Process parameter adjustment method based on high multi-layer board compression bubble defect detection feedback

Abstract

The application provides a process parameter adjustment method based on high multilayer board pressing bubble defect detection feedback, comprising: adopting high-resolution X-ray phase contrast imaging combined with adaptive sampling density to perform bubble detection and three-dimensional reconstruction on the full-board multilayer structure of the PCB, extracting bubble clusters based on eight-neighborhood analysis, fusing the interlayer penetration depth of the bubble cluster and the glass cloth tightness matching degree by using a process knowledge graph, calculating dynamic process weights and feeding back the comprehensive strength according to the gridded unit to realize parameter differentiation adjustment of the pressing mold partition; through infrared hot spot monitoring and self-calibration mechanism, the closed-loop optimization of the weight model is realized, and the feedback and regulation are completed in the temperature rising window, the application can effectively improve the local defect recognition accuracy, the real-time performance and the pertinence of parameter adjustment, and promote the reduction of bubble defects and the improvement of process stability.

Description

Technical Field

[0001] This invention relates to the field of digital detection and dynamic parameter feedback control technology for lamination processes in printed circuit board manufacturing, and particularly to a method for adjusting process parameters based on feedback from the detection of bubble defects in high-multilayer board lamination. Background Technology

[0002] Currently, in the printed circuit board (PCB) manufacturing process, bubble defect detection and process parameter feedback control technology in the high-multilayer board lamination stage have become key aspects for improving product reliability and process yield. Mainstream industry technologies typically employ offline X-ray or ultrasonic imaging to screen for defects in the laminated product, combining historical lamination process parameters with empirical formulas to set parameter fine-tuning ranges, thus indirectly optimizing the process. Some high-end equipment has introduced real-time online monitoring of lamination pressure-temperature curves, coupled with simple feedback loops, to improve bubble elimination rates through synchronous correction of global parameters. Current development trends focus on process digitization, refined control, and adaptive defect identification, emphasizing the construction of a closed-loop automated system of detection-feedback-control. Among these, technologies such as automatic image segmentation, parameter correlation research based on defect voxel features, and multimodal data fusion analysis are gradually being validated and applied. However, the detection and feedback control technologies currently widely used in the lamination process of multilayer boards still have the following shortcomings and technical defects: (1) Traditional methods generally adopt a feedback mechanism based on the total volume or area of ​​the plate, which fails to establish differentiated discrimination criteria for the spatial distribution structure and aggregation characteristics of bubble defects. This results in no effective distinction between "dense cluster high-risk area" and "diffuse distribution low-risk area" in the feedback weight, which may lead to an imbalance in parameter resource allocation. (2) Existing mainstream feedback strategies mostly rely on fixed weights and static priorities, and cannot dynamically adjust parameter allocation according to the real-time severity of defects, distribution density, and changes in the local physical environment. This results in a lag in the system's response to newly emerging critical defects and insufficient efficiency in eliminating local residual bubbles. (3) There is no technology that couples and quantifies the topological features of defects (such as local entropy, spatial connectivity, non-uniformity, etc.) with key process parameters (such as penetration depth, glass cloth tightness) to guide more accurate parameter fine-tuning. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for adjusting process parameters based on feedback from the detection of bubble defects in the lamination of multilayer boards.

[0004] The technical solution of this invention is implemented as follows: a process parameter adjustment method based on feedback from the detection of bubble defects in multilayer board lamination, comprising: S1: Preset high-resolution X-ray phase-contrast imaging scanning path, covering the entire surface of the printed circuit board and performing intensified scanning on areas of abrupt changes in the stacked structure to obtain a three-dimensional voxel-level bubble density distribution map. S2: Perform eight-neighborhood connectivity analysis on the three-dimensional voxel-level bubble density distribution map to extract all independent bubble clusters, and calculate the local topological entropy of each bubble cluster based on the local density level distribution of voxels within the cluster. The local density level is divided into 5 levels according to the number of bubble voxels in the voxel neighborhood, and the local topological entropy reflects the degree of non-uniformity of density distribution within the bubble cluster. S3: For the bubble cluster, the interlayer penetration depth label and the glass cloth tightness matching degree are integrated with the process knowledge graph. The interlayer penetration depth label is determined based on the layer position of the bubble cluster, and the glass cloth tightness matching degree is obtained by querying the pre-stored glass cloth porosity-resin wetting rate mapping table. S4: Based on the local topological entropy, interlayer penetration depth label, and glass cloth tightness matching degree of each bubble cluster, calculate the dynamic weight value. This dynamic weight value is obtained by multiplying the local topological entropy by the first process sensitivity coefficient, (3 minus the interlayer penetration depth label) by the second process sensitivity coefficient, and (1 minus the glass cloth tightness matching degree) by the third process sensitivity coefficient, and then summing them up. The first process sensitivity coefficient is not less than 0.5 to ensure that the topological entropy dominates the weight allocation, and the first, second, and third process sensitivity coefficients are process sensitivity coefficients that can be calibrated online. S5: Divide the entire printed circuit board into N×M process control cells, calculate the weighted average of the dynamic weight values ​​of all bubble clusters in each process control cell, and use it as the comprehensive feedback intensity value of that cell. S6: Determine whether the comprehensive feedback intensity value of each process control cell is greater than the first threshold. If it is greater than the first threshold, trigger the micro-pressure increase and extended heat preservation time of the corresponding pressing mold zone of that cell; if the comprehensive feedback intensity value is greater than the second threshold and less than or equal to the first threshold, only the micro-pressure increase is triggered; otherwise, maintain the baseline pressing curve, where the second threshold is 70% of the first threshold. S7: Collect infrared hot spot images of the pressing process after local parameter adjustment, calculate the hot spot area shrinkage rate. If the hot spot area shrinkage rate is less than 15%, increase the first process sensitivity coefficient by 0.05 to achieve closed-loop self-calibration of the weight model. S8: Complete the weight update and parameter instruction generation of all process control cells within 300 seconds before the end of the pressing and heating stage to ensure that the feedback control meets the resin softening point window requirements.

[0005] The process parameter adjustment method based on feedback from the detection of bubble defects in multilayer board lamination provided by this invention has the following beneficial effects: (1) This invention introduces a topological entropy quantification mechanism based on three-dimensional voxel-level X-ray phase contrast imaging, which realizes a refined characterization of the spatial configuration complexity of bubble clusters during PCB lamination. Combined with eight-neighborhood connectivity analysis and local density level probability distribution calculation, the topological entropy can sensitively reflect the disorder of the internal structure of the bubble cluster, so that high-entropy defects that have interlayer expansion potential and are prone to short circuits or delamination failures can be accurately identified, significantly improving the relevance and reliability of defect assessment, and providing a scientific basis for subsequent differentiated process intervention. (2) This invention constructs a dynamic weighted feedback model that integrates process knowledge graphs, and couples the comprehensive strength function dominated by topological entropy with key process parameters such as interlayer penetration depth and glass cloth tightness matching degree to form an interpretable control logic oriented towards physical mechanisms. This avoids the generalization instability and untraceable decision-making problems that are common in black box data-driven models in industrial scenarios. By setting the first process sensitivity coefficient to be greater than or equal to 0.5, it ensures that the topological structure properties dominate the weight allocation, and supports online calibration of the second and third process sensitivity coefficients to adapt to different board systems and laminate designs. This gives the system good process adaptability and on-site adjustability. Based on the cell weighted average strength, a graded response strategy is implemented, triggering micro-pressure boosting and / or delayed heat preservation only in the Top-30% high-risk area. Under the premise of ensuring the overall pressing quality, this significantly reduces the energy consumption increase and cycle extension caused by excessive intervention, and achieves synergistic optimization between control accuracy and production efficiency.

