A prepreg thickness uniformity analysis method based on deep learning

Abstract

This application provides a deep learning-based method for analyzing the thickness uniformity of prepregs, comprising: acquiring fiber orientation interlacing data during the multilayer lamination of prepregs, collecting fiber angle distribution images, and generating a first fiber distribution map; extracting fiber orientation interlacing features from the first fiber distribution map, separating regions at different angles, and obtaining fiber angle distributions; analyzing resin flow unevenness based on fiber angle distributions, and generating a resin flow distribution map including the locations of tiny resin extrusion protrusions; based on the resin flow distribution map, and combined with data on the gradual adjustment of prepreg multilayer lamination pressure from low to high, simulating the influence of pressure changes on interlayer adhesion, and generating a first adhesion distribution map incorporating resin flow information; and based on the defect location and severity distribution map, and combined with real-time thickness measurement data, predicting the thickness distribution deviation caused by the interaction between tiny protrusions and air bubbles, and generating a first thickness distribution map.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for analyzing the thickness uniformity of prepregs based on deep learning. Background Technology

[0002] Prepreg multilayer lamination technology directly affects the stability of material properties and product reliability, especially in high-precision industries such as aerospace and automotive, where thickness uniformity is a core indicator of product quality. Traditional testing methods mainly rely on single physical measurements or simple image analysis, which are insufficient to handle the dynamic changes of complex interlayer interfaces. These methods often lack the ability to dynamically capture microscopic features when faced with complex situations such as interlaced fiber orientations and variations in interlayer adhesion, resulting in insufficient testing accuracy, particularly in predicting interface defects and thickness distribution. The complexity of interlayer interfaces is a core challenge in this field. First, the interlacing of fiber orientations between upper and lower sheets significantly affects the formation of interface features. For example, during lamination, differences in fiber angles can lead to uneven resin flow, resulting in microscopic protrusions or residual bubbles, which directly affect thickness uniformity. Second, as lamination pressure increases, changes in interlayer adhesion further exacerbate the complexity of interface features. For example, excessive pressure may extrude too much resin, forming irregular protrusions, while insufficient pressure may lead to residual bubbles and thickness fluctuations. These factors are interconnected; the formation of microscopic protrusions amplifies the impact of residual bubbles, making thickness distribution prediction more difficult. Specifically, in actual production, as the pressure is gradually adjusted from low to high, the detection system needs to identify in real time the interaction between the microscopic protrusions formed by resin extrusion and the thickness fluctuations caused by bubbles. Existing methods struggle to dynamically adapt to such multi-scale characteristic changes. Therefore, how to dynamically capture the microscopic interface features jointly caused by changes in fiber orientation interlacing and interlayer adhesion under varying lamination pressure, and accurately identify defects and predict thickness distribution, becomes a key issue. Summary of the Invention

[0003] This invention provides a deep learning-based method for analyzing the thickness uniformity of prepregs, mainly comprising:

[0004] Data on fiber orientation interlacing during the multilayer lamination of prepreg is acquired, and fiber angle distribution images are collected to generate a first fiber distribution map. Fiber orientation interlacing features are extracted from the first fiber distribution map, and different angle regions are separated to obtain fiber angle distributions. Based on the fiber angle distributions, resin flow unevenness is analyzed to generate a resin flow distribution map. Based on the resin flow distribution map, the influence of pressure changes on interlayer adhesion is simulated to generate a first adhesion distribution map incorporating resin flow information. Micro-protrusion regions and bubble distribution information are extracted from the first adhesion distribution map, and a convolutional neural network is used to identify the interaction regions of micro-protrusions and bubble thickness fluctuations to obtain a first interface feature set. Based on the first interface feature set, the coupling effect of fiber orientation interlacing and interlayer adhesion changes is analyzed through multi-scale feature extraction, and a deep learning network is used for defect classification to output defect location and severity distribution maps. Based on the defect location and severity distribution maps, thickness measurement data is used to predict the thickness distribution deviation caused by the interaction between micro-protrusions and bubbles, generating a first thickness distribution map. The lamination pressure parameters are adjusted according to the deviation distribution of the first thickness distribution map to regenerate a second thickness distribution map.

[0005] Furthermore, the step of acquiring fiber orientation interlacing data during the multilayer lamination of the prepreg, collecting fiber angle distribution images, and generating a first fiber distribution map includes:

[0006] The coordinates of the semi-cured sheet sample boundary are determined by the positioning device of the optical scanning equipment, and the fiber orientation interlacing region is identified based on the change in fiber weaving density. For the fiber orientation interlacing region, a line scan camera with a coaxial light source is used to scan line by line, and a grayscale image sequence is obtained based on the difference in fiber surface reflectivity. Based on the grayscale image sequence, Fourier transform is used to extract the frequency domain features of the fiber texture, and the fiber orientation angle is calculated based on the position of the main peak of the spectrum to generate a first fiber distribution map.

[0007] Furthermore, the step of extracting fiber orientation interlacing features from the first fiber distribution map and separating regions at different angles to obtain the fiber angle distribution includes:

[0008] The fiber angle values ​​of pixels are read from the first fiber distribution map. The angle gradient is calculated by the angle difference between adjacent pixels. Boundary points where the angle gradient exceeds the threshold are marked. Based on the boundary points, the first fiber distribution map is segmented using the watershed algorithm. The standard deviation of the pixel angle values ​​within the segmented area is calculated. Regions with a standard deviation less than the threshold are retained to obtain a set of angle-consistent regions. The average fiber angle of the set of angle-consistent regions is calculated to construct an interleaved feature matrix and output the fiber angle distribution.

[0009] Furthermore, the step of analyzing resin flow unevenness based on fiber angle distribution and generating a resin flow distribution map includes:

[0010] Based on the angle difference between adjacent regions in the fiber angle distribution, the change rate of void volume at fiber intersections is calculated to determine the local permeability coefficient and obtain the resin flow resistance distribution. Based on the flow resistance distribution, the resin pressure field is calculated using the finite difference method, and the coordinates of resin-rich regions where the pressure gradient exceeds a threshold are recorded. Based on the ratio of the cumulative volume of the resin-rich region to the fiber gap capacity, the resin rise height is calculated, and a resin flow distribution map including the locations of tiny protrusions is generated.

[0011] Furthermore, the step of simulating the effect of pressure changes on interlayer adhesion based on the resin flow distribution map to generate a first adhesion distribution map incorporating resin flow information includes:

[0012] The coordinates and height data of the protrusions are extracted from the resin flow distribution map. The compression amount of the protrusions is calculated based on the stacking pressure time series data to obtain the height distribution of the protrusions after compression. Based on the height distribution of the protrusions after compression, the distance field between the upper and lower surface layers is calculated, and the ratio of the contact point area to the total area is statistically analyzed to obtain the initial adhesion value. Based on the initial adhesion value and the lateral flow distance of the resin, the adhesion value of the overlapping area is updated to generate the first adhesion distribution map.

[0013] Furthermore, the step of extracting the micro-protrusion region and bubble distribution information from the first fit distribution map, and using a convolutional neural network to identify the interaction region of the micro-protrusion and bubble thickness fluctuations, yields a first interface feature set, including:

[0014] The fit gradient is calculated from the first fit distribution map using a gradient operator. Candidate protrusions with gradient magnitudes exceeding a threshold and bubble positions with fit less than a threshold are marked to obtain a protrusion coordinate set and a bubble contour set. Based on the protrusion coordinate set and bubble contour set, a neural network containing multi-scale convolutional kernels is constructed, and feature maps with different receptive fields are concatenated to generate a multi-scale feature map. According to the multi-scale feature map, the spatial distance matrix between protrusions and bubbles is calculated, the interaction region is determined, the thickness change rate is calculated, an interaction intensity index is constructed, and a first interface feature set containing protrusion positions, bubble contours, and interaction intensity values ​​is output.

