PCB board defect detection method and system based on CT three-dimensional reconstruction

Abstract

This invention specifically relates to a PCB board defect detection method and system based on CT 3D reconstruction, belonging to the field of image data processing technology. It includes: clustering prior exception regions to mark them as 3D voxel clusters, i.e., defect candidate regions; extracting 3D morphological parameter sets from iterative intermediate data round by round and calculating the comprehensive change in morphological parameters; outputting 3D morphological parameter sets as an exemption confirmation instruction for stable regions and a constraint recovery instruction for fluctuating regions. In this invention, the projection data of each viewpoint are differentially weighted according to the per-view signal-to-noise characteristic parameters and the per-view penetration path length, and homogeneous constraints are applied only within the same structural layer according to the per-voxel layer partitioning. Simultaneously, in each reconstruction iteration, the prior violation degree of each voxel is calculated from the projection residual components and the intra-layer deviation, causing the structural constraint strength to continuously weaken as the prior violation degree increases. This ensures that sparse artifacts in normal structural regions are fully suppressed by structural constraints and prevents defect signals in suspected defect regions from being smoothed out by the constraint mechanism.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method and system for detecting defects in PCB boards based on CT three-dimensional reconstruction. Background Technology

[0002] Printed circuit boards (PCBs), as the core interconnecting carriers of electronic products, are developing towards higher layer counts, higher density, and finer pitch. Defects such as solder joint voids, interlayer delamination, microcracks, and bridging are deeply embedded within the board and cannot be directly observed through visual inspection. X-ray CT tomography and 3D reconstruction technology can obtain three-dimensional information about the internal structure of PCBs without damaging the board, and has become the mainstream method for non-destructive testing of internal defects in PCBs.

[0003] A PCB board is a flat object composed of copper foil traces, solder, and resin dielectric layers. The penetration path of rays incident along the board surface is much larger than that incident along the normal direction, resulting in significant differences in the signal-to-noise levels of projection data from different viewpoints. The time constraints of online inspection limit the number of sampling viewpoints, and 3D reconstruction under sparse sampling inevitably involves a large number of artifacts. These objective constraints cause minute defect signals and artifact signals to mix together in the reconstructed 3D volume data, making direct interpretation difficult.

[0004] In existing technologies, one approach filters tomographic images layer by layer and compares them with a standard template to identify defects. However, layer-by-layer analysis cannot utilize the three-dimensional spatial morphology information of defects, making it difficult to distinguish between real defects and reconstruction artifacts. Another approach introduces structural prior constraints to improve reconstruction quality, but these constraints are applied indiscriminately to all areas. While the structural priors suppress artifacts, they also gradually smooth out real defects, failing to differentiate between suspected defect areas. Furthermore, both approaches employ a unified detection mode after reconstruction convergence, making the detection cycle time difficult to adapt to the needs of online production lines. Therefore, existing technologies have limited effectiveness in detecting minute internal defects in PCB boards and cannot meet the requirements for high-precision online inspection of high-density PCB boards. Summary of the Invention

[0005] The purpose of this invention is to propose a PCB board defect detection method and system based on CT three-dimensional reconstruction in order to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: PCB board defect detection methods based on CT 3D reconstruction include: The corrected projection data sequence, per-view signal-to-noise ratio parameters, per-view penetration path length and scanning geometry calibration parameters of sparsely scanned PCB boards are obtained. The pre-reconstructed volume data is reconstructed from the corrected projection data sequence and scanning geometric calibration parameters, and the set of layer interface locations is identified, volumetric layer partition markers and material attenuation reference values ​​are generated from it. Iterative reconstruction is performed using pre-reconstructed volume data. The fidelity weight of the data is set according to the signal-to-noise ratio parameter and the penetration path length of the data per view. Intra-layer homogeneity constraints are applied according to the voxel layer partitioning labels. In each round, the prior violation degree is calculated by the projection residual component of the corrected projection data sequence and the intra-layer deviation of the voxel layer partitioning labels. The larger the value, the smaller the voxel exemption weight. Prior exception region labels are generated for those exceeding the exception judgment level, and iterative intermediate volume data is output. Clustering prior exception regions are marked as three-dimensional voxel clusters, i.e. defect candidate regions. Three-dimensional morphological parameter sets are extracted from the intermediate data of the iteration round by round, and the comprehensive change of morphological parameters is calculated. Stable regions are granted exemption confirmation instructions and output three-dimensional morphological parameter sets, while fluctuating regions are granted constraint recovery instructions. The defect category is determined by the three-dimensional morphological parameter set, material attenuation benchmark value and layer interface location set, and the three-dimensional location and size are output in advance; the re-converging area after artifact identification and the marked area without prior exception are treated as normal areas and structural constraints are added to accelerate convergence. After termination, the inspection report is output.

[0007] Preferably, the sparse scanning PCB board includes: The PCB board is clamped with its surface parallel to the rotation axis of the turntable. Sampling angles are extracted at equal angular intervals in the full circumference for projection acquisition. The signal-to-noise ratio parameters for each angle are determined by the noise statistics of the air region in the corrected projection data sequence and the signal statistics of the PCB board projection coverage area, which are used to characterize the reliability of the projection data of each angle. The penetration path length for each angle is determined by the intersection of the initial equation of the board plane and the central ray of each sampling angle. Based on the penetration path length per view, the sampling view is divided into the normal high-quality view range, the intermediate transition view range, and the grazing-limited view range. For the sampling view that meets the preset signal-to-noise requirements, multi-frame acquisition of the same view is performed, and the average of the multi-frame projection is used as the source of the corrected projection data for the corresponding view.

[0008] Preferably, the step of identifying the set of layer interface locations from the pre-reconstructed volume data and generating voxel-by-voxel layer partition markers includes: The pre-reconstructed volume data is resampled to the plate coordinate system. Sampling columns are arranged throughout the plate surface, running through the plate thickness direction. The median value of each sampling column is taken for each plate thickness coordinate to obtain the plate thickness attenuation profile. The peak positions in the rate of change sequence of the plate thickness attenuation profile that exceed the adaptive rate of change level are taken as the layer interface position set. Based on the layer interface position set, the volume data is divided into layer partitions on a voxel-by-voxel basis, and voxel-by-voxel layer partition labels are generated. Interface transition labels are added to voxels whose distance from the layer interface is less than the width of the interface transition band.

[0009] Preferably, the estimated material attenuation benchmark value includes: The values ​​of the high-level plateaus in the thickness-direction attenuation profile are sorted by amplitude and divided at the point where the difference between adjacent values ​​is the largest. The high amplitude class corresponds to the attenuation reference value of solder material, the medium amplitude class corresponds to the attenuation reference value of copper material, and the low plateau value corresponds to the attenuation reference value of resin dielectric material. Each material attenuation reference value is accompanied by a tolerance band, which is determined according to the dispersion of the values ​​within the corresponding plateau.

[0010] Preferably, the iterative reconstruction solves for the refined reconstructed volume data by minimizing the objective function, which is: ; in, The data represents the refined reconstructed volume obtained by minimizing the objective function; Represents the current volume data during the iteration process; Represents the total number of actual sampled viewpoints; Representing the Per-view data fidelity weights for each sampling view; Representing the The system projection matrix corresponding to each sampling viewpoint; Representing the Corrected projection data for each sampling viewpoint; Represents the overall strength coefficient of the structural constraints; Representative voxels Voxel exemption weights at the location; Representative action on voxels Structural constraint functionals; Intralayer homogeneous constraints are constructed using the following formula: ; in, Representative voxels The set of voxels in the spatial neighborhood; Representing neighborhood voxels Volumetric layer partitioning markers; Representative voxels Volumetric layer partitioning markers; Representing neighborhood voxels The current attenuation reconstruction value; Representative voxels The current attenuation reconstruction value.

[0011] Preferably, the prior violation degree and voxel exemption weight are determined according to the following calculation chain: First, calculate the prior violation of the voxel: ; in, Representative voxels a priori violation degree; Representative voxels The projected residual components; Representative voxels Intra-layer deviation; The weighting coefficients representing the projected residual components; The weighting coefficients representing the deviation within a layer. and The sum is 1; Then, the prior violation degree Substituting into the following formula, the voxel exemption weights are obtained through mapping. ; in, Representative voxels Voxel exemption weights at the location; Represents the baseline level of violation, which is a preset multiple of the median of the prior violation values ​​of all voxels in the current iteration.

[0012] Preferably, each iteration of the iterative reconstruction solves the objective function using an alternating optimization approach: First, perform a perspective-by-perspective algebraic updates on each sampled perspective in order of perspective index, correcting the direction of the solution set of the projection equation of the current volume data onto each perspective, and scaling the correction step size according to the perspective-by-perspective data fidelity weight of the corresponding sampled perspective; then perform gradient correction of the structural constraint terms, and scale the correction amount of each voxel according to the voxel exemption weight of each voxel.

[0013] Preferably, the three-dimensional morphological parameter set includes the defect candidate region volume, equivalent diameter, sphericity, anisotropy ratio in the principal direction, span in the plate thickness direction, and attenuation contrast. The overall change in morphological parameters is calculated using the following formula: ; in, Representing the The overall change in morphological parameters across monitoring rounds; This represents the total number of parameters included in the 3D morphology parameter group; Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The minimum measurement standard for the morphological parameters of an item; The morphological parameters of the defect candidate region are determined to be stable if the total change in the morphological parameters is below the stable level for a consecutive preset number of rounds, and unstable if the change is continuously above the fluctuation level.