[0006] (3) The present invention establishes a closed-loop self-calibration mechanism, using the shrinkage rate of infrared hot spot area after pressing as an external verification signal, dynamically adjusting the topological entropy sensitivity coefficient, so that the weight model can be corrected online with actual disturbances such as material batch fluctuations and equipment state drift, which significantly enhances the robustness and time response capability of the system; all calculations and instruction generation are limited to be completed within 300 seconds before the end of the heating stage, which fully matches the rheological response timing of the resin in the softening point window period, ensuring that the control measures take effect when the material is still plastic, and effectively improving the effectiveness of intervention. Attached Figure Description

[0007] Figure 1 This is a flowchart of the process parameter adjustment method based on the detection feedback of bubble defects in high multilayer board lamination according to the present invention; Figure 2 This is a sub-flowchart of the process parameter adjustment method based on feedback from the detection of bubble defects in multilayer board lamination according to the present invention. Figure 3 This is another sub-flowchart of the process parameter adjustment method based on feedback from the detection of bubble defects in multilayer board lamination according to the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides a method for adjusting process parameters based on feedback from the detection of bubble defects during lamination of multilayer boards, specifically including: S1: Preset high-resolution X-ray phase-contrast imaging scanning path, covering the entire surface of the printed circuit board and performing intensified scanning on areas of abrupt changes in the stacked structure to obtain a three-dimensional voxel-level bubble density distribution map. S2: Perform eight-neighborhood connectivity analysis on the three-dimensional voxel-level bubble density distribution map to extract all independent bubble clusters, and calculate the local topological entropy of each bubble cluster based on the local density level distribution of voxels within the cluster. The local density level is divided into 5 levels according to the number of bubble voxels in the voxel neighborhood, and the local topological entropy reflects the degree of non-uniformity of density distribution within the bubble cluster. S3: For the bubble cluster, the interlayer penetration depth label and the glass cloth tightness matching degree are integrated with the process knowledge graph. The interlayer penetration depth label is determined based on the layer position of the bubble cluster, and the glass cloth tightness matching degree is obtained by querying the pre-stored glass cloth porosity-resin wetting rate mapping table. S4: Based on the local topological entropy, interlayer penetration depth label, and glass cloth tightness matching degree of each bubble cluster, calculate the dynamic weight value. This dynamic weight value is obtained by multiplying the local topological entropy by the first process sensitivity coefficient, (3 minus the interlayer penetration depth label) by the second process sensitivity coefficient, and (1 minus the glass cloth tightness matching degree) by the third process sensitivity coefficient, and then summing them up. The first process sensitivity coefficient is not less than 0.5 to ensure that the topological entropy dominates the weight allocation, and the first, second, and third process sensitivity coefficients are process sensitivity coefficients that can be calibrated online. S5: Divide the entire printed circuit board into N×M process control cells, calculate the weighted average of the dynamic weight values ​​of all bubble clusters in each process control cell, and use it as the comprehensive feedback intensity value of that cell. S6: Determine whether the comprehensive feedback intensity value of each process control cell is greater than the first threshold. If it is greater than the first threshold, trigger the micro-pressure increase and extended heat preservation time of the corresponding pressing mold zone of that cell; if the comprehensive feedback intensity value is greater than the second threshold and less than or equal to the first threshold, only the micro-pressure increase is triggered; otherwise, maintain the baseline pressing curve, where the second threshold is 70% of the first threshold. S7: Collect infrared hot spot images of the pressing process after local parameter adjustment, calculate the hot spot area shrinkage rate. If the hot spot area shrinkage rate is less than 15%, increase the first process sensitivity coefficient by 0.05 to achieve closed-loop self-calibration of the weight model. S8: Complete the weight update and parameter instruction generation of all process control cells within 300 seconds before the end of the pressing and heating stage to ensure that the feedback control meets the resin softening point window requirements.

[0011] Step S1: A high-resolution X-ray phase-contrast imaging scanning path is preset, covering the entire surface of the printed circuit board and performing intensified scanning on areas of abrupt changes in the stacked structure to obtain a three-dimensional voxel-level bubble density distribution map. Specifically, this includes: S1.1: Based on the stacked structure parameter data in the process knowledge graph, a region identification algorithm is executed to define a set of stacked structure abrupt change regions. The input conditions are the inner layer copper thickness distribution data and the prepreg flow boundary parameters stored in the process knowledge graph. The interlayer jump detection technology based on threshold segmentation is used to extract the abrupt change points of the parameter data, and the output is a set of stacked structure abrupt change region locations, including the geographic coordinate information of the inner layer copper thickness jump region and the prepreg flow boundary zone. Based on the stack-up structure parameter data stored in the process knowledge graph, a parameter parsing algorithm (input conditions: inner layer copper thickness distribution data, prepreg flow boundary parameters) is used to deconstruct the physical layer properties of different regions of the multilayer board. Furthermore, a multi-dimensional parameter matching method (parameters: copper thickness gradient threshold, resin flow boundary displacement threshold) is used to generate regional feature vectors, which serve as the data basis for subsequent analysis of changes in the stacked structure. Furthermore, a threshold-based interlayer transition detection technique (parameters: copper thickness change rate threshold r1, adhesive flow boundary displacement rate threshold r2) is adopted to perform index screening of abnormal change points on the parsed parameter matrix and obtain a set of potential mutation points. Furthermore, through a spatial positioning algorithm (parameters: origin O of the three-dimensional coordinate system on the board surface, index table of stacked structure parameters), the set of abrupt change points is converted into a set of geographic coordinate information, and a coordinate matrix containing the inner copper thickness jump zone and the flow boundary zone of the semi-cured sheet is generated. Furthermore, a set merging method is employed (parameter: minimum region merging threshold). This allows for the regional merging of continuous or adjacent abrupt change points in the geographic coordinate matrix, resulting in a set of abrupt change regions with a layered structure. The set of abrupt change regions in the stacked structure output by the region recognition algorithm is used to transform the parameter change detection results of the previous step into a set of spatial locations, thereby providing the coordinate basis of the high-resolution scanning region for the subsequent generation of the encrypted scanning path point set. For example, on a high-multilayer PCB board with a thickness of 2.8 mm and a total of 20 layers, the inner layer copper thickness distribution data ranges from 0.5 oz to 2.0 oz per layer. Prepreg boundary parameters record the resin flow distance versus time curve. This embodiment uses a copper thickness change rate threshold. Set as Flow boundary displacement rate threshold Set as After threshold segmentation, the number of mutation points was calculated to be 42. Using the origin O of the three-dimensional coordinate system on the board surface as the location point in the lower left corner, the coordinate matrix output by the spatial positioning algorithm contains 42 three-dimensional coordinate points. These points are then processed using a minimum region merging threshold. Set as After mm², the regions are merged into 6 abrupt change regions in the stacked structure. These output coordinates are used for the increased sampling density in step S1.3, improving the scanning resolution of the abrupt change regions and thus providing more accurate positioning conditions for detecting high-risk bubble regions. S1.2: Apply a uniform grid partitioning algorithm to the full-plane design file data of the printed circuit board to generate a basic scan path point set; the input condition is the Gerber format geometric contour data of the board design file, and a two-dimensional grid generation operation is performed on the board surface using an equal-spacing sampling strategy. The output is a basic scan path point set, which contains an initial scan coordinate sequence uniformly distributed across the entire board surface. The input condition is a Gerber format geometric contour data file from the entire surface of the printed circuit board, which includes geometric layer information such as the board outline, through holes and solder mask openings, and is accurately described in a two-dimensional coordinate system. The contour analysis algorithm (parameter: supports Gerber RS-274X standard instruction set) is used to extract the geometric boundary coordinates and standardize them to a unified plate surface reference coordinate system to realize the structured expression of plate contour data; Furthermore, the integrity of the boundary data is verified by using a two-dimensional boundary closure detection algorithm (parameter: tolerance threshold 0.01mm), and the corrected two-dimensional plate surface contour matrix is ​​output as the input reference for mesh generation. Furthermore, a uniform grid partitioning algorithm is adopted (parameter: grid cell side length). , (Calculated from the ratio of target scanning resolution to board size), the standardized two-dimensional board contour matrix is ​​decomposed into regularity to generate a two-dimensional grid node matrix with equal spacing in the horizontal and vertical directions, and grid nodes that exceed the board contour are filtered out to realize the initial construction of the basic scanning path coordinate set; Furthermore, by using a coordinate sequence sorting algorithm (parameter: sorting criterion is row priority or column priority), the basic scanning path point set is arranged sequentially to ensure the trajectory continuity and data acquisition synchronization of subsequent scanning devices when executing path loading. By using uniform grid division and sequence sorting, the Gerber geometric data from the previous step is transformed into an initial scan coordinate sequence covering the entire plate surface, thereby achieving a spatially uniform distribution of the basic scan path and the executability of the scan plan. For example, on a PCB board measuring 250mm × 300mm, an input Gerber design file containing outline layer, drill layer, and solder mask data is used. The outline coordinates are extracted using a contour parsing algorithm within the range of... , The coordinates are standardized to the reference coordinate system (0,0) to (250,300) on the board surface. A uniform mesh generation algorithm is used, with the target scanning resolution set to 5mm, and the side length of the mesh cells is calculated. =5mm, the board surface was decomposed into a mesh of 50×60 nodes, generating a total of 3000 candidate scan coordinate points. Approximately 120 coordinate points located outside the polygonal boundaries were removed using a boundary filtering algorithm, resulting in a complete basic scan path coordinate set of 2880 points. A row-first sorting algorithm was used to arrange the coordinate sequence in ascending order of Y-coordinate, and within the same column, in ascending order of X-coordinate, forming a path sequence that the scanning equipment can directly load, covering the entire board surface. In actual scanning, this path significantly improved the coverage of bubble defect detection, kept the scanning time stably within the allowable cycle range of the equipment specifications, and seamlessly superimposed high-density sampling points in abrupt change areas during the subsequent S1.3 encrypted path generation process to achieve process optimization. S1.3: Based on the location set of the abrupt change region of the stacked structure and the basic scan path point set, perform adaptive sampling density adjustment processing to generate an encrypted scan path point set; wherein, the input conditions are the geographic coordinate information of the location set of the abrupt change region of the stacked structure and the initial coordinate sequence of the basic scan path point set, and use the local density enhancement algorithm to perform sampling point density doubling operation on the abrupt change region, and output the encrypted scan path point set, which includes the fused coordinate sequence of the basic path of the whole board and the high-resolution encrypted path of the abrupt change region; Based on the input conditions of the location set of abrupt change regions in the stacked structure and the basic scan path point set, a local density enhancement algorithm (parameters: density multiplication coefficient K=3, boundary expansion radius R=2.5mm) is adopted to achieve density multiplication processing of sampling points in abrupt change regions, so as to ensure that subsequent imaging data obtains higher spatial resolution in defect-prone areas. Furthermore, by using a region coverage detection algorithm (parameters: abrupt region coordinate set, basic scan path coordinate sequence), the spatial overlap between the abrupt region and the basic path is determined, and a path index list within the abrupt region is generated to form a high-density sampling candidate path set; Furthermore, an interpolation resampling algorithm is employed (parameters: bilinear interpolation, sampling interval). =0.25mm), to achieve local sampling point interpolation refinement processing of the above candidate path set, and generate a density-multiplying path segment coordinate sequence; Furthermore, the basic scan path point set and the encrypted path segment coordinate sequence are merged into a unified coordinate sequence by using a path fusion sorting algorithm (parameters: fusion priority: basic path priority = 1, encrypted path priority = 2; sorting rule: increasing by X coordinate), and the continuity and non-repetition of the coordinate sequence in spatial distribution are ensured. Furthermore, a path sequence validity verification algorithm (parameters: minimum spacing threshold between paths = 0.2 mm, maximum spacing threshold = 1.5 mm) is used to perform spacing rule verification on the fused coordinate sequence, remove points that do not meet the constraints, and output the final encrypted scanning path point set for the high-resolution X-ray phase-contrast imaging system to perform point-by-point scanning operations. By combining the local density enhancement algorithm and the path fusion sorting algorithm, the abrupt region identification results of the previous step are transformed into a scan coordinate sequence that integrates the basic path and the high-resolution encrypted path, achieving a balance between improving the acquisition accuracy of abrupt regions and covering the entire board space. For example, on a high-multilayer PCB board with dimensions of 610mm × 510mm, the location set of abrupt change regions in the stack-up structure includes 8 copper thickness transition regions and 5 prepreg flow boundary bands, with a geographic coordinate accuracy of 0.1mm. The basic scan path point set is generated by a uniform grid partitioning algorithm with a sampling interval of 1mm and a total of 310,000 points. In the local density enhancement algorithm, the sampling point density multiplication factor K in the abrupt change region is set to 3, and the boundary expansion radius R is set to 2.5mm, resulting in a candidate path set of approximately 45,000 points. An interpolation resampling algorithm is used to refine the candidate path set, with a sampling interval of... Setting the margin to 0.25mm, approximately 135,000 points of the encrypted path segment coordinate sequence were generated. The basic path and the encrypted path were then merged using a path fusion and sorting algorithm, with the sorting rule increasing according to the X-coordinate. The total number of points in the merged coordinate sequence was approximately 400,000. Finally, after verification by a path sequence validity algorithm, approximately 5,000 invalid points with spacing less than 0.2mm or greater than 1.5mm were removed, resulting in a final encrypted scan path point set of 395,000 coordinate points. In the subsequent point-by-point imaging in step S1.4, the imaging clarity in abrupt change areas was significantly improved compared to the basic path scheme, and the resulting 3D voxel-level bubble density distribution map showed a substantial improvement in detail representation in high-risk areas.

