. The prepreg manufacturing process includes continuous stages such as impregnation, drying, and cooling. A set of temporal boundary constraint parameters corresponding to the duration of each stage is defined; for example, the drying stage lasts for T1 seconds, and the cooling stage lasts for T2 seconds. The current blank to be inspected is after a known process, such as the end of drying. The focus is on defect evolution over a preset time period, which may span multiple stages such as cooling and subsequent placement. Based on the above temporal boundary constraint parameters, the time axis of interest is divided into multiple temporal partitions; for example, temporal partition 1 corresponds to the beginning of the cooling stage, and temporal partition 2 corresponds to the end of the cooling stage or the placement stage. For each temporal partition, valid distribution data within the time window of that temporal partition is extracted from the energy accumulation tendency intensity distribution map and the stress concentration nucleus germination probability distribution map. This validity can be inferred based on a physical model, for example, considering the modulation effect of the temperature field on the energy accumulation tendency intensity distribution map at that stage. Thus, the temporal partition energy accumulation tendency intensity distribution map and the temporal partition stress concentration nucleus germination probability distribution map associated with each process stage are obtained.

[0064] Step 1344: Perform spatiotemporal correlation modeling on the temporal partition energy accumulation tendency intensity distribution map and the temporal partition stress concentration nucleus germination probability distribution map. Spatially align and superimpose the energy accumulation tendency intensity distribution and stress concentration nucleus germination probability density distribution within the same temporal partition to generate a spatiotemporal evolution model of energy accumulation tendency that reflects the linkage evolution relationship between energy accumulation tendency and stress concentration nucleus germination probability under temporal boundary constraints. The spatiotemporal evolution model of energy accumulation tendency includes the energy accumulation tendency intensity evolution trajectory and the stress concentration nucleus germination probability density evolution trajectory under the temporal sequence.

[0065] For each temporal partition, the energy accumulation tendency intensity distribution map and the stress concentration nucleus initiation probability distribution map belonging to that temporal partition are aligned and superimposed on the same spatial coordinate grid. At each grid node, two values ​​are simultaneously held: the energy accumulation tendency intensity from the first map and the stress concentration nucleus initiation probability density from the second map. The spatial pairing of these two values ​​is itself a static spatial correlation.

[0066] Furthermore, the spatial distributions of the above pairs under different temporal partitions are concatenated in the order of the temporal partitions. For a fixed spatial location, one evolutionary trajectory of energy accumulation tendency intensity changing with time (i.e., the temporal partition sequence) and one evolutionary trajectory of stress concentration nucleus germination probability density are extracted. The combination of these two evolutionary trajectories constitutes a local spatiotemporal evolution model for that spatial point. Summarizing the above models for all spatial points yields the spatiotemporal evolution model of energy accumulation tendency in the Rayleigh instability perturbation connected domain. This model quantitatively describes how energy driving forces and stress concentration phenomena co-evolve under the joint constraints of time and space.

[0067] Step 1345: Integrate the spatiotemporal evolution patterns of energy accumulation tendency of all Rayleigh unstable perturbation connected domains to generate a semantic feature description of defect evolution. The semantic feature description of defect evolution uses Rayleigh unstable perturbation connected domains as basic units and stores the temporal sequence energy accumulation tendency intensity distribution map and the temporal sequence stress concentration kernel germination probability distribution map of each Rayleigh unstable perturbation connected domain.

[0068] Traverse all Rayleigh unstable perturbation connected components generated in step 133, and perform steps 1341 to 1344 on each connected component to obtain its corresponding spatiotemporal evolution pattern of energy accumulation tendency. Finally, systematically organize and package the processing results of all connected components to form a high-level semantic data construct, namely, a semantic feature description of defect evolution.

[0069] In this description, each Rayleigh instability perturbation connected component is assigned a unique identifier as a basic unit. Under each basic unit, in the order of the temporal partition index, the energy accumulation tendency intensity distribution map and the stress concentration kernel initiation probability distribution map of the connected component in the corresponding temporal state are arranged sequentially. This structured data arrangement provides a complete spatiotemporal four-dimensional information panorama for subsequent steps of global risk mapping and prediction.

[0070] Step 140: Perform defect emergence risk mapping processing based on the defect evolution semantic feature description to generate a defect emergence risk cloud map. The defect emergence risk cloud map includes a defect emergence probability value distribution and a defect category tendency coding distribution associated with spatial coordinates.

[0071] Based on the physical field evolution information provided by the semantic feature description of defect evolution, this step transforms it into a more direct risk measurement indicator, namely the probability and approximate category of defect emergence, and integrates it into a unified "risk cloud map". The above step 140 can be further implemented through steps 141 to 147.

[0072] Step 141: Perform temporal fusion processing on the temporal sequence energy accumulation tendency intensity distribution map and the temporal sequence stress concentration kernel germination probability distribution map of each Rayleigh unstable perturbation connected domain in the defect evolution semantic feature description, and extract the energy accumulation tendency intensity distribution and stress concentration kernel germination probability density distribution of each Rayleigh unstable perturbation connected domain in the current temporal partition and adjacent future temporal partitions to obtain the current temporal energy accumulation tendency intensity distribution map, the current temporal stress concentration kernel germination probability distribution map, the future temporal energy accumulation tendency intensity distribution map, and the future temporal stress concentration kernel germination probability distribution map.

[0073] Please refer to the following: Figure 3 In the semantic feature description of defect evolution, the temporal sequence data stored in each Rayleigh instability perturbation connected component is located. The temporal partition corresponding to the process stage to which the current moment belongs is defined as Tcur, and its next predefined future temporal partition is defined as Tnext. Data from the above two defined partitions is extracted from the dataset. Thus, for the same connected component, the following results are obtained: Figure 3 The four distribution maps shown are: the energy accumulation tendency intensity distribution map Ecur and the stress concentration nucleus germination probability distribution map Scur under the current time zone Tcur, and the energy accumulation tendency intensity distribution map Enext and the stress concentration nucleus germination probability distribution map Snext under the adjacent future time zone Tnext. These four distribution maps constitute the data basis for subsequent current risk assessment and future risk prediction.

[0074] Step 142: Perform joint probability mapping processing on the current temporal energy accumulation tendency intensity distribution map and the current temporal stress concentration nucleus germination probability distribution map. By performing a spatial dot product operation on the current temporal energy accumulation tendency intensity value and the current temporal stress concentration nucleus germination probability density value, a current defect germination risk probability distribution map reflecting the defect germination risk at the current moment is generated.

[0075] In this embodiment of the invention, the current temporal energy accumulation tendency intensity value refers to the scalar value obtained by spatial weighted accumulation at each spatial grid node of the Rayleigh unstable perturbation connected domain under the current temporal partition Tcur. It characterizes the degree of perturbation energy accumulation around the node and the magnitude of the interface energy driving force. Its calculation method follows the feature association logic after decoupling of micro-region physical perturbation in this embodiment of the invention: First, scan outward radially from the geometric center of the Rayleigh unstable perturbation connected domain. The preset scan step size is 0.5 times the side length of the micro-region unit. For any position on the scan line, extract the local interface free energy fluctuation feature value (i.e., the mean component in the six-dimensional feature vector) of all 32×32 sampling point micro-region units within the preset step size range before and after the position. Use the cosine of the angle between the spatial gradient direction of the interface free energy fluctuation feature value of each micro-region unit and the scan line direction as the weight to perform weighted accumulation on all feature values. The accumulation result is the energy accumulation tendency intensity value at that position.

[0076] The current temporal stress concentration kernel emergence probability density value refers to the stress concentration kernel emergence probability density at each spatial grid node in the Rayleigh unstable perturbation connected domain under the current temporal partition Tcur, obtained by estimating the stress concentration kernel emergence probability density through nonparametric Gaussian kernel density. The calculation method is to normalize the stress concentration kernel tendency characteristic value (i.e. the maximum convergence density of the connected domain) of the micro-region cells around the node by superimposing Gaussian kernel functions. The kernel function bandwidth parameter is taken as the average value of the nearest neighbor distance between the center points of all micro-region cells in the connected domain.

[0077] On the current temporal partition Tcur cross section, the aforementioned distribution maps Ecur and Scur, belonging to the same connected domain, are jointly analyzed. The joint analysis is implemented using a point-by-point dot product operation. A spatial coordinate grid is established, where the coordinates of each node are represented by a two-dimensional index (x, y). For each node (xi, yj) in the grid, the intensity value Ecur(xi, yj) of that point is read from the current temporal energy accumulation tendency intensity distribution map Ecur, and the probability density value Scur(xi, yj) of that point is read from the current temporal stress concentration nucleus initiation probability distribution map Scur. Then, the current defect initiation risk probability value Pcur(xi, yj) at that node is generated through the following relationship: Pcur(xi, yj) = Ecur(xi, yj) × Scur(xi, yj).

[0078] Where i and j represent the row index and column index of the node in the spatial coordinate grid established on the two-dimensional plane of the semi-cured sheet blank surface, respectively, both of which are consecutive integers starting from 0. The spatial coordinate grid takes the bottom left corner of the blank as the origin, the main longitudinal texture direction (length direction) of the blank as the x-axis, and the main latitudinal texture direction (width direction) as the y-axis. The spatial resolution of the grid in both the x-axis and y-axis directions is uniformly set to 0.1 mm. Therefore, the actual physical spatial coordinates of the grid node can be obtained by the formula (x=i×Δx, y=j×Δy), where Δx and Δy are both 0.1 mm.

[0079] This invention only predicts defects on the surface and within 0.1 mm of the prepreg blank. All physical quantity measurements (interfacial tension field distribution), micro-region division, feature extraction, and spatiotemporal evolution modeling are completed in a two-dimensional plane. The material property differences in the thickness direction of the blank have been incorporated into the inversion calculation of the interfacial tension field through orthogonal anisotropic linear elastic constitutive relations. There is no need to introduce a separate three-dimensional z-axis coordinate dimension. Therefore, there is no problem of dimensional mismatch between three-dimensional spatial coordinates and two-dimensional coordinates.

[0080] The above multiplication operation is performed by traversing all grid nodes, resulting in a spatial distribution map with the risk probability value Pcur as the pixel gray level. This is the current defect initiation risk probability distribution map Pcur. The physical meaning of this distribution map is that only spatial locations with a strong tendency for energy accumulation and a high tendency for stress concentration will be assigned a high probability of defect initiation.

