A copper-clad plate impregnation process intelligent visual monitoring method

Abstract

The application provides an intelligent visual monitoring method for a copper-clad plate impregnation process, comprising: analyzing the uneven degree of glue liquid accumulation thickness and glue liquid distribution uniformity to obtain a thickness uneven characteristic value, and judging whether the thickness uneven characteristic value exceeds a preset threshold; analyzing the obstruction degree of the thickness uneven characteristic value to glue liquid diffusion to a center area of a plate to obtain a diffusion obstruction distribution map, combining segmentation boundary accuracy and thickness distribution data, and calculating a local glue liquid distribution imbalance characteristic value; according to the local glue liquid distribution imbalance characteristic value, combining a new segmentation boundary, analyzing glue liquid diffusion speed and glue liquid diffusion obstruction to the center area, obtaining a diffusion obstruction characteristic value, and judging whether the diffusion obstruction characteristic value exceeds a preset threshold; and performing feature enhancement processing on a real-time collected copper-clad plate edge image through a new detection parameter, analyzing glue liquid accumulation thickness and glue liquid diffusion speed, and determining a dynamic monitoring result of impregnation quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a kind of copper-clad plate impregnation process intelligent visual monitoring method. BACKGROUND

[0002] Copper-clad plate manufacturing is a crucial link in the electronics industry, and its quality directly affects the performance and reliability of the circuit board. As a core step in the production of copper-clad plate, the impregnation process determines the uniformity of the glue on the plate and the solidification effect, which plays a decisive role in the electrical performance and mechanical strength of the final product. However, the existing monitoring method for the impregnation process has significant shortcomings when faced with dynamic changes. Traditional visual monitoring systems often rely on fixed detection parameters, making it difficult to adapt to changes in impregnation speed or the geometry of the copper-clad plate edge, resulting in the inability to effectively identify abnormal phenomena such as glue accumulation in the edge area. This limitation is mainly reflected in the system's lack of adaptability to complex dynamic environments, especially when production conditions change, the monitoring accuracy decreases, and the stability of quality control is affected. In the impregnation process, the glue accumulation problem in the edge area of the copper-clad plate is particularly prominent. Changes in edge geometry can cause uneven flow resistance of the glue, leading to uneven distribution of the accumulation thickness. This unevenness further hinders the diffusion of the glue to the center area of the plate, resulting in local imbalance in glue distribution. For example, at high impregnation speeds, the edge area may experience excessive or insufficient glue accumulation due to changes in fluid mechanics effects, and existing monitoring systems are unable to adjust detection strategies in real time to accurately capture these dynamic changes. The unevenness of glue accumulation not only affects the quality of the edge area, but also can spread through diffusion barriers to the entire plate, leading to a decline in overall impregnation quality. Therefore, how to dynamically adjust the detection strategy of the visual monitoring system to adapt to the uneven glue accumulation and diffusion barrier caused by changes in impregnation speed and edge geometry has become a key issue in improving the monitoring effect of copper-clad plate impregnation quality. SUMMARY

[0003] The present application provides a kind of copper-clad plate impregnation process intelligent visual monitoring method, mainly includes:

[0004] Collecting real-time images of the edge area of the copper-clad plate, obtaining and preprocessing image data of glue distribution uniformity and glue accumulation thickness, extracting features of glue accumulation thickness and edge geometry, using image segmentation algorithm to separate the edge accumulation area and the center area, and obtaining thickness distribution data;

[0005] Analyzing the unevenness of glue accumulation thickness and glue distribution uniformity to obtain thickness unevenness characteristic value, and determining whether the thickness unevenness characteristic value exceeds the preset threshold;

[0006] If the thickness unevenness feature value exceeds the preset threshold, the segmentation threshold and region weight of the image segmentation algorithm are adjusted according to the impregnation speed control and the changing trend of edge geometry. The new segmentation boundary is determined according to the adjusted image segmentation algorithm. If the thickness unevenness feature value does not exceed the preset threshold, the current image segmentation algorithm parameters are kept unchanged.

[0007] The degree of resistance of thickness unevenness characteristics to the diffusion of adhesive to the central area of ​​the board is analyzed to obtain a diffusion resistance distribution map. Combined with the segmentation boundary accuracy and thickness distribution data, the local adhesive distribution imbalance characteristic value is calculated.

[0008] Based on the local adhesive distribution imbalance characteristic value, combined with the new segmentation boundary, analyze the adhesive diffusion rate and the degree of adhesive diffusion obstruction to the central region, obtain the diffusion obstruction characteristic value, and determine whether the diffusion obstruction characteristic value exceeds the preset threshold.

[0009] If the diffusion obstruction characteristic value exceeds the preset threshold, then based on the thickness unevenness characteristic value, the diffusion obstruction characteristic value, and the real-time monitored ambient temperature, the flow path and diffusion rate of the adhesive are predicted and optimized to generate new detection parameters.

[0010] By using new detection parameters to perform feature enhancement processing on the real-time acquired copper-clad laminate edge images, the adhesive accumulation thickness and adhesive diffusion rate are analyzed to determine the dynamic monitoring results of the adhesive impregnation quality.

[0011] Furthermore, the process involves acquiring real-time images of the copper-clad laminate edge region, obtaining and preprocessing image data on the uniformity of adhesive distribution and adhesive buildup thickness, extracting features of adhesive buildup thickness and edge geometry, and using image segmentation algorithms to separate the edge buildup region from the central region to obtain thickness distribution data, including:

[0012] Real-time image data of the copper-clad laminate edge region is acquired. A filtering algorithm is used to remove noise from the real-time image data. An edge enhancement algorithm is applied to calculate the edge gradient magnitude of the real-time image data, and preprocessed image data is generated based on the edge gradient magnitude. For the adhesive accumulation area in the preprocessed image data, pixels with changing grayscale values ​​are selected as seed points. The grayscale statistics of pixels surrounding the seed points are calculated, and the estimated adhesive thickness and uniformity marker data are determined based on the grayscale statistics. Based on the estimated adhesive thickness and the uniformity marker data, a region growing algorithm is executed to separate the edge accumulation area from the central area, and the boundary pixel coordinates are recorded to obtain a segmentation boundary coordinate sequence. Based on the segmentation boundary coordinate sequence, grayscale profile data of the boundary points is extracted and converted into adhesive accumulation thickness values ​​through the grayscale-thickness mapping relationship, generating continuous thickness distribution data.

[0013] Further, the analysis of the glue liquid accumulation thickness and the uniformity of the glue liquid distribution unevenness degree, obtain the thickness unevenness characteristic value, judge whether the thickness unevenness characteristic value exceeds the preset threshold value, including:

[0014] The thickness distribution matrix is constructed, the thickness difference value of adjacent positions in the thickness distribution matrix is calculated, and the average thickness change amount is generated; the thickness distribution matrix and the average thickness change amount are used as inputs, and the local thickness change feature is extracted by using the neural network, the thickness distribution matrix is processed through the convolution layer and the pooling layer, and the thickness unevenness characteristic value representing the glue liquid distribution unevenness degree is generated.

