A point cloud data processing method for printed circuit board detection

Abstract

The present application relates to the technical field of image analysis, in particular to a point cloud data processing method for printed circuit board detection, comprising: acquiring original three-dimensional point cloud, coaxial two-dimensional spectral reflectance image and penetrating imaging data, combining design entity data constraint to convert the penetrating data into geometric compensation point cloud and register to form a fusion data field; using local geometric feature to identify material physical boundary, dividing the fusion field into corresponding full physical layer attribute area; determining design zero reference for each level, and constructing three-dimensional tolerance body according to preset tolerance band normal offset; extracting substrate layer feature to fit nonlinear dynamic reference surface to correct original point cloud, calculating normal displacement of corrected point cloud relative to zero reference, and generating deformation distribution field; mapping the deformation distribution field to the three-dimensional tolerance body to determine candidate defects, combining original data distribution density before correction, spectral abnormal feature and overflow vector pointing consistency to perform verification, and outputting detection results through level correlation review.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, specifically to a point cloud data processing method for printed circuit board inspection. Background Technology

[0002] Traditional PCB inspection often relies on single optical imaging or surface point cloud scanning. When faced with complex vertical stacked structures, the limitations of visible light wavelength and occlusion effects cause severe projection collapse and information loss of the topological features of the target object in depth space, making it difficult for algorithms to construct a complete geometric representation in feature space. At the same time, secondary reflection interference caused by highly reflective metal sidewalls creates inconsistent artifact noise in the image gradient domain, leading to ambiguity and false enlargement in the semantic extraction of target boundaries by edge detection operators, severely limiting the sub-pixel positioning accuracy of linewidth measurement. In addition, due to the influence of global nonlinear deformation caused by production stress, traditional rigid or affine transformation registration models cannot fit complex geometric distortions. This non-rigid manifold deviation causes consistency drift in the feature matching pairs between the reference template and the image under inspection, resulting in measurement benchmark failure and algorithm misjudgment.

[0003] How to solve the problems of artifact interference, semantic attribute misjudgment, and reference failure caused by deformation in PCB 3D measurement under complex multi-material and structural occlusion environments.

[0004] To address this, a point cloud data processing method for printed circuit board inspection is proposed. Summary of the Invention

[0005] This invention aims to provide a point cloud data processing method for printed circuit board inspection. By fusing multi-source data and dynamic tolerance constraints, it enables the measurement of PCB entities at all layers and with multiple materials, and determines defect attributes based on vector deviation fields.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The system acquires the original three-dimensional point cloud data, coaxial two-dimensional spectral reflectance image, and penetration imaging data for the occluded layer of the printed circuit board under test, and retrieves the corresponding design entity data. Through preset cross-source calibration parameters and combined with the spatial position constraints of the design entity data, the penetration imaging data is converted into a geometrically compensated point cloud characterizing the inner layer morphology, and registered together with the original three-dimensional point cloud data and the two-dimensional spectral reflectance image into the reference coordinate system to form a fused data field. Based on the local geometric features of the fused data field, the physical boundaries of different materials are identified; combining the physical boundaries with the design entity data, the fused data field is divided into different point cloud regions (substrate layer, conductor layer, solder layer, solder mask layer, silkscreen layer and via plug layer) according to the process level. For each level of point cloud region, the corresponding design zero reference is determined according to the geometric configuration; according to the preset tolerance zone, the outward offset along the normal of the contour surface is performed to generate a three-dimensional tolerance body that encloses the design entity; the normal displacement of each point in the point cloud region relative to the design zero reference is calculated to construct the deformation distribution field. The deformation distribution field is mapped onto the three-dimensional tolerance volume for defect detection.

[0007] Preferably, the step of forming the fused data field includes: The standard template is synchronously detected in advance, and the relative position deflection value and scaling value between the penetration imaging device, the three-dimensional point cloud acquisition device and the spectral imaging device are calculated based on the common feature points on the standard template to obtain the cross-source calibration parameters. The pixels in the penetrating imaging data are spatially translated and rotated using the cross-source calibration parameters, and their depth coordinates are assigned according to the thickness level determined by the design entity data to generate the geometric compensation point cloud. The generated geometrically compensated point cloud, the original three-dimensional point cloud data, and the two-dimensional spectral reflectance image are spatially superimposed with the coordinate origin according to a unified ratio determined by the cross-source calibration parameters to obtain the fused data field.

[0008] Preferably, the step of dividing the fused data field into different point cloud regions according to the process level includes: Using the spectral reflectance values ​​of each sampling point in the fused data field, point clouds that conform to the preset metal spectral reflectance characteristics are labeled as metal regions, and point clouds that conform to the preset non-metal spectral reflectance characteristics are labeled as non-metal regions. Within the non-metallic region, the height change relative to the theoretical plane in the design entity data is extracted, and the point cloud with the height rise edge change within a first preset range is classified as solder resist layer, and the remaining point cloud is classified as substrate layer. The spatial inclusion relationship between the metal region and the conductor outline in the design entity data is verified. Point clouds located within the conductor outline are classified as conductor layers, and point clouds that extend beyond the conductor outline and exhibit local bulging features are classified as solder layers. The point cloud with a height higher than the solder resist layer and a spectral reflectance that conforms to the high diffuse reflectance characteristics is compared with the character positions in the design entity data to obtain the silkscreen layer; the recessed feature points in the non-metallic area that conform to the hole position coordinates in the design entity data are extracted to construct the hole plugging layer.

[0009] Preferably, after dividing the point cloud into different regions, the process also includes inter-layer boundary trimming: Extracting constraint boundaries: Extract the window outlines corresponding to the process junctions of each level from the design entity data, and project them along the normal direction perpendicular to the substrate plane to construct a three-dimensional virtual cutting surface; Interlayer attribution verification: Identify the overlapping point cloud at the junction of the solder mask layer and the conductor layer, and calculate the spatial positional relationship of each overlapping sampling point relative to the three-dimensional virtual cutting surface; Physical region redistribution: Overlapping sampling points located inside the three-dimensional virtual cutting surface are merged into the conductor layer, and overlapping sampling points located outside are merged into the solder mask layer, eliminating the regional attribute aliasing caused by material surface overflow.

[0010] Preferably, after the step of dividing the fused data field into point cloud regions, the method further includes: The surface normal vector and curvature distribution of the substrate layer are extracted to construct a nonlinear dynamic reference surface that fits the deformation of the printed circuit board surface in real time, thereby decoupling the height measurement deviation caused by physical deformation. Calculate the rate of change of point cloud curvature at the junction of conductor layer and solder layer to obtain the wetting creep gradient of the two in three-dimensional space; combine the pad opening size determined in the design entity data to perform spatial volume integration on the overflow part of solder layer and extract the volume wetting rate feature that characterizes the welding quality. Using the nonlinear dynamic reference surface as a correction reference, the system offset of the original three-dimensional point cloud data in the corresponding coordinates is corrected to obtain the corrected point cloud.

[0011] Preferably, the step of determining the corresponding design zero baseline includes: For entities with a centrally symmetric structure, the missing symmetric displacement is identified by performing self-mirror overlap calculation on the measured point cloud, and the corrected measured geometric center is used as the zero reference. For non-centrosymmetric surface levels, deviation points that conform to the foreign object reflection characteristic region are excluded during the alignment process, and the sampling points are weighted according to the spectral reflectance characteristics, with the fitted steady-state contact surface as the zero reference. Based on the corrected measured geometric center and steady-state contact surface, an independent local relative coordinate system is established for each process level, serving as the spatial origin for calculating the deformation distribution field.

[0012] Preferably, the step of constructing the deformation distribution field includes: Extract the surface contour lines of the design entity data, perform outward equidistant offset along the contour normal direction, and construct a three-dimensional tolerance body enclosed by the offset surface mesh; The curvature change rate of the conductor layer sidewall region is calculated to determine the sidewall steepness. Vector contraction correction is performed on the sidewall sampling points affected by secondary reflection interference to restore the lateral position coordinates. The normal displacement of the corrected point cloud relative to the design zero reference is interpolated in piecewise space using radial basis functions to generate a piecewise smooth deviation gradient field covering the solid surface, which serves as the deformation distribution field.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing penetrating imaging and spectral reflectance images, the problem of missing information in the identification of occluded layers and materials of a single 3D point cloud is solved. The dynamic reference surface fitting technology effectively decouples the macroscopic warping of the carrier plate from local process deviations, providing a highly robust physical measurement coordinate reference for complex hierarchical structures and improving the integrity of the data source.

[0014] 2. To address the measurement noise caused by secondary reflections from the conductor sidewalls, the true spatial coordinates of the metal edge are physically restored by calculating the sidewall steepness and performing normal vector contraction correction. Combined with radial basis function piecewise spatial interpolation, a high-precision deviation gradient field is generated while preserving the edge abrupt change characteristics, thus enabling the characterization of the microscopic deformation trends at each process level of the PCB.