[0081] Step 143: Perform joint probability mapping processing on the future temporal energy accumulation tendency intensity distribution map and the future temporal stress concentration kernel germination probability distribution map. By performing a spatial dot product operation on the future temporal energy accumulation tendency intensity value and the future temporal stress concentration kernel germination probability density value, a future defect germination risk probability distribution map reflecting the future defect germination risk is generated.

[0082] In this embodiment of the invention, the future temporal energy accumulation tendency intensity value refers to the energy accumulation intensity of each spatial grid node in the Rayleigh unstable perturbation connected domain under the adjacent future temporal partition Tnext, after being modulated by the temporal boundary constraint parameters of the corresponding process stage. Its calculation method is based on the current temporal energy accumulation tendency intensity distribution, combined with the influence coefficient of the physical parameters of the prepreg manufacturing process stage (such as drying and cooling section, cooling section) corresponding to the future temporal partition on the energy accumulation rate for time domain extrapolation: First, the duration and temperature field change curve of the future temporal partition are determined according to the current batch process parameters. The energy dissipation coefficient is calculated by introducing the dynamic equation of the resin curing degree changing with time. Then, the energy accumulation tendency intensity value of each node in the current temporal is multiplied by the coefficient and combined with the energy gradient propagation direction for spatial extrapolation, and finally the energy accumulation tendency intensity distribution of the future temporal is obtained.

[0083] In addition, the probability density value of stress concentration nucleus germination in the future temporal region refers to the probability density of stress concentration nucleus germination at each spatial grid node of the Rayleigh instability perturbation connected domain under the adjacent future temporal region Tnext, after being modulated by the temporal boundary constraint parameters. The calculation method is based on the current temporal probability density distribution, combined with the resin stress relaxation coefficient corresponding to the future temporal region (determined by the difference between the temperature of this stage and the glass transition temperature of the resin) for temporal extrapolation, while also considering the spatial diffusion effect of stress concentration nuclei.

[0084] Consistent with the logic of step 142, the input data is replaced with the energy accumulation tendency intensity distribution map Enext and the stress concentration nucleus initiation probability distribution map Snext under the future temporal partition Tnext. Using the same spatial dot product operation, for grid node (xi, yj), the intensity value Enext(xi, yj) and probability density value Snext(xi, yj) of the future temporal are read. The future defect initiation risk probability value Pnext(xi, yj) corresponding to this node is generated by the relationship Pnext(xi, yj) = Enext(xi, yj) × Snext(xi, yj).

[0085] This operation is performed by traversing all grid nodes, and the resulting new spatial distribution map is the probability distribution map Pnext of future defect initiation risk. This distribution map depicts the spatial distribution pattern of defect risk within the connected domain in subsequent process stages if no intervention is applied.

[0086] Step 144: Based on the current defect spawning risk probability distribution map and the future defect spawning risk probability distribution map, perform spatiotemporal propagation analysis of risk probability, calculate the gradient propagation direction and propagation rate of defect spawning risk probability in space, and obtain the defect spawning risk probability propagation trend vector field. The defect spawning risk probability propagation trend vector field includes the propagation direction vector and propagation rate value of each spatial coordinate point.

[0087] To analyze the propagation dynamics of risk from the current moment to future moments, it is necessary to analyze two risk probability maps at different time states: Pcur and Pnext. First, the spatial gradient field of the risk probability value is calculated. For each node position (xi, yj) in the current defect priming risk probability distribution map Pcur, its spatial gradient vector is calculated using the central difference method or the Sobel operator. This gradient vector can be decomposed into a component Gx(xi, yj) along the x-direction and a component Gy(xi, yj) along the y-direction. The gradient magnitude is denoted as Gmag(xi, yj), and its direction is the direction in which the risk probability increases most rapidly.

[0088] Secondly, the time change of the risk probability value is calculated. At the same node position (xi, yj), the difference between the future probability value and the current probability value is calculated, i.e., ΔP(xi, yj) = Pnext(xi, yj) - Pcur(xi, yj). The spatial gradient vector is combined with the time change ΔP to estimate the risk propagation vector. The propagation direction vector Dvec(xi, yj) is defined as the unit vector of the spatial gradient, with components Dx = Gx / (Gmag + ε) and Dy = Gy / (Gmag + ε), where ε is a small positive constant to prevent division by zero errors. The propagation rate Vrate(xi, yj) is proportional to the absolute value of the time change ΔP and inversely proportional to the spatial gradient magnitude Gmag, expressed as Vrate(xi, yj) = |ΔP(xi, yj)| / (Gmag(xi, yj) × ΔT + ε). Here, ΔT is the nominal time interval between the temporal partitions Tcur and Tnext. This calculation is performed on all nodes in the mesh to generate a vector field jointly defined by the propagation direction vector Dvec and the propagation rate value Vrate at each node. This vector field is the vector field representing the propagation trend of the probability of defect initiation risk.

[0089] Step 145: Based on the propagation trend vector field of the defect germination risk probability, perform propagation extension processing on the current defect germination risk probability distribution map, and spatially extrapolate the defect germination risk probability at the current moment along the propagation direction vector with the propagation rate value to generate a defect germination risk probability propagation distribution map that covers the continuous distribution of the defect germination risk probability in the time period between the current moment and the future moment.

[0090] Using the propagation trend vector field obtained in step 144, the probability distribution of defect germination risk at the current moment is extrapolated and interpolated in the time dimension to reconstruct the continuous risk field over the entire time period. For any intermediate time point Tmid between the current moment Tcur and the future moment Tnext, the risk probability value Pprop(xi, yj, Tmid) at the spatial node (xi, yj) is obtained by backtracking along the reverse path of the propagation vector.

[0091] The specific construction method is as follows: For a node (xi, yj) and an intermediate time Tmid, define the backtracking vector Bvec = Vrate(xi, yj) × (Tmid - Tcur) × Dvec(xi, yj). Then the backtracked source position coordinates (xsrc, ysrc) can be given by xsrc = xi - Bx and ysrc = yj - By, where Bx and By are the components of the backtracking vector Bvec in the x and y directions, respectively. The value of Pprop(xi, yj, Tmid) is obtained by bilinear interpolation at the source position coordinates (xsrc, ysrc) in the current distribution map Pcur. By performing the above spatial extrapolation and interpolation operations on a series of densely sampled intermediate times, a set of temporally quasi-continuous risk probability snapshot sequences can be generated. By spatially mosaicking and combining this set of sequences, a defect priming risk probability propagation distribution map Pprop, which contains propagation path information and reflects the dynamic evolution of risk throughout the entire time period, is formed.

[0092] Step 146: Perform defect category tendency mapping processing on the defect emergence risk probability propagation distribution map, match the defect emergence risk probability value of each spatial coordinate point in the defect emergence risk probability propagation distribution map with the preset defect category association rules, determine the defect category tendency code corresponding to each spatial coordinate point, and generate a defect category tendency code distribution map.

[0093] In addition to assessing the probability of risk occurrence, it is also necessary to predict the most likely type of defect that the risk point will develop into, such as microcracks or pores. To this end, a set of defect category association rules is pre-established. This rule can be a decision tree classification model trained with historical defect cases. The input feature vector of this model contains attributes extracted from multi-source physical fields at a certain node location, such as the average value of the defect initiation risk probability value Pprop(xi, yj, Tmid) at that node, the consistency metric parameter between the gradient direction of the energy accumulation tendency intensity map at that node and the direction of the stress concentration kernel, and the kurtosis statistics of the interface free energy fluctuation characteristics, etc. After the decision tree model is trained, each leaf node corresponds to a defect category tendency code, such as CODE_C representing microcracks and CODE_P representing pores.

[0094] During the mapping process, for each spatial coordinate point on the defect spawning risk probability propagation distribution map Pprop, the aforementioned input features are extracted for that point and fed into the decision tree model. The model assigns the point to a leaf node based on the feature's decision path and outputs the defect category propensity code corresponding to that leaf node. This matching process is completed across all spatial coordinate points in the entire map, ultimately forming a defect category propensity code distribution map Ctype that is perfectly aligned with the Pprop space and filled with category codes.

[0095] Step 147: Spatial coordinate alignment and fusion of the defect germination risk probability propagation distribution map and the defect category tendency coding distribution map. Using the spatial coordinate grid as a reference, the defect germination risk probability value and defect category tendency coding value of each grid node are combined and stored to generate a defect germination risk cloud map. The defect germination risk cloud map includes the defect germination probability value distribution and defect category tendency coding distribution indexed by spatial coordinates.

[0096] This step is the final data fusion and encapsulation operation. A unified spatial coordinate grid covering the entire billet surface is established. For a node on this grid with a location index of (xi, yj), its corresponding risk probability value is indexed from the defect initiation risk probability propagation distribution map Pprop, denoted as Pprop(xi, yj); simultaneously, its corresponding category code is indexed from the defect category tendency coding distribution map Ctype, denoted as Ctype(xi, yj). These two values ​​are combined into a binary data set {Pprop(xi, yj), Ctype(xi, yj)}. The collection of such spatiotemporally aligned binary data sets for all grid nodes together constitutes a complete multidimensional defect initiation risk cloud map. This cloud map is essentially a lookup table or multidimensional array that can be directly addressed by spatial coordinates. By inputting the spatial coordinates (xi, yj), both the defect initiation probability value at that point and the most likely defect category code at that point can be returned simultaneously, realizing the integrated expression of multimodal risk information.

[0097] Step 150: Combine the defect initiation risk cloud map and the spatial change rate of the interface tension gradient in the interface tension field distribution information to determine the defect evolution probability distribution within a preset time period, and generate a defect prediction instruction set based on the defect evolution probability distribution.

[0098] This step is the final decision-making stage. It combines the risk cloud map with the rate of change of the original macroscopic physical quantity of interface tension field to make the final prediction of the probability of defect evolution, and encodes the prediction information into executable machine instructions. The implementation of step 150 includes steps 151 to 155.

[0099] Step 151: Extract the defect emergence probability value distribution data from the defect emergence risk cloud map, perform spatial clustering processing on the defect emergence probability value distribution data, identify spatially continuous regions where the defect emergence probability value exceeds the preset risk concern threshold, and obtain a risk defect emergence region set. Each risk defect emergence region has a region outline boundary and a defect emergence probability value distribution within the region.