[0015] Further, if the thickness unevenness characteristic value exceeds the preset threshold value, the segmentation threshold value and the region weight of the image segmentation algorithm are adjusted according to the change trend of the edge geometry shape, the new segmentation boundary is determined according to the adjusted image segmentation algorithm, including:

[0016] The displacement amount of the glue liquid edge position is calculated, and the glue liquid flow speed value is generated; the speed change rate is calculated according to the glue liquid flow speed value, and the local curvature value of the edge geometry shape is extracted; the segmentation threshold value is adjusted according to the speed change rate, and the weight coefficient is determined according to the local curvature value; the segmentation threshold value and the weight coefficient are used to process the real-time image through the image segmentation algorithm, and the new segmentation boundary is generated.

[0017] Further, the analysis of the thickness unevenness characteristic value on the glue liquid diffusion to the center area of the plate is hindered, and the diffusion hindering distribution map is obtained, the local glue liquid distribution imbalance characteristic value is calculated by combining the segmentation boundary accuracy and the thickness distribution data, including:

[0018] The surface of the copper-clad plate is divided into a plurality of sub-regions, the average thickness value of the sub-regions is calculated, and the thickness gradient value is generated; the diffusion hindering coefficient is generated according to the thickness gradient value, and the diffusion hindering coefficient matrix is arranged, the diffusion hindering distribution map is generated by interpolation processing, the region with the boundary accuracy lower than the threshold value is marked, the thickness difference ratio of the sub-regions is calculated, and the local glue liquid distribution imbalance characteristic value is generated.

[0019] Further, according to the local glue liquid distribution imbalance characteristic value, the glue liquid diffusion speed and the glue liquid diffusion resistance to the center area are analyzed by combining the new segmentation boundary, and the diffusion resistance characteristic value is obtained, including:

[0020] According to the local glue solution distribution imbalance characteristic value, a seed point is selected in the new segmentation boundary, a region is expanded through a region growing algorithm, the number of newly added pixels is recorded, and a transient diffusion speed value is calculated; a smooth speed sequence is generated for the transient diffusion speed value, and a speed change rate is calculated; distance segments are divided according to the distance from the seed point to the center of the plate, the speed change rate of the distance segments is counted, an average speed attenuation index is generated, a resistance density coefficient and a resistance coverage coefficient are calculated, and a diffusion resistance characteristic value is generated.

[0021] Further, if the diffusion resistance characteristic value exceeds a preset threshold value, then according to the thickness unevenness characteristic value, the diffusion resistance characteristic value and the real-time monitored ambient temperature, the glue solution flow path and the diffusion speed are predicted and optimized, new detection parameters are generated, including:

[0022] According to the diffusion resistance characteristic value and the ambient temperature value, an input feature vector is generated, a machine learning algorithm is used to output a diffusion speed prediction value and a flow direction angle, a diffusion optimization parameter group is formed; the speed prediction value in the diffusion optimization parameter group is used to adjust the monitoring frequency value, the flow direction angle is used to determine a key monitoring area, the feature extraction density is adjusted, and new detection parameters are generated.

[0023] Further, the real-time collected copper-clad plate edge image is processed for feature enhancement through the new detection parameters, the glue solution accumulation thickness and the glue solution diffusion speed are analyzed, and a dynamic monitoring result of the impregnation quality is determined, including:

[0024] The real-time image is processed for sharpening through the new detection parameters, the gradient amplitude and the texture parameters of the glue solution edge are extracted, a feature vector is generated; the displacement amount of the feature vector is calculated, the glue solution diffusion speed and the thickness change amount are generated, the accumulation rate and the glue solution solidification time are calculated; the sharpening parameters and the feature extraction interval are adjusted according to the accumulation rate and the glue solution solidification time, a parameter adjustment rule set is generated; the image sequence is processed according to the parameter adjustment rule set, the uniformity index and the thickness ratio of the glue solution thickness are calculated, a quality evaluation value is generated, and the dynamic monitoring result of the impregnation quality is determined.

[0025] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0026] The application discloses a kind of copper-clad plate impregnation process intelligent visual monitoring method, solve the uneven thickness of glue liquid accumulation, edge geometry changes and diffusion speed obstruction caused quality unstable business scene problem.The present application is by high-resolution camera real-time acquisition copper-clad plate edge image, after denoising and edge enhancement pretreatment, extract glue liquid accumulation thickness and edge geometry feature, using image segmentation algorithm separates edge and center area, in combination with convolutional neural network analysis uneven thickness characteristic value, judge whether it exceeds threshold value based on historical data.If it is over the limit, the present application dynamically adjusts segmentation threshold and area weight, optimizes segmentation boundary;Meanwhile, through region growing algorithm analysis glue liquid diffusion obstruction situation, in combination with environmental temperature and other factors, use machine learning to predict glue liquid flow path, generate optimization parameter, adjust monitoring frequency and feature extraction efficiency.The present application finally realizes the accurate monitoring of glue liquid accumulation change trend and solidification time, dynamically updates image processing parameter, ensures that impregnation quality is stable, improves production efficiency and product consistency. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A flow chart of the copper-clad plate impregnation process intelligent visual monitoring method of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application.

[0029] As Figure 1 , the copper-clad plate impregnation process intelligent visual monitoring method of the present embodiment can specifically include:

[0030] S101, collect real-time image of copper-clad plate edge area, obtain and pretreat image data of glue liquid distribution uniformity and glue liquid accumulation thickness, extract glue liquid accumulation thickness and edge geometry feature, use image segmentation algorithm to separate edge accumulation area and center area, obtain thickness distribution data.

[0031] The real-time image data of the edge region of the copper-clad plate is collected by a high-resolution camera, the median filter algorithm is used to remove image noise, the Sobel operator is used for edge enhancement processing, the edge gradient amplitude of the processed image is calculated, and according to the comparison result of the gradient amplitude and the preset definition threshold, if the gradient amplitude is greater than the preset definition threshold, the next step is executed, otherwise the filter parameters are adjusted again for processing, and the preprocessed image data meeting the definition requirement is obtained. For the glue liquid accumulation region in the preprocessed image data, the pixel points with obvious gray value change are selected as the initial seed points, the gray mean and variance of the pixels around the seed points are calculated, the initial estimated value of the glue liquid thickness is constructed according to the gray mean, and the uniformity of the glue liquid distribution in the region is judged by the variance, if the variance is less than the preset uniformity threshold, it is marked as a uniformly distributed region, and the initial estimated value of the glue liquid thickness and the uniformity marking data of each region are obtained. Based on the initial estimated value of the glue liquid thickness and the uniformity marking data, the region growing algorithm is executed from the initial seed point, the region expansion is performed according to the gray difference between adjacent pixels and the seed region, when the gray difference is less than the preset growth threshold, the pixel is merged into the growing region, the edge accumulation region and the center region are separated through the iterative growing process, the boundary pixel coordinates in the growing process are recorded, and the segmentation boundary coordinate sequence of the edge accumulation region and the center region is obtained. According to the segmentation boundary coordinate sequence, the gray profile data of each boundary point is extracted along the normal direction of the boundary, the gray value is converted into the glue liquid accumulation thickness value through the pre-established gray thickness mapping relationship table, the mapping relationship table is obtained by measuring the gray of the standard thickness sample, and the converted thickness value is arranged according to the boundary position to form discrete thickness distribution data. The discrete data is processed by cubic spline interpolation, and the continuous glue liquid accumulation thickness distribution data and edge geometric shape parameters of the edge region of the copper-clad plate are obtained.

