Chip soldering inspection method

Abstract

The application discloses a chip welding detection method and relates to the technical field of electronic manufacturing detection. The method comprises the following steps: collecting visual images, X-ray projections, thermoelectric sequences, process data and geometric priors to generate a solder joint evidence; completing registration, calibration and alignment correction according to the geometric priors to obtain a correction evidence package, an uncertainty and a calibration parameter; generating a structure embedding based on the correction evidence package self-supervised learning; combining the correction X-ray projection, the calibration parameter, the geometric prior and the structure embedding to generate a voxel field and a voxel uncertainty; generating a mechanism free energy, a residual error, a conflict degree and an evidence coupling potential energy coefficient to determine a mechanism according to a preset defect mechanism set; and generating a risk interval and determining the risk interval according to the mechanism, the voxel field, the uncertainty and the potential energy coefficient. Through the geometric prior constraint multi-source evidence alignment and the few-view three-dimensional reconstruction, the risk interval is generated and determined in combination with the mechanism free energy and the evidence coupling potential energy coefficient, so that the consistency and controllability of the solder joint detection are improved.

Description

Technical Field

[0001] This invention relates to the field of electronic manufacturing inspection technology, specifically a chip welding inspection method. Background Technology

[0002] In the surface mount technology (SMT) manufacturing process, chip devices are connected by solder joints through printing, mounting, and reflow soldering. The geometry and internal defects of these solder joints directly affect the electrical connectivity and service reliability of the devices. With the development of device miniaturization, high-density packaging, and array interconnection, the spacing between solder joints has decreased, the number of solder balls has increased, and the size and shape of solder joint defects have become smaller and more complex. Moreover, some defects are located on the bottom of the device or in the internal structure, exhibiting invisible or weakly visible characteristics, which significantly increases the difficulty of inspection.

[0003] Current welding inspection methods typically employ automated optical inspection, X-ray inspection, online electrical testing, or manual review. Automated optical inspection relies on visual information and can achieve high-speed full inspection, but its detectability of defects such as component-obscured areas, bottom solder joints, internal voids, and cracks is limited. When there is surface reflection, flux residue, shadowing, or significant changes in board texture, the inspection results are prone to false alarms or missed detections. X-ray inspection can provide internal structural information, but common projection imaging is a two-dimensional observation. Factors such as defect and structure overlap, scattering and beam hardening artifacts, thick plate attenuation, and the number of viewing angles limited by the time frame can easily introduce structural ambiguity. Under limited viewing angle conditions, the reconstruction or interpretation process often relies on empirical thresholds or fixed rules, making it difficult to stably adapt to different board types, batches, and process windows. Online electrical testing can detect connectivity anomalies, but it is usually difficult to provide solder joint-level spatial location and defect mechanism explanation, and its sensitivity to early reliability issues is limited by test excitation and measurement conditions.

[0004] In the practice of combining multiple sources, the registration, calibration, and temporal alignment of cross-modal data are crucial. Existing methods often employ staged processing or simple correlation strategies, which are prone to deviations in weld point indexing, clipping windows, viewpoint consistency, and time window consistency, making it difficult to form a closed loop of multi-source evidence for the same weld point. Simultaneously, detection algorithms often lack a unified characterization of observational and structural uncertainties, making it difficult to incorporate the credibility of different modalities into a consistent inference framework. For situations where multiple defect mechanisms coexist or overlap, such as voids, cracks, insufficient wetting, bridging, and headrest defects, existing technologies often substitute single classification outputs or simple scoring for mechanism discrimination, lacking a mechanism-level measure of separability, resulting in unclear defect type boundaries and unstable verification strategies.

[0005] Furthermore, existing technologies mostly use "point value judgment" to determine whether a chip is qualified or not, making it difficult to form a traceable risk range expression. For data distribution drift caused by different batches, different equipment states, and different imaging conditions, model maintenance usually relies on manual parameter tuning or repeated annotation training, resulting in high engineering costs. For these reasons, there is still a need for a chip welding inspection method that can achieve consistent organization of multi-source evidence at the solder joint granularity, complete registration and alignment correction under geometric prior constraints, and achieve computable inference and risk range output at the defect mechanism level, in order to improve the reproducibility, traceability, and engineering adaptability of the inspection process. Summary of the Invention

[0006] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a chip welding inspection method to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a chip welding inspection method, comprising: S1: Acquire visual images, X-ray projections, thermoelectric sequences, process data, and geometric priors to generate weld point evidence; S2: Perform registration, calibration, and alignment correction on the weld point evidence based on geometric priors, and generate a correction evidence package, uncertainty, and calibration parameter set; S3: Generate structural embeddings through self-supervised learning based on the correction evidence package; S4: Generate voxel fields and voxel uncertainties based on the resampled weld X-ray projection fragment sequence, calibration parameters, geometric priors, and structural embeddings in the correction evidence package; S5: Preset a set of defect mechanisms, generate mechanism free energy, residuals and conflict degree based on correction evidence, voxel field and voxel uncertainty, generate evidence coupling potential energy coefficient based on mechanism free energy, residuals and conflict degree, and determine the weld joint mechanism conclusion; S6: After generating the risk range based on the weld joint mechanism conclusion, voxel field, voxel uncertainty and evidence coupling potential coefficient, output the judgment.

[0008] The present invention is further configured such that the generation of solder joint evidence includes: Based on the geometric prior, the solder joint identifier and nominal location are determined, and the solder joint index relationship is established. The geometric prior generates the local geometric prior of the solder joint according to the solder joint identifier. Based on the nominal location of the solder joint, extract the solder joint image fragment from the visual image; based on the nominal location of the solder joint, extract the solder joint projection fragment sequence from the X-ray projection; based on the solder joint identifier, extract the solder joint time fragment from the thermoelectric sequence. Extract solder joint process data items from the process data based on the solder joint identifier; Solder joint evidence is generated by associating solder joint identifiers, solder joint image fragments, solder joint projection fragment sequences, solder joint time segments, solder joint process data items, and local geometric priors of the solder joint.

[0009] The present invention is further configured such that the registration, calibration, and alignment correction of the solder joint evidence based on geometric priors includes: Based on the prior analysis of the local geometry of the solder joint, the nominal center of the solder joint, the local orientation base, the boundary of the solder pad, and the boundary of the solder mask opening, a standard mesh for the solder joint is established. A visual reference template is generated based on the prior local geometry of the weld joint. Prior constraint registration is performed on the visual fragments of the weld joint based on the visual reference template to obtain the visual registration parameters. The visual fragments of the weld joint are then resampled to the weld joint specification mesh. Based on the local geometric prior of the weld joint, a nominal three-dimensional prior volume of the weld joint is generated. Based on the nominal three-dimensional prior volume of the weld joint, the viewpoint geometry is calibrated on the X-ray projection segment sequence of the weld joint to obtain the viewpoint calibration parameters. The X-ray projection segment sequence of the weld joint is resampled to the weld joint standard mesh and a cross-viewpoint consistent index is generated. Perform event window alignment on the thermoelectric time segment of the solder joint to obtain time alignment parameters, and resample the thermoelectric time segment of the solder joint to the standard time axis.

[0010] The present invention is further configured such that the generation of the correction evidence package, uncertainty, and calibration parameter set includes: The resampled solder joint visual fragments, the resampled solder joint X-ray projection fragment sequence, and the resampled solder joint thermoelectric time fragments are encapsulated into a correction evidence package; Visual registration parameters, view calibration parameters, time alignment parameters, and cross-view consistency index are encapsulated into a set of calibration parameters; Visual uncertainty, projection uncertainty, and thermoelectric uncertainty are generated based on the residual information from the registration, calibration, and alignment correction processes.

[0011] The present invention is further configured such that the self-supervised learning to generate structural embeddings based on the correction evidence package includes: Solder joint samples were constructed based on the correction evidence package. The solder joint samples included visual segments, X-ray projection segment sequences, and thermoelectric time segments. Mask reconstruction constraints are applied to the solder joint samples to generate visual intermediate representations, X-ray intermediate representations, and thermoelectric intermediate representations. Perform cross-modal uniform constraints on visual intermediate characterization, X-ray intermediate characterization and thermoelectric intermediate characterization to generate visual embedding, X-ray embedding and thermoelectric embedding; Structural embedding is generated based on visual embedding, X-ray embedding, and thermoelectric embedding.

[0012] The present invention is further configured such that the voxel field generated based on the resampled weld point X-ray projection segment sequence, calibration parameters, geometric priors, and structural embeddings in the correction evidence package includes: The three-dimensional voxel mesh and support mask of the solder joint are determined based on the local geometric prior of the solder joint. The support mask defines the feasible occupied area within the voxel mesh. Based on the perspective geometry parameters in the calibration parameters, construct the forward projection relationship of each perspective, and organize the X-ray projection segment sequence of the weld point into an observation sequence by cross-perspective consistent indexing. Observation weights are generated based on the uncertainty of the X-ray projection segment sequence of the weld joint, and data consistency constraints are formed based on the observation weights, forward projection relationship, and observation sequence. Based on the structural embedding and the local geometry prior of the weld points, conditional prior constraints are generated. Based on the data consistency constraints and conditional prior constraints, a voxel field is generated within the voxel mesh defined by the support mask.