[0100] First, the defect initiation probability distribution data is separated from the defect initiation risk cloud map. A preset risk concern threshold, TRisk, is set, which can be determined by operators based on quality standards or historical statistical analysis. All grid nodes in the probability distribution data with a probability value P greater than TRisk are marked as high-risk points. Then, spatial clustering is performed on these discrete high-risk points. For example, a density-based noisy spatial clustering algorithm is used, which aggregates high-risk points spatially separated by less than one neighborhood radius Eps into clusters, each cluster corresponding to a risky defect initiation region. The algorithm automatically delineates the region's boundary outline and extracts the defect initiation probability values ​​of all grid points within that outline, forming the internal defect initiation probability distribution of that region.

[0101] Step 152: For each risk defect origination area, extract the defect category tendency code distribution data corresponding to the area from the defect origination risk cloud map, count the defect category tendency code value that appears most frequently in the area, take it as the dominant defect category tendency code of the risk defect origination area, and calculate the frequency ratio of the dominant defect category tendency code as the category consistency.

[0102] For each risk defect initiation region identified in step 151, its outline boundary is used as a mask to extract all category codes within that region from the defect category tendency coding distribution layer of the defect initiation risk cloud map. Frequency statistics are performed on this local code set, and the codes are sorted in descending order of occurrence. The code with the highest frequency, such as CODE_C, is designated as the dominant defect category tendency code for that risk defect initiation region.

[0103] Simultaneously, the ratio of the frequency of occurrence of the dominant code CODE_C to the total number of grid points in the region is calculated; this ratio represents the category consistency. High consistency indicates that the potential defect type in the region is relatively singular and definite, while low consistency indicates that multiple defects may coexist or there is significant uncertainty.

[0104] Step 153: For each risk defect initiation area, extract the spatial change rate of the interface tension gradient corresponding to the area from the interface tension field distribution information, perform overlay analysis on the spatial distribution of the interface tension gradient spatial change rate and the regional contour boundary of the risk defect initiation area, calculate the statistical mean and degree of variation of the spatial change rate of the interface tension gradient within the region, and obtain the regional interface tension gradient driving strength index.

[0105] The spatial rate of change of the interfacial tension gradient is a macroscopic mechanical factor driving defect evolution. Returning to the interfacial tension field distribution information obtained in step 110, the amplitude of the interfacial tension gradient G(x, y) at various points on the entire billet surface is calculated using spatial difference operations.

[0106] Then, the contour boundaries of each risk defect initiation region are spatially overlaid to extract the gradient magnitude value G of each grid point within that region. The statistical mean G of these G values ​​is calculated as a measure of the average driving force intensity of the region. Simultaneously, its coefficient of variation is calculated, for example, the standard deviation of the G values ​​divided by the G mean, as a measure of the degree of variation. The G mean and the coefficient of variation are combined according to preset weights to form a comprehensive index, which is the regional interfacial tension gradient driving intensity index for that region. This index comprehensively reflects the potential energy and uniformity of defect evolution driven by the macroscopic stress field in that region.

[0107] Step 154: Based on the distribution of defect emergence probability values ​​within each risk defect emergence area, the dominant defect category tendency coding, category consistency, and the regional interface tension gradient driving intensity index, construct a defect evolution dynamics prediction mapping relationship. Using a defect evolution dynamics prediction mapping function pre-established based on historical defect cases, generate the defect evolution probability value and defect evolution category evolution trend of the risk defect emergence area within a preset time period, and obtain the regional defect evolution probability distribution.

[0108] This step is the core of the prediction. A defect evolution dynamics prediction mapping function is trained in advance using a large number of labeled historical defect case samples. The model architecture of this function can be a GradientBoosting ensemble tree model.

[0109] In application, the input feature vector of the function consists of four parts: statistical features extracted from the distribution of defect initiation probability values ​​within the region (such as the maximum, mean, and quantile of P); a one-hot vector encoding the dominant defect category tendency; a category consistency value; and a regional interfacial tension gradient-driven intensity index. The output of this mapping function includes two parts: the first part is a scalar representing the probability of defect occurrence in the region within a preset time period; the second part is a vector reflecting the category evolution trend, for example, [0.80, 0.15, 0.05], representing the probability distribution of evolution into microcracks, pores, or other categories, respectively. For each risk defect initiation region, after organizing the above input features, feeding them into the mapping function, we can obtain its specific regional defect initiation probability value and defect initiation category evolution trend. The above results for all regions are summarized to form the regional defect initiation probability distribution.

[0110] Step 155: Based on the regional defect evolution probability distribution, determine the defect evolution probability distribution prediction map, and use the defect evolution probability distribution prediction map to determine the spatial positioning coordinate sequence of the risk area, the defect category identification code sequence, and the confidence rating identification sequence, and encapsulate them into a defect prediction instruction set.

[0111] The above step 155 implements the final instruction generation, which is specifically completed by steps 1551 to 1554.

[0112] Step 1551: Spatially summarize the regional defect evolution probability distribution of all risk defect initiation areas, and spatially stitch together the defect evolution probability values ​​and defect evolution category evolution trends of each region according to the regional outline boundary to generate a defect evolution probability distribution prediction map covering the entire surface of the semi-cured sheet blank.

[0113] The discrete prediction results for different risk defect initiation regions calculated in step 154 ​​are mapped back to the spatial coordinate grid of the entire billet. On a blank full-surface grid map aligned with the billet surface, the defect evolution probability value of each region is filled into the grid nodes covered by the region's outline. For the category evolution trend vector, the values ​​of each dimension are also filled into the different channels of the corresponding grid nodes. For background grid nodes not covered by any risk regions, their defect evolution probability value is set to zero, and the category evolution trend is set to background or "no risk" category. This generates a spatially continuous and complete defect evolution probability distribution prediction map covering the entire surface of the prepreg billet.

[0114] Step 1552: Perform risk level classification processing on the defect evolution probability distribution prediction map, divide the surface of the semi-cured sheet blank into multiple risk level blocks according to the distribution range of the defect evolution probability value, and extract the spatial positioning coordinate sequence of each risk level block to generate the risk area spatial positioning coordinate sequence.

[0115] Based on several preset risk level thresholds, such as T1 for medium risk and T2 for high risk, the probability values ​​on the defect evolution probability distribution prediction map are divided into intervals. The range from 0 to T1 is low risk, T1 to T2 is medium risk, and above T2 is high risk. The probability distribution prediction map is then binarized or segmented, and connected component analysis is used to find the complete outline of each risk level (especially medium and high risk) block.

[0116] Then, the spatial coordinates of each contour boundary point, or the coordinates of the polygon vertices, are extracted in vector form to form a spatial positioning coordinate sequence representing the risk level area. Thus, several risk areas of different levels are assigned their precise contour coordinate sequences.

[0117] Step 1553: Assign defect category identifier codes according to the expected defect category codes corresponding to the spatial positioning coordinate sequence of each risk area, and assign confidence rating identifiers according to the defect evolution probability value of the area, thereby obtaining defect category identifier code sequences and confidence rating identifier sequences.

[0118] For each risk region identified in step 1552, locate the set of grid nodes covering that region in the defect evolution probability distribution prediction map. Within this node set, determine which dimension in the category evolution trend vector has the highest average probability value; the defect category corresponding to that dimension is then encoded as the expected defect category for that region. Assign a formal defect category identifier code to this region, such as ID_C.

[0119] Simultaneously, the average defect evolution probability value of all nodes within the region is extracted. This average value represents the overall risk severity of the region and is used as the basis for assigning confidence rating labels. For example, based on a mapping table, the rating label for an average probability P value in the interval [0.8, 1.0] can be set as "Level A", [0.6, 0.8] as "Level B", and so on. The label codes and rating labels for all regions are arranged according to spatial correspondence, forming a defect category label code sequence and a confidence rating label sequence, respectively.

[0120] Step 1554: Encapsulate the spatial positioning coordinate sequence of the risk area, the defect category identification code sequence, and the confidence rating identification sequence into an instruction format to generate a defect prediction instruction set. Each instruction in the defect prediction instruction set corresponds to a risk area and includes the spatial positioning coordinates, defect category identification code, and confidence rating identification of the risk area.

[0121] Iterate through all identified risk areas, generating one independent pre-judgment instruction for each area. Each instruction is a structured data object, such as a JSON object, containing three key-value pairs: the key "coord" represents the spatial coordinate sequence of the risk area; the key "defect_type" represents the defect category identifier for the area; and the key "confidence" represents the confidence rating for the area. All risk area instruction objects are packaged into an array or list. This array or list data set constitutes the final defect pre-judgment instruction set, which can be used to drive downstream sorting equipment, issue alarms, or provide precise spatial guidance for process parameter adjustments.

[0122] Based on steps 110-150, as an optional embodiment, the method further includes: Step 161: Obtain the spatial distribution dataset of process conditions for subsequent processes that have a temporal connection with the current process during the manufacturing of copper clad laminate prepreg.

[0123] In a preferred embodiment, defect prediction is not limited to the current process but also proactively considers the impact of subsequent processes on defect evolution. To this end, it is necessary to obtain a spatial distribution dataset of process conditions for subsequent processes, which can be obtained from a manufacturing execution system or a process simulation system. For example, for the pressing process, this dataset contains the temperature distribution matrix T at various locations on the blank surface applied by the pressing template. 压 (x, y) and pressure distribution matrix P 压 (x, y), this dataset details the intensity and distribution of physical conditions acting on different spatial locations of the billet in subsequent processes.

[0124] Step 162: Transform the spatial positioning coordinate sequence of each risk area marked in the defect prediction instruction set from the current process coordinate system to the equivalent spatial coordinate system of the subsequent process to obtain the spatial alignment result of the risk area across processes.

[0125] Because the billet may undergo thermal deformation, traction stretching, or changes in positional reference between different processes, spatial coordinate transformation is required. An affine transformation matrix is ​​established between the coordinate system of the current process and the coordinate system of subsequent processes. This matrix can be calculated by measuring the coordinate pairs of several known reference points on the billet in the two processes. Then, the spatial positioning coordinate sequence of the risk area carried by each instruction in the defect prediction instruction set is transformed one by one through this affine transformation matrix, thereby obtaining the new coordinates of the risk area in the space of the subsequent processes. This is the spatial alignment result of the risk area across processes, ensuring that the physical position of the risk area is accurately tracked between different processes.