[0015] Furthermore, based on the first interface feature set, the coupling effect of fiber orientation interlacing and interlayer bonding variation is analyzed through multi-scale feature extraction, and defect classification is performed using a deep learning network, outputting a distribution map of defect location and severity, including:

[0016] Interaction intensity and spatial location information are extracted from the first interface feature set. The coupling coefficient between fiber orientation angle difference and fit value is calculated. Multi-scale feature vectors are extracted by pyramid pooling. Based on the multi-scale feature vectors, a residual neural network is constructed to extract deep features. The data is input into a classifier to determine the defect type and severity. Based on the defect type and severity, the data is interpolated to the original image size. The defect category and severity values ​​are labeled. Adjacent defect regions of the same type are merged, and a defect distribution map is output.

[0017] Furthermore, the process of generating a first thickness distribution map based on the defect location and severity distribution map, combined with thickness measurement data, to predict the thickness distribution deviation caused by the interaction between micro-bumps and bubbles, includes:

[0018] Defect coordinates are extracted from the defect location and severity distribution map, and a weighted thickness value is calculated by combining the thickness measurement data. Based on the weighted thickness value, the overlapping areas of protrusions and bubbles are identified, the local thickness change is calculated, and a thickness deviation field is generated by interpolation. According to the thickness deviation field, the original thickness data is superimposed, the thickness values ​​of the protrusion and bubble positions are adjusted, and a first thickness distribution map is generated.

[0019] Furthermore, the step of adjusting the stacking pressure parameters based on the deviation distribution of the first thickness distribution map and regenerating the second thickness distribution map includes:

[0020] The difference between the thickness value and the target thickness is calculated from the first thickness distribution map, the coordinates of the deviation exceeding the limit area are extracted, and the pressure control zone number is determined; based on the pressure control zone number, the pressure compensation amount is calculated and the pressure parameters are updated; using the updated pressure parameters, the resin flow state, fit, defect distribution and thickness deviation are recalculated to generate the second thickness distribution map.

[0021] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0022] This invention discloses a deep learning-based method for analyzing the thickness uniformity of prepreg sheets. Addressing business scenarios where uneven resin flow due to interlaced fiber orientation, thickness deviations caused by interactions between micro-protrusions and air bubbles, and interlayer bonding defects are identified, the method uses high-resolution optical scanning to acquire fiber angle distribution, employs image segmentation algorithms to extract orientation features, and combines pressure adjustment data to simulate resin flow and bonding changes. A convolutional neural network with multi-scale feature fusion is used to identify areas where micro-protrusions and air bubbles interact, thereby analyzing the coupling effect between fiber orientation and bonding, classifying defects, and predicting thickness deviations. When the thickness deviation exceeds the standard, the lamination pressure parameters are dynamically adjusted, iteratively optimizing and generating a second thickness distribution map. This invention, through multi-scale feature fusion and dynamic pressure adjustment, significantly improves the lamination quality of prepreg sheets, reduces the defect rate, optimizes the uniformity of thickness distribution, and ensures the manufacturing precision and reliability of high-performance composite materials. Attached Figure Description

[0023] Figure 1 This is a flowchart of a deep learning-based method for analyzing the thickness uniformity of prepregs according to the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0025] like Figure 1 This embodiment of a deep learning-based method for analyzing the thickness uniformity of prepregs may specifically include:

[0026] Step S101: Obtain fiber orientation interlacing data during the multilayer lamination of the prepreg, collect fiber angle distribution images, and generate the first fiber distribution map.

[0027] The initial position information of the prepreg sample is obtained, and the sample boundary coordinates are determined using the positioning device of a high-resolution optical scanning device. The fiber orientation interlacing regions are identified based on changes in fiber weave density, and a scanning path planning scheme is determined. For the fiber orientation interlacing regions, a linear CCD camera with a coaxial light source is used for line-by-line scanning. Grayscale image sequences are obtained based on differences in fiber surface reflectivity, and image registration is performed using the correspondence between feature points of adjacent scan lines. Based on the grayscale image sequences, Fourier transform is used to extract the frequency domain features of the fiber texture. The fiber orientation angle is calculated based on the position of the main peak in the spectrum, and the fiber angle distribution data for each region is obtained through angle distribution histogram statistics, generating a first fiber distribution map containing fiber angle distribution information.

[0028] In one embodiment, the initial position information of the prepreg sample is obtained through a laser positioning device of a high-resolution optical scanning device. The laser positioning device projects a cross-shaped reference line onto the sample stage and identifies the prepreg outline using an edge detection algorithm. When the fiber weaving density exceeds a preset threshold, the area is determined to be a fiber orientation interlacing region. The fiber weaving density is quantified by the number of fiber bundle intersections per unit area; areas with a density value greater than 15 intersections per square centimeter are marked as key scanning areas.

[0029] Specifically, the scanning process of the linear CCD camera employs a push-broom imaging method. A coaxial light source illuminates the surface of the prepreg perpendicularly. The light source intensity is adjusted to a suitable range based on the refractive index of the fiber material; for carbon fiber, the light intensity is set to 1.2 times the standard value, and for glass fiber, it is set to 0.8 times the standard value. The camera moves laterally at a constant speed, acquiring one line of pixel data at a time. The line spacing is determined based on the fiber bundle width, typically set to one-third of the average fiber bundle width. Adjacent scan lines in the grayscale image sequence are registered using SURF feature point extraction. Feature point matching employs the nearest neighbor distance ratio method; a single registration is completed when the number of matched point pairs exceeds a preset threshold.

[0030] In one possible implementation, the Fourier transform process for processing grayscale image sequences includes performing a two-dimensional fast Fourier transform on each image patch, with the patch size set to 128×128 pixels. In the transformed spectrogram, the dominant frequency component corresponding to the fiber texture is represented as bright stripes extending from the origin, with the stripe direction perpendicular to the fiber orientation. Straight-line features in the spectrogram are detected using a Hough transform; the angle of the straight line plus 90 degrees represents the fiber orientation angle of the corresponding region. The angle distribution histogram is statistically analyzed in 5-degree intervals, with the frequency of each interval reflecting the fiber distribution density within that angle range.

[0031] Preferably, angle interpolation is incorporated into the generation of the first fiber distribution map. For scanning blind spots or areas with indistinct features, a bilinear interpolation method is used to estimate the fiber orientation at that point based on known angle values ​​from the surrounding area. The interpolation weight is determined inversely proportional to the distance to ensure the continuity of angle transitions. The generated first fiber distribution map is displayed in pseudo-color, with different colors representing different fiber orientation angle ranges: red represents 0-30 degrees, green represents 30-60 degrees, and blue represents 60-90 degrees, providing an intuitive visualization of the fiber orientation distribution.

[0032] Step S102: Extract fiber orientation interlacing features from the first fiber distribution map, separate regions at different angles, and obtain fiber angle distribution.

[0033] The fiber angle values ​​of each pixel in the first fiber distribution map are read. The angle difference is obtained by subtracting the angle values ​​of adjacent pixels as the angle gradient. When the absolute value of the angle gradient exceeds a preset threshold, it is marked as a potential interlacing boundary point. An initial boundary line is formed by connecting eight neighbors. Based on the initial boundary line, the watershed algorithm is used to segment the fiber distribution map into regions. The standard deviation of the angle values ​​of all pixels in each segmented region is calculated to determine the homogeneity of the region. If the standard deviation is less than the preset threshold, the region is retained; otherwise, the region is segmented again until the homogeneity requirement is met, resulting in a set of angle-consistent regions. For the set of angle-consistent regions, the average fiber angle and area of ​​each region are calculated. An interlacing feature matrix is ​​constructed by the absolute value of the difference between the average angles of adjacent regions. The element in the i-th row and j-th column of the matrix is ​​the angle difference between the i-th region and the j-th region. The output is a fiber angle distribution containing the region number, average angle, and interlacing relationship.