[0014] Preferably, the defect category is determined as follows: Based on the attenuation contrast sign, the candidate defect region is divided into a low attenuation branch and a high attenuation branch. In the low attenuation branch, those with sphericity higher than the level of dense agglomerates are identified as void defects, those with sphericity lower than the level of lamellar extension and a thickness span not greater than the distance between adjacent layer interfaces are identified as delamination defects, and those with anisotropy ratio in the main direction higher than the level of strip extension and a thickness span crossing at least one layer interface in the set of layer interface locations are identified as crack defects. In the high attenuation branch, defects whose attenuation values ​​are compatible with the material attenuation reference values ​​of copper or solder and whose three-dimensional voxel clusters cross the insulation spacing of adjacent traces are identified as bridging defects. For void defects located in the solder joint area, the void rate is output as the ratio of the volume of the defect candidate area to the volume of the outer contour area of ​​the solder joint. The outer contour area of ​​the solder joint is the set of connected voxels of the solder joint determined by the voxel layer partitioning mark and the material attenuation reference value, along with the void voxels inside it. The method is applied to the online inspection of PCB boards. The whole board inspection report records the total number of sampling perspectives, the total number of iterations, the number of iterations at which each defect candidate area stabilizes, and the number of batches of the normal area confirmation list feedback.

[0015] A PCB board defect detection system based on CT 3D reconstruction includes: Module 1: Obtain the corrected projection data sequence, per-view signal-to-noise ratio parameters, per-view penetration path length, and scanning geometric calibration parameters for sparse scanning of PCB boards; Module 2 reconstructs pre-reconstructed volume data from the corrected projection data sequence and scanning geometric calibration parameters, identifying the set of layer interface locations, generating volumetric pixel partition markers and material attenuation reference values. Module 3 uses pre-reconstructed volume data for iterative reconstruction. It sets the fidelity weight of the data per view according to the signal-to-noise ratio parameters and the penetration path length per view, and applies intra-layer homogeneity constraints according to the voxel layer partitioning labels. In each round, the prior violation is calculated by the projection residual component of the corrected projection data sequence and the intra-layer deviation of the voxel layer partitioning labels. The larger the value, the smaller the voxel exemption weight. Those exceeding the exception judgment level generate prior exception region labels and output iterative intermediate volume data. Module 4: Clustering prior exception regions are marked as three-dimensional voxel clusters, i.e. defect candidate regions. The three-dimensional morphological parameter set is extracted from the intermediate data of the iteration round by round, and the comprehensive change of morphological parameters is calculated. Stable regions are given back exemption confirmation instructions and output three-dimensional morphological parameter sets, while fluctuating regions are given back constraint recovery instructions. Module 5 determines the defect type and outputs the three-dimensional location and size in advance based on the three-dimensional morphological parameter set, material attenuation benchmark value and layer interface location set; it also adds structural constraints to the re-converged area after artifact identification and the area marked with no prior exception as normal area to accelerate convergence, and outputs the inspection report after termination.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention differentiates the projection data of each viewpoint by weighting it according to the signal-to-noise ratio parameters and the penetration path length of each viewpoint, and applies homogeneous constraints only within the same structural layer by marking the voxel layer partitions. At the same time, in each reconstruction iteration, the prior violation of each voxel is calculated by the projection residual components and the deviation within the layer, so that the strength of the structural constraint continuously weakens as the prior violation increases. This ensures that the sparse artifacts in the normal structural region are fully suppressed by the structural constraint, and also prevents the defect signals in the suspected defect region from being smoothed out by the constraint mechanism.

[0017] 2. This invention extracts three-dimensional morphological parameter sets in three-dimensional voxel clusters formed by clustering prior exceptional regions in round by round, and performs cross-iteration stability verification on each defect candidate region based on the comprehensive change of morphological parameters. Regions with stable morphology are classified in advance and their three-dimensional position and size are output, while regions with continuously fluctuating morphology are identified as artifacts and their constraint suppression is restored. Attached Figure Description

[0018] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0020] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0021] Example 1

[0022] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0023] In this embodiment, it includes: Step 1: Acquisition of multi-view projection scan data: Processing target and data interface: The goal of this step is to acquire a multi-view tomographic projection data sequence covering all thickness directions of the PCB board under test. During the acquisition process, the per-view signal-to-noise characteristics and per-view penetration path length of each view are extracted simultaneously. This ensures that the subsequent reconstruction process not only has the projection data itself, but also understands the reliability and geometric properties of each projection data set, thus providing a direct basis for the per-view data fidelity weighting in step three.

[0024] The inputs for this step are the clamping completion status of the PCB board under test and the readiness status of the scanning device. The output includes four types of data: One is to correct the projection data sequence; Secondly, it consists of a set of sampled viewpoints that correspond one-to-one with the corrected projection data sequence; Thirdly, it is a set of signal-to-noise characteristic parameters for each viewpoint; Fourthly, it is the set of scanning geometry calibration parameters.

[0025] Of the four types of outputs mentioned above, the corrected projection data sequence and the scanning geometric calibration parameters are simultaneously sent to steps two and three, the per-view signal-to-noise characteristic parameters are sent to step three, and the sampled view set is sent to steps two and three as the geometric reference for reconstruction.

[0026] Geometric configuration of the scanning device: The scanning device adopts a cone-beam CT geometry consisting of a microfocus X-ray source and a flat panel array detector. The PCB board under test is fixed on a precision turntable by a fixture.

[0027] The clamping attitude is set as follows: The PCB board surface is parallel to the rotation axis of the turntable, that is, the PCB board is clamped in a vertical and side-mounted manner, so that during one rotation of the turntable, the ray beam can pass through the normal incident posture and the grazing incident posture of the board surface in turn.

[0028] The reason for using vertical side clamping is: PCB boards are flat objects. When a ray is incident along the normal direction of the board surface, the penetration path of the ray within the board is the shortest, resulting in moderate attenuation of the projected signal and a good signal-to-noise ratio. However, when the ray is incident along the direction of the board surface, the penetration path can be tens of times longer than that in the case of the normal direction, leading to severe attenuation or even failure of the projected signal. Vertical side clamping ensures that all sampling angles are symmetrically distributed about the normal direction of the board surface, facilitating subsequent classification and management of angles based on the penetration path length of each angle.

[0029] The focal spot size of the X-ray source is selected based on the minimum target size of the defect to be inspected. The focal spot size should not be greater than half of the minimum target defect size to ensure that the defect projection after geometric magnification still has resolvable sharpness. The geometric magnification factor is determined by the ratio of the source-object distance to the source-detector distance. The magnification factor is selected so that the projection range of the PCB board on the detector occupies two-thirds to four-fifths of the detector's effective imaging surface, taking into account both resolution and field of view integrity.

[0030] Sparse sampling perspective planning: To meet the cycle time requirements of industrial online inspection, this step employs a sparse angle sampling method to acquire projection data. Assuming the number of viewing angles required for conventional full-circumference sampling is the baseline number of viewing angles, this step uniformly extracts actual sampling viewing angles from the full circumference according to the extraction ratio. The actual number of sampling viewing angles... The ratio of the reference angle number to the reference angle number is controlled between one-fifth and one-half.

[0031] The rules for determining the total number of actual sampling angles are as follows: when the total copper layer thickness is no greater than 0.5 mm and the minimum target defect size is no less than 100 micrometers, the sampling ratio is one-fifth; when the total copper layer thickness is greater than 0.5 mm but no greater than 1.5 mm, or the minimum target defect size is no less than 50 micrometers but less than 100 micrometers, the sampling ratio is one-third; when the total copper layer thickness is greater than 1.5 mm, or the minimum target defect size is less than 50 micrometers, the sampling ratio is one-half. The total copper layer thickness is obtained from the process documents of the PCB board under test or from actual measurement of the board edge slices, and the minimum target defect size is specified by the inspection requirements. The sampling angles are distributed at equal angular intervals throughout the circumference, and the angular interval between adjacent sampling angles is... .

[0032] The angular interval and the total number of viewing angles satisfy the following conditions. The relationship between the viewpoint indices. The set of viewpoint indices is denoted as... Each viewpoint corresponds to a projection angle. .

[0033] The reconstruction artifacts caused by incomplete projection data due to sparse sampling are compensated by the structural prior constraint mechanism in step three. This step only needs to ensure the uniformity of the sampling viewpoint distribution and the accurate recording of the viewpoint position.

[0034] Projection acquisition execution and preprocessing: During acquisition, the X-ray source tube voltage is set based on the penetration requirements of the thickest copper layer on the PCB board. The product of the tube current and the single-view exposure time is set according to the utilization ratio of the detector's full-well capacity, ensuring that the dynamic range utilization of the projected image is in the middle of the detector's linear response range, avoiding quantization loss in weak signal areas and saturation cutoff in strong signal areas. For critical viewing angles with high signal-to-noise requirements, the projection signal-to-noise level can be improved by averaging multiple frames acquired from the same viewing angle. The default number of frames for multi-frame acquisition is 4.

[0035] The original projection images acquired from each viewpoint are sequentially subjected to dark field correction, gain correction, and bad pixel repair: Dark field correction uses the dark field image pre-stored before acquisition to subtract from the original projection; gain correction uses the flat field image of the empty field scan to normalize the projection after subtracting the dark field. Bad pixel repair is performed based on the bad pixel table calibrated at the detector's factory, replacing the bad pixel value with the median of the neighboring valid pixels. After the above correction is completed, a logarithmic transformation is performed on the air transmittance projection to obtain corrected projection data that is proportional to the linear integral of the ray attenuation. .

[0036] in, Representing the Detector coordinates at each sampling perspective The corrected projection data values ​​at the location; The row coordinates of the detector pixels; The column coordinates represent the detector pixels.

[0037] Per-view signal-to-noise characteristic parameter extraction: After the projection data preprocessing is completed, the per-view signal-to-noise ratio (SNR) parameters are extracted for the projection image at each viewpoint. The extraction method is as follows: An air region not covering the PCB board projection area is selected in the projection image. The air region is a rectangular area at the four corners of the projection image excluding the board projection. The mean and standard deviation of the pixel projection values ​​within the air region are calculated. The standard deviation of the air region projection values ​​represents the noise level of the viewpoint. Simultaneously, the distribution range of projection values ​​within the PCB board projection coverage area is statistically analyzed, and the per-view SNR parameters are calculated using the following formula: ; in, Representing the Per-view signal-to-noise characteristics parameters for each sampling viewpoint; Representing the The arithmetic mean of all pixel values ​​in the PCB board projection coverage area of ​​the sampled viewpoint correction projection data. The coverage area is defined by the set of pixels in the projection image whose projection values ​​are significantly higher than three times the standard deviation of the air area mean. Representing the Each sampling perspective corrects the sample standard deviation of the projection data for all pixels in the air region.