[0126] Step 163: Based on the spatial alignment result, extract the local temperature intensity sequence and local pressure intensity sequence corresponding to each risk region from the spatial distribution dataset of process conditions to generate the thermodynamic driving condition features specific to the risk region.

[0127] Using the spatial alignment result obtained in step 170 as a spatial filter or query mask, the spatial distribution dataset T of process conditions is applied. 压 (x, y) and P 压 The query is performed at (x, y). For a risk area, the temperature values ​​of all grid points within the area's contour are extracted to form a local temperature sequence. This temperature sequence is then subjected to feature engineering, such as calculating its spatial mean, maximum spatial gradient, and the percentage of area exceeding a certain critical temperature threshold, to form a multi-dimensional local temperature intensity feature vector. Similarly, the pressure value sequence of the same area is extracted, and its spatial mean and pressure non-uniformity are calculated to form a local pressure intensity feature vector. The two feature vectors are then concatenated to obtain the thermodynamic driving condition features specific to this risk area.

[0128] Step 164: Retrieve the spatiotemporal evolution pattern of energy accumulation tendency of Rayleigh instability perturbation connected domains associated with each risk region in the semantic feature description of defect evolution.

[0129] In the semantic feature description of defect evolution, each risk region has a Rayleigh instability perturbation connected component from which it originates. By using pre-established tracing relationships or spatial inclusion relationships, the identifiers of one or more connected components associated with each risk region are precisely located. Then, from the defect evolution semantic feature description dataset, the energy accumulation tendency intensity evolution trajectory and stress concentration kernel germination probability density evolution trajectory of the corresponding connected component stored in the temporal sequence are directly retrieved, i.e., its spatiotemporal evolution mode of energy accumulation tendency.

[0130] Step 165: Perform inter-process defect evolution driving force coupling analysis on the thermodynamic driving condition characteristics and the spatiotemporal evolution mode of energy accumulation tendency to generate a distribution map of the enhancement or weakening tendency of defect evolution potential in each risk area under the influence of subsequent processes. Based on the distribution map of the enhancement or weakening tendency of defect evolution potential, adjust the gain or attenuation of the corresponding confidence rating label in the defect prediction instruction set and generate additional prediction instructions. Merge the additional prediction instructions with the defect prediction instruction set to obtain a prediction instruction group for multi-process serial manufacturing.

[0131] The core of coupling analysis is to determine whether the thermodynamic driving force of subsequent processes exacerbates or inhibits defect evolution. This can be achieved through a pre-trained inter-process coupling impact assessment network. The network's input consists of several key time-node states extracted from the spatiotemporal evolution patterns of thermodynamic driving conditions and energy accumulation tendencies. The output is a strengthening or weakening tendency value, with positive values ​​representing strengthening and negative values ​​representing weakening, and the magnitude representing the degree of influence. This assessment is performed on all risk areas to generate a distribution map of the strengthening or weakening tendency of defect evolution potential.

[0132] Based on the trend distribution map, the confidence rating labels in the original defect prediction instruction set are dynamically adjusted. For example, for a risk area where the enhancement trend value is greater than the threshold T for enhancement 0, its confidence rating label is upgraded from "Level B" to "Level A", and the original instruction is marked as needing to be added. For the above-mentioned risk areas that need to be added, additional prediction instructions are generated, which include the updated confidence rating label and the subsequent process label that caused the adjustment. Finally, the original defect prediction instruction set and the newly generated additional prediction instructions are merged to form a prediction instruction group with richer content and coverage of multiple processes.

[0133] In yet another alternative embodiment, after step 150, the method further includes: Step 210: Obtain a set of historical defect prediction instructions for multiple batches of copper-clad laminate prepregs that have been manufactured. Each set of historical defect prediction instructions is associated with a historical defect evolution probability distribution prediction map.

[0134] In another preferred embodiment, to incorporate batch experience, a historical case library is first established. Historical data from multiple previous manufacturing batches is extracted from the database. This data includes two parts: first, a set of defect prediction instructions generated at that time, serving as labels for the samples; and second, a predicted defect evolution probability distribution map generated at that time, associated with these instructions, serving as the feature distribution of the samples.

[0135] Step 220: Spatial normalization processing is performed on the prediction maps of the historical defect evolution probability distributions to map the risk distributions of different sheet surfaces onto a unified standard spatial grid, thereby obtaining a set of historical standard probability distribution maps.

[0136] Different sheet materials may have slight differences in size and shape. Therefore, for each historical defect evolution probability distribution prediction map, spatial normalization methods such as affine transformation or thin-plate spline interpolation are used to map its risk probability distribution pattern onto a standardized spatial grid of a fixed size, such as a 1000×1000 grid. All the mapped, standardized images constitute a historical standard probability distribution map set.

[0137] Step 230: In the unified standard spatial grid, the historical standard probability distribution map is subjected to agglomerative clustering processing of spatial distribution morphology. The center of each morphological cluster formed by the clustering is defined as the common spatial pattern prototype of defect evolution, and a common pattern prototype set is generated.

[0138] In a standardized spatial grid, each standardized historical probability map is considered a high-dimensional vector. Agglomerated clustering algorithms, such as Ward's agglomerated clustering, are used to cluster these vectors. The algorithm starts with each sample as an independent cluster, merging the two clusters that minimize the increase in total variance each time, until a preset number of clusters or a merging distance threshold is reached. After clustering, for each resulting morphological cluster, the average of all sample vectors within that cluster is calculated and used as the cluster center. This center vector, when reconstructed onto the spatial grid, represents a typical probability distribution pattern, defined as a prototype of a common spatial pattern for defect evolution. The set of all prototypes constitutes the common pattern prototype set, with each prototype representing a historically frequent and typical defect risk distribution pattern.

[0139] Step 240: Use the defect prediction instruction set to backtrack and obtain the current defect evolution probability distribution prediction map of the current semi-cured sheet blank to be tested, and transform the current defect evolution probability distribution prediction map to the unified standard space grid to generate the current standard probability distribution map.

[0140] The current defect prediction result of the sheet is traced back to the source data on which the above instructions were based, namely the defect evolution probability distribution prediction map. Then, using the same spatial normalization transformation method as in step 220, the map is also mapped onto the same 1000×1000 unified standard spatial grid to obtain the current standard probability distribution map of the current sheet.

[0141] Step 250: Perform spatial morphological similarity measurement between the current standard probability distribution map and each defect evolution public spatial pattern prototype in the public pattern prototype set, and determine the matching spatial pattern prototype that is most consistent with the current defect evolution distribution morphology.

[0142] The current standard probability distribution vector is compared with each of the prototype vectors in the common pattern prototype set for similarity measurement. A correlation-based method can be used, such as calculating the Pearson correlation coefficient. The closer the coefficient is to 1, the more similar the spatial patterns of the two distributions are. The prototype with the highest correlation coefficient is selected as the best-matching spatial pattern prototype for the current sheet material.

[0143] Step 260: Extract the confidence distribution bias features and defect category transfer path features of the historical risk area corresponding to the matching spatial pattern prototype from the historical defect prediction instruction set sample.

[0144] In historical data, we trace all historical billet samples belonging to the matching spatial pattern prototype. From their corresponding historical defect prediction instruction sets, we compile and statistically analyze the distribution tendency of their confidence ratings in similar risk areas, such as whether they have historically been rated as high-risk or medium-risk; this is the confidence distribution bias characteristic. Simultaneously, we examine the probability of similar defect categories evolving into other categories (such as from microcracks to delamination) in subsequent processes or final product inspections within these historical samples; this is the defect category transfer path characteristic.

[0145] Step 270: Statistically correct the confidence rating identifier sequence and defect category tendency coding sequence of each risk area in the defect prediction instruction set according to the confidence distribution bias feature and defect category transfer path feature, and add the identifier code of the matching space pattern prototype to the defect prediction instruction set to generate a compensated defect prediction instruction set that integrates batch experience.

[0146] By leveraging trend information extracted from historical experience, current predictions are revised. For each instruction in the current defect prediction instruction set, based on its location, the confidence distribution bias characteristics are examined, and the confidence rating can be fine-tuned to align with historical bias levels, such as increasing or decreasing it by one level. Simultaneously, based on the defect category trend code for that region, defect category transfer path characteristics are examined. If a category has a high probability of transferring to another defined category, the potential target category code is appended to the defect category label of the instruction. Finally, an additional field is added to the metadata of the instruction set to record the identifier code of the matching spatial pattern prototype. The final instruction set formed after these revisions and additions is the compensated defect prediction instruction set that incorporates batch experience, resulting in more robust and accurate predictions.

[0147] In one embodiment of the present invention, the defect evolution dynamics prediction mapping function mentioned in step 154 ​​above, which is pre-established based on historical defect cases, specifically adopts a gradient boosting tree ensemble model in its core architecture, for example, using the XGBoost framework. This model is composed of multiple weak regression trees stacked sequentially, with each new tree fitting along the residual descent gradient direction of the previous tree.

[0148] Key implementations within the model include: using mean squared error as the loss function for splitting nodes, setting the maximum tree depth parameter to 7, setting the minimum sample weight of child nodes to 1.5 to control overfitting, and setting the column sampling ratio to 0.8 to increase sub-model diversity. During training, the dataset used is derived from several batches of copper-clad laminate prepreg samples accumulated on the production line, with a scale of no less than tens of thousands of groups. Each input sample is a multi-dimensional feature vector, whose fields consist of: the gradient mean and coefficient of variation in the regional interface tension gradient-driven intensity index obtained in step 153; the one-hot encoded vector of the dominant defect category tendency encoding obtained in step 152 and the corresponding category consistency; and the statistics of the distribution of defect germination probability values ​​within the region obtained in step 151, including probability maxima, probability mean, and probability quantiles. The output label of the sample contains a scalar defect evolution probability value and a multi-dimensional vector representing the category evolution trend. The Adam optimizer was used during training, with the initial learning rate set between 0.01 and 0.1, the maximum number of boost epochs set to 150, and an early stopping mechanism enabled. Training was terminated when the mean absolute error on the validation set did not decrease within 15 consecutive epochs.