[0015] 3. By verifying the density alignment of the dual-track point cloud data before and after correction, and combining the spectral reflectance anomaly features with the local pointing consistency analysis of the overflow vector, a composite judgment mechanism from the underlying physical properties to the high-level logical association was constructed. This effectively shielded the false interference of algorithm correction artifacts and isolated noise points, and achieved effective identification of composite defects such as inter-layer alignment deviation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a point cloud data processing method for printed circuit board inspection according to the present invention. Figure 2 This is a flowchart illustrating the interlayer boundary trimming logic of the present invention. Figure 3 This is a schematic diagram of the multidimensional composite defect determination process of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1 to 3 This invention provides a point cloud data processing method for printed circuit board inspection, referring to... Figure 1 The flowchart and technical solution are as follows: The system acquires the original three-dimensional point cloud data, coaxial two-dimensional spectral reflectance image, and penetration imaging data for the occluded layer of the printed circuit board under test, and retrieves the corresponding design entity data. Through preset cross-source calibration parameters and combined with the spatial position constraints of the design entity data, the penetration imaging data is converted into a geometrically compensated point cloud characterizing the inner layer morphology, and registered together with the original three-dimensional point cloud data and the two-dimensional spectral reflectance image into the reference coordinate system to form a fused data field. Based on the local geometric features of the fused data field, the physical boundaries of different materials are identified; combining the physical boundaries with the design entity data, the fused data field is divided into different point cloud regions according to the process level. For each level of point cloud region, the corresponding design zero reference is determined according to the geometric configuration; according to the preset tolerance zone, the outward offset along the normal of the contour surface is performed to generate a three-dimensional tolerance body that encloses the design entity; the normal displacement of each point in the point cloud region relative to the design zero reference is calculated to construct the deformation distribution field. The deformation distribution field is mapped onto the three-dimensional tolerance volume for defect detection.

[0019] Example 1:

[0020] This embodiment uses point cloud data processing for printed circuit board inspection. Through collaborative acquisition and spatial fusion of multi-source heterogeneous data, combined with point cloud region division based on physical hierarchy attributes, dynamic tolerance constraints, and vector deviation field analysis, it achieves complete measurement and reliable judgment of process defects at all levels of the printed circuit board.

[0021] First, the original 3D point cloud data, coaxial 2D spectral reflectance image, and penetration imaging data for the occluded layer of the printed circuit board under test are acquired, and the corresponding design entity data is retrieved. Based on the grayscale projection representing the occluded layer structure in the penetration imaging data, the vertical height information corresponding to the grayscale projection is assigned according to the thickness stacking order in the design entity data, generating a geometrically compensated point cloud with depth coordinates. Using the common feature points determined by the design entity data as positioning anchor points, the geometrically compensated point cloud and the original 3D point cloud data are spatially overlapped through coordinate axis alignment and spatial rotation to fill the scanning blind spots formed by occlusion in the original 3D point cloud. The spectral reflectance values ​​in the coaxial 2D spectral reflectance image are mapped to the surface of the spatially overlapped point cloud according to the corresponding pixel positions, giving the fused data field the optical property characteristics representing the material type.

[0022] Specifically, after the 3D scan is completed, a 3D coordinate matrix covering the entire surface of the board under test is output. Each element in the matrix records the 3D spatial coordinates of the corresponding position. This matrix is ​​written into the coordinate channel of the memory buffer in binary point cloud format as the spatial skeleton for all subsequent calculations. The main control computer synchronously sends a query request to the production management system to retrieve the corresponding design entity data file using the current board number as the index. The file stores the theoretical contour polygon list of each process layer, the nominal thickness value of each layer, and the hole position coordinate table in a hierarchical structure. After reading, it is parsed into a hierarchical object array in memory for direct use in depth assignment and subsequent registration steps.

[0023] A hyperspectral linear array camera sharing the same optical axis with a 3D scanning device synchronously acquires the reflectance spectrum at each position during the scanning stroke, covering the visible to near-infrared bands. Each scanning position outputs a multi-dimensional spectral reflectance numerical vector. After scanning, a multi-channel spectral image is formed that is completely corresponding to the 3D coordinate matrix in pixel coordinates. This image is written to the spectral channels of the memory buffer in floating-point format and shares row and column indices with the 3D coordinate matrix, ensuring that two types of data can be directly associated and read at any pixel coordinate through the same index.

[0024] The penetrating imaging device performs penetrating irradiation on the test plate and outputs a transmission projection image that reflects the attenuation intensity of the entire layer along the vertical direction at each position. This image is written to the memory buffer for later use. At this time, the transmission image only carries the pixel coordinate information in the horizontal plane and does not yet have the vertical depth dimension. The depth value needs to be assigned in conjunction with the design entity data.

[0025] In the depth assignment stage, the pixel matrix of the transmission image and the hierarchical object array of the design entity data are used as dual inputs. The program traverses each inner layer hierarchical object in the design entity data arranged in the process stacking order, reads the nominal thickness value corresponding to each layer, and calculates the absolute height coordinates of each layer in the order of accumulation from the bottom surface of the substrate upwards. Taking the height of the bottom surface of the substrate as the zero reference value, the height coordinate of the first inner layer is equal to the substrate thickness, the height coordinate of the second inner layer is equal to the sum of the substrate thickness and the thickness of the first inner layer, and so on. For grayscale values ​​in the transmission image that exceed a set threshold (the set threshold in the transmission image is dynamically calculated), ... Method determined as follows: Extract the average gray value of the background area without physical obstruction in the transmission image, and set a threshold of 1.5-2.5 times the average gray value of the background; or in the 8-bit quantized image, set an empirical fixed value of 120-180 for the pixels, find the corresponding level in the design entity data based on its horizontal coordinate, and assign the absolute height coordinate of the corresponding level to the depth field of the pixel. After the assignment is completed, each effective pixel carries three fields: horizontal coordinate, vertical coordinate and height coordinate, forming a geometric compensation point cloud with complete three-dimensional coordinates, and writes it into the compensation point cloud channel of the memory buffer.

[0026] The registration stage uses geometrically compensated point cloud and original 3D point cloud data as dual inputs, and the theoretical coordinates of common feature points marked in the design entity data as the anchor point set. The common feature points are selected as the reference hole centers at the four corners and the center of the plate surface. The program locates the measured coordinates of the corresponding hole centers in the geometrically compensated point cloud and the original 3D point cloud respectively, and calculates the rigid transformation parameters between the two sets of hole center coordinate sets, including the translation along the horizontal dual axes and the rotation angle around the vertical axis. The calculated transformation parameters are applied to all sampling points of the geometrically compensated point cloud to make its coordinate system completely aligned with the coordinate system of the original 3D point cloud. After alignment, the program searches for the missing grid cells of effective sampling points in the original 3D point cloud, inserts the sampling points of the geometrically compensated point cloud that fall into the missing cells into the original point cloud, completes the blind spot filling, and writes the merged point cloud after filling into the fusion coordinate channel of the memory buffer. In this stage, no overwrite operation is performed on the coordinates of the existing sampling points in the original 3D point cloud.

[0027] The spectral attribute mapping stage uses the merged point cloud and multi-channel spectral image as dual inputs. Based on the coaxial mounting relationship, the program establishes a one-to-one correspondence index table between the horizontal coordinates of the point cloud and the pixel coordinates of the spectral image. For each sampling point in the merged point cloud, the program looks up its corresponding spectral image pixel position through the index table, reads the spectral reflectance vector of that pixel, and writes it into the attribute channel field of the sampling point's structure. For blind sampling points filled by geometrically compensated point clouds, since their positions are not covered by corresponding coaxial spectral images, the program reads the material type label of the level to which the point belongs from the design entity data, assigns it initial spectral attributes based on the pre-stored standard spectral template vectors of each material, and adds a confidence flag to the attribute field, marking it as an inferred assignment rather than a measured assignment, for subsequent hierarchical division stages to perform weight reduction processing on this type of point. After attribute mapping is completed, each sampling point in the merged point cloud carries both three-dimensional coordinates and a spectral reflectance vector, which are written as a whole into the fused data field storage area in structure array format, forming a unified input data source for all subsequent analysis steps.

[0028] Furthermore, the steps for forming the fused data field include: pre-detecting the standard template synchronously, and using the common feature points on the standard template as a reference, calculating the relative position deflection value and scaling value between the penetrating imaging device, the three-dimensional point cloud acquisition device and the spectral imaging device to obtain the cross-source calibration parameters; Specifically, the cross-source calibration parameters include: the three-dimensional translation vector of the penetrating imaging device relative to the reference coordinate system, the rotation angle vector around the three axes of the reference coordinate system, the scaling factor of the pixel size of the penetrating imaging device relative to the length unit of the reference coordinate system, and the two-dimensional translation vector and scaling factor of the spectral imaging device relative to the reference coordinate system. The cross-source calibration parameters are obtained in advance through the following steps: using a standard template as a common detection object, the surface of the template is provided with common feature points that can be stably detected in all three types of acquired images; using the coordinate system of the three-dimensional point cloud acquisition device as the reference coordinate system, the relative position deflection value and scaling value between the penetrating imaging device and the reference coordinate system, and between the spectral imaging device and the reference coordinate system are calculated respectively.

[0029] The pixels in the penetrating imaging data are spatially translated and rotated using the cross-source calibration parameters, and their depth coordinates are assigned according to the thickness level determined by the design entity data to generate the geometric compensation point cloud. The generated geometrically compensated point cloud, the original three-dimensional point cloud data, and the two-dimensional spectral reflectance image are spatially superimposed with the coordinate origin according to a unified ratio determined by the cross-source calibration parameters to obtain the fused data field.

[0030] Specifically, in the calibration phase, a standard template is used as the common detection object. The surface of the template has common feature points that can be stably detected in all three types of acquired images. The feature points are composed of several geometrically regular and clearly defined reference marks. The theoretical coordinates of each mark have been measured with high precision and recorded in the calibration data file. The three types of acquisition devices perform a complete synchronous detection on the standard template without moving the test board. Each device outputs raw data containing feature point response information. This raw data is stored in the memory buffer in the format of each device's own pixel coordinates or point cloud coordinates, serving as the raw input for calibration calculation.