[0032] Specifically, in the copper-clad plate edge glue coating quality detection, the collection process of the high-resolution camera is crucial.

[0033] Specifically, the camera needs to be configured with a CMOS sensor of at least 2048*2048 pixels, and a telecentric lens is used to eliminate perspective distortion. During collection, the edge region is illuminated from multiple angles by a ring-shaped LED light source, so that the height difference caused by the glue liquid accumulation shows obvious gray change in the image. The median filter algorithm performs well in the noise removal process, which replaces the gray value of each pixel point with the median value of all pixel gray values in its neighborhood. This method is particularly effective in suppressing salt and pepper noise, while maintaining the definition of the edge profile. The Sobel operator, as a classic edge detection operator, enhances edge features by calculating the gradient of the image in the horizontal and vertical directions.

[0034] In one possible implementation, the Sobel operator uses a 3x3 convolution kernel to calculate the gradient components in the X and Y directions respectively, and then obtains the edge intensity information through the calculation of the gradient magnitude. The regions with larger gradient magnitude correspond to the positions of sharp changes in the thickness of the glue solution, which are the key areas for subsequent analysis.

[0035] It should be noted that the selection of the initial seed point directly affects the effect of region growing. By analyzing the gray histogram of the preprocessed image, the pixels in the gray value range with typical glue solution accumulation characteristics are selected as candidate seed points. These seed points are usually located in the region with higher gray value, because the thicker the glue solution accumulation, the greater the reflected light intensity, and the higher the corresponding gray value. The average gray value of the pixels around the seed point reflects the average thickness level of the region, while the variance represents the dispersion degree of the thickness distribution. The execution process of the region growing algorithm embodies the analysis idea from the local to the whole.

[0036] Preferably, the algorithm starts from the seed point and gradually includes the adjacent pixels that meet the growth criteria into the growing region. The setting of the growth criteria needs to consider the physical properties of the glue solution. Since the glue solution has fluidity before solidification, the thickness of the adjacent region usually changes gradually, so the setting of the gray difference threshold should reflect this gradual change characteristic. In this way, the edge accumulation region and the center flat region can be accurately separated, laying the foundation for subsequent thickness quantification analysis. The establishment of the gray-thickness mapping table is a key link to realize quantitative analysis.

[0037] In one embodiment, a laser thickness gauge is used to accurately measure standard glue solution samples with different thicknesses, while the corresponding image gray values are collected, establishing the correspondence between the gray values and the actual thickness. This mapping relationship usually presents a nonlinear characteristic, because there is a complex optical relationship between the reflected intensity of light and the thickness of the material. The application of cubic spline interpolation enables the discrete measurement points to form a continuous and smooth thickness distribution curve, which is of great significance for evaluating the uniformity of glue solution distribution and identifying abnormal accumulation regions.

[0038] S102, analyze the glue solution accumulation thickness and the unevenness of the glue solution distribution uniformity to obtain a thickness unevenness characteristic value, and determine whether the thickness unevenness characteristic value exceeds a preset threshold value.

[0039] A thickness distribution matrix is constructed, each element of the matrix representing the thickness value of the corresponding position of the glue liquid, the absolute value of the thickness difference between adjacent positions is calculated, the sum of all difference values is divided by the total number of sampling points to obtain the average thickness variation, the deviation square sum of the thickness value of each sampling point and the average thickness is calculated, and the standard deviation value is obtained by dividing the total number of sampling points and taking the square root. The thickness standard deviation value and the thickness distribution matrix are used as input data, and a convolutional neural network is used for feature extraction. The convolutional layer uses multiple size convolution kernels to scan the thickness distribution matrix to extract local thickness variation patterns. The pooling layer down-samples the convolution feature map to reduce the parameter amount. The fully connected layer maps the pooled feature vector to a single value. The thickness unevenness feature value representing the degree of unevenness of the glue liquid distribution is output. Based on the thickness unevenness feature value, the thickness unevenness feature value record of the qualified product batch is read from the historical production database, the statistical mean and standard deviation of the historical feature value are calculated, the judgment threshold is determined according to the mean plus the preset multiple standard deviation, and if the thickness unevenness feature value of the current batch exceeds the judgment threshold, it is determined that the degree of unevenness of the glue liquid distribution is out of standard.

[0040] Specifically, the construction of the thickness distribution matrix is the basis for realizing the quantitative evaluation of the unevenness of the glue liquid.

[0041] Specifically, the matrix organizes thickness data in a grid manner, and each matrix element corresponds to the thickness measurement value of the glue liquid at a specific position on the edge of the copper-clad plate. The number of rows and columns of the matrix is determined according to the detection accuracy requirement, and a specification of 64x64 or 128x128 is usually used, which ensures sufficient spatial resolution and avoids excessive computational burden. The calculation of the thickness difference between adjacent positions uses a four-neighbor or eight-neighbor method, which reflects the local thickness variation by calculating the thickness difference between the center point and the surrounding adjacent points. The two statistical indicators, average thickness variation and thickness standard deviation, represent the uniformity of the glue liquid distribution from different angles.

[0042] In one possible implementation, the average thickness variation is obtained by accumulating the absolute values of the thickness differences between all adjacent points and then dividing by the total number of differences. This indicator directly reflects the degree of thickness variation in space. The thickness standard deviation is evaluated from the perspective of overall distribution, which first calculates the average thickness of all sampling points, then calculates the deviation square sum of each point from the average value, and finally takes the square root and normalizes to obtain the standard deviation. These two indicators, in combination, can comprehensively reflect the unevenness characteristics of the glue liquid distribution. Convolutional neural networks exhibit unique advantages when processing thickness distribution matrices.

[0043] It should be noted that the convolutional layer uses multiple sizes of convolution kernels, such as 3x3, 5x5 and 7x7, to capture different spatial scale thickness variation patterns. Small size convolution kernels are good at detecting local sharp thickness changes, while large size convolution kernels can identify larger range of gradual trends. Each convolution kernel extracts specific thickness distribution pattern features by performing convolution operations with the thickness values of the corresponding region when scanning the thickness distribution matrix. The pooling layer uses maximum pooling or average pooling to reduce the spatial dimension of the convolution feature map by half, which not only reduces the subsequent calculation amount, but also retains the most significant feature information. The design of the fully connected layer reflects the conversion process from local features to overall evaluation.

