Computer vision-based real-time evaluation method for copper plating uniformity of PCB hole wall

Abstract

The present application relates to a computer vision-based real-time evaluation method for PCB hole wall copper plating uniformity. The core is to use a multi-angle, multi-polarization state pulse ring LED array light source and a high-resolution camera to cooperate, collect and preprocess the hole wall reflected light intensity sequence, combine with known parameters such as Young's modulus, establish a fine response model representing the correlation between hole wall micro-deformation and reflectivity, and generate a three-dimensional local physical fingerprint accordingly, realize automatic segmentation and texture feature extraction in nano, sub-micron and micron multi-scale sensitive domain. Through graph convolution embedding and time series adaptive regression mapping, a dynamic mapping of texture-process parameters is established, the process response delay is compensated, and adaptive optimization of copper plating process fluctuation is effectively realized. The application has the beneficial effects of high microstate detection precision, fast process response and strong model adaptive ability.

Description

Technical Field

[0001] This invention relates to the field of printed circuit board (PCB) surface texture feature detection and process adaptive optimization technology, and in particular to a computer vision-based method for real-time evaluation of PCB hole wall copper uniformity. Background Technology

[0002] In the field of printed circuit board (PCB) manufacturing, especially in the through-hole copper deposition process, achieving high-precision real-time detection of the uniformity of copper deposition and surface texture on the hole walls has become a key aspect of improving production yield and intelligent process control. To achieve this goal, mainstream detection technologies in the industry can be broadly categorized into two types: one is image processing and detection algorithms based on fixed-scale texture features, and the other is identification methods that combine multi-scale texture fusion with process parameter criteria. Existing image detection schemes are generally based on two-dimensional or three-dimensional reflectivity imaging technology, utilizing fixed-scale mathematical transformations such as the Laplacian operator, Gabor wavelets, and multi-directional gradient spatial filtering to extract texture features and perform principal component analysis on the hole wall surface. These technologies can, to a certain extent, identify obvious defects such as plating undulations, micropores, and roughness, and can provide auxiliary basis for quantifying copper deposition quality by establishing gray-level distribution histograms and GLCM texture matrices. Some cutting-edge solutions further introduce strategies such as multi-scale image pyramids, feature pyramid networks (FPNs), and multi-branch attention mechanisms, hoping to improve the integration capability of texture information at different spatial frequencies, thereby improving the resolution and robustness of defect detection. In the area of ​​process feedback optimization, some studies have explored the direct mapping between process parameters and characterization features. For example, they have directly correlated copper plating bath temperature and current density with texture statistical features, using machine learning to automatically compensate for process fluctuations and improve the automation level of process control. Another approach attempts to fuse multi-sensor data, comparing visual features with ultrasonic, infrared, and even electrochemical sensor outputs to construct more comprehensive physical feature criteria, thereby reducing the detection blind spots of single signals.

[0003] However, existing multi-scale texture analysis schemes generally face the following prominent problems: First, traditional fixed-scale decomposition or weighted fusion methods often fail to take into account the physical discriminative nature of texture information. Due to the lack of tight coupling with intrinsic physical parameters such as the actual material state and elastic response mechanism of the hole wall, these schemes usually rely on scale weights obtained from prior or experimental statistics to empirically weight texture features at different scales. As a result, they often can only capture relatively obvious coating fluctuations at the macro scale, and there are obvious detection blind spots and errors in the detection of local non-uniform defects at the nanometer to submicron level that determine the final copper plating density and flatness. Second, existing fusion algorithms generally focus on the mathematical and statistical characteristics of image space, deviating from the actual physical roots of the material mechanics and optical coupling in the copper plating process. This makes the detection results susceptible to external factors such as the test environment (e.g., lighting conditions, acquisition angle) and equipment configuration, making it difficult to achieve dynamic and adaptive texture recognition for the local real physical state of PCB vias, thus limiting the accuracy and response speed of process closed-loop optimization control.

[0004] Furthermore, current mainstream technologies face practical challenges in detecting subtle non-uniformities, including low signal-to-noise ratio, large result fluctuations, and insufficient sensitivity. Multi-scale pyramids and attention heatmaps, which are typically based on mathematical paradigms such as hierarchical convolution or weighted aggregation, struggle to make high-confidence physical causal judgments on texture changes in real physical regions. On automated production lines with high process tolerance requirements, stability and timeliness are difficult to guarantee. Summary of the Invention

[0005] This application provides a computer vision-based real-time evaluation method for the uniformity of copper plating on PCB hole walls, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0006] The real-time evaluation method for PCB hole wall copper uniformity based on computer vision provided in this application specifically includes: S1: Using a ring-shaped LED array light source with controllable polarization angle and microsecond-level pulse modulation capability, the inner wall of the PCB via is excited by a multi-angle, multi-polarization state sequence. Simultaneously, a high-resolution line scan camera is triggered to acquire and record the sequence of reflected light intensity from the via wall, including the incident angle of the light source, polarization direction, and pulse timing markers.