[0031] The program locates the measured coordinates of common feature points from the outputs of the three types of devices. Using the coordinate system of the 3D point cloud acquisition device as the reference coordinate system, it calculates the relative position deflection and scaling values ​​between the penetration imaging device and the reference coordinate system, and between the spectral imaging device and the reference coordinate system. The deflection value represents the rotation angle of each device's coordinate axis relative to the reference coordinate axis, and the scaling value represents the conversion coefficient between the pixel size of each device and the length unit of the reference coordinate system. The above parameters together constitute the cross-source calibration parameter set, which is written to the configuration file in structured text format for persistent storage. Subsequent batch testing processes can directly read and reuse this file without having to repeat the calibration operation for each board under test.

[0032] The geometric compensation point cloud generation stage uses the pixel matrix of the penetrating imaging data and the cross-source calibration parameter set as dual inputs. The program reads the deflection value and scaling value corresponding to the penetrating imaging device, and performs spatial translation and rotation processing on each effective pixel in the penetrating image in sequence, mapping it from the pixel coordinate system of the penetrating imaging device to the horizontal plane of the reference coordinate system. After completing the coordinate alignment in the horizontal plane, the program reads the nominal thickness values ​​of each occluded layer in the design entity data file, and accumulates them layer by layer from the bottom of the substrate upwards according to the process stacking order, assigning the absolute height coordinate value of the corresponding pixel in the reference coordinate system to the corresponding pixel of each layer. After the three-step processing of translation, rotation and depth assignment, all penetrating pixels that originally only had two-dimensional projection coordinates obtain complete three-dimensional coordinates, forming a geometric compensation point cloud. This point cloud is written into the compensation point cloud channel of the memory buffer in the same structure array format as the original three-dimensional point cloud, for direct reading in the next stage of spatial stacking.

[0033] In the spatial overlay stage, geometrically compensated point cloud, original 3D point cloud data, and 2D spectral reflectance image are used as three inputs. The program reads the unified scale coefficient and coordinate origin definition determined by the cross-source calibration parameter set, and uniformly converts the three input data to the same scale and the same spatial origin. Since the original 3D point cloud data is based on the reference coordinate system, no additional transformation is required. The geometrically compensated point cloud has already completed coordinate alignment in the previous stage. The spectral reflectance image completes the mapping of pixel coordinates to reference coordinates according to the calibration parameters corresponding to the spectral imaging device. After the three data are aligned, the program uses the original 3D point cloud as the main body, inserts the sampling points that fill the blind spots in the geometrically compensated point cloud into the main point cloud, and then writes the reflectance vector of each pixel in the spectral image into the attribute field of the merged point cloud point by point according to the coordinate correspondence. Finally, it outputs a fused data field in which each sampling point carries both 3D coordinates and spectral reflectance vector, and writes the fused data field in memory in the form of a structure array as a unified input data source for subsequent hierarchical division and defect detection.

[0034] By pre-calculating the spatial transformation parameters across devices using a standard template, three types of heterogeneous data sources are unified into the same coordinate system, achieving precise alignment and complete fusion of multi-source data in geometric and attribute dimensions, and eliminating the systematic interference of equipment installation deviations on subsequent measurement results.

[0035] Further, the step of dividing the fused data field into different point cloud regions according to the process level includes: using the spectral reflectance values ​​of each sampling point in the fused data field, marking point clouds that conform to preset metal spectral reflectance characteristics (the metal spectral reflectance characteristic template is characterized by: reflectance greater than 70% in the near-infrared band, such as 800-1000nm, while reflectance is less than 40% in the visible short-wave band, exhibiting typical high specular reflectance gradient characteristics; the non-metal spectral reflectance characteristic template, corresponding to media such as epoxy resin, is characterized by: a smooth spectral response distribution in the entire visible to near-infrared band, a reflectance fluctuation variance of less than 0.05, and an overall diffuse reflectance stable in the range of 20%-40%) as metal regions, and marking point clouds that conform to preset non-metal spectral reflectance characteristics as metal regions, and marking point clouds that conform to preset non-metal spectral reflectance characteristics as metal regions. The point cloud of the feature is labeled as a non-metallic region; within the non-metallic region, the height change relative to the theoretical plane in the design entity data is extracted, and the point cloud with the height rise change within a first preset range is classified as a solder resist layer, and the remaining point cloud is classified as a substrate layer; the spatial inclusion relationship between the metallic region and the conductor contour in the design entity data is verified, and the point cloud located within the conductor contour is classified as a conductor layer, and the point cloud exceeding the conductor contour and exhibiting local bulge features is classified as a solder layer; the point cloud with a height higher than the solder resist layer and a spectral reflectance conforming to the high diffuse reflectance feature is compared with the character position in the design entity data to obtain the silkscreen layer; the recessed feature points in the non-metallic region that conform to the hole position coordinates in the design entity data are extracted to construct the via plugging layer.

[0036] Specifically, this stage takes the array of structures in the fused data field storage area as input. Each sampling point carries two types of attributes: three-dimensional coordinates and spectral reflectance vector. The hierarchical division program performs attribute judgment on each sampling point in turn. The judgment result is written back to the structure of the sampling point in the form of a hierarchical label field. After all sampling points are labeled, six types of point cloud region arrays are output according to the label values, corresponding to the substrate layer, conductor layer, solder layer, solder resist layer, silkscreen layer and via plug layer, respectively. Each group array is stored independently in the hierarchical point cloud buffer in memory as the classification input for subsequent tolerance body construction and deformation field calculation.

[0037] In the initial classification stage of metals and non-metals, the spectral reflectance vector of each sampling point is used as input. The program reads the metal spectral reflectance feature template and the non-metal spectral reflectance feature template stored in the configuration file. The metal feature template is based on the high specular reflectance distribution of typical conductor materials such as copper and tin in the visible to near-infrared band. The non-metal feature template is based on the diffuse reflectance distribution of typical media materials such as epoxy resin substrate and solder resist ink. The program calculates the normalized correlation coefficient between the spectral reflectance vector of each sampling point and the two types of templates. The category label corresponding to the side with the higher correlation coefficient is written into the initial classification field of that point. The output result divides all sampling points into two subsets: the metal region point set and the non-metal region point set. The two subsets enter the subsequent subdivision and classification process respectively.

[0038] The non-metallic region subdivision stage uses the non-metallic region point set and the theoretical substrate plane parameters in the design entity data as dual inputs. The program calculates the height change of each sampling point in the non-metallic region relative to the theoretical substrate plane, and calculates the difference of the height change of adjacent sampling points along the horizontal direction of the board surface to obtain the height rise edge change of each point. The program reads the first preset range set in the configuration file, which corresponds to the height step amplitude value generated by the typical thickness range of the solder resist ink layer. Sampling points whose height rise edge change falls within the first preset range are classified as solder resist layers and written with corresponding labels. Sampling points whose height change is close to zero or does not meet the rise edge criterion are classified as substrate layers and written with corresponding labels. The two types of labeling results are stored in the solder resist layer point cloud buffer and the substrate layer point cloud buffer, respectively. Corresponding to the standard coating thickness of the typical solder resist ink layer in the printed circuit board manufacturing process, it can effectively eliminate the small roughness fluctuations of the substrate itself and the interference of thick surface mount components.

[0039] The metal region subdivision stage uses the metal region point set and the conductor contour polygon list in the design entity data as dual inputs. The program performs a judgment on the inclusion relationship of the horizontal coordinate of each sampling point in the metal region within the polygon. It traverses all conductor contour polygons in the design entity data. If the horizontal coordinate of the sampling point falls inside or on the boundary of any conductor contour polygon, it is classified as a conductor layer. If the horizontal coordinate of the sampling point exceeds the range of all conductor contour polygons, the program further checks the height distribution of the point and its neighboring point set. If the height coordinate in the neighborhood shows local maximum characteristics, that is, the height of the point is significantly higher than the surrounding non-metallic region sampling points and forms a continuous raised surface, it is classified as a solder layer. The two types of annotation results are stored in the conductor layer point cloud buffer and the solder layer point cloud buffer, respectively.

[0040] In the silkscreen layer extraction stage, the program uses all the sampling points that have completed the initial classification and the character position annotations in the design entity data as dual inputs. The program filters out the set of points whose height coordinates are higher than the average height of the sampling points in the solder resist layer point cloud buffer. In this height-filtered subset, sampling points whose spectral reflectance vectors meet the high diffuse reflectance characteristics are further extracted. The high diffuse reflectance characteristics are based on the broadband uniform reflectance distribution of silkscreen ink in the visible light band. The selected subset is spatially overlapped with the character contour area recorded in the design entity data. Sampling points whose horizontal coordinates fall within the character contour range are classified as silkscreen layers. The annotation results are stored in the silkscreen layer point cloud buffer.

[0041] In the hole-filling layer construction stage, the non-metallic region point set and the hole position coordinate table in the design entity data are used as dual inputs. The program reads the center coordinates and nominal hole diameter of each hole in the hole position coordinate table, and retrieves sampling points whose horizontal coordinates fall within the range of each hole position in the non-metallic region point set. The difference between the height coordinate of this type of sampling point and the average height of the surrounding substrate layer is calculated. Sampling points with negative height difference values, i.e., showing concave features, are classified as hole-filling layers. The labeling results are stored in the hole-filling layer point cloud buffer. At this point, all sampling points in the fused data field have completed the hierarchical labeling. The six types of hierarchical point cloud buffers are passed to the subsequent processing flow as complete classification output.