[0044] Preferably, the multi-dimensional feature vector after pooling is first fused by a hidden layer containing 256 neurons using a ReLU activation function to enhance non-linear expression capability. The final output layer only contains one neuron, which maps the fused high-dimensional features to a single thickness unevenness feature value. The larger this value, the more uneven the glue distribution, providing a quantitative basis for subsequent quality judgment. Statistical analysis of historical production data provides a scientific basis for threshold setting. By collecting a large number of thickness unevenness feature values of qualified products, a statistically significant reference data set is formed. These historical data usually exhibit approximately normal distribution characteristics, i.e. the feature values of most qualified products are concentrated around the mean, and only a few are close to the quality boundary. The choice of preset multiple directly affects the strictness of quality control. The larger the multiple, the higher the allowed thickness unevenness, but the product quality risk also increases accordingly. Through this statistical-based threshold determination method, an objective judgment of glue distribution quality is achieved.

[0045] S103, if the thickness unevenness feature value exceeds the preset threshold, adjust the segmentation threshold and region weight of the image segmentation algorithm according to the change trend of the impregnation speed control and the edge geometry, determine the new segmentation boundary according to the adjusted image segmentation algorithm, if the thickness unevenness feature value does not exceed the preset threshold, keep the current image segmentation algorithm parameters unchanged.

[0046] If the uneven thickness feature value exceeds the preset threshold, a plurality of continuous frames of images are extracted from the real-time image sequence of the edge region of the copper-clad plate, the displacement of the glue liquid edge position between adjacent frames is calculated, the glue liquid flow speed value at each time is obtained by dividing the displacement by the time interval between frames, the least square method is used to linearly fit the speed value to obtain the speed change rate, the edge contour coordinate point sequence is extracted from the preprocessed image, and the local curvature value of each point is obtained by calculating the included angle formed by adjacent three points. According to the positive and negative of the speed change rate, the acceleration or deceleration state of the glue liquid flow is judged, and based on the positive and negative of the local curvature value, the convex or concave feature of the edge is identified. When the speed change rate is positive, the original segmentation threshold is multiplied by a preset increase coefficient, and when the speed change rate is negative, the original segmentation threshold is multiplied by a preset decrease coefficient. The weight coefficient of each position is determined according to the ratio of the absolute value of the local curvature value to the preset curvature reference value. The adjusted segmentation threshold and the weight coefficient of each position are used to process the current image by using the adaptive threshold segmentation method, the pixel attribution is determined according to the comparison result of the gray value of each pixel point and the corresponding position segmentation threshold, the accurate boundary position is obtained by weighted average calculation of the pixels near the boundary through the weight coefficient, and the updated segmentation boundary coordinate sequence is output. If the uneven thickness feature value does not exceed the preset threshold, the original segmentation parameter is maintained unchanged.

[0047] Specifically, real-time monitoring of the glue liquid flow speed is the key basis for dynamically adjusting the segmentation parameters.

[0048] In a possible implementation, the edge region images are continuously collected by a high-speed camera at a rate of 30 frames per second, and the time interval between each adjacent two frames is 33.3 milliseconds. In each frame of image, the glue liquid edge appears as a sudden boundary of gray value, and the accurate position coordinates of the glue liquid front are determined by edge detection. The instantaneous flow speed of the point can be obtained by dividing the coordinate difference of the corresponding positions in adjacent frames by the time interval. Since the glue liquid is subjected to the combined action of surface tension, viscosity and substrate adsorption force during impregnation, its flow speed presents a non-uniform speed characteristic. The least square method plays an important role in the calculation of the speed change rate.

[0049] Specifically, the speed values at a plurality of continuous time points are taken as the vertical coordinates, and the corresponding time points are taken as the horizontal coordinates to construct a speed-time scatter plot. The least square method finds the best fitting straight line by minimizing the sum of the squares of the perpendicular distances of all data points to the fitting straight line. The slope of the straight line is the speed change rate, and a positive value indicates that the glue liquid flow is accelerating, and a negative value indicates deceleration. This linear fitting method can effectively filter out measurement noise and obtain a stable speed change trend. The calculation of the edge curvature reveals the geometric characteristics of the glue liquid accumulation.

[0050] It should be noted that the local curvature is measured by the included angle formed by the adjacent three points. Selecting three consecutive points on the edge contour, calculating the perpendicular distance from the middle point to the line connecting the two end points, and then combining the arc length between the three points, the curvature value of the point can be obtained. Positive curvature indicates that the edge protrudes outward, which usually corresponds to the area of excessive glue accumulation; negative curvature indicates inward concave, which may be the area of insufficient glue. The greater the absolute value of the curvature, the more severe the change in the edge shape, and more accurate segmentation processing is required. The dynamic adjustment of the segmentation threshold embodies the adaptive processing idea. When the glue flow is detected to accelerate, it means that new glue is being quickly replenished, at this time, the segmentation threshold is appropriately increased to avoid misjudging the thin layer of glue in the flow as background. The preset increasing coefficient is usually set to be between 1.1 and 1.3, and the specific value is determined according to the viscosity characteristics of the glue. On the contrary, when the flow slows down, the glue tends to be stable accumulation, and reducing the segmentation threshold helps to accurately identify the glue boundary. The decreasing coefficient is generally selected in the range of 0.7 to 0.9. The determination of the weight coefficient realizes the differential treatment of different curvature regions.

[0051] Preferably, the preset curvature reference value is set as the curvature reference value of the flat edge, which is usually close to zero. The ratio of the actual curvature value to the reference value reflects the shape complexity of the position. The greater the ratio, the more special the edge shape of the region, and a higher weight should be given when determining the boundary. Through weighted average calculation, the shape complexity region contributes more to the final boundary position, thereby improving the segmentation accuracy of the special shape region. The application of the adaptive threshold segmentation method ensures the stability of the segmentation effect. Instead of using a single global threshold, this method dynamically determines the segmentation threshold according to the local features of each pixel position. Combined with the adjustment coefficient and weight coefficient calculated as described above, each pixel has its own exclusive segmentation criterion. This fine segmentation strategy is particularly suitable for processing edge regions with gradually changing glue thickness, and can accurately capture the subtle changes in glue distribution, providing reliable boundary data for subsequent quality evaluation.

[0052] S104, analyze the hindering degree of the thickness unevenness characteristic value to the diffusion of the glue to the center region of the plate, obtain a diffusion hindering distribution map, and combine the segmentation boundary accuracy and the thickness distribution data to calculate a local glue distribution imbalance characteristic value.