[0007] S2: Based on the reflected light intensity sequence of the hole wall and the known Young's modulus, Poisson's ratio, copper thickness parameters of copper material and geometric curvature radius of the hole wall, a hole wall micro-deformation response model is constructed that maps the micron-level elastic displacement distribution of the hole wall surface under light pressure to a local reflectivity modulation function.

[0008] S3: Using the three-dimensional reconstructed mesh of the aperture wall as a reference frame, the reflected light intensity sequence of the aperture wall is input into the micro-deformation response model of the aperture wall for inverse decoupling. The slope, saturation threshold and hysteresis interval of the response curve of each mesh unit under different illumination conditions are fitted to generate a response fingerprint that characterizes the comprehensive physical state of local copper layer compactness and micro-flatness.

[0009] S4: Based on the steepness, monotonicity, and repeatability of the response fingerprint, three physically sensitive scale domains are defined, corresponding to the high steepness domain, medium slope domain, and significant hysteresis domain, respectively, for nanoscale grain orientation differences, submicron-level pore distribution, and micron-level coating undulations.

[0010] S5: Perform trend consistency determination and signal-to-noise ratio threshold comparison on the response fingerprints of each physical sensitive scale domain acquired in three consecutive frames, and generate a gating enable signal characterizing whether each scale domain meets the activation conditions; use the gating enable signal to filter the response fingerprints, and independently perform Gabor direction response, local contrast gradient histogram and Fourier frequency domain energy spectrum extraction only on the scale domains that meet the activation conditions, and generate multi-scale texture feature channels output by nonlinear gating.

[0011] S6: The multi-scale texture feature channels are fed into a graph convolutional network. Using the hole wall mesh nodes as graph structure vertices and the response fingerprint similarity as edge weights, a cross-node texture propagation aggregation operation is performed to generate a hole wall texture embedding vector with spatial physical consistency.

[0012] S7: Based on the nonlinear mapping relationship between the hole wall texture embedding vector and the copper plating process parameters, calculate the dynamic optimization factor used to compensate for the delay in image features and process control response, and generate process correction instructions for fluctuations in the copper plating process.

[0013] S8: If the confidence level of the dynamic tuning factor is detected to be lower than the preset stability threshold, the issuance of the process correction instruction is paused and the model self-update process is triggered; otherwise, the process correction instruction is sent to the copper plating control system to perform closed-loop feedback optimization.

[0014] The computer vision-based real-time evaluation method for PCB hole wall copper uniformity provided in this application has the following advantages: (1) By introducing a ring LED array light source with controllable polarization angle and microsecond-level pulse modulation, and combining it with a high-resolution line scan camera to simultaneously acquire the reflected light intensity sequence of the hole wall under multiple conditions, and establishing a micro-deformation response model based on the physical properties of copper material and the geometric parameters of the hole wall, the traditional static texture analysis is transformed into modeling the local reflectivity dynamic modulation process caused by the elastic displacement of the hole wall under the action of light pressure. This effectively overcomes the information limitation problem caused by the image feature extraction under a single illumination or fixed polarization condition in the existing technology. This method enables the gray-scale response difference at the same location under different excitation conditions to sensitively reflect the comprehensive state of the coating density and micro-flatness, significantly improving the recognition sensitivity of defects from nanometer to micrometer scale. In particular, it shows stronger discrimination ability in distinguishing cross-scale structural features such as grain orientation, submicron pores and macro-coating undulations, and solves the problem of feature redundancy and semantic ambiguity caused by the lack of physical mechanism guidance in traditional methods.

[0015] (2) Based on the slope characteristics, monotonicity and inter-frame stability of the "response fingerprint" of each grid cell, three physical sensitive scale domains of fine, medium and macro are automatically divided. The corresponding texture feature extraction channel is activated only when the trend consistency and high signal-to-noise ratio are met, realizing the paradigm shift from passive full calculation to active condition triggering. After each channel independently extracts features such as Gabor direction response, local contrast gradient histogram and Fourier frequency domain energy spectrum, nonlinear gating output is further performed through the confidence of the response fingerprint, avoiding the scale conflict and information submersion phenomenon caused by manually setting weights in the traditional weighted fusion strategy. This mechanism not only greatly reduces the invalid calculation overhead and improves the real-time response capability of the system, but also ensures the strong correlation between the extracted features and their underlying physical causes, and enhances the interpretability and robustness of feature expression. It is especially suitable for the quality assessment scenario of non-steady coating under complex process fluctuations.