[0042] By applying multi-dimensional joint criteria, including spectral reflectance characteristics, elevation rise time variation, spatial inclusion relationship, and design entity data, layer by layer, the precise classification of point clouds at six process levels is achieved, avoiding misclassification caused by a single criterion in material overlapping regions.

[0043] Furthermore, after dividing the data into different point cloud regions, the process also includes interlayer boundary trimming: extracting the window outlines corresponding to the process boundaries of each layer from the design entity data and projecting them along the normal direction perpendicular to the substrate plane to construct a three-dimensional virtual cutting surface; identifying the overlapping point clouds at the interface between the solder mask layer and the conductor layer, calculating the spatial positional relationship of each overlapping sampling point relative to the three-dimensional virtual cutting surface; merging the overlapping sampling points located inside the three-dimensional virtual cutting surface to the conductor layer, merging the overlapping sampling points located outside to the solder mask layer, and eliminating the regional attribute aliasing caused by material surface overflow.

[0044] Reference Figure 2The flowchart of the interlayer boundary trimming logic of the present invention is shown below. Specifically, in this stage, the six types of layer point cloud buffers after the layer division is completed and the design entity data file are used as dual inputs. The interlayer boundary trimming program reads the sampling point structure array in the conductor layer point cloud buffer and the solder mask layer point cloud buffer in sequence, as well as the list of polygon vertex coordinates of the window contour lines at each process interface recorded in the design entity data. The trimming result is written back to the structure of the corresponding sampling point with the updated layer label field. After the trimming is completed, the contents of the two types of buffers are refreshed synchronously, and the conductor layer and solder mask layer point clouds after boundary trimming are output for subsequent tolerance body construction process to call.

[0045] The constraint boundary construction stage takes the window outlines at the process junctions of each layer in the design entity data as input. The window outlines are stored in the format of vertex coordinate sequence of closed polygons. The horizontal position of the solder mask window edge relative to the substrate plane is recorded. The program reads the vertex coordinate sequence of each window outline and performs a stretching projection operation on each outline along the normal direction perpendicular to the substrate plane. The stretching height covers the complete height range of the conductor layer and solder mask in the vertical direction, so that the outline forms a closed vertical curved surface in three-dimensional space, that is, a three-dimensional virtual cutting surface. This cutting surface is stored in the cutting surface buffer in memory in the format of triangular mesh facets. Its horizontal cross-sectional outline is completely consistent with the corresponding window outline in the design entity data, and serves as the spatial reference for subsequent overlapping point cloud assignment judgment.

[0046] The interlayer attribution verification stage uses the conductor layer point cloud buffer, the solder mask layer point cloud buffer, and the cutting surface buffer as three inputs. The program searches for spatially adjacent sampling point pairs between the conductor layer point cloud and the solder mask layer point cloud. Specifically, for each sampling point in the conductor layer point cloud buffer, a neighborhood search range centered on its horizontal coordinate is established. Within this range, sampling points existing in the solder mask layer point cloud buffer are searched. The set of sampling points with corresponding points in the neighborhood range of the two types of point clouds is marked as the overlapping point cloud at the boundary. The overlapping point cloud set is stored in memory in the form of a temporary list. For each sampling point in the overlapping point cloud list, the program calculates the spatial positional relationship of its three-dimensional coordinates relative to the triangular mesh facet in the cutting surface buffer. The criterion is the sign of the signed distance value obtained after the sampling point is projected onto the cutting surface along the normal direction. A positive value indicates that it is located outside the cutting surface, and a negative value indicates that it is located inside the cutting surface. The judgment result is written into the temporary attribute of the sampling point as an inside / outside flag field.

[0047] The physical region redistribution stage takes a temporary list of overlapping point clouds carrying internal and external flags as input. The program traverses all sampling points in the list, updates the layer label field of sampling points with internal and external flags set to the conductor layer, and updates the layer label field of sampling points with internal and external flags set to the solder mask layer. The update operation is directly written back to the structure of the corresponding sampling point in the fused data field storage area, and the contents of the conductor layer point cloud buffer and the solder mask layer point cloud buffer are refreshed simultaneously. After the refresh is completed, there is no longer a situation in the two types of buffers where the same sampling point is simultaneously assigned to two layers. The material aliasing area attribute error caused by solder mask ink overflow at the conductor edge is corrected at the physical level. The two types of point cloud buffers after correction are passed to the subsequent processing flow as the final output of this stage.

[0048] By constructing a three-dimensional virtual cutting surface using the window outline in the design entity data as a spatial reference, and performing signed distance judgment on the overlapping point clouds at the intersection and forcibly reassigning the hierarchical affiliation, the property aliasing caused by material overflow is eliminated from the physical definition level, ensuring the geometric accuracy of the boundary between the conductor layer and the solder mask layer.

[0049] Further, after dividing the fused data field into point cloud regions, the method further includes: extracting the surface normal vector and curvature distribution of the substrate layer, constructing a nonlinear dynamic reference surface that fits in real time with the deformation of the printed circuit board surface, and decoupling the height measurement deviation caused by physical deformation; calculating the curvature change rate of the point cloud at the junction of the conductor layer and the solder layer to obtain the wetting and creep gradient of the two in three-dimensional space; combining the pad opening size determined in the design entity data, performing spatial volume integration on the overflow portion of the solder layer to extract the volume wetting rate feature characterizing the soldering quality; using the nonlinear dynamic reference surface as a correction reference to correct the system offset of the original three-dimensional point cloud data in the corresponding coordinates to obtain the corrected point cloud.

[0050] Specifically, this stage uses the six types of layer point cloud buffers after the interlayer boundary trimming is completed and the design entity data file as joint input. The feature extraction program sequentially reads the sampling point structure arrays in the substrate layer point cloud buffer, conductor layer point cloud buffer and solder layer point cloud buffer. The intermediate calculation results of each stage are written into the feature storage area in memory as independent fields. Finally, the nonlinear dynamic reference surface parameter file, volume wettability feature array and corrected point cloud structure array are output. The three types of outputs are stored in the corresponding buffers for subsequent tolerance body construction and deformation field calculation.

[0051] In the nonlinear dynamic reference surface construction stage, all sampling points in the substrate layer point cloud buffer are used as input. The program calculates the local surface normal vector for each sampling point in the substrate layer point cloud. Specifically, several adjacent sampling points are selected in the neighborhood of the sampling point as the center, and local plane fitting is performed on the neighborhood point set. The normal direction of the fitted plane is written as the surface normal vector of the sampling point and written into the normal vector field. After obtaining the normal vector of each sampling point, the program further calculates the curvature value of each sampling point. The curvature is calculated by extracting the divergence of the normal vector field in the neighborhood of the point. Regions with large absolute divergence values ​​correspond to warped regions with high local curvature, and regions with divergence close to zero correspond to locally flat regions. The normal vector field and curvature distribution together constitute the geometric feature field describing the macroscopic deformation state of the substrate layer, which is written into the reference surface feature buffer in memory in the format of a floating-point array.

[0052] The program takes the normal vector field and curvature distribution in the reference surface feature buffer as input, and uses radial basis functions to perform global piecewise interpolation fitting on the height coordinates of the substrate sampling points. The influence radius of the interpolation kernel function is adaptively adjusted with the curvature value of each sampling point as a weight. The influence radius is reduced in areas with high curvature to preserve local deformation details, while the influence radius is expanded in areas with low curvature to maintain the overall smoothness of the fitted surface. The interpolation calculation is performed point by point on the grid nodes covering the entire board surface. Each grid node outputs a fitted height value. The fitted height values ​​of all grid nodes constitute a nonlinear dynamic reference surface. This reference surface is stored in the reference surface buffer in a regular grid height map format. Its spatial resolution is consistent with the sampling interval of the fused data field. Its physical meaning is the theoretical reference height distribution determined only by the macroscopic warpage of the carrier board after eliminating the influence of process layer superposition.

[0053] Specifically, the nonlinear dynamic reference surface uses multiple quadratic radial basis functions to perform overall block interpolation and fitting of the height coordinates of points collected on the substrate layer. This method constructs a fitting model based on the position of each collection point and its corresponding vertical displacement. The entire surface is calculated by superimposing the forces acting on each point under the multiple quadratic radial basis functions. The function contains shape parameters used to control the shape of the surface. These shape parameters are adaptively adjusted according to the surface curvature at each collection point. Based on the basic shape parameters, the shape parameters of the current point are dynamically determined by combining the absolute curvature of the surface at that point and the preset weight ratio.

[0054] The resulting adjustment relationship aims to achieve the following: in areas of severe surface curvature, automatically reduce the influence range of that point on the surrounding area to accurately preserve local deformation details; while in areas of gentle surface curvature, automatically expand the influence range to maintain the overall smoothness of the final fitted surface.

[0055] The wetting climb gradient and volumetric wetting rate extraction stages use the conductor layer point cloud buffer and the solder layer point cloud buffer as dual inputs. The program first retrieves spatially adjacent sampling point pairs between the conductor layer and the solder layer to form a boundary region point set. For each sampling point in the boundary region point set, the rate of change of point cloud curvature along the solder climb direction is calculated. Specifically, the first-order difference of the curvature value is calculated along the path from the conductor layer sidewall to the solder layer surface. The difference result reflects the change of the climb slope of the solder at the conductor edge. This difference value is defined as the wetting climb gradient at that location and written into the wetting gradient field of the feature storage area in scalar array format. The spatial distribution of the wetting climb gradient describes the three-dimensional morphological characteristics of the solder climbing along the conductor sidewall.