[0053] The watershed algorithm is used to divide the surface of the copper-clad plate into multiple sub-regions, the average thickness value of all pixel points in each sub-region is calculated as the representative thickness of the region, for two adjacent sub-regions, the absolute value of the thickness difference is calculated, the Euclidean distance between the geometric center points of the two sub-regions is measured, and the thickness gradient value between the two regions is obtained by dividing the thickness difference by the center distance. The thickness gradient values of all edge regions and their adjacent inner regions are compared, and the decreasing rate of the gradient value along the radial direction is calculated as the gradient change rate. According to the gradient change rate of each sub-region, an initial resistance coefficient is generated, and the gradient change rate multiplied by a preset conversion factor obtains the corresponding diffusion resistance coefficient. The diffusion resistance coefficients of each sub-region are arranged according to their spatial positions to form a diffusion resistance coefficient matrix, and a bilinear interpolation method is used to perform spatial interpolation processing on the discrete coefficient values in the matrix to generate a continuous diffusion resistance distribution map covering the entire detection region. Based on the diffusion resistance distribution map and the segmentation boundary precision data, regions with a boundary precision lower than a preset precision threshold are marked as regions to be analyzed, the average thickness value of each sub-region is extracted from the thickness distribution data, the thickness difference between adjacent regions is calculated and divided by the sum of the average thicknesses of the two regions to obtain the thickness difference ratio, and the diffusion resistance coefficients of each adjacent region pair are normalized as the weight values corresponding to the thickness difference ratio. The weighted thickness difference ratios of all adjacent regions are summed to obtain a local glue distribution imbalance feature value representing the degree of uneven glue distribution.

[0054] Specifically, the watershed algorithm has unique advantages in the division of the surface area of the copper-clad plate.

[0055] Specifically, the algorithm regards the gray value in the image as the terrain height, and the area with larger glue thickness corresponds to the mountain peak, and the area with smaller thickness corresponds to the valley. The algorithm simulates the process of water rising from each local minimum point, and when water flows from different sources meet, the meeting position forms a watershed, thereby realizing the natural segmentation of the region. This segmentation method based on the principle of geomorphology is particularly suitable for processing the gradual change of glue thickness and can accurately identify the natural boundary of thickness distribution. The calculation of the thickness gradient value reveals the spatial characteristics of glue diffusion.

[0056] In one possible implementation, for two adjacent sub-regions, the coordinates of the respective geometric center points are first calculated. The geometric center is obtained by the arithmetic average of all pixel coordinates in the region, and this method is also applicable to irregularly shaped regions. The calculation of the Euclidean distance uses the standard straight-line distance formula, which reflects the actual spatial interval between the centers of the two regions. Dividing the thickness difference by this distance, the gradient value obtained has a clear physical meaning, that is, the thickness change amount per unit distance. The larger the gradient value, the more intense the thickness change of the glue in that direction. The decreasing property of the gradient change rate reflects the dynamic process of glue diffusion.

[0057] It should be noted that from the edge to the center direction, the normal glue liquid diffusion should show a trend of gradually decreasing gradient. By calculating the difference between the gradient values of adjacent layers, the degree of this decrease can be quantified. When the gradient change rate of a certain area abnormally increases, it often means that there is a diffusion barrier at this place, which may be caused by uneven substrate surface properties or local abnormality of glue viscosity. This abnormality will hinder the normal flow of glue liquid to the center of the plate. The construction of the diffusion barrier coefficient matrix realizes the quantitative expression of the degree of hindrance. The preset conversion factor is usually determined according to the type of glue and the characteristics of the substrate, and is generally between 0.1 and 1.0. After the gradient change rate is multiplied by the conversion factor, the diffusion barrier coefficient obtained has the normalized characteristic, which is convenient for subsequent processing. The application of bilinear interpolation method ensures the smooth transition between discrete coefficient values. This method generates a continuous and smooth diffusion barrier distribution map by weighted average of the coefficient values of the four nearest neighbor points, and the weight is determined by the distance from the interpolation point to each neighbor point. The comprehensive analysis of boundary accuracy and thickness distribution embodies the multi-dimensional evaluation idea.

[0058] Preferably, when the positioning accuracy of the segmentation boundary is low, it indicates that the glue distribution in this area may have complex changes and needs to be paid attention to. Marking these low-precision areas and conducting in-depth analysis combined with their thickness distribution characteristics. The calculation of thickness difference ratio uses relative value instead of absolute value, eliminating the influence of different area reference thickness difference, making the evaluation result more comparable. Normalization processing and weighted summation realize the comprehensive quantification of unbalanced characteristics. The normalization of diffusion barrier coefficient ensures that all weight values are between 0 and 1, avoiding the dominance of a very large coefficient on the overall result. The greater the diffusion barrier, the more significant the contribution of the thickness difference to the overall imbalance.

[0059] S105, according to the local glue distribution imbalance characteristic value, combined with the new segmentation boundary, analyze the glue diffusion speed and the diffusion resistance of glue to the center area, get the diffusion resistance characteristic value, judge whether the diffusion resistance characteristic value exceeds the preset threshold.

[0060] According to the local glue liquid distribution imbalance characteristic value, the position with the highest imbalance characteristic value in the new segmentation boundary is selected as the starting seed point of the region growing algorithm, and the growth criterion is set as the difference between the gray value of the adjacent pixel and the average gray value of the current region being less than the preset growth threshold. Starting from the seed point, the region is gradually expanded according to the growth criterion, and the number of newly covered pixels is recorded every fixed time interval. The instantaneous diffusion speed value is obtained by dividing the number of newly added pixels by the time interval and the physical size of the pixel. For the continuously obtained instantaneous diffusion speed value sequence, the arithmetic mean of the speed values in each window is calculated using a sliding window, and the window size is a preset time window length, to obtain the smoothed diffusion speed time sequence. The speed change rate at each time is obtained by dividing the difference between the speed values at adjacent time by the time interval. According to the Euclidean distance from the seed point to the geometric center of the plate, the diffusion path is equally divided into multiple distance segments. The speed change rate at each time in each distance segment is counted, and the average value is calculated as the average speed attenuation index of the distance segment. When the average speed attenuation index of a distance segment exceeds the attenuation reference value determined according to historical data, the segment is marked as a diffusion blocked segment, and the number of blocked segments and the actual length value of each segment are recorded. The blocked density coefficient is calculated according to the proportion of the number of blocked segments to the total number of distance segments, and the blocked coverage coefficient is calculated according to the proportion of the total length of the blocked segments to the total length of the diffusion path. The blocked density coefficient is multiplied by the preset density weight, and the blocked coverage coefficient is multiplied by the preset coverage weight to obtain the comprehensive diffusion blocking characteristic value. Whether the diffusion blocking characteristic value exceeds the preset threshold is judged.

[0061] Specifically, the application of the region growing algorithm in the glue liquid diffusion analysis embodies the analysis idea from the local to the whole.

[0062] Specifically, selecting the position with the highest imbalance characteristic value as the seed point has a clear physical meaning. These positions are usually the areas with the most uneven glue liquid distribution, and are also the key positions that are most likely to affect the overall diffusion effect. The setting of the growth criterion is based on the optical properties of the glue liquid. When the gray value of the adjacent pixel is similar to the average value of the current region, it means that they belong to the same glue liquid coverage area. The preset growth threshold is usually determined according to the type of glue liquid and imaging conditions. If it is too small, it will cause the region to be fragmented, and if it is too large, it may incorrectly merge regions of different thicknesses. The calculation of the instantaneous diffusion speed needs to consider the spatial resolution of the image.