[0149] In the inference application phase, for a new risk defect initiation region, the statistical characteristics of the defect initiation probability distribution within the region, the dominant defect category tendency encoding, category consistency, and the regional interface tension gradient driving strength index are concatenated into an input feature vector in the same field order as during training. This feature vector is then fed into the pre-trained XGBoost model. The model traverses all regression trees, accumulates and sums the scores, and finally obtains the scalar value of the defect evolution probability in the region from the model's output. Simultaneously, the multi-objective output is normalized using the softmax function to obtain the probability vector of the defect category evolution trend. This model architecture and application method ensure the interpretability of the prediction process and effectively capture the nonlinear complex relationships between multi-source heterogeneous features, meeting the stringent requirements for accuracy and stability in industrial defect prediction scenarios.

[0150] Please refer to the following: Figure 1 and Figure 2 The method for predicting defects in copper-clad laminate prepregs based on interface semantic features disclosed in this invention can be referred to in its overall technical approach. Figure 2 To understand this, the overall solution begins with obtaining information on the interfacial tension field distribution on the surface of the prepreg blank to be tested, such as... Figure 2 As shown, this is the data source of the entire prediction process. The interface tension field distribution information is obtained through a non-contact sensing unit and structured into a two-dimensional dataset containing the spatial coordinates of each sampling point, the amplitude of the interface tension, and the direction angle.

[0151] Subsequently, this distribution information is sent to the "micro-area physical disturbance decoupling" processing stage. For example... Figure 2 As shown, this step aims to separate microscopic perturbation information of latent defect precursors from the macroscopic interfacial tension field. Specifically, the surface is divided into an array of micro-units with spatial location indices based on the surface texture direction of the billet. For each micro-unit, a local weighted regression algorithm is used to construct a steady-state reference surface representing the ideal distribution of the interfacial tension without perturbation. Then, by performing amplitude and direction difference calculations between the original interfacial tension vectors at each sampling point and the steady-state reference surface, local perturbation residual components are decomposed. Based on this, a correlation mapping is performed between the perturbation amplitude residual and the surface tension coefficient to generate local interfacial free energy fluctuation features representing energy fluctuations and aggregation tendencies. Simultaneously, by calculating the spatial divergence of the perturbation direction residual field, the convergence region of the perturbation direction where the vector field converges is identified, and stress concentration kernel tendency features are generated based on its area connectivity and convergence density. These two types of features are associated and encapsulated with the spatial location indices of the micro-units, and all micro-units are traversed to aggregate and generate a set of semantic features of latent precursors in the micro-regions, thus completing the first feature abstraction from the original physical field to structured semantic features.

[0152] After obtaining the semantic feature set of latent precursors in micro-regions, the technical solution proceeds... Figure 2 The "Rayleigh instability texture" processing stage, as indicated in the text, focuses on establishing a spatial correlation between energy and stress perturbations across the boundaries of isolated micro-regions. For example... Figure 2 As shown, this stage integrates discrete micro-units into Rayleigh unstable perturbation connected domains with unified perturbation growth conditions by mining implicit associations between adjacent micro-units that meet the Rayleigh instability perturbation growth criterion. Specifically, a micro-unit adjacency graph is first constructed based on the spatial adjacency relationships of the micro-units. Then, for each pair of micro-units connected by edges in the graph, it is determined whether the difference in their interface free energy fluctuations and the consistency of the stress concentration kernel direction simultaneously satisfy a preset criterion. Edges that meet the conditions are marked as Rayleigh unstable micro-unit associated edges. The above associated edges are then aggregated into multiple non-overlapping spatially continuous connected domains using a connected component extraction algorithm. For each connected domain, cumulative processing of energy accumulation tendency and kernel density estimation of stress concentration kernel germination probability are performed, generating their respective energy accumulation tendency intensity distribution map and stress concentration kernel germination probability distribution map. Furthermore, temporal boundary constraint parameters corresponding to the manufacturing process stages are introduced, and the two distribution maps are divided into temporal partitions on the time axis. The two maps in the same temporal partition are spatially aligned and superimposed and connected in multiple temporal phases to construct a spatiotemporal evolution model of energy accumulation tendency that reflects the linkage and evolution of energy and stress on the spatiotemporal scale. Finally, the models of all connected domains are integrated to form a semantic feature description of defect evolution, realizing the second abstraction from static semantic features to dynamic spatiotemporal evolution features.

[0153] Immediately afterwards, such as Figure 2 As shown in the flowchart, the semantic feature description of defect evolution is passed to the "defect germination risk cloud map" generation stage. This step transforms the spatiotemporal evolution pattern of the physical field into probability and category information that can be directly used for risk decision-making. First, temporal fusion is performed on each connected component to extract its energy and stress distribution maps in the current temporal state and adjacent future temporal states. Then, through dot product operations on each spatial point, the current defect germination risk probability distribution map and the future defect germination risk probability distribution map are generated respectively. On this basis, spatiotemporal propagation analysis of risk probability is performed. By calculating the relationship between spatial gradient and time change, a defect germination risk probability propagation trend vector field containing propagation direction vector and propagation rate value is constructed. This vector field is then used to spatially extrapolate the current risk probability to generate a defect germination risk probability propagation distribution map covering the entire time period. Finally, each spatial point in the propagation distribution map is mapped to a category using preset defect category association rules, and spatial coordinates are aligned and fused with the risk probability value to generate a defect germination risk cloud map that simultaneously contains the distribution of defect germination probability values ​​and the distribution of defect category tendency codes, completing the third abstraction of risk decision-making information.

[0154] The final predictive decision-making stage, such as Figure 2As shown at the end, the "defect initiation risk cloud map" is fused with the macroscopic mechanical information from the original "interfacial tension field distribution" to determine the defect evolution probability distribution and generate predictive instructions. The specific process is as follows: probability values ​​are extracted from the defect initiation risk cloud map and spatial clustering is performed to identify risk defect initiation regions, determining the dominant defect category code and category consistency for each region. Simultaneously, the spatial change rate of the interfacial tension gradient is extracted from the interfacial tension field distribution information, and the gradient driving strength index for each region is calculated through overlay analysis. These features are input into a defect evolution dynamics prediction mapping function pre-established based on historical defect cases to generate the defect evolution probability value and category evolution trend for each risk region within a preset time period, thus obtaining the regional defect evolution probability distribution. Based on this, through a series of operations such as spatial aggregation, risk level classification, code allocation, and instruction format encapsulation, a set of defect predictive instructions is finally generated, containing a sequence of spatial location coordinates for risk regions, a sequence of defect category identification codes, and a sequence of confidence rating identifiers, providing accurate and executable decision-making basis for quality control in the manufacturing process.

[0155] To facilitate a clearer and more complete implementation of the embodiments of the present invention, the above-mentioned parts are further described in detail below.

[0156] In the specific implementation of acquiring interfacial tension field distribution information, when using an optical measurement mode based on digital image correlation, a random speckle pattern is pre-prepared on the surface of the prepreg blank to be tested. This speckle pattern is formed by randomly distributing titanium dioxide particles with a particle size of 0.5 to 5 micrometers on the blank surface using a vapor deposition method, with the speckle particle density controlled to be no less than 200 particles per square millimeter. Using a high-speed industrial camera with a resolution of 2448×2048 pixels, and a telecentric lens with a magnification of 2x to 5x, digital image sequences of the blank surface are continuously acquired at a acquisition rate of 30 frames per second under uniform illumination conditions of a ring LED cold light source. For two adjacent frames of images, a subset matching is performed using the zero-mean normalized cross-correlation criterion, with the subset size set to 29×29 pixels and the matching step size set to 1 pixel, to obtain the in-plane displacement field between the two adjacent frames. The in-plane strain field is obtained by numerical differentiation of the displacement field using the central difference method, with the differentiation step size consistent with the subset matching step size.

[0157] Based on the obtained in-plane strain field, the interfacial tension field is inverted using an orthogonal anisotropic linear elastic constitutive relation. In this relation, the meridional elastic modulus, latitudinal Poisson's ratio, latitudinal Poisson's ratio, and in-plane shear modulus of the reinforcing material are pre-calibrated through independent uniaxial and off-axis tensile tests. These parameters are substituted into the incremental form of the constitutive equation, and the interfacial tension vector is solved point-by-point. Its amplitude F represents the maximum principal stress at that point, and its direction angle θ represents the principal direction angle of that maximum principal stress. The interfacial tension vectors of all sampling points are arranged in spatial coordinates to constitute the interfacial tension field distribution information. When using an acoustic measurement mode based on ultrasonic surface wave reflection, a point-focusing ultrasonic transducer with a center frequency of 10 MHz to 50 MHz is used to perform two-dimensional grid scanning on the surface of the billet at a preset scanning step interval. The time-domain waveform of the surface wave reflection signal is acquired at each scanning point. The near-surface residual stress is inferred by calculating the change in wave velocity of the surface wave propagating on the surface of the billet. Then, the wave velocity change is mapped to the interfacial tension amplitude by combining the acoustoelastic coefficient of the billet material. The direction of the interfacial tension is determined by the acoustoelastic relationship between the principal stress direction and the surface wave propagation direction. This acoustoelastic coefficient is obtained in advance by measuring the surface wave velocity under different stress states on a calibration sample made of the same material and process as the billet to be tested.

[0158] When performing micro-region physical perturbation decoupling processing on the interfacial tension field distribution information, the specific implementation method of dividing the micro-region units according to the surface texture direction of the prepreg blank in step 121 is as follows: First, calculate the gradient direction of each pixel point in the blank surface image, and generate a gradient direction histogram. The histogram divides the range from 0 degrees to 180 degrees into 36 equal intervals. Identify the directions corresponding to the two peaks in the histogram as the main longitudinal and latitudinal texture directions. Establish an orthogonal grid coordinate system with the main longitudinal texture direction as the row direction and the main latitudinal texture direction as the column direction. Preset the micro-region scale to include 32 sampling points in the row direction and 32 sampling points in the column direction. Divide the entire sampling point layout of the blank into multiple adjacent micro-region units with overlapping boundaries according to this scale. The actual size of the edge micro-region units with less than 32 sampling points at the boundary is retained. Assign a spatial position index consisting of row-order integers and column-order integers to each micro-region unit.