[0056] The program reads the opening contour polygon and nominal opening size of the corresponding pad from the design entity data. It marks the sampling points in the solder layer point cloud buffer whose horizontal coordinates exceed the range of the pad opening contour as an overflow point set. It performs spatial volume integration on the overflow point set. The integration method is to project the overflow point set onto the horizontal plane to form an overflow contour region. For each grid cell in this region, the difference between the height value of its corresponding sampling point and the height value of the pad opening edge is taken as the local height. Multiplying the local height by the area of ​​the grid cell yields a local volume element. The summation of all the elements yields the solder overflow volume. The solder overflow volume is then divided by the nominal solder volume of the pad to obtain the volume wettability feature value that characterizes the soldering quality. This value is written in floating-point format to the wettability field of the feature storage area according to the pad number for subsequent defect judgment stage reading.

[0057] The point cloud correction stage uses the nonlinear dynamic datum height map in the datum buffer and the original 3D point cloud data as dual inputs. For each sampling point in the original 3D point cloud, the program interpolates the corresponding grid node's fitted height value in the nonlinear dynamic datum height map based on its horizontal coordinates. The difference between the fitted height value and the theoretical design datum height is defined as the system offset at that location. The system offset reflects the height measurement error introduced by the macroscopic warping of the carrier plate at that location. The program subtracts the corresponding system offset from the height coordinate of the sampling point in the original 3D point cloud to obtain the corrected height value after eliminating the influence of carrier plate warping. The correction operation is performed point by point on all sampling points in the original 3D point cloud, and the corrected point cloud structure array is output. The height coordinate of each sampling point has been restored to the true deviation that only reflects the local process deformation. The corrected point cloud is written into the correction point cloud buffer in the same structure array format as the original point cloud as the direct input for subsequent deformation distribution field calculation.

[0058] By performing adaptive radial basis function fitting on the normal vector field and curvature distribution of the substrate point cloud, the macroscopic warpage of the carrier plate is decoupled and removed from the height measurement data. At the same time, the wetting creep gradient and volume wetting rate of the welding area are extracted as quantitative characteristics of the welding quality, providing a high-precision correction data basis after eliminating systematic errors for subsequent deformation field analysis.

[0059] Furthermore, the step of determining the corresponding design zero reference includes: for the entity level with a centrally symmetric structure, identifying the symmetric missing displacement by performing self-mirror overlap calculation on the measured point cloud, and using the corrected measured geometric center as the zero reference; for the non-centrally symmetric surface level, excluding deviation points that conform to the foreign object reflection characteristic region during the alignment process, and assigning weights to the sampling points according to the spectral reflectivity characteristics, using the fitted steady-state contact surface as the zero reference; and establishing an independent local relative coordinate system for each process level based on the corrected measured geometric center and the steady-state contact surface, serving as the spatial origin for calculating the deformation distribution field.

[0060] Specifically, in this stage, the point clouds of each level in the corrected point cloud buffer and the design entity data file are used as joint inputs. The zero reference determination program divides the six levels into two branches based on the geometric configuration attributes of each level: centrosymmetric entity level and non-centrosymmetric surface level. The centrosymmetric entity level includes the via layer and the conductor via region with a regular rotating body contour. The non-centrosymmetric surface level includes the planar contact region of the substrate layer, solder mask layer, silkscreen layer and solder layer. The processing results of both branches are written into the zero reference parameter file in the form of the local coordinate system origin coordinates and coordinate axis direction vectors. The zero reference buffer in memory is indexed by the level number and stored as needed for subsequent deformation distribution field calculation.

[0061] In the stage of determining the zero reference for the centrally symmetric entity level, the corrected point cloud of the corresponding level and the theoretical symmetry axis parameters of the same level in the design entity data are used as dual inputs. The program takes the theoretical geometric center coordinates recorded in the design entity data as the initial reference, performs mirror flipping of the measured point cloud of the level along the two orthogonal axes in the horizontal plane to generate a mirror point cloud corresponding to the original point cloud. The mirror point cloud and the original point cloud are spatially overlapped. The symmetry missing displacement is taken as the 3D coordinate difference vector between the corresponding sampling points of the two. The direction of the symmetry missing displacement points to the offset direction of the measured geometric center relative to the theoretical symmetry axis, and the magnitude reflects the spatial distance of the offset. The program translates the initial theoretical geometric center coordinates along the opposite direction of the symmetry missing displacement by the corresponding magnitude to obtain the corrected measured geometric center coordinates. The coordinates are written into the zero reference buffer as the design zero reference of the level. The physical meaning of the corrected measured geometric center is the spatial point that best represents the true center position of the symmetric structure after eliminating the actual processing eccentricity.

[0062] In the non-centrosymmetric surface level zero reference determination stage, the corrected point cloud of the corresponding level and the spectral reflectance vector carried by the sampling points of that level are used as dual inputs. The program first performs similarity calculation between the spectral reflectance vector of each sampling point in the point cloud of that level and a pre-stored foreign matter reflection feature template (the foreign matter reflection feature template is established based on typical foreign contaminants, such as copper oxide film and flux residue, and its specific spectral characteristics are: in a specific oxidation characteristic absorption peak band, such as 500-600nm, its reflectance is more than 30% lower than that of the normal reference metal, producing a significant spectral absorption notch). The foreign matter reflection feature template is established based on the reflection characteristics of typical foreign matter such as foreign contaminants, oxide film and flux residue in the spectral domain. Sampling points with similarity exceeding the set threshold are marked as foreign matter deviation points and removed from subsequent fitting calculations. The removal operation is achieved by resetting the participation weight of the corresponding sampling point to zero, without deleting the original data. The similarity threshold uses cosine similarity, with a specific value of 0.85. This is based on the spectral statistical analysis of a large number of normal benchmark and typical foreign object samples. The optimal classification boundary was determined by plotting ROC curves and optimizing between the trace foreign object false alarm rate and the normal tolerance false alarm rate (maximizing the F1 score).

[0063] The program assigns fitting weights to the remaining sampling points that have passed the foreign object screening based on the degree of matching between their spectral reflectance vectors and the standard material template of that level. Sampling points with a high degree of matching receive a weight value close to 1, while sampling points with a low degree of matching receive a weight value close to zero. The weight values ​​are written to the temporary weight field of each sampling point in floating-point format. The program performs weighted least squares plane fitting with the set of sampling points carrying the weight field as input. During the fitting process, the weight value of each sampling point is used as a scaling factor for its coordinate error contribution. Sampling points with high weights have a stronger constraint on the fitting results, while sampling points with low weights have a minimal impact on the fitting results. After the weighted fitting converges, it outputs a spatial plane parameter that best represents the real physical contact surface of that level, i.e., the steady-state contact surface. The normal vector of the steady-state contact surface and the coordinates of the in-plane reference point are written to the zero reference buffer as the design zero reference of that level.

[0064] In the local coordinate system establishment stage, the zero-datum parameters of all levels in the zero-datum buffer are used as input. The program establishes an independent local relative coordinate system for each level according to its zero-datum type. For the centrosymmetric entity level, the corrected measured geometric center is used as the origin of the coordinate system, and the axis direction of the structure in the design entity data is used as the principal axis direction. Two orthogonal auxiliary axes are established in the plane perpendicular to the principal axis to form a complete three-axis local coordinate system. For the non-centrosymmetric surface level, the normal vector of the steady-state contact surface is used as the vertical axis direction, and the in-plane reference point is used as the origin of the coordinate system. The direction parallel to the long side of the plate surface is selected in the steady-state contact surface as the horizontal principal axis to establish a three-axis local coordinate system. The local coordinate system parameters of each level are appended to the zero-datum buffer in the form of origin coordinates and three-axis direction vectors according to the level number to form a complete level coordinate system parameter table, which serves as the spatial origin definition for the calculation of the normal displacement of each sampling point in the subsequent deformation distribution field calculation.

[0065] By performing self-mirror overlap to identify eccentricity at the centrosymmetric level and performing spectral weighted steady-state surface fitting at the non-centrosymmetric level, the interference of processing eccentricity and foreign matter contamination on zero-reference positioning is eliminated, thus establishing an independent spatial reference system with clear physical meaning and strong anti-interference ability for deformation field calculation at each process level.

[0066] Further, the step of constructing the deformation distribution field includes: extracting the surface contour lines of the design entity data, performing outward equidistant offset along the contour normal direction, and constructing a three-dimensional tolerance body enclosed by the offset surface mesh; calculating the rate of curvature change of the conductor layer sidewall region to determine the sidewall steepness, performing vector contraction correction pointing towards the entity sidewall sampling points affected by secondary reflection interference, and restoring the lateral position coordinates; and using radial basis functions to perform piecewise spatial interpolation on the normal displacement of the corrected point cloud relative to the design zero reference to generate a piecewise smooth deviation gradient field covering the entity surface, which serves as the deformation distribution field.

[0067] Specifically, this stage uses the design entity data file, the local coordinate system parameters of each level in the zero reference buffer, and the corrected point cloud structure array in the corrected point cloud buffer as three joint inputs. The deformation distribution field construction program sequentially completes three sub-processes: three-dimensional tolerance body generation, conductor sidewall sampling point vector shrinkage correction, and radial basis function piecewise interpolation. Finally, it outputs a piecewise smoothed deviation gradient field with reference to each level of local coordinate system, which is written into the deformation distribution field storage area in grid scalar field format according to the level number, as the direct input for the subsequent defect detection stage.

[0068] The local curvature of the conductor layer sidewall region is calculated to assess the steepness of the sidewall, and positional contraction correction is performed towards the interior of the solid for the sidewall sampling points affected by secondary reflection interference.