[0063] In one possible implementation, the physical size corresponding to each pixel is obtained through camera calibration, which is usually in the order of microns. The number of newly added pixels is multiplied by the physical area of a single pixel to obtain the actual diffusion area. Dividing the diffusion area by the time interval gives the diffusion speed with physical meaning, which is in square microns per second. This speed calculation based on the actual physical size makes the measurement results at different magnifications comparable. The sliding window average processing is a classic method in signal processing, which is used here to eliminate the influence of measurement noise.

[0064] It should be noted that the selection of window size needs to balance the smoothing effect and the time resolution. Larger window can filter out random fluctuations better, but will reduce the response ability to rapid changes in speed. The preset time window length usually contains 5 to 10 sampling points, so that the curve can be effectively smoothed, and the details of the speed change will not be excessively blurred. The segmented statistics of the speed change rate reveals the spatial characteristics of the diffusion process. On the path from the seed point to the center of the plate, the glue diffusion speed should remain relatively stable or slightly decrease under normal circumstances. When a certain distance segment shows a significant speed attenuation, it usually means that there is a diffusion barrier in that area. Such barriers may be caused by differences in the microstructure of the substrate surface, local temperature changes, or the influence of previous coating layers. By dividing the distance segment at equal intervals, the spatial location of the diffusion obstruction can be accurately located. The determination of the attenuation reference value depends on the statistical analysis of historical production data.

[0065] Preferably, the diffusion speed data of a large number of qualified products are collected, and the speed attenuation rate distribution of each distance segment is calculated. The upper quartile of the distribution is taken as the attenuation reference value, which can tolerate normal speed fluctuations while identifying abnormal diffusion obstructions. When the measured attenuation rate exceeds the reference value, it indicates that the diffusion resistance in that area is abnormally increased. The double quantification of the obstruction feature reflects the comprehensiveness of the evaluation. The obstruction density coefficient reflects the frequency of obstacles appearing on the diffusion path, and the obstruction coverage coefficient represents the spatial range of the obstacle influence. The two coefficients describe the severity of the diffusion obstruction from different dimensions. The preset density weight and coverage weight are determined according to the product quality requirements. For high-precision applications, the density coefficient is usually given a higher weight because frequent small-range obstructions can cause discontinuity in glue distribution. Through weighted integration, the diffusion obstruction feature value can comprehensively reflect the degree of obstruction of the glue diffusion to the center.

[0066] S106, if the diffusion obstruction feature value exceeds the preset threshold value, the glue flow path and diffusion speed are predicted and optimized according to the thickness unevenness feature value, the diffusion obstruction feature value and the real-time monitored environmental temperature, the diffusion optimization parameters are generated, the real-time monitoring frequency and feature extraction efficiency are adjusted, and the new detection parameters are generated. If the diffusion obstruction feature value does not exceed the preset threshold value, the existing detection parameter configuration is maintained.

[0067] If the diffusion hindrance characteristic value exceeds the preset threshold, the thickness unevenness characteristic value, the diffusion hindrance characteristic value and the ambient temperature value measured by the temperature sensor are obtained, the three values are normalized to form an input feature vector, a random forest algorithm is used to establish a mapping relationship between the feature vector and the glue liquid diffusion speed, the algorithm outputs a diffusion speed prediction value and an optimal flow direction angle through the voting mechanism of multiple decision trees, and a diffusion optimization parameter group containing the speed value and the direction angle is formed. According to the speed prediction value in the diffusion optimization parameter group, the ratio of the speed prediction value to the historical average diffusion speed is taken as a speed adjustment coefficient, the current image acquisition interval time is divided by the speed adjustment coefficient to obtain a new monitoring frequency value, the flow direction angle in the diffusion optimization parameter group is used to determine a key monitoring area, and the feature extraction density of the area is increased to the original density multiplied by a preset enhancement coefficient. The new monitoring frequency value is used to update the image acquisition control parameter, the spatial distribution weight of the feature extraction is adjusted according to the position of the key monitoring area, the feature extraction density of the non-key area is reduced to the original density multiplied by a preset attenuation coefficient, and the new detection parameter configuration is generated by combining the updated monitoring frequency value, the feature extraction density distribution and the spatial weight parameter. If the diffusion hindrance characteristic value does not exceed the preset threshold, the existing detection parameter configuration is maintained.

[0068] Specifically, the normalization processing of the feature vector is a key prerequisite for establishing an effective prediction model.

[0069] In a possible implementation, the thickness unevenness characteristic value usually varies between 0.1 and 10, the diffusion hindrance characteristic value ranges from 0 to 100, and the ambient temperature fluctuates between 15 and 35 degrees Celsius. The dimensions and numerical ranges of the three parameters differ greatly, and direct use will cause the temperature and other values to dominate the model training. The normalization processing maps each feature to a unified interval of 0 to 1, and the specific method is to subtract the historical minimum value of the feature from the current value, and then divide by the difference between the historical maximum value and the minimum value. After this processing, the three features have the same importance weight. The random forest algorithm exhibits unique advantages when processing glue liquid diffusion prediction.

[0070] Specifically, the algorithm constructs multiple decision trees, and each tree uses different subsets of training samples and features. When predicting the diffusion speed, each decision tree reaches a leaf node through a series of judgment nodes according to the input normalized feature vector, and outputs a speed prediction value. The prediction values of all decision trees are obtained through an average or voting mechanism to get the final result. This integrated method can effectively reduce the risk of overfitting and improve the stability and accuracy of prediction. The prediction of the flow direction angle reflects the directional characteristics of the glue liquid diffusion.

[0071] It should be noted that the diffusion of glue solution on the surface of copper-clad plate is not uniformly expanded to all directions, but is affected by surface tension, substrate texture and existing glue solution distribution, showing a specific dominant direction. Random forest can predict the most likely flow direction of glue solution under current conditions by learning the diffusion pattern in historical data. This direction is expressed in terms of angle relative to the center of the plate, ranging from 0 to 360 degrees. The calculation of the speed adjustment coefficient realizes dynamic monitoring frequency adjustment. The historical average diffusion speed is obtained by statistical data of past normal production batches, representing the diffusion rate under standard working conditions. When the predicted speed is higher than the historical average, it means that the glue solution diffuses faster, and the monitoring frequency needs to be increased to capture rapid changes; otherwise, the monitoring frequency can be reduced to save computing resources. The speed adjustment coefficient, as a ratio, directly reflects the deviation of the current diffusion speed from the normal level. The determination of the key monitoring area is based on the prediction results of the flow direction.