[0038] The higher the value of the per-view signal-to-noise ratio parameter, the more reliable the projection data characterizing the viewpoint. For viewpoints adjacent to the PCB board surface in the normal direction, the ray penetration path is shorter. The value is relatively large; for grazing angles near the plate surface, the ray penetration path is long and the attenuation is severe. The value is relatively small.

[0039] Per-view signal-to-noise characteristic parameter set The data is output along with the corrected projection data sequence and used as the basis for setting the fidelity weights for each viewpoint in step three. This ensures that the fidelity weights change monotonically with the viewpoint signal-to-noise level, and that projection data from inferior viewpoints do not dominate the reconstruction results.

[0040] Per-view penetration path length calculation: Based on the design geometry of the fixture and the zero-position calibration results of the turntable, the initial spatial equation of the PCB board plane is established, and the per-view penetration path length of the central ray of the beam through the PCB board body is calculated at each sampling angle. The central ray is defined as the straight line connecting the focal point of the ray source and the geometric center of the detector's imaging plane.

[0041] Representing the The length of the penetration path of the central ray within the PCB board under each sampling perspective.

[0042] The method for calculating the penetration path length per viewpoint is as follows: based on the initial equation of the plate plane and the central ray direction vector corresponding to the viewpoint, the central ray direction vector is obtained by subtracting the coordinates of the ray source focus and the geometric center coordinates of the detector imaging surface and normalizing them. Both coordinates are taken from the scanning geometric calibration parameters. The two intersection points of the central ray with the upper and lower surfaces of the plate are obtained, and the spatial distance between the two intersection points is the penetration path length per viewpoint.

[0043] The per-view penetration path length serves two purposes: first, it acts as a basis for viewpoint classification, dividing all sampled views into three categories: normal high-quality viewpoint range, intermediate transition viewpoint range, and grazing-limited viewpoint range. The classification results are used as a reference in step three when weighting the data for fidelity. Second, it serves as the initial value for the board attitude estimation in step two, shortening the search range for refining the board plane equation in step two. The accurate calculation of the per-view penetration path length is updated once with the estimation result after the board attitude estimation is completed in step two. The updated accurate value is fed back to step three to replace the initial value.

[0044] Geometric calibration and output data list: The scanning geometric calibration parameters include the distance from the X-ray source focal point to the turntable's rotation axis, the distance from the X-ray source focal point to the detector's imaging surface, the physical size of the detector pixels, the offset of the projection position of the rotation axis on the detector's imaging surface, and the zero-position deviation of the turntable's angular position encoder. These parameters are obtained through scanning a calibration phantom before acquisition. The calibration phantom is an array of spheres with known precise dimensions. The geometric parameters are inversely solved by the fitting residuals between the projected centroid trajectories of the spheres and the theoretical trajectories.

[0045] The final output list for this step is as follows: Corrected projection data sequence Sampling perspective set A set of signal-to-noise characteristic parameters per view Set of path lengths for each perspective In addition, a set of scanning geometry calibration parameters is provided. The above data is output through a unified data interface, in which each set of corrected projection data is accompanied by its viewpoint index, ensuring that the one-to-one correspondence between data and viewpoint is not lost when called in subsequent steps.

[0046] Step 2, Structural Prior Modeling and Pre-Reconstruction: Processing target and data interface: The goal of this step is to use the sparse sampling correction projection data output from step one to quickly obtain pre-reconstructed volume data with low computational overhead, and automatically extract the layered structure prior of the PCB board from the pre-reconstructed volume data to assemble a structural prior model that can be directly called in step three.

[0047] The structural prior model is the core data foundation of this method, and its contents include: PCB board coordinate system, layer interface location set, number of layers, voxel-by-voxel partition markers, and material attenuation benchmark values. The input of this step is the corrected projection data sequence, sampling viewpoint set, per-viewpoint penetration path length, and scanning geometric calibration parameters output from step one; the output is the pre-reconstructed volume data and all components of the structural prior model. Among them, the pre-reconstructed volume data serves as the initial value for the iterative reconstruction in step three, the voxel-by-voxel partition markers and material attenuation benchmark values ​​serve as the basis for applying structural constraints in step three, and the layer interface location set and voxel-by-voxel partition markers are simultaneously transmitted to step four as spatial guidance for defect candidate regions.

[0048] Projection downsampling and fast pre-reconstruction: The goal of rapid pre-reconstruction is not to obtain high-precision reconstruction results, but to obtain low-resolution volumetric data capable of resolving the PCB board's layer stack-up structure in the shortest possible time. To this end, the corrected projection data output from step one is first downsampled: adjacent pixels are merged at a fixed ratio along the detector pixel dimension, with a default merging ratio of 2×2 pixels. After merging, the spatial resolution of the projection data is reduced, and the data volume is reduced by the square of the merging ratio. The downsampled projection data is then input into the rapid analytical reconstruction process, with the voxel size of the reconstructed volumetric data set to 2 to 4 times the size of the final fine reconstruction target voxel.

[0049] Due to the sparse sampling perspective, strip-shaped sparse artifacts are inevitable in the analytical reconstruction results. However, the attenuation difference along the thickness direction of the PCB board's stack-up structure is significant, and the layer interfaces can still be identified despite artifact interference. The time consumption of fast pre-reconstruction is controlled to be less than one-tenth of the total time consumption of fine reconstruction in step three. The pre-reconstructed volume data is denoted as... , This represents the pre-reconstructed volume data obtained through rapid pre-reconstruction.

[0050] Plate attitude estimation and plate coordinate system establishment: Using the initial equation of the plate plane corresponding to the per-view penetration path length output in step one as the search starting point, in the pre-reconstructed volume data Precise estimation of the attitude of the execution board.

[0051] The precise estimation method is as follows: The distribution of high-attenuation voxels in the statistical volume data layer by layer along the candidate board plane normal is analyzed. When the candidate board plane is parallel to the real board plane, the high-attenuation voxels (corresponding to the copper layer) exhibit the highest spatial aggregation in the two-dimensional slice parallel to the candidate board plane. The board plane normal and position are searched based on the maximum aggregation degree to obtain the accurate board plane equation.

[0052] After precise estimation and convergence, establish a plate coordinate system: the origin of the coordinate system is located at the geometric center of the plate plane. The axis is along the normal direction of the plate plane, i.e., the plate thickness direction. shaft and The axes lie within the board plane and are parallel to the main routing directions of the PCB design. After establishing the board coordinate system, the pre-reconstructed volume data is resampled to the board coordinate system, ensuring that the board thickness direction aligns with the volume data in all subsequent processing. Dimensions are strictly aligned, and the layer interface is represented in the volume data as follows The plane has approximately constant coordinates. The plate attitude estimation results are synchronously fed back to the view-by-view penetration path length update process described in step one.

[0053] Thickness-direction attenuation profile extraction: In the pre-reconstructed volume data in the plate coordinate system, the attenuation profile along the plate thickness direction is extracted. Specifically, within the plate plane, a number of sampling columns are selected using a uniform grid, with a minimum of 100 columns. If the area of ​​a single column (the plate area divided by 100) is greater than the projected area of ​​the minimum target defect within the plate surface, columns are selected using a finer grid until the area of ​​a single column is no greater than the projected area of ​​the minimum target defect. Each sampling column is along... A slender column of voxels running through the plate, recording the attenuation coefficient along each sampling column. The sequence of coordinate values; merging the value sequences of all sampled columns into the same... Taking the median at the coordinates, we obtain the attenuation profile in the thickness direction of the plate. .

[0054] in, Represents the thickness direction of the plate Median value of the attenuation profile in the thickness direction at the coordinate point; The coordinates along the thickness direction of the plate in the plate coordinate system.

[0055] The reason for using the median instead of the mean is that the sampling column may pass through local high-attenuation structures such as solder joints, vias, and wiring, or local low-attenuation structures such as defects and voids. Median statistics are not sensitive to these local anomalies. The resulting profile represents the typical attenuation variation along the thickness direction of the layered substrate, rather than the accidental fluctuations of local structures. It presents a clear alternation of high and low positions, with each high platform corresponding to a copper layer or solder layer, and each low platform corresponding to a resin dielectric layer.

[0056] Layer interface identification and layer number determination: Attenuation profile in the thickness direction Calculate along The rate of change sequence is calculated by averaging forward and backward differences to suppress one-sided errors. Local maxima in the rate of change sequence correspond to sharp rising or falling edges of the profile, i.e., the physical interface between the copper and dielectric layers. The interface identification criterion is: values ​​in the rate of change sequence exceeding the adaptive rate of change level. And the spacing is not less than the minimum layer thickness constraint. The peak position was identified as the layer interface position.

[0057] in, The rate of change in the representation layer interface recognition is adaptively leveled. This represents the minimum layer thickness constraint.

[0058] Rate of change adaptive level The method for determining this is as follows: Three times the median of the absolute values ​​of the rate of change sequence is taken as the default value; the median is statistically robust to profile noise, and a three-fold level can suppress spurious peaks caused by noise. Minimum layer thickness constraint. The determination method is as follows: take the smaller value between the thinnest copper layer thickness and the thinnest dielectric layer thickness specified in the manufacturing process document of the PCB board being tested; if the process document cannot be obtained, use the default value of 12 micrometers. Confirm all interfaces by clicking... The coordinates are arranged in ascending order to form the set of interface positions of the layer. number of floors It is determined by adding 1 to the number of interfaces.

[0059] in, Representing the Each interface in the plate coordinate system coordinate; This represents the total number of PCB board structural layers identified.