[0034] In one embodiment, the angle gradient is calculated using the central difference method. For a pixel with coordinates (x, y) in the first fiber distribution map, its horizontal angle gradient is obtained by calculating the angle difference between the positions (x+1, y) and (x-1, y) and then dividing by 2. The vertical angle gradient is calculated by the angle difference between (x, y+1) and (x, y-1).

[0035] It should be noted that the fiber angle value cycles within the range of 0 to 180 degrees. When two angle values ​​cross the boundary between 0 and 180 degrees, the smaller angle difference is selected as the actual gradient value.

[0036] Specifically, the eight-neighborhood connection determination process is implemented by scanning each marked potential intersecting boundary point. For each boundary point, it checks whether its eight surrounding adjacent pixels are also marked as boundary points. If a continuous sequence of boundary points exists, a complete boundary line is formed using a chain code tracing method. Broken boundary line segments are connected by determining the distance between their endpoints; when the distance between two endpoints is less than a preset number of pixels, they are considered to belong to the same boundary.

[0037] Preferably, in the implementation of the watershed algorithm, the fiber distribution map is first subjected to distance transformation processing. The distance transformation calculates the Euclidean distance from each non-boundary pixel to the nearest boundary line, forming a topographic map. Local minima are found in the topographic map as seed points, and the segmentation is expanded outward from the seed points until the watershed ridge is encountered, completing the initial segmentation. For regions with too small an area, they are merged with the adjacent most similar regions, and the similarity is determined by the average angle difference between the two regions.

[0038] In one possible implementation, the secondary segmentation for region homogeneity is performed recursively. When the angular standard deviation of a region exceeds a threshold, edge detection is re-executed within that region to find secondary staggered boundaries. The detection threshold for secondary boundaries is set to 0.7 times the initial threshold to ensure that weak angular changes can be identified. The segmented sub-regions are then subjected to homogeneity checks again until all regions meet the requirements or reach the minimum region area limit.

[0039] For example, in constructing the interlacing feature matrix, in addition to recording the angle difference, the length of the shared boundary between regions is also considered. The matrix element values ​​are weighted by multiplying the angle difference by the boundary length; the longer the shared boundary between adjacent regions, the more pronounced their interlacing characteristics. The output fiber angle distribution data structure contains three parts: regional geometric information recording the boundary coordinates and area of ​​each region, angle statistics recording the average angle and angle distribution range, and interlacing relationship information storing the degree of interlacing between adjacent regions through a sparse matrix.

[0040] Step S103: Analyze the resin flow unevenness based on the fiber angle distribution and generate a resin flow distribution map that includes the locations of tiny protrusions in the resin extrusion.

[0041] Based on the angle difference between adjacent regions in the fiber angle distribution, the void volume change rate at fiber intersections is calculated. When the fiber angle difference exceeds a preset threshold, the local permeability coefficient is determined based on the ratio of void volume to fiber bundle spacing, thus obtaining the resin flow resistance distribution in different directions. Based on this flow resistance distribution, the continuity equation is solved using the finite difference method. The resin pressure field is obtained by iterative calculation of the pressure values ​​at the grid nodes. When the local pressure gradient exceeds a critical value, resin enrichment is determined in that region, and the coordinates of the enriched region and the resin accumulation volume are recorded. For the resin enriched region and accumulation volume, the overflow degree is determined based on the ratio of the accumulation volume to the fiber gap capacity. If the ratio is greater than a preset threshold, the resin climbing height on the fiber surface is calculated based on the difference between the resin surface tension and the fiber surface energy, obtaining the height distribution data of the micro-protrusions. Based on the height distribution data of the micro-protrusions, a continuous resin thickness field is constructed using interpolation methods. Bicubic interpolation is used in densely interwoven fiber regions to preserve detailed features, while linear interpolation is used in parallel fiber regions to improve computational efficiency. The coordinates of the protrusion positions and the thickness values ​​are merged to generate a resin flow distribution map containing the positions of the resin extrusion micro-protrusions.

[0042] In one embodiment, the correlation between fiber angle distribution and resin flow resistance is achieved through multiphysics coupling. The rate of change of void volume at fiber interlacing depends on the relative angle between the upper and lower fiber layers. When two fiber layers interlaced at 90 degrees, the resulting rhomboid voids are minimized, resulting in maximum resin flow resistance. When the fibers are arranged in parallel, the resulting rectangular channel voids are maximized, resulting in minimum flow resistance. The void volume is calculated using a trigonometric function relationship between fiber bundle width, thickness, and interlacing angle, specifically by multiplying the void area by the unit thickness, where the void area equals the square of the fiber bundle spacing multiplied by the sine of the interlacing angle.

[0043] It should be noted that the determination of the local permeability coefficient involves porous media theory. The permeability coefficient is directly proportional to the cube of the pore volume and inversely proportional to the square of the fiber surface area. During the prepreg stacking process, the compression of the fiber bundles changes the pore structure, and the permeability coefficient dynamically changes with pressure. After obtaining the baseline permeability coefficient through experimental calibration, it is corrected based on the ratio of the actual pore volume to the baseline pore volume. The flow resistance distribution forms a two-dimensional tensor field, with the principal direction along the fiber alignment direction, and the resistance in the perpendicular direction is typically 3 to 5 times that in the principal direction.

[0044] Specifically, the finite difference method for solving the continuity equation employs staggered grid discretization. Pressure nodes are located at the grid center, and velocity components are defined at the midpoints of the grid boundaries. The continuity equation is expressed as the conservation of resin mass, meaning the mass of resin flowing into the control volume per unit time equals the mass flowing out plus the change in resin mass within the control volume. After discretization, a system of linear equations is formed, with each internal node corresponding to one equation, and boundary nodes set according to actual boundary conditions. Iterative solutions use an over-relaxation method, with the relaxation factor determined based on grid density and permeability distribution, typically between 1.2 and 1.8. The convergence criterion is that the maximum relative error of the pressure field between two adjacent iterations is less than a preset threshold. The pressure gradient is calculated by dividing the pressure difference between adjacent nodes by the node spacing; when the gradient value exceeds the critical driving pressure for resin flow, it is marked as a potential enrichment zone.

[0045] Preferably, the determination of resin enrichment incorporates a time-cumulative effect. The resin accumulation volume depends not only on the instantaneous flow rate but also on the duration. Within each time step, the velocity field is calculated based on the pressure gradient and permeability coefficient. The displacement is obtained by multiplying the velocity by the time step. The net accumulation volume is obtained by summing the resin inflow and subtracting the outflow from the resin inflow at each time step. The fiber gap capacity is calculated based on the fiber volume fraction. The fiber volume fraction of the prepreg is typically 0.5 to 0.6, with the remaining space being the volume that the resin can occupy.

[0046] In one possible implementation, the resin climbing height is calculated based on the principle of capillary action. The resin climbing on the fiber surface is driven by surface tension; the climbing height is directly proportional to the surface tension and inversely proportional to the product of resin density and gravitational acceleration. Surface tension is measured using the pendant drop method, and fiber surface energy is obtained using a contact angle meter. When the interfacial tension between the resin and fiber is less than the resin's cohesive force, the resin spreads on the fiber surface and climbs upwards. The curvature of the meniscus formed during the climbing process determines the final protrusion shape, and the radius of curvature is related to the pressure difference through the Laplace equation.

[0047] For example, statistical methods are used to process the height distribution data of minute protrusions. Gaussian fitting is performed on the protrusion height of each enriched region to obtain the mean height and standard deviation. Abnormally high protrusions are identified using a 3-standard-deviation criterion; these locations typically correspond to fiber breakage or misalignment defects. The lateral extent of the protrusions is determined by the gradient change in height, with the protrusion boundary considered reached when the height decreases to 10% of the peak value. Furthermore, the interpolation method is chosen based on local fiber arrangement characteristics. Bicubic interpolation constructs a cubic polynomial surface using 16 neighboring points, maintaining the continuity and smoothness of the height field. In densely interwoven fiber regions, sampling points are set at the fiber bundle scale to ensure the capture of detailed variations. Linear interpolation maintains basic continuity while requiring less computation and is suitable for regions with gradual changes. Interpolation weights are determined inversely proportional to distance, with the influence of nearest neighbors being greater than that of far neighbors.