[0159] In step 122, when constructing the steady-state reference surface of interfacial tension using the local weighted regression algorithm, the weighting function is a cubic kernel function, which has the form w(r) = (1-r^3)^3, where r is the normalized geometric distance from other points within the window to the center point. The half-width of the window is set to 0.4 times the smaller value of the number of rows and columns of sampling points within the micro-area unit, and the order of the fitting polynomial is selected as second order. The calibration method for the surface tension coefficient gamma used in step 1241 is as follows: using the contact angle measurement method, a semi-cured sheet sample from the same batch as the blank to be tested is selected, and three standard test liquids with different surface tensions are dropped onto its surface. The static contact angle of each liquid on the sample surface is measured, and the dispersive component and polar component of the surface free energy of the sample are calculated according to the Owens-Wendt method. The sum of the two parts is taken as the value of the surface tension coefficient gamma. The generated local interface free energy fluctuation feature vector contains the following statistics: the mean, standard deviation, skewness, kurtosis, full width at half maximum (FWHM), and the proportion of high-value regions in the micro-region whose values ​​exceed the mean plus twice the standard deviation. These are concatenated in this order to form a six-dimensional feature vector.

[0160] In step 1242, when calculating the divergence of the disturbance direction residual vector field, a discrete divergence operator based on central difference is used. The specific method for setting the divergence threshold is to take the mean of the divergence values ​​of all sampling points within the micro-region cell minus one standard deviation. The connected region area is calculated based on the number of sampling points contained within the connected region. The convergence density is defined as the negative divergence integral of the unit vector of the direction residual of all sampling points within the connected region divided by the area of ​​the connected region. The stress concentration kernel tendency feature vector contains the maximum area, maximum convergence density, and total number of connected regions identified within the micro-region cell. If no convergence region is identified within the micro-region cell, all three items are set to zero.

[0161] When performing latent feature association processing of Rayleigh instability texture on the semantic feature set of latent precursors of micro-regions, the neighborhood rule of the micro-region adjacency relationship graph in step 131 adopts 4-neighborhood, that is, for a micro-region unit with index (u, v), its neighborhood set consists of micro-region units with indices (u-1, v), (u+1, v), (u, v-1), and (u, v+1). The specific criteria for determining the Rayleigh instability perturbation growth criterion in step 132 are as follows: Firstly, the mean component of the six-dimensional local interface free energy fluctuation feature vectors of each of the two adjacent micro-units is extracted, and the absolute value of the difference between them is calculated. If this absolute value exceeds a preset energy difference threshold, the energy difference condition is satisfied. This preset energy difference threshold is the upper quartile of the energy mean difference among all micro-units. Secondly, the maximum convergence density is extracted from the stress concentration kernel tendency feature vectors of each of the two adjacent micro-units. If both maximum convergence densities are not zero, the principal stress concentration directions are calculated. These principal stress concentration directions are taken as the vector average direction of the residuals of all perturbation directions within the corresponding connected domain. The angle between the two principal stress concentration directions is further calculated. If this angle is less than 45 degrees, the direction consistency condition is satisfied. When both the energy difference condition and the direction consistency condition are satisfied simultaneously, the pair of adjacent micro-units is determined to satisfy the Rayleigh instability perturbation growth criterion.

[0162] In step 133, the connected component extraction is implemented using a graph algorithm based on depth-first search. In step 1342, the nonparametric kernel density estimation uses a Gaussian kernel function, specifically in the form K(d) = (1 / √(2π))exp(-0.5d^2), where d is the Euclidean distance from a point in space to a data point divided by the bandwidth parameter, and the bandwidth parameter is the average of the nearest neighbor distances between the center points of all micro-cells within the Rayleigh unstable perturbation connected domain.

[0163] In step 1343, the temporal boundary constraint parameters are divided as follows: the prepreg manufacturing process is divided into five consecutive stages from the end of impregnation: drying and heating stage, drying and constant temperature stage, drying and cooling stage, cooling stage, and placement stage. The duration of each stage is determined according to the actual process parameter settings of the current batch. The end time of the current process stage of the blank to be tested is taken as the temporal base point. The preset time period covered after the temporal base point is divided into corresponding temporal partitions according to the boundaries of each process stage. When the preset time period spans multiple process stages, the temporal partitions are divided according to the actual start and end times of each stage. The energy accumulation tendency intensity distribution map and stress concentration nucleus germination probability distribution map in each temporal partition are obtained by time-weighted averaging and fusion of the effective distribution data within the time window of the temporal partition. The time weighting weight is proportional to the duration of each sub-time period within the temporal partition.

[0164] When performing defect initiation risk mapping based on the semantic feature description of defect evolution, the specific process of calculating the probability propagation trend vector field of defect initiation risk in step 144 is as follows: First, the spatial gradient components Gx and Gy in the x and y directions of each grid node are calculated using a 3×3 Sobel operator on the current defect initiation risk probability distribution map Pcur, with gradient magnitude Gmag = sqrt(Gx^2 + Gy^2); at the same grid node position of the current defect initiation risk probability distribution map Pcur and the future defect initiation risk probability distribution map Pnext, the time change ΔP = Pnext - Pcur is calculated; the components of the propagation direction vector Dvec are Dx = Gx / (Gmag + ε) and Dy = Gy / (Gmag + ε), where ε is 1.0 × 10^-8; the propagation rate Vrate = |ΔP| / (Gmag × ΔT + ε), where ΔT is the nominal time interval between the current time partition and the adjacent future time partition; the above calculations are performed point by point on all grid nodes to obtain the probability propagation trend vector field of defect initiation risk.

[0165] In step 145, when propagating and expanding the current defect priming risk probability distribution map, the intermediate time Tmid is taken as M time points uniformly sampled between the current time and future time, with M being 10. For each Tmid, the backtracking vector Bvec = Vrate(xi, yj) × (Tmid - Tcur) × Dvec(xi, yj) is calculated for the grid node (xi, yj), and the backtracking source position coordinates xsrc = xi - Bx and ysrc = yj - By are the components of the backtracking vector in the x and y directions, respectively. The value of Pprop(xi, yj, Tmid) is obtained by bilinear interpolation of the source position coordinates (xsrc, ysrc) in the current distribution map Pcur. The risk probability distribution snapshots of the M intermediate times are arranged in chronological order together with Pcur and Pnext to form the defect priming risk probability propagation distribution map.

[0166] In step 146, the defect category association rule is implemented using the C4.5 decision tree classification model. The input feature vector of this decision tree model consists of five attributes extracted from the multi-source physical field at the spatial coordinate point: the average risk probability value of the point across all time snapshots in the defect initiation risk probability propagation distribution map; the cosine of the angle between the gradient direction of the energy accumulation tendency intensity distribution map and the gradient direction of the stress concentration kernel initiation probability distribution map at the point; the kurtosis statistics of the interface free energy fluctuation metric at the point; the amplitude of the interface tension gradient at the point; and the total number of micro-region units within the Rayleigh instability perturbation connected domain where the point is located. During the training phase, the decision tree model uses a training dataset constructed from historical defect cases. The input feature vector of each sample in the training dataset corresponds to the above five attributes, and the sample label is the defect category code, where CODE_C corresponds to microcracks, CODE_P corresponds to pores, and CODE_D corresponds to stratification. In the application phase, the above five attribute values ​​are extracted for each spatial coordinate point on the defect initiation risk probability propagation distribution map and input into the decision tree model. The model outputs the defect category tendency code corresponding to that point. In step 147, when the probability propagation distribution map of defect initiation risk and the defect category tendency coding distribution map are fused together by spatial coordinate alignment, the resolution of the unified spatial coordinate grid in both the x and y directions is set to 0.1 mm.

[0167] When determining the probability distribution of defect evolution by combining the defect initiation risk cloud map and the spatial change rate of interface tension gradient, the risk concern threshold TRisk in step 151 is set to 0.35, the spatial clustering process adopts the density-based DBSCAN algorithm, the neighborhood radius Eps is set to 5 times the resolution of the unified spatial coordinate grid, i.e., 0.5 mm, and the minimum number of contained points MinPts is set to 10. The calculation method for the spatial change rate of the interfacial tension gradient in step 153 is as follows: return to the interfacial tension field distribution information obtained in step 110, and use the central difference method to calculate the gradient in the x-direction and the gradient in the y-direction for the interfacial tension amplitude F of each sampling point. The gradient amplitude G = sqrt((dF / dx)^2 + (dF / dy)^2); statistically analyze the gradient amplitude values ​​of all sampling points or grid nodes within the contour boundary of each risk defect initiation area, and calculate their statistical mean G_mean and coefficient of variation. The coefficient of variation is the standard deviation of the gradient amplitude value divided by the mean G. The calculation formula for the regional interfacial tension gradient driving strength index is: driving strength index = α × mean G / G_ref + (1-α) × coefficient of variation, where G_ref is the median of the gradient amplitude of the entire surface used for normalization, and the weighting coefficient α is 0.65.

[0168] The defect evolution dynamics prediction mapping function in step 154 ​​is implemented using the XGBoost gradient boosting tree ensemble model. This model consists of 200 regression trees stacked sequentially, with a maximum depth of 7 for each tree, a minimum sample weight sum of 1.5 for child nodes, and a column sampling ratio of 0.8. During model training, mean squared error is used as the loss function, the learning rate is set to 0.05, the maximum number of boosting rounds is set to 150, and an early stopping mechanism is enabled. Training terminates when the mean absolute error on the validation set does not decrease within 15 consecutive rounds. The training sample set is constructed from historical defect cases accumulated on the production line. The input feature vector fields for each sample group consist of: the gradient mean and coefficient of variation in the regional interface tension gradient-driven intensity index obtained in step 153; the one-hot encoded vector of the dominant defect category tendency encoding obtained in step 152 and the corresponding category consistency; and the distribution of the probability values ​​of defect germination within the region obtained in step 151. The measurement includes probability maxima, probability mean, and probability quantiles. The output label of the sample contains a scalar defect evolution probability value and a three-dimensional vector representing the category evolution trend. This three-dimensional vector corresponds to the evolution probability distribution of three defect categories: microcracks, pores, and delamination. These three defect categories are consistent with the definitions of the aforementioned defect category tendency codes CODE_C, CODE_P, and CODE_D, respectively. In inference applications, for a new risk defect initiation area, the statistical characteristics of the defect initiation probability distribution within the area, the one-hot encoding vector of the dominant defect category tendency code, the category consistency, and the values ​​of the regional interface tension gradient driving strength index are concatenated into an input feature vector in the same field order as during training. This vector is then fed into the trained XGBoost model for score accumulation and summation to obtain the scalar value of the defect evolution probability. Finally, the multi-objective output is normalized using the softmax function to obtain the probability vector of the defect category evolution trend.