[0012] S1.4: Using a high-resolution X-ray phase-contrast imaging system, perform point-by-point scanning operations according to the encrypted scanning path point set to acquire a transmission image data sequence; wherein, the input condition is the fused coordinate sequence of the encrypted scanning path point set, and X-ray projection data acquisition and processing are performed on the scanning path through phase-contrast imaging technology, and the output transmission image data sequence includes a two-dimensional projection image set containing multi-angle bubble scattering characteristics. S1.5: Perform 3D reconstruction and voxelization on the transmission image data sequence to generate a 3D voxel-level bubble density distribution map; wherein, the input condition is a 2D projection image set of the transmission image data sequence, after performing image reconstruction operation using a filtered back projection algorithm, voxel space partitioning and bubble density calculation are performed, and the output is a 3D voxel-level bubble density distribution map, which includes the quantized bubble density value of each voxel in the voxel grid of the entire printed circuit board. For a two-dimensional projection image set of a transmission image data sequence, a filtered back projection algorithm (parameters: back projection angle step size 1°, filter type Ram-Lak) is used to realize the inversion and reconstruction function of multi-view projection data. Furthermore, by using a voxel space partitioning method (parameter: voxel side length 0.1mm), the three-dimensional reconstruction results are discretized in the Cartesian coordinate system, and three-dimensional voxel mesh data with fixed spatial resolution is obtained. Furthermore, a bubble density calculation algorithm (parameters: neighborhood radius of 2 voxels, statistical bubble voxel percentage) is adopted to analyze the bubble content of each voxel and generate a quantified voxel-level bubble density value. Furthermore, by performing density normalization (parameter: the maximum density reference value is taken as the 95th percentile of the total number of bubble elements in the plate), the comparability of bubble densities in different regions on a unified metric scale is achieved, and a normalized density distribution map is obtained. By using voxelization and density calculation processing, the two-dimensional projection image set from the previous step is transformed into a three-dimensional voxel-level bubble density distribution map containing the voxel-quantized bubble density value of each voxel in the voxel grid of the entire printed circuit board, thereby achieving a global and accurate characterization of the spatial distribution of bubble defects. For example, during the lamination inspection of a 400mm × 300mm multilayer PCB, the transmission image data sequence was acquired by an X-ray phase-contrast imaging system, with an acquisition angle covering 0° to 179° and a step size of 1°, forming 180 two-dimensional projection images. When using a filtered back-projection algorithm, a Ram-Lak filter was selected. Each projection was frequency-domain filtered before back-projection to obtain the primary 3D reconstruction data volume. The voxel space was partitioned using a cubic grid with a side length of 0.1mm, forming a voxel grid of 4000 × 3000 × the number of layers corresponding to the thickness direction. For each voxel, the proportion of bubble voxels within a neighborhood of 2 voxels was statistically analyzed, with bubble voxels defined as those with a gray level lower than the board mean minus two standard deviations. After calculating the density values, normalization was performed to obtain a normalized density distribution. The three-dimensional voxel-level bubble density distribution map output by this process was used in the subsequent S2 step for eight-neighbor connectivity analysis to identify high-density clustering regions. This effectively supports the spatial topological entropy analysis and dynamic weight feedback control of bubble defects in multilayer boards, enabling targeted adjustment of pressing parameters in high-entropy risk areas.

