Prepreg thickness uniformity analysis method based on deep learning

Abstract

The invention provides a prepreg thickness uniformity analysis method based on deep learning, and the method comprises the steps: obtaining fiber orientation staggering data in a prepreg multi-layer lamination process, collecting a fiber angle distribution image, and generating a first fiber distribution diagram; extracting fiber orientation staggering characteristics from the first fiber distribution diagram, and separating different angle regions to obtain fiber angle distribution; analyzing the non-uniform resin flow condition according to the fiber angle distribution, and generating a resin flow distribution diagram containing the positions of resin extrusion tiny bulges; based on the resin flow distribution diagram, in combination with gradual adjustment data of lamination pressure of multiple layers of prepregs from low to high, simulating the influence of pressure change on the interlayer fitting degree, and generating a first fitting degree distribution diagram fused with resin flow information; based on the defect position and severity distribution map, combining real-time thickness measurement data to predict thickness distribution deviation caused by interaction of tiny protrusions and bubbles, and generating a first thickness distribution map;

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a prepreg thickness uniformity analysis method based on deep learning. BACKGROUND

[0002] The prepreg multi-layer lamination technology directly affects the stability of material performance and product reliability, especially in high-precision industries such as aerospace and automotive. Thickness uniformity is a core indicator of product quality. Traditional detection methods mainly rely on single physical measurement or simple image analysis, which is difficult to cope with the dynamic changes of complex interlayer interfaces. These methods often lack the ability to dynamically capture micro features when facing complex situations such as fiber orientation interlacing and interlayer adhesion changes, resulting in insufficient detection accuracy, especially in the prediction of interface defects and thickness distribution. The complexity of the interlayer interface is a core challenge in this field. First, the fiber orientation interlacing of the upper and lower sheets significantly affects the formation of interface features. For example, during the lamination process, the difference in fiber angle can cause uneven resin flow, which in turn forms micro protrusions or bubble residues. These microstructures directly affect the thickness uniformity. Second, the change in interlayer adhesion when the lamination pressure increases further exacerbates the complexity of the interface features. For example, excessive pressure can squeeze out too much resin, forming irregular protrusions, while insufficient pressure can cause bubble residues, causing thickness fluctuations. These factors are interrelated, and the formation of micro protrusions can amplify the impact of bubble residues, making it more difficult to predict thickness distribution. Specifically, in actual production, when the pressure is gradually adjusted from low to high, the detection system needs to identify the interactive effects between the micro protrusions caused by resin extrusion and the thickness fluctuations caused by bubbles in real time. However, existing methods are difficult to dynamically adapt to such multi-scale feature changes. Therefore, how to dynamically capture the micro interface features caused by fiber orientation interlacing and interlayer adhesion changes under varying lamination pressure, and accurately identify defects and predict thickness distribution, has become a key problem. SUMMARY

[0003] The present application provides a prepreg thickness uniformity analysis method based on deep learning, mainly including:

[0004] The fiber orientation interlacing data in the prepreg multi-layer lamination process is acquired, the fiber angle distribution image is collected, and a first fiber distribution map is generated; the fiber orientation interlacing features are extracted from the first fiber distribution map, different angle regions are separated, and a fiber angle distribution is obtained; resin flow unevenness is analyzed according to the fiber angle distribution, and a resin flow distribution map is generated; based on the resin flow distribution map, the influence of pressure change on interlayer adhesion is simulated, and a first adhesion distribution map fused with resin flow information is generated; the micro convex region and bubble distribution information are extracted from the first adhesion distribution map, the convolutional neural network is used to identify the interaction region of the micro convex and bubble thickness fluctuation, and a first interface feature set is obtained; based on the first interface feature set, the coupling effect of fiber orientation interlacing and interlayer adhesion change is analyzed through multi-scale feature extraction, the defect classification is combined with the deep learning network, and the defect position and severity distribution map are output; based on the defect position and severity distribution map, the thickness distribution deviation caused by the interaction of micro convex and bubbles is predicted combined with the thickness measurement data, and a first thickness distribution map is generated; the lamination pressure parameters are adjusted according to the deviation distribution of the first thickness distribution map, and a second thickness distribution map is regenerated.

[0005] Further, the acquisition of the fiber orientation interlacing data in the prepreg multi-layer lamination process, the collection of the fiber angle distribution image, and the generation of the first fiber distribution map include:

[0006] The boundary coordinates of the prepreg sample are determined by the positioning device of the optical scanning equipment, and the fiber orientation interlacing region is identified according to the change of fiber weaving density; for the fiber orientation interlacing region, a line array camera is used to scan row by row with a coaxial light source, and a gray scale image sequence is acquired according to the difference in fiber surface reflectivity; based on the gray scale image sequence, the frequency domain features of the fiber texture are extracted by using Fourier transform, the fiber orientation angle is calculated according to the main peak position of the spectrum, and the first fiber distribution map is generated.

[0007] Further, the extraction of the fiber orientation interlacing features from the first fiber distribution map, the separation of different angle regions, and the obtaining of the fiber angle distribution include:

[0008] The fiber angle values of the pixel points are read from the first fiber distribution map, the angle gradient is calculated by the angle difference value of the adjacent pixels, and the boundary points with the angle gradient exceeding the threshold value are marked; based on the boundary points, the first fiber distribution map is segmented by using the watershed algorithm, the standard deviation of the pixel angle values in the segmented region is calculated, the region with the standard deviation less than the threshold value is retained, and a set of angle consistency regions is obtained; the average fiber angle of the set of angle consistency regions is counted, the interlacing feature matrix is constructed, and the fiber angle distribution is output.

[0009] Further, the analysis of resin flow unevenness according to the fiber angle distribution and the generation of the resin flow distribution map include:

[0010] According to the angle difference of adjacent regions in the fiber angle distribution, the void volume change rate at the fiber intersection is calculated, the local permeability coefficient is determined, and the resin flow resistance distribution is obtained; based on the flow resistance distribution, the finite difference method is used to calculate the resin pressure field, and the coordinates of the resin enrichment area with a pressure gradient exceeding the threshold value are recorded; according to the ratio of the cumulative volume of the resin enrichment area to the fiber gap capacity, the resin climbing height is calculated, and a resin flow distribution map containing the position of the micro protrusion is generated.

[0011] Further, based on the resin flow distribution map, the influence of pressure change on the interlayer adhesion is simulated, and a first adhesion distribution map containing resin flow information is generated, including:

[0012] The protrusion position coordinates and height data are extracted from the resin flow distribution map, the protrusion compression amount is calculated according to the time series data of the lamination pressure, and the post-compression protrusion height distribution is obtained; based on the post-compression protrusion height distribution, the surface distance field of the upper and lower layers is calculated, the area ratio of the contact points to the total area is calculated, and the initial adhesion value is obtained; according to the initial adhesion value, the resin transverse flow distance is combined to update the adhesion value of the overlapping area, and the first adhesion distribution map is generated.