[0060] Perform a reasonableness check on the identification results: check whether the number of layers is within the range achievable by the PCB board process, check whether the spacing between adjacent interfaces is not less than the minimum layer thickness constraint, and check whether the attenuation level of the high-level platform of the profile is compatible with the copper material attenuation benchmark value. If the check fails, the adaptive rate of change level is reduced by 0.5 times each time and the identification is re-identified. The lower limit of the reduction is 1.5 times. If the check still fails after reaching the lower limit, a manual review is prompted.

[0061] Voxel layer partitioning marker generation: Once the set of layer interface locations is determined, a voxel-by-voxel layer partition label is generated for each voxel in the volume data. The labeling rule is: voxel... The coordinates fall at the th The interface and the first When there are multiple interfaces, the voxel's per-voxel layer partitioning marker value is taken as follows: .for Voxels whose coordinates are less than the width of the interface transition zone at a given layer are assigned an interface transition label in addition to a per-voxel layer partition label, indicating that the voxel is located in the interlayer transition region. The per-voxel layer partition label is denoted as... .

[0062] in, Representative voxels The volumetric partitioning markers range from 1 to [value missing]. Integers; Voxel indexes in the representative volume data.

[0063] The width of the interface transition zone is set based on the spatial resolution and interface recognition accuracy of the pre-reconstructed volume data, ranging from 1 to 3 times the voxel size, with a default of 2 times. Interface transition markers are used for interface protection in step three: voxels in the transition zone do not participate in intra-layer homogeneous constraints, preventing the constraints from smoothing out the actual inter-layer steps and also preventing excessive smoothing of the interface regions where delamination defects are located. The generation of voxel-by-voxel layer partition markers is completed in the plate coordinate system. When the resolution of the fine-reconstructed volume data differs from that of the pre-reconstructed volume data, the voxel-by-voxel layer partition markers are mapped to the fine-reconstructed mesh according to the correspondence of voxel spatial positions.

[0064] Material attenuation baseline estimation: Attenuation profile based on plate thickness Based on the voxel value distribution within each layer partition, the material attenuation baseline value of the materials contained in the PCB board is estimated. The estimation method is as follows: all values ​​of the high-level platform of the profile are sorted in ascending order of amplitude, and the point with the largest difference between adjacent values ​​is taken as the dividing point. The high-level platform values ​​are divided into a one-dimensional segmentation of high amplitude class and medium amplitude class. The high amplitude class corresponds to the material attenuation baseline value of the solder material. The medium amplitude category corresponds to the material attenuation benchmark value of copper materials. The value of the low-level platform in the profile corresponds to the material attenuation reference value of the resin dielectric material. .

[0065] Each material attenuation benchmark value is accompanied by a tolerance band. The tolerance band is determined by: statistically analyzing the sample standard deviation of all values ​​within the corresponding category platform, taking twice the sample standard deviation as the half-width of the tolerance band, and using the mean of the values ​​within the category platform as the material attenuation benchmark value. The tolerance band is used to characterize the confidence range of the material attenuation benchmark value.

[0066] The material attenuation reference value has two uses: Firstly, in step three, the deviation between the reconstructed values ​​of voxels within the layer and the attenuation benchmark value of the dominant material in the layer, which serves as the reference level for homogeneous constraints within the layer in the structural prior model, is used in the calculation of the prior violation degree. Secondly, the material classification criteria used in step five for defect category determination, for example, defect candidate areas located in the solder joint area and whose attenuation is significantly lower than the solder material attenuation benchmark value, tend to be classified as solder voids.

[0067] Structural prior model assembly and output: The final assembled prior structural model in this step includes five elements: plate coordinate system parameters, layer interface location set, etc. Number of floors Voxel layer partitioning labeling The model includes interface transition markers, material attenuation baseline values, and their tolerance bands. The core assumptions underlying the structural prior model are: the PCB board exhibits approximate homogeneity within layers, meaning that the voxel attenuation coefficients within the same layer partition tend to be similar in the absence of defects; and the PCB board exhibits discontinuity between layers, meaning that a step change in the voxel attenuation coefficient is allowed at cross-layer interfaces. These two assumptions together constitute the physical basis for the structural constraints in step three. Defects, however, are precisely the local disruptions to the approximate homogeneity within layers. This contradiction between disruption and constraint is addressed by the violation-driven exemption mechanism in step three.

[0068] The output data list for this step is: Pre-reconstructed volume data. The parameters include plate coordinate system parameters, layer interface location set, number of layers, volumetric layer partition markers and interface transition markers, material attenuation baseline values, and tolerance bands. The pre-reconstructed volume data also serves as the initial volume data for the iterative reconstruction in step three, avoiding starting the iteration from zero initial values ​​and thus increasing the number of convergence rounds.

[0069] Step 3, Violation-driven exemption constraint reconstruction: Processing target and data interface: The goal of this step is to apply structural constraints to the iterative reconstruction using the structural prior model assembled in step two, under the condition that the sparse sampling correction projection data is incomplete, in order to suppress sparse artifacts. At the same time, the constraint strength is dynamically allocated by calculating the prior violation degree per voxel, so that the constraints only act on normal areas that conform to the layered structure law, while areas suspected of destroying the layered structure are exempted step by step, so as to avoid the real defects being smoothed out by the constraint mechanism.

[0070] The inputs for this step include: the corrected projection data sequence, sampling view set, per-view signal-to-noise characteristic parameters, per-view penetration path length and scanning geometry calibration parameters output from step one; and the pre-reconstructed volume data, plate coordinate system parameters, layer interface location set, per-volumetric layer partition markers and interface transition markers, material attenuation reference values ​​and tolerance bands output from step two. After each iteration, this step outputs the current iteration intermediate data, the voxel-by-voxel prior violation distribution, and prior exception region markers to step four. It also receives the exemption adjustment instructions from step four and the normal region confirmation information from step five. The iteration continues until all defect candidate regions have been determined by step four and the reconstruction convergence condition is met, at which point the process terminates, and finally, the refined reconstructed volume data is output.

[0071] The overall structure of the iterative reconstruction objective function: Iterative reconstruction solves for the refined reconstructed volume data by minimizing the following objective function: ; in, The data represents the refined reconstructed volume obtained by minimizing the objective function; Represents the current volume data during the iteration process; Represents the total number of actual sampled viewpoints; Representing the Per-view data fidelity weights for each sampling view; Representing the The system projection matrix corresponding to the sampling viewpoint, where each element represents the contribution of each voxel to the attenuation line integral of the projection ray, is calculated as follows: For the ... For each detector ray from each sampling angle, the equation of the straight line from the ray source focus to the detector pixel is calculated based on the scanning geometry calibration parameters. The intersection length of the line through each voxel is obtained for each voxel. The intersection length is the value of the corresponding matrix element. The element corresponding to the voxel where the rays do not intersect is 0. Representing the The corrected projection data from each sampling viewpoint participates in the calculation as a vector in the objective function, and this vector is defined in step one. The values ​​of all pixels are arranged row by row according to the detector coordinates; Represents the overall strength coefficient of the structural constraints; Representative voxels The voxel exemption weight at the location, with a value ranging from 0 to 1; Representative action on voxels Structural constraint functionals.

[0072] The first term of the objective function is a view-by-view weighted data fidelity term, ensuring consistency between the reconstruction result and the actual measured projection; the second term is a voxel exemption weight modulated structural constraint term, ensuring that the reconstruction result conforms to the prior knowledge of the PCB board's layered structure within the undisputed region. The relative weights between the two types of terms are determined by the overall structural constraint strength coefficient. balance.

[0073] The relationship is set as follows: ; in, This represents the baseline strength coefficient, which is set to 0.2 by default. During implementation, it is fine-tuned based on the reconstruction effect of the first sample. It is adjusted upwards when there are obvious residual sparse artifacts and downwards when the defect edges are over-sharpened and damaged. The reference viewing angle number is determined by the sampling theorem based on the detector pixel count and the PCB board projection width, i.e., the effective pixel count of the detector in the board surface direction multiplied by... Round up; This represents the total number of actual sampling perspectives.

[0074] Setting the per-view data fidelity weights: The per-view data fidelity weights are set based on the per-view signal-to-noise ratio parameters output in step one, and the setting relationship is as follows: ; in, Representing the Per-view signal-to-noise characteristics parameters for each sampling viewpoint; This represents the maximum value among the signal-to-noise ratio parameters for all sampled viewpoints.

[0075] After normalization, the per-view data fidelity weight for high-quality normal views is close to 1, while the per-view data fidelity weight for grazing-limited views is significantly less than 1. For grazing-limited views where the per-view penetration path length exceeds 10 times the plate thickness and the per-view signal-to-noise ratio is lower than the median value of all views, the per-view data fidelity weight is multiplied by an attenuation coefficient based on the normalization result. The attenuation coefficient ranges from 0.1 to 0.5, with a default value of 0.3. Through per-view differential weighting, the data error of inferior views is limited to a secondary position in the fidelity term and will not be amplified into structural artifacts during the iteration process.

[0076] Construction of structurally constrained functionals: The structural constraint functional is constructed as an intralayer homogeneous constraint based on the volumetric layer partitioning label: ; in, Representative voxels The set of spatial neighborhood voxels, where the neighborhood can be one of 6, 18 or 26 neighborhoods, with 6 neighborhoods as the default. Representing neighborhood voxels Volumetric layer partitioning markers; Representative voxels Volumetric layer partitioning markers; Representing neighborhood voxels The current attenuation reconstruction value; Representative voxels The current attenuation reconstruction value.

[0077] The constraint functional only penalizes neighboring voxel pairs with consistent voxel layer partitioning labels, preventing them from converging. Voxel pairs at cross-layer interfaces are not penalized, thus protecting the interlayer decay step from smoothing effects and preserving the interlayer discontinuity characteristics. Voxels with interface transition labels are neither subject to constraint nor used as neighboring references, keeping interlayer regions with high rates of layer defects in a constraint-neutral state. The physical meaning of structural constraints is that the material composition within the same structural layer is approximately homogeneous, and the voxel decay values ​​within the layer should tend to be similar in the absence of defects. However, the presence of defects causes local voxel decay values ​​to deviate from the average level within the layer. If the constraint mechanism is applied indiscriminately, it will smooth out the deviation step by step. The violation-driven exemption mechanism is designed to address this contradiction.