[0069] The specific calculation logic for the correction magnitude is as follows: based on the deviation between the actual measured position and the theoretical design position of the sampling point, multiply by a correction intensity coefficient determined through statistical experiments along the vertical direction inward from the sidewall to obtain the final shrinkage amount. The magnitude of the correction intensity coefficient establishes an adaptive linkage mechanism with the steepness of the sidewall: for high-steepness areas where the steepness exceeds the preset limit, a larger correction intensity is assigned; conversely, for low-steepness areas, a smaller correction intensity is assigned. This adaptive mechanism dynamically assigns a constrained correction intensity of 0.5 (a slight correction to preserve the true shape) to 0.9 (a depth correction to strongly eliminate reflection artifacts) based on the local steepness calculated in real time. The trigger threshold of the preset limit is derived from physical statistical experiments. By comparing the optical scanning measurement values ​​with the physical true values ​​obtained by high-precision equipment, the critical inflection point where the error surges due to secondary reflection is located can be determined. The specific threshold can be adjusted according to the actual situation.

[0070] In the 3D tolerance body construction stage, the program takes the polygon vertex coordinate sequence of the surface contour lines of each level of entity in the design entity data as input. The program reads the preset tolerance zone values ​​corresponding to each level. The tolerance zone values ​​are pre-written into the configuration file based on the manufacturing tolerance specifications of each process level. The program extracts the local normal direction at each vertex of each surface contour line. The normal direction is calculated in the contour plane based on the angle bisector of the adjacent edge vectors. According to the preset tolerance zone, the program performs an outward offset along the surface normal to generate a 3D tolerance body that envelops the design entity. The preset tolerance zone values ​​are determined according to the manufacturing tolerance specifications of each process level in the design entity data. Specifically, they include the substrate layer tolerance zone, conductor layer tolerance zone, solder layer tolerance zone, and solder mask layer tolerance zone. The tolerance values ​​of each layer are determined by the IPC standard or design documents. The program reads the preset tolerance zone values ​​corresponding to each level, extracts the local normal direction at each vertex of each surface contour line, and translates each vertex outward along its normal direction by the distance corresponding to the tolerance zone value to generate the offset contour vertex coordinates. Each vertex is translated outward along its normal direction by the distance corresponding to the tolerance zone value, generating the offset profile vertex coordinates. The profile vertices before and after the offset are connected according to the corresponding relationship to form a triangular mesh patch. After the top and bottom surfaces are closed, a three-dimensional tolerance body that completely envelops the design entity is formed. This three-dimensional tolerance body is stored in the tolerance body buffer in a triangular mesh format with hierarchical numbering. Its inner surface corresponds to the design zero reference profile, and its outer surface corresponds to the tolerance boundary. The closed space between the two layers is the spatial range in which the sampling point should fall under the qualified process state.

[0071] In the conductor sidewall sampling point vector shrinkage correction stage, the corrected sampling points in the conductor layer point cloud buffer and the sidewall contour parameters of the conductor entity in the design entity data are used as dual inputs. The program extracts the local curvature change rate of each point in the sampling point subset located in the sidewall region of the conductor layer point cloud. Specifically, it calculates the first-order difference of the curvature value sequence along the sidewall normal direction in the neighborhood of each sampling point in the subset. Sampling points with large absolute difference values ​​correspond to regions with high sidewall steepness, while sampling points with absolute difference values ​​close to zero correspond to regions with gentle sidewall slopes. The program uses the sidewall steepness threshold as a criterion to filter out the high-steepness sidewall sampling point subset with a high probability of secondary reflection interference. Secondary reflection interference in the high-steepness sidewall region manifests as the sampling point coordinates shifting outward along the sidewall normal direction, resulting in the measured sidewall contour being wider than the actual conductor width. The sidewall steepness threshold is determined based on the material's optical properties and statistical experiments, corresponding to the physical inflection point where the incident angle of the laser or structured light from the 3D scanning equipment exceeds the critical angle, causing the secondary reflection artifact error to increase exponentially.

[0072] The program calculates the unit vector pointing to the interior of the entity for each of the selected high-steepness sidewall sampling points. This unit vector is based on the inner normal direction of the designed entity sidewall contour at that point. The product of the curvature change rate of that point and the preset correction ratio coefficient is the shrinkage amplitude. The sampling point coordinates are translated along the inner normal direction according to the shrinkage amplitude to obtain the corrected lateral position coordinates. The correction operation pulls the false coordinates that have been offset to the outside of the entity contour back to the spatial position that matches the actual conductor sidewall. The correction result is written back to the structure coordinate field of the sampling point in the updated three-dimensional coordinates. After all high-steepness sidewall sampling points have completed vector shrinkage correction, the updated conductor layer point cloud structure array is written to the sidewall correction point cloud buffer for the next sub-process to read.

[0073] In the radial basis function piecewise space interpolation stage, the sampling point structure array in the sidewall correction point cloud buffer and the local coordinate system parameters of each level in the zero reference buffer are used as dual inputs. The program performs the following processing for each level: First, the three-dimensional coordinates of all sampling points in the level are transformed from the reference coordinate system to the local relative coordinate system of the level. After the transformation, the vertical axis coordinate component of each sampling point is the normal displacement value of the point relative to the design zero reference. The normal displacement value is represented by a signed floating-point number. A positive value indicates that the sampling point is located outside the design zero reference, and a negative value indicates that it is located inside. The normal displacement value array is written into the normal displacement temporary buffer with the sampling point number as the index.

[0074] The program takes the sampling point positions and normal displacement values ​​in the temporary buffer of normal displacement as input. It uses radial basis functions to perform piecewise spatial interpolation on the surface of the hierarchical entity. The piecewise interpolation method is based on the geometric partitioning of the hierarchical entity. Independent interpolation subdomains are established for the planar and curved regions respectively. The radial basis functions in each subdomain use the horizontal coordinates of all sampling points within that subdomain as the basis function center and the corresponding normal displacement value as the fitting target. The weight coefficients of each basis function are determined by solving a system of linear equations. The equation system is solved using a direct decomposition method. After the solution is complete, each subdomain outputs a continuous normal value. The displacement interpolation function can output the corresponding normal displacement interpolation at any spatial location within the subdomain. The interpolation results of adjacent subdomains at the common boundary are weighted and mixed to ensure a continuous transition. The interpolation functions of all subdomains constitute a piecewise smooth deviation gradient field on the surface of the entity at this level. The deviation gradient field is written into the deformation distribution field storage area in a regular grid scalar field format with the same spatial resolution as the tolerance buffer. The physical meaning is the continuous spatial distribution of the normal deviation of each position on the surface of the entity at this level relative to the design zero datum, which serves as the input data source for the overflow vector calculation in the defect detection stage.

[0075] Furthermore, during the construction of the deformation distribution field, the definitions of the normal displacement references corresponding to different geometric configurations of the process layers are fundamentally different. The plugging layer with a body of revolution profile uses the radial direction of the axis as the normal displacement reference, while the substrate layer with a planar profile uses the normal direction of the curved surface as the normal displacement reference. Directly splicing the two types of references at the junction of the plugging edge and the substrate surface will introduce gradient jumps in the deviation field. To eliminate the above jumps, after the radial basis function piecewise interpolation is completed, the program performs continuity processing on the boundary regions where there is a reference definition switch in adjacent piecewise subdomains, and constructs a gradual transition from the boundary line to both sides. In the buffer zone, Gaussian weighted coefficients are constructed for each grid node within the buffer zone based on its distance from the boundary line. The interpolation results of the two types of reference subdomains are then weighted and mixed to ensure a smooth transition of the normal displacement value from one type of reference definition to the other. After mixing, bilateral filtering is further applied to the nodes within the boundary buffer zone, using spatial proximity distance and similarity of normal displacement value as dual weight kernels to smooth residual high-frequency jumps while preserving the abrupt change characteristics of the true deformation gradient. After the above two-stage processing, the deformation distribution field satisfies the gradient continuity requirement across the entire domain and does not introduce false overflow vectors at the boundary due to the switching of reference definitions.

[0076] By performing inner normal direction vector contraction correction on the high steepness region of the conductor sidewall, the virtual expansion error of the sidewall coordinate introduced by the secondary reflection is eliminated. Combined with the piecewise adaptive interpolation strategy of the radial basis function, a physically continuous deviation gradient field is generated while preserving the edge abrupt change characteristics, thus realizing a high-precision spatial characterization of the micro deformation trend of each process level.

[0077] Furthermore, the defect detection steps specifically include: determining candidate defects by calculating the overflow vector of each point in the deformation distribution field relative to the boundary of the three-dimensional tolerance body; verifying the validity of the candidate defects by combining the distribution density of the original three-dimensional point cloud data before correction and the abnormal features in the coaxial two-dimensional spectral reflectance image, and calculating the local pointing consistency of the overflow vector within the candidate defect region that has passed the verification to distinguish between isolated noise points and solid deformation defects; performing a consistency review by combining the increase or decrease attributes of the candidate defects with the hierarchical correlation of the design entity data to determine whether there are composite defects and output the detection results.

[0078] The orientation consistency calculation involves averaging the unit direction vectors of each overflow vector within the candidate defect region and calculating the standard deviation of the angle between each vector and the mean direction. The consistency threshold is set to a standard deviation of 10 degrees; values ​​below this threshold are considered highly consistent in orientation and identified as solid deformation defects; otherwise, they are identified as isolated noise. Extensive scanning of calibration plates known to contain real continuous deformation defects was performed to extract the angle distribution of local overflow vectors. Statistical results show that within the real physical defect region, over 95% of the overflow vectors deviate from their mean direction by less than 10 degrees.