[0013] Further, the micro protrusion region and bubble distribution information are extracted from the first adhesion distribution map, and a convolutional neural network is used to identify the interaction region of the micro protrusion and the bubble thickness fluctuation to obtain a first interface feature set, including:

[0014] The adhesion gradient is calculated from the first adhesion distribution map by a gradient operator, the protrusion candidate points with a gradient amplitude exceeding a threshold value and the bubble positions with an adhesion less than a threshold value are marked, and the protrusion coordinate set and the bubble contour set are obtained; based on the protrusion coordinate set and the bubble contour set, a neural network containing a multi-scale convolution kernel is constructed, the feature maps of different receptive fields are spliced, and a multi-scale feature map is generated; according to the multi-scale feature map, the spatial distance matrix of the protrusion and the bubble is calculated, the interaction region is determined, the thickness change rate is calculated, the interaction intensity index is constructed, and the first interface feature set containing the protrusion position, the bubble contour and the interaction intensity value is output.

[0015] Further, based on the first interface feature set, the coupling effect of fiber orientation intersection and interlayer adhesion change is analyzed by multi-scale feature extraction, and defect classification is performed combined with a deep learning network to output a defect position and severity distribution map, including:

[0016] The interaction intensity value and the spatial position information are extracted from the first interface feature set, a coupling coefficient of a fiber orientation angle difference and a fit degree value is calculated, and a multi-scale feature vector is extracted through pyramid pooling; based on the multi-scale feature vector, a residual neural network is constructed to extract deep features, and a classifier is input to determine the defect type and severity; based on the defect type and severity, an interpolation is mapped to the original image size, a defect category and a severity value are labeled, adjacent defect regions of the same category are merged, and a defect distribution map is output.

[0017] Further, based on the defect position and severity distribution map, the thickness distribution deviation caused by the interaction of the micro protrusions and the bubbles is predicted by combining the thickness measurement data, and a first thickness distribution map is generated, including:

[0018] Defect coordinates are extracted from the defect position and severity distribution map, and a weighted thickness value is calculated by combining the thickness measurement data; based on the weighted thickness value, an overlapping area of the protrusions and the bubbles is identified, a local thickness change amount is calculated, and a thickness deviation field is generated through interpolation; according to the thickness deviation field, the original thickness data is superimposed, the thickness value of the protrusion and bubble position is adjusted, and the first thickness distribution map is generated.

[0019] Further, the deviation distribution of the first thickness distribution map is adjusted to adjust the lamination pressure parameters, and a second thickness distribution map is regenerated, including:

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

[0021] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0022] The present application discloses a semi-cured sheet thickness uniformity analysis method based on deep learning, which is aimed at the business scenario problems of resin flow unevenness caused by fiber orientation stagger, thickness deviation and interlayer fit defects caused by the interaction of micro protrusions and bubbles, acquires fiber angle distribution through high-resolution optical scanning, extracts orientation features by using image segmentation algorithm, simulates resin flow and fit degree change by combining pressure adjustment data, identifies the interaction area of micro protrusions and bubbles by using multi-scale feature fusion convolutional neural network, and then analyzes the coupling effect of fiber orientation and fit degree, classifies defects and predicts thickness deviation. When the thickness deviation is out of standard, the lamination pressure parameters are dynamically adjusted, and the second thickness distribution map is iteratively optimized. The present application significantly improves the semi-cured sheet lamination quality, reduces the defect rate, optimizes the uniformity of thickness distribution, and guarantees the manufacturing precision and reliability of high-performance composite materials by multi-scale feature fusion and dynamic pressure adjustment. 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 array CCD camera adopts a push-broom imaging mode. A coaxial light source vertically irradiates the surface of the prepreg, and the intensity of the light source is adjusted to an appropriate range according to the refractive index of the fiber material. The intensity of the carbon fiber material is set to 1.2 times the standard value, and the intensity of the glass fiber material is set to 0.8 times the standard value. The camera moves at a constant speed in the transverse direction, and each time a row of pixel data is collected. The row spacing is determined according to the width of the fiber bundle, and is usually set to one-third of the average width of the fiber bundle. Adjacent scanning rows in the gray-scale image sequence are registered by SURF feature point extraction, and the nearest neighbor distance ratio method is used for feature point matching. When the number of matched point pairs exceeds a predetermined threshold, a single registration is completed.

[0030] In one possible implementation, the specific process of Fourier transform processing the gray-scale image sequence includes performing two-dimensional fast Fourier transform on each image block, and the size of the image block is set to 128x128 pixels. In the transformed frequency spectrum, the main frequency component corresponding to the fiber texture appears as a bright stripe extending from the origin, and the stripe direction is perpendicular to the fiber orientation. Straight line features in the frequency spectrum are detected by Hough transform, and the angle of the straight line plus 90 degrees is the fiber orientation angle of the corresponding region. The angle distribution histogram is counted in intervals of 5 degrees, and the frequency of each interval reflects the distribution density of the fibers in that angle range.

[0031] Preferably, an angle interpolation process is introduced in the generation process of the first fiber distribution map. For scanning blind areas or regions with unclear features, a bilinear interpolation method is used to estimate the fiber orientation at that point according to the surrounding known angle values. The interpolation weight is determined according to the inverse distance to ensure the continuity of the angle transition. The generated first fiber distribution map is displayed in pseudo-color form, 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, achieving intuitive visualization of the fiber orientation distribution.

[0032] Step S102, extracting the fiber orientation staggered feature from the first fiber distribution map, separating different angle regions, and obtaining the fiber angle distribution.

[0033] The fiber angle value of each pixel point is read from the first fiber distribution map, the angle difference value is obtained by subtracting the angle value of adjacent pixels, and the angle gradient is taken as the angle difference value. When the absolute value of the angle gradient exceeds a preset threshold, it is marked as a potential interleaving boundary point, and an initial boundary line is formed by eight-neighbor connection judgment. Based on the initial boundary line, a watershed algorithm is used to perform region segmentation on the fiber distribution map, the standard deviation of the angle value of all pixels in each segmented region is calculated to judge the homogeneity of the region, and if the standard deviation is less than a preset threshold, the region is retained, otherwise the region is subjected to secondary segmentation until the homogeneity requirement is met, and a set of angle consistency regions is obtained. For the set of angle consistency regions, the average fiber angle and the area of each region are counted, the absolute value of the difference between the average angles of adjacent regions is used to construct an interleaving feature matrix, and 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 fiber angle distribution containing the region number, the average angle and the interleaving relationship is output.

[0034] In an embodiment, the calculation of the angle gradient is realized by using the central difference method. For a pixel point with coordinates (x, y) in the first fiber distribution map, the horizontal direction angle gradient is obtained by calculating the angle difference between the positions (x+1, y) and (x-1, y) and then dividing by 2, and the vertical direction 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 is cyclic in the range of 0 degrees to 180 degrees, and when two angle values cross the 0 degree or 180 degree boundary, the smaller angle difference is selected as the actual gradient value.

[0036] Specifically, the eight-neighbor connection judgment process is realized by scanning each labeled potential interleaving boundary point. For each boundary point, it is checked whether the eight adjacent pixels around it are also labeled as boundary points. If there is a continuous sequence of boundary points, a complete boundary line is formed by a chain code tracking method. The broken boundary line segments are connected by endpoint distance judgment, and 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 process 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 to form a terrain map. Local minimum points are found in the terrain map as seed points, and the initial segmentation is completed by expanding outward from the seed points until the watershed ridge line is encountered. For regions with an area that is too small, the region is merged with the most similar adjacent region, and the similarity is judged by the average angle difference between the two regions.