[0078] Calculation of voxel-by-voxel prior violation: In each iteration, the prior violation is calculated on a voxel-by-voxel basis: ; in, Representative voxels The higher the value of the prior violation, the more likely the voxel is to be a real defect rather than a reconstruction underconvergence or noise. Representative voxels The projected residual components; Representative voxels Intra-layer deviation; The weighting coefficients representing the projected residual components; The weighting coefficients representing the deviation within a layer. and The sum is 1, and each is 0.5 by default.

[0079] The projection residual components are obtained as follows: An orthographic projection operation is performed on the current volume data to obtain a simulated projection. The difference between the simulated projection and the corrected projection data is calculated view-by-view to obtain the projection residual. The projection residuals of each view are then weighted according to the view-by-view data fidelity weights and back-projected onto the volume data space, and voxels are taken. The back-projection value at a given location, after normalization by the maximum absolute value of the back-projection values ​​of all voxels, is used as... The projection residual component characterizes the degree of discrepancy between the current voxel value and the actual measurement data, with the discrepancy concentrated at locations where the true structure has not been correctly reconstructed.

[0080] The method for obtaining the intra-layer deviation is as follows: Get voxels The median of the current values ​​of all voxels with the same label within the layer partition is used as the intra-layer reference level. The absolute value of the difference between the voxel value and the intra-layer reference level is calculated, and then normalized by the tolerance band width of the material attenuation reference value for that layer. .

[0081] The intralayer deviation characterizes the degree to which voxel values ​​violate the intralayer homogeneity assumption. The two components measure the degree of doubt of voxels from two independent aspects: the consistency of measurement data and the consistency of prior structural knowledge. A high value for either component increases the degree of violation, and a high value for both components simultaneously increases the degree of violation significantly.

[0082] Voxel exemption weights and prior exception region markings: The actual strength of the structural constraint applied to each voxel is the product of the overall strength coefficient of the structural constraint and the voxel exemption weight. The voxel exemption weight is obtained by mapping the prior violation degree through a decay function. ; in, The benchmark level for violation is determined as follows: take 1 times the median of the distribution of prior violation values ​​of all voxels in the current iteration as the default value, i.e., the median itself. The adjustment range is allowed to be 1 to 2 times. When the dispersion of the prior violation distribution is large, i.e. the standard deviation exceeds half of the median, the adjustment factor is increased.

[0083] The selection of the voxel exemption weighting function causes the constraint strength to continuously decrease as the prior violation increases: when the prior violation is much lower than the baseline level, the voxel exemption weight is close to 1, and the structural constraint is applied in full; when the prior violation is close to the baseline level, the constraint strength drops to about two-thirds to one-half of the full amount; when the prior violation is much higher than the baseline level, the constraint strength approaches 0, and the voxel is only governed by the data fidelity term. When the prior violation exceeds the exception determination level... Generate prior exception region markers on the voxels. The remaining voxels .

[0084] in, This represents the exception handling level, with a value higher than the benchmark level for violation. Default value 2 times; Representative voxels The prior exception region marker, with a value of 1, indicates that the voxel is exempt from structural constraints.

[0085] The separation of the exception judgment level and the constraint strength attenuation curve forms a two-level processing: voxels with moderate to high prior violation are only subject to reduced constraints, while voxels with significantly high prior violation are completely exempted and marked as suspected defect carriers, directly serving as the source of monitoring objects in step four.

[0086] In the first iteration, the voxel exemption rights of all voxels are reset to 1 and the prior exception region marker is set to 0. From the second iteration onwards, the voxel exemption rights are updated round by round according to the above mechanism.

[0087] Iterative update process and response to external commands: Each iteration solves the objective function using an alternating optimization approach. The solution process is as follows: First, a view-by-view algebraic update is performed along the data fidelity terms. The current volume data is corrected sequentially according to the view index, moving towards the solution set of the projection equations for each view. The correction amount is equal to the value of the projection residual after back-projection through the system projection matrix. The correction step size is scaled according to the view-by-view data fidelity weights for that view. Second, gradient correction is performed on the structural constraint terms. The gradient of the structural constraint functional with respect to the voxel value is calculated for each voxel. The correction amount is the gradient multiplied by the step size factor and then by the voxel exemption weight for that voxel. The step size factor is set to 0.3 by default. These two alternating steps constitute a complete iteration. After each iteration completes the volume data update, external data interaction is performed in the following order: orthographic projection and projection residual calculations are performed, updating the voxel-by-voxel projection residual components and intra-layer deviations to obtain a new round of prior violation distribution. Subsequently, the voxel exemption weight distribution and prior exception region markers are updated based on the prior violation distribution. Finally, the intermediate volume data of the current iteration is... Prior violation distribution Prior exception region marker Output to step four.

[0088] in, Represents the iteration round number; Representing the Intermediate iteration data when a round of iterations is completed.

[0089] The exemption adjustment instructions fed back in step four include two types: One is the exemption confirmation instruction. For defect candidate regions where the morphological parameters tend to be stable, the instruction maintains the voxels in the region as fully exempted and no longer participates in the subsequent prior violation recalculation. Processing resources are no longer allocated to such regions. The second is the constraint recovery instruction. For defect candidate regions with continuously fluctuating morphological parameters, the instruction resets the voxel exemption weight of the voxels in the region to 1 and clears the prior exception region marker, so that the structural constraints are reapplied in full to suppress the false deviations caused by under-convergence of the iteration.

[0090] The normal region confirmation information fed back in step five applies to the third location: the confirmed large normal structural region, whose structural constraint strength is within... The normal region is multiplied by an acceleration factor, which ranges from 1.5 to 3, with a default value of 2. This allows the normal region to converge at a faster rate, and iterative computing resources are concentrated on the undetermined defect candidate regions.

[0091] Reconstruction convergence and termination conditions: The reconstruction can only be terminated if two conditions are met simultaneously: First, all defect candidate regions fed back in step four have completed stability determination and the corresponding determination results have been output in step five. There are no defect candidate regions in the monitoring state. Secondly, the relative change in the residual energy of the fidelity term for all prime data over five consecutive iterations is lower than the convergence level, which ranges from [value missing]. to When both conditions are met, the final detailed reconstructed volume data is output and archived as a spatial carrier of the defect detection results for future reference.

[0092] Step 4, cross-iteration stability monitoring of candidate regions: Processing target and data interface: The goal of this step is to extract candidate defect regions and monitor cross-iteration stability by performing the prior exception region markers output in step three round by round. By utilizing the temporal difference that the morphology of real defects tends to stabilize in continuous iterations, while the morphology of false deviations caused by under-convergence of reconstruction continues to fluctuate, the authenticity of the candidate defect regions is verified in the time dimension. The verification results are then fed back to step three to adjust the exemption strategy and to step five to trigger early judgment, so that defect detection and reconstruction iteration can proceed in parallel without waiting for the reconstruction to fully converge.

[0093] The inputs to this step are the intermediate iteration data, prior violation distribution and prior exception region markers output from each iteration of step three, as well as the set of layer interface locations, volumetric layer partition markers and material attenuation benchmark values ​​output from step two. The outputs of this step include the stability confirmation conclusion and the three-dimensional morphological parameter set of the defect candidate region sent to step five, and the exemption confirmation instruction and constraint recovery instruction fed back to step three.

[0094] Generation of monitoring objects: After receiving the output from step three in each round, three-dimensional connectivity clustering is performed on the set of voxels with a priori exception region marker value of 1: under the 26-neighborhood connectivity rule, interconnected exception voxels are aggregated into independent three-dimensional voxel clusters, and each three-dimensional voxel cluster constitutes a defect candidate region. After clustering, size filtering is performed: three-dimensional voxel clusters containing voxels with a number of voxels smaller than the minimum candidate size are removed, and the priori exception region markers of voxels in the removed clusters are retained, but no monitoring files are created in this round; the minimum candidate size is set according to the ratio of the minimum target defect size to the voxel size, and by default, it is taken as half of the number of voxels covered by the volume of the sphere corresponding to the equivalent diameter of the target defect.

[0095] A monitoring profile is established for each surviving 3D voxel cluster. The profile records the voxel list of the 3D voxel cluster, the voxel layer partition marker to which it belongs, whether it carries an interface transition marker, and the 3D morphological parameters added in subsequent rounds. The correspondence between 3D voxel clusters across rounds is established through spatial overlap matching: when the spatial overlap volume between a new round of 3D voxel cluster and an existing archive cluster exceeds half of the larger of the two volumes, it is determined to be a continuation of the same defect candidate region, and the morphological parameters of the new 3D voxel cluster are added to the existing archive; new 3D voxel clusters without matching archives are created as new archives; archives that have not been matched for several consecutive rounds are put into a suspended state, with a default suspension round count of 5 rounds. After 5 rounds of suspension, the constraint recovery command is applied.

[0096] Extraction of 3D morphological parameter sets: For each candidate defect region for which a monitoring file is established, a set of three-dimensional morphological parameters is extracted in each round. The parameter set includes six parameters: candidate defect region volume, equivalent diameter, sphericity, anisotropy ratio in the principal direction, span in the thickness direction, and attenuation contrast. (Candidate defect region volume) The product of the number of voxels contained in a three-dimensional voxel cluster and the physical volume of the voxels; equivalent diameter The diameter of a sphere with the same volume as a three-dimensional voxel cluster; sphericity Calculated according to the following relationship: ; in, The three-dimensional volume representing the candidate defect region; The surface area of ​​the three-dimensional voxel cluster representing the defect candidate region is calculated as follows: traverse all boundary voxels of the three-dimensional voxel cluster. Boundary voxels are cluster voxels with external voxels in their 26 neighborhoods. Count the total number of facets that are adjacent to and in contact with the external voxels of the cluster. Multiply the total number of facets by the area of ​​a single facet to obtain the surface area. The area of ​​a single facet is the square of the voxel size. The sphericity of the defect candidate region is represented by a value closer to 1, indicating that the three-dimensional voxel cluster is closer to a dense mass morphology, while a smaller value indicates that the three-dimensional voxel cluster tends to be more strip-shaped or sheet-like.

