An AOI defect compensation method, device and computer equipment

Abstract

This application relates to an AOI defect compensation method, apparatus, and computer equipment. The method includes: performing neural radiation field inversion on AOI multi-source data and multi-exposure real-world image data of the circuit board to be inspected to obtain physical characterization data of the board surface; constructing a condition set based on the physical characterization data of the board surface, using upstream process data of the circuit board to be inspected; performing condition-driven reasoning on the normative characterization synthesis model of the circuit board to be inspected based on the condition set to obtain a defect-free normative characterization; performing residual signature analysis on the physical characterization data of the board surface and the defect-free normative characterization to obtain defect analysis results; and performing closed-loop compensation control analysis on the compensation control object corresponding to the circuit board to be inspected based on the defect analysis results to obtain compensation action control data. This method can improve production line yield.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to an AOI defect compensation method, apparatus and computer equipment. Background Technology

[0002] In traditional AOI defect compensation methods, after positioning and alignment, the acquired image is first normalized for brightness / contrast, and background subtraction or simple filtering is performed to suppress illumination and noise fluctuations. Then, binarization with fixed or segmented thresholds is used to extract suspected defect areas. Combined with template comparison (such as differential / correlation matching with the "gold sample" image) or preset criteria such as size, grayscale, aspect ratio, and edge strength, the defect area is corrected, the boundary is expanded or shrunken, holes are filled, and isolated points are removed. Regular compensation is achieved through manually set shielding areas / tolerance zones (for reflective, screen printing, shadow, or process-permissible areas). However, traditional techniques rely heavily on human experience and fixed thresholds / rules, and are not robust enough to changes in illumination, material reflection, batch differences, or complex defect morphologies. They are prone to over-compensation or under-compensation, leading to false positives and false negatives, resulting in low production line yield. Summary of the Invention

[0003] Therefore, it is necessary to provide an AOI defect compensation method, apparatus, and computer equipment that can improve production line yield in response to the above-mentioned technical problems.

[0004] Firstly, this application provides an AOI defect compensation method, including: Neural radiation field inversion was performed on the AOI multi-light source data and multi-exposure real-shot image data of the circuit board under inspection to obtain physical characterization data of the board surface; Based on the physical characterization data of the board surface, a condition set is constructed for the upstream process data of the circuit board to be inspected, and a condition set is obtained. Based on the set of conditions, condition-driven reasoning is performed on the normative characterization synthesis model of the circuit board to be inspected to obtain a defect-free normative characterization. Residual signature analysis is performed on the physical characterization data of the plate surface and the defect-free normative characterization to obtain the defect analysis results; Based on the defect analysis results, closed-loop compensation control analysis is performed on the compensation control object corresponding to the circuit board under inspection to obtain compensation action control data.

[0005] Secondly, this application also provides an AOI defect compensation device, comprising: The multi-source data inversion module is used to perform neural radiation field inversion on the AOI multi-source data and multi-exposure real-shot image data of the circuit board under inspection to obtain physical characterization data of the board surface. The condition set construction module is used to construct a condition set for the upstream process data of the circuit board under inspection based on the physical characterization data of the board surface, and obtain the condition set. The condition-driven reasoning module is used to perform condition-driven reasoning on the normative characterization synthesis model of the circuit board to be inspected based on the set of conditions, so as to obtain a defect-free normative characterization. The residual signature analysis module is used to perform residual signature analysis on the physical characterization data of the board surface and the defect-free normative characterization to obtain the defect analysis results. The compensation control analysis module is used to perform closed-loop compensation control analysis on the compensation control object corresponding to the circuit board under inspection based on the defect analysis results, and obtain compensation action control data.

[0006] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of an AOI defect compensation method.

[0007] The aforementioned AOI defect compensation method, apparatus, and computer equipment obtain physical characterization data of the circuit board surface by performing neural radiation field inversion on AOI multi-source data and multi-exposure real-shot image data of the circuit board to be inspected. Based on this physical characterization data, a condition set consistent with upstream process data is constructed, thereby driving a normative characterization synthesis model to generate a defect-free normative characterization consistent with the current working conditions and imaging conditions. Finally, residual signature analysis of the physical characterization and the defect-free normative characterization replaces the traditional template difference / fixed threshold judgment and outputs the defect analysis results. The defect analysis results are mapped in a closed loop to the compensation control objects of AOI judgment boundary parameters and / or upstream process parameters to generate executable compensation action control data. This achieves decoupling and suppression of imaging drift and process drift, and interpretable and quantifiable identification of real defects. It significantly reduces false alarms and false negatives and improves detection stability across batches, machines, and lighting conditions. At the same time, the detection results are directly converted into dual closed-loop control of judgment compensation and process compensation, reducing the cost of re-judgment and re-inspection, shortening the process convergence time, and suppressing the spread of defect trends, thereby improving production line yield and detection efficiency. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1This is a diagram illustrating the application environment of the AOI defect compensation method in one embodiment. Figure 2 This is a flowchart illustrating an AOI defect compensation method in one embodiment; Figure 3 This is a structural block diagram of an AOI defect compensation device in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] The AOI defect compensation method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0012] In one exemplary embodiment, such as Figure 2 As shown, an AOI defect compensation method is provided, which is applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein:

[0013] Step 202: Perform neural radiation field inversion on the AOI multi-source data and multi-exposure real-shot image data of the circuit board to be inspected to obtain physical characterization data of the board surface.

[0014] Step 204: Based on the physical characterization data of the board surface, construct a condition set for the upstream process data of the circuit board to be inspected, and obtain the condition set.

[0015] Step 206: Based on the condition set, perform condition-driven reasoning on the ideal characterization synthesis model of the circuit board to be inspected to obtain the defect-free ideal characterization.

[0016] Step 208: Perform residual signature analysis on the physical characterization data of the plate surface and the expected characterization of the defect-free state to obtain the defect analysis results.

[0017] Step 210: Based on the defect analysis results, perform closed-loop compensation control analysis on the compensation control object corresponding to the circuit board under inspection to obtain compensation action control data.

[0018] Among them, the circuit board to be inspected is the target PCB (which can be a single board or a unit board in a panel) that is in the AOI inspection station and needs to be judged for appearance / solder joint / assembly quality.

[0019] Among them, AOI multi-source data is a set of AOI images and their corresponding source configuration parameters acquired under the same field of view by different source channels / different incident directions / different illumination combinations.

[0020] Among them, the multi-exposure real-shot image data is a set of multiple AOI real-shot images and their exposure parameters acquired under the same lighting conditions through different exposure times / gains and other imaging parameters.

[0021] Among them, neural radiation field inversion is a computational process that uses differentiable rendering consistency and neural implicit field representation to inversely calculate the scene radiation and reflection response from multi-source / multi-exposure observations in order to achieve decoupling of imaging factors.

[0022] Among them, the physical characterization data of the board surface is obtained by inversion and is a physical characterization of the board surface reflectance, normal, and specular shadow separation, which are relatively independent of the changes in illumination and exposure and can be used for stable contrast.

[0023] Among them, the upstream process data are the process measurement data and process log data generated by the equipment in the manufacturing processes of the circuit board before AOI (such as SPI printing inspection, chip mounting, reflow, etc.).

[0024] Among them, the construction of the condition set is a process of associating upstream process data, design prior data and imaging system parameters on the same board, cleaning and encoding them and organizing them into usable condition inputs for the model.

[0025] Among them, the condition set is a unified conditional input data structure (vector / tensor / field set) used to describe the current plate structure, current working condition drift and current imaging conditions.

[0026] Among them, the normative characterization synthesis model is a predictive / synthetic model that generates "the characterization that the plate should present under defect-free conditions" under the input of structural constraints and working / imaging conditions.

[0027] Among them, condition-driven reasoning is the process of using a set of conditions as control input to perform reasoning calculations on a synthetic model of the expected representation of the target in order to output the expected representation of the target.

[0028] Among them, the defect-free expected characterization is the defect-free appearance / physical characterization result generated by the model under given working conditions and imaging conditions, which is comparable to the physical characterization of the plate surface in the same domain.

[0029] Among them, residual signature analysis is an analytical process that extracts, encodes and forms a discernible "signature" from the structured differences between the measured physical characterization of the plate surface and the expected characterization of the defect-free plate surface.

[0030] The defect analysis results are defect output information inferred from residual signatures, including at least the defect category, defect location, and / or defect severity and its confidence level.

[0031] The compensation control object is a set of control variables that can be adjusted to achieve compensation, including at least AOI decision boundary parameters and / or upstream process equipment process parameters.

[0032] Closed-loop compensation control analysis is a control analysis process that uses the results of defect analysis to attribute, constrain, solve, and verify the compensation control object in order to form a closed-loop correction strategy.

[0033] Among them, the compensation action control data is the compensation output data that can be directly used for execution, including the adjustment amount of the judgment boundary such as AOI formula / threshold and / or the correction amount of the process parameters of upstream equipment / action sequence.