[0048] Understandably, the data structure of the resin flow distribution map is organized in layers. The bottom layer is a regular grid storing the basic thickness field, the middle layer is a sparse matrix recording the protrusion positions and height increments, and the top layer is vector data describing the protrusion contours.

[0049] Step S104: Based on the resin flow distribution map and combined with the data of gradually adjusting the pressure of the prepreg multilayer lamination from low to high, simulate the effect of pressure change on interlayer adhesion and generate a first adhesion distribution map that incorporates the resin flow information.

[0050] The coordinates and height data of the protrusions are extracted from the resin flow distribution map. A pressure-time curve is established based on the stacking pressure time series data. The compression amount of the protrusions under different pressures is calculated using the pressure-deformation relationship in Hertzian contact theory. The post-compression protrusion height distribution is obtained by subtracting the compression amount from the initial protrusion height. Based on this post-compression protrusion height distribution, the distance field between the upper and lower prepreg surfaces is calculated. When the post-compression protrusion height is greater than the distance between the upper and lower surfaces, it is recorded as a contact point. The ratio of the area of ​​all contact points to the total area is calculated to obtain the initial adhesion value. For this initial adhesion value, the lateral flow distance of the resin from the top of the protrusion to the surrounding area under pressure is considered. This lateral flow distance is determined by the ratio of resin viscosity to pressure gradient. If the resin-covered area overlaps with the covered area of ​​an adjacent protrusion, the adhesion value of the overlapping area is updated to a fully adhered state, resulting in a corrected adhesion field. Based on the modified fit field, the resin thickness value at each location is extracted from the resin flow distribution map. A comprehensive fit feature is constructed by weighting the fit value and the resin thickness value. The weighting coefficient is dynamically adjusted according to the pressure magnitude, and a first fit distribution map integrating resin flow information is output.

[0051] In one embodiment, the application of Hertzian contact theory requires consideration of the elastic modulus and Poisson's ratio of the prepreg material. The resin protrusion can be viewed as a spherical indenter in an elastic half-space; when vertical pressure is applied, a circular indentation forms in the contact area. The relationship between compression and pressure follows a power function, with the exponent depending on the geometry of the protrusion. For a hemispherical protrusion, compression is proportional to the 2 / 3 power of the pressure; for a conical protrusion, compression is proportional to the 1 / 2 power of the pressure. The actual protrusion shape lies between these two values ​​and is corrected for using a shape factor.

[0052] It should be noted that the pressure-time curve is established based on actual lamination process parameters. The prepreg lamination process is typically divided into three stages: the initial contact stage, where the pressure gradually increases from zero to 30% of the working pressure; the holding stage, where the pressure is maintained at a constant level; and the depressurization stage, where the pressure decreases slowly. The duration and pressure change rate of each stage are determined based on material properties and product requirements. Pressure sensors collect pressure data at a fixed frequency, with the time interval typically set to 0.1 seconds, ensuring that the dynamic process of pressure changes is captured.

[0053] Specifically, the distance field is calculated using the Euclidean distance transformation method. For each point on the surface of the upper and lower prepreg layers, the vertical distance to the nearest surface feature point is calculated. The upper surface uses its bottom surface as a reference, and the lower surface uses its top surface as a reference; the sum of the two distance fields is the interlayer distance field. In fiber-interlaced regions, the local distance field may fluctuate due to the superposition of fiber thicknesses. The resolution of the distance field is matched to the scanning accuracy, typically 10 sampling points per millimeter. When the height of the bulge after pressing is greater than the distance field value at the corresponding location, it indicates that actual contact has occurred at that location. The contact area is identified through binarization processing to form a contact mask. The contact area is obtained by counting the number of pixels with a value of 1 in the mask image; the ratio to the total number of pixels represents the fit.

[0054] Preferably, the calculation of the resin lateral flow distance incorporates the theory of extrusion flow. Under pressure, the resin at the top of the protrusion is extruded, generating radial flow. The flow velocity is directly proportional to the pressure gradient and inversely proportional to the resin viscosity. The pressure gradient is obtained by dividing the difference between the pressure at the center and edge of the protrusion by the protrusion radius. The resin viscosity varies with temperature and shear rate, described using a power-law fluid model. The flow distance is equal to the integral of the flow velocity over time, with the upper limit of the integral being the pressure application time. When the resin flow regions of two adjacent protrusions overlap, the resin thickness in the overlapping region increases, forming a continuous resin layer. This phenomenon is called resin bridging, which can significantly improve local bonding strength.

[0055] In one possible implementation, the updated fit field is updated using an iterative method. The initial fit field is calculated based on geometric contact, and the fit value is updated according to resin flow in each iteration. The update rule is: if the original fit at a certain location is less than 1 and is covered by resin, the fit is increased by an increment value, which is proportional to the resin thickness. Iteration continues until the fit field converges or the maximum number of iterations is reached. The convergence criterion is that the maximum difference between the fit fields of two adjacent iterations is less than a preset threshold.

[0056] For example, the construction process of the comprehensive adhesion feature needs to consider the weighting of multiple factors. The adhesion degree value reflects the degree of geometric contact and ranges from 0 to 1; the resin thickness value reflects the physical thickness of the adhesive layer and is also mapped to the range of 0 to 1 after normalization. The weighting coefficient is determined based on the pressure magnitude. At low pressure, the resin thickness has a higher weight because the adhesion mainly relies on resin bonding; at high pressure, the adhesion degree has a higher weight because a tight mechanical contact is formed. The weighting coefficient varies between two extreme values ​​through linear interpolation. When the pressure is zero, the resin thickness weight is 0.7 and the adhesion degree weight is 0.3; when the pressure reaches its maximum value, the resin thickness weight is 0.3 and the adhesion degree weight is 0.7. In one embodiment, the first adhesion degree distribution map contains multiple channels (adhesion degree, resin thickness, pressure), each channel reflecting a different physical state. The weighting coefficient provides a mechanism to weight and fuse these different dimensions of information (geometric contact, material thickness) into a unified comprehensive evaluation index. Furthermore, the data structure of the first adhesion degree distribution map adopts a multi-channel image format. The first channel stores the fit value, the second channel stores the resin thickness information, and the third channel stores the pressure distribution data. This multi-channel representation facilitates subsequent processing and visualization; each channel can be displayed independently or in combination.

[0057] Understandably, the spatial resolution of the fit distribution map needs to match the requirements of the actual application. For aerospace-grade composite materials, a resolution of 0.1 mm is required; for automotive applications, the resolution can be relaxed to 0.5 mm. High resolution can capture minute defects, but it increases the amount of data processing and storage requirements.

[0058] Step S105: Simultaneously extract the distribution information of the micro-protrusion region and the bubble from the first fit distribution map, and use a multi-scale feature fusion convolutional neural network to identify the interaction region of the micro-protrusion and bubble thickness fluctuation, so as to obtain the first interface feature set containing the interaction features.