[0038] In one possible implementation, the secondary segmentation of region homogeneity is performed in a recursive manner. When the angle standard deviation of a region exceeds the threshold, edge detection is re-performed within the region to find secondary interleaving boundaries. The detection threshold of secondary boundaries is set to 0.7 times the initial threshold, ensuring that weaker angle changes can be identified. The segmented sub-regions are again subjected to homogeneity test until all regions meet the requirements or reach the minimum region area limit.

[0039] Exemplarily, in the construction process of the interleaving feature matrix, in addition to recording the angle difference, the length of the shared boundary between regions is also considered. The matrix element value is weighted by the product of the angle difference and the boundary length, and the longer the shared boundary between adjacent regions, the more obvious the interleaving feature. The output fiber angle distribution data structure includes three parts: region geometry information records the boundary coordinates and area of each region, angle statistical information records the average angle and angle distribution range, and interleaving relationship information stores the interleaving degree value between adjacent regions through a sparse matrix.

[0040] Step S103, according to the resin flow distribution analysis of the fiber angle distribution, a resin flow distribution map containing the position of the resin extrusion micro-protrusion is generated.

[0041] According to the angle difference between adjacent regions in the fiber angle distribution, the void volume change rate at the fiber interleaving position is calculated, when the fiber angle difference exceeds the preset threshold, the local permeability coefficient is determined according to the ratio of the void volume to the fiber spacing, and the flow resistance distribution of the resin in different directions is obtained. Based on the flow resistance distribution, the continuity equation is solved by using the finite difference method, and the resin pressure field is obtained by iterative calculation of the pressure value on the grid node. When the local pressure gradient exceeds the critical value, it is determined that the resin enrichment occurs in the region, and the enrichment region coordinates and the resin cumulative volume are recorded. For the resin enrichment region and the cumulative volume, the overflow degree is determined according to the ratio of the cumulative volume to the fiber gap capacity, and if the ratio is greater than the preset threshold, the climbing height of the resin on the fiber surface is calculated according to the difference between the surface tension of the resin and the surface energy of the fiber, and the height distribution data of the micro-protrusion is obtained. Based on the height distribution data of the micro-protrusion, a continuous resin thickness field is constructed by an interpolation method, a bicubic interpolation is used in the fiber interleaving dense region to maintain the detail features, a linear interpolation is used in the fiber parallel region to improve the calculation efficiency, and the protrusion position coordinates and the thickness value are combined to generate a resin flow distribution map containing the position of the resin extrusion micro-protrusion.

[0042] In one embodiment, the correlation between fiber angle distribution and resin flow resistance is achieved by multi-physics coupling. The rate of change of void volume at fiber intersections depends on the relative angles of the upper and lower fibers, when the two layers of fibers are 90 degrees staggered, the rhombus-shaped voids formed are the smallest, and the resin flow resistance is the largest; when the fibers are arranged in parallel, the rectangular channel voids formed are the largest, and the flow resistance is the smallest. The void volume is calculated by the trigonometric function relationship of the fiber bundle width, thickness and stagger angle, specifically the void area multiplied by the unit thickness, wherein the void area is equal to the square of the fiber bundle spacing multiplied by the sine value of the stagger angle.

[0043] It should be noted that the determination process of the local permeability coefficient involves porous medium theory. The permeability coefficient is proportional to the cube of the void volume and inversely proportional to the square of the fiber surface area. During the process of prepreg lamination, the compression of the fiber bundle changes the void structure, and the permeability coefficient changes dynamically with the pressure. After obtaining the reference permeability coefficient through experimental calibration, the correction is made according to the ratio of the actual void volume to the reference void volume. The flow resistance distribution forms a two-dimensional tensor field, the main direction is along the fiber arrangement direction, and the resistance in the vertical direction is usually 3 to 5 times that in the main direction.

[0044] Specifically, the implementation process of the finite difference method for solving the continuity equation adopts staggered grid discretization. The pressure nodes are located at the center of the grid, and the velocity components are defined at the midpoint of the grid boundary. The continuity equation is expressed as the conservation of resin mass, i.e. the mass of resin flowing into the control volume per unit time is equal to the mass flowing out plus the change of resin mass in the control volume. After discretization, a linear equation system is formed, each internal node corresponds to an equation, and the boundary nodes are set according to the actual boundary conditions. The iterative solution adopts the over-relaxation method, the relaxation factor is determined according to the grid density and the permeability coefficient distribution, and the value is usually between 1.2 and 1.8. The convergence criterion is that the maximum relative error of the pressure field of the adjacent two iterations is less than a preset threshold. The pressure gradient is calculated by the pressure difference between adjacent nodes divided by the node spacing, and when the gradient value exceeds the critical driving pressure of resin flow, it is marked as a potential enrichment area.

[0045] Preferably, the determination of resin enrichment introduces a time accumulation effect. The resin accumulation volume depends not only on the instantaneous flow rate but also on the duration. In each time step, the flow rate field is calculated according to the pressure gradient and the permeability coefficient, the flow rate is multiplied by the time step to obtain the displacement amount, the net accumulation volume is obtained by accumulating the resin inflow and subtracting the outflow in each time step. The fiber gap capacity is calculated according to the fiber volume fraction, and the fiber volume fraction of the prepreg is usually 0.5 to 0.6, and the remaining space is the volume that can be occupied by the resin.

[0046] In one possible implementation, the calculation of the resin climbing height is based on the principle of capillary phenomenon. The climbing of resin on the fiber surface is driven by surface tension, and the climbing height is proportional to the surface tension and inversely proportional to the product of the resin density and the acceleration of gravity. The surface tension is determined by the pendant drop method, and the fiber surface energy is obtained by the contact angle measuring instrument. When the interfacial tension between the resin and the fiber is less than the cohesive force of the resin, the resin spreads on the fiber surface and climbs upward. The meniscus curvature formed during the climbing process determines the final convex shape, and the curvature radius is related to the pressure difference through the Laplace equation.

[0047] Exemplarily, the processing of the micro-convex height distribution data adopts a statistical method. Gaussian fitting is performed on the convex height of each enrichment area to obtain the average height and standard deviation. Abnormally high convexities are identified by a 3-sigma criterion, and these positions usually correspond to fiber breakage or misalignment defects. The lateral extension range of the convexity is determined by the gradient change of the height, and the height is considered to reach the convex boundary when it drops to 10% of the peak value. Further, the selection of the interpolation method is based on the local fiber arrangement characteristics. The bicubic interpolation constructs a cubic polynomial surface through 16 adjacent points, maintaining the continuity and smoothness of the height field. In the fiber interlaced dense area, the sampling points are set according to the fiber bundle scale to ensure the capture of detail changes. Linear interpolation maintains basic continuity while reducing computational complexity, and is suitable for areas with gentle changes. The interpolation weight is determined inversely proportional to the distance, and the influence of the near neighbor point is greater than that of the far end point.