[0034] Specifically, at the AOI station, AOI multi-source data and multi-exposure real-world image data are collected from the same circuit board under inspection and the same field of view according to a preset scheme (ensuring that camera intrinsics, source numbers, exposure parameters, and acquisition timestamps are traceable). Basic corrections are performed on the images (dark field / flat field, bad pixels, distortion, and inter-frame alignment). Then, the "image-source-exposure-viewpoint (which can be fixed)" set is input into the neural radiation field inversion module. With the goal of differentiable rendering consistency, the radiation / reflection response of the scene is jointly estimated so that observations under different lighting and exposures can be explained by the same implicit field. After the inversion converges, physical characterization data of the board surface decoupled from the imaging response is derived from this implicit field, such as reflectivity / material characterization maps, normal or micro-surface roughness characterizations, and specular / shadow separation characterizations, etc., and these characterizations are saved one-to-one with the current board's field of view coordinate system.

[0035] The physical characterization data of the board surface is bound to the board-level information (such as board barcode / serial number, panel position, work order number, timestamp, and camera field of view coordinates). This bound information is used to retrieve the associated records of the same board in the upstream process in the MES / equipment log, forming a "traceable same-board data package". Then, the retrieved upstream process data is cleaned and structured by type. During the cleaning and structuring process by type, the solder paste volume / height / area / offset and its statistics are extracted from the SPI side, the placement side is extracted from the placement deviation / nozzle and material station status / compensation table version, etc., and the reflow side is extracted from the key features of the furnace temperature curve (peak value, TAL, heating slope, temperature difference of each zone, etc.). At the same time, the imaging system parameters (light source configuration, exposure gain, camera calibration version) related to this AOI imaging and the design priors related to this board type (label, pad / package geometry, target solder joint shape constraints) are also organized. Finally, these heterogeneous data are encoded (normalization, discrete feature embedding, temporal feature aggregation, missing value imputation markers, etc.) to output a unified data structure, namely a set of conditions, in which each condition can be traced back to the original device / design source.

[0036] The condition set is divided into three types of input channels: structural constraint channel (from design priors, specifying the geometry and spatial relationship of devices / pads), operating condition channel (from SPI / surface mount / reflow, describing the current process window and drift situation), and imaging condition channel (describing the changes in appearance caused by camera / light source / exposure). These three channels are organized into structural condition tensors and condition vectors according to model requirements. Then, the normative representation synthesis model is called for inference. During inference, the normative representation synthesis model first establishes a "morphological skeleton of a defect-free appearance" under structural constraints, and then modulates the latent space using operating and imaging conditions to produce a result that represents "the defect-free appearance that should appear under the current operating and imaging conditions." To ensure comparability with the physical representation of the board surface in step 202, the output of the normative representation synthesis model is not directly the original RGB image, but a defect-free normative representation consistent with the physical representation (e.g., normative reflectivity / normative normal / normative specular component, etc.), and is strictly aligned with the view coordinates.

[0037] In the same coordinate system, the measured physical characterization of the board surface is paired with the expected characterization of a defect-free surface. Multi-channel residuals (e.g., reflectivity residuals, normal differences, texture / spectral differences) are calculated for each device bit / pad region. The residuals are then expressed in a structured manner, that is, the "intensity difference" and "morphological difference" are measured separately (e.g., boundary offset, connectivity changes, topological features of abnormal protrusions / depressions, geometric inconsistencies of solder joints, etc.), forming residual signatures that can be used for discrimination. Subsequently, defect analysis is performed in the residual signature space, that is, candidate abnormal regions are located (screened according to the spatial sparsity and consistency of the residuals). The signatures are then matched or classified with the defect prior template / discriminator to output the defect analysis results, which at least include the defect category (e.g., insufficient solder / bridging / offset / tombstone, etc.), defect location (tack number / pad coordinates / pixel region), and defect severity (continuous quantity level or risk score).

[0038] The defect analysis results are mapped to a "controllable object space," which is then categorized into two types: one is the decision boundary parameters on the AOI side (thresholds, confidence thresholds, tolerance bands, re-determination trigger conditions, etc.), and the other is the process parameters of upstream processes (executable setpoints for equipment such as printing / mounting / reflow). Within the same control framework, a closed-loop control analysis of "attribution—constraint—solution—verification" is then performed. During the analysis, the cause path is first attributed and sensitivity is assessed based on the defect type, location distribution, and severity (determining whether it's more like process drift, imaging drift, or a true defect trend). Then, a feasible control domain is generated under equipment parameter boundaries, safety constraints, and stability constraints. Subsequently, compensation actions (which can be single actions or action sequences) are solved within the feasible domain. Simultaneously, risk sensitivity verification is performed on these actions (e.g., assessing false alarm / missed alarm risks, overcompensation risks, and equipment out-of-bounds risks). The verified actions are then solidified as compensation action control data. Based on the compensated action control data, the AOI side outputs updatable formula / judgment boundary adjustment amounts, outputs executable parameter correction amounts or maintenance instructions (such as calibration / replacement / cleaning triggers) to the upstream equipment, and writes back the actions, the board's defect signature, condition set, and execution results to form the learning and self-consistency update basis for the next closed-loop cycle.

[0039] In the aforementioned AOI defect compensation method, neural radiation field inversion is performed on the AOI multi-source data and multi-exposure real-shot image data of the circuit board to be inspected to obtain physical characterization data of the board surface. Based on the physical characterization data of the board surface, a set of conditions consistent with the upstream process data is constructed, thereby driving the normative characterization synthesis model to generate a defect-free normative characterization consistent with the current working conditions and imaging conditions. Then, the residual signature analysis of the physical characterization of the board surface and the defect-free normative characterization replaces the traditional template difference / fixed threshold judgment and outputs the defect analysis results. Finally, the defect analysis results are mapped in a closed loop to the compensation control object of the AOI judgment boundary parameters and / or the upstream process process parameters to generate executable compensation action control data. This achieves decoupling and suppression of imaging drift and process drift, and interpretable and quantifiable identification of real defects. It significantly reduces false alarms and false negatives and improves the detection stability across batches, machines, and lighting conditions. At the same time, the detection results are directly converted into a dual closed-loop control of judgment compensation and process compensation, reducing the cost of re-judgment and re-inspection, shortening the process convergence time, and suppressing the spread of defect trends, thereby improving the production line yield and detection efficiency.

[0040] In an exemplary embodiment, condition-driven reasoning is performed on the synthetic model of the expected characterization of the circuit board to be inspected based on a set of conditions to obtain a defect-free expected characterization, including steps 302 to 308. Wherein:

[0041] Step 302: Perform cross-domain latent variable inversion processing on the condition set to obtain the implicit state vector.

[0042] Step 304: Based on the implicit state vector, perform causal temporal unwrapping on the condition set to obtain the decoupled condition set.

[0043] Step 306: Perform structural constraint tensor quantization on the decoupling condition set to obtain the structural condition tensor.

[0044] Step 308: Based on the structural condition tensor and the implicit state vector, perform comparative conditional reasoning on the normative representation synthesis model to obtain a defect-free normative representation.

[0045] Among them, cross-domain latent variable inversion processing is a process of jointly inverting conditional data from multiple sources such as structural domain, process domain and imaging domain, and finding latent state variables in a shared latent space that can consistently interpret the observations of each domain.

[0046] Among them, the implicit state vector is a low-dimensional vector obtained by cross-domain latent variable inversion, which is used to compactly characterize the comprehensive implicit state of the board under the coupling of operating condition drift, equipment state and imaging conditions.

[0047] Among them, causal temporal untangling is a process that uses implicit state vectors to decompose the temporal and causal aliasing of conditional data, and separates the system drift component that is not related to defects from the abnormal component that is related to defects.

[0048] The decoupling condition set is the set of condition inputs obtained after causal temporal untangling, in which drift components and anomalous components have been separated and reorganized in a form that can be used independently by the model.

[0049] Among them, structural constraint tensor quantization is a process of mapping and encoding structural constraints such as design priors, pad geometry, and layout topology into tensor representations aligned with the AOI view, so that structural constraints can be directly imposed on the model during inference.

[0050] Among them, the structural condition tensor is a multi-channel tensor input to the output of the structural constraint tensor. It contains the spatial / topological constraints of the structural prior and their alignment encoding with the operating conditions, which is used to constrain and modulate the generation of the normative characterization.

[0051] Among them, the contrastive conditional reasoning introduces a contrastive gating / modulation mechanism when reasoning in the normative characterization synthesis model. This mechanism retains the influence of defect-independent drift while suppressing defect-related abnormal components, thereby generating a reasoning method for defect-free normative characterization.

[0052] Specifically, the condition set is split into three domains based on the source domain: structural domain (design prior reference designators / pad geometry / layout constraints), process domain (SPI / surface mounting / reflow measurement and timing characteristics), and imaging domain (camera calibration, light source configuration, exposure gain, etc.). Data from each domain is then uniformly scaled, masked for missing fields, and time-aligned. The multi-domain condition inputs are then fed into a cross-domain latent variable inversion module. This module employs an inversion framework of "encoder—latent variable—consistency constraint," establishing intra-domain encoders for the structural, process, and imaging domains to extract domain features. These features are then fused into the desired latent variables using an aggregator in a shared latent space. Iterative inversion is performed with cross-domain consistency (consistent interpretation of latent variables across domains under the same board and operating conditions), reversibility (latent variables can reconstruct key statistics for each domain), and prior regularization (smooth, sparse, or physically interpretable latent variables) as optimization objectives. After the inversion converges, an implicit state vector is output. This vector, in a low-dimensional form, jointly represents the "comprehensive state" of the current board under the coupling of process drift, equipment status, and imaging conditions.