[0013] Step S2: Perform eight-neighborhood connectivity analysis on the three-dimensional voxel-level bubble density distribution map to extract all independent bubble clusters. Calculate the local topological entropy of each bubble cluster based on the local density level distribution of voxels within the cluster. The local density level is divided into 5 levels according to the number of bubble voxels in the voxel's neighborhood. The local topological entropy reflects the degree of non-uniformity of the density distribution within the bubble cluster. Specifically, this includes: S2.1: Based on the number of bubble voxels in the neighborhood of a voxel, a 5-level local density classification model is constructed. The range of bubble voxel counts in the neighborhood is divided into 5 levels from low density to high density by a preset threshold, and represented by standardized voxel density features. S2.2: Apply the eight-neighborhood connectivity analysis algorithm to the three-dimensional voxel-level bubble density distribution map to identify all connected bubble regions and extract them as independent bubble clusters in order to separate different defect aggregates; S2.3: For each extracted independent bubble cluster, the density level distribution of each voxel within the cluster is statistically analyzed based on a 5-level local density level division model, and the probability of the proportion of voxels in the cluster at each density level is calculated to generate a density level distribution probability vector. For the set of independent bubble clusters output by step S2.2, the input conditions include a three-dimensional voxel-level bubble density distribution map and a 5-level local density classification model constructed in step S2.1. A voxel density level statistical method (parameters: number of bubble voxels in the local neighborhood, density level threshold) is used to classify the density level of each voxel within each bubble cluster and form a density level label matrix of voxels within the cluster. Furthermore, by using a group counting algorithm (parameters: density level label matrix, level value range 1~5), the voxel count statistics corresponding to each density level within the cluster are realized, and the counting vector is obtained; Furthermore, a normalized probability calculation method (parameters: counting vector, total number of primes in the cluster) is adopted to calculate the probability of the proportion of voxels at each density level and generate a density level distribution probability vector. Furthermore, by employing a probability vector rearrangement algorithm (parameters: rank order, cumulative probability verification), the consistency of the order of the probability vectors of the density rank distribution and the sum of the total probabilities are achieved. The verification process ensures that the data conforms to the input specifications for subsequent entropy calculations. By performing probability normalization and rearrangement, the result of the previous step is transformed into a structured density level distribution probability vector, thereby achieving a quantitative representation of the density distribution inside the defect space. For example, in the defect analysis of a high-multilayer PCB lamination process, an independent bubble cluster A is selected, with a total prime number of 500. The voxel counts for low-density level L1 are 50, level L2 is 100, level L3 is 150, level L4 is 120, and level L5 is 80. A grouped counting algorithm is used to obtain the counting vector [50, 100, 150, 120, 80]. Based on the normalized probability calculation method, the probability values ​​for each level are calculated respectively: , , , , The output probability vector is [0.1, 0.2, 0.3, 0.24, 0.16]. The sum of probabilities is verified using a rearrangement algorithm. The density level distribution probability vector is obtained by maintaining a consistent order of levels and finally satisfying the input specifications. In the subsequent entropy calculation in S2.4, this probability vector will serve as the core input, which can significantly improve the sensitivity of topological entropy assessment to defect density inhomogeneity. S2.4: Based on the density level distribution probability vector of each bubble cluster, calculate the local topological entropy to quantify the non-uniformity of the density distribution. This calculation outputs an entropy value based on the information entropy theory, reflecting the degree of randomness of the density level within the cluster. S2.5: Output the local topological entropy value of each bubble cluster, which serves as a key parameter for the topological features of the defect space and is used as input for the subsequent dynamic weight calculation module.

[0014] like Figure 2 As shown, step S3 involves: for the bubble cluster, integrating the interlayer penetration depth label from the process knowledge graph with the glass cloth tightness matching degree. The interlayer penetration depth label is determined based on the layer location of the bubble cluster, and the glass cloth tightness matching degree is obtained by querying a pre-stored glass cloth porosity-resin wetting rate mapping table. Specifically, this includes: S3.1: Analyze the structural parameters of the lamination process to obtain the total number of layers and the stacking order information; perform region boundary determination processing based on the analysis results to divide the stacked structure into surface region, intermediate layer region and core layer region; generate an interlayer penetration depth level mapping table based on the region division results, where the surface region corresponds to label 0, the intermediate layer region corresponds to label 1 and the core layer region corresponds to label 2. S3.2 Based on the resin impregnation experimental data, a porosity-resin impregnation rate mapping table is constructed. The interlayer penetration depth level mapping table and the porosity-resin impregnation rate mapping table are integrated into a process knowledge graph and stored as a queryable database structure to provide a basis for labeling bubble cluster process parameters. S3.3: Perform region attribution determination on the layer location information of the bubble cluster. Based on the layer region boundaries in the process knowledge graph, determine the layer region where the bubble cluster is located. Query the process knowledge graph according to the layer region attribution result to obtain interlayer penetration depth level labels, where interlayer penetration depth level label 0 represents the surface region, label 1 represents the middle layer region, and label 2 represents the core layer region. Perform validity verification on the interlayer penetration depth level labels to ensure that the label values ​​are consistent with the physical location of the stacked structure. Output the verified interlayer penetration depth labels to reflect the depth location characteristics of the bubble cluster in the stacked structure. S3.4: Extract parameters from the glass cloth specifications in the area where the bubble clusters are located to obtain the glass cloth porosity and density parameters. Query the resin wetting rate based on the porosity-resin wetting rate mapping table in the process knowledge graph. Calculate the glass cloth density matching degree based on the resin wetting rate and the preset standard rate. Perform normalization on the glass cloth density matching degree to ensure that the value range is between 0 and 1, representing the degree of matching. Output the standardized glass cloth density matching degree to quantify the compatibility level between resin flow and glass cloth characteristics.

[0015] like Figure 3 As shown, step S4 involves calculating a dynamic weight value based on the local topological entropy, interlayer penetration depth label, and glass cloth tightness matching degree of each bubble cluster. This dynamic weight value is obtained by multiplying the local topological entropy by a first process sensitivity coefficient, (3 minus the interlayer penetration depth label) by a second process sensitivity coefficient, and (1 minus the glass cloth tightness matching degree) by a third process sensitivity coefficient, and then summing the results. The first process sensitivity coefficient is not less than 0.5 to ensure that the topological entropy dominates the weight allocation, and the first, second, and third process sensitivity coefficients are process sensitivity coefficients that can be calibrated online. Specifically, this includes: S4.1: Obtain the local topological entropy, interlayer penetration depth label and glass cloth tightness matching degree for each bubble cluster, and verify whether the first process sensitivity coefficient is not less than 0.5 to ensure that the topological entropy dominates the weight allocation, and output the verified process parameter set; The input conditions include the local topological entropy value of each bubble cluster output from step S3, the interlayer penetration depth label value, and the normalized value of the glass cloth tightness matching degree, as well as the online parameter cache of the current process sensitivity coefficient set; A parameter verification algorithm (constraint: first process sensitivity coefficient ≥ 0.5) is used to check the threshold validity of the first process sensitivity coefficient. This algorithm reads the first process sensitivity coefficient. Value, and A comparison is performed, and if the comparison result is less than the threshold, the online calibration module is invoked to correct the parameters. The value is raised to the threshold value or higher. Furthermore, by using a data consistency check method (rule set: all three parameters must be within the legal range), the local topological entropy is checked. Interlayer penetration depth label Matching degree of glass cloth tightness Consistency verification is performed, and a list of data records that failed verification is generated as an anomaly input filtering condition; Furthermore, through parameter set integration processing (input: This enables the structured combination of three process characteristic parameters and their corresponding sensitivity coefficients to form a process parameter set, ensuring that subsequent weighted component calculations can directly receive this structured data. Furthermore, through the index mapping program (index key: bubble cluster ID), a two-way binding between the parameter set and the unique identifier of the bubble cluster is achieved, resulting in a bound parameter data table, which is used to support the operation calls from S4.2 to S4.4; By using parameter verification and structured combination processing, the annotation results of the previous step S3 are transformed into a set of process parameters that have been validated, thus ensuring the prerequisite for topological entropy to dominate weight allocation in dynamic weight calculation. For example, in a certain high-multilayer PCB lamination batch, the input local topology entropy for Interlayer penetration depth label for Glass cloth tightness matching degree for The sensitivity coefficient α of the first process was obtained as follows: Second process sensitivity coefficient for The third process sensitivity coefficient for The parameter verification algorithm is executed as follows: Read value Comparison rules ≥ Not satisfied, will Corrected to Confirmation during consistency testing Within the range of defect density entropy values Inside, legitimate, Valid; structured combination generates parameter sets It also binds the bubble cluster ID#C102. The execution result is that the weight value can be directly calculated in S4.2. Performance verification shows that the proportion of the topological entropy component in the weight allocation is significantly improved, supporting the generation of subsequent pressure and insulation adjustment commands; S4.2: Based on the verified local topological entropy and the first process sensitivity coefficient, perform a multiplication operation to generate the first weighted component; The input conditions are the set of local topological entropy values ​​verified by step S4.1 and the set of first process sensitivity coefficient parameters, which are used as the direct calculation objects of this sub-step. A floating-point precision preservation algorithm (parameters: double precision 64-bit, rounding mode is rounding) is adopted to achieve accurate numerical loading of local topological entropy and the first process sensitivity coefficient, ensuring the stability and repeatability of subsequent multiplication operations; Furthermore, through the arithmetic multiplication operation module (parameters: multiplier type is pipeline floating-point multiplication, throughput is one result per cycle), the point-by-point product calculation of the local topological entropy corresponding to each bubble cluster and its first process sensitivity coefficient is realized, and the initial product dataset is obtained; Furthermore, a numerical overflow detection algorithm (parameter: detection threshold is used) is employed. This allows for the validity review of the initial product dataset, excluding product results with numerical anomalies, and generating a product result set that has passed the validity screening. Furthermore, a result normalization method (parameters: normalization interval is [0,1], method is minimum-maximum linear normalization) is adopted to achieve interval standardization of the product result set after validity screening, so that the weighted components between different bubble clusters have consistent dimensions; Through the above multiplication and standardization operations, the local topological entropy value and the first process sensitivity coefficient from the previous step are transformed into a dimensionless first weighted component, thereby realizing the quantification and weighting of defect space information and laying the computational foundation for the differentiated fusion of subsequent depth compensation value and matching compensation value.