[0048] It can be understood that the data structure of the resin flow distribution map adopts a hierarchical organization. The bottom layer stores the basic thickness field in a regular grid, the middle layer records the convex position and height increment in a sparse matrix, and the top layer describes the convex profile in a vector data.

[0049] Step S104, based on the resin flow distribution map, gradually adjusts the data from low to high in combination with the multilayer lamination pressure of the prepreg, simulates the influence of pressure change on the interlayer adhesion, and generates a first adhesion distribution map that integrates the resin flow information.

[0050] The protrusion position coordinates and height data are extracted from the resin flow distribution map, a pressure-time curve is established according to the lamination pressure time series data, the compression amount of the protrusion under different pressures is calculated through the relationship between pressure and deformation in the Hertz contact theory, and the post-compression protrusion height distribution is obtained by subtracting the compression amount from the initial protrusion height. Based on the post-compression protrusion height distribution, the distance field of the upper and lower prepreg surfaces is calculated, and 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 initial bonding degree value is obtained by calculating the ratio of the area of all contact points to the total area. For the initial bonding degree value, combined with the lateral flow distance of the resin from the top of the protrusion to the surrounding area under the action of pressure, the lateral flow distance is determined by the ratio of the resin viscosity and the pressure gradient. If the covered area of the resin after flowing overlaps with the covered area of the adjacent protrusion, the bonding degree value of the overlapping area is updated to the fully bonded state, and the corrected bonding degree field is obtained. Based on the corrected bonding degree field, the resin thickness value at each position is extracted from the resin flow distribution map, and the comprehensive bonding feature is constructed by the weighted sum of the bonding degree value and the resin thickness value. The weight coefficient is dynamically adjusted according to the pressure size, and the first bonding degree distribution map integrating the resin flow information is output.

[0051] In one embodiment, the application of Hertz contact theory needs to consider the elastic modulus and Poisson's ratio of the prepreg material. The resin protrusion can be regarded as a spherical indenter on an elastic half-space, and when a vertical pressure is applied, a circular indentation is formed in the contact area. The relationship between compression and pressure follows a power function rule, and the index value depends on the geometry of the protrusion. For a hemispherical protrusion, the compression amount is proportional to the 2 / 3 power of the pressure; for a conical protrusion, the compression amount is proportional to the 1 / 2 power of the pressure. The actual protrusion shape is between the two, and is corrected by a shape factor.

[0052] It should be noted that the establishment of the pressure-time curve is based on the actual lamination process parameters. The prepreg lamination process is usually divided into three stages: the initial contact stage, the pressure gradually increases from zero to 30% of the working pressure, the pressure maintaining stage, the pressure is maintained constant, and the pressure releasing stage, the pressure slowly decreases. The duration of each stage and the pressure change rate are determined according to the material properties and product requirements. The pressure sensor collects pressure data at a fixed frequency, and the time interval is usually set to 0.1 seconds to ensure that the dynamic process of pressure change is captured.

[0053] Specifically, the calculation of the distance field adopts the Euclidean distance transform method. For each point on the upper and lower prepreg surfaces, the vertical distance to the nearest surface feature point is calculated. The sum of the two distance fields, with the bottom surface of the upper layer as the reference and the top surface of the lower layer as the reference, is the interlayer distance field. In the fiber interlaced region, due to the superposition of fiber thickness, the local distance field will fluctuate. The resolution of the distance field matches the scanning accuracy, which is usually 10 sampling points per millimeter. When the post-press protrusion height value is greater than the distance field value at the corresponding position, it indicates that actual contact occurs at the position. The contact area is identified by binary processing to form a contact mask map. The contact area is obtained by counting the number of pixels with a value of 1 in the mask map, and the ratio of the total number of pixels is the fit degree.

[0054] Preferably, the calculation of the resin lateral flow distance introduces the extrusion flow theory. Under the action of pressure, the resin at the top of the protrusion is extruded, generating radial flow. The flow velocity is proportional to the pressure gradient and inversely proportional to the resin viscosity. The pressure gradient is obtained by dividing the difference between the center pressure and the edge pressure of the protrusion by the protrusion radius. The resin viscosity varies with temperature and shear rate, which is described by a power-law fluid model. The flow distance is equal to the integral of the flow velocity over time, and the upper limit of the integral is the pressure action 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 the local fit strength.

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

[0056] Exemplarily, the construction process of the comprehensive fit feature needs to consider the weight distribution of multiple factors. The fit value reflects the degree of geometric contact, with a value range of 0 to 1; the resin thickness value reflects the physical thickness of the bonding layer, which is also mapped to the range of 0 to 1 after normalization processing. The determination of the weight coefficient is based on the pressure size. When the pressure is low, the resin thickness weight is higher, because at this time it mainly relies on resin bonding; when the pressure is high, the fit weight is higher, because at this time a close mechanical contact is formed. The weight coefficient changes between two extreme values through linear interpolation. When the pressure is zero, the resin thickness weight is 0.7 and the fit weight is 0.3; when the pressure reaches the maximum value, the resin thickness weight is 0.3 and the fit weight is 0.7. In one embodiment, the first fit distribution map contains multiple channels (fit, resin thickness, pressure), each channel reflecting different physical states. The weight coefficient provides a mechanism to weight and integrate these different dimensional information (geometric contact, material thickness) into a unified comprehensive evaluation index. Further, the data structure of the first fit 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, and each channel can be displayed independently or in combination.

[0057] It can be understood that the spatial resolution of the fit distribution map needs to match the actual application requirements. For aerospace-grade composite materials, the resolution requirement reaches the level of 0.1 millimeter; for automotive industry applications, the resolution can be relaxed to the level of 0.5 millimeter. High resolution can capture small defects, but increases data processing and storage requirements.

[0058] Step S105, simultaneously extract the micro-bump region and bubble distribution information from the first fit distribution map, use a multi-scale feature fusion convolutional neural network to identify the interactive region of micro-bumps and bubble thickness fluctuations, and obtain a first interface feature set containing interactive features.

[0059] The horizontal and vertical direction gradients of the fit degree are calculated from the first fit degree distribution map by Sobel gradient operator, and the positions with gradient amplitude exceeding a preset threshold are marked as convex candidate points, and at the same time, the area with fit degree less than a preset lower limit value is identified as a bubble position by eight-connected domain marking method to obtain a convex coordinate set and a bubble contour set. Based on the convex coordinate set and the bubble contour set, a convolutional neural network containing three parallel branches is constructed, the branches respectively adopt 3×3, 5×5 and 7×7 convolution kernels to extract features of different receptive fields, and the output feature maps of the three branches are spliced in the channel dimension to obtain a fused multi-scale feature map. For the multi-scale feature map, the Euclidean distances between the convex region center point and each point on the bubble contour are calculated to form a spatial distance matrix, and when the minimum distance is less than a preset threshold, it is determined as a potential interaction area, and the thickness change rate is calculated by the thickness difference value of adjacent pixels divided by the pixel spacing in the interaction area. Based on the thickness change rate, the convex height and the bubble depth are normalized and multiplied, and then the thickness change rate is weighted and summed to construct an interaction intensity index, and according to the interaction intensity value, the interaction area is divided into three categories of strong, medium and weak, and the first interface feature set containing the convex position, the bubble contour, the interaction area category and the interaction intensity value is output.