[0053] Based on the drift direction and amplitude reflected in the implicit state vector, the timestamps, batch windows, and equipment states of processes such as SPI → patching → reflow are aligned to make the mismatch caused by different sampling frequencies and delays of different equipment explicit. In the unwrapping module, a "causal-temporal" decomposition is constructed, dividing the changing components in the condition set into defect-independent drift components and defect-related anomaly components. In practical applications, the decomposition uses an implicit state vector-driven gating / attention mechanism as the decomposition operator to separate upstream operating condition features. Simultaneously, temporal consistency constraints ensure smooth evolution of drift components within adjacent boards / adjacent time windows, and anomaly sparsity constraints ensure locality of anomaly components in the spatial / attribute dimensions. After separation, the conditions of the structural domain, process domain, and imaging domain are reorganized in the manner of "stable components with drift removed + anomaly components retained for explaining defects," outputting a decoupled condition set. This allows the model inference to focus more on defect-related factors and avoid being misled by systematic drift.

[0054] The design prior structural information (including device reference designators, pad geometry parameters, layout topology, and target solder joint region definitions) is extracted from the decoupling condition set and mapped to a spatial coordinate frame consistent with the AOI view. A hierarchical index of reference designator—pad—spatial location is used to establish the correspondence between structural elements and the image mesh. Then, the structural information is encoded into tensor form according to the principle of "hard constraint computability," such as generating geometric masks for pads / devices, boundary distance fields, target contour fields, or projection channels for layout adjacency relationships. The operating condition components associated with local locations or reference designators in the decoupling condition set are injected into the structural tensor as additional condition channels, allowing structural constraints and condition modulation to be consumed end-to-end by the model within the same tensor domain. Finally, the resulting multi-channel structural tensor is scaled and aligned, outputting the structural condition tensor.

[0055] In the contrastive conditional reasoning stage, the structural condition tensor is used as the "structural constraint input," and the implicit state vector is used as the "integrated state input for operating conditions / imaging." A contrastive reasoning mechanism is then enabled in the normative representation synthesis model. The normative representation synthesis model internally sets up a contrastive control channel, using the implicit state vector to generate contrastive gating (e.g., channel gating / modulation parameters for latent space or intermediate features) to suppress the direct influence of defect-related anomalous components on the generated results, while retaining the reasonable influence of defect-independent systematic drift on the appearance. Subsequently, the normative representation synthesis model performs conditional reasoning under the constraints of the structural condition tensor to obtain the normative result of "under the current operating and imaging conditions but without defects." To ensure the effectiveness of the comparative reasoning, structural consistency constraints (the output must satisfy the pad geometry and layout topology) and state consistency constraints (the output appearance change is consistent with the drift direction indicated by the implicit state vector) can be applied simultaneously during the reasoning process. This will enable the generated defect-free normative characterization to not only conform to the design structure but also match the current production line state. Finally, the reasoning completes the output and provides an initial defect-free normative characterization (e.g., normative reflectivity / normative normal / normative specular component, etc.) that is comparable to the physical characterization of the board surface in the same domain.

[0056] In this embodiment, an implicit state vector that can uniformly characterize the coupling state of working condition drift and imaging conditions is obtained by cross-domain latent variable inversion of the condition set. The implicit state vector is then used to perform causal temporal unwrapping of the condition set to separate the defect-independent drift component and the defect-related anomaly component. The decoupled condition set is then subjected to structural constraint tensor quantization to form a structural condition tensor that can be directly applied by the model. Finally, under the joint constraints of the structural condition tensor and the implicit state vector, a comparative conditional inference is performed on the applied characterization synthesis model to generate a defect-free normative characterization. This suppresses the interference of systematic drift on the judgment at the generation stage and strengthens the controllable separation of structural prior constraints and anomaly factors, significantly improving the intra-domain consistency and interpretability of the defect-free normative characterization, enhancing the stability of residual signature analysis, and reducing the risk of false positives and false negatives.

[0057] In an exemplary embodiment, based on the structural condition tensor and the implicit state vector, a contrastive conditional inference is performed on the normative representation synthesis model to obtain a defect-free normative representation, including steps 402 to 408. Wherein:

[0058] Step 402: Based on the implicit state vector, perform a control gating transformation on the control control variables of the natural characterization synthesis model to obtain the control gating vector.

[0059] Step 404: Perform feature projection fusion on the structural condition tensor to obtain the inference input tensor.

[0060] Step 406: Based on the inference input tensor and the reference gating vector, perform energy constraint sampling inference on the normative characterization synthesis model to obtain candidate defect-free normative characterizations.

[0061] Step 408: Perform double-constrained differentiable feasible domain projection calibration on the candidate defect-free normative characterization to obtain the defect-free normative characterization.

[0062] Among them, the control variable is the control input variable used in the normative characterization synthesis model to implement the control reasoning, which is used to modulate the latent space or intermediate features to suppress defect-related anomalous components.

[0063] Among them, the control gating transformation is a transformation process that maps the implicit state vector to the gating parameters available for the control variable, so as to generate a control quantity that has a "suppression / retention" effect on the generation process.

[0064] Among them, the contrast gating vector is a set of vectorized gating parameters output by the contrast gating transformation, which is used to gating and modulate model features by channel / space / level during inference.

[0065] Feature projection fusion is a process that maps the structural condition tensor to the model feature domain through projection and fuses it with the backbone inference features, so as to inject structural constraints into the model in a computable form.

[0066] Among them, the inference input tensor is the tensorized conditional input after feature projection fusion, which carries the structural prior and its alignment encoding and serves as the main conditional constraint for subsequent sampling inference.

[0067] Among them, energy-constrained sampling reasoning is a reasoning method that iteratively samples and updates the random latent variables in the latent space under conditional energy constraints in order to generate the normative representation that satisfies structural constraints and gating constraints.

[0068] Among them, the candidate defect-free normative characterization is an intermediate generation result obtained in the energy constraint sampling reasoning stage. It is a defect-free normative output that meets the main constraints but has not yet undergone final feasible domain projection calibration.

[0069] Among them, the dual-constraint differentiable feasible region projection calibration is a calibration process that maps the candidate defect-free normative characterization to the feasible region jointly defined by structural constraints and gated suppression constraints through a differentiable projection operator, thereby obtaining the final consistent defect-free normative characterization.

[0070] Specifically, when the server detects a task that triggers the contrastive conditional inference, the implicit state vector is input into the contrastive gating conversion module. The contrastive gating conversion module first normalizes and scales the implicit state vector to eliminate amplitude differences caused by different batches / devices. Then, it converts it into gating parameters of the contrastive control variables (such as channel gating coefficients, scale and bias of feature modulation, or latent space suppression coefficients) through a gating generation network or parameter mapping operator. The gating parameters are then subject to constraints such as nonnegativity, sparsity, and smoothness to give them an interpretable "suppression-preservation" effect. After the conversion is completed, the contrastive gating vector is output.

[0071] The structural condition tensor is fed as a structural prior input into the feature projection fusion module. This module first aligns the resolution and scale of each channel of the structural condition tensor (e.g., geometric mask, distance field, target contour field, layout topology projection channel, and additional condition channel) to match the feature scale of the backbone network of the normative representation synthesis model. Then, it maps the structural condition tensor to the semantic feature domain of the model's latent space using linear projection, convolutional projection, or attention projection, and fuses them in a "structural constraint channel + condition modulation channel" manner (e.g., splicing before projection, or injecting structural channels as key-value pairs into the backbone features using attention). During the fusion process, the spatial location information and topological relationships of the tag / pad are preserved, ensuring that subsequent generation does not deviate from the design constraints structurally. Finally, the inference input tensor is output.

[0072] Using the inference input tensor as a fixed constraint on structure and conditions, and the contrast-gated vector as a dynamic modulation signal during the inference process, an energy constraint term is introduced into the model inference objective of the normative representation synthesis model. This ensures that the generated result simultaneously satisfies structural consistency (consistency with the geometric / topological constraints encoded by the inference input tensor) and contrast-gated suppression constraints (suppression of defect-related anomalous components in the latent space). Random latent variables are initialized using reparameterized sampling, and the contrast-gated vector is used to modulate the score direction or update step size during sampling iterations, causing the latent variables to evolve along the direction of decreasing conditional energy. With iterative updates, the normative representation synthesis model continuously generates and evaluates candidate representations until convergence criteria (energy decrease, low structural constraint error, satisfaction of gating suppression, etc.) are met, outputting a candidate defect-free normative representation. This candidate result has been formed under "structural constraints + gating constraints," but still retains a small amount of boundary errors or local inconsistencies that can be further corrected by projection calibration.