[0097] The anisotropy ratio along the principal direction is obtained through second-moment analysis of the voxel coordinates of the three-dimensional voxel cluster: the covariance matrix of the voxel coordinate distribution of the three-dimensional voxel cluster is calculated, and the square root of the ratio of its largest to smallest eigenvalue is taken as the anisotropy ratio. This parameter characterizes the degree of geometric disparity between the three-dimensional voxel cluster along the principal extension direction and the shortest direction. The span along the plate thickness direction is the voxel of the three-dimensional voxel cluster in the plate coordinate system. The range of coordinates along the axis represents the order of magnitude of the penetration of three-dimensional voxel clusters through the structural layers.

[0098] The attenuation contrast is the difference between the average attenuation of voxels within a 3D voxel cluster and the average attenuation of background voxels in the same layer surrounding the cluster. This is normalized by the tolerance band width of the material attenuation reference value in the layer in which the cluster is located. The delineation method for the background voxels in the same layer surrounding the cluster is as follows: A shell region is formed by extending five voxels outward from the circumscribed cuboid of the three-dimensional voxel cluster along each of the three coordinate axes. Voxels within the shell region that have the same per-voxel layer partitioning label as the three-dimensional voxel cluster are selected as background voxels. A negative attenuation contrast value indicates that the attenuation of the three-dimensional voxel cluster is lower than that of the background, corresponding to typical manifestations of voids and layered defects; a positive value with a large amplitude corresponds to high attenuation anomalies of inclusions.

[0099] Cross-iteration stability criterion: For each candidate defect region, the monitoring file records the evolution sequence of its three-dimensional morphological parameters across consecutive iterations, and calculates the comprehensive change in morphological parameters according to the following relationship: ; in, Represents the monitoring round sequence number; This represents the total number of parameters in the three-dimensional morphological parameter group; in this method, it is set to 6. Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The minimum measurement benchmark for the morphological parameters is used to prevent the relative change from diverging when the parameter value is close to zero. Its values ​​for each parameter are: [The text abruptly ends here, so the translation stops as well.] Take the physical volume of a single voxel; the equivalent diameter parameter and the span parameter in the thickness direction of the plate. Voxel size; sphericity parameters, principal direction anisotropy ratio parameters, and attenuation contrast parameters. All values ​​are taken as 0.01; Representing the The overall change in morphological parameters across all monitoring rounds.

[0100] The stability criterion is: when the comprehensive change in the morphological parameters of the defect candidate region is continuous... The number of monitoring rounds was below the stable level. When the morphology of the defect candidate region is stable, it is determined that the overall change is consistently higher than the fluctuation level. At that time, the morphology of the candidate defect region is determined to be unstable.

[0101] in, This represents the number of consecutive monitoring rounds required to confirm stability; the default value is 3. The stability level is defined as follows: In the middle of the iteration, when the number of iterations reaches half of the expected total number of iterations, the median of the absolute value of the single-cycle update of all elements is calculated. The relative change level corresponding to this median is used as the reference value. The stability level is twice the reference value and is limited to the range of 0.02 to 0.05. If twice the reference value exceeds the range, the endpoint value of the range is taken. This represents the level of volatility, and its value is higher than the stable level. The default value is 3 times the stable level.

[0102] The value of the stability level is determined based on the single-round voxel change calibrated in step three iterations to ensure that the criterion matches the iteration convergence rhythm: in the early stage of iteration, the voxel update amplitude is large and the fluctuation of morphological parameters is a normal phenomenon, and the criterion does not easily give a stable conclusion at this stage; in the middle and late stages of iteration, the voxel update amplitude decreases, and the morphological parameters of the real defects tend to be constant, and the criterion has the conditions for recognition.

[0103] Classification and feedback of stability conclusions: Based on the conclusions of the stability criterion, the candidate defect regions are classified.

[0104] The first type is a morphologically stable defect candidate region: the exemption confirmation instruction is fed back to step three to maintain the fully exempted state of the voxel in the region and exempt it from subsequent prior violation recalculation; at the same time, the three-dimensional morphological parameter set and spatial coordinates of the defect candidate region are sent to step five to trigger the early defect judgment in step five without waiting for the overall reconstruction to converge.

[0105] The second category is candidate regions with persistently unstable morphology: these are determined to be spurious deviations caused by iterative underconvergence or sparse artifacts. Constraint recovery instructions are fed back to step three, and voxel exemption weights are assigned to the voxels within the region. Reset to 1, clear the prior exception region marker, reapply the full amount of structural constraints, so that false deviations are suppressed and eliminated in subsequent iterations; close the monitoring file for this type of defect candidate region.

[0106] The third category is defect candidate regions that are between stable and unstable: the current voxel exemption weights are kept unchanged, and monitoring rounds are continued until a clear conclusion is obtained or the reconstruction termination condition is approaching. When the conservative strategy is adopted, the voxel exemption weights are maintained but the early judgment is not triggered. After the reconstruction converges, the data is reviewed on the fine reconstruction volume data.

[0107] Utilizing the temporal morphological differences between real defects and artifacts: The physical basis for distinguishing between genuine and fake artifacts in this step lies in the fundamental difference in the morphological evolution of sparse reconstruction artifacts and real defects over the iterative time sequence. The strip-shaped artifacts caused by sparse angular sampling extend along the geometric distribution direction of the sampling viewpoint, and their anisotropy ratio is significantly higher than that of dense defects. Moreover, the artifact strips usually extend across multiple layer partitions. This penetrating feature can be identified by comparing the span in the plate thickness direction with the volumetric layer partition markings. More importantly, as iterations progress and structural constraints are gradually applied, the attenuation amplitude and volume of the artifact stripes continue to decrease, and the overall change in their morphological parameters is higher than the fluctuation level in the long term.

[0108] As a real physical structure, the projection evidence of a real defect is fixedly distributed at a consistent position in the projection residuals from various viewpoints. During iterative progress, the attenuation value and geometric boundary of the defect region rapidly solidify. Parameters such as sphericity, volume, and attenuation contrast change only slightly between consecutive iterations, and the overall change in morphological parameters remains below a stable level. Utilizing the aforementioned temporal differences, this step cross-combines the spatial dimension judgment of whether step three violates the structural prior with the temporal dimension judgment of whether the violation of morphology crosses iterative stability: those that only violate spatially but are temporally unstable are identified as artifacts and their constraints are restored; those that violate spatially but are temporally stable are confirmed as real defect carriers and their judgments are output in advance.

[0109] This cross mechanism ensures that the exemption mechanism in step three will not produce artifact residues, and also naturally limits the monitoring targets in this step to a small number of areas after the prior violation screening, keeping the monitoring computation overhead to less than 10% of the overall processing time.

[0110] Monitoring rhythm and output data list: The monitoring rhythm is synchronized with the iteration rhythm of step three. A monitoring cycle is executed after each iteration output is received. To control computational overhead, the morphological parameter extraction of the defect candidate region is only calculated for the voxel increment within the three-dimensional voxel cluster voxel list, without involving the traversal of the entire data.

[0111] The output data list for this step is as follows: The data sent to step five includes the three-dimensional morphological parameter group of the morphologically stable defect candidate region, the three-dimensional voxel cluster spatial coordinates, the voxel layer partition marker and attenuation contrast symbol attribute; the data fed back to step three includes the exemption confirmation instruction, the constraint recovery instruction and the voxel exemption weight maintenance instruction. The instruction includes the defect candidate region identifier and the voxel list to ensure that step three accurately adjusts the exemption status of the corresponding region according to the instruction.

[0112] Step 5, Defect Judgment Output and Reconstruction Acceleration: Processing target and data interface: The goal of this step is to perform defect category determination and three-dimensional parameter quantization on the morphologically stable defect candidate regions delivered in step four, generate detection results containing the three-dimensional location, size, and category of the defects, and output them in advance; at the same time, normal region confirmation information is generated and fed back to step three based on the monitoring progress, so that reconstruction computing resources are concentrated on undetermined regions, and the detection conclusions accelerate the reconstruction convergence in reverse.

[0113] The input for this step is: Step four outputs the three-dimensional morphological parameter set of the candidate region for morphologically stable defects, the three-dimensional voxel cluster spatial coordinates, the voxel layer partition marker and attenuation contrast symbol attribute, the layer interface location set, layer number, voxel layer partition marker and interface transition marker, material attenuation reference value and tolerance band output in step two, and the fine reconstructed volume data output in step three after reconstruction is terminated; the output of this step is the defect-by-defect judgment result record and summary test report, as well as the normal area confirmation information fed back to step three.

[0114] The rule system for determining defect categories: Defect category determination employs a deterministic rule system based on three-dimensional morphological parameter sets and layer information, without relying on any machine learning or deep learning models. The rule system first divides all candidate defect regions into two main branches based on the sign attribute of attenuation contrast: a low-attenuation branch and a high-attenuation branch. Defect candidate regions in the low-attenuation branch have attenuation values ​​lower than the background at the same layer, corresponding to material missing defects; defect candidate regions in the high-attenuation branch have attenuation values ​​higher than the background at the same layer or form abnormally high-attenuation connections, corresponding to material anomaly defects.