[0059] The horizontal and vertical gradients of the fit are calculated from the first fit distribution map using the Sobel gradient operator. Locations with gradient magnitudes exceeding a preset threshold are marked as candidate protrusion points. Simultaneously, regions with fit less than a preset lower limit are identified as bubble locations using an 8-connected domain labeling method, resulting in a set of protrusion coordinates and a set of bubble contours. Based on these sets, a convolutional neural network with three parallel branches is constructed. Each branch uses 3×3, 5×5, and 7×7 convolutional kernels to extract features from different receptive fields. The output feature maps of the three branches are concatenated along the channel dimension to obtain a fused multi-scale feature map. For this multi-scale feature map, the Euclidean distance between the center point of the protrusion region and each point on the bubble contour is calculated to form a spatial distance matrix. When the minimum distance is less than a preset threshold, it is identified as a potential interaction zone. Within the interaction zone, the thickness change rate is calculated by dividing the thickness difference between adjacent pixels by the pixel spacing. Based on the thickness change rate, the protrusion height and bubble depth are normalized and multiplied, and then weighted and summed with the thickness change rate to construct an interaction intensity index. The interaction area is divided into three categories: strong, medium and weak according to the interaction intensity value, and the first interface feature set containing the protrusion position, bubble outline, interaction area category and interaction intensity value is output.

[0060] In one embodiment, the application of the Sobel gradient operator involves convolution operations in two directions. A 3×3 convolution kernel of [-1,0,1; -2,0,2; -1,0,1] is used in the horizontal direction, and a convolution kernel of [-1,-2,-1; 0,0,0; 1,2,1] is used in the vertical direction. Convolution operations are performed on each pixel location of the first fit distribution map to obtain the horizontal gradient Gx and the vertical gradient Gy. The gradient magnitude is calculated by taking the square root of the sum of the squares of Gx and Gy, and the gradient direction is determined by the arctangent function of Gy and Gx. When the gradient magnitude exceeds a preset threshold, the location is marked as a potential raised edge point.

[0061] It should be noted that the eight-connected region labeling method employs a two-scan strategy. The first scan starts from the top left corner, checking each pixel row by row and column by column. If the current pixel's fit value is less than the bubble determination threshold, the labeling status of its four neighboring pixels (top left, top, top right, and left) is checked. If a labeled pixel exists in the neighborhood, the current pixel inherits the smallest label value; otherwise, a new label value is assigned. The second scan addresses the equivalent labeling problem, unifying pixels belonging to the same connected region but with different labels into the same label. Each independent connected region represents a potential bubble, and the spatial characteristics of the bubble are determined by calculating the region's bounding rectangle and centroid position.

[0062] Specifically, the multi-scale convolutional neural network employs a parallel branch structure to achieve multi-scale feature extraction. The first branch uses a 3×3 convolutional kernel with a small receptive field, primarily capturing local detail features such as the sharp edges of protrusions and the subtle textures of bubbles. The second branch uses a 5×5 convolutional kernel with a medium receptive field, capturing medium-scale structural features such as the overall shape of protrusions and the basic outline of bubbles. The third branch uses a 7×7 convolutional kernel with a larger receptive field, capturing global contextual information such as the distribution pattern of protrusion groups and the spatial arrangement of bubbles. Each branch contains two convolutional layers: the first layer performs feature extraction, and the second layer performs feature enhancement. Batch normalization layers and ReLU activation functions are applied after the convolutional operations to maintain the non-linear expressiveness of the features. The output feature maps of the three branches are concatenated along the channel dimension to form a fused feature map containing multi-scale information. The number of channels in the concatenated feature map is the sum of the number of channels in the outputs of the three branches. Channel dimensionality reduction is performed using 1×1 convolutions to reduce computational complexity while preserving key features.

[0063] Preferably, the calculation process of the spatial distance matrix takes into account the geometric characteristics of the protrusions and bubbles. For each protrusion region, its geometric center is first determined by the arithmetic mean of all pixel coordinates within the region. For each bubble contour, sampling points on the contour are extracted, and the sampling interval is adaptively adjusted according to the contour perimeter. The element D[i,j] of the distance matrix represents the minimum Euclidean distance from the center of the i-th protrusion to all sampling points on the contour of the j-th bubble. The Euclidean distance is calculated by taking the square root of the sum of the squares of the coordinate differences. When D[i,j] is less than the interaction determination threshold, the i-th protrusion and the j-th bubble are considered to have a spatial interaction relationship.

[0064] In one possible implementation, the calculation of the thickness change rate needs to consider the thickness distribution of the local neighborhood. For each pixel within the interaction area, the thickness value in its 3×3 neighborhood is extracted. A local plane is fitted using the least squares method, and the gradient coefficient of the plane equation is the thickness change rate at that point. The direction of the thickness change rate indicates the direction of the fastest increase in thickness, and the magnitude represents the degree of change. At the boundary between the bump and the bubble, the thickness change rate usually exhibits a local maximum.

[0065] For example, the construction process of the interaction intensity index involves the normalization and weighted fusion of multiple parameters. The protrusion height is mapped to the 0-1 range through maximum value normalization, and the bubble depth is also normalized. The product of the two reflects the degree of coupling of morphological features. The thickness change rate is normalized after logarithmic transformation to reduce the influence of extreme values. The interaction intensity is calculated through weighted summation, with weight coefficients determined based on the material properties of the prepreg. For carbon fiber materials, the protrusion height weight is set to 0.4, the bubble depth weight to 0.3, and the thickness change rate weight to 0.3; for glass fiber materials, the corresponding weights are 0.35, 0.35, and 0.3, respectively. Furthermore, the classification of interaction regions uses a threshold segmentation method. Interaction intensity values ​​in the range of 0-0.3 are marked as weak interaction, mainly characterized by the independent existence of protrusions and bubbles with minimal mutual influence. Intensity values ​​in the range of 0.3-0.7 are marked as moderate interaction, where the edges of protrusions partially overlap with the boundaries of bubbles, and a transition zone appears in the thickness distribution. A strength value greater than 0.7 is marked as strong interaction, where the protrusion completely covers the bubble or the bubble surrounds the protrusion, forming a composite defect structure.

[0066] Understandably, the data organization of the first interface feature set adopts a structured storage method. The protrusion position is recorded in the form of center coordinates and bounding box, the bubble outline is stored as a sequence of polygon vertices, the interaction area is represented by a binary mask image, and the interaction intensity value is stored as a floating-point matrix.

[0067] Step S106: Based on the interaction features in the first interface feature set, the coupling effect of fiber orientation interlacing and interlayer bonding change is analyzed by multi-scale feature extraction method. The defects of the prepreg are classified by combining the multi-scale feature fusion strategy of deep learning network, and the distribution map of defect location and severity is output.

[0068] Interaction intensity values ​​and spatial location information are extracted from the first interface feature set. The fiber orientation angle difference and fit value are normalized and multiplied to obtain the coupling coefficient. Local features are extracted on different scale grids using pyramid pooling to obtain a multi-scale feature vector set containing coupling relationships. Based on the multi-scale feature vector set, a residual neural network is constructed for deep feature extraction. The network contains four residual blocks. Each residual block retains the original feature information and learns the residual mapping through skip connections. Global average pooling compresses the feature map into a fixed-length feature vector. For the feature vector, a softmax classifier is input to determine the defect type. Defects are classified into four categories: layering, porosity, wrinkles, and hybrid. The severity of the defect is determined by comparing the maximum classification probability with a preset threshold. A probability value less than the first threshold is considered mild, between the first and second thresholds is moderate, and greater than the second threshold is severe. Based on the defect type and severity level, bilinear interpolation maps the classification results back from the feature map size to the original image size. The defect category and severity value are labeled at each pixel location. Adjacent defect regions of the same type are merged using morphological closing operations, outputting a defect distribution map containing defect location coordinates, category labels, and severity values.

[0069] In one embodiment, the calculation of the coupling coefficient requires comprehensive consideration of the interaction between two dimensions: fiber orientation and fit. The fiber orientation angle difference is determined by the angle between the main fiber directions of adjacent regions, ranging from 0 to 90 degrees. When the angle difference is 0 degrees, the fibers are parallel, resulting in uniform interlayer stress transfer; when the angle difference is 90 degrees, the fibers are orthogonally aligned, forming stress concentration points. The angle difference is normalized by dividing by 90 degrees, mapping to the 0-1 range. The fit value itself is already within the 0-1 range and is used directly. The coupling coefficient obtained by multiplying the two values ​​reflects the degree of structural heterogeneity; a larger value indicates that the region is more prone to defects.