[0013] The present invention is further configured such that the uncertainty of the generated voxel includes: Voxel information parameters are generated based on forward projection relationships, observation weights, and voxel fields. Prior information parameters are generated based on the conditional prior constraints and the voxel field. The voxel information parameters and the prior information parameters are then associated with the solder joint identifier. Voxel uncertainty is generated based on voxel information parameters and prior information parameters.

[0014] The present invention is further configured such that the generation mechanism free energy, residual, and conflict degree include: A set of pre-defined defect mechanisms is used to generate a sequence of mechanistic free energy under each mechanism condition in the set of defect mechanisms, based on correction evidence, voxel field, and voxel uncertainty. Based on the X-ray projection segment sequence of the weld joint and the voxel field, a projection-consistent residual is generated; based on the geometric prior and the voxel field, a geometrically consistent residual is generated; based on the corrected thermoelectric sequence and the mechanism conditions, a temporally consistent residual is generated; and based on the projection-consistent residual, the geometrically consistent residual, and the temporally consistent residual, a physical residual intensity is generated. The conflict degree is generated based on the distribution of mechanistic evidence on the defect mechanism set, including visual evidence, X-ray evidence, and thermoelectric evidence. The distribution of mechanistic evidence is generated by the mechanistic free energy of the corresponding mode.

[0015] The present invention is further configured such that the generation of evidence coupling potential energy coefficients to determine the solder joint mechanism conclusions includes: The optimal and suboptimal defect mechanisms are determined based on the mechanism free energy sequence, and the free energy gap is generated based on the optimal and suboptimal defect mechanisms. The evidence coupling potential coefficient is generated based on the free energy gap, physical residual strength, and conflict degree. The conclusion of the solder joint mechanism is generated based on the optimal defect mechanism.

[0016] The present invention is further configured such that the output judgment after generating the risk interval includes: Based on the solder joint mechanism conclusion, select the mechanism sensitive operator corresponding to the solder joint mechanism conclusion from the preset mechanism sensitive operator set, and map the voxel field to generate the mechanism sensitive field based on the mechanism sensitive operator; Based on the mechanism-sensitive field and voxel uncertainty, generate mechanism-damage surrogate quantity and surrogate quantity uncertainty; Based on the conclusions of the weld joint mechanism, the surrogate quantity of the mechanism damage and the uncertainty of the surrogate quantity, risk point estimates and model uncertainties are generated. The interval scale parameter is generated based on the risk point estimate, model uncertainty, surrogate quantity uncertainty, and evidence coupling potential coefficient. The risk interval is then generated based on the interval scale parameter. The judgment label is output based on the risk range and the preset judgment threshold set.

[0017] This invention provides a chip soldering inspection method. The method generates solder joint evidence by acquiring visual images, X-ray projections, thermoelectric sequences, process data, and geometric priors. Based on the geometric priors, the solder joint evidence is registered, calibrated, and aligned to generate a calibration evidence package, uncertainty, and calibration parameter set. Self-supervised learning is performed on the calibration evidence package to generate a structural embedding. Based on the resampled solder joint X-ray projection segment sequence, calibration parameters, geometric priors, and structural embedding in the calibration evidence package, voxel fields and voxel uncertainties are generated. A set of defect mechanisms is preset, and based on the calibration evidence, voxel fields, and voxel uncertainties, mechanism free energy, residuals, and conflict degrees are generated. Based on the mechanism free energy, residuals, and conflict degrees, an evidence coupling potential energy coefficient is generated to determine the solder joint mechanism conclusion. Based on the solder joint mechanism conclusion, voxel fields, voxel uncertainties, and evidence coupling potential energy coefficient, a risk interval is generated, and a judgment is output. The beneficial effects include: 1. Consistent organization and traceable transfer of multi-source evidence: Based on geometric priors, weld point identifiers and local priors are generated. Visual segments, X-ray projection segment sequences and thermoelectric time segments are extracted in a unified manner to form weld point evidence and run through a continuous data chain from the correction evidence package, structural embedding, voxel field to the risk interval, reducing cross-modal index drift and improving the consistency and traceability of the detection process. 2. Few-view structural reconstruction and voxel uncertainty constraint: Based on the X-ray projection segment sequence of weld points, calibration parameters and structural embedding, a voxel field is generated, and voxel uncertainty is generated simultaneously to characterize the strength of structural layer constraints. This enables the internal structural information under few-view conditions to have verifiable support domains and uncertainty characterization, reduces structural ambiguity caused by underdetermined reconstruction, and improves the stability of subsequent mechanism inference. 3. Mechanism inference and potential energy-driven risk range determination: under the defect mechanism set, generate mechanism free energy, residual and conflict degree, form evidence coupling potential energy coefficient and output weld joint mechanism conclusion; under the constraint of mechanism sensitive operator, combine voxel uncertainty and potential energy coefficient to generate risk range and determine, realize closed-loop output from mechanism layer to risk layer, and improve the controllability of determination.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart illustrating a chip welding inspection method as an exemplary embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0023] Example 1: A chip soldering inspection method, such as Figure 1 As shown, it includes: S1: Acquire visual images, X-ray projections, thermoelectric sequences, process data, and geometric priors to generate weld point evidence; S2: Perform registration, calibration, and alignment correction on the weld point evidence based on geometric priors, and generate a correction evidence package, uncertainty, and calibration parameter set; S3: Generate structural embeddings through self-supervised learning based on the correction evidence package; S4: Generate voxel fields and voxel uncertainties based on the resampled weld X-ray projection fragment sequence, calibration parameters, geometric priors, and structural embeddings in the correction evidence package; S5: Preset a set of defect mechanisms, generate mechanism free energy, residuals and conflict degree based on correction evidence, voxel field and voxel uncertainty, generate evidence coupling potential energy coefficient based on mechanism free energy, residuals and conflict degree, and determine the weld joint mechanism conclusion; S6: After generating the risk range based on the weld joint mechanism conclusion, voxel field, voxel uncertainty and evidence coupling potential coefficient, output the judgment.

[0024] The present invention is further configured such that the generation of solder joint evidence includes: Based on geometric priors, solder joint identifiers and nominal locations are determined, and a solder joint index relationship is established. The geometric prior generates a local geometric prior for each solder joint based on its identifier. Specifically, the geometric prior inputs use circuit board design data and package library data, which are parsed to obtain device reference bases, solder joint numbering rules, pad center coordinate tables, pad outlines, and solder mask window outlines. Solder joint identifiers are generated according to the solder joint numbering rules, and an index relationship is established between these identifiers and the pad center coordinates, forming a directly searchable solder joint index table. The local geometric prior is formed by cropping the pad outline and solder mask window outline. The local range is determined by extending the pad outline to a safety boundary, which is constrained by solder joint spacing and imaging resolution, ensuring that the local range covers the effective area of ​​the solder joint and does not cross adjacent solder joints. This process fixes the solder joint location and boundary constraints on the data side, ensuring consistent indexing and support domains for subsequent multimodal evidence extraction. Based on the nominal location of the solder joint, solder joint image fragments are extracted from the visible image. Based on the nominal location of the solder joint, a sequence of solder joint projection fragments is extracted from the X-ray projection. Based on the solder joint identifier, a time fragment of the solder joint is extracted from the thermoelectric sequence. Specifically, on the visible image side, the conversion relationship between camera calibration parameters and board coordinates to pixel coordinates is read, the center coordinates of the solder pad are mapped to image coordinates, and the solder joint image fragments are obtained by cropping according to the local range given by the local geometry prior of the solder joint. At the same time, the acquisition configuration such as the cropping start point, cropping scale, imaging resolution, exposure and gain are recorded as fragment metadata. On the X-ray projection side, the projection image is read according to the viewpoint number. Based on the initial values ​​of the device geometry and the initial mapping from board coordinates to projection coordinates, the center position of the solder joint in each viewpoint is obtained. The solder joint projection fragment sequence is obtained by cropping according to the local range consistent with the visible side. The sequence is organized by viewpoint number and the viewpoint number list and cropping scale are recorded to ensure the consistency of the evidence support domain between viewpoints. The thermoelectric sequence side reads the mapping relationship between solder joint identifiers and channel numbers or measurement area numbers in the test configuration, determines the starting point of the sampling window based on the trigger timestamp, and extracts the solder joint time segment based on the sampling rate and preset window length. The sampling rate, window starting point, and window length are recorded as time segment metadata. This process uses geometric priors to drive the pruning and uses configuration metadata to solidify the evidence source, enabling cross-modal evidence to have a traceable and consistent organizational form. Solder joint process data items are extracted from the process data based on solder joint identifiers. Specifically, the process data input includes printing inspection results, placement offset records, key feature quantities of reflow curves, and batch equipment status markers. First, component-level association is completed, and then solder joint-level mapping is performed based on the solder joint identifiers. When the process output is a gridded field, the corresponding field is read after determining the grid cell based on the pad center coordinates. When the process output is a solder joint-level field, it is directly indexed and read based on the solder joint identifier. Each solder joint process data item forms at least a structured field set, including solder paste volume deviation level, placement offset direction and offset level, reflow critical temperature zone arrival time level and peak temperature level, and batch status markers. Each level field is obtained by conversion using a preset classification table, which is given by the process specification and is consistent with the batch records. Solder joint evidence is generated by associating solder joint identifiers, solder joint image segments, solder joint projection segment sequences, solder joint time segments, solder joint process data items, and local geometric priors. Specifically, using the solder joint identifier as the primary key, associations are established for solder joint image segments, solder joint projection segment sequences, solder joint time segments, solder joint process data items, and local geometric priors, and the resulting evidence is encapsulated. The solder joint evidence synchronously includes device identifiers, pad center coordinates, clipping element information, a list of viewpoint numbers, channel numbers, timestamps, and acquisition configuration markers, forming a complete data transmission carrier. Missing data, invalid frames, or window truncation are recorded in visible segments, projection segment sequences, or time segments, and a missing status marker is established between the missing status marker and the corresponding segment to ensure consistent processing strategies for subsequent registration, calibration, and alignment correction. A circuit board reference coordinate system is set up as a unified reference coordinate system. This reference coordinate system is established based on the positioning holes, reference pads, or alignment marks in the geometric prior. On the visual imaging side, pixel coordinates are mapped to the circuit board reference coordinate system through camera calibration parameters. On the X-ray imaging side, projection coordinates are mapped to the circuit board reference coordinate system through the constraint that the initial geometric values ​​of the equipment are consistent with the geometric prior projection. On the thermoelectric side, solder joint identifiers and measurement point channel numbers or measurement area numbers are mapped to the geometric prior measurement point coordinate table, forming a unified index binding for solder joints in the three modes. Time scale differences are resolved through event window normalization. The event window is defined by the test configuration or process trigger signal, including excitation start, steady state arrival, and excitation end. Thermoelectric time segments are resampled to the standard time axis according to the event window. The length of the standard time axis and the sampling rate are constrained by the allowable deviation of the process step, keeping the event sequence monotonic and limiting the maximum stretch ratio. Registration error is uniformly represented by residual statistics. The visual residual is obtained from the offset between the reference template boundary and the corrected visual segment boundary. The projection residual is obtained from the difference between the predicted projection and the corrected projection segment of the nominal 3D prior body under the viewpoint geometry configuration. The temporal residual is obtained from the morphological difference between the mechanistic condition thermal prediction and the corrected thermoelectric segment within the event window. The residuals are monotonically mapped to generate uncertainty and participate in subsequent weighting. The upper limit of the allowable deviation is determined by the geometric prior constraints. The smaller value between the solder joint spacing grading threshold and the minimum feature size grading threshold of the solder pad is preferably taken as the registration allowable deviation. When the registration error exceeds the upper limit of the allowable deviation, the uncertainty of the corresponding solder joint is increased and a re-registration or over-sampling strategy is triggered. Re-registration prioritizes local boundary refinement, and over-sampling prioritizes adding a small number of viewpoints or improving local exposure consistency.