[0060] In an embodiment, the application of Sobel gradient operator involves convolution operation in two directions. The 3×3 convolution kernel of [-1, 0, 1; -2, 0, 2; -1, 0, 1] is used in the horizontal direction, and the convolution kernel of [-1, -2, -1; 0, 0, 0; 1, 2, 1] is used in the vertical direction. Convolution operation is performed on each pixel position of the first fit degree distribution map respectively to obtain horizontal gradient Gx and vertical gradient Gy. The gradient amplitude is calculated by the square sum of Gx and Gy, and the gradient direction is determined by the inverse tangent function of Gy and Gx. When the gradient amplitude exceeds a preset threshold, the position is marked as a potential convex edge point.

[0061] It should be noted that the implementation process of the eight-connected domain marking method adopts a two-scan strategy. The first scan starts from the top-left corner and checks each pixel row by row and column by column. If the fit degree value of the current pixel is less than the bubble determination threshold, the marking conditions of the left-up, up, right-up and left four neighboring pixels of the current pixel are checked. If there is a marked pixel in the neighboring area, the current pixel inherits the smallest marking value; if there is no marked pixel in the neighboring area, a new marking value is assigned. The second scan solves the equivalent marking problem and unifies the pixels belonging to the same connected region but with different markings to the same marking. Each independent connected region represents a potential bubble, and the spatial features of the bubble are determined by calculating the circumscribed rectangle and the centroid position of the region.

[0062] Specifically, the construction of the multi-scale convolutional neural network adopts a parallel branch structure to realize multi-scale feature extraction. The first branch uses a 3x3 convolution kernel, with a small receptive field, mainly capturing local detailed features such as sharp edges of protrusions and fine textures of bubbles. The second branch uses a 5x5 convolution kernel, with a medium receptive field, capturing medium-scale structural features such as the overall morphology of protrusions and the basic outline of bubbles. The third branch uses a 7x7 convolution kernel, with a large receptive field, capturing global context information such as the distribution pattern of protrusion groups and the spatial arrangement of bubbles. Each branch contains two convolution layers, the first layer for feature extraction and the second layer for feature enhancement. After convolution operation, a batch normalization layer and a ReLU activation function are connected to maintain the non-linear representation ability of the features. The output feature maps of the three branches are concatenated in the channel dimension to form a fusion feature map containing multi-scale information. The number of channels of the concatenated feature map is the sum of the output channel numbers of the three branches, and a 1x1 convolution is performed for channel dimension reduction to reduce the 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, first determine its geometric center by taking the arithmetic mean of all pixel coordinates within the region. For each bubble contour, extract the sampling points on the contour, and adjust the sampling interval adaptively according to the contour perimeter. The element D[i,j] of the distance matrix represents the minimum Euclidean distance from the i-th protrusion center to all sampling points on the j-th bubble contour. The Euclidean distance is calculated by the square root of the sum of coordinate differences. When D[i,j] is less than the interaction judgment 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 variation rate needs to consider the thickness distribution of the local neighborhood. For each pixel within the interaction region, extract the thickness values within its 3x3 neighborhood. Fit a local plane by least squares method, and the gradient coefficient of the plane equation is the thickness variation rate of that point. The direction of the thickness variation rate indicates the direction of the fastest thickness increase, and the amplitude represents the degree of change. At the interface between the protrusion and the bubble, the thickness variation rate usually presents a local maximum value.

[0065] Exemplarily, 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 interval by maximum normalization, and the bubble depth is also normalized. The product of the two reflects the coupling degree of the morphological characteristics. The thickness variation rate is normalized after logarithmic transformation to reduce the influence of extreme values. The interaction intensity is calculated by weighted summation, and the weight coefficient is determined according to the material properties of the prepreg. The protrusion height weight of carbon fiber material is set to 0.4, the bubble depth weight is set to 0.3, and the thickness variation rate weight is set to 0.3; the corresponding weights of glass fiber material are 0.35, 0.35 and 0.3 respectively. Further, the classification of the interaction region uses a threshold segmentation method. The interaction intensity value in the range of 0-0.3 is marked as weak interaction, mainly showing the independent existence of protrusions and bubbles, and the mutual influence is small. The intensity value in the range of 0.3-0.7 is marked as moderate interaction, the protrusion edge partially overlaps with the bubble boundary, and the thickness distribution appears a transition zone. The intensity value greater than 0.7 is marked as strong interaction, the protrusion is completely covered by the bubble or the bubble surrounds the protrusion, forming a composite defect structure.

[0066] It can be understood that 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 contour is stored in the form of polygon vertex sequence, the interaction region is represented by a binary mask map, and the interaction intensity value is saved in the form of a floating-point number matrix.

[0067] Step S106, based on the interaction features in the first interface feature set, the coupling effect of fiber orientation interlacing and interlayer adhesion change is analyzed by a multi-scale feature extraction method, the defect classification of the prepreg is carried out combined with the multi-scale feature fusion strategy of the deep learning network, and the defect position and severity distribution map are output.

[0068] The interaction intensity value and the spatial position information are extracted from the first interface feature set, the fiber orientation angle difference and the fit degree value are normalized and multiplied to obtain a coupling coefficient, local features are extracted on different scale grids through pyramid pooling to obtain a multi-scale feature vector group containing coupling relationship. Based on the multi-scale feature vector group, a residual neural network is constructed for deep feature extraction, the network includes four residual blocks, each residual block retains the original feature information while learning residual mapping through jump connection, and the feature map is compressed into a fixed length feature vector through global average pooling. For the feature vector, a softmax classifier is input to determine the defect type, the defect is classified into four categories of layering, porosity, wrinkle and mixed, and the defect severity is determined according to the comparison of the maximum value of the classification probability and the preset threshold value, when the probability value is less than the first threshold value, it is slight, between the first and second threshold values, it is moderate, and greater than the second threshold value, it is serious. Based on the defect type and severity level, the classification result is mapped back to the original image size through bilinear interpolation, the defect category and severity value are labeled at each pixel position, and adjacent defect regions of the same type are merged through morphological closing operation, and a defect distribution map containing defect position coordinates, category label and severity value is output.

[0069] In an embodiment, the calculation of the coupling coefficient needs to consider the interaction of the fiber orientation and the fit degree in two dimensions. The fiber orientation angle difference is determined by the angle between the principal fiber directions of adjacent regions, and the value range is 0 to 90 degrees. When the angle difference is 0 degrees, the fibers are arranged in parallel, and the interlayer stress is uniformly transmitted; when the angle difference is 90 degrees, the fibers are arranged orthogonally, forming a stress concentration point. The angle difference is normalized by dividing by 90 degrees, which is mapped to the 0-1 interval. The fit degree value is directly used as it is in the 0-1 range. The coupling coefficient obtained by multiplying the two reflects the degree of structural heterogeneity, and the larger the value, the more likely it is to produce defects in the region.