[0115] The low-attenuation branch is further subdivided into the following levels: Defect candidate regions with sphericity higher than the dense agglomerate level and volume within the void size range are classified as void defects; defect candidate regions with sphericity lower than the lamellar extension level, thickness span not exceeding a single interlayer spacing, and three-dimensional voxel clusters carrying interface transition markers are classified as delamination defects; defect candidate regions with anisotropy ratio in the principal direction higher than the strip extension level, thickness span crossing at least one interlayer interface, and sphericity in the intermediate range are classified as crack defects. The method for setting the three judgment levels—dense agglomerate level, lamellar extension level, and strip extension level—is as follows: Before implementing this method, PCB board samples with known defect types are selected. The three-dimensional morphological parameter sets of various known defects are obtained according to the processing flow from step one to step four. The sphericity distribution of void-type defect samples is statistically analyzed, and its lower quantile value is taken as the level of dense agglomerates. The sphericity distribution of layered defect samples is statistically analyzed, and its upper quantile value is taken as the level of lamellar extension. The anisotropy ratio distribution of crack-type defect samples is statistically analyzed, and its lower quantile value is taken as the level of strip extension. The lower quantile value refers to the value at which the cumulative proportion in the sample distribution reaches one-quarter, and the upper quantile value refers to the value at which the cumulative proportion in the sample distribution reaches three-quarters. When no sample is available, the default reference interval is used, with the level of dense agglomerates being 0.7 to 0.85, the level of lamellar extension being 0.3 to 0.5, and the level of strip extension being 3 to 6.

[0116] Within the high attenuation branch, the following levels are used for further subdivision: If the attenuation value of the defect candidate region is compatible with the material attenuation benchmark value of copper or solder, and the three-dimensional voxel cluster extends across the insulation distance between two adjacent traces, it is identified as a bridging defect; if the attenuation value of the defect candidate region exceeds the upper limit of the tolerance band of all known material attenuation benchmark values, it is identified as a high attenuation inclusion defect; if the attenuation value of the defect candidate region is significantly lower than the attenuation benchmark value of the dominant material in the same layer but higher than the void value level, and the attenuation distribution inside the three-dimensional voxel cluster is uneven, it is identified as a resin inclusion or low-density inclusion defect.

[0117] Each judgment conclusion is accompanied by a matching confidence level: the proportion of the number of parameters that meet the category rules among the six morphological parameters of the defect candidate region to the total number of parameters is used as the matching confidence level. The lower limit of the matching confidence level is set by default to two-thirds, that is, the judgment conclusion is valid when at least four of the six morphological parameters meet the category rules. Judgment conclusions with a confidence level lower than the set lower limit are marked in the report as pending manual review.

[0118] Quantification of the three-dimensional location and size of defects: The quantization method for the three-dimensional location of defects is as follows: calculate the centroid coordinates of the three-dimensional voxel clusters in the defect candidate region in the plate coordinate system, and the centroid coordinates are... The components are compared with the set of layer interface positions output in step two to determine the structural layer number or the layer interface number that the defect spans. The in-plane components of the centroid coordinates directly correspond to the planar position coordinates in the PCB design file, for use in process traceability. The defect size is quantified by the volume of the defect candidate area. With equivalent diameter The maximum extension length of the defect is taken as the maximum chord length in the main extension direction of the three-dimensional voxel cluster, which serves as the dimensional reference parameter for the defect.

[0119] in, The three-dimensional volume representing the candidate defect region; The equivalent diameter of a sphere with the same volume as the defect candidate region.

[0120] For void-type defects located in the solder joint area, the void ratio is further calculated: the outer contour region of the solder joint, determined by the voxel layer partition marking and the material attenuation reference value, is used as the denominator volume, and the volume of the defect candidate region is used as the numerator volume. The ratio of the two is used as the output of the solder joint void ratio. The outer contour region of the solder joint is determined as follows: within the voxel layer partition marking range corresponding to the solder layer, the set of connected voxels whose attenuation values ​​are within the tolerance band of the solder material attenuation reference value, along with the void voxels within them, are included in the denominator.

[0121] The void ratio is a direct quantitative indicator for evaluating the quality of welding processes. Its three-dimensional calculation method overcomes the inherent defects of two-dimensional slice estimation of void ratio, such as sensitivity to slice position and estimation bias.

[0122] Early output and verification mechanism for judgment results: Whenever step four sends a request to determine a candidate region of a morphologically stable defect, this step completes a category determination and parameter quantization. The determination result is recorded and written to the result cache and output immediately, without waiting for the overall reconstruction to converge.

[0123] For the candidate defect regions monitored using the conservative strategy in step four—that is, candidate defect regions that are between stable and unstable and have not triggered early judgment—after the reconstruction termination condition in step three is met, the extraction of three-dimensional morphological parameters and category determination are re-executed on the refined reconstruction volume data. The verification conclusion is then merged with the existing records in the cache. The refined reconstruction volume data also serves as the spatial carrier archive for all judgment results. Each defect in the judgment result record stores its voxel index range in the refined reconstruction volume data, ensuring that the correspondence between the detection results and the reconstruction data is traceable.

[0124] Generation and feedback of normal area confirmation information: The reconstruction acceleration function in this step is achieved through the generation and feedback of normal region confirmation information. As steps three and four proceed in parallel, all voxels of the fine-reconstructed volume data are divided into four states at any given time: confirmed defect regions, i.e., defect candidate regions that have been judged in step five; candidate monitoring regions, i.e., defect candidate regions that are being monitored in step four but have not yet reached a conclusion; identified artifact regions, i.e., voxel regions that have been judged as false deviations in step four and have converged again under the action of structural constraints after the constraint recovery command has been fed back; and normal regions that have not violated, i.e., voxel regions whose prior violation degree is lower than the exception judgment level from beginning to end.

[0125] This step generates an incremental list of confirmed normal regions at fixed iteration intervals, with the generation rhythm synchronized with the monitoring rounds in step four; that is, an incremental list is generated after each iteration. The list includes two types of voxels. One type consists of newly converged voxels in identified artifact regions. Convergence in these regions is determined by the criterion that the relative change in voxel values ​​over two consecutive iterations is lower than the voxel convergence level. The voxel convergence level is the convergence criterion applied to the relative change in voxel values, and its value range is consistent with the convergence level applied to the residual energy described in step three. to Another category consists of normal region voxels that have not produced any prior exception region markers for three consecutive rounds.

[0126] After the list is fed back to step three, step three applies structural constraint enhancement to the voxels in the list, that is, it increases the overall strength coefficient of the structural constraints. Multiplying this by an acceleration factor, which ranges from 1.5 to 3 and defaults to 2, allows these regions to approach a homogeneous state within the layer at a faster rate, and the iteration update amplitude decays rapidly. Correspondingly, the computational resources for iterative solutions naturally concentrate on candidate monitoring regions that have not yet been determined, thereby improving the reconstruction and refinement rate of defect candidate regions, which in turn shortens the number of monitoring rounds required to obtain a stable conclusion in step four.

[0127] in, This represents the overall strength coefficient of the structural constraints.

[0128] This feedback mechanism allows detection results to directly participate in the scheduling of the reconstruction process: the early determination of defective regions removes the data portion that has already been processed, and the confirmation of normal regions compresses the convergence time of the remaining reconstruction, forming a mutually reinforcing relationship between the number of determinations completed and the reconstruction convergence progress. Compared to the serial mode of unified detection after complete reconstruction convergence, the detection output time of this method is dispersed and advanced as the morphological stabilization time of each defect candidate region is reached. The total time for whole-board detection is reduced from the sum of reconstruction convergence time and detection time to the processing time corresponding to the stabilization time of the last defect candidate region, thereby improving the online detection cycle time. Furthermore, since the constraint enhancement of normal regions only applies to voxels that have been doubly confirmed, the acceleration process does not introduce the risk of defect smoothing.

[0129] Generation of summary test reports: Once the reconstruction termination conditions of step three are met and all candidate defect regions have received a judgment, this step summarizes all judgment results and generates a full-board inspection report. The inspection report includes: the total number of defects; the category, confidence level, three-dimensional coordinates of the centroid, structural layer number, volume, equivalent diameter, and maximum extension length of each defect; additional void rate indicators for solder joint void defects; statistical distribution of defects grouped by structural layer; and statistical count of the number of defects grouped by defect category.

[0130] All location information in the report is marked with both board coordinates and design document coordinates, allowing inspectors to directly pinpoint any defect to a specific layer and component on the PCB. The report concludes with the processing parameters for this inspection, including the total number of sampling angles, the total number of iterations, the stabilization time of each defect candidate region, and the number of feedback batches for the normal region confirmation list, facilitating inspection cycle analysis and process traceability.

[0131] The final output data list for this step is as follows: The defect determination results are recorded, including category, confidence level, three-dimensional location, and size parameters; the whole board inspection report; the data archive index of the fine reconstruction volume; and the feedback of normal area confirmation information from step three, including the incremental normal area confirmation list and its voxel range.

[0132] At this point, the complete processing flow from projection data acquisition to defect detection report output is completed, with data from each stage seamlessly connected: corrected projection data supports pre-reconstruction and structural prior modeling, structural prior supports violation-driven exemption constraint reconstruction, iterative intermediate data supports cross-iteration stability monitoring of defect candidate regions, monitoring conclusions support defect judgment and accelerated reconstruction, judgment results are summarized into a detection report, and all intermediate data can be traced back to the original projection data.

[0133] Example 2

[0134] The PCB board defect detection system based on CT 3D reconstruction includes the following components: Module 1: Obtain the corrected projection data sequence, per-view signal-to-noise ratio parameters, per-view penetration path length, and scanning geometric calibration parameters for sparse scanning of PCB boards; Module 2 reconstructs pre-reconstructed volume data from the corrected projection data sequence and scanning geometric calibration parameters, identifying the set of layer interface locations, generating volumetric pixel partition markers and material attenuation reference values. Module 3 uses pre-reconstructed volume data for iterative reconstruction. It sets the fidelity weight of the data per view according to the signal-to-noise ratio parameters and the penetration path length per view, and applies intra-layer homogeneity constraints according to the voxel layer partitioning labels. In each round, the prior violation is calculated by the projection residual component of the corrected projection data sequence and the intra-layer deviation of the voxel layer partitioning labels. The larger the value, the smaller the voxel exemption weight. Those exceeding the exception judgment level generate prior exception region labels and output iterative intermediate volume data. Module 4: Clustering prior exception regions are marked as three-dimensional voxel clusters, i.e. defect candidate regions. The three-dimensional morphological parameter set is extracted from the intermediate data of the iteration round by round, and the comprehensive change of morphological parameters is calculated. Stable regions are given back exemption confirmation instructions and output three-dimensional morphological parameter sets, while fluctuating regions are given back constraint recovery instructions. Module 5 determines the defect type and outputs the three-dimensional location and size in advance based on the three-dimensional morphological parameter set, material attenuation benchmark value and layer interface location set; it also adds structural constraints to the re-converged area after artifact identification and the area marked with no prior exception as normal area to accelerate convergence, and outputs the inspection report after termination.