[0025] The present invention is further configured such that the registration, calibration, and alignment correction of the solder joint evidence based on geometric priors includes: Based on the local geometric prior analysis of the solder joint, the nominal center, local orientation base, pad boundary, and solder mask window boundary are analyzed to establish a standardized solder joint mesh. Specifically, the local geometric prior reads the nominal center position, local orientation information, pad boundary contour, and solder mask window boundary contour. The nominal center is used to determine the local coordinate origin of the solder joint, and the local orientation information is used to determine the orientation of the local coordinate axes, ensuring a consistent orientation definition for different solder joints under the same rules. The pad boundary contour and the solder mask window boundary contour are used to determine the effective local support range. The support range is generated by the joint constraint of the pad shape and the window shape, with an outward safety boundary to cover the solder extension area. The safety boundary is constrained by the solder joint spacing and pixel resolution to avoid crossing adjacent solder joints. The standardized mesh is established with the nominal center as the mesh center, the local orientation as the mesh coordinate axes, and the support range as the mesh coverage boundary. The mesh size, mesh resolution, and the mapping relationship between the mesh and board coordinates are recorded as unified target coordinates for subsequent resampling. This process ensures that the visible segments, projected segment sequences, and thermoelectric segments have a consistent spatial support framework before entering subsequent processing, reducing alignment deviations caused by geometric differences between solder joints. A visual reference template is generated based on the prior local geometry of the solder joints. Prior constraint registration is then performed on the visible segments of the solder joints based on this template to obtain visual registration parameters. The visible segments are then resampled to the solder joint's standard mesh. Specifically, visual registration employs a two-stage strategy: first, an initial mapping from the board surface to the image is obtained using geometric matching of reference feature points; then, local refinement is performed within the local window of the solder joint using the pad boundaries and solder mask window boundaries. The initial mapping is solved using robust estimation, and outlier removal is based on feature point reprojection deviation exceeding a preset pixel threshold. Local refinement uses boundary consistency optimization to minimize the positional error of the pad boundaries within the standard mesh. After visual registration, the visual registration parameters are output, along with the residual statistics of the reference feature points. These residual statistics include at least the mean square deviation and the maximum deviation. The visual reference template is rendered within the standard mesh using the pad boundary contours and solder mask window boundary contours. The template includes the pad outline boundary, window boundary, and nominal center marker, and uses the same resolution and coordinate system as the standard mesh during template rendering. The prior constraint registration employs a robust geometric registration strategy: targeting the boundary curve of the template and observing the structural edges in the visible segments, a multi-scale iterative approach is used to first solve the coarse registration and then the fine registration. In the coarse registration stage, only translational and rotational degrees of freedom are allowed to avoid overfitting when texture noise is high. In the fine registration stage, slight affine adjustments are introduced while maintaining geometric consistency constraints to absorb lens distortion and slight warping of the board surface. Robust loss suppression is used for outlier matching points during registration, with outlier identification based on the degree of boundary consistency deviation, preventing reflections, stains, or screen printing interference from having a dominant impact on the registration results. After registration, the visible segments are mapped to a standard mesh according to the registration parameters and resampling is performed. Resampling uses an edge-preserving interpolation strategy to retain boundary geometric information, while simultaneously recording the registration parameters, resampling interpolation method, and sampling scale, forming visual correction evidence and visual registration parameter output. This process uses geometric boundaries as the core constraint, ensuring a direct correspondence between the registration target and the solder joint structure consistency, reducing the risk of drift solely based on texture-related factors. Based on the local geometric prior of the solder joint, a nominal 3D prior volume of the solder joint is generated. Viewpoint geometric calibration is then performed on the X-ray projection segment sequence of the solder joint based on this nominal 3D prior volume, yielding viewpoint calibration parameters. The X-ray projection segment sequence is then resampled to the solder joint's standard mesh, and a cross-viewpoint consistent index is generated. Specifically, the X-ray calibration employs a priori projection consistency constraints: using the nominal 3D prior volume of the solder joint generated by the geometric prior as a reference, the viewpoint geometric parameters are adjusted to ensure that the reference projection matches the metal contour and pad boundary projection positions of the actual projection segment. The calibration consistency metric includes both contour consistency and attenuation trend consistency terms, avoiding false matching caused by using only grayscale differences. For each viewpoint, the system outputs viewpoint geometric parameters and viewpoint calibration residuals. The viewpoint calibration residuals include at least two types of statistics: contour deviation and intensity deviation. A cross-viewpoint consistency index records the viewpoint number, clipping boundary identifier, and resampling scale identifier, ensuring that the observation sequence is organized within the same solder joint and support domain. The nominal 3D prior is generated from the local geometric prior of the solder joint and the package structure rules. The generated content includes at least the spatial location of the solder pad, the 3D support volume allowed for solder, and sidewall constraints consistent with the solder pad boundary. The 3D prior is used to generate predictive projections at each viewpoint. Predictive projection generation uses the device's initial geometry as a starting point, iteratively adjusting the viewpoint geometric parameters to ensure consistency between the predicted projection and the actual projected segment in terms of structural contour and grayscale attenuation distribution. The viewpoint geometric parameters include the relative pose of the source detector corresponding to the viewpoint number, the projection scale, and the principal point offset. The adjustment process is constrained by the device's reachability to prevent the solution from falling into unrealizable geometric configurations. To avoid spurious matching based solely on grayscale differences, the calibration consistency metric includes both contour consistency and strength consistency terms: the contour consistency term constrains the coincidence of pad boundaries in the projection, while the strength consistency term constrains the attenuation distribution to align with the projection thickness trend. After obtaining the viewpoint geometric calibration parameters, the projected segments of each viewpoint are mapped to a normalized grid for resampling, and a cross-viewpoint consistency index is generated in order of viewpoint number. The consistency index includes at least the viewpoint number, resampling scale, clipping boundary identifier, and grid mapping identifier to ensure that subsequent multi-view sequences are organized within the same spatial support domain at the same solder joint. This process eliminates geometric deviations between viewpoints through the link of prior volume—predicted projection—geometric calibration—normalized resampling, providing a consistent spatial alignment basis for observations under conditions of few viewpoints. Event window alignment is performed on the thermoelectric time segment of the solder joint to obtain time alignment parameters. The thermoelectric time segment of the solder joint is then resampled to the standard time axis. Specifically, the thermoelectric time segment alignment establishes a standard time axis centered on the event window. The event window is defined by key trigger events in the process record or test configuration. Events include heating start, reaching steady state, stopping heating, or load change points. The alignment adopts a dynamic time warping strategy: using the temperature change pattern within the event window as a constraint, a monotonic mapping from the original time axis to the standard time axis is solved, absorbing time drift caused by different batches or different trigger delays, while maintaining the event sequence. Monotonicity constraints and maximum stretch constraints are set during the alignment process. The monotonicity constraint ensures that the time mapping does not reverse, and the maximum stretch constraint is determined by the sampling rate and the allowable deviation of the process step to prevent excessive stretching of the mapping, which would lead to distortion of the physical meaning. After obtaining the time alignment parameters, the thermoelectric time segment is resampled to the standard time axis. The resampling uses conformal interpolation to maintain the relative relationship between the peak position and the rise and fall slopes, and the standard time axis length, sampling rate, and alignment parameters are recorded as part of the calibration parameters. This process makes thermoelectric evidence comparable in terms of time scale, providing consistent temporal support for subsequent mechanism inferences and avoiding cross-batch misalignment issues caused by using a fixed truncation starting point.