[0070] It should be noted that the implementation of pyramid pooling adopts a spatial pyramid structure to realize multi-scale feature extraction. The input feature map is divided into grids of different granularities, the first layer is a 1x1 grid, and the global feature is extracted; the second layer is a 2x2 grid, and the region features of the four quadrants are extracted; the third layer is a 4x4 grid, and the local features of 16 sub-regions are extracted. The maximum pooling operation is performed in each grid to retain the most significant feature response. The pooling results of different levels are restored to the same size through upsampling, and then spliced in the channel dimension to form a feature representation containing multi-scale information.

[0071] Specifically, the construction of the residual neural network follows the identity mapping principle, ensuring effective gradient propagation. Each residual block contains two paths: the main path and the skip path. The main path contains two 3x3 convolution layers, with batch normalization and ReLU activation functions inserted in between. The first convolution maintains the number of channels, 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, and when the input and output dimensions do not match, a 1x1 convolution is used for dimension adjustment. The outputs of the two paths are fused by element-wise addition and then passed through the ReLU activation. Four residual blocks are connected in sequence, with channel numbers of 64, 128, 256, and 512, respectively, gradually extracting more abstract feature representations. The global average pooling at the end of the network averages the feature maps of each channel, compressing the spatial dimension to 1x1, and outputs 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 implementation process of defect classification requires pre-defining the feature patterns of defect categories. Hierarchical defects appear as continuous regions with local zero adhesion, usually in the form of strips or patches; porosity defects appear as scattered small-area low-adhesion points, randomly distributed; wrinkle defects appear as periodic adhesion fluctuations, accompanied by regular changes in fiber orientation; mixed defects have multiple defect characteristics and are difficult to classify into a single category. The input of 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 the score of a defect type. The softmax function converts the score into a probability distribution, and the class with the highest probability is taken as the prediction result.

[0073] In one possible implementation, the severity determination uses a hierarchical 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, it indicates that the defect features are not obvious, and it is determined to be a minor level; the probability is between 0.6 and 0.8, the defect features are relatively clear, and it is determined to be a moderate level; the probability is greater than 0.8, the defect features are very obvious, and it is determined to be a severe level. This hierarchical method takes into account the correlation between classification confidence and defect significance.

[0074] Exemplarily, the application of bilinear interpolation ensures the spatial correspondence of the classification results with the original image. After multiple pooling and convolution, 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 through the weighted average of the four adjacent points, and the weight is determined according to the inverse distance. The interpolation process is carried out step by step, and each time the resolution is doubled until it is restored to the original size. Further, a morphological closing operation is used to optimize the connectivity of the defect region. The closing operation first performs a dilation operation and then a corrosion operation using the same structural element. The dilation operation expands the boundary of the defect region, fills the internal cavities and connects adjacent regions; the corrosion operation shrinks the boundary to restore the original size. The structural element is selected as a square kernel of 3x3 or 5x5, which is determined according to the typical size of the defect.

[0075] It can be understood that the output format of the defect distribution map adopts a multi-layer structure organization. The bottom layer is a position coordinate matrix, which records the bounding box of each defect region; the middle layer is a category label matrix, which stores the corresponding defect type code for each pixel position; and the top layer is a severity matrix, which stores the quantized severity value.

[0076] In step S107, based on the defect position and severity distribution maps, the thickness distribution deviation caused by the interaction of the micro protrusion and the bubble is predicted by combining the real-time thickness measurement data, and a first thickness distribution map is generated.

[0077] The defect type and spatial coordinates are extracted from the defect position and severity distribution maps, the thickness data in the defect region and the surrounding area are collected by an ultrasonic thickness gauge, the influence weight is calculated by the inverse of the Euclidean distance between the measurement point and the defect center, the closer the distance, the greater the weight, and the weighted thickness value of each measurement point is obtained. Based on the weighted thickness value, the spatial overlap area of the micro protrusion and the bubble is identified, the protrusion height is obtained from the resin flow distribution map, the bubble depth is determined by the depth of the low value area in the adhesion distribution map, and the difference between the two forms a local thickness change value. The continuous thickness deviation field is obtained by spatial interpolation of the discrete change value through the Kriging interpolation method. For the thickness deviation field, a deviation amplification coefficient is given according to the defect severity, the severe defect coefficient is taken as the upper limit of the preset range, the moderate defect is taken as the middle value of the range, and the slight defect is taken as the lower limit of the range. The weighted deviation distribution is obtained by multiplying the deviation field and the amplification coefficient point by point. Based on the weighted deviation distribution, the original thickness data measured in real time is superimposed, the deviation value is added at the protrusion position and subtracted at the bubble position, the thickness boundary is smoothed by a Gaussian filter, and a first thickness distribution map containing the interaction influence of the micro protrusion and the bubble is output.

[0078] In one embodiment, the data acquisition of the ultrasonic thickness gauge adopts the principle of pulse-echo method. The ultrasonic pulse emitted by the probe penetrates the prepreg and produces reflected echoes on the upper and lower surfaces. By measuring the time difference between the transmitted pulse and the received echo, combined with the sound speed in the material, the thickness value is calculated. The arrangement of measurement points follows the principle of gridding. In the central area of the defect, a dense grid is used with a grid spacing of 2 mm; in the edge area of the defect, a sparse grid is used with a spacing of 5 mm; in the defect-free area, the spacing can reach 10 mm. This variable-density sampling strategy balances the requirements of measurement accuracy and efficiency.

[0079] It should be noted that the calculation of the weight is based on the principle of spatial attenuation. The Euclidean distance is obtained by measuring the square sum of the difference between the point coordinates and the defect center coordinates and taking the square root. 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 center of the defect, the maximum weight value is set to 0.5 to avoid abnormality caused by excessive weight of a single point. For measurement points with a distance exceeding the defect influence radius, the weight value is set to zero and does not participate in subsequent calculations. The weighted thickness value is obtained by multiplying and accumulating the thickness of each measurement point and the corresponding weight.

[0080] Specifically, the acquisition of 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 position relative to the reference surface. The bubble depth is obtained by analyzing the low value area in the fit degree distribution map. The vertical distance from the center point of the area with a fit degree close to zero to the surface is the bubble depth. In the spatial overlap area, the protrusion may partially fill the bubble space, and the actual thickness change needs to consider the relative position relationship between the two. When the protrusion is completely located 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 edge contact, the change is determined by geometric intersection calculation.

[0081] Preferably, the implementation of Kriging interpolation method requires the construction of variogram model. The variogram describes the correlation of thickness change in space, and the empirical variogram is obtained by calculating the semivariance of sample point pairs at different distances. Common theoretical models include spherical model, exponential model and Gaussian model, and the appropriate model is selected for fitting according to the shape of the empirical variogram. The model parameters include nugget value, sill value and range, which respectively reflect the measurement error, spatial variation intensity and correlation distance. The interpolation process obtains the weight coefficient of the estimated point by solving the Kriging equation set, and the coefficient matrix of the equation set is composed of the variogram values between the known points. The interpolation result not only provides the estimated value, but also gives the estimated variance, which is used to evaluate the interpolation accuracy. Near the defect boundary, the interpolation accuracy will decrease due to uneven distribution of sample points, and at this time, boundary constraints need to be increased.