[0016]

[0017] in, As the first weighted component, This is the local topological entropy value. The first process sensitivity coefficient; For example, in a certain high-multilayer PCB lamination process, the local topological entropy value of a certain bubble cluster is... The first process sensitivity coefficient is set to The initial product is obtained after performing double-precision multiplication. After overflow detection, the value is within the valid range and enters the normalization stage, based on the minimum value of the global product result set. With the maximum value According to the minimum-maximum normalization formula Calculations were performed to obtain the normalized result. This component serves as the first weighted component for the bubble cluster. When subsequently fused with the depth compensation and matching compensation values, this component significantly enhances the contribution to the dynamic weight value, ensuring that the priority of the topological entropy-dominated defect severity representation in parameter adjustment decisions is not weakened. S4.3: Based on the verified interlayer penetration depth label, calculate the difference between 3 and the interlayer penetration depth label as the depth compensation value, and then multiply the depth compensation value by the second process sensitivity coefficient to generate the second weighted component; Based on the first weighted component generated in the preceding sub-step S4.2 and the process parameter set of the currently processed bubble cluster, the depth compensation value is calculated using the difference operation method (parameters: constant 3, interlayer penetration depth label); Furthermore, through the arithmetic subtraction processing module (parameter source: interlayer penetration depth label in the process parameter set), the operation between constant 3 and interlayer penetration depth label is performed to obtain the depth compensation value, and a numerical representation related to the vertical distribution position of the bubble cluster is established. Furthermore, a coefficient multiplication operation method (parameters: depth compensation value, second process sensitivity coefficient) is adopted to realize the coefficientization of the depth compensation value and obtain the weighted component of the sensitivity to the depth position of the pressing process; Furthermore, the multiplication unit performs a one-to-one correspondence operation between the depth compensation value and the second process sensitivity coefficient to generate a second weighted component, providing depth sensitivity input for the multidimensional combination of dynamic weights, calculated as follows:

[0018] in, This is a numerical representation of the interlayer penetration depth label, with a value ranging from 0 to 2; This is the second process sensitivity coefficient, which can be calibrated online to adapt to the response requirements of defects at different levels; By using a weighted component generation algorithm, the depth compensation value is transformed into a numerical index that can be directly superimposed, thereby achieving a priority response effect to deep defects in the process. For example, during PCB lamination, the interlayer penetration depth label of a certain bubble cluster is 2, corresponding to the core layer position, the input constant is 3, and the second process sensitivity coefficient β is 0.45. After performing the difference calculation, the depth compensation value = = Multiplying this depth compensation value by β yields the second weighted component. = This value significantly improves the response priority of the corresponding process control cell to core layer defects in dynamic weight calculation, effectively realizing the strategy of precise pressure compensation and extended heat preservation time in the case of deep defects, and ensuring a significant improvement in the removal capacity of bubble defects in key layers; S4.4: Based on the verified glass cloth tightness matching degree, calculate the difference between 1 and the glass cloth tightness matching degree as the matching compensation value, and then multiply the matching compensation value by the third process sensitivity coefficient to generate the third weighted component; Numerical reading processing is performed on the glass cloth tightness matching degree parameter (input condition: glass cloth tightness matching degree after S3.3 normalization, value range is 0 to 1), and the difference calculation algorithm is used (parameter: the baseline matching degree is set to 1) to realize the quantification of the degree of missing matching degree; Furthermore, the matching compensation value is output through the difference calculation algorithm, and the calculation formula is as follows:

[0019] in, The difference in the glass cloth tightness matching degree is used to characterize the extent of the mismatch between resin flow and glass cloth properties in the region where the bubble cluster is located. Furthermore, through the coefficient multiplication operation module (parameter: third process sensitivity coefficient) (Can be calibrated online) to achieve linear weighting of matching compensation value and process sensitivity coefficient, generating a third weighted component; Furthermore, the multiplication operation between the matching compensation value and the third process sensitivity coefficient is numerically verified to ensure that the calculation accuracy meets the process feedback strength requirements. The verification formula is as follows:

[0020] in, The third process sensitivity coefficient is multiplied, and the result is the final value of the third weighted component. Through the above coefficient weighting algorithm, the missing matching degree information is transformed into a quantified third weighted component, thereby achieving an accurate representation of the contribution of material compatibility influencing factors in the dynamic weight value. For example, in a certain multilayer board lamination process, the glass cloth tightness matching degree is 0.82, and the matching compensation value is obtained by difference calculation. = Let the sensitivity coefficient of the third process be... = Then the third weighted component is = In this embodiment, the value is directly used for the weighted summation calculation in step S4.5, and together with the first weighted component and the second weighted component, constitutes a dynamic weight value. After subsequent cell feedback intensity calculation, it is verified that the contribution of the matching degree gap significantly improves the influence of material compatibility factors on local parameter adjustment decisions, and achieves effective resin flow compensation in areas with low glass cloth porosity. S4.5: Sum the first weighted component, the second weighted component, and the third weighted component to generate a dynamic weight value, and output the dynamic weight value for subsequent calculation of the comprehensive feedback intensity of the process control cell.

[0021] Step S5: Divide the entire printed circuit board into N×M process control cells, and calculate the weighted average of the dynamic weight values ​​of all bubble clusters within each process control cell, which serves as the comprehensive feedback intensity value for that cell. Specifically, this includes: S5.1: Parse the physical dimension data of the PCB board, perform the division number calculation based on the preset resolution threshold, determine the number of rows N and columns M of the process control cells, and generate a space uniform cell division scheme covering the entire board surface; Perform high-precision analytical processing on the physical dimension data of the PCB board (parameters: length and width dimensions, unit: mm) to obtain a physical quantitative expression of the geometric boundary and total area of ​​the board surface; By using a boundary rectangularization algorithm (parameter: minimum outer rectangle approximation accuracy ≤ 0.01mm), a unified coordinate system mapping of the plate boundary is achieved, and a basic size dataset compatible with process control mesh generation is generated. Calculate the number of partitions based on a preset resolution threshold (parameter: resolution threshold). (unit: mm²), which realizes the theoretical value of converting the ratio of board area to resolution threshold into the total number of cells; Furthermore, by using a two-dimensional factorization algorithm (parameter: prioritize keeping the row-to-column ratio close to the design aspect ratio), the theoretical total number of cells is decomposed into two sets of integer solutions: the number of rows N and the number of columns M, ensuring that the value of N×M is close to the expected total number and meets the process equipment control step requirements; For the number of rows N and columns M obtained from the parsing, execute an integer adjustment optimization algorithm (parameter: adjustment range ≤ ±1 row or column) to optimize the balance between the cell aspect ratio and the total area coverage accuracy, and generate the final grid row and column pairs; Furthermore, the cell side length parameter is calculated using the grid spacing calculation formula, which is:

[0022] and

[0023] in The length of the cell's row direction. For the column length of the cell, For PCB board width, For PCB board height, and The number of rows and columns; Through the above algorithm or processing method, the physical size data of the previous step is transformed into a spatial uniform cell division scheme, realizing the structured partitioning of the entire board surface, and providing an accurate spatial basis for subsequent bubble cluster mapping and comprehensive feedback intensity calculation. For example, on a PCB board with a length of 480mm and a width of 320mm, an area of ​​153600mm² is obtained through resolution analysis, and the boundary coordinate set is output after rectangular mapping. A resolution threshold is then set. With an area of ​​1600 mm² and an area-to-threshold ratio of 96, the two-dimensional factorization algorithm outputs row and column pairs (N=12, M=8). Integer adjustment optimization maintains the row and column ratio consistent with the design aspect ratio, resulting in a final row and column count scheme that remains unchanged. Using the grid spacing calculation formula, we can obtain... =40mm, =40mm, achieving consistency control between cell size and board size. Specifically, the entire board is divided into 96 cells, each with an area of ​​40mm×40mm. This ensures that the spatial resolution of bubble defect mapping meets the process control accuracy requirements in subsequent steps and significantly improves the spatial matching of local defect feedback. S5.2: Based on the number of divisions N and M, generate a grid coordinate system of N×M process control cells, covering the entire PCB board, to define the spatial location boundary and coordinate range of each process control cell; After obtaining the parameter conditions of the number of rows N and the number of columns M of the process control cell, input the physical size data of the PCB board and the configuration information of the division quantity of N and M into the grid coordinate system generation module as the basic data source for coordinate definition calculation; A two-dimensional Cartesian mesh generation algorithm (parameters: board length L, board width W, number of rows N, number of columns M) is used to divide the physical surface of the PCB board into N rows and M columns with equal spacing in the horizontal and vertical directions, respectively, to realize the construction of a node coordinate matrix with alternating rows and columns; Furthermore, the coordinates of the four vertices of each cell are calculated using the boundary calculation method (parameters: L, W, N, M), namely the absolute positions of the top left, top right, bottom left, and bottom right, to generate the corresponding geometric boundary set for spatial location constraint. Furthermore, a coordinate range mapping method (parameter: cell vertex coordinate set) is used to represent the coverage area of ​​each cell as a horizontal x-axis interval. and the interval perpendicular to the y-axis And establish a regional index table to ensure the rapid location of the geometric center coordinates of subsequent defect clusters; Furthermore, the validity of the generated grid coordinate system is verified by using a precision constraint verification algorithm (parameters: N, M, minimum cell size threshold), which removes non-standard cells caused by size errors and maintains the uniformity and integrity of grid coverage. Through the above two-dimensional Cartesian grid generation and boundary coordinate mapping process, the row and column division results are transformed into N×M process control cell spatial location boundary and coordinate range data, realizing the expected technical effect of spatial positioning and area definition of process control cells on the entire plate surface. For example, in a scenario where the physical dimensions of the PCB are 420mm long and 280mm wide, the number of rows N is set to 14 and the number of columns M is set to 9. A two-dimensional Cartesian grid generation algorithm is used to divide the board surface into 14×9 control cells, totaling 126 cells. For the cell at row (5,4), the boundary calculation method yields the coordinates of the top left vertex as (120mm, 90mm), the top right vertex as (160mm, 90mm), the bottom left vertex as (120mm, 110mm), and the bottom right vertex as (160mm, 110mm). The corresponding horizontal x-axis interval is [120mm, 160mm], and the vertical y-axis interval is [90mm, 110mm]. In the accuracy constraint verification algorithm, the minimum cell size threshold is set to 20mm. The width and height of the cell are both found to meet the constraint conditions, so the cell is retained in the region index table. Finally, a standardized coordinate system covering the entire board surface is generated, which provides an accurate spatial positioning basis for subsequent spatial mapping and feedback intensity value calculation. S5.3: Based on the geometric center coordinates of each bubble cluster and the grid coordinate system of the process control cell, execute the spatial mapping algorithm to assign the bubble clusters to the corresponding process control cells, so as to establish an association list between cells and bubble clusters; Based on the geometric center coordinates of each bubble cluster and the established process control cell grid coordinate system, a spatial envelope determination algorithm (parameters: cell boundary rectangle coordinate range set, bubble cluster geometric center coordinate set) is used to realize the spatial ownership determination function of bubble clusters and cells; Furthermore, by using a dual-axis interval comparison method (parameters: X and Y components of the bubble cluster center coordinates, row and column boundary values ​​of the cell coordinate system), the landing point of the bubble cluster in the two-dimensional grid is verified, and the row and column index data of the bubble cluster is obtained. Furthermore, a two-dimensional index mapping algorithm (parameters: row and column indexes, cell unique identifiers) is used to generate the binding relationship between bubble clusters and cell identifiers, and to generate an initial list of associated pairs; Furthermore, by optimizing the data structure (parameters: list of related pairs, threshold for the number of bubble clusters in a cell), a standardized storage structure for the list of related relationships is constructed, and a mapping table adapted to subsequent weight calculation is generated. By using spatial mapping and index binding, the geometric coordinate data of the bubble clusters from the previous step are transformed into cell attribution determination results, achieving precise association between cells and bubble clusters and supporting efficient local defect intensity calculation. For example, on a PCB board with physical dimensions of 400mm × 500mm, the process control cells are divided into 20mm × 20mm units, forming a 20 × 25 grid coordinate system. The geometric center coordinates of bubble cluster A are (185.6mm, 312.4mm). The cell boundaries are sorted by column number in the X-axis direction as the 10th column boundary [180mm, 200mm], and sorted by row number in the Y-axis direction as the 16th row boundary [320mm, 340mm].

[0024] A spatial envelope determination algorithm is used to compare the X and Y coordinates of bubble cluster A with the column and row boundary conditions to obtain the landing point verification results of column index 10 and row index 15. A two-dimensional index mapping algorithm is then used to combine the row and column indices into a unique identifier Cell(15,10), and an entry (A, Cell(15,10)) is created in the association pair list. During data structure optimization, based on the maximum threshold of 50 associated bubble clusters per cell, the association list is stored in a matrix, forming a 20×25 matrix mapping table. Cell(15,10) registers bubble cluster A and the set of other bubble cluster identifiers belonging to that cell. This process ensures that position determination does not need to be recalculated during subsequent weighted average calculations, significantly improving mapping efficiency and data access performance. S5.4: For each process control cell, collect the dynamic weight values ​​and voxel counts of all bubble clusters within it to prepare the input data set required for weighted average calculation; S5.5: Calculate the weighted average of the dynamic weight values ​​within each process control cell, with the weight being the number of voxels in the bubble cluster. This value serves as the comprehensive feedback intensity of the process control cell, quantifying the overall impact of local defects and supporting subsequent parameter adjustment decisions. Based on the dynamic weight values ​​and voxel counts of bubble clusters within the process control cell output from step S5.4, a weighted average calculation method (parameter: weighting coefficient is the number of voxels in the bubble cluster) is used to quantify and integrate the degree of influence of different bubble clusters on cell defects. Furthermore, by using a weighted sum calculation method (parameter: the sum of the product of the dynamic weight value of the bubble cluster and the corresponding number of voxels), the weighted sum of the bubble clusters in the cell is obtained, and the unnormalized local influence value is obtained. Furthermore, a normalization method is adopted (parameter: total weight divided by the total number of elements in the bubble clusters in the cell) to convert the weighted weight sum into a comprehensive feedback intensity value and generate a numerical result that meets the requirements of process control accuracy. Furthermore, a data validity verification algorithm (parameter: remove bubble clusters with missing dynamic weight values ​​or abnormal voxel counts) is applied to ensure the statistical reliability of the comprehensive feedback intensity value and to exclude abnormal data that may affect subsequent pressing parameter adjustments. By using the above weighted average calculation and normalization process, the dynamic weight result of the previous step is transformed into a comprehensive feedback intensity value that characterizes the degree of local defect aggregation, thereby achieving the technical effect of precise execution of the subsequent pressing parameter triggering strategy. For example, in a 600mm × 500mm PCB board, the entire board is divided into 20 × 20 process control cells, each cell covering an area of ​​30mm × 25mm. A single cell contains 3 bubble clusters with dynamic weight values ​​of 2.8, 1.5, and 3.2, corresponding to 120, 80, and 150 voxels respectively. A weighted sum calculation formula is used:

[0025] The numerator is the sum of the products of the dynamic weight values ​​of each bubble cluster and the number of voxels, while the denominator is the cumulative number of voxels in the cell. According to the formula, the total weighted weight is 336 + 120 + 480 = 936, the total number of voxels is 350, and the overall feedback intensity is 2.674. After data validity testing confirmed the absence of outliers, this overall feedback intensity value was used for the dual threshold interval determination in step S6. In this implementation scenario, the calculation results significantly improved the ability to distinguish local defects from overall pressing parameter decisions, making the pressing compensation processing of high-risk cells more accurate.