[0026] The present invention is further configured such that the generation of the correction evidence package, uncertainty, and calibration parameter set includes: The resampled visible solder joint segments, the resampled X-ray projection segment sequence of solder joints, and the resampled thermoelectric time segment of solder joints are encapsulated into a calibration evidence package. Specifically, the resampled visible solder joint segments, the resampled X-ray projection segment sequence of solder joints, and the resampled thermoelectric time segment of solder joints are aggregated and encapsulated according to solder joint identifiers to form a calibration evidence package. The calibration evidence package refers to the collection of resampled visible segments, X-ray projection segment sequences of solder joints, and thermoelectric time segments encapsulated according to solder joint identifiers. The calibration parameter set refers to the collection of visual registration parameters, viewpoint calibration parameters, time alignment parameters, and cross-viewpoint consistency indexes that are bound one-to-one with the calibration evidence package. During encapsulation, a structured organization method is adopted, with solder joint identifiers as the primary key, evidence content as the payload, and metadata as additional fields: the evidence content includes three types of resampled segment bodies, and the metadata includes canonical grid identifiers, resampled resolution, clipping scale, viewpoint number list, canonical time axis length, and sampling rate. To ensure consistency in cross-modal data transmission, the standardized mesh identifier and standardized time axis identifier are enforced to be unique within the evidence package, and the viewpoint number list is written in a one-to-one correspondence with the projected segment sequence to avoid viewpoint order disorder or scale inconsistency in subsequent processing stages. This encapsulation method allows subsequent self-supervised learning, 3D reconstruction, and mechanism inference to directly use the corrected evidence package as input without repeated alignment and resampling processing, reducing untraceable errors caused by repeated calculations and implicit transformations at the data link level; Visual registration parameters, viewpoint calibration parameters, time alignment parameters, and cross-viewpoint consistency indexes are encapsulated into a calibration parameter set. Specifically, these parameters are encapsulated into a calibration parameter set based on weld point identifiers. Visual registration parameters at least include the geometric mapping relationship from the visual segment to the canonical mesh, the mapping degree of freedom configuration, and the iterative convergence marker. Viewpoint calibration parameters at least include the geometric pose configuration, projection scale configuration, and principal point offset configuration for each viewpoint, corresponding one-to-one with the viewpoint number. Time alignment parameters at least include a monotonic mapping description from the original time axis to the canonical time axis, a maximum stretch constraint marker, and an alignment convergence marker. The cross-viewpoint consistency index at least includes the viewpoint number, resampling scale identifier, clipping boundary identifier, and canonical mesh mapping identifier, used to ensure that the projected segment sequence maintains the same spatial support domain in subsequent 3D reconstruction stages. The calibration parameter set and the correction evidence package are bound together through weld point identifiers, canonical mesh identifiers, and canonical time axis identifiers, ensuring that any correction evidence can be traced back to the registration, calibration, and alignment configurations upon which it was generated. This binding mechanism enables the backtracking of specific calibration links and corresponding perspectives and time periods when anomalies occur in subsequent stages, improving the auditability and operability of the process. Visual uncertainty, projection uncertainty, and thermoelectric uncertainty are generated based on the residual information from the registration, calibration, and alignment correction processes. Specifically, the residual information from registration, calibration, and alignment is extracted from the core consistency constraints of each process, forming three types of residual sources: visual residual, projection residual, and temporal residual. The visual residual is obtained based on the degree of alignment deviation between the boundary of the visual reference template and the boundary of the resampled visual segment. The degree of deviation is described by the average and maximum offsets from the boundary points to the boundary curves, avoiding the masking of local mismatches by a single average quantity. The projection residual is obtained based on the difference between the predicted projection of the nominal 3D prior body under the obtained viewpoint geometry and the resampled projected segment. The difference measure includes both structural contour deviation and intensity attenuation trend deviation to distinguish between geometric mismatch and imaging noise. The temporal residual is obtained based on the deviation of the morphological consistency between the thermoelectric segment after time mapping and the standard time axis. The degree of deviation is described by the alignment error of key event points and the morphological difference within the event window, and is corrected by the degree of violation of monotonicity and maximum stretch constraints. When generating uncertainty based on residuals, a monotonic mapping strategy is used to convert the residuals into a reliable scale of the same dimension. The mapping strategy satisfies the condition that uncertainty increases with increasing residuals and saturates within a reasonable range. The upper limit of saturation is jointly determined by the upper limit of equipment noise and the allowable fluctuation range of the process, avoiding the numerical dominance of abnormal residuals in subsequent fusion. Visual uncertainty is bound to visual resampling segments, projection uncertainty is bound to projection segments at each viewpoint, and thermoelectric uncertainty is bound to thermoelectric resampling segments. These are written into the corresponding fields in the correction evidence package according to the weld point identifier and viewpoint number. This uncertainty generation method explicitly encodes the correction difficulty into an available quantity. Subsequent voxel reconstruction can use this to weight and organize observations, and mechanism inference can use this to distinguish the credibility differences of evidence, thereby reducing the interference of weak evidence on mechanism conclusions and risk ranges.

[0027] The present invention is further configured such that the self-supervised learning to generate structural embeddings based on the correction evidence package includes: Solder joint samples are constructed based on the correction evidence package. These samples include visual segments, X-ray projection segment sequences, and thermoelectric time segments. Specifically, the correction evidence package reads the visual segments, X-ray projection segment sequences, and thermoelectric time segments according to the solder joint identifier to form solder joint samples. The samples retain standardized grid identifiers, viewpoint number lists, and standardized time axis identifiers to ensure traceability and consistency of the three types of inputs in both the spatial and temporal support domains. Visual segments undergo intensity normalization and boundary-fidelity rescaling, with the normalization range determined by the camera's dynamic range. Projection segment sequences are stacked in viewpoint number order to form a unified tensor representation, retaining viewpoint position information as sequence position encoding during stacking to avoid feature drift caused by viewpoint order perturbations. Thermoelectric time segments undergo amplitude normalization and sampling rate consistency according to the standardized time axis, with the normalization benchmark given by the test configuration. This consistency processing allows subsequent self-supervised objectives to focus on solder joint structure rather than dimensional differences, reducing the risk of the network mislearning exposure, gain, or sampling scale as structural differences. Masking reconstruction constraints are applied to solder joint samples to generate visual intermediate representations, X-ray intermediate representations, and thermoelectric intermediate representations. Specifically, a masking reconstruction task is performed on the solder joint samples, employing a structure-sensitive masking strategy: in visual segments, priority is given to masking the neighborhood of solder pad boundaries and areas suspected of being covered by solder; in projection segment sequences, priority is given to masking areas with significant changes in projection thickness and boundary transition areas; and in thermoelectric time segments, priority is given to masking slope abrupt change segments within key event windows. The mask ratio is set within a preset range and adaptively adjusted according to sample quality. Sample quality is determined by the uncertainty output during the calibration phase; when the uncertainty is high, the mask ratio is reduced to ensure the reconstruction task is learnable, while when the uncertainty is low, the mask ratio is increased to enhance the task difficulty. The encoder uses a three-branch structure to extract features from the three types of inputs and output intermediate representations. The decoder reconstructs the intermediate representations modally, calculating the reconstruction error only for the masked areas. The reconstruction error uses a cost metric that is more sensitive to boundaries to avoid only fitting a smooth background. Through the above masking reconstruction, visual intermediate representations, X-ray intermediate representations, and thermoelectric intermediate representations are obtained, and the intermediate representations are associated with solder joint identifiers. This process, driven by missing data completion, forces the representation to capture recoverable structural components, reducing its dependence on random noise and local textures. Cross-modal consistency constraints are applied to visual intermediate representations, X-ray intermediate representations, and thermoelectric intermediate representations to generate visual embeddings, X-ray embeddings, and thermoelectric embeddings. Specifically, cross-modal consistency constraints are introduced based on the intermediate representations, and a contrastive learning mechanism is used to construct positive and negative sample relationships: the correspondences between visual and projection, visual and thermoelectric, and projection and thermoelectric under the same solder joint identifier are considered positive pairs, while the correspondences between different solder joint identifiers are considered negative pairs. A hard example mining strategy is used for negative pair sampling, prioritizing solder joints with similar adjacency relationships, similar pad sizes, or similar process data items in the geometric prior, to avoid the model relying solely on low-difficulty cues such as position or size for differentiation. Contrastive learning outputs visual embeddings, X-ray embeddings, and thermoelectric embeddings with a unified dimension. The embedding vectors are represented under unit norm constraints, which are used to stabilize scale drift across different batches. This consistency constraint aligns the representations of the same solder joint structure across different modalities in the semantic space while maintaining separability for nearest-neighbor solder joints, providing cross-modal consistent structural condition inputs for subsequent 3D reconstruction. Structural embeddings are generated based on visual embeddings, X-ray embeddings, and thermoelectric embeddings. Specifically, these embeddings are input into a fusion network to generate structural embeddings. The fusion network employs a gated attention mechanism: a reliability gating weight is calculated for each modal embedding, driven by both the uncertainty of the corresponding modality and the sequence integrity label. The weight of the modality is reduced when the uncertainty is high or a missing label is triggered. Under the influence of these weights, cross-modal attention convergence is performed to obtain the structural embedding. Redundancy removal constraints are applied to the structural embeddings to suppress the separability of the embeddings based on differences in acquisition configurations. Redundancy removal constraints are achieved through adversarial discrimination of acquisition configuration labels within training batches, making the structural embeddings more focused on expressing the weld joint structural information. A one-to-one correspondence is established between the structural embeddings and the weld joint identifiers, serving as conditional inputs and outputs for subsequent voxel reconstruction. Simultaneously, the three types of modal embeddings are retained as inputs for subsequent generation of mechanism conflict degrees. This fusion method allows the structural embeddings to be compensated by other modalities even when evidence is missing or a single modality degenerates, forming a stable structural expression under uncertainty constraints.