[0082] In a possible implementation, the determination of the amplification coefficient of the deviation is based on the influence degree of the defect on the product performance. The amplification coefficient of a serious defect is set to a range of 1.5 to 2.0, indicating that the actual deviation can reach twice the measured value; the coefficient of a moderate defect is in a range of 1.2 to 1.5, reflecting moderate uncertainty; and the coefficient of a slight defect is in a range of 1.0 to 1.2, basically keeping the original measured value. The specific value is determined by a weighted average of the area proportion and the depth proportion of the defect, and the area proportion is weighted by 0.6 and the depth proportion is weighted by 0.4.

[0083] Exemplarily, the application of the Gaussian filter realizes a smooth transition of the thickness boundary. The size of the filter kernel is determined according to the spatial frequency of the thickness variation, and a two-dimensional Gaussian kernel of 5x5 or 7x7 is usually used. The standard deviation parameter controls the degree of smoothing, and a larger standard deviation produces a stronger smoothing effect but may lose details, and a smaller standard deviation retains more details but has a limited smoothing effect. At the junction of the protrusion and the normal region, the filtering process eliminates the step change of the thickness, forming a gradual transition. At the edge of the bubble, the filtering slows down the sharp drop in thickness, making the thickness distribution more consistent with the actual physical characteristics. Further, the data organization of the first thickness distribution map adopts a hierarchical structure. The basic layer stores the original measured thickness, keeping the integrity of the measurement data; the deviation layer records the thickness deviation value of each position, facilitating the tracing of the deviation source; and the fusion layer contains the final thickness distribution result, which is used for quality evaluation and visual display.

[0084] It can be understood that the accuracy verification of the thickness distribution map is realized by comparison with a standard sample. The standard sample with a known thickness distribution is selected, the same measurement and processing process is applied, and the difference between the predicted result and the actual thickness is compared. When the root mean square value of the prediction error is less than the allowable tolerance, it is considered that the thickness distribution map meets the accuracy requirement.

[0085] In step S108, if the thickness deviation in the first thickness distribution map exceeds a preset threshold, the prepreg lamination pressure parameter is dynamically adjusted according to the deviation distribution of the first thickness distribution map by analyzing the lamination pressure insufficient position corresponding to the maximum area of the thickness deviation, 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 position and the design target thickness is calculated from the first thickness distribution map. If the absolute value of the maximum deviation exceeds the preset threshold, the spatial coordinates of the deviation out-of-limit region are extracted, and the pressure control partition number to which the region belongs is determined through the pre-established correspondence between the coordinates and the pressure partition. Based on the pressure control partition number, the current pressure value applied in the partition is read, and the required pressure compensation is calculated according to the deviation value multiplied by the inverse of the material compression modulus, which is pre-determined through material testing. For the pressure compensation, the new pressure parameter is obtained by superimposing the compensation on the original pressure value. If the new pressure parameter exceeds the upper limit of the pressure allowed by the equipment, the upper limit value is taken. The pressure setting value of the corresponding partition 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, the interlayer adhesion degree, the interface defect distribution, and the 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 the preset improvement threshold, the second thickness distribution map is output as the corrected thickness distribution prediction result.

[0087] In one embodiment, the determination of the design target thickness is based on product specification requirements and material properties. For aerospace-grade composites, the target thickness is usually set as the number of fiber layers multiplied by the theoretical thickness of a single layer, taking into account the influence of resin content. The single-layer thickness of carbon fiber prepreg is generally 0.125 mm, and that of glass fiber prepreg is 0.2 mm. The contribution of resin content to thickness is calculated by volume fraction. When the resin volume fraction is 40%, the thickness increase coefficient is 1.1. The setting of the deviation threshold takes into account the use requirements of the product and the manufacturing tolerance. The thickness deviation threshold of an aerospace structure is usually ±5% of the target thickness, and that of an automotive component can be relaxed to ±8%.

[0088] It should be noted that the establishment of the correspondence between the coordinates and the pressure partitions involves the physical structure of the lamination equipment. The pressure application of the semi-cured sheet laminator is realized through multiple independently controlled hydraulic cylinders, each covering a specific working area. The pressure partitions are divided according to the distribution of the hydraulic cylinders, and a typical configuration is a 4x4 or 6x6 matrix arrangement. The boundary of each partition is determined by the perpendicular bisector of the center line of adjacent hydraulic cylinders. The coordinate mapping is realized through a lookup table, with the input being the (x, y) position in the workpiece coordinate system and the output being the corresponding partition number. The boundary region may be affected by multiple partitions, in which case a distance weighting method is used to determine the main control partition.

[0089] Specifically, the material compression modulus is measured following the standard test method. The specimen is prepared using the same material and process as the actual product, with a size of 100 x 100 x 10 mm. The compression test is performed on a universal material testing machine with a loading rate of 1 mm / min. The thickness change under different pressures is recorded, and the stress-strain curve is plotted. The compression modulus is determined by the slope of the linear segment of the curve, and the compression modulus of carbon fiber composite material is usually in the range of 8-12 GPa. The calculation of pressure compensation takes into account the nonlinear characteristics of the material, and when the strain exceeds 2%, a nonlinear correction coefficient needs to be introduced. The compensation calculation formula is Pressure compensation = thickness deviation x compression modulus x nonlinear correction coefficient, where the nonlinear correction coefficient is obtained by fitting the test data, with a value range of 0.8-1.2.

[0090] Preferably, the pressure adjustment of the hydraulic system is realized through closed-loop control of proportional relief valves and pressure sensors. Each pressure partition is equipped with an independent hydraulic circuit, including a hydraulic pump, a proportional relief valve, a pressure sensor, and an execution hydraulic cylinder. The proportional relief valve receives a control signal to adjust 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 a closed-loop control. When the pressure needs to be adjusted, the control system calculates a new control voltage value, which is output to the proportional relief valve through a digital-to-analog converter. The response time of pressure adjustment depends on the dynamic characteristics of the hydraulic system, and the typical value is 2-5 seconds to reach a stable state. The upper limit of the device pressure is determined by the rated pressure of the hydraulic pump and the set value of the safety valve, usually 10 MPa.

[0091] In one possible implementation, the implementation of the recalculation process adopts a parameter update mechanism. The recalculation of the resin flow state is based on the updated pressure field, and the increase in pressure leads to an increase in resin flow rate and an increase in flow distance. The degree of interlayer adhesion is evaluated by the change in contact area, and the increase in pressure causes more micro convexities to be flattened, increasing the contact area. The update of the interface defect distribution considers the compression effect of pressure on bubbles, and the volume of bubbles decreases or even disappears under high pressure. The thickness deviation prediction considers the changes of the above factors, and calculates the new thickness distribution by numerical integration.

[0092] Exemplarily, the evaluation of the improvement rate takes the relative error reduction as an index. 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 the process requirements, and is usually 30%. If the improvement rate does not reach the threshold, further iteration optimization may be needed. The number of iterations is limited to 3 to avoid infinite loop. The pressure adjustment amplitude of each iteration is gradually reduced, and a decay factor of 0.7 is used to ensure convergence. Further, the verification of the second thickness distribution map is completed by comparison with the measured data. In actual production, representative positions are selected for thickness measurement, and the number of measurement points is not less than 20, which are uniformly 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] It can be understood that the optimization of the pressure parameter not only improves the thickness uniformity, but also affects the microstructure of the material. Proper pressure can promote the penetration of resin between fibers, reduce porosity, and improve interlaminar shear strength. Excessive pressure may cause excessive resin loss, fiber volume fraction exceeding the standard, and thus reduce the mechanical properties.