[0073] The structural feasible region (e.g., pad boundaries, device contours, and allowable morphological deviations) is determined based on the structural condition tensor, and the perturbation suppression feasible region (e.g., the upper limit of allowable residual components related to defects and the suppression threshold of abnormal channels) is determined based on the reference gating vector. These two types of constraints are collectively defined as the dual-constraint feasible region. Then, the candidate defect-free normative characterization is used as the variable to be projected, and a differentiable projection operator is used to calibrate it, ensuring that it strictly satisfies both types of constraints while maintaining the closest possible approximation to the candidate results (this can be implemented as a proximal operator, Lagrange multiplier, or implicit differentiable optimization layer). During the projection process, the continuity of the output in the spatial and channel domains is maintained to avoid introducing artifacts. After projection, the final defect-free normative characterization is obtained, which exhibits stronger structural consistency and gating suppression consistency compared to the candidate characterization.

[0074] In this embodiment, by generating a reference gating vector from the implicit state vector and projecting and fusing the structural condition tensor to form the inference input tensor, the normative characterization synthesis model is simultaneously subject to anomaly suppression and structural constraint control during the inference stage. Then, energy constraint condition sampling inference is used and combined with double-constraint differentiable feasible domain projection calibration to generate a defect-free normative characterization, thereby improving the co-domain consistency and stability of the normative benchmark and reducing false alarms and false negatives in subsequent defect determination.

[0075] In an exemplary embodiment, based on the inference input tensor and the reference gating vector, energy-constrained sampling inference is performed on the normative representation synthesis model to obtain candidate defect-free normative representations, including steps 502 to 510. Wherein:

[0076] Step 502: Align the inference input tensor with the information geometry domain to obtain the manifold-aligned inference tensor.

[0077] Step 504: Based on the manifold aligned inference tensor and the reference gating vector, construct the energy field parameterization of the inference objective of the probability characterization synthesis model to obtain the conditional energy field.

[0078] Step 506: Perform reparameterization sampling initialization on the conditional energy field to obtain the initial random latent variables.

[0079] Step 508: Implicitly differentiable solution is performed on the conditional score field of the conditional energy field with respect to the initial random latent variables to obtain the energy guiding field.

[0080] Step 510: Based on the energy guiding field and the reference gating vector, perform predictive-corrective energy sampling update of the initial random latent variables with differentiable annealing path constraints to obtain the candidate defect-free normative characterization.

[0081] Among them, information geometric domain alignment is a differentiable alignment transformation of the statistical distribution of inference input under information geometric metric to eliminate domain offset caused by cross batch / cross machine / cross operating conditions.

[0082] Among them, the manifold-aligned inference tensor is the inference input tensor obtained after alignment with the information geometry domain. It is consistent with the reference distribution on the statistical manifold and maintains the availability of structural constraint information.

[0083] The reasoning objective is the objective function or posterior criterion that the normative characterization synthetic model optimizes / approaches during the reasoning phase, used to define the constraints and priorities that the "defect-free normative solution" should satisfy.

[0084] Among them, the parameterization of the energy field is the process of representing the inference objective as a differentiable conditional energy function and learning / networking its form using model parameters.

[0085] Among them, the conditional energy field is the energy function terrain defined under the given manifold aligned inference tensor and the contrast gating vector. The lower the energy, the more it conforms to the defect-free normative constraint.

[0086] Among them, reparameterized sampling initialization generates initial values ​​of latent variables from the reference noise distribution through differentiable reparameterized transformation, so that sampling inference can be backpropagated and stably initialized.

[0087] The initial random latent variable is the latent space random state variable obtained by reparameterized sampling initialization, which serves as the starting point for subsequent energy sampling iteration updates.

[0088] Among them, the conditional score field is the field representation of the score function (usually the gradient or equivalent guide) of the conditional energy field with respect to the latent variables in the latent variable space, which is used to indicate the descent / correction direction.

[0089] Implicit differentiable solution is a solution process that solves for the objective quantity through implicit equations / fixed-point iteration and uses implicit differentiation to maintain global differentiability.

[0090] Among them, the energy guiding field is a guiding quantity used to drive the update of latent variables, which is obtained by implicit differentiability solution and can be equivalent to the conditional score field or its equivalent guiding representation.

[0091] Among them, the predictive-corrective energy sampling update of differentiable annealing path constraints is a sampling process that iteratively updates latent variables along the differentiable annealing schedule by "predictive advancement + corrective constraint" in order to balance between exploration and convergence and satisfy gating / structural constraints.

[0092] Specifically, the inference input tensor is constructed into a statistical description (e.g., first / second moments, local covariance, or divergence statistics of embedded features) in the channel dimension and spatial local block dimension, and its distribution differences are characterized by information geometry measures, enabling the inference inputs under different batches, machines, and operating conditions to be compared on the same statistical manifold. A differentiable alignment operator is applied to the inference input tensor on this statistical manifold, mapping it from the current distribution point along the geodesic to a preset reference distribution (or a sliding window self-updating reference distribution) through recalibration in the sense of natural gradients. This absorbs the domain offset primarily into the alignment transformation while preserving the topological morphology of the structural constraint channels. After alignment, a manifold-aligned inference tensor is output, which, while retaining the original structural constraint information and spatial positioning relationships, achieves statistical geometric consistency in the distribution of the inference input.

[0093] The manifold-aligned inference tensor is used as a deterministic input to the structural and conditional constraints, providing "ought-to-be" a priori boundaries such as pad / label geometry, spatial topology, and permissible morphological range. Simultaneously, a contrast-gated vector is used as the modulation signal for the inference objective, explicitly suppressing defect-related anomaly directions while allowing defect-independent drift to influence generation in a controlled manner. Subsequently, a differentiable conditional energy form is defined in the ought-to-be synthetic model, writing the inference objective as a scalar energy about random latent variables or intermediate representations. A parameterized energy head / energy network merges the "structural consistency penalty term (limited by the manifold-aligned inference tensor)" and the "gating suppression penalty term (modulated by the contrast-gated vector)" into a unified energy output, ensuring that the energy in the low-value region corresponds to a "defect-free ought-to-be solution that satisfies structural constraints and suppresses anomalies," ultimately yielding the conditional energy field.

[0094] After the conditional energy field is determined, a random latent variable that can be iteratively updated needs to be initialized in the latent space of the model for conditional sampling inference. Therefore, the prior family and dimension of the latent variable are determined in the latent space of the normative representation synthesis model so that it can bear the main degrees of freedom of the flawless normative representation. Then, the basic random quantity is obtained by sampling from the baseline noise distribution, and the basic random quantity is mapped to the initial value of the latent variable through a differentiable reparameterization transformation. The distribution parameters of the reparameterization transformation are given by the conditional energy field under the constraints of the current manifold aligned inference tensor and the contrast gate vector (e.g., the conditional mean and scale derived from the energy field or the initialization statistics output by the conditional network). This makes the initial value of the latent variable naturally close to the low-energy region and reduces the convergence burden of subsequent sampling iterations. Finally, amplitude boundaries and stability regularization are applied to the obtained initial value of the latent variable to avoid numerical divergence, and the initial random latent variable is output.

[0095] The conditional energy field is considered as a differentiable posterior terrain defined with respect to latent variables, and the "conditional score field" is defined as the score function of this energy with respect to latent variables (i.e., a guide quantity in the latent space indicating the direction and intensity of energy descent). Then, during the inference phase, an implicit solution form satisfying stability and gating suppression constraints is constructed. For example, the score field is written as a fixed-point relationship or implicit equation determined by the conditional energy gradient and the control gating modulation term. A differentiable implicit solver (such as a fixed-point iterative / implicit layer) is used to iteratively solve this implicit equation until convergence, ensuring that the sub-directions maintain a consistent descent property with respect to structural constraints while suppressing defect-related anomalous components. To guarantee end-to-end differentiability, after convergence, the solution process is backpropagated using an implicit function theorem or an equivalent differentiable backpropagation mechanism to obtain a solution with differentiable model parameters, ultimately outputting the energy guide field.

[0096] Based on a preset or adaptive annealing path, continuously differentiable scheduling of temperature, noise injection intensity, and step size is generated at each iteration step, allowing the sampling process to gradually transition from high-temperature exploration to low-temperature convergence. In each iteration step, a prediction update is performed, advancing the latent variables along the energy guiding field to enter lower-energy candidate regions, while simultaneously introducing random perturbations matched to the temperature to avoid getting trapped in local minima. A correction update is then performed, using a control gating vector to suppress the predicted latent variables in defect-related directions and pull back structural consistency deviations, thus achieving a balance between "energy descent" and "gating constraints / stability constraints." When the energy convergence index and constraints meet the stopping condition, the updated latent variables are input into the decoding end of the normative characterization synthesis model, outputting a candidate defect-free normative characterization. This allows the candidate result to possess stronger defect-free consistency and comparability under the combined effects of the annealing path, energy guidance, and control gating.