[0016] (3) A lightweight graph convolutional network is adopted, with the mesh nodes of the 3D reconstruction of the hole wall as the vertices of the graph structure and the edge weights constructed by the response fingerprint similarity, to realize the cross-node texture feature propagation and aggregation, and generate the hole wall texture embedding vector with spatial physical consistency. This breaks through the problem of spatial context breakage caused by the traditional method relying on manual design of global statistical features or local window operation. The graph structure modeling method fully respects the mechanical consistency constraint of continuous deformation of the hole wall surface, so that the feature perturbation of the abnormal area can be reasonably diffused and aggregated and enhanced between neighboring nodes, thereby improving the detection rate of weak defects. The entire technical path abandons the general frameworks such as pyramid decomposition, attention mechanism, GLCM parameter adaptive adjustment, and multi-sensor data fusion, and instead endogenizes image analysis into the mechanical-optical coupling behavior of the object under test, forming a new detection paradigm driven by physical laws rather than dominated by mathematical transformation. It has good scalability and process adaptability, and can be stably deployed without repeated parameter tuning for different PCB types.

[0017] The above technical solution, by constructing a closed-loop analysis link of "excitation-response-decoupling-gating-aggregation", elevates the quality inspection of the inner wall of PCB vias from surface texture recognition to the level of physical response mechanism identification for the first time. It achieves accurate separation of cross-scale features and clear physical meaning of feature expression, significantly enhancing detection accuracy and system robustness while ensuring computational efficiency, and providing a new and reliable technical path for online quality monitoring of high-density interconnected printed circuit boards. Attached Figure Description

[0018] Figure 1 This is the main flowchart of a computer vision-based real-time evaluation method for the uniformity of copper plating on PCB hole walls.

[0019] Figure 2 This is a sub-flowchart of a computer vision-based real-time evaluation method for the uniformity of copper plating on PCB hole walls.

[0020] Figure 3 This is another sub-flowchart of a computer vision-based real-time evaluation method for the uniformity of copper plating on PCB hole walls. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this application provides a computer vision-based real-time evaluation method for the uniformity of copper plating on PCB via walls, specifically including: S1: Using a ring-shaped LED array light source with controllable polarization angle and microsecond-level pulse modulation capability, the inner wall of the PCB via is excited by a multi-angle, multi-polarization state sequence. Simultaneously, a high-resolution line scan camera is triggered to acquire and record the sequence of reflected light intensity from the via wall, including the incident angle of the light source, polarization direction, and pulse timing markers.

[0024] S2: Based on the reflected light intensity sequence of the hole wall and the known Young's modulus, Poisson's ratio, copper thickness parameters of copper material and geometric curvature radius of the hole wall, a hole wall micro-deformation response model is constructed that maps the micron-level elastic displacement distribution of the hole wall surface under light pressure to a local reflectivity modulation function.

[0025] S3: Using the three-dimensional reconstructed mesh of the aperture wall as a reference frame, the reflected light intensity sequence of the aperture wall is input into the micro-deformation response model of the aperture wall for inverse decoupling. The slope, saturation threshold and hysteresis interval of the response curve of each mesh unit under different illumination conditions are fitted to generate a response fingerprint that characterizes the comprehensive physical state of local copper layer compactness and micro-flatness.

[0026] S4: Based on the steepness, monotonicity, and repeatability of the response fingerprint, three physically sensitive scale domains are defined, corresponding to the high steepness domain, medium slope domain, and significant hysteresis domain, respectively, for nanoscale grain orientation differences, submicron-level pore distribution, and micron-level coating undulations.

[0027] S5: Perform trend consistency determination and signal-to-noise ratio threshold comparison on the response fingerprints of each physical sensitive scale domain acquired in three consecutive frames, and generate a gating enable signal characterizing whether each scale domain meets the activation conditions; use the gating enable signal to filter the response fingerprints, and independently perform Gabor direction response, local contrast gradient histogram and Fourier frequency domain energy spectrum extraction only on the scale domains that meet the activation conditions, and generate multi-scale texture feature channels output by nonlinear gating.

[0028] S6: The multi-scale texture feature channels are fed into a graph convolutional network. Using the hole wall mesh nodes as graph structure vertices and the response fingerprint similarity as edge weights, a cross-node texture propagation aggregation operation is performed to generate a hole wall texture embedding vector with spatial physical consistency.

[0029] S7: Based on the nonlinear mapping relationship between the hole wall texture embedding vector and the copper plating process parameters, calculate the dynamic optimization factor used to compensate for the delay in image features and process control response, and generate process correction instructions for fluctuations in the copper plating process.

[0030] S8: If the confidence level of the dynamic tuning factor is detected to be lower than the preset stability threshold, the issuance of the process correction instruction is paused and the model self-update process is triggered; otherwise, the process correction instruction is sent to the copper plating control system to perform closed-loop feedback optimization.