[0079] Specifically, this stage uses five inputs: the piecewise smoothing deviation gradient field in the deformation distribution field storage area, the three-dimensional tolerance body mesh in the tolerance body buffer, the original three-dimensional point cloud data, the coaxial two-dimensional spectral reflectance image in the fused data field, and the design entity data file. The defect detection program sequentially completes three sub-processes: overflow vector calculation and candidate defect calibration, validity verification and orientation consistency analysis, and hierarchical correlation verification and result output. Finally, it outputs a list of detection result structures indexed by defect number. Each structure contains fields for defect level, spatial location, overflow amplitude, defect type, and composite attribute flags, which are written to the detection result storage area for subsequent report generation module to read.

[0080] In the overflow vector calculation and candidate defect calibration stage, the deviation gradient field mesh at each level in the deformation distribution field storage area and the corresponding level of the 3D tolerance body triangular mesh in the tolerance body buffer are used as dual inputs. For each mesh node in the deviation gradient field, the program determines whether the node is outside the tolerance range defined by the 3D tolerance body based on its normal displacement value in the local coordinate system. Specifically, the determination method is to calculate the signed distance from the 3D coordinates of the node to the nearest triangular facet on the outer surface of the 3D tolerance body. A positive signed distance indicates that the node has exceeded the outer surface of the tolerance body, that is, the measured deformation of the node exceeds the tolerance body's tolerance range. The upper limit of the allowable tolerance is given. When the signed distance is negative and its absolute value exceeds the thickness of the tolerance body, it means that the node falls within the inner surface of the tolerance body. That is, the measured deformation of the node is lower than the lower limit of the allowable tolerance. Both cases are judged as overflow. The program uses the direction vector of the node position pointing to the nearest point on the tolerance body boundary and the magnitude of the signed distance to form the overflow vector of the node. The direction of the overflow vector represents the spatial direction of the deformation offset, and the magnitude represents the degree of exceeding the tolerance. The overflow vectors of all overflow nodes are written into the candidate defect temporary buffer in the format of a structure array to form a set of candidate defects to be verified.

[0081] Specifically, the signed distance from the node to be judged to the nearest triangular facet on the outer surface of the 3D tolerance body is calculated. The specific distance determination logic is as follows: First, calculate the distance from the node along the normal pointing outwards from the tolerance body to the plane containing the nearest triangular facet. If the projection of the node onto the aforementioned plane falls directly into the interior region of the triangular facet, then the aforementioned planar distance is taken as the final signed distance; if the projection falls outside the interior region of the triangular facet, then the final signed distance is taken as the shortest spatial straight-line distance from the node to each boundary line of the triangular facet, and is assigned a corresponding positive or negative sign according to its relative position to indicate the direction.

[0082] The thickness of the tolerance body is set based on the preset tolerance zone size, and defect screening is performed accordingly: When the calculated signed distance is positive, it indicates that the node protrudes from the outer surface of the tolerance body and is extracted as a candidate defect point; when the calculated signed distance is negative and the absolute value of its deviation exceeds the thickness of the tolerance body, it indicates that the node is recessed too deeply and crosses the inner boundary of the tolerance body, and is also extracted as a candidate defect point.

[0083] The validity verification stage takes the overflow node set in the candidate defect temporary buffer, the original 3D point cloud data, and the coaxial 2D spectral reflectance image as three inputs. The program retrieves the number of sampling points in the neighborhood of the corresponding spatial location of each overflow node in the candidate defect set in the original 3D point cloud data, and calculates the point cloud distribution density in the neighborhood. If the number of sampling points in the neighborhood is significantly lower than the average point cloud density of the normal area at the same level, it is determined that the area corresponding to the overflow node has data sparsity anomaly in the original scan. The overflow signal of such nodes may be caused by scan missing rather than entity deformation. The program sets its validity flag to pending confirmation and enters the spectral verification stage. Overflow nodes with normal point cloud density directly pass the density verification.

[0084] The spectral verification step takes overflow nodes that have passed density verification or are in a pending confirmation state as input. The program reads the spectral reflectance vector of the corresponding pixel position in the coaxial two-dimensional spectral reflectance image and performs a similarity calculation with the standard spectral template of the normal material at that level. If the similarity is lower than a set threshold, it is determined that there are spectral abnormal features at that position. The spectral abnormal features indicate that the material properties of the region are inconsistent with the normal process state, which is consistent with the overflow vector signal. The program updates the validity flag of the node that has both spectral abnormality and point cloud overflow to confirmed valid, updates the validity flag of the node with point cloud density abnormality and no spectral abnormality to confirmed invalid, and removes it from the candidate defect set. The set of candidate defect nodes that have passed the validity verification is written into the valid candidate defect buffer in the updated structure array format.

[0085] The consistency analysis phase takes the set of overflow nodes in the valid candidate defect buffer as input. The program clusters spatially adjacent overflow nodes in this set. The clustering method uses the three-dimensional coordinates of each overflow node as input to perform connected component labeling based on spatial proximity. Overflow nodes with a spatial distance less than a set neighborhood radius are grouped into the same connected component. Each connected component is processed independently as a candidate defect region. The program performs local consistency calculations on the overflow vector directions of all overflow nodes within each candidate defect region. The consistency calculation method involves averaging the unit direction vectors of all overflow vectors within the region and then calculating the standard deviation of the angle between each vector and the mean direction. Regions with small standard deviations indicate that the overflow vectors point in a highly consistent manner, corresponding to overall offset defects caused by continuous deformation of the entity. Regions with large standard deviations indicate that the overflow vectors point in a discrete manner, corresponding to randomly distributed isolated noise points. The program uses a set consistency threshold as a criterion to mark candidate defect regions with consistency below the threshold as noise and remove them from the valid candidate defect buffer. Candidate defect regions with consistency above the threshold are confirmed as entity deformation defects. At the same time, based on the main direction of the overflow vector, each confirmed defect region is labeled with additive or subtractive properties. Overflow vectors pointing outside the tolerance body indicate additive properties, and those pointing inside the tolerance body indicate subtractive properties. The labeling results are written into the defect structure field corresponding to the region.

[0086] The hierarchical correlation verification stage uses the array of confirmed defect structures labeled with addition and subtraction attributes and the hierarchical correlation rules in the design entity data as dual inputs. The design entity data predefines correlation rules between levels that establish physical constraints in the manufacturing process. For example, if an additive defect in the conductor layer and a subtractive defect in the adjacent solder layer highly overlap in spatial location, they are physically correlated. Individually classifying them as two independent defects might lead to misjudgment, as they are actually the same solder bridging composite defect. The program iterates through all confirmed defect structures and performs consistency verification on defects that overlap across layers in spatial location. The verification logic involves judging two adjacent layers... If the combination of added and removed attributes of a level defect matches a known composite defect pattern, the two are merged into a single composite defect record, and the composite type field is written into the merged defect structure. If they do not match, the independent defect records remain unchanged. After all checks are completed, the program sorts the final defect structure list by level number and spatial location coordinates and writes it into the detection result storage area. Each record includes the defect level, spatial boundary coordinates, average overflow amplitude, added and removed attributes, consistency score, and composite defect flag field. The complete list in the detection result storage area is passed to the report generation module as the final output of this method.

[0087] By employing a four-level progressive judgment mechanism—including overflow vector calibration, dual-track data validity verification, pointing consistency noise elimination, and cross-layer correlation verification—a complete defect confirmation link is constructed from the underlying physical signal to the upper-level process semantics. This effectively suppresses false alarms caused by algorithm correction artifacts and isolated noise, and enables synchronous and reliable identification of single-layer deformation defects and cross-layer composite defects.

[0088] This embodiment is based on a standard single-layer board inspection scenario. The object under test is a rigid printed circuit board produced using conventional stacking technology. By sequentially performing multi-source heterogeneous data fusion, process-level point cloud division based on physical properties, dynamic reference surface correction and three-dimensional tolerance volume constraint, and multi-level defect judgment based on overflow vector field on the board under test, a complete closed-loop processing link from raw data acquisition to defect detection conclusion is established. This covers the full-element measurement needs of different process levels and solves the problem of missing information in occluded areas and material identification when using single point cloud scanning. It also achieves synchronous and reliable identification of single-layer deformation defects and composite defects.

[0089] Example 2:

[0090] This embodiment addresses the problem that cross-layer misalignment composite defects caused by the large number of stacked layers and the extremely small interlayer alignment tolerance in high-density interconnect printed circuit boards are difficult to identify effectively; During the interlayer boundary trimming stage, based on the construction of the 3D virtual cutting surface, a cross-layer topology connectivity verification mechanism is further introduced. The point cloud buffers of all completed boundary trimming layers and the theoretical alignment tolerances between each layer recorded in the design entity data are used as joint inputs. According to the layer stacking order defined in the design entity data, the program performs pairwise verification of the positional correspondence of point clouds between adjacent layers in the vertical direction in the horizontal plane. The verification method is to project the upper and lower layer point clouds into contour polygons in the horizontal plane, calculate the centroid coordinate offset of the two contour polygons, and compare the offset with the theoretical alignment tolerance of the corresponding layer pair in the design entity data. Layer pairs whose offset exceeds the theoretical alignment tolerance are marked as having interlayer alignment candidate anomalies. The spatial location of the candidate anomaly, the layer number involved, and the offset magnitude are written into the interlayer anomaly temporary buffer in structure format for direct call by the cross-layer correlation verification in the defect detection stage.