[0026] Step S6: Determine whether the comprehensive feedback intensity value of each process control cell is greater than the first threshold. If it is greater than the first threshold, trigger a micro-pressure increase and extended holding time for the corresponding pressing mold zone of that cell; if the comprehensive feedback intensity value is greater than the second threshold and less than or equal to the first threshold, only a micro-pressure increase is triggered; otherwise, maintain the baseline pressing curve, where the second threshold is 70% of the first threshold. Specifically, this includes: S6.1: Based on the historical pressing defect dataset and the statistical distribution of comprehensive feedback intensity values ​​in the pre-stored process knowledge graph, perform defect elimination rate inflection point analysis to determine the first threshold T1, where the first threshold T1 corresponds to the critical intensity value at which the bubble defect elimination rate is significantly improved, and generate the first threshold parameter as the benchmark input for subsequent comparisons; The pre-stored process knowledge graph is called to access the defect data parsing module (parameters: historical pressing defect dataset, statistical distribution of comprehensive feedback intensity value) to fit the functional relationship between the full sample defect elimination rate and the comprehensive feedback intensity, and to obtain the original scatter distribution data. A multinomial regression fitting method (parameter: fitting order no higher than the third order to avoid overfitting) is used to continuously model the defect elimination rate-intensity value curve and obtain the derivative function of the curve for inflection point detection. Furthermore, by using an inflection point analysis algorithm (parameter: derivative zero-point search precision 0.001), the second derivative of the curve is calculated, and the coordinates of the inflection point position within the interval of the second derivative sign change are obtained; Furthermore, a critical intensity selection strategy is adopted (parameter: the improvement in defect elimination rate at the inflection point exceeds a set incremental threshold) to extract the comprehensive feedback intensity value corresponding to the horizontal coordinate at the inflection point and generate a candidate first threshold set; By using a robustness verification method (parameter: more than 5 historical data multi-cycle verification samples), the candidate first threshold set is compared with the defect elimination rate performance of different pressing batches. The intensity value that shows a significant increase in defect elimination rate in most batches is selected to determine the final first threshold T1. By using the inflection point analysis algorithm, the curve fitting result of the previous step is transformed into a stable and usable first threshold parameter, thereby setting the benchmark input for the subsequent double threshold determination process. For example, in a high-multilayer board lamination process optimization, the input historical lamination defect dataset contains 1200 samples, with a comprehensive feedback intensity value ranging from 0.2 to 1.0, corresponding to a defect elimination rate varying between 15% and 85%. The system calls the defect data parsing module, using the intensity value of the historical data as the independent variable and the defect elimination rate as the dependent variable, and performs third-order polynomial regression fitting to obtain the curve. Form, in which This represents the overall feedback strength value. (Regarding the function...) Perform first and second derivative calculations, accurately locate the interval where the second derivative changes from positive to negative, and determine the coordinates of the zero point of the second derivative. Detect the change in defect elimination rate at this coordinate. The value was 0.31, exceeding the preset threshold of 0.25. Therefore, the coordinate value of 0.78 was included in the candidate threshold set. Multi-cycle robustness verification showed that this intensity value corresponded to a significant improvement in defect elimination rate in the last seven batches. Therefore, the first threshold T1 was finally determined to be 0.78. This value was used as the upper limit threshold in subsequent dual-threshold interval determination, effectively improving the accuracy of pressing parameter adjustment for high-risk bubble clusters. S6.2: Obtain the set of comprehensive feedback intensity values ​​of the process control cells output by step S5. This set contains the weighted average intensity value data of N×M cells. Perform intensity value data caching processing to generate a real-time intensity value input queue as the operation object for threshold comparison. S6.3: Based on the first threshold T1 parameter, perform the second threshold T2 generation process, where the second threshold T2 is equal to 70% of the first threshold. The second threshold parameter is generated through the proportional coefficient calculation unit to ensure the correlation between the second threshold T2 and the first threshold T1 in the output. S6.4: For each cell data in the process control cell comprehensive feedback intensity value input queue, perform dual threshold interval judgment processing, compare the comprehensive feedback intensity value with the first threshold T1 and the second threshold T2, and generate intensity value interval classification indicator, wherein the classification indicator includes three types of state output: greater than the first threshold, between the second threshold and the first threshold, and less than the second threshold; S6.5: Based on the strength value range classification indicator, perform differential triggering processing of pressing parameters. When the indicator is greater than the first threshold, generate a combination command of micro-pressurization and extended holding time; when the indicator is between the second threshold and the first threshold, generate a single micro-pressurization command; when the indicator is less than the second threshold, generate a baseline pressing curve maintenance command, so as to output the final process control command set to achieve precise adjustment of local parameters. Based on the strength value range classification indicator output by step S6.4, a parameter mapping parsing algorithm (parameters: classification indicator, process parameter reference table) is used to realize the mapping process from classification status to pressing adjustment command; Furthermore, by using a combined instruction generation algorithm (parameters: micro-pressure boost amplitude range 0.15~0.3MPa, heat preservation extension time range 12~25s), the micro-pressure boost and heat preservation extension combined instruction is constructed when the classification indicator corresponds to the "greater than the first threshold" state, and an instruction data structure containing the pressing mold partition identifier, micro-pressure boost amplitude, and extended heat preservation time is obtained. Furthermore, through a single instruction generation algorithm (parameter: micro-boost amplitude range 0.15~0.3MPa), a single micro-boost instruction is constructed when the classification indicator corresponds to the state "between the second threshold and the first threshold", and an instruction data structure containing the pressing mold partition identifier and the boost amplitude is generated. Furthermore, through the baseline curve output algorithm (parameter: baseline pressing curve template), the baseline curve maintenance instruction is constructed when the classification indicator corresponds to the "less than the second threshold" state, and an instruction data structure containing mold partition identifier and curve template reference is generated; By integrating instruction sets, the instruction data structures generated under the above different states are aggregated into a unified format process control instruction set, thereby realizing the execution input required for precise adjustment of local parameters; For example, in the lamination process of a multilayer board, the N×M cells are divided into 20×15 cells. The feedback strength value of cell (8,10) corresponding to the mold partition is 42.5. The first threshold T1 is set to 40, and the second threshold T2 is set to 28. A parameter mapping parsing algorithm is used to map states with strength values ​​greater than T1 to combined adjustment categories. Combined with the combined instruction generation algorithm, a micro-pressure increase amplitude of 0.25 MPa is selected, and the holding time is extended by 20 seconds, forming the instruction: Partition (8,10) → Pressure Increase 0.25 MPa + Delay 20 seconds. For partition (5,7) with a strength value of 36.4, a single instruction generation algorithm is used, with a micro-pressure increase amplitude of 0.18 MPa, forming the instruction: Partition (5,7) → Pressure Increase 0.18 MPa. For partition (2,3) with a strength value of 22.0, based on the baseline curve output algorithm, the standard lamination curve template number C-2023 is referenced, forming the maintenance instruction: Partition (2,3) → Curve C-2023. In this scenario, the final instruction set contains three types of instructions, all of which are generated into a unified XML format through instruction set integration and processing and pushed to the pressing control system. This effectively achieves differentiated compensation for high-risk defect areas and stable operation in low-risk areas, significantly improving the ability to suppress local defects and maintaining the stability of the overall process.

[0027] Step S7: Acquire infrared hot spot images of the pressing process after local parameter adjustment, calculate the hot spot area shrinkage rate. If the hot spot area shrinkage rate is less than 15%, increase the first process sensitivity coefficient by 0.05 to achieve closed-loop self-calibration of the weight model. Specifically, this includes: S7.1: Based on the historical process data repository, retrieve the infrared hot spot image of the previous cycle of the target process control cell; perform image thresholding processing on the infrared hot spot image to identify the hot spot region boundary; calculate the pixel area statistics of the hot spot region and generate the baseline hot spot area value. S7.2: Perform infrared thermal imaging scanning on the target process control cell after local parameter adjustment in the current cycle to acquire real-time infrared hot spot images; perform morphological opening operations on the infrared hot spot images to eliminate noise interference; extract the effective pixel area of ​​the hot spot region to generate the current hot spot area value; S7.3: Input the reference hot spot area value and the current hot spot area value into the area shrinkage rate calculation module; perform area difference operation to obtain the area difference value; calculate the hot spot area shrinkage rate percentage based on the quotient of the area difference value and the reference hot spot area value; S7.4: Compare the generated hot spot area shrinkage percentage with a preset threshold of 15%; if the hot spot area shrinkage percentage is less than the threshold, output a high-priority calibration trigger signal; otherwise, maintain the low-priority calibration trigger signal state. Based on the output of the area shrinkage rate calculation module, the hot spot shrinkage threshold comparison algorithm (parameter: preset shrinkage rate threshold = 15%) is called to realize the numerical judgment of the percentage of hot spot area shrinkage rate of the current cycle target process control cell; Furthermore, through the threshold comparison operation unit (parameter: comparison type = less than judgment), the relationship between the shrinkage rate percentage and the threshold is judged, and a Boolean judgment signal is generated. A Boolean value of true indicates that the current shrinkage rate is lower than the preset threshold. Furthermore, a signal priority mapping method is adopted (parameters: high priority flag = 1, low priority flag = 0) to convert the Boolean judgment signal into a priority calibration trigger flag and obtain a binary priority state value; Furthermore, by utilizing the status value encoding conversion module (parameter: output format = trigger signal instruction set), the encoding mapping from priority markers to trigger signals is realized, generating trigger signal data packets that conform to the interface protocol of the pressing control system; By using threshold comparison and priority mapping, the area shrinkage rate determination result is converted into a high or low priority calibration trigger signal, thereby realizing the automated trigger control of the weighted model calibration process. For example, in a certain pressing cycle, the area shrinkage rate calculation module outputs a shrinkage rate percentage of 10%. The preset threshold parameter of the threshold comparison operation unit is: ,implement The decision operation outputs a Boolean decision signal of true. The signal priority mapping method assigns the higher priority flag value corresponding to the Boolean value of true. The data is passed to the status value encoding conversion module, where it is encoded and mapped to generate code values. As a high-priority calibration trigger signal data packet, it is sent to the pressing control system interface. Based on this, the system performs a weighted self-calibration action with a sensitivity coefficient increment adjustment of 0.05, and significantly improves the elimination rate of bubble defects in the high-entropy region during this period. S7.5: Perform incremental adjustment of the first process sensitivity coefficient according to the high-priority calibration trigger signal; increase the first process sensitivity coefficient by 0.05; write the updated first process sensitivity coefficient into the process parameter database to realize the closed-loop self-calibration of the weight model.