[0031] Step S1: Using a ring-shaped LED array light source with controllable polarization angle and microsecond-level pulse modulation capability, the inner wall of the PCB via is excited sequentially from multiple angles and polarization states. Simultaneously, a high-resolution line scan camera is triggered to acquire and record the sequence of reflected light intensity from the via wall, including the incident angle, polarization direction, and pulse timing markers. Specifically, this includes: S1.1: Initialize and configure the control parameters of the ring LED array light source, set the multi-angle incident angle sequence, multi-polarization state direction sequence and microsecond-level pulse timing mark, and generate a light source excitation instruction set for exciting the inner wall of the PCB through hole.

[0032] The firmware interface parameters of the ring LED array light source control system are initialized and called, and the light source position matrix and adjustable polarization angle range data are loaded to establish the basic state mapping table for light source control.

[0033] Based on the geometric center coordinates and aperture size data of the PCB through-hole, the light source geometry layout calculation module is called to generate a multi-angle incident angle sequence to ensure that the laser vector direction covers the entire circumference of the hole wall and the incident angle variation step size meets the optical uniformity constraint.

[0034] Based on the optical anisotropy characteristics of the copper layer and the polarization state sensitive response curve, a multi-polarization state direction sequence is set, the polarization direction is discretized into several equally distributed angles, and mapped to the polarization adjustment control register of the light source array driving unit.

[0035] Utilizing the high-frequency pulse modulation capability of the light source driving circuit, and based on the synchronization requirements of the imaging system for exposure time, microsecond-level pulse timing markers are generated. The timing values ​​are calculated using the following formula: Where Δt is the duration of a single pulse and f is the pulse modulation frequency, ensuring that the pulse duration is consistent with the trigger interval of the acquisition system.

[0036] The generated multi-angle incident angle sequence, multi-polarization state direction sequence, and microsecond-level pulse timing mark are combined into a structured control instruction set. The instruction set is stored in a hierarchical coding structure to facilitate subsequent sequence excitation execution and synchronous calling of the light source drive.

[0037] Through the above processing method, the result of the previous step is transformed into an executable set of light source excitation instructions, enabling precise control of the laser incident angle, polarization state, and pulse time.

[0038] S1.2: Drive and control the ring LED array light source based on the light source excitation instruction set, perform multi-angle, multi-polarization state sequence excitation operations, and generate a modulated excitation beam carrying specific physical state information.

[0039] Using the light source excitation instruction set output from the preceding steps as the driving input, the multi-channel driving control interface of the ring LED array light source is invoked. The corresponding angle setpoint is loaded onto the incident angle motor adjustment module of each channel, and closed-loop position control is executed to ensure that the beam pointing accurately conforms to the spatial vector requirements of the instruction set. For the polarization modulation components of each channel, the phase delay parameters of the liquid crystal polarization elements are adjusted according to the polarization direction sequence in the instruction set. The specified polarization state output is achieved through voltage modulation, ensuring that polarization state switching is completed within microseconds. The pulse modulation power supply module is driven by voltage according to the microsecond-level pulse timing marks in the instruction set. A high-frequency pulse signal with variable width is used to control the light emission period of the LED array and precisely adjust the peak power to form a beam with stable energy characteristics and accurate time resolution. Real-time status detection is performed on the above multi-angle, multi-polarization state, and pulse parameters. The execution errors of each adjustment module are collected and fed back to the driving control module. The output is corrected based on a proportional-integral-derivative adjustment strategy to eliminate execution errors. After all channels have been executed, the beams from each channel are synthesized to form a multi-angle, multi-polarization state modulated excitation beam, which carries specific physical state information corresponding to the instruction set. Through the above-mentioned driving control method, the light source excitation instruction set of the previous step is transformed into a physically realizable modulated excitation beam with precise spatial, polarization and temporal characteristics, thereby realizing controllable optical excitation of the physical state of the inner wall of the PCB via.

[0040] For example, in a PCB copper plating quality inspection scenario, a ring LED array light source includes 16 independently controllable light-emitting channels with incident angle sequences of 0°, 22.5°, 45°, and 67.5°, polarization direction sequences of 0°, 45°, 90°, and 135°, and microsecond-level pulse widths set to 3μs, 5μs, 7μs, and 10μs. During drive control, the closed-loop accuracy of angle control for each channel motor is ±0.05°, the voltage modulation range of the liquid crystal polarization element is 0-5V with an accuracy of 0.01V, and the peak power of the pulse power supply is 1.2W, 1.5W, 1.8W, and 2.0W corresponding to different pulse widths. Status detection revealed a maximum skew angle error of 0.03° and a maximum polarization direction error of 0.5°. After PID correction, these errors were reduced to 0.01° and 0.1°, respectively, and the pulse peak power error