[0072] Preferably, a certain angular range is expanded to both sides with the predicted flow direction as the center, forming a fan-shaped key monitoring area. This area is the direction in which the glue solution is most likely to spread quickly, and requires more intensive feature extraction to accurately track the changes in the glue solution boundary. The preset enhancement coefficient is usually set to 1.5 to 2.0, meaning that the feature extraction density in the key area is increased to 1.5 to 2 times the original. The adjustment of the spatial distribution weight realizes the optimal allocation of computing resources. In non-key areas, the glue solution diffuses relatively slowly or has stabilized, and too high a feature extraction density will cause waste of computing resources. By introducing a preset attenuation coefficient, the feature extraction density in these areas is reduced to 0.5 to 0.7 times the original. This differentiated processing strategy ensures the monitoring accuracy of key areas and improves the overall system efficiency. The new detection parameter configuration integrates the monitoring frequency in the time dimension and the feature extraction distribution in the spatial dimension, realizing accurate dynamic monitoring of the glue solution diffusion process.

[0073] S107, perform feature enhancement processing on the real-time collected copper-clad plate edge images through the new detection parameters, analyze the glue solution accumulation thickness and glue solution diffusion speed, and determine the dynamic monitoring result of the glue dipping quality.

[0074] The Laplace sharpening processing is performed on the real-time collected copper-clad plate edge image by a new detection parameter to enhance the contrast of the edge profile, and the histogram equalization is used to improve the uniformity of the overall brightness distribution of the image to obtain a processed image with enhanced features. The gradient amplitude, gradient direction and gray level co-occurrence matrix texture parameters of the glue solution edge are extracted from the processed image to form a feature vector. The scale-invariant feature transform algorithm is used to match the current feature vector with the previous feature vector to calculate the displacement of the feature points to obtain the glue solution diffusion speed. The thickness change amount is calculated according to the gray value difference between adjacent time points, and the accumulation rate is obtained by dividing the thickness change amount by the time interval. The solidification time of the glue solution in the stable state is estimated according to the attenuation law of the accumulation rate. Based on the accumulation rate and the solidification time, when the accumulation rate is higher than the preset rate threshold, the convolution kernel size of the subsequent image sharpening is increased, and when the solidification time is less than the preset time threshold, the time interval for feature extraction is shortened. The adjusted convolution kernel size and time interval parameters are used as updated image processing parameters to form a parameter dynamic adjustment rule set. The image sequence collected according to the parameter dynamic adjustment rule set is processed, the standard deviation of the glue solution thickness of each sub-region is calculated as the uniformity index, the ratio of the maximum thickness to the average thickness of the edge is calculated, and the quality evaluation value is obtained by multiplying the reciprocal of the uniformity index and the thickness ratio. When the quality evaluation value exceeds the preset qualified threshold, it is determined to be qualified, and the dynamic monitoring result of the glue dipping quality including the quality evaluation value and the qualified determination is generated.

[0075] Specifically, the Laplace sharpening processing plays a key role in the glue solution edge enhancement.

[0076] Specifically, the algorithm detects the areas with sharp changes in gray level in the image through a second-order differential operator. In the application of copper-clad plate edge detection, the Laplace operator uses a 3x3 convolution kernel, with a center weight of -8 and a weight of 1 for the surrounding 8 positions. When the convolution kernel scans the image, a large response value will be generated at the glue solution edge because of the significant difference in gray level on both sides of the edge. This enhancement effect makes the subsequent feature extraction more accurate, especially for the weak edges of thin glue solution. As a classic method of texture analysis, the gray level co-occurrence matrix can quantify the texture features of the glue solution surface.

[0077] In one possible implementation, the co-occurrence matrix is constructed by counting the frequency of the gray level value combinations of pixel pairs at a specific distance in the image. Four key parameters, contrast, correlation, energy and homogeneity, can be extracted from the matrix. The contrast reflects the roughness of the glue solution surface, the correlation represents the directionality of the texture, the energy reflects the uniformity of the texture, and the homogeneity reflects the similarity of the local area. The feature vector formed by the combination of these parameters provides a multi-dimensional quantitative basis for glue quality evaluation. The scale-invariant feature transform algorithm has unique advantages in dynamic monitoring.

[0078] It should be noted that the SIFT algorithm first detects key points in different scale spaces, which correspond to the significant feature positions in the image. For each key point, the algorithm calculates the gradient direction histogram of the region around it to form a 128-dimensional feature descriptor. In the matching of images at adjacent time points, the corresponding points are found by comparing the Euclidean distance between the feature descriptors. The spatial displacement of the matching points reflects the flow of the glue solution, and the diffusion velocity is obtained by dividing the displacement by the time interval. The decay law of the accumulation rate reveals the physical process of the solidification of the glue solution. In the initial stage, the glue solution has good flowability and a high accumulation rate. As the solvent evaporates and the chemical reaction proceeds, the viscosity of the glue solution gradually increases, and the accumulation rate shows an exponential decay trend. By continuously monitoring the accumulation rate at multiple time points, the decay curve can be fitted. When the rate decreases to less than 10% of the initial value, the glue solution is considered to be solidified. This time point is the estimated solidification time, which is of great significance for the setting of subsequent process parameters. The establishment of the parameter dynamic adjustment rule set realizes adaptive monitoring.

[0079] Preferably, when a high accumulation rate is detected, it indicates that the glue solution flows violently, and more delicate edge detection is needed. At this time, the Laplacian convolution kernel size is expanded from 3x3 to 5x5, and the enhanced spatial resolution can capture more edge details. At the same time, the time interval for feature extraction is shortened from 100 milliseconds to 50 milliseconds, and the time resolution is improved to track rapid changes. This real-time state-based parameter adjustment ensures that the monitoring system always works in the best state. The comprehensive calculation of the quality evaluation value embodies the idea of multi-index fusion. The uniformity index is obtained by calculating the standard deviation of the thickness of each sub-region, and the smaller the standard deviation, the more uniform the distribution. Taking its reciprocal makes the product with higher uniformity score higher. The thickness ratio reflects the severity of the edge accumulation phenomenon, and in the ideal case, the ratio should be close to 1. The quality evaluation value obtained by multiplying the two indexes takes into account the overall uniformity and focuses on the key defect of edge accumulation. By comparing with the threshold set by the historical qualified product data, objective and quantitative quality judgment is realized.

[0080] The above is only an example and description of the structure of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present claims, and they should belong to the protection scope of the present application.