[0169] In summary, this invention, by acquiring interfacial tension field distribution information and performing micro-region physical perturbation decoupling processing, decomposes the original tension vector data into steady-state baseline components and local perturbation residual components. From these, it extracts local interfacial free energy fluctuation characteristics and stress concentration kernel tendency characteristics, forming a set of semantic features for latent precursors in micro-regions. This achieves the physical separation and quantitative characterization of hidden defect precursors. Based on this, by performing latent feature association processing of Rayleigh instability texture on the semantic feature set of latent precursors in micro-regions, it constructs a spatiotemporal evolution model of energy accumulation tendency under temporal boundary constraints and obtains a semantic feature description of defect evolution, coupling isolated micro-region features into a defect embryo expression with spatiotemporal evolution attributes. Furthermore, it generates a defect initiation risk cloud map through defect initiation risk mapping processing, and combines it with the spatial change rate of interfacial tension gradient to determine the defect evolution probability distribution, ultimately generating a set of defect prediction instructions. The entire process, from macroscopic field distribution to microscopic perturbation decoupling, and then to cross-micro-region correlation and spatiotemporal evolution modeling, forms a multi-level, multi-physical field fusion defect prediction link. This overcomes the shortcomings of existing methods in being insensitive to latent defects and lacking spatiotemporal evolution prediction capabilities, and improves the accuracy, timeliness, and interpretability of defect initiation risk prediction.

[0170] It should be noted that the acquisition of the interfacial tension field distribution information described in this embodiment does not rely directly on real-time inversion using optical or acoustic sensors in actual production line applications. Instead, it is indirectly achieved by constructing a "process parameter-interfacial tension field" proxy model. Specifically, the temperature field and resin rheological parameters collected by sensors widely equipped in actual production lines, such as infrared thermal imagers and online resin content detectors, are used as inputs. A pre-trained deep neural network model is used to infer the corresponding interfacial tension field distribution information in real time. The training data for this proxy model comes from offline high-precision destructive tests. By collecting semi-cured sheet blanks under different process formulations, micro-Raman spectrometry is used to calibrate the residual stress in micro-areas. The meridional stress, latitudinal stress, and in-plane shear stress together constitute a supervised learning label dataset, which corresponds one-to-one with the process parameter data, thus providing a real and stable data source for defect prediction.

[0171] Furthermore, regarding the implementation of the "defect evolution dynamics prediction mapping function" black-box model, this invention abandons the pure machine learning black-box approach that relies on massive amounts of unknown labeled data. Instead, it employs a deep neural network guided by physical information. The front end of this network is a spatiotemporal convolutional module that takes the semantic features of micro-region latent precursors as input, while the middle layer embeds a physical constraint layer containing the curing dynamics equations and elastic constitutive relations of the prepreg resin. During training, the loss function not only includes the deviation from the labeled region of historical defect cases but also introduces the energy balance residual term from Rayleigh instability theory, forcing the network to converge under the joint supervision of data and physical laws. This architecture ensures that even when historical defect samples are scarce, the model can still output a defect evolution probability distribution that conforms to physical laws based on the underlying materials science principles, thereby transforming unquantifiable semantic features into a reproducible technical solution with a physical core.

[0172] Please see Figure 4 The figure is a schematic diagram of the basic structure of a copper-clad laminate prepreg defect prediction server 200 provided in an embodiment of this application. The copper-clad laminate prepreg defect prediction server 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the copper-clad laminate prepreg defect prediction methods based on interface semantic features.

[0173] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0174] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the copper-clad laminate prepreg defect prediction server reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the copper-clad laminate prepreg defect prediction server to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0175] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A method for predicting defects in copper-clad laminate prepregs based on interface semantic features, characterized in that, include: To obtain the interfacial tension field distribution information on the surface of the prepreg blank to be tested during the manufacturing process of copper clad laminate prepreg; The interface tension field distribution information is subjected to micro-region physical perturbation decoupling processing. The interface tension vector data of each micro-region unit in the interface tension field distribution information is decomposed into steady-state interface tension reference component and local perturbation residual component. Based on the local perturbation residual component, the local interface free energy fluctuation characteristics and stress concentration kernel tendency characteristics are extracted to obtain the micro-region latent precursor semantic feature set. The latent precursor semantic feature set of the micro-region is subjected to Rayleigh instability texture latent feature association processing to construct the spatiotemporal evolution mode of energy accumulation tendency under temporal boundary constraints, and obtain the semantic feature description of defect evolution. The semantic feature description of defect evolution includes the energy accumulation tendency intensity distribution map and the stress concentration nucleus germination probability distribution map. Based on the semantic feature description of defect evolution, a defect emergence risk mapping process is performed to generate a defect emergence risk cloud map. The defect emergence risk cloud map includes a distribution of defect emergence probability values ​​and a defect category tendency coding distribution associated with spatial coordinates. By combining the defect initiation risk cloud map and the spatial change rate of the interface tension gradient in the interface tension field distribution information, the defect evolution probability distribution within a preset time period is determined, and a defect prediction instruction set is generated based on the defect evolution probability distribution.

2. The method according to claim 1, characterized in that, The process involves decoupling the interfacial tension field distribution information by performing micro-region physical perturbation, decomposing the interfacial tension vector data of each micro-region unit in the interfacial tension field distribution information into a steady-state interfacial tension reference component and a local perturbation residual component. Based on the local perturbation residual component, local interfacial free energy fluctuation features and stress concentration kernel tendency features are extracted to obtain a set of semantic features for latent precursors in the micro-region, including: The interfacial tension field distribution information is processed by spatial micro-region division. Based on the surface texture direction of the prepreg blank and the preset micro-region scale, the surface of the prepreg blank is divided into multiple adjacent micro-region units with overlapping boundaries. A spatial position index is assigned to each micro-region unit to obtain the micro-region unit division result. The interface tension vector data in each micro-unit is fitted with a steady-state trend surface. The interface tension steady-state reference surface is constructed in the micro-unit using a local weighted regression algorithm to obtain the micro-unit steady-state interface tension reference component. The interface tension steady-state reference surface represents the ideal distribution of the interface tension field in the micro-unit under undisturbed conditions. Based on the steady-state interface tension reference components of the micro-region, residual separation processing is performed on the interface tension vector data of each sampling point in the micro-region unit. The interface tension amplitude and interface tension direction angle of each sampling point are respectively compared with the steady-state amplitude and steady-state direction at the corresponding positions on the steady-state reference surface of the interface tension to obtain the local disturbance residual components of the micro-region. The local disturbance residual components of the micro-region include disturbance amplitude residual and disturbance direction residual. Based on the residual components of the local perturbation in the micro-region, the interface free energy fluctuations of the micro-region units are quantified and the stress concentration kernel tendency analysis is performed to obtain the local interface free energy fluctuation characteristics and the stress concentration kernel tendency characteristics. The local interface free energy fluctuation characteristics and the stress concentration kernel tendency characteristics are combined to generate a set of semantic features for the latent precursors of the micro-region.

3. The method according to claim 2, characterized in that, The process involves quantifying the interface free energy fluctuations and performing stress concentration kernel tendency analysis on the residual components of the local perturbation in the micro-region, resulting in local interface free energy fluctuation characteristics and stress concentration kernel tendency characteristics. Combining these characteristics, a set of semantic features for latent precursors in the micro-region is generated, including: Based on the disturbance amplitude residual in the local disturbance residual component of the micro-region, the interfacial free energy fluctuation quantification processing is performed on each micro-region unit. The square of the disturbance amplitude residual is correlated and mapped with the surface tension coefficient of the semi-cured sheet blank to generate local interfacial free energy fluctuation characteristics. The local interfacial free energy fluctuation characteristics characterize the degree of fluctuation of interfacial free energy in space and the tendency of energy accumulation in the micro-region unit. Based on the perturbation direction residual in the local perturbation residual component of the micro-region, stress concentration kernel tendency analysis is performed on each micro-region unit. The perturbation direction convergence region within the micro-region unit is determined by calculating the spatial divergence of the perturbation direction residual. Stress concentration kernel tendency features are generated based on the area connectivity and directional convergence density of the perturbation direction convergence region. The stress concentration kernel tendency features characterize the probability and approximate spatial range of stress concentration kernel initiation within the micro-region unit. The local interface free energy fluctuation features and stress concentration core tendency features are semantically encapsulated, and the local interface free energy fluctuation features and stress concentration core tendency features are associated and combined with the spatial location index of the corresponding micro-region unit to obtain the latent precursor semantic feature record of the micro-region unit. Traverse all micro-region units and aggregate the latent precursor semantic feature records of each micro-region unit to generate a micro-region latent precursor semantic feature set. Each record in the micro-region latent precursor semantic feature set includes a micro-region unit spatial location index field, an interface free energy fluctuation feature field, and a stress concentration nucleus germination tendency feature field.

4. The method according to any one of claims 1-3, characterized in that, The method involves performing latent feature association processing on the semantic feature set of latent precursors in the micro-region using Rayleigh instability textures to construct a spatiotemporal evolution model of energy accumulation tendency under temporal boundary constraints, thereby obtaining a semantic feature description of defect evolution, including: The spatial adjacency relationship is constructed by performing spatial adjacency relationship construction on the semantic feature set of the latent precursor of the micro-region. Based on the spatial location index of each micro-region unit, the set of neighboring micro-region units that are spatially adjacent to the micro-region unit is determined, and the spatial adjacency topology of the micro-region is established to obtain the micro-region adjacency relationship graph. In the micro-region adjacency graph, Rayleigh instability condition determination is performed on each pair of adjacent micro-region units. Based on the difference in the local interface free energy fluctuation characteristics of adjacent micro-region units and the consistency of the direction of the stress concentration kernel tendency characteristics, adjacent micro-region unit pairs that satisfy the Rayleigh instability perturbation growth criterion are identified, and Rayleigh unstable micro-region associated edge set is obtained. The connected component extraction process is performed on the associated edge set of the Rayleigh unstable micro-regions, and the associated edges with shared micro-region units are merged to generate multiple non-overlapping Rayleigh unstable perturbation connected domains. Each Rayleigh unstable perturbation connected domain contains a set of spatially continuous micro-region units with uniform perturbation growth conditions. For each Rayleigh instability perturbation connected domain, an energy accumulation tendency intensity distribution map and a stress concentration kernel initiation probability distribution map are generated, and the semantic feature description of defect evolution is determined by temporal partitioning and spatiotemporal correlation modeling.