[0028] Step S8: Within 300 seconds before the end of the pressing and heating stage, complete the weight update and parameter instruction generation of all process control cells to ensure that the feedback control meets the resin softening point window requirements. Specifically, this includes: S8.1: Collect and process real-time time data during the pressing and heating stage to calculate the time difference between the current time and the end point of heating, and generate a remaining time monitoring signal as the input basis for time constraint management. S8.2: Perform a threshold comparison operation based on the remaining time monitoring signal. When the time difference is less than or equal to 300 seconds, generate a weight update trigger signal as the decision basis for starting the weight update process. S8.3: Perform hot spot area shrinkage rate calculation on the infrared hot spot monitoring data of the process control cell, and perform incremental update operation of the first process sensitivity coefficient based on the judgment result of whether the shrinkage rate is less than 15% to complete the closed-loop self-calibration of the weight model and output the updated weight parameters. S8.4: Based on the updated weight parameters and the comprehensive feedback intensity value of the process control cell, execute the parameter instruction generation algorithm with preset threshold rules to generate the corresponding pressing parameter adjustment instruction as the execution input of the pressing control system; S8.5: Output the generated pressing parameter adjustment command to the pressing control system, and record the command generation timestamp for window effect verification processing to confirm that the feedback control meets the resin softening point window requirements and output an effect confirmation signal.

[0029] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adjusting process parameters based on feedback from the detection of bubble defects during lamination of multilayer boards, characterized in that, Includes the following steps: S1: Preset high-resolution X-ray phase-contrast imaging scanning path to perform intensified scanning of abrupt changes in the stacked structure and obtain a three-dimensional voxel-level bubble density distribution map. S2: Perform eight-neighbor connectivity analysis on the three-dimensional voxel-level bubble density distribution map to extract all independent bubble clusters, and calculate the local topological entropy of each bubble cluster based on the local density level distribution of voxels within the cluster. S3: For the bubble cluster, the interlayer penetration depth label is integrated with the glass cloth tightness matching degree by the process knowledge graph annotation; S4: Calculate the dynamic weight value based on the local topological entropy of each bubble cluster, the interlayer penetration depth label, and the glass cloth tightness matching degree; S5: Divide the entire printed circuit board into process control cells, calculate the weighted average of the dynamic weight values ​​of all bubble clusters in each process control cell, and use it as the comprehensive feedback intensity value of that cell. S6: Determine whether the comprehensive feedback intensity value of each process control cell is greater than the first threshold. If it is greater than the first threshold, trigger the micro-pressure increase and heat preservation time extension of the corresponding pressing mold partition of the cell. If the overall feedback strength value is greater than the second threshold and less than or equal to the first threshold, then only micro-boost is triggered; otherwise, the baseline compression curve is maintained. S7: Collect infrared hot spot images of the pressing process after local parameter adjustment, calculate the hot spot area shrinkage rate, and if the hot spot area shrinkage rate is less than the preset threshold, increase the first process sensitivity coefficient.

2. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, Following step S7, the following is also included: S8: Complete the weight update and parameter instruction generation of all process control cells within 300 seconds before the end of the pressing and heating stage to ensure that the feedback control meets the resin softening point window requirements.

3. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, The abrupt change region of the stacked structure is generated by threshold segmentation and spatial coordinate positioning of copper thickness distribution data and semi-cured sheet flow boundary parameters, and includes the geographic coordinate information of the inner copper thickness jump region and the semi-cured sheet flow boundary zone.

4. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, The local density level is divided into 5 levels based on the number of bubble voxels in the neighborhood of the voxel, and the local topological entropy reflects the degree of non-uniformity of density distribution within the bubble cluster.

5. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, Step S3 specifically includes: The layer structure parameters of the lamination process are analyzed to obtain the total number of layers and the stacking order information. Based on the analysis results, the region boundary determination process is performed to divide the stacked structure into surface region, intermediate layer region and core layer region. Based on the region division results, an interlayer penetration depth level mapping table is generated. Based on resin impregnation experimental data, a porosity-resin impregnation rate mapping table is constructed, and the interlayer penetration depth level mapping table and the porosity-resin impregnation rate mapping table are integrated into a process knowledge graph. The layer location information of the bubble cluster is processed for regional attribution determination. Based on the layer region boundary in the process knowledge graph, the layer region where the bubble cluster is located is determined. According to the layer region attribution result, the process knowledge graph is queried to obtain the interlayer penetration depth level label. The validity verification process is performed on the interlayer penetration depth level label, and the verified interlayer penetration depth label is output. The specification information of the glass cloth in the area where the bubble cluster is located is processed by parameter extraction to obtain the porosity and tightness parameters of the glass cloth. Based on the porosity-resin wetting rate mapping table in the process knowledge graph, the resin wetting rate is queried. The glass cloth tightness matching degree is calculated according to the resin wetting rate and the preset standard rate. The glass cloth tightness matching degree is normalized and the standardized glass cloth tightness matching degree is output.

6. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 5, characterized in that, In the interlayer penetration depth level mapping table, the surface layer area corresponds to label 0, the intermediate layer area corresponds to label 1, and the core layer area corresponds to label 2.

7. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, Step S4 specifically includes: The local topological entropy, interlayer penetration depth label, and glass cloth tightness matching degree of each bubble cluster are obtained, and the first process sensitivity coefficient is verified to be no less than 0.

5. The verified process parameter set is then output. Based on the local topological entropy of the verified process parameter set and the first process sensitivity coefficient, a multiplication operation is performed to generate the first weighted component. Based on the interlayer penetration depth labels in the verified process parameter set, a depth compensation value is calculated, and the depth compensation value is multiplied by a second process sensitivity coefficient to generate a second weighted component. Based on the glass cloth tightness matching degree in the verified process parameter set, a matching compensation value is calculated, and the matching compensation value is multiplied by a third process sensitivity coefficient to generate a third weighted component. The first weighted component, the second weighted component, and the third weighted component are summed to generate a dynamic weight value.

8. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 7, characterized in that, The depth compensation value is 3 minus the difference between the interlayer penetration depth labels, and the matching compensation value is 1 minus the difference between the glass cloth tightness matching degree.

9. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, Step S5 specifically includes: The physical dimension data of the PCB board is parsed, and the number of divisions is calculated based on the preset resolution threshold. The number of rows and columns of the process control cells are determined, and a spatial uniform cell division scheme covering the entire board surface is generated. Based on the number of rows and columns, a grid coordinate system for the process control cells is generated, covering the entire PCB board, to define the spatial location boundary and coordinate range of each process control cell; Based on the geometric center coordinates of each bubble cluster and the network coordinate system of the process control cell grid, a spatial mapping algorithm is executed to assign the bubble clusters to the corresponding process control cells and establish an association list between cells and bubble clusters. For each process control cell, collect the dynamic weight values ​​and voxel counts of all bubble clusters within it; Calculate the weighted average of the dynamic weight values ​​within each process control cell, where the weight is the number of voxels, and use this average as the overall feedback intensity value for that process control cell.

10. The method for adjusting process parameters based on feedback from the detection of bubble defects in multilayer board lamination according to claim 1, characterized in that, The first threshold corresponds to the critical intensity value at which the bubble defect elimination rate is significantly improved, and the second threshold is 70% of the first threshold.

Images