Claims

1. A method for intelligent visual monitoring of the copper clad laminate impregnation process, characterized in that, The method includes: Real-time images of the edge region of the copper clad laminate are acquired, and image data of adhesive distribution uniformity and adhesive accumulation thickness are obtained and preprocessed. Features of adhesive accumulation thickness and edge geometry are extracted, and image segmentation algorithm is used to separate the edge accumulation region from the center region to obtain thickness distribution data. Analyze the unevenness of adhesive accumulation thickness and adhesive distribution uniformity to obtain thickness unevenness characteristic values, and determine whether the thickness unevenness characteristic values ​​exceed preset thresholds; If the thickness unevenness feature value exceeds the preset threshold, the segmentation threshold and region weight of the image segmentation algorithm are adjusted according to the impregnation speed control and the changing trend of edge geometry, and a new segmentation boundary is determined according to the adjusted image segmentation algorithm. The degree of resistance of thickness unevenness characteristics to the diffusion of adhesive to the central area of ​​the board is analyzed to obtain a diffusion resistance distribution map. Combined with the segmentation boundary accuracy and thickness distribution data, the local adhesive distribution imbalance characteristic value is calculated. Based on the local adhesive distribution imbalance characteristic value, combined with the new segmentation boundary, analyze the adhesive diffusion rate and the degree of adhesive diffusion obstruction to the central region, obtain the diffusion obstruction characteristic value, and determine whether the diffusion obstruction characteristic value exceeds the preset threshold. If the diffusion obstruction characteristic value exceeds the preset threshold, then based on the thickness unevenness characteristic value, the diffusion obstruction characteristic value, and the real-time monitored ambient temperature, the flow path and diffusion rate of the adhesive are predicted and optimized to generate new detection parameters. By using new detection parameters to perform feature enhancement processing on the real-time acquired copper-clad laminate edge images, the adhesive accumulation thickness and adhesive diffusion rate are analyzed to determine the dynamic monitoring results of the adhesive impregnation quality.

2. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, The process involves acquiring real-time images of the copper-clad laminate edge region, obtaining and preprocessing image data on the uniformity of adhesive distribution and adhesive buildup thickness, extracting features of adhesive buildup thickness and edge geometry, and using image segmentation algorithms to separate the edge buildup region from the central region to obtain thickness distribution data, including: Real-time image data of the copper-clad laminate edge region is acquired. A filtering algorithm is used to remove noise from the real-time image data. An edge enhancement algorithm is applied to calculate the edge gradient magnitude of the real-time image data, and preprocessed image data is generated based on the edge gradient magnitude. For the adhesive accumulation area in the preprocessed image data, pixels with changing grayscale values ​​are selected as seed points. The grayscale statistics of pixels surrounding the seed points are calculated, and the estimated adhesive thickness and uniformity marker data are determined based on the grayscale statistics. Based on the estimated adhesive thickness and the uniformity marker data, a region growing algorithm is executed to separate the edge accumulation area from the central area, and the boundary pixel coordinates are recorded to obtain a segmentation boundary coordinate sequence. Based on the segmentation boundary coordinate sequence, grayscale profile data of the boundary points is extracted and converted into adhesive accumulation thickness values ​​through the grayscale-thickness mapping relationship, generating continuous thickness distribution data.

3. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, The analysis examines the unevenness of the adhesive buildup thickness and adhesive distribution to obtain thickness unevenness characteristic values. It then determines whether these thickness unevenness characteristic values ​​exceed a preset threshold, including: A thickness distribution matrix is ​​constructed, and the thickness difference between adjacent positions in the thickness distribution matrix is ​​calculated to generate an average thickness change. Using the thickness distribution matrix and the average thickness change as input, a neural network is used to extract local thickness change features. The thickness distribution matrix is ​​then processed through convolutional and pooling layers to generate thickness unevenness feature values ​​that characterize the degree of unevenness in adhesive distribution.

4. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, If the thickness unevenness feature value exceeds a preset threshold, the segmentation threshold and region weight of the image segmentation algorithm are adjusted according to the impregnation speed control and the changing trend of edge geometry. A new segmentation boundary is then determined based on the adjusted image segmentation algorithm, including: Calculate the displacement at the edge of the adhesive to generate an adhesive flow velocity value; calculate the velocity change rate based on the adhesive flow velocity value and extract the local curvature value of the edge geometry; adjust the segmentation threshold based on the velocity change rate and determine the weight coefficient based on the local curvature value; use the segmentation threshold and the weight coefficient to process the real-time image through an image segmentation algorithm to generate a new segmentation boundary.

5. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, The analysis of the thickness unevenness characteristic value hinders the diffusion of adhesive towards the center region of the board, resulting in a diffusion obstacle distribution map. Combining the segmentation boundary accuracy and thickness distribution data, the local adhesive distribution imbalance characteristic value is calculated, including: The surface of the copper-clad laminate is divided into multiple sub-regions. The average thickness value of the sub-regions is calculated to generate a thickness gradient value. A diffusion barrier coefficient is generated based on the thickness gradient value. The coefficients are arranged to form a diffusion barrier coefficient matrix. A diffusion barrier distribution map is generated through interpolation. Regions with boundary accuracy lower than a threshold are marked. The thickness difference ratio of the sub-regions is calculated to generate a local adhesive distribution imbalance feature value.

6. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, The process involves analyzing the adhesive diffusion rate and the degree of diffusion obstruction towards the central region based on the local adhesive distribution imbalance characteristic value and the new segmentation boundary, to obtain diffusion obstruction characteristic values, including: Based on the local adhesive distribution imbalance characteristic value, seed points are selected within the new segmentation boundary, the region is expanded using a region growth algorithm, the number of newly added pixels is recorded, and the instantaneous diffusion velocity value is calculated; a smooth velocity sequence is generated for the instantaneous diffusion velocity value, and the velocity change rate is calculated; distance segments are divided according to the distance from the seed point to the center of the board, the velocity change rate of the distance segments is statistically analyzed, an average velocity attenuation index is generated, the obstruction density coefficient and obstruction coverage coefficient are calculated, and diffusion obstruction characteristic values ​​are generated.

7. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, If the diffusion obstruction characteristic value exceeds a preset threshold, then based on the thickness unevenness characteristic value, the diffusion obstruction characteristic value, and the real-time monitored ambient temperature, the adhesive flow path and diffusion rate are predicted and optimized to generate new detection parameters, including: Based on the diffusion obstruction characteristic value and the ambient temperature value, an input feature vector is generated. A machine learning algorithm is used to output the diffusion velocity prediction value and the flow direction angle to form a diffusion optimization parameter set. The monitoring frequency value is adjusted according to the velocity prediction value in the diffusion optimization parameter set, the key monitoring area is determined according to the flow direction angle, the feature extraction density is adjusted, and new detection parameters are generated.

8. The intelligent visual monitoring method for the copper clad laminate impregnation process according to claim 1, characterized in that, The process involves performing feature enhancement processing on the real-time acquired copper-clad laminate edge images using new detection parameters, analyzing the adhesive buildup thickness and adhesive diffusion rate, and determining the dynamic monitoring results of the adhesive impregnation quality, including: The real-time image is sharpened using the new detection parameters, and the gradient magnitude and texture parameters of the adhesive edge are extracted to generate a feature vector. The displacement of the feature vector is calculated to generate the adhesive diffusion rate and thickness change, and the accumulation rate and adhesive curing time are calculated. The sharpening parameters and feature extraction interval are adjusted according to the accumulation rate and adhesive curing time to generate a parameter adjustment rule set. The image sequence is processed according to the parameter adjustment rule set to calculate the uniformity index and thickness ratio of the adhesive thickness, generate a quality assessment value, and determine the dynamic monitoring results of the adhesive impregnation quality.

Images