[0097] In this embodiment, the conditional distribution offset caused by cross-batch / cross-machine is weakened by aligning the information geometric domain of the inference input tensor. Combined with the comparison gating vector, the conditional energy field parameterization of the inference target is constructed, so that the generation of the normative characterization is transformed into energy inference driven by both structural constraints and anomaly suppression. Subsequently, the latent variables are initialized in a reparameterized manner, the energy guidance field is solved implicitly and differentiably, and the prediction-correction sampling update is performed along the differentiable annealing path to obtain the candidate defect-free normative characterization. This improves the convergence stability and intra-domain consistency of the generation process and reduces the contamination of the defect-free benchmark by operating conditions and imaging drift.

[0098] In an exemplary embodiment, residual signature analysis is performed on the physical characterization data of the plate surface and the expected characterization of the defect-free state to obtain defect analysis results, including steps 602 to 606. Wherein:

[0099] Step 602: Perform a joint embedding of the physical characterization data of the plate surface and the defect-free normative characterization with the same manifold to obtain a shared characterization space comparison characterization pair.

[0100] Step 604: Based on the shared characterization space comparison characterization pair, perform optimal transport alignment residual analysis on the physical characterization data of the plate surface and the defect-free expected characterization to obtain the transport residual map.

[0101] Step 606: Perform persistent topology analysis signature on the transmission residual map to obtain the defect analysis results.

[0102] Among them, the co-embedding of the same manifold is a process that maps the physical characterization data of the plate surface and the defect-free expected characterization to the same embedding manifold through the contrast constraint, so that the corresponding features at the same location can be directly measured and compared in the shared space.

[0103] Among them, the shared representation space contrastive representation pair is a one-to-one correspondence representation pair composed of "measured embedding representation" and "ought-to-be embedding representation" in the shared embedding space, used to characterize the contrastive relationship at the same position.

[0104] Among them, the optimal transmission alignment residual analysis is an analytical process that uses the optimal transmission to solve the optimal alignment mapping between the measured and the expected in the shared characterization space, and calculates the structured alignment residual to suppress the effects of mismatch and deformation.

[0105] Among them, the transmission residual map is a graph formed by the rasterization reconstruction of the spatialized residual field obtained after optimal transmission alignment, which is used to intuitively represent the structured difference distribution of the measured relative defect-free norm.

[0106] Among them, persistent topology analysis signature is a process that involves constructing multi-scale topology filtering and calculating persistent homology from the transport residual graph, encoding topological invariants into discriminative signatures to output defect features.

[0107] Specifically, under the same field-of-view coordinate system, the physical representation data of the board surface and the defect-free expected representation are precisely registered and sliced ​​according to the tag number / pad region. Channel uniformity (dimensional normalization, noise robust filtering, and local contrast correction) is performed on the multi-channel physical representation to ensure that subsequent embedding comparisons are based on "physical differences" rather than imaging scale differences. A joint embedding network is then constructed, inputting the physical representation data of the board surface and the defect-free expected representation into the domain encoder to extract hierarchical features of local texture, boundary morphology, and topology. These features are then mapped to the same embedding manifold through a shared projector and trained jointly with contrast constraints and manifold constraints. The joint training with contrast constraints and manifold constraints requires applying a convergence constraint to measured-expected feature pairs at the same spatial location and a distance constraint to feature pairs at different locations / tag numbers. Simultaneously, local neighborhood preservation and curvature uniformity constraints are applied to both embeddings to ensure they have a consistent nearest-neighbor structure in the shared space. After training or online adaptive convergence, the measured embedding and expected embedding in the shared representation space are output for each region, forming a contrast representation pair in the shared representation space.

[0108] Based on the weights of the corresponding representation points within the inference region according to the shared representation space, marginal quality allocation is performed on the measured and expected ends respectively, thus explicitly encoding "which locations / features should be aligned and focused on" into the source and target marginal quality distributions. Then, using the alignment metric implicitly constrained by the comparison relationship in the shared representation space as the distance basis, entropy regularization optimal transmission iteration is performed on the source and target edges to obtain the optimal transmission coupling matrix, so that the measured representation points match the expected representation points with the minimum "alignment cost" under the constraint of quality conservation. Next, the centroid alignment mapping reconstruction is performed on the defect-free expected representation according to the coupling matrix to obtain the aligned expected representation in the same alignment domain as the measured end, and the aligned expected representation and the physical representation data of the board are spatially subtracted to generate an alignment residual field. Finally, the residual field is reconstructed into a transmission residual map by pixel / gridization, so that it intuitively presents the "structured difference between the measured and the defect-free expected" and its intensity distribution in the spatial domain.

[0109] A multi-scale filtering sequence is constructed for the transport residual map (e.g., using sub-level sets / upper-level sets with residual intensity thresholds from low to high, or using the multi-scale smoothed residual field as a filter). The residual structure at each scale is represented as a topological complex to calculate the "birth-death" process of topological features such as connected components and holes as the scale changes. Then, persistent cohomology is calculated to obtain a persistent map / persistent barcode, which is further encoded into a discriminative topological signature (e.g., persistent spectrum, statistical vectorization of the persistent map, or topological summary features). This signature is stable to noise and local deformation but sensitive to changes in defect morphology. Finally, the topological signature is input into the defect discrimination module for semantic mapping and level evaluation, and the defect analysis results are output. The defect analysis results include at least one of the following: defect category, defect location, and / or defect severity corresponding to the candidate defect region, and may include confidence scores for risk constraints in closed-loop compensation control.

[0110] In this embodiment, by comparing and embedding the physical characterization data of the board surface with the expected characterization of the defect-free state into the same manifold, a shared characterization space comparison characterization pair is obtained, so that the measured and the expected state can establish a stable correspondence under a unified metric. Then, based on the comparison characterization pair, optimal transmission alignment residual analysis is performed to generate a transmission residual map, thereby replacing simple difference with structured alignment residuals and suppressing artifacts caused by pose, scale and local deformation. Finally, persistent topological analysis signature is performed on the transmission residual map to output the defect analysis results, making the defect judgment more robust to noise and local disturbances and more sensitive to changes in morphology and connectivity, thereby improving the interpretability of defect identification.

[0111] In an exemplary embodiment, a persistent topology analysis signature is performed on the transport residual map to obtain the defect analysis result, including steps 702 to 708. Wherein:

[0112] Step 702: Perform multi-scale topological complex construction on the transmission residual map to obtain the filtered complex sequence.

[0113] Step 704: Threshold segmentation is performed on the residual intensity data in the transmission residual map to extract the connected components and obtain the candidate defect region.

[0114] Step 706: Based on the filtered complex sequence, perform persistent homology multi-scale statistical analysis on the candidate defect region to obtain the topological analysis signature.

[0115] Step 708: Based on the topological analysis signature, perform joint topological semantic analysis on the defect category and defect degree corresponding to the candidate defect region to obtain the defect analysis results.

[0116] Among them, multi-scale topological complex construction transforms the transport residual map into a computable simple complex or cubic complex at different thresholds / scales to characterize the topological morphology of the residual structure as the scale changes.

[0117] Among them, the filtered complex sequence is a set of complexes constructed from multi-scale topological complexes, which reflects the gradual growth and evolution of the topological structure according to the threshold or scale.

[0118] Among them, residual intensity data is the residual amplitude or energy value of each pixel / grid position in the transmission residual map, which is used to represent the strength of local differences.

[0119] Among them, threshold segmentation connected component extraction is a process that thresholds the residual intensity data and marks connected regions based on adjacency relationships in order to extract spatially continuous outlier regions from the residual map.

[0120] Among them, the candidate defect region is a set of spatial regions that may correspond to real defects, which are extracted by threshold segmentation of connected components. It is usually represented by a region mask or bounding box.

[0121] Among them, persistent homology multiscale statistical analysis is an analysis process that calculates persistent homology on filtered complex sequences and statistically aggregates the birth-death patterns of topological features to obtain stable multiscale topological features.

[0122] Among them, the topology analysis signature is a discriminative feature representation formed by vectorizing / encoding persistent homology statistics, which is used to characterize the topological features of candidate defect regions.

[0123] Among them, the defect category is the result of determining the defect type to which the candidate defect area belongs (e.g., bridging, insufficient tin, offset, etc.).

[0124] Among them, the defect severity is a quantitative result or rating of the severity / impact of the candidate defect area, used to distinguish between minor and severe defects.

[0125] Among them, topological semantic joint analysis is an analysis process that integrates topological analysis signatures with process context / location information and performs classification and severity assessment to output defect categories and defect severity.

[0126] Specifically, the transport residual map is treated as a scalar field defined on a regular grid (the residual intensity is a function value), and a filtering method is selected to generate a multi-scale topological evolution sequence. In practical applications, this typically involves sub-level set / upper-level set filtering with residual intensity thresholds increasing from low to high, or scale-space filtering combined with Gaussian smoothing. Subsequently, at each filtering scale, grid points / pixels that meet the threshold conditions are taken as 0-similarity, and 1-similarity (edges) and 2-similarity (faces) are progressively completed according to grid adjacency relationships to form a computable simplex or cubic complex. Simultaneously, boundary connectivity rules and hole formation rules are consistently defined to ensure cross-scale comparability. As the threshold or scale parameters increase, a series of complex sets that gradually "grow / merge / fill" in a topological sense are obtained, ultimately outputting a filtered complex sequence.