[0135] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A PCB board defect detection method based on CT three-dimensional reconstruction, characterized in that, include: The corrected projection data sequence, per-view signal-to-noise ratio parameters, per-view penetration path length and scanning geometry calibration parameters of sparsely scanned PCB boards are obtained. The pre-reconstructed volume data is reconstructed from the corrected projection data sequence and scanning geometric calibration parameters, and the set of layer interface locations is identified, volumetric layer partition markers and material attenuation reference values ​​are generated from it. Iterative reconstruction is performed using pre-reconstructed volume data. The fidelity weight of the data is set according to the signal-to-noise ratio parameter and the penetration path length of the data per view. Intra-layer homogeneity constraints are applied according to the voxel layer partitioning labels. In each round, the prior violation degree is calculated by the projection residual component of the corrected projection data sequence and the intra-layer deviation of the voxel layer partitioning labels. The larger the value, the smaller the voxel exemption weight. Prior exception region labels are generated for those exceeding the exception judgment level, and iterative intermediate volume data is output. The clustering prior exception regions are marked as three-dimensional voxel clusters, i.e., defect candidate regions. The three-dimensional morphological parameter groups are extracted from the iterative intermediate data in round by round, and the comprehensive change of morphological parameters is calculated. The stable entity grants back the exemption confirmation instruction and outputs the three-dimensional morphological parameter set; the volatile entity grants back the constraint recovery instruction. Defect categories are determined based on the three-dimensional morphological parameter set, material attenuation reference value and layer interface location set, and the three-dimensional location and size are output in advance; The regions that re-converge after artifact identification and the regions marked with no prior exceptions are treated as normal regions. Structural constraints are added to accelerate convergence, and a detection report is output after termination.

2. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, Sparse scanning PCBs include: The PCB board is clamped with its surface parallel to the rotation axis of the turntable. Sampling angles are extracted at equal angular intervals in the full circumference for projection acquisition. The signal-to-noise ratio parameters for each angle are determined by the noise statistics of the air region in the corrected projection data sequence and the signal statistics of the PCB board projection coverage area, which are used to characterize the reliability of the projection data of each angle. The penetration path length for each angle is determined by the intersection of the initial equation of the board plane and the central ray of each sampling angle. Based on the penetration path length per view, the sampling view is divided into the normal high-quality view range, the intermediate transition view range, and the grazing-limited view range. For the sampling view that meets the preset signal-to-noise requirements, multi-frame acquisition of the same view is performed, and the average of the multi-frame projection is used as the source of the corrected projection data for the corresponding view.

3. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, Identifying the set of layer interface locations from pre-reconstructed volumetric data and generating voxel-by-voxel layer partition markers includes: The pre-reconstructed volume data is resampled to the plate coordinate system. Sampling columns are arranged throughout the plate surface, running through the plate thickness direction. The median value of each sampling column is taken for each plate thickness coordinate to obtain the plate thickness attenuation profile. The peak positions in the rate of change sequence of the plate thickness attenuation profile that exceed the adaptive rate of change level are taken as the layer interface position set. Based on the layer interface position set, the volume data is divided into layer partitions on a voxel-by-voxel basis, and voxel-by-voxel layer partition labels are generated. Interface transition labels are added to voxels whose distance from the layer interface is less than the width of the interface transition band.

4. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, The estimated baseline values ​​for material degradation include: The values ​​of the high-level plateaus in the thickness-direction attenuation profile are sorted by amplitude and divided at the point where the difference between adjacent values ​​is the largest. The high amplitude class corresponds to the attenuation reference value of solder material, the medium amplitude class corresponds to the attenuation reference value of copper material, and the low plateau value corresponds to the attenuation reference value of resin dielectric material. Each material attenuation reference value is accompanied by a tolerance band, which is determined according to the dispersion of the values ​​within the corresponding plateau.

5. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, Iterative reconstruction solves for the refined reconstructed volume data by minimizing the objective function, which is: ； in, The data represents the refined reconstructed volume obtained by minimizing the objective function; Represents the current volume data during the iteration process; Represents the total number of actual sampled viewpoints; Representing the Per-view data fidelity weights for each sampling view; Representing the The system projection matrix corresponding to each sampling viewpoint; Representing the Corrected projection data for each sampling viewpoint; Represents the overall strength coefficient of the structural constraints; Representative voxels Voxel exemption weights at the location; Representative action on voxels Structural constraint functionals; Intralayer homogeneous constraints are constructed using the following formula: ； in, Representative voxels The set of spatial neighborhood voxels; Representing neighborhood voxels Volumetric layer partitioning markers; Representative voxels Volumetric layer partitioning markers; Representing neighborhood voxels The current attenuation reconstruction value; Representative voxels The current attenuation reconstruction value.

6. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 5, characterized in that, The prior violation and voxel exemption weights are determined by the following calculation chain: First, calculate the prior violation of the voxel: ； in, Representative voxels a priori violation degree; Representative voxels The projected residual components; Representative voxels Intra-layer deviation; The weighting coefficients representing the projected residual components; The weighting coefficients representing the deviation within a layer. and The sum is 1; Then, the prior violation degree Substituting into the following formula, we obtain the voxel exemption weights. ； in, Representative voxels Voxel exemption weights at the location; Represents the baseline level of violation, which is a preset multiple of the median of the prior violation values ​​of all voxels in the current iteration.

7. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 5, characterized in that, In each iteration of the iterative reconstruction, the objective function is solved using an alternating optimization approach: First, perform a per-view algebraic update on each sampled view in order of view index, correct the solution set direction of the projection equation of the current volume data onto each view, and scale the correction step size according to the per-view data fidelity weight of the corresponding sampled view. Then, the gradient correction of the structural constraint terms is performed, and the correction amount of each voxel is scaled according to the voxel exemption weight of each voxel.

8. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, The three-dimensional morphological parameter set includes the volume of the defect candidate region, equivalent diameter, sphericity, anisotropy ratio in the principal direction, span in the thickness direction, and attenuation contrast. The overall change in morphological parameters is calculated using the following formula: ； in, Representing the The overall change in morphological parameters across monitoring rounds; This represents the total number of parameters included in the 3D morphology parameter group; Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The first monitoring round extracted the first Item morphological parameter values; Representing the The minimum measurement standard for the morphological parameters of an item; The morphological parameters of the defect candidate region are determined to be stable if the total change in the morphological parameters is below the stable level for a consecutive preset number of rounds, and unstable if it is continuously above the fluctuation level.

9. The PCB board defect detection method based on CT three-dimensional reconstruction according to claim 1, characterized in that, The defect categories include: Based on the attenuation contrast sign, the candidate defect region is divided into a low attenuation branch and a high attenuation branch. In the low attenuation branch, those with sphericity higher than the level of dense agglomerates are identified as void defects, those with sphericity lower than the level of lamellar extension and a thickness span not greater than the distance between adjacent layer interfaces are identified as delamination defects, and those with anisotropy ratio in the main direction higher than the level of strip extension and a thickness span crossing at least one layer interface in the set of layer interface locations are identified as crack defects. In the high attenuation branch, defects whose attenuation values ​​are compatible with the material attenuation reference values ​​of copper or solder and whose three-dimensional voxel clusters cross the insulation spacing of adjacent traces are identified as bridging defects. For void defects located in the solder joint area, the void rate is output as the ratio of the volume of the defect candidate area to the volume of the outer contour area of ​​the solder joint. The outer contour area of ​​the solder joint is the set of connected voxels of the solder joint determined by the voxel layer partitioning mark and the material attenuation reference value, along with the void voxels inside it. The method is applied to the online inspection of PCB boards. The whole board inspection report records the total number of sampling perspectives, the total number of iterations, the number of iterations at which each defect candidate area stabilizes, and the number of batches of the normal area confirmation list feedback.

10. A PCB board defect detection system based on CT three-dimensional reconstruction, wherein the PCB board defect detection method based on CT three-dimensional reconstruction according to any one of claims 1-9 is characterized in that, include: Module 1: Obtain the corrected projection data sequence, per-view signal-to-noise ratio parameters, per-view penetration path length, and scanning geometric calibration parameters for sparse scanning of PCB boards; Module 2 reconstructs pre-reconstructed volume data from the corrected projection data sequence and scanning geometric calibration parameters, identifying the set of layer interface locations, generating volumetric layer partition markers and material attenuation reference values; Module 3 uses pre-reconstructed volume data for iterative reconstruction. It sets the fidelity weight of the data per view according to the signal-to-noise ratio parameters and the penetration path length per view, and applies intra-layer homogeneity constraints according to the voxel layer partitioning labels. In each round, the prior violation is calculated by the projection residual component of the corrected projection data sequence and the intra-layer deviation of the voxel layer partitioning labels. The larger the value, the smaller the voxel exemption weight. Those exceeding the exception judgment level generate prior exception region labels and output iterative intermediate volume data. Module 4: Clustering prior exception regions are marked as three-dimensional voxel clusters, i.e. defect candidate regions. The three-dimensional morphological parameter set is extracted from the intermediate data of the iteration round by round, and the comprehensive change of morphological parameters is calculated. Stable regions are given back exemption confirmation instructions and output three-dimensional morphological parameter sets, while fluctuating regions are given back constraint recovery instructions. Module 5 determines the defect type and outputs the three-dimensional location and size in advance based on the three-dimensional morphological parameter set, material attenuation benchmark value and layer interface location set; The regions that re-converge after artifact identification and the regions marked with no prior exceptions are treated as normal regions. Structural constraints are added to accelerate convergence, and a detection report is output after termination.

Images