[0070] It should be noted that pyramid pooling employs a spatial pyramid structure to achieve multi-scale feature extraction. The input feature map is divided into grids of different granularities: the first layer is a 1×1 grid, extracting global features; the second layer is a 2×2 grid, extracting regional features in four quadrants; and the third layer is a 4×4 grid, extracting local features from 16 sub-regions. Max pooling is performed within each grid to retain the most salient feature responses. The pooling results from different levels are upsampled to the same size and then concatenated along the channel dimension to form a feature representation containing multi-scale information.

[0071] Specifically, the residual neural network is constructed following the identity mapping principle to ensure effective gradient propagation. Each residual block contains two paths: a main path and a skip path. The main path contains two 3×3 convolutional layers, with batch normalization and ReLU activation functions inserted in between. The first convolution keeps the number of channels constant, while the second convolution can change the number of channels to adapt to different feature dimension requirements. The skip path directly passes the input to the output; when the input and output dimensions do not match, dimensionality adjustment is performed through a 1×1 convolution. The outputs of the two paths are fused element-wise and then activated by ReLU. Four residual blocks are connected sequentially, with channel numbers of 64, 128, 256, and 512, progressively extracting more abstract feature representations. Global average pooling at the end of the network averages the feature maps of each channel, compressing the spatial dimension to 1×1, and outputting a feature vector with a length equal to the number of channels. This pooling method reduces the number of parameters while maintaining robustness to changes in input size.

[0072] Preferably, the defect classification process requires pre-defining the feature patterns of defect categories. Layered defects are characterized by continuous regions with locally zero fit, typically distributed in bands or sheets; porosity defects are characterized by scattered small areas of low fit, randomly distributed; wrinkled defects are characterized by periodic fluctuations in fit, accompanied by regular changes in fiber orientation; mixed defects possess multiple defect features simultaneously, making them difficult to classify into a single category. The input to the softmax classifier is the feature vector output by the residual network, which is mapped to a 4-dimensional space through a fully connected layer, with each dimension corresponding to a score for a defect type. The softmax function converts the scores into a probability distribution, and the category with the highest probability is used as the prediction result.

[0073] In one possible implementation, severity is determined using a tiered threshold method. The first threshold is set to 0.6, and the second threshold is set to 0.8. When the maximum classification probability is less than 0.6, the defect characteristics are not obvious, and it is classified as a minor level; when the probability is between 0.6 and 0.8, the defect characteristics are relatively clear, and it is classified as a moderate level; when the probability is greater than 0.8, the defect characteristics are very obvious, and it is classified as a severe level. This tiered approach considers the correlation between classification confidence and defect significance.

[0074] For example, the application of bilinear interpolation ensures the spatial correspondence between the classification result and the original image. After multiple pooling and convolutions, the spatial resolution of the feature map is reduced to 1 / 32 or 1 / 16 of the original size. Bilinear interpolation calculates the value of the interpolated point by weighted averaging of four adjacent points, with the weights determined inversely proportional to the distance. The interpolation process is performed incrementally, increasing the resolution by a factor of 2 each time until the original size is restored. Furthermore, morphological closing operations are used to optimize the connectivity of the defect region. The closing operation first performs dilation and then erosion, using the same structuring element. The dilation operation expands the boundary of the defect region, fills internal holes, and connects adjacent regions; the erosion operation shrinks the boundary, restoring the original size. The structuring element is a 3×3 or 5×5 square kernel, determined based on the typical size of the defect.

[0075] Understandably, the output format of the defect distribution map adopts a multi-layered structure. The bottom layer is a location coordinate matrix, recording the bounding box of each defect region; the middle layer is a category label matrix, storing the corresponding defect type code for each pixel position; and the top layer is a severity matrix, storing the quantified severity values.

[0076] Step S107: Based on the defect location and severity distribution map, and combined with real-time thickness measurement data, predict the thickness distribution deviation caused by the interaction between the micro-bumps and bubbles, and generate the first thickness distribution map.

[0077] Defect type and spatial coordinates are extracted from the defect location and severity distribution map. Thickness data is collected in the defect area and its surroundings using an ultrasonic thickness gauge. The influence weight is calculated by the reciprocal of the Euclidean distance between the measurement point and the defect center; the closer the distance, the greater the weight, resulting in a weighted thickness value for each measurement point. Based on these weighted thickness values, the spatial overlap region of micro-protrusions and bubbles is identified. The protrusion height is obtained from the resin flow distribution map, and the bubble depth is determined by the depth of the low-value region in the fit distribution map. The difference between the two forms a local thickness variation. A continuous thickness deviation field is obtained by spatially interpolating the discrete variation using Kriging interpolation. For the thickness deviation field, a deviation amplification coefficient is assigned according to the defect severity. The coefficient for severe defects is taken as the upper limit of a preset range, for moderate defects as the median of the range, and for minor defects as the lower limit of the range. The deviation field is multiplied point by point by the amplification coefficient to obtain a weighted deviation distribution. Based on the weighted deviation distribution, the original thickness data measured in real time is superimposed. The deviation value is added to the protrusion location and subtracted from the bubble location. The thickness boundaries are smoothed using a Gaussian filter, outputting a first thickness distribution map that includes the interaction between micro-protrusions and bubbles.

[0078] In one embodiment, the ultrasonic thickness gauge uses the pulse-echo method for data acquisition. Ultrasonic pulses emitted by the probe penetrate the prepreg, generating reflected echoes on the upper and lower surfaces. By measuring the time difference between the emitted pulse and the received echo, and combining this with the sound velocity in the material, the thickness value is calculated. The measurement points are arranged according to a gridding principle: a dense grid with a spacing of 2 mm is used in the defect center region; a sparse grid with a spacing of 5 mm is used in the defect edge region; and the spacing can reach 10 mm in the defect-free region. This variable-density sampling strategy balances the requirements for measurement accuracy and efficiency.

[0079] It should be noted that the weighting calculation is based on the principle of spatial attenuation. The Euclidean distance is obtained by taking the square root of the sum of the squares of the differences between the coordinates of the measurement point and the coordinates of the defect center. The weight value is equal to the normalized result of the reciprocal of the distance, ensuring that the sum of all weights is 1. When the measurement point is exactly located at the defect center, a maximum weight value of 0.5 is set to avoid anomalies caused by excessively large weights at a single point. For measurement points whose distance exceeds the defect's influence radius, the weight value is set to zero and they are not included in subsequent calculations. The weighted thickness value is obtained by multiplying the thickness of each measurement point by its corresponding weight and summing the results.

[0080] Specifically, obtaining the protrusion height and bubble depth involves multi-source data fusion. The protrusion height is directly read from the resin flow distribution map, which records the height value of each protrusion location relative to the reference surface. The bubble depth is obtained through analysis of low-value regions in the fit distribution map; the vertical distance from the center point of the region with near-zero fit to the surface is the bubble depth. In spatially overlapping regions, protrusions may partially fill the bubble space, and the actual thickness change needs to consider the relative positional relationship between the two. When the protrusion is completely inside the bubble, the thickness change is the protrusion height; when the protrusion partially enters the bubble, the change is the distance from the top of the protrusion to the bottom of the bubble; when the two only have edge contact, the change is determined by geometric intersection calculations.

[0081] Preferably, the implementation of Kriging interpolation requires the construction of a variogram model. The variogram describes the spatial correlation of thickness variation, and the empirical variogram is obtained by calculating the semivariance of sample point pairs at different distances. Commonly used theoretical models include the spherical model, the exponential model, and the Gaussian model. A suitable model is selected for fitting based on the shape of the empirical variogram. Model parameters include nugget value, sill value, and range, reflecting measurement error, spatial variability intensity, and correlation distance, respectively. The interpolation process obtains the weighting coefficients of the points to be estimated by solving the Kriging equations. The coefficient matrix of the equations consists of the variogram values ​​between known points. The interpolation results not only provide the estimated value but also the estimated variance, used to evaluate the interpolation accuracy. Near the defect boundary, due to the uneven distribution of sample points, the interpolation accuracy will decrease, requiring the addition of boundary constraints.