5. The method according to claim 4, characterized in that, For each Rayleigh instability perturbation connected domain, an energy accumulation tendency intensity distribution map and a stress concentration kernel initiation probability distribution map are generated. Furthermore, a semantic feature description of defect evolution is determined through temporal partitioning and spatiotemporal correlation modeling, including: The local interface free energy fluctuation characteristics within each Rayleigh unstable perturbation connected domain are processed by accumulating the energy accumulation tendency. The interface free energy fluctuation characteristic values ​​of all micro-units within the Rayleigh unstable perturbation connected domain are spatially weighted and accumulated. The spatial gradient direction of the interface free energy fluctuation characteristic values ​​is used as the weighting direction to generate the energy accumulation tendency intensity distribution map of the Rayleigh unstable perturbation connected domain. The stress concentration kernel tendency characteristics within each Rayleigh unstable perturbation connected domain are subjected to probability density estimation. Based on the nonparametric kernel density estimation, a spatially continuous probability density field is constructed for the stress concentration kernel emergence tendency characteristic values ​​of each micro-unit within the Rayleigh unstable perturbation connected domain, thereby generating a probability distribution map of stress concentration kernel emergence within the Rayleigh unstable perturbation connected domain. For each Rayleigh instability perturbation connected domain, the energy accumulation tendency intensity distribution map and the stress concentration kernel germination probability distribution map are subjected to temporal boundary segmentation. Based on the time process stage of the prepreg manufacturing process, temporal boundary constraint parameters corresponding to the process stage of the prepreg blank are introduced. The energy accumulation tendency intensity distribution map and the stress concentration kernel germination probability distribution map are divided into temporal partitions corresponding to different process stages, resulting in temporal partition energy accumulation tendency intensity distribution map and temporal partition stress concentration kernel germination probability distribution map. Spatiotemporal correlation modeling is performed on the temporal partition energy accumulation tendency intensity distribution map and the temporal partition stress concentration nucleus germination probability distribution map. The energy accumulation tendency intensity distribution and stress concentration nucleus germination probability density distribution within the same temporal partition are spatially aligned and superimposed to generate a spatiotemporal evolution model of energy accumulation tendency that reflects the linkage between energy accumulation tendency and stress concentration nucleus germination probability under temporal boundary constraints. The spatiotemporal evolution model of energy accumulation tendency includes the energy accumulation tendency intensity evolution trajectory and stress concentration nucleus germination probability density evolution trajectory under the temporal sequence. The spatiotemporal evolution patterns of energy accumulation tendency of all Rayleigh unstable perturbation connected domains are integrated to generate a semantic feature description of defect evolution. The semantic feature description of defect evolution uses Rayleigh unstable perturbation connected domains as basic units and stores the temporal sequence energy accumulation tendency intensity distribution map and the temporal sequence stress concentration kernel germination probability distribution map of each Rayleigh unstable perturbation connected domain.

6. The method according to claim 1, characterized in that, The step of performing defect emergence risk mapping processing based on the defect evolution semantic feature description to generate a defect emergence risk cloud map includes: The temporal sequence energy accumulation tendency intensity distribution map and the temporal sequence stress concentration kernel germination probability distribution map of each Rayleigh unstable perturbation connected domain in the defect evolution semantic feature description are subjected to temporal fusion processing. The energy accumulation tendency intensity distribution and stress concentration kernel germination probability density distribution of each Rayleigh unstable perturbation connected domain in the current temporal partition and adjacent future temporal partition are extracted to obtain the current temporal energy accumulation tendency intensity distribution map, the current temporal stress concentration kernel germination probability distribution map, the future temporal energy accumulation tendency intensity distribution map, and the future temporal stress concentration kernel germination probability distribution map. The current temporal energy accumulation tendency intensity distribution map and the current temporal stress concentration nucleus germination probability distribution map are subjected to joint probability mapping processing. By performing a spatial dot product operation on the current temporal energy accumulation tendency intensity value and the current temporal stress concentration nucleus germination probability density value, a current defect germination risk probability distribution map reflecting the defect germination risk at the current moment is generated. The future temporal energy accumulation tendency intensity distribution map and the future temporal stress concentration nucleus germination probability distribution map are subjected to joint probability mapping processing. By performing a spatial dot product operation on the future temporal energy accumulation tendency intensity value and the future temporal stress concentration nucleus germination probability density value, a future defect germination risk probability distribution map reflecting the future defect germination risk is generated. Based on the current defect spawning risk probability distribution map and the future defect spawning risk probability distribution map, a spatiotemporal propagation analysis of risk probability is performed to calculate the gradient propagation direction and propagation rate of defect spawning risk probability in space, thereby obtaining a defect spawning risk probability propagation trend vector field. The defect spawning risk probability propagation trend vector field includes the propagation direction vector and propagation rate value of each spatial coordinate point. Based on the propagation trend vector field of the defect germination risk probability, the current defect germination risk probability distribution map is propagated and extended. The defect germination risk probability at the current moment is spatially extrapolated along the propagation direction vector with a propagation rate value to generate a defect germination risk probability propagation distribution map that covers the continuous distribution of the defect germination risk probability within the time period between the current moment and the future moment. The defect priming risk probability propagation distribution map is subjected to defect category tendency mapping processing. The defect priming risk probability value of each spatial coordinate point in the defect priming risk probability propagation distribution map is matched with the preset defect category association rules to determine the defect category tendency code corresponding to each spatial coordinate point and generate a defect category tendency code distribution map. The defect germination risk probability propagation distribution map and the defect category tendency coding distribution map are spatially aligned and fused. Using the spatial coordinate grid as a reference, the defect germination risk probability value and defect category tendency coding value of each grid node are combined and stored to generate a defect germination risk cloud map. The defect germination risk cloud map includes the defect germination probability value distribution and defect category tendency coding distribution indexed by spatial coordinates.

7. The method according to any one of claims 1-3 and 6, characterized in that, The method combines the defect initiation risk cloud map and the spatial change rate of the interface tension gradient in the interface tension field distribution information to determine the defect evolution probability distribution within a preset time period, and generates a defect prediction instruction set based on the defect evolution probability distribution, including: Defect germination probability value distribution data is extracted from the defect germination risk cloud map. Spatial clustering processing is performed on the defect germination probability value distribution data to identify spatially continuous regions where the defect germination probability value exceeds a preset risk concern threshold, thereby obtaining a risk defect germination region set. Each risk defect germination region has a region outline boundary and a defect germination probability value distribution within the region. For each risk defect origination area, the defect category tendency code distribution data corresponding to the area is extracted from the defect origination risk cloud map. The defect category tendency code value with the highest frequency in the area is counted as the dominant defect category tendency code of the risk defect origination area, and the frequency ratio of the dominant defect category tendency code is calculated as the category consistency. For each risk defect initiation area, the spatial change rate of the interface tension gradient corresponding to the area is extracted from the interface tension field distribution information. The spatial distribution of the interface tension gradient spatial change rate is overlaid with the regional contour boundary of the risk defect initiation area. The statistical mean and degree of variation of the spatial change rate of the interface tension gradient inside the region are calculated to obtain the regional interface tension gradient driving strength index. Based on the distribution of defect emergence probability values ​​within each risk defect emergence area, the dominant defect category tendency coding, category consistency, and the regional interface tension gradient driving intensity index, a defect evolution dynamics prediction mapping relationship is constructed. Using a defect evolution dynamics prediction mapping function pre-established based on historical defect cases, the defect evolution probability value and defect evolution category evolution trend of the risk defect emergence area within a preset time period are generated, and the regional defect evolution probability distribution is obtained. Based on the regional defect evolution probability distribution, a defect evolution probability distribution prediction map is determined. The defect evolution probability distribution prediction map is used to determine the spatial positioning coordinate sequence of the risk area, the defect category identification code sequence, and the confidence rating identification sequence, and encapsulates them into a defect prediction instruction set.

8. The method according to claim 7, characterized in that, The method involves determining a defect evolution probability distribution prediction map based on the regional defect evolution probability distribution, using the defect evolution probability distribution prediction map to determine the spatial positioning coordinate sequence of the risk area, the defect category identification code sequence, and the confidence rating identification sequence, and encapsulating them into a defect prediction instruction set, including: Spatially summarize the regional defect evolution probability distribution of all risk defect initiation areas, and spatially stitch together the defect evolution probability values ​​and defect evolution category evolution trends of each region according to the regional outline boundary to generate a defect evolution probability distribution prediction map covering the entire surface of the semi-cured sheet blank. The defect evolution probability distribution prediction map is processed by risk level classification. The surface of the semi-cured sheet blank is divided into multiple risk level blocks according to the distribution range of the defect evolution probability value. The spatial positioning coordinate sequence of each risk level block is extracted to generate the spatial positioning coordinate sequence of the risk area. Defect category identifier codes are assigned based on the expected defect category codes corresponding to the spatial positioning coordinate sequence of each risk area, and confidence rating identifiers are assigned based on the defect evolution probability value of the area, resulting in defect category identifier code sequences and confidence rating identifier sequences. The spatial positioning coordinate sequence of the risk area, the defect category identification code sequence, and the confidence rating identification sequence are encapsulated in an instruction format to generate a defect prediction instruction set. Each instruction in the defect prediction instruction set corresponds to a risk area and includes the spatial positioning coordinates, defect category identification code, and confidence rating identification of the risk area.

9. A server for predicting defects in copper-clad laminate prepregs, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the method for predicting defects in copper-clad laminate prepregs based on interface semantic features as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the method for predicting defects in copper-clad laminate prepregs based on interface semantic features as described in any one of claims 1-8.

Images