[0127] The residual intensity data in the transport residual map is selected and robustly thresholded (a fixed threshold, quantile threshold, or locally adaptive threshold can be used to adapt to different noise levels). Pixels / mesh values ​​above the threshold are marked as anomalous foreground, and the rest are marked as background. Connectivity labeling is performed on the foreground region (according to 4-neighborhood or 8-neighborhood rules), and morphological cleanup is performed on the connected regions (removing small noise domains, filling small holes, and merging nearest-neighbor fragments) to obtain several spatially continuous candidate anomalous regions. Finally, the bounding box, area, perimeter, and spatial location index of each candidate region are calculated and associated with the tag number / pad region to output each candidate defect region.

[0128] Each candidate defect region is mapped to its corresponding complex substructure (i.e., the complex substructure falling within the candidate region is extracted at each scale), ensuring that the statistical analysis is performed on the "local candidate defect" rather than the global background. Then, persistent cohomology is calculated for the complex subsequence of each candidate defect region, tracking the birth and death of different topological features during scale evolution (e.g., merging of 0-dimensional connected components, appearance and filling of 1-dimensional holes), resulting in a persistent map / barcode. Next, multi-scale statistical convergence and vectorized encoding are performed on the persistent map, such as extracting persistent lifetime distributions, persistent entropy, the number and scale position of key bars, and cross-scale stability indices in different dimensions. This transforms the "connectivity / porosity / fragmentation of residual morphology" into a numerical topological analysis signature, ultimately outputting the topological analysis signature as a highly robust morphological feature representation of the candidate defect region.

[0129] The topology analysis signature is fused with the location information of the candidate defect region (tack designation, pad type, and position relative to the pad boundary) and necessary residual statistics (peak value, energy, directionality, etc.) to enable joint judgment of "topology morphology + process context" at the semantic level. The fused features are then input into a defect discrimination module for classification and severity regression. This module can be a rule base and threshold tree (e.g., distinguishing between bridging / insufficient solder / cold solder joints based on rules such as the number of connected components, hole persistence, and area ratio) or a learned classification regressor (e.g., inputting the topology signature vector into a multilayer perceptron / graph network to output category and severity score), and can output confidence as a risk indicator. Finally, the discrimination results are backfilled into the candidate defect region and summarized to form the defect analysis results. These results include at least one or more of the following: the defect category, defect severity, and spatial location index of the candidate defect region.

[0130] In this embodiment, a multi-scale topological complex is constructed from the transport residual map to form a filtered complex sequence, thereby explicitly modeling the connectivity, porosity, and scale evolution of the residual. Then, candidate defect regions are extracted by threshold segmentation of connected components to focus on real anomalies and suppress background noise interference. Furthermore, persistent homology multi-scale statistical analysis is performed on the candidate defect regions to obtain topological analysis signatures, thereby obtaining stable features that are robust to local deformation and imaging perturbations and sensitive to changes in defect morphology. Finally, based on the topological analysis signatures, joint topological semantic analysis is performed to output defect analysis results of defect category and defect degree, thereby improving the consistency of defect identification across operating conditions and reducing false alarms and false negatives.

[0131] In an exemplary embodiment, optimal transport alignment residual analysis is performed on the physical characterization data of the plate surface and the defect-free expected characterization based on the shared characterization space comparison characterization pair to obtain a transport residual map, including steps 802 to 810. Wherein:

[0132] Step 802: Based on the shared characterization space comparison characterization pair, perform comparison weight inference on the physical characterization data of the plate surface and the expected characterization of the defect-free surface to obtain the comparison weight.

[0133] Step 804: Based on the corresponding weights, perform marginal quality adaptive allocation on the representation points in the physical representation data of the plate surface and the representation points in the defect-free normative representation to obtain the source marginal quality distribution and the target marginal quality distribution.

[0134] Step 806: Based on the implicitly constrained alignment metric in the shared representation space and the comparison representation pair, perform entropy-normalized optimal transmission iterative solution on the source marginal quality distribution and the target marginal quality distribution to obtain the optimal transmission coupling matrix.

[0135] Step 808: Based on the optimal transmission coupling matrix, the defect-free normative characterization is reconstructed by centroid alignment mapping to obtain the aligned normative characterization.

[0136] Step 810: Perform alignment residual field rasterization reconstruction on the alignment normative characterization and plate physical characterization data to obtain the transmission residual map.

[0137] Among them, the row-to-row correspondence weight inference is based on the processing of the soft correspondence strength between the representation points of the estimated physical representation data of the plate surface and the representation points of the defect-free normative representation, which is based on the shared representation space.

[0138] Among them, the corresponding weight is the weight matrix or weight field inferred from the corresponding weight, which is used to represent the matching confidence and corresponding strength between two sets of characterization points.

[0139] Among them, the characterization point is the basic element used for alignment and calculation in the physical characterization data of the plate surface or the defect-free normative characterization. It can be a multi-channel feature vector of pixel / grid position or a feature point in the embedding space.

[0140] Among them, the marginal quality adaptive allocation is a process that assigns quality weights to source / target representation points based on the corresponding weights and structural priors to form the optimal transmission marginal constraints.

[0141] Among them, the source marginal quality distribution is the quality weight distribution allocated on the physical characterization data point set of the plate surface, which is used as the source-end marginal constraint for optimal transmission solution.

[0142] Among them, the target marginal quality distribution is the quality weight distribution obtained by allocating it on the set of characteristic points of the defect-free normative characterization, and is used as the target end marginal constraint for optimal transmission solution.

[0143] The implicitly defined alignment metric is a point-to-point distance / similarity calculation rule determined by the shared representation space and the representation pairs. It is used to define the alignment cost between the source and target points in the solution without the need for an explicit cost tensor.

[0144] Among them, the entropy-regularized optimal transport iterative solution is a solution process that iteratively updates the coupling under the entropy regularization constraint to satisfy the marginal quality distribution of the source / target, thereby obtaining a stable optimal transport solution.

[0145] Among them, the optimal transmission coupling matrix is ​​the mass transport scheme matrix obtained by solving the optimal transmission, which gives the optimal alignment and allocation relationship between the source representation point and the target representation point.

[0146] Among them, the centroid alignment mapping reconstruction uses the optimal transmission coupling matrix to weight the centroid aggregation of the target representation and map it to the source domain coordinates, thereby realizing the alignment reconstruction process.

[0147] Among them, the aligned normative characterization is the aligned version of the defect-free normative characterization obtained after centroid alignment mapping reconstruction in the source domain (plate surface physical characterization data domain), which can be compared with the measured point by point.

[0148] Among them, the alignment residual field rasterization reconstruction is the process of calculating the difference between the alignment expected characterization and the physical characterization data of the plate as a residual field and mapping it to a regular grid to form a transmission residual map.

[0149] Specifically, the shared representation space is organized into a set of "measured embedding point set - expected embedding point set" by tag number / pad region. Within the shared representation space, the similarity between each measured embedding point of the board physical representation data and each expected embedding point of the defect-free expected representation is calculated (e.g., based on cosine similarity, Mahalanobis distance, or geodesic distance on the shared manifold). A contrast consistency constraint is introduced to reweight the similarity, ensuring smooth correspondences within the same spatial neighborhood while suppressing correspondences across structural boundaries. Local structural priors (such as pad boundaries and device contours) can be combined to penalize cross-region matching. Finally, the reweighted similarity is normalized (e.g., softmax or double random normalization) to convert it into a correspondence probability or correspondence weight matrix, outputting the contrast correspondence weights to explicitly express "which measured representation points should align with which expected representation points" and their confidence distribution.

[0150] Based on the corresponding weights, an allocatable total "mass" is defined for the representation points in the physical characterization data and the representation points in the defect-free normative characterization. The corresponding weights are interpreted as an indicator of the importance of mass on the point set. Points with high corresponding weights, stable structures, and those related to defect-sensitive locations should be assigned higher quality. Subsequently, based on the joint results of the local residual priors of the points (such as the local texture gradient, boundary strength, and pad critical region weights of the physical characterization) and the corresponding weights, the source point set and the target point set are normalized and assigned respectively, forming source marginal mass distributions and target marginal mass distributions that satisfy mass conservation. At the same time, a minimum mass lower limit or temperature smoothing is applied to extreme uncertain points to avoid overly sharp assignments that could lead to unstable solutions. Finally, the source marginal mass distribution and the target marginal mass distribution are output.

[0151] Using the reference-representation pairs in the shared representation space as a metric generator, the alignment cost calculation rule between the source and target representation points is defined. Specifically, in each iteration, the point-to-point distance is calculated instantaneously based on a combination of the reference embedding distance, the local neighborhood preservation term, and the boundary consistency penalty, thus eliminating the need to explicitly introduce a separate transmission cost tensor. Under this alignment metric constraint, the coupling matrix is ​​initialized, and an entropy regularization term is introduced to ensure the smoothness and numerical stability of the coupling. The coupling matrix is ​​normalized through alternating scaling iterations to gradually satisfy the marginal constraints of the source and target marginal quality distributions. Simultaneously, entropy regularization is used during iteration to suppress overly sharp matches, enhancing robustness against noise and local mismatches. When the marginal constraint error and the objective function change converge to a preset threshold, the optimal transmission coupling matrix is ​​output, providing the optimal quality transport scheme between the source and target.