[0082] In one possible implementation, the deviation amplification factor is determined based on the degree of impact of the defect on product performance. The amplification factor for severe defects is set to range from 1.5 to 2.0, indicating that the actual deviation may reach twice the measured value; the factor for moderate defects ranges from 1.2 to 1.5, reflecting a moderate level of uncertainty; and the factor for minor defects ranges from 1.0 to 1.2, essentially maintaining the original measured value. The specific value is determined by a weighted average of the defect area ratio and depth ratio, with the area ratio having a weight of 0.6 and the depth ratio a weight of 0.4.

[0083] For example, a Gaussian filter is used to achieve a smooth transition at thickness boundaries. The size of the filter kernel is determined based on the spatial frequency of the thickness variation, typically using a 5×5 or 7×7 two-dimensional Gaussian kernel. The standard deviation parameter controls the degree of smoothing; a larger standard deviation produces a stronger smoothing effect but may lose detail, while a smaller standard deviation retains more detail but has a limited smoothing effect. At the boundary between the bulge and the normal region, the filtering process eliminates abrupt changes in thickness, forming a gradual transition. At the bubble edge, the filtering slows down the sharp drop in thickness, making the thickness distribution more consistent with actual physical characteristics. Furthermore, the data organization of the first thickness distribution map adopts a hierarchical structure. The base layer stores the original measured thickness, maintaining the integrity of the measurement data; the deviation layer records the thickness deviation values ​​at each location, facilitating the tracing of the deviation source; the fusion layer contains the final thickness distribution result for quality assessment and visualization.

[0084] Understandably, the accuracy of the thickness distribution map is verified by comparison with a standard sample. A standard sample with a known thickness distribution is selected, and the same measurement and processing procedures are applied to compare the difference between the predicted results and the actual thickness. When the root mean square value of the prediction error is less than the allowable tolerance, the thickness distribution map is considered to meet the accuracy requirements.

[0085] Step S108: If the thickness deviation in the first thickness distribution map exceeds a preset threshold, the prepreg stacking pressure parameters are dynamically adjusted by analyzing the insufficient stacking pressure location corresponding to the area with the largest thickness deviation based on the deviation distribution of the first thickness distribution map, and an optimized second thickness distribution map is generated as the corrected thickness distribution prediction result.

[0086] The difference between the thickness value at each location and the design target thickness is calculated from the first thickness distribution map. If the absolute value of the maximum deviation exceeds a preset threshold, the spatial coordinates of the deviation-exceeding area are extracted. The pressure control zone number to which the area belongs is determined by the pre-established coordinate-pressure zone correspondence. Based on the pressure control zone number, the pressure value currently applied to that zone is read. The required pressure compensation is calculated by multiplying the deviation value by the reciprocal of the material's compressive modulus, which is pre-determined through material testing. For the pressure compensation, the compensation is added to the original pressure value to obtain a new pressure parameter. If the new pressure parameter exceeds the upper limit of the equipment's allowable pressure, the upper limit value is taken. The pressure setting value of the corresponding zone is updated by adjusting the output pressure of the hydraulic system, and the adjusted pressure parameter is recorded. Based on the adjusted pressure parameter, the resin flow state, interlayer bonding degree, interface defect distribution, and thickness deviation prediction are recalculated using the new pressure value to obtain a second thickness distribution map. The ratio of the maximum deviation of the second thickness distribution map to the maximum deviation of the first thickness distribution map is calculated. If the ratio is less than a preset improvement threshold, the second thickness distribution map is output as the corrected thickness distribution prediction result.

[0087] In one embodiment, the target thickness is determined based on product specifications and material properties. For aerospace-grade composites, the target thickness is typically set as the number of fiber layers multiplied by the theoretical thickness per layer, taking into account the effect of resin content. The typical thickness per layer for carbon fiber prepreg is 0.125 mm, and for glass fiber prepreg, it is 0.2 mm. The contribution of resin content to thickness is calculated using volume fraction; a resin volume fraction of 40% results in a thickness increase factor of 1.1. Deviation thresholds are set considering product usage requirements and manufacturing tolerances; the thickness deviation threshold for aerospace structural components is typically ±5% of the target thickness, while for automotive parts it can be relaxed to ±8%.

[0088] It should be noted that establishing the correspondence between coordinates and pressure zones involves the physical structure of the laminating equipment. Pressure application in the prepreg laminating machine is achieved through multiple independently controlled hydraulic cylinders, each covering a specific working area. Pressure zones are defined according to the distribution of the hydraulic cylinders, typically arranged in a 4×4 or 6×6 matrix. The boundary of each zone is determined by the perpendicular bisector of the line connecting the centers of adjacent hydraulic cylinders. Coordinate mapping is achieved through a lookup table; the input is the (x, y) position in the workpiece coordinate system, and the output is the corresponding zone number. Since boundary areas may be affected by multiple zones, a distance-weighted method is used to determine the primary control zone.

[0089] Specifically, the determination of the material's compressive modulus follows standard test methods. Specimens are prepared using the same materials and processes as the actual product, with dimensions of 100×100×10 mm. Compression tests are conducted on a universal testing machine at a loading rate of 1 mm / min. Thickness variations under different pressures are recorded, and stress-strain curves are plotted. The compressive modulus is determined by the slope of the linear segment of the curve; the compressive modulus of carbon fiber composites is typically in the range of 8-12 GPa. The calculation of the pressure compensation takes into account the material's nonlinear characteristics; when the strain exceeds 2%, a nonlinear correction factor needs to be introduced. The compensation formula is: Pressure Compensation = Thickness Deviation × Compressive Modulus × Nonlinear Correction Factor, where the nonlinear correction factor is obtained through fitting experimental data and ranges from 0.8 to 1.2.

[0090] Preferably, pressure adjustment of the hydraulic system is achieved through closed-loop control of a proportional relief valve and a pressure sensor. Each pressure zone is equipped with an independent hydraulic circuit, including a hydraulic pump, a proportional relief valve, a pressure sensor, and an actuating hydraulic cylinder. The proportional relief valve receives a control signal and adjusts the circuit pressure. The control signal is an analog voltage of 0-10V, corresponding to a linear relationship from 0 to the maximum working pressure. The pressure sensor monitors the actual pressure in real time, and the feedback signal is compared with the set value to form closed-loop control. When pressure adjustment is required, the control system calculates a new control voltage value and outputs it to the proportional relief valve via a digital-to-analog converter. The response time for pressure adjustment depends on the dynamic characteristics of the hydraulic system, typically reaching a steady state in 2-5 seconds. The upper limit of the equipment pressure is determined by the rated pressure of the hydraulic pump and the safety valve setpoint, typically 10MPa.

[0091] In one possible implementation, the recalculation process employs a parameter update mechanism. The recalculation of the resin flow state is based on the updated pressure field; increased pressure leads to higher resin flow velocity and a longer flow distance. The degree of interlayer adhesion is assessed through changes in contact area; increased pressure flattens more microscopic protrusions, increasing the contact area. The update of the interface defect distribution considers the compressive effect of pressure on bubbles; under high pressure, bubble volume decreases or even disappears. Thickness deviation prediction comprehensively considers the changes in the above factors and calculates the new thickness distribution through numerical integration.