[0091] Furthermore, for each candidate anomaly record in the interlayer anomaly temporary buffer, the program extracts a subset of edge sampling points located in the centroid offset direction from the point clouds of the corresponding two layers. The point cloud density gradient along the offset direction is calculated for each subset. The density gradient is represented by the first difference of the inverse sequence of the distance between adjacent sampling points. A subset with a monotonically increasing density gradient in the offset direction indicates that there is an inward shrinkage of the entity edge in that direction at that layer, while a subset with a monotonically decreasing density gradient indicates that there is an outward expansion of the entity edge. The program determines the physical cause type of the interlayer alignment deviation based on the combination of the density gradient directions of the two layer subsets, and writes the determination result into the corresponding record in the interlayer anomaly temporary buffer as an offset type label field. The offset type label distinguishes between three types: single-layer slip, double-layer opposite offset, and double-layer same-direction offset, providing a pre-classification basis for the qualitative classification of composite defects in the subsequent defect detection stage.

[0092] In the defect detection phase, to address the issue of interference between adjacent overflow vector fields caused by the high spatial density of multiple candidate defects in high-density interconnect boards, an adaptive neighborhood radius strategy is introduced based on the connected component clustering method. The program uses the set of overflow nodes in the effective candidate defect buffer and the local point cloud density statistics of the layer to which each node belongs as dual inputs. For each overflow node in the effective candidate defect buffer, the program calculates the adaptive neighborhood radius based on the local point cloud density of its layer region. Regions with high local point cloud density correspond to fine routing areas with extremely small line widths and spacing in the high-density interconnect board, and the adaptive neighborhood radius of such regions is reduced accordingly to avoid adjacent independent defects being mistakenly merged into the same connected component. Regions with low local point cloud density correspond to wide-spacing routing areas, and the adaptive neighborhood radius of such regions is expanded accordingly to ensure that sparse overflow nodes can be correctly classified into the same defect region. The adaptive neighborhood radius value is written to the temporary radius field of each overflow node in floating-point format. During clustering, the average of the adaptive neighborhood radii of two nodes is used as the distance threshold for determining whether they are connected. After clustering, the calculation process for the consistency of overflow vector pointing of each connected component is consistent with that in Example 1.

[0093] Furthermore, during the hierarchical correlation verification stage, the program performs spatial location matching between the candidate anomaly records in the interlayer anomaly temporary buffer and the confirmed defect structure array. For defect records with overlapping spatial locations and interlayer candidate anomaly records, a joint verification is performed. The joint verification logic is to determine whether the increase or decrease attribute of the confirmed defect at that spatial location and the offset type label of the interlayer candidate anomaly are physically consistent. For example, if the confirmed defect of a conductor layer at a certain location is a subtractive property and the offset type of the interlayer candidate anomaly is single-layer slip, then the two are highly consistent in physical cause. The program determines that the defect at that location is a composite defect of conductor layer displacement caused by interlayer alignment deviation, updates the composite type field to the interlayer alignment deviation class, and merges it into a single composite defect record. After all joint verifications are completed, the program outputs a final detection result list containing detailed annotations of composite defect types and writes it to the detection result storage area.

[0094] This embodiment introduces cross-layer topological connectivity verification and density gradient cause analysis during the interlayer boundary trimming stage, pre-establishes type labels for interlayer alignment candidate anomalies, and solves the problem of mis-merging of defects in high-density areas with an adaptive neighborhood radius clustering strategy during the defect detection stage. Finally, through physical self-consistency joint verification, interlayer alignment deviation type composite defects are accurately distinguished and identified from ordinary deformation defects, effectively improving the detection rate and qualitative accuracy of composite defects in high-density interconnect printed circuit boards under complex stack-up structures.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A point cloud data processing method for printed circuit board inspection, characterized in that, Includes the following steps: The system acquires the original three-dimensional point cloud data, coaxial two-dimensional spectral reflectance image, and penetration imaging data for the occluded layer of the printed circuit board under test, and retrieves the corresponding design entity data. Through preset cross-source calibration parameters and combined with the spatial position constraints of the design entity data, the penetration imaging data is converted into a geometrically compensated point cloud characterizing the inner layer morphology, and registered together with the original three-dimensional point cloud data and the two-dimensional spectral reflectance image into the reference coordinate system to form a fused data field. Based on the local geometric features of the fused data field, the physical boundaries of different materials are identified; combining the physical boundaries with the design entity data, the fused data field is divided into different point cloud regions according to the process level. For point cloud regions at each level, the corresponding design zero reference is determined based on the geometric configuration; and an outward offset along the normal of the contour surface is performed according to the preset tolerance zone to generate a three-dimensional tolerance body that envelops the design entity. Calculate the normal displacement of each point within the point cloud region relative to the design zero reference, and construct the deformation distribution field; The deformation distribution field is mapped onto the three-dimensional tolerance volume for defect detection.

2. The point cloud data processing method for printed circuit board inspection according to claim 1, characterized in that, The steps for forming the fused data field include: A standard template is pre-detected synchronously, and the relative positional deflection and scaling values ​​between the penetrating imaging device, the 3D point cloud acquisition device, and the spectral imaging device are calculated based on the common feature points on the standard template to obtain the cross-source calibration parameters. The pixels in the penetrating imaging data are spatially translated and rotated using the cross-source calibration parameters, and depth coordinates are assigned according to the thickness level determined by the design entity data to generate the geometric compensation point cloud. The generated geometric compensation point cloud, the original 3D point cloud data, and the 2D spectral reflectance image are spatially superimposed according to the unified scale and coordinate origin determined by the cross-source calibration parameters to obtain the fused data field.

3. The point cloud data processing method for printed circuit board inspection according to claim 1, characterized in that, The step of dividing the fused data field into different point cloud regions according to the process level includes: Using the spectral reflectance values ​​of each sampling point in the fused data field, point clouds that conform to the preset metal spectral reflectance characteristics are labeled as metal regions, and point clouds that conform to the preset non-metal spectral reflectance characteristics are labeled as non-metal regions. Within the non-metal regions, the height change relative to the theoretical plane in the design entity data is extracted, and point clouds whose height rise edge change is within a first preset range are classified as solder resist layers, and the remaining point clouds are classified as substrate layers. The spatial inclusion relationship between the metal region and the conductor outline in the design entity data is verified. Point clouds located within the conductor outline are classified as conductor layers, and point clouds that extend beyond the conductor outline and exhibit local bulging features are classified as solder layers. The point cloud with a height higher than the solder resist layer and a spectral reflectance that conforms to the high diffuse reflectance characteristics is compared with the character positions in the design entity data to obtain the silkscreen layer; the recessed feature points in the non-metallic area that conform to the hole position coordinates in the design entity data are extracted to construct the hole plugging layer.

4. The point cloud data processing method for printed circuit board inspection according to claim 3, characterized in that, After dividing the point cloud into different regions, the process also includes trimming the interlayer boundaries: Extracting constraint boundaries: Extract the window outlines corresponding to the process junctions of each level from the design entity data, and project them along the normal direction perpendicular to the substrate plane to construct a three-dimensional virtual cutting surface; Interlayer attribution verification: Identify the overlapping point cloud at the junction of the solder mask layer and the conductor layer, and calculate the spatial positional relationship of each overlapping sampling point relative to the three-dimensional virtual cutting surface; Physical region redistribution: Overlapping sampling points located inside the three-dimensional virtual cutting surface are merged into the conductor layer, and overlapping sampling points located outside are merged into the solder mask layer, eliminating the regional attribute aliasing caused by material surface overflow.

5. The point cloud data processing method for printed circuit board inspection according to claim 1, characterized in that, After dividing the fused data field into point cloud regions, the method further includes: The surface normal vector and curvature distribution of the substrate layer are extracted to construct a nonlinear dynamic reference surface that is fitted in real time with the deformation of the printed circuit board surface, thereby decoupling the height measurement deviation caused by physical deformation; the curvature change rate of the point cloud at the junction of the conductor layer and the solder layer is calculated to obtain the wetting and creep gradient of the two in three-dimensional space. Based on the pad opening size determined in the design entity data, the overflow portion of the solder layer is spatially integrated to extract the volumetric wetting rate feature that characterizes the soldering quality; the nonlinear dynamic reference surface is used as a correction reference to correct the system offset of the original three-dimensional point cloud data in the corresponding coordinates, resulting in the corrected point cloud.

6. The point cloud data processing method for printed circuit board inspection according to claim 1, characterized in that, The steps for determining the corresponding design zero baseline include: For entities with a centrally symmetric structure, the missing symmetric displacement is identified by performing self-mirror overlap calculation on the measured point cloud, and the corrected measured geometric center is used as the zero reference. For non-centrosymmetric surface levels, deviation points that conform to the foreign object reflection characteristic region are excluded during the alignment process, and the sampling points are weighted according to the spectral reflectance characteristics, with the fitted steady-state contact surface as the zero reference. Based on the corrected measured geometric center and steady-state contact surface, an independent local relative coordinate system is established for each process level, serving as the spatial origin for calculating the deformation distribution field.

7. The point cloud data processing method for printed circuit board inspection according to claim 1, characterized in that, The steps for constructing the deformation distribution field include: The surface contour lines of the design entity data are extracted, and an outward equidistant offset is performed along the contour normal direction to construct a three-dimensional tolerance body enclosed by the offset surface mesh. The curvature change rate of the conductor layer sidewall region is calculated to determine the sidewall steepness. Vector contraction correction is performed on the sidewall sampling points affected by secondary reflection interference to restore the lateral position coordinates. The normal displacement of the corrected point cloud relative to the design zero reference is piecewise spatially interpolated using radial basis functions to generate a piecewise smooth deviation gradient field covering the entity surface as the deformation distribution field.

Images