[0028] The present invention is further configured such that the voxel field generated based on the resampled weld point X-ray projection segment sequence, calibration parameters, geometric priors, and structural embeddings in the correction evidence package includes: Based on the local geometric prior of the solder joint, a 3D voxel mesh and a support mask are determined. The support mask defines the feasible occupied area within the voxel mesh. Specifically, the local geometric prior of the solder joint reads the pad boundary, solder mask window boundary, nominal center, and local direction information. Based on the nominal center and local direction, a local coordinate system for the solder joint is established in 3D space. Under this coordinate system, the spatial range of the 3D voxel mesh is determined by the joint expansion of the pad shape and the window shape, and covers the height range where solder accumulation and interface transition can occur in the thickness direction. The voxel resolution is set according to the principle that the lateral resolution is consistent with the projected pixel resolution, and the longitudinal resolution is matched with the effective resolution of the ray penetration direction, so that the voxel scale is on the same order of magnitude as the observation resolution, avoiding the increase of underdetermined degrees of freedom due to excessive voxel density or the smoothing of small defects due to excessive voxel coarsness. The support mask is generated based on local geometric priors: the solder pad entity corresponds to the allowed area, the solder mask cover and forbidden area correspond to the disallowed area, and the non-solder area within the extended safety boundary corresponds to the low-confidence allowed area. The allowed area, low-confidence allowed area, and forbidden area are written using hierarchical mask encoding, which corresponds to strong constraints, soft constraints, and hard constraints in subsequent iterations, respectively. This construction method explicitly embeds the geometrically feasible domain into the solution domain, reducing the probability of geometric out-of-bounds structures during reconstruction and providing a clear spatial support range for subsequent data consistency constraints. Based on the viewpoint geometry parameters in the calibration parameter set, the forward projection relationship for each viewpoint is constructed. The corrected X-ray projection segments are then organized into an observation sequence according to the cross-viewpoint uniformity index. Specifically, the calibration parameter set reads the viewpoint geometry parameters and the cross-viewpoint uniformity index. The viewpoint geometry parameters include the relative pose of the X-ray source and detector, the projection scale, and the principal point offset. The cross-viewpoint uniformity index is used to ensure that the projection segments for each viewpoint are organized under the same standard grid scale. When constructing the forward projection relationship based on the viewpoint geometry parameters, a discrete ray integral model is used to establish the mapping from voxels to projected pixels: each projected pixel corresponds to a set of ray paths traversing the voxel, and the path length serves as the contribution weight of that voxel to that pixel; the ray path is determined by the pose configuration, and the contribution weight is determined by the voxel scale and the ray traversal length. The corrected X-ray projection segments are organized into an observation sequence according to the cross-viewpoint uniformity index order, and the viewpoint number, projection clipping boundary, and resampling scale are retained in the sequence, allowing subsequent solutions to trace back to any viewpoint observation. This organization method ensures that observations under conditions with few views still have a consistent geometric interpretation path, avoiding mistaking local clippings from different views as repeated observations from the same viewpoint. Observation weights are generated based on the uncertainties of the resampled solder joint X-ray projection segment sequence from the correction evidence package. Data consistency constraints are then formed based on these observation weights, the forward projection relationship, and the observation sequence. Specifically, the uncertainty input uses projection uncertainty, which is obtained by monotonically mapping the projection residuals from the previous correction stage and corresponds one-to-one with the viewpoint number. Observation weights are generated based on projection uncertainty, using a mapping rule of low weight for high uncertainty and high weight for low uncertainty. A lower limit for the weights is also set to prevent a single viewpoint from being completely shielded, which would cause the reconstruction to lose observational support. After the weights are generated, a predicted projection is generated for the current voxel candidate structure based on the forward projection relationship, and the consistency deviation is obtained by comparing it with the observation sequence view by view. The consistency deviation forms data consistency constraints under the weights, and the constraints include two layers: the first is an overall consistency constraint, used to control the deviation between the predicted projection and the observed projection in terms of intensity decay trends; the second is a structural consistency constraint, used to control the deviation of key structural positions such as solder pad boundaries and solder transition areas in the projection domain. The key structural positions are determined by the mapping of local geometric priors in the projection domain. This constraint design focuses on physical projection consistency while explicitly introducing uncertainty to prevent weakly reliable observations from dominating the solution and reduce bias propagation caused by noise, scattering, or beam hardening. Based on structural embedding and local geometric priors of the solder joints, conditional prior constraints are generated. Voxel fields are then generated within the voxel mesh defined by the support mask, based on data consistency constraints and conditional prior constraints. Specifically, structural embedding, as a conditional input, works with the local geometric priors of the solder joints to construct conditional prior constraints, which are implemented using a generative prior model. In the offline phase, the generative prior model is trained on the voxel structures of the solder joints. The training inputs are the structural embedding and the local geometric priors, and the training objective is to output a voxel structure distribution that satisfies the geometrically feasible region. Training samples are derived from high-quality reconstruction results of labeled samples or a statistical voxel structure library of process benchmark samples, thus enabling the prior model to express solder joint structure patterns. In the online solution phase, the prior model outputs a prior consistency metric for the current voxel candidate structure based on the structural embedding. This metric penalizes candidate structures with geometric out-of-bounds errors, interface discontinuities, or structures that do not conform to typical solder joint patterns, and employs soft penalties in the low-confidence allowance region to preserve expressiveness under process anomalies. The voxel field generation employs a plug-and-play iterative strategy: in each iteration, the voxel candidate structure is first updated based on data consistency constraints, gradually approximating the observed sequence through predicted projections; then, the voxel candidate structure is corrected based on conditional prior constraints, ensuring the structure conforms to the structural patterns expressed by geometric priors and structural embedding; subsequently, support mask constraints are applied, forcing voxels in forbidden regions to zero, keeping voxels in allowed regions non-negative, and applying soft gating to voxels in low-confidence allowed regions to suppress drift without observational support. The iteration termination condition uses a multi-condition joint determination, including the decrease in data consistency deviation reaching a preset threshold, the voxel update magnitude reaching a preset threshold, or reaching the maximum number of iterations, and the termination reason is recorded for subsequent traceable verification. This solution process, through the synergy of observational consistency driving, conditional prior constraints, and geometric mask feasible regions, can still obtain a structurally interpretable, spatially feasible, and observationally consistent voxel field under limited viewpoint conditions, providing a structural-level evidence basis for subsequent mechanism inference.