[0092] For example, the improvement rate is evaluated using the reduction in relative error as the indicator. The maximum deviation of the first thickness distribution map is denoted as D1, and the maximum deviation of the second thickness distribution map is denoted as D2. The improvement rate is defined as (D1-D2) / D1×100%. The preset improvement threshold is set according to process requirements, typically 30%. If the improvement rate does not reach the threshold, further iterative optimization may be necessary. The number of iterations is limited to 3 to avoid infinite loops. The pressure adjustment range of each iteration gradually decreases, using a decay factor of 0.7 to ensure convergence. Furthermore, the verification of the second thickness distribution map is completed by comparing it with measured data. In actual production, representative locations are selected for thickness measurement, with no fewer than 20 measurement points evenly distributed on the workpiece surface. When the correlation coefficient between the measured value and the predicted value is greater than 0.9, the prediction model is considered effective.

[0093] Understandably, optimizing pressure parameters not only improves thickness uniformity but also affects the material's microstructure. Appropriate pressure promotes resin penetration between fibers, reduces porosity, and increases interlaminar shear strength. Excessive pressure, however, can lead to excessive resin loss and excessive fiber volume fraction, ultimately reducing mechanical properties.

[0094] Based on the embodiments of the present invention described above, and through the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for analyzing the thickness uniformity of prepregs based on deep learning, characterized in that, include: Data on fiber orientation interlacing during the multilayer lamination of prepreg is obtained, and images of fiber angle distribution are acquired to generate a first fiber distribution map. The fiber orientation interlacing features are extracted from the first fiber distribution map, and regions at different angles are separated to obtain the fiber angle distribution; Based on the analysis of the fiber angle distribution, resin flow unevenness is analyzed, and a resin flow distribution map is generated. This includes: calculating the change rate of void volume at fiber intersections based on the angle difference between adjacent regions in the fiber angle distribution, determining the local permeability coefficient, and obtaining the resin flow resistance distribution; calculating the resin pressure field using the finite difference method based on the flow resistance distribution, and recording the coordinates of resin-rich regions where the pressure gradient exceeds a threshold; calculating the resin climbing height based on the ratio of the cumulative volume of the resin-rich region to the fiber gap capacity, and generating a resin flow distribution map including the locations of micro-protrusions; simulating the impact of pressure changes on interlayer adhesion based on the resin flow distribution map, and generating a first adhesion distribution map that integrates resin flow information; extracting micro-protrusion regions and bubble distribution information from the first adhesion distribution map, and using a convolutional neural network to identify the interaction regions of micro-protrusions and bubble thickness fluctuations, obtaining a first interface feature set; and based on the first interface feature set, analyzing the coupling effect between fiber orientation interlacing and interlayer adhesion changes through multi-scale feature extraction, combining a deep learning network for defect classification, and outputting a distribution map of defect location and severity. Based on the defect location and severity distribution map, and combined with thickness measurement data, the thickness distribution deviation caused by the interaction between micro-protrusions and bubbles is predicted, and a first thickness distribution map is generated; the stacking pressure parameters are adjusted according to the deviation distribution of the first thickness distribution map, and a second thickness distribution map is regenerated.

2. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The process of acquiring fiber orientation interlacing data during the multilayer lamination of prepreg, collecting fiber angle distribution images, and generating a first fiber distribution map includes: determining the boundary coordinates of the prepreg sample using the positioning device of an optical scanning device, identifying fiber orientation interlacing regions based on changes in fiber weaving density; for the fiber orientation interlacing regions, using a line scan camera in conjunction with a coaxial light source to scan line by line, acquiring a grayscale image sequence based on the difference in fiber surface reflectivity; based on the grayscale image sequence, using Fourier transform to extract the frequency domain features of the fiber texture, calculating the fiber orientation angle based on the position of the main peak in the spectrum, and generating a first fiber distribution map.

3. The method for analyzing the thickness uniformity of prepregs based on deep learning according to claim 1, characterized in that, The step of extracting fiber orientation interlacing features from the first fiber distribution map and separating different angle regions to obtain the fiber angle distribution includes: reading the fiber angle values ​​of pixels from the first fiber distribution map, calculating the angle gradient through the angle difference between adjacent pixels, and marking boundary points where the angle gradient exceeds a threshold; based on the boundary points, segmenting the first fiber distribution map using a watershed algorithm, calculating the standard deviation of pixel angle values ​​within the segmented region, retaining regions with a standard deviation less than a threshold to obtain a set of angle-consistent regions; statistically analyzing the average fiber angle of the set of angle-consistent regions, constructing an interlacing feature matrix, and outputting the fiber angle distribution.

4. The method for analyzing the thickness uniformity of prepregs based on deep learning according to claim 1, characterized in that, The method of simulating the effect of pressure changes on interlayer adhesion based on the resin flow distribution map and generating a first adhesion distribution map that incorporates resin flow information includes: extracting the coordinates and height data of the protrusion positions from the resin flow distribution map; calculating the protrusion compression amount based on the stacking pressure time series data to obtain the post-compression protrusion height distribution; calculating the distance field between the upper and lower layer surfaces based on the post-compression protrusion height distribution; calculating the ratio of the contact point area to the total area to obtain the initial adhesion value; and updating the adhesion value of the overlapping area based on the initial adhesion value and the resin lateral flow distance to generate the first adhesion distribution map.

5. The method for analyzing the thickness uniformity of prepregs based on deep learning according to claim 1, characterized in that, The process of extracting micro-protrusion regions and bubble distribution information from the first fit distribution map, and using a convolutional neural network to identify the interaction regions of micro-protrusions and bubble thickness fluctuations to obtain a first interface feature set includes: calculating the fit gradient from the first fit distribution map using a gradient operator, marking candidate protrusion points with gradient magnitudes exceeding a threshold and bubble positions with fit less than a threshold, and obtaining a protrusion coordinate set and a bubble contour set; constructing a neural network containing multi-scale convolutional kernels based on the protrusion coordinate set and bubble contour set, concatenating feature maps with different receptive fields to generate a multi-scale feature map; calculating the spatial distance matrix between protrusions and bubbles based on the multi-scale feature map, determining the interaction region, calculating the thickness change rate, constructing an interaction intensity index, and outputting a first interface feature set containing protrusion positions, bubble contours, and interaction intensity values.

6. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The process, based on a first interface feature set, analyzes the coupling effect between fiber orientation interlacing and interlayer fit variation through multi-scale feature extraction, and classifies defects using a deep learning network to output a defect location and severity distribution map. This includes: extracting interaction intensity values ​​and spatial location information from the first interface feature set; calculating the coupling coefficient between fiber orientation angle difference and fit value; extracting multi-scale feature vectors using pyramid pooling; constructing a residual neural network based on the multi-scale feature vectors to extract deep features; inputting these features into a classifier to determine the defect type and severity; and based on the defect type and severity, mapping the data to the original image size through interpolation, labeling the defect category and severity values, merging adjacent defect regions of the same type, and outputting a defect distribution map.

7. The deep learning-based prepreg thickness uniformity analysis method of claim 1, wherein, The method of predicting the thickness distribution deviation caused by the interaction between micro-bumps and bubbles based on the defect location and severity distribution map and thickness measurement data to generate a first thickness distribution map includes: extracting defect coordinates from the defect location and severity distribution map and calculating a weighted thickness value based on the thickness measurement data; identifying overlapping areas of bumps and bubbles based on the weighted thickness value, calculating local thickness changes, and generating a thickness deviation field through interpolation; and adjusting the thickness values ​​of the bump and bubble positions by superimposing the original thickness data according to the thickness deviation field to generate the first thickness distribution map.

8. The deep learning-based prepreg thickness uniformity analysis method of claim 1, wherein, The step of adjusting the lamination pressure parameters according to the deviation distribution of the first thickness distribution map and regenerating the second thickness distribution map includes: calculating the difference between the thickness value and the target thickness from the first thickness distribution map, extracting the coordinates of the deviation exceeding the limit area, and determining the pressure control zone number; calculating the pressure compensation amount based on the pressure control zone number and updating the pressure parameters; and using the updated pressure parameters to recalculate the resin flow state, fit, defect distribution, and thickness deviation to generate the second thickness distribution map.

Images