[0094] With the above examples according to the embodiments of the present application as inspiration, through the above description, relevant personnel can certainly make various changes and modifications without deviating from the scope of the technical idea of the present application. The technical scope of the present application is not limited to the contents of the specification, and must be determined according to the scope of the claims.

Claims

1. A prepreg thickness uniformity analysis method based on deep learning, characterized by, The method comprises the following steps: Obtain fiber orientation stagger data in the prepreg multi-layer lamination process, collect fiber angle distribution images, and generate a first fiber distribution map; Extract fiber orientation stagger features from the first fiber distribution map, separate different angle regions, and obtain a fiber angle distribution; analyze resin flow unevenness according to the fiber angle distribution, and generate a resin flow distribution map; based on the resin flow distribution map, simulate the influence of pressure change on interlayer adhesion, and generate a first adhesion distribution map fused with resin flow information; extract micro convex region and bubble distribution information from the first adhesion distribution map, use a convolutional neural network to identify the interaction region of micro convex and bubble thickness fluctuation, and obtain a first interface feature set; based on the first interface feature set, analyze the coupling effect of fiber orientation stagger and interlayer adhesion change through multi-scale feature extraction, combine a deep learning network for defect classification, and output a defect position and severity distribution map; Based on the defect position and severity distribution map, combine thickness measurement data to predict the thickness distribution deviation caused by the interaction of micro convex and bubbles, and generate a first thickness distribution map; adjust the lamination pressure parameters according to the deviation distribution of the first thickness distribution map, and regenerate a second thickness distribution map.

2. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The method comprises the following steps: Determine the prepreg sample boundary coordinates through the positioning device of the optical scanning equipment, and identify the fiber orientation stagger region according to the fiber weaving density change; for the fiber orientation stagger region, use a line array camera to scan row by row with a coaxial light source, obtain a gray image sequence according to the fiber surface reflectivity difference; based on the gray image sequence, use Fourier transform to extract the frequency domain features of the fiber texture, calculate the fiber orientation angle according to the spectrum main peak position, and generate a first fiber distribution map.

3. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The method comprises the following steps: Read the fiber angle value of the pixel point from the first fiber distribution map, calculate the angle gradient through the angle difference value of the adjacent pixels, and mark the boundary points whose angle gradient exceeds the threshold value; based on the boundary points, use the watershed algorithm to segment the first fiber distribution map, calculate the standard deviation of the pixel angle value in the segmented region, retain the region whose standard deviation is less than the threshold value, obtain a set of angle consistency regions, and construct an interlaced feature matrix by statistically analyzing the average fiber angle of the set of angle consistency regions, and output the fiber angle distribution.

4. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The method comprises the following steps: According to the angle difference value of the adjacent regions in the fiber angle distribution, calculate the void volume change rate at the fiber stagger, determine the local permeability coefficient, obtain the resin flow resistance distribution; based on the flow resistance distribution, calculate the resin pressure field by using the finite difference method, and record the coordinates of the resin enrichment region whose pressure gradient exceeds the threshold value; according to the ratio of the cumulative volume of the resin enrichment region to the fiber gap capacity, calculate the resin climbing height, and generate a resin flow distribution map containing the micro convex position.

5. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The resin flow distribution map is used to simulate the influence of pressure change on the interlayer bonding degree, and a first bonding degree distribution map integrating resin flow information is generated, including: The convex position coordinates and height data are extracted from the resin flow distribution map, the convex compression amount is calculated according to the lamination pressure time sequence data, and the post-compression convex height distribution is obtained; based on the post-compression convex height distribution, the upper and lower layer surface distance field is calculated, the contact point area to total area ratio is counted, and the initial bonding degree value is obtained; according to the initial bonding degree value, the bonding degree value of the overlapping area is updated in combination with the resin transverse flow distance, and the first bonding degree distribution map is generated.

6. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The first bonding degree distribution map is used to extract micro-convex region and bubble distribution information, and a convolutional neural network is used to identify the interaction region of micro-convex and bubble thickness fluctuation to obtain a first interface feature set, including: The bonding degree gradient is calculated from the first bonding degree distribution map by using a gradient operator, the convex candidate points with gradient amplitude exceeding a threshold value and the bubble positions with bonding degree less than a threshold value are marked, the convex coordinate set and the bubble contour set are obtained; based on the convex coordinate set and the bubble contour set, a neural network containing multi-scale convolution kernels is constructed, the feature maps of different receptive fields are spliced, and a multi-scale feature map is generated; according to the multi-scale feature map, the spatial distance matrix of convex and bubble is calculated, the interaction region is determined, the thickness change rate is calculated, the interaction intensity index is constructed, and the first interface feature set containing convex position, bubble contour and interaction intensity value is output. 7.The deep learning-based prepreg thickness uniformity analysis method of claim 1, wherein, The first interface feature set is used to analyze the coupling effect of fiber orientation intersection and interlayer bonding degree change through multi-scale feature extraction, and a deep learning network is used for defect classification to output a defect position and severity distribution map, including: The interaction intensity value and spatial position information are extracted from the first interface feature set, the coupling coefficient of fiber orientation angle difference and bonding degree value is calculated, and a multi-scale feature vector is extracted through pyramid pooling; based on the multi-scale feature vector, a residual neural network is constructed to extract deep features, which are input into a classifier to determine the defect type and severity; based on the defect type and severity, the defect category and severity value are labeled by interpolation mapping to the original image size, adjacent defect regions of the same type are merged, and a defect distribution map is output.

8. The deep learning-based prepreg thickness uniformity analysis method according to claim 1, characterized by, The defect position and severity distribution map is used to predict the thickness distribution deviation caused by the interaction of micro-convex and bubbles in combination with the thickness measurement data to generate a first thickness distribution map, including: The defect coordinates are extracted from the defect position and severity distribution map, and the weighted thickness value is calculated in combination with the thickness measurement data; based on the weighted thickness value, the overlapping area of convex and bubble is identified, the local thickness change amount is calculated, and the thickness deviation field is generated by interpolation; according to the thickness deviation field, the original thickness data is superimposed, the thickness value of the convex and bubble position is adjusted, and the first thickness distribution map is generated. 9.The prepreg thickness uniformity analysis method based on deep learning according to claim 1, wherein, The lamination pressure parameters are adjusted according to the deviation distribution of the first thickness distribution map, and a second thickness distribution map is regenerated, including: A difference between a thickness value and a target thickness is calculated from the first thickness distribution map, a deviation overrun region coordinate is extracted, and a pressure control partition number is determined; based on the pressure control partition number, a pressure compensation amount is calculated, and a pressure parameter is updated; using the updated pressure parameter, a resin flow state, a fit degree, a defect distribution, and a thickness deviation are recalculated, and a second thickness distribution map is generated.

Images