[0152] The optimal transfer coupling matrix is ​​interpreted as a soft correspondence between the set of representation points of the defect-free normative representation and the set of representation points of the physical representation data. Each representation point (or grid position) of the physical representation data is used as a reconstruction anchor point. Then, for each anchor point, the weight distribution of its corresponding defect-free normative representation point is read from the optimal transfer coupling matrix, and the weighted centroid is calculated. The representation values ​​of the defect-free normative representation on each channel are weighted and aggregated according to this weight, thereby transferring the defect-free normative representation to the same alignment domain as the physical representation data. To avoid excessive mixing across pad boundaries or device structures, locality and boundary preservation constraints are applied to the weight support domain during centroid aggregation (e.g., restricting weights from falling within the same reference number / geometric mask or penalizing weights across boundaries). The aggregation result is then subjected to necessary smoothing-sharpening balance to preserve fine structures and edge clarity. Finally, the aligned normative representation is output, ensuring a one-to-one correspondence with the physical representation data in spatial location, scale, and structural semantics.

[0153] The alignment residual field is obtained by calculating the residuals of the alignment-predicted characterization and the physical characterization data of the board surface under the same grid / same channel semantics. This alignment residual field can contain multi-channel residuals (such as reflectivity residuals, normal difference, texture / spectral residuals) and their fused residual intensities. Subsequently, the alignment residual field is rasterized and reconstructed according to the pixel grid of the AOI view or a preset analysis grid. If the residuals are calculated on point sets / sparse sampling, they are mapped to a regular grid through interpolation / weighted scattering, while preserving boundaries and local extrema to avoid smoothing out defects. Finally, the rasterized residual map is subjected to necessary scale unification and dynamic range compression to highlight the defect area without amplifying noise, and a transmission residual map is output.

[0154] In this embodiment, the weights of the comparison between the inferred physical representation data of the plate surface and the defect-free normative representation are determined by the comparison representation based on the shared representation space. Based on this, the marginal quality of the representation points on both sides is adaptively allocated to form a controllable optimal transmission marginal constraint, so that the alignment process can highlight the key structural regions and suppress uncertain matching. Furthermore, the optimal transmission coupling matrix is ​​obtained by performing entropy regular optimal transmission iterative solution under the implicitly constrained alignment metric, so as to complete the soft alignment in a smooth, stable and differentiable manner and reduce the mismatch risk caused by noise and local deformation. Subsequently, the centroid alignment mapping is used to reconstruct the defect-free normative representation using the coupling matrix to obtain the aligned normative representation, and the alignment residual field is rasterized and reconstructed with the physical representation data of the plate surface to output the transmission residual map. Thus, the structured alignment residual is used to replace the simple difference, improving the spatial consistency and interpretability of the residual representation.

[0155] In an exemplary embodiment, based on the defect analysis results, closed-loop compensation control analysis is performed on the compensation control object corresponding to the circuit board under inspection to obtain compensation action control data, including steps 902 to 908. Wherein:

[0156] Step 902: Based on the defect analysis results, perform causal attribution calculation on the defect cause path of the circuit board under inspection to obtain the defect driving factor vector.

[0157] Step 904: Based on the defect driving factor vector, construct the feasible domain constraint for the compensation control object to obtain the constraint control domain.

[0158] Step 906: Based on the constrained control domain, perform reverse policy solution on the sequence of each compensation action of the compensation control object to obtain candidate compensation policies.

[0159] Step 908: Based on the candidate compensation strategy, perform risk-sensitive closed-loop verification and screening for each compensation action to obtain compensation action control data.

[0160] Among them, the defect cause path is to trace the defect phenomenon back along the manufacturing process to the causal chain that may affect the links and key parameters, in order to locate the main source of the defect.

[0161] Among them, causal attribution solution is a calculation process that estimates and decomposes the causal contribution of each potential causal factor based on the defect analysis results and process data.

[0162] Among them, the defect driving factor vector is the result of vectorizing the contribution, direction of action and uncertainty of each causal factor, and is used to guide the objectives and priorities of compensation control.

[0163] Among them, feasible domain constraint construction is the process of organizing constraints such as device boundaries, safety and stability in mathematical form into an executable action space.

[0164] Among them, the constraint control domain is a feasible set of compensation control obtained by constructing the feasible domain constraints, which limits the allowable adjustment range and combination relationship of the compensation control object.

[0165] Among them, the compensation action is a specific adjustment operation or parameter update (single or sequential) performed on the compensation control object to suppress the influence of defect driving factors.

[0166] Among them, the reverse strategy solution is a solution process that starts from the improvement effect of the target defect, reverses the process within the constraint control domain, and searches for a sequence of compensation actions that satisfy the target.

[0167] Among them, the candidate compensation strategy is a set of optional compensation action sequences obtained by solving the inverse strategy, which includes information such as adjustment direction, magnitude, order and period.

[0168] Among them, risk-sensitive closed-loop verification screening is the process of evaluating, eliminating and selecting candidate compensation strategies and their actions under closed-loop feedback and risk indicator constraints to output executable control data.

[0169] Specifically, based on the defect analysis results (including at least defect category, defect location / tag distribution, defect severity, and confidence level), the spatial distribution of defects on the circuit board under inspection is mapped to traceable process units (e.g., corresponding to SPI pads, mounted devices, reflow temperature zones, AOI recipe items, etc.), and the upstream process records and equipment status logs of the board are linked in chronological order to form a "defect cause path candidate chain". Then, a set of potential driving factors (such as solder paste volume deviation, printing offset, mounting deviation, nozzle status, reflow peak / heating slope, AOI threshold drift, light source attenuation, etc.) is defined in the candidate chain. Using the defect results as the observation output, an interpretable causal attribution model is constructed (which can use a structural causal graph / path model or a causal scoring model with intervention items). The contribution and directionality of each driving factor to the current defect are estimated on the same board data and historical window data. Finally, the contribution, directionality, and uncertainty are summarized and encoded into a defect driving factor vector, which simultaneously expresses "which factors are most likely to cause this defect", "whether the impact is positive or negative", and "the confidence level".

[0170] Server 104 clearly defines the set of variables and operational granularity of the compensation control object, including at least AOI-side judgment boundary parameters (thresholds, confidence thresholds, tolerance bands, re-judgment triggers, etc.) and / or upstream process parameters (printing pressure / speed / wiping strategy, mounting bias compensation, reflow temperature zone settings, etc.). For each type of control variable, it reads the hard boundaries (upper and lower limits, resolution, update frequency, mutual exclusion relationships) and safety boundaries (risk intervals that are not allowed to be introduced) of the equipment / formula. Then, it transforms the defect driving factor vector into constraints, setting stricter stability constraints (action amplitude, action frequency, rate of change) for control variables corresponding to "high contribution" driving factors, and setting more conservative action upper limits for variables corresponding to "low contribution or uncertain" factors to prevent overcompensation. Simultaneously, it incorporates quality and risk constraints (false alarm / missed alarm risk upper limit, yield fluctuation upper limit, equipment alarm / downtime risk upper limit, etc.). Finally, it unifies the above hard boundaries, safety constraints, stability constraints, and goal-oriented constraints into a feasible domain (e.g., a set of inequalities, a domain defined by differentiable penalties or projection operators), outputting the constraint control domain.

[0171] The compensation strategy is formalized as a sequence of actions on the compensation control object (AOI decision boundary parameters and / or upstream process parameters) within several closed-loop cycles. The "defect effect to be suppressed" corresponding to the defect driving factor vector is defined as the reverse solution objective (e.g., reduction of defect severity, convergence of residual signature to the defect-free domain, or control of key risk indicators). Using the constraint control domain as the feasible solution space, hard boundaries, stability, and safety constraints are embedded into the sequence solution process. The expected amount of compensation for the defect driving factor at each step is deduced from the objective end through reverse programming / constraint optimization. Then, the action sequence that satisfies the objective and has the minimum cost is searched within the feasible domain. The coupling between actions and the temporal effect are explicitly considered in the search (e.g., first adjust the AOI boundary to reduce the risk of misjudgment, then gradually correct the process parameters to eliminate the real defect, or the reverse order to accelerate process convergence). When the degree of objective achievement, constraint satisfaction, and convergence criterion meet the preset thresholds, candidate compensation strategies are output. The candidate compensation strategies include at least the adjustment direction, adjustment magnitude, execution order, and execution cycle of each control variable.