[0029] The present invention is further configured such that the uncertainty of the generated voxel includes: Voxel information parameters are generated based on forward projection relationships, observation weights, and the voxel field. Specifically, the forward projection relationship is based on the ray traversal relationship from the voxel to the projected pixel, the observation weights reflect the confidence level of the projection from each viewpoint, and the voxel field represents the currently reconstructed 3D structure. The generation of voxel information parameters employs a local observability measurement strategy: for each voxel, the number of times it is covered by rays in each viewpoint and the length of the coverage path are counted, and the coverage contribution is weighted and accumulated according to the observation weights of the corresponding viewpoints. The number of coverages characterizes how many observation constraints the voxel is under, the path length characterizes the voxel's sensitivity to the projection response, and the weighted accumulation is used to suppress spurious increases in observability from low-confidence viewpoints. To avoid bias from relying solely on the number of coverages, a structural sensitivity screening is introduced. This means that only voxels in the voxel field located in the solder-interface transition region, areas where voids may accumulate, or the neighborhood of geometric boundaries receive increased information content. These regions are jointly determined by the boundary mapping of the local geometric prior of the solder joint and the gradient changes in the voxel field, thus making the information parameters more focused on structural locations related to defect identification. The generated voxel information parameters are written one-to-one with the solder joint identifier and voxel coordinate index, ensuring that subsequent voxel uncertainties can be directly located and verified in voxel space. This process explicitly quantifies the strength of the observation constraint on the voxel, enabling the identification of underdetermined regions with few viewing angles and their conservative handling in subsequent mechanism inference and risk assessment. Prior information parameters are generated based on conditional prior constraints and voxel fields. These parameters are then associated with solder joint identifiers. Specifically, the conditional prior constraints are driven by structural embedding and local geometric priors of the solder joints, with the voxel field serving as the current candidate structure. The generation of prior information parameters employs a prior local rigidity metric strategy: at each voxel position in the voxel field, the consistency contribution of that voxel to the conditional prior model output is evaluated, i.e., the magnitude of change in the prior consistency metric when the voxel undergoes a small perturbation. A larger magnitude indicates a stronger prior constraint on the voxel position, while a smaller magnitude indicates that the prior allows for greater structural freedom for the voxel. To ensure feasibility, the magnitude of the prior consistency change is evaluated using a discrete perturbation method: the voxel value is subjected to finite-amplitude forward and reverse perturbations, and the difference in the prior consistency metric is calculated. This difference magnitude is used as the prior information intensity of the voxel. The perturbation amplitude is limited by the range of voxel field values ​​and the upper bound of the reconstruction noise to prevent excessive perturbation from causing nonlocal structural changes. For prohibited regions within the support mask, the prior information is forcibly set to a high value to reflect hard constraints; for low-confidence allowable regions, the prior information is attenuated according to the reliability gating of the structural embedding, ensuring that necessary structural representation space is retained even in the event of process anomalies. A one-to-one correspondence is established between the generated prior information parameters and the solder joint identifiers and voxel coordinate indices, and these are stored aligned with the voxel information parameters within the same voxel grid. This process independently characterizes the strength of prior constraints on voxels, enabling subsequent uncertainties to distinguish between two sources: insufficient data constraints and insufficient prior constraints. Voxel uncertainty is generated based on voxel information parameters and prior information parameters. Specifically, the generation of voxel uncertainty adopts an inverse information-proportional mapping and robust fusion strategy: when both voxel information and prior information are high, the voxel uncertainty is set to a lower level; when one of the two types of information is low, the voxel uncertainty is increased according to a more conservative rule. During fusion, lower and upper limits are set for voxel information and prior information, respectively. The lower limit is determined by the numerical stability requirements, and the upper limit is determined by the number of observation perspectives and the saturation range of the prior model, to prevent extreme values ​​from causing abnormal contraction or amplification of uncertainty. The voxel uncertainty output is in the form of a three-dimensional array with the same resolution as the voxel field, and is bound to the weld point identifier and voxel grid index. At the same time, the observation weight version number and the prior model version number used during generation are recorded to ensure consistent reference in subsequent calculations of mechanism free energy, residuals, and conflict degree. This generation method makes voxel uncertainty a direct carrier of the credibility of structural layer evidence. Subsequent mechanism inferences can reduce the explanatory weight of underdetermined regions based on this, and risk interval generation can expand the conservative boundary based on this, thereby improving the controllability and traceability of the judgment.

[0030] The present invention is further configured such that the generation mechanism free energy, residual, and conflict degree include: A pre-defined set of defect mechanisms is used. Based on calibration evidence, voxel fields, and voxel uncertainties, mechanistic free energy sequences are generated under various mechanistic conditions within this set. Specifically, the defect mechanism set is pre-configured as an enumerable list of mechanisms, with each mechanism bound to a mechanism parameter template and a mechanism condition generation model. The mechanism parameter template defines the morphological priors, interface constraints, and thermal response priors for that mechanism, while the mechanism condition generation model generates interpretable observational predictions under those conditions. Calibration evidence inputs include visual segments, calibration projection segment sequences, and calibration thermoelectric time segments from the calibration evidence package. Structural evidence inputs use voxel fields, and credible evidence inputs use voxel uncertainties. Voxel uncertainties are weighted to form a voxel credible weight field, causing the contribution of underdetermined voxels to the mechanism assessment to decay according to their credibility. For each mechanism, mechanism condition inference is performed: first, the allowable defect configuration region and interface position constraints are determined based on geometric priors; then, structural and thermal predictions under this mechanism are generated based on mechanism parameter templates, and the observation interpretation cost is calculated under the constraint of a voxel confidence weight field; simultaneously, the mechanism prior distribution is generated based on process data items and geometric priors, and the deviation cost between the inferred mechanism posterior and the mechanism prior is included in the mechanism score. The observation interpretation cost and the prior deviation cost are synthesized to obtain the mechanism free energy of the mechanism, which is then aggregated in the order of the mechanism list to generate a mechanism free energy sequence, and the mechanism free energy sequence is bound to the weld point identifier. This process uses the mechanism condition generation model as a unified semantic carrier, enabling different mechanisms to be compared at the same scale, and voxel uncertainty participates in weighting, reducing the misleading effect of underdetermined regions on mechanism ranking, and improving the repeatability and auditability of mechanism inference. Based on the resampled weld point X-ray projection segment sequence and voxel field in the calibration evidence package, a projection consistency residual is generated. Based on the geometric prior and voxel field, a geometric consistency residual is generated. Based on the calibrated thermoelectric sequence and mechanistic conditions, a temporal consistency residual is generated. Based on the projection consistency residual, geometric consistency residual, and temporal consistency residual, a physical residual intensity is generated. Specifically, when generating the projection consistency residual, the viewpoint geometric parameters and cross-viewpoint consistency index in the calibration parameter set are read, and a predicted projection is generated based on the forward projection of the voxel field at each viewpoint. The predicted projection and the calibrated projection segment are compared viewpoint by viewpoint to obtain the viewpoint residual. Based on the projection uncertainty, a viewpoint confidence weight is generated. A confidence weighted convergence is performed on the viewpoint residual to obtain the projection consistency residual intensity and the viewpoint residual record. During geometrically consistent residual generation, the pad boundaries, solder mask window boundaries, and support mask definitions from the geometric prior are read, and a feasible region verification is performed on the voxel field: hard violation costs are incurred for non-zero voxel occupancy within the prohibited occupancy region, and soft violation costs are incurred for anomalous connectivity or anomalous expansion within the low-confidence allowable region. The hard and soft violation costs are then combined to obtain the geometrically consistent residual strength. During temporally consistent residual generation, the thermal response prediction output by the model is generated based on mechanistic conditions. After morphological alignment of the corrected thermoelectric time segments, the difference strength is calculated. The difference metric simultaneously covers the timescale deviation of key event points and the slope deviation within the event window. A confidence weighting is applied to the difference strength based on thermoelectric uncertainty to obtain the temporally consistent residual strength. The projection-consistent residual strength, geometrically consistent residual strength, and temporally consistent residual strength are aggregated according to preset weights to generate the physical residual strength. The weights are jointly managed by the number of viewpoints, voxel uncertainty statistics, and thermoelectric uncertainty, increasing the proportion of projection channels when observations are sufficient and increasing the proportion of geometric and temporal channels when underdetermined conditions are significant. This process forms a three-channel consistency verification of structure, geometry and time, with clear sources of residuals and constraints, and can be directly traced back to the corresponding viewpoint, corresponding voxel region and corresponding time window, reducing the risk of misjudgment caused by single-channel distortion. The conflict degree is generated based on the distribution of mechanistic evidence (visual, X-ray, and thermoelectric) across the defect mechanism set. The mechanistic evidence distribution is generated from the mechanistic free energy of the corresponding mode. Specifically, the mechanistic free energy sequence is decomposed into modal mechanistic free energies: the visual channel uses only visual segments and mechanistic condition appearance prediction to calculate the interpretation cost, superimposed with a priori deviation cost; the projection channel uses only corrected projection segment sequences and mechanistic condition projection prediction to calculate the interpretation cost, superimposed with a priori deviation cost; and the thermoelectric channel uses only corrected thermoelectric time segments and mechanistic condition thermal prediction to calculate the interpretation cost, superimposed with a priori deviation cost. All three modal mechanistic free energies are calculated under voxel confidence weight fields and corresponding modal uncertainty weight constraints, ensuring that the influence of low-confidence evidence on distribution sharpness is suppressed. The mechanistic free energy of each mode is mapped to the mechanistic evidence distribution through temperature smoothing: mechanisms with smaller free energies receive higher probabilities. The temperature parameter is adaptively scheduled by the uncertainty of that mode; when the uncertainty is high, the temperature increases to reduce overconfidence in the distribution, and when the uncertainty is low, the temperature decreases to enhance mechanism separability. The conflict degree is generated by the divergence among the distributions of the three-modal mechanistic evidence: using the average distribution of the three modalities as a reference, the divergence between each modal distribution and the average distribution is calculated and averaged. Symmetric information divergence is used to ensure that the divergence measure is independent of modal order. Simultaneously, distribution degradation processing is performed on missing modalities, distributing them into a uniform distribution and reducing their weight during conflict degree convergence to avoid artificially inflated conflict degrees due to missing modalities. The conflict degree is output and bound to the solder joint identifier for use in the subsequent generation of evidence coupling potential coefficients. This process uses the divergence of the mechanistic evidence distribution as the definition of the conflict degree, ensuring strict consistency between the conflict characterization and the semantics of the mechanism set. Furthermore, the uncertainty scheduling mechanism guarantees that the conflict degree reflects genuine cross-modal contradictions while suppressing spurious conflicts caused by weakly credible evidence.

[0031] The present invention is further configured such that the generation of evidence coupling potential energy coefficients to determine the solder joint mechanism conclusions includes: The optimal and suboptimal defect mechanisms are determined based on the mechanistic free energy sequence. Free energy gaps are then generated based on these mechanisms. Specifically, the input mechanistic free energy sequence uses a score list arranged by mechanism from the previous stage. This score is a synthesis of the observational interpretation cost and the prior deviation cost; a smaller value indicates greater consistency in interpreting multi-source evidence under that mechanism. The mechanistic free energy sequence is sorted, and the mechanism with the smallest score is selected as the optimal defect mechanism. Simultaneously, the mechanism with the smallest score (excluding the optimal defect mechanism) is selected as the suboptimal defect mechanism. The original indices and corresponding scores of both mechanisms in the mechanism set are retained for traceability. The free energy gap is obtained from the difference in mechanistic free energy between the optimal and suboptimal mechanisms, and normalized according to the discrete scale of the mechanistic free energy sequence to maintain comparability across batch scale variations. To avoid incomparability of gaps due to scale variations between different batches or samples, the free energy gap is normalized according to the discrete scale of the mechanistic free energy sequence. The discrete scale is jointly determined by the median and the dispersion amplitude of the mechanistic free energy sequence, giving the gap metric cross-batch stability. This process uses the separability of the mechanism as the core input for subsequent coupling strength, avoiding the direct output of the mechanism based solely on the minimum value while ignoring the proximity of the suboptimal competing mechanisms, thereby improving the reliability of the mechanistic conclusions. Evidence coupling potential coefficients are generated based on free energy gap, physical residual strength, and conflict degree. Specifically, the input of evidence coupling potential coefficients adopts three types of quantities: free energy gap, physical residual strength, and conflict degree. These three quantities respectively characterize mechanism separability, mechanism explanation feasibility, and cross-modal consistency. The evidence coupling potential coefficient is obtained by fusing the free energy gap, physical residual strength, and conflict degree. The fusion rule satisfies that the potential coefficient increases with the increase of the free energy gap, decreases with the increase of the physical residual strength, and decreases with the increase of the conflict degree. After fusion, it is limited to the range of zero to one by monotonically compressing mapping, and saturation limit is triggered for extreme residuals or extreme conflicts to prevent the potential coefficient from falling into the unusable range. The potential coefficient is generated using a monotonically fusion strategy: the potential coefficient increases with the increase of the free energy gap, decreases with the increase of the physical residual strength, and decreases with the increase of the conflict degree. Scale alignment and saturation boundary are set during fusion. Scale alignment is determined by historical calibration samples or reference intervals given by process specifications, so that the three types of quantities have comparable influence strength in the same numerical domain. The saturation boundary is used to limit the excessive suppression of the potential coefficient by extreme residuals or extreme conflicts, and to avoid the potential coefficient being pulled into the unusable range by individual viewpoint noise or single mode degradation. To enhance traceability, the potential energy coefficient is simultaneously output with its component contribution markers. These markers record the weight configuration version numbers of the free energy gap channel, residual channel, and conflict channel, as well as the status of whether saturation limits have been triggered. This allows subsequent risk interval generation to reuse the same scheduling caliber and support verification. As a unified carrier of evidence coupling strength, this potential energy coefficient can be directly used for subsequent risk interval-scale scheduling, achieving continuous quantity transfer from mechanism inference to judgment output and reducing instability caused by threshold-based hard switching. The solder joint mechanism conclusion is generated based on the optimal defect mechanism. Specifically, the solder joint mechanism conclusion uses the optimal defect mechanism as the mechanism output, and simultaneously binds the evidence coupling potential energy coefficient and free energy gap to the solder joint identifier, forming the mechanism conclusion output structure. To ensure sufficient disclosure and engineering reproducibility, the output structure retains the identifiers, corresponding scores, physical residual strength, and conflict degree of both the optimal and suboptimal defect mechanisms as accompanying evidence for the mechanism conclusion. This accompanying evidence does not participate in the repeated generation of subsequent structure calculations; it is only used for verification and traceability. This encapsulation method allows subsequent stages to directly call the mechanism conclusion and potential energy coefficient for scale scheduling when generating risk intervals. It also allows for locating the source of mechanism uncertainty when verification is required, such as insufficient mechanism separability, insufficient physical consistency, or enhanced cross-modal conflict, thereby improving the interpretability and auditability of the judgment chain.

[0032] The present invention is further configured such that the output judgment after generating the risk interval includes: Based on the solder joint mechanism conclusions, mechanism-sensitive operators corresponding to the solder joint mechanism conclusions are selected from a pre-defined set of mechanism-sensitive operators. The voxel field is then mapped using these operators to generate a mechanism-sensitive field. Specifically, the set of mechanism-sensitive operators is pre-defined as an indexable list based on the defect mechanism. Each mechanism corresponds to a set of targeted operator chains, which include at least three types of processing: spatial support constraint, structural enhancement, and topological metric. Spatial support constraint is determined by geometric priors and is calculated only within the feasible occupied areas of the pad boundaries and window boundaries, avoiding the introduction of voxel perturbations outside the support domain into the risk assessment. Structural enhancement selects a combination of three-dimensional multi-scale filtering and morphological reconstruction based on the mechanism differences. For example, for void-type mechanisms, emphasis is placed on enhancing the connectivity of low-density clusters; for crack-type mechanisms, emphasis is placed on enhancing the directional consistency of slender structures; for insufficient wetting mechanisms, emphasis is placed on boundary enhancement of interface continuity; and for bridging mechanisms, emphasis is placed on enhancing the path continuity across pad gaps. The topological metric employs algorithms such as voxel adjacency graph-based connectivity decomposition, geodesic distance-based shortest path search, and skeletonized slender structure extraction to map the voxel field into a mechanism-sensitive field. This mechanism-sensitive field is output with an index consistent with the voxel grid, ensuring that subsequent uncertainty propagation and region-weighted aggregation can directly reuse the same spatial coordinate system. This process enables risk assessment to use mechanism-driven structural quantities as the entry point, reducing reliance on general statistics and ensuring semantic consistency between mechanism conclusions and risk metrics. Mechanism damage surrogate quantity and surrogate quantity uncertainty are generated based on the mechanism-sensitive field and voxel uncertainty. Specifically, the mechanism damage surrogate quantity is generated on the mechanism-sensitive field using a key region weighted aggregation strategy. The key region is jointly determined by the pad boundary, interface neighborhood bandwidth, and adjacent solder joint interval given by geometric priors. Within the region, a combination of quantile aggregation and connectivity-weighted aggregation is performed on the mechanism-sensitive field: quantile aggregation emphasizes extreme defect signs rather than average levels, while connectivity-weighted aggregation emphasizes the spatial continuity of defects rather than scattered noise. Voxel uncertainty participates in aggregation in the form of reliable weights. The weight mapping satisfies the condition that the weight decreases as the uncertainty increases, while a lower limit for the weight is set to avoid underdetermined regions being completely ignored, leading to a systematic underestimation of mechanism damage. The surrogate quantity uncertainty is generated using a consistent propagation caliber: for voxel regions assigned high weights during aggregation, structural uncertainty is formed based on the accumulation of voxel uncertainty; for the path search and skeleton extraction process of connected structures, algorithm-side uncertainty is formed based on path stability and multi-scale consistency; the two types of uncertainty are output as the surrogate quantity uncertainty in a conservative fusion manner. This process explicitly transfers structural underdeterminacy to the risk interval construction, avoiding overly narrow risk conclusions when the quality of evidence is insufficient. Based on the weld joint mechanism conclusions, mechanistic damage surrogate quantity, and surrogate quantity uncertainty, risk point estimates and model uncertainties are generated. Specifically, risk point estimates are generated according to the mechanism, and a mechanism-conditional risk mapper is used to map the mechanistic damage surrogate quantity to risk points. The risk mapper is pre-calibrated offline, and the training inputs use the mechanistic damage surrogate quantity, surrogate quantity uncertainty, and key condition fields from the process data. The training labels use quality judgment results or reliability indicators. The model family uses regressors with probability output capabilities, which at least satisfy the requirement of outputting two types of information: center estimates and uncertainty characterization. To distinguish between instability caused by data sparsity and model extrapolation, model uncertainty is obtained using Bayesian ensemble or Gaussian process methods: Bayesian ensemble measures extrapolation risk through multi-submodel consistency, while Gaussian process methods measure prediction reliability through similar sample density. When the uncertainty in the correction stage is high or the evidence coupling potential coefficient is low, the model uncertainty adopts an adaptive amplification rule to avoid risk points showing overconfidence in weak evidence. This processing ensures that risk points not only reflect defect intensity but also carry model-level reliability boundaries, providing necessary input for subsequent interval-based judgment. Interval scale parameters are generated based on risk point estimation, model uncertainty, surrogate uncertainty, and evidence coupling potential coefficient. Risk intervals are then generated based on these parameters. Specifically, the interval scale parameters are jointly generated by model uncertainty, surrogate uncertainty, and evidence coupling potential coefficient. Scale fusion employs a component separation and robust superposition strategy: model uncertainty characterizes the instability of the mapper's extrapolation from surrogate to risk points, while surrogate uncertainty characterizes voxel structure underdetermination and algorithm propagation error. Both are first scale-aligned within a unified scaling domain, with the alignment benchmark provided by the offline calibration set. The evidence coupling potential coefficient is incorporated into the scale parameters as a monotonic scheduling factor. When the potential coefficient decreases, the scale parameters expand according to conservative rules, with the expansion upper limit limited by the maximum uncertainty interval allowed by the process specifications, preventing the scale from expanding indefinitely and losing its judgment value. Risk intervals are constructed jointly by risk points and interval scale parameters. The interval coverage level is determined by a pre-set confidence level, which is provided by the quality strategy or verification specifications and bound to the coverage consistency constraints of the offline calibration set, ensuring that the intervals have verifiable statistical meaning. This process enables the transformation from point estimation to interval representation, allowing the judgment to explicitly reflect the strength of evidence coupling and the degree of structural indeterminacy. The system outputs a judgment label based on the risk range and a preset set of judgment thresholds. Specifically, the preset set of judgment thresholds includes at least two levels, corresponding to the release boundary and the disposal boundary, respectively. These thresholds are derived from quality standards or process control specifications. Judgment is not based on direct comparison of risk points, but rather on the relative position of the upper and lower boundaries of the risk range with respect to the threshold set: a release label is output when the overall risk range is below the release boundary, a disposal label is output when the overall risk range is above the disposal boundary, and a review label is output in other cases. When evidence is missing or the potential coefficient triggers conservative amplification, the review label is output first to ensure process consistency. The output results are linked to the solder joint identifier, along with the solder joint mechanism conclusion, evidence coupling potential coefficient, risk range boundary, and confidence level marker used for range construction. This allows subsequent reviews to directly pinpoint the source of risk as an anomaly in the mechanism-sensitive field, high voxel uncertainty, high model uncertainty, or insufficient evidence coupling. This judgment method replaces point-based decision-making with range-based decision-making, ensuring that uncertainty is explicitly used at the decision-making level, thereby improving the consistency and controllability of the judgment.

[0033] 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 chip soldering inspection method, characterized in that, include: S1: Acquire visual images, X-ray projections, thermoelectric sequences, process data, and geometric priors to generate weld point evidence; S2: Perform registration, calibration, and alignment correction on the weld point evidence based on geometric priors, and generate a correction evidence package, uncertainty, and calibration parameter set; S3: Generate structural embeddings through self-supervised learning based on the correction evidence package; S4: Generate voxel fields and voxel uncertainties based on the resampled weld point X-ray projection fragment sequence, calibration parameter set, geometric priors, and structural embeddings in the correction evidence package; S5: Preset a set of defect mechanisms, generate mechanism free energy, residuals and conflict degree based on correction evidence, voxel field and voxel uncertainty, generate evidence coupling potential energy coefficient based on mechanism free energy, residuals and conflict degree, and determine the weld joint mechanism conclusion; S6: After generating the risk range based on the weld joint mechanism conclusion, voxel field, voxel uncertainty and evidence coupling potential coefficient, output the judgment.

2. The chip welding inspection method according to claim 1, characterized in that, Evidence of solder joints includes: Based on the geometric prior, the solder joint identifier and nominal location are determined, and the solder joint index relationship is established. The geometric prior generates the local geometric prior of the solder joint according to the solder joint identifier. Based on the nominal location of the solder joint, extract the solder joint image fragment from the visual image; based on the nominal location of the solder joint, extract the solder joint projection fragment sequence from the X-ray projection; based on the solder joint identifier, extract the solder joint time fragment from the thermoelectric sequence. Extract solder joint process data items from the process data based on the solder joint identifier; Solder joint evidence is generated by associating solder joint identifiers, solder joint image fragments, solder joint projection fragment sequences, solder joint time segments, solder joint process data items, and local geometric priors of the solder joint.

3. The chip welding inspection method according to claim 2, characterized in that, Performing registration, calibration, and alignment correction based on weld point evidence according to geometric priors includes: Based on the prior analysis of the local geometry of the solder joint, the nominal center of the solder joint, the local orientation base, the boundary of the solder pad, and the boundary of the solder mask opening, a standard mesh for the solder joint is established. A visual reference template is generated based on the prior local geometry of the weld joint. Prior constraint registration is performed on the visual fragments of the weld joint based on the visual reference template to obtain the visual registration parameters. The visual fragments of the weld joint are then resampled to the weld joint specification mesh. Based on the local geometric prior of the weld joint, a nominal three-dimensional prior volume of the weld joint is generated. Based on the nominal three-dimensional prior volume of the weld joint, the viewpoint geometry is calibrated on the X-ray projection segment sequence of the weld joint to obtain the viewpoint calibration parameters. The X-ray projection segment sequence of the weld joint is resampled to the weld joint standard mesh and a cross-viewpoint consistent index is generated. Perform event window alignment on the thermoelectric time segment of the solder joint to obtain time alignment parameters, and resample the thermoelectric time segment of the solder joint to the standard time axis.

4. The chip welding inspection method according to claim 3, characterized in that, The generation of the correction evidence package, uncertainty, and calibration parameter set includes: The resampled solder joint visual fragments, the resampled solder joint X-ray projection fragment sequence, and the resampled solder joint thermoelectric time fragments are encapsulated into a correction evidence package; Visual registration parameters, view calibration parameters, time alignment parameters, and cross-view consistency index are encapsulated into a set of calibration parameters; Visual uncertainty, projection uncertainty, and thermoelectric uncertainty are generated based on the residual information from the registration, calibration, and alignment correction processes.

5. The chip welding inspection method according to claim 1, characterized in that, Self-supervised learning to generate structural embeddings based on the corrected evidence package includes: Solder joint samples were constructed based on the correction evidence package. The solder joint samples included visual fragments, X-ray projection fragment sequences, and thermoelectric time fragments. Mask reconstruction constraints are applied to the solder joint samples to generate visual intermediate representations, X-ray intermediate representations, and thermoelectric intermediate representations. Perform cross-modal uniform constraints on visual intermediate characterization, X-ray intermediate characterization and thermoelectric intermediate characterization to generate visual embedding, X-ray embedding and thermoelectric embedding; Structural embedding is generated based on visual embedding, X-ray embedding, and thermoelectric embedding.

6. The chip welding inspection method according to claim 1, characterized in that, Based on the resampled weld joint X-ray projection fragment sequence, calibration parameters, geometric priors, and structure embedding generated voxel fields in the correction evidence package, including: The three-dimensional voxel mesh and support mask of the solder joint are determined based on the local geometric prior of the solder joint. The support mask defines the feasible occupied area within the voxel mesh. Based on the perspective geometry parameters in the calibration parameters, construct the forward projection relationship of each perspective, and organize the X-ray projection segment sequence of the weld point into an observation sequence by cross-perspective consistent indexing. Observation weights are generated based on the uncertainty of the X-ray projection segment sequence of the weld joint, and data consistency constraints are formed based on the observation weights, forward projection relationship, and observation sequence. Based on the structural embedding and the local geometry prior of the weld points, conditional prior constraints are generated. Based on the data consistency constraints and conditional prior constraints, a voxel field is generated within the voxel mesh defined by the support mask.

7. The chip welding inspection method according to claim 6, characterized in that, The uncertainty of voxel generation includes: Voxel information parameters are generated based on forward projection relationships, observation weights, and voxel fields. Generate prior information parameters based on conditional prior constraints and voxel fields; Voxel uncertainty is generated based on voxel information parameters and prior information parameters.

8. The chip welding inspection method according to claim 1, characterized in that, The generation mechanism free energy, residuals, and conflict degree include: A set of pre-defined defect mechanisms is used to generate a sequence of mechanistic free energy under each mechanism condition in the set of defect mechanisms, based on correction evidence, voxel field, and voxel uncertainty. Based on the X-ray projection segment sequence of the weld joint and the voxel field, a projection-consistent residual is generated; based on the geometric prior and the voxel field, a geometrically consistent residual is generated; based on the corrected thermoelectric sequence and the mechanism conditions, a temporally consistent residual is generated; and based on the projection-consistent residual, the geometrically consistent residual, and the temporally consistent residual, a physical residual intensity is generated. The conflict degree is generated based on the distribution of mechanistic evidence on the defect mechanism set, including visual evidence, X-ray evidence, and thermoelectric evidence. The distribution of mechanistic evidence is generated by the mechanistic free energy of the corresponding mode.

9. A chip welding inspection method according to claim 8, characterized in that, The conclusions drawn from generating the evidence coupling potential coefficient and determining the solder joint mechanism include: The optimal and suboptimal defect mechanisms are determined based on the mechanism free energy sequence, and the free energy gap is generated based on the optimal and suboptimal defect mechanisms. The evidence coupling potential coefficient is generated based on the free energy gap, physical residual strength, and conflict degree. The conclusion of the solder joint mechanism is generated based on the optimal defect mechanism.

10. A chip welding inspection method according to claim 9, characterized in that, After generating the risk range, the output judgment includes: Based on the solder joint mechanism conclusion, select the mechanism sensitive operator corresponding to the solder joint mechanism conclusion from the preset mechanism sensitive operator set, and map the voxel field to generate the mechanism sensitive field based on the mechanism sensitive operator; Based on the mechanism-sensitive field and voxel uncertainty, generate mechanism-damage surrogate quantity and surrogate quantity uncertainty; Based on the conclusions of the weld joint mechanism, the surrogate quantity of the mechanism damage and the uncertainty of the surrogate quantity, risk point estimates and model uncertainties are generated. The interval scale parameter is generated based on the risk point estimate, model uncertainty, surrogate quantity uncertainty, and evidence coupling potential coefficient. The risk interval is then generated based on the interval scale parameter. The judgment label is output based on the risk range and the preset judgment threshold set.

Images