[0172] Candidate compensation strategies are decomposed into a sequence of compensation actions arranged in execution order. A calculable risk and benefit assessment term is established for each compensation action in the sequence. Risks include at least the risk of equipment parameters exceeding limits, stability risks caused by action amplitude / frequency, false alarm / missed alarm changes, and the risk of defect type migration due to overcompensation. During the verification process, closed-loop consistency checks are performed on each compensation action step by step. This involves hard-constraint verification of the action's executability within the constraint control domain, and assessment of whether the action acts on high-contribution driving factors and whether its direction of action is correct, combined with the defect driving factor vector. Simultaneously, the improvement in defect severity and residual convergence by the action is predicted using a fast surrogate model / empirical response curve or online small-step trial feedback. Based on successful action-level verification, sequence-level verification of candidate compensation strategies is further conducted to evaluate the combined impact of action coupling and temporal cumulative effects on risk and benefit indicators. Candidate compensation strategies are then ranked and eliminated according to the "benefit-risk" criterion. Finally, the verified and risk-controlled compensation action sequences are selected and solidified into compensation action control data. The compensation action control data includes at least the parameter update amount, execution order, execution cycle, monitoring threshold, and conditions for triggering rollback for each compensation action.

[0173] In this embodiment, a defect driving factor vector is obtained by performing causal attribution calculation on the defect cause path based on the defect analysis results. This transforms the compensation decision from experience-based parameter tuning to directional control based on the contribution of the cause. Then, a constraint control domain is constructed based on the defect driving factor vector to construct the compensation control object. Equipment boundaries, safety risks, and stability requirements are explicitly included in the feasible solution space to avoid over-compensation and exceeding limits. Furthermore, inverse strategy solving is performed on the compensation action sequence within the constraint control domain to generate candidate compensation strategies. This ensures that the action sequence and magnitude converge around the defect driving factors to achieve target-oriented minimum cost convergence. Finally, risk-sensitive closed-loop verification and screening are performed on each compensation action in the candidate compensation strategies to output compensation action control data. This accelerates defect convergence, suppresses defect propagation, and reduces yield fluctuations caused by mis-tuning while ensuring that risks are controlled.

[0174] Based on the same inventive concept, this application also provides an AOI defect compensation device for implementing the above-mentioned AOI defect compensation method. For example... Figure 3 As shown, it includes: a multi-source data inversion module, a condition set construction module, a condition-driven reasoning module, a residual signature analysis module, and a compensation control analysis module. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more AOI defect compensation device embodiments provided below can be found in the limitations of an AOI defect compensation method described above, and will not be repeated here.

[0175] The modules in the aforementioned AOI defect compensation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0176] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces.

[0177] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0178] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0179] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0180] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0181] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0182] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0183] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0184] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An AOI defect compensation method, characterized in that, The method includes: Neural radiation field inversion was performed on the AOI multi-source data and multi-exposure real-shot image data of the circuit board under inspection to obtain physical characterization data of the board surface; Based on the physical characterization data of the board surface, a condition set is constructed for the upstream process data of the circuit board to be inspected, and a condition set is obtained. Based on the set of conditions, condition-driven reasoning is performed on the normative characterization synthesis model of the circuit board to be inspected to obtain a defect-free normative characterization. Residual signature analysis is performed on the physical characterization data of the plate surface and the defect-free normative characterization to obtain the defect analysis results; Based on the defect analysis results, closed-loop compensation control analysis is performed on the compensation control object corresponding to the circuit board under inspection to obtain compensation action control data.

2. The method according to claim 1, characterized in that, The step of performing condition-driven reasoning on the normative characterization synthesis model of the circuit board under inspection based on the set of conditions to obtain a defect-free normative characterization includes: The condition set is subjected to cross-domain latent variable inversion processing to obtain the implicit state vector; Based on the implicit state vector, the condition set is subjected to causal temporal unwrapping to obtain a decoupled condition set; The structural constraint tensor is obtained by performing structural constraint tensor quantization on the decoupling condition set; Based on the structural condition tensor and the implicit state vector, a comparative conditional inference is performed on the normative characterization synthesis model to obtain the defect-free normative characterization.

3. The method according to claim 2, characterized in that, The step of performing comparative conditional reasoning on the normative characterization synthesis model based on the structural condition tensor and the implicit state vector to obtain the defect-free normative characterization includes: Based on the implicit state vector, the control variables of the normative characterization synthesis model are subjected to control gating transformation to obtain the control gating vector; The structural condition tensor is subjected to feature projection fusion to obtain the inference input tensor; Based on the inference input tensor and the reference gating vector, energy constraint sampling inference is performed on the normative characterization synthesis model to obtain candidate defect-free normative characterizations. The candidate defect-free normative characterization is subjected to double-constrained differentiable feasible domain projection calibration to obtain the defect-free normative characterization.

4. The method according to claim 3, characterized in that, The step of performing energy-constrained sampling inference on the normative characterization synthesis model based on the inference input tensor and the reference gating vector to obtain candidate defect-free normative characterizations includes: The inference input tensor is aligned with the information geometry domain to obtain the manifold-aligned inference tensor; Based on the manifold aligned inference tensor and the reference gating vector, the inference objective of the normative characterization synthesis model is constructed by energy field parameterization to obtain the conditional energy field; The conditional energy field is reparameterized and sampled to initialize it, thus obtaining initial random latent variables; The conditional score field of the conditional energy field with respect to the initial random latent variable is implicitly differentiable to obtain the energy guiding field; Based on the energy guiding field and the control gating vector, the initial random latent variable is updated by predictive-corrective energy sampling with differentiable annealing path constraints to obtain the candidate defect-free normative characterization.

5. The method according to any one of claims 1 to 4, characterized in that, The residual signature analysis performed on the physical characterization data of the plate surface and the defect-free expected characterization yields the defect analysis results, including: The physical characterization data of the plate surface and the defect-free normative characterization are subjected to a joint embedding of the same manifold to obtain a shared characterization space comparison characterization pair. Based on the shared characterization space comparison characterization pair, the optimal transmission alignment residual analysis is performed on the physical characterization data of the plate surface and the defect-free normative characterization to obtain the transmission residual map. Persistent topology analysis signature is performed on the transmission residual graph to obtain the defect analysis results.

6. The method according to claim 5, characterized in that, The process of performing persistent topology analysis and signature on the transmission residual map to obtain the defect analysis result includes: The transmission residual map is subjected to multi-scale topological complex construction to obtain a filtered complex sequence; Threshold segmentation is performed on the residual intensity data in the transmission residual map to extract the connected components and obtain candidate defect regions; Based on the filtered complex sequence, persistent homology multi-scale statistical analysis is performed on the candidate defect region to obtain the topological analysis signature; Based on the topological analysis signature, a joint topological semantic analysis is performed on the defect category and defect severity corresponding to the candidate defect region to obtain the defect analysis result.

7. The method according to claim 5, characterized in that, The step of performing optimal transport alignment residual analysis on the physical characterization data of the plate surface and the defect-free normative characterization based on the shared characterization space comparison characterization pair to obtain a transport residual map includes: Based on the shared characterization space comparison characterization pair, the physical characterization data of the plate surface and the defect-free normative characterization are compared and weighted to infer the corresponding weights, and the corresponding weights are obtained. Based on the corresponding weights, the marginal quality adaptive allocation is performed on the representation points in the physical representation data of the plate surface and the representation points in the defect-free normative representation to obtain the source marginal quality distribution and the target marginal quality distribution. Based on the implicitly defined alignment metric in the shared representation space and the corresponding representation pair, the source marginal quality distribution and the target marginal quality distribution are solved by entropy regularization optimal transmission iteration to obtain the optimal transmission coupling matrix. Based on the optimal transmission coupling matrix, the defect-free normative characterization is reconstructed by centroid alignment mapping to obtain the aligned normative characterization. The alignment-predicted characterization and the physical characterization data of the plate surface are reconstructed by rasterization of the alignment residual field to obtain the transmission residual map.

8. The method according to claim 1, characterized in that, Based on the defect analysis results, closed-loop compensation control analysis is performed on the compensation control object corresponding to the circuit board under inspection to obtain compensation action control data, including: Based on the defect analysis results, the cause path of the defects in the circuit board under inspection is calculated by causal attribution to obtain the defect driving factor vector. Based on the defect driving factor vector, a feasible domain constraint is constructed for the compensation control object to obtain the constraint control domain; Based on the constrained control domain, the sequence of each compensation action of the compensation control object is solved by reverse strategy to obtain candidate compensation strategies; Based on the candidate compensation strategy, risk-sensitive closed-loop verification and screening are performed on each compensation action to obtain the compensation action control data.

9. An AOI defect compensation device, characterized in that, The device includes: The multi-source data inversion module is used to perform neural radiation field inversion on the AOI multi-source data and multi-exposure real-shot image data of the circuit board under inspection to obtain physical characterization data of the board surface. The condition set construction module is used to construct a condition set for the upstream process data of the circuit board under inspection based on the physical characterization data of the board surface, and obtain the condition set. The condition-driven reasoning module is used to perform condition-driven reasoning on the normative characterization synthesis model of the circuit board to be inspected based on the set of conditions, so as to obtain a defect-free normative characterization. The residual signature analysis module is used to perform residual signature analysis on the physical characterization data of the board surface and the defect-free normative characterization to obtain the defect analysis results. The compensation control analysis module is used to perform closed-loop compensation control analysis on the compensation control object corresponding to the circuit board under inspection based on the defect analysis results, and obtain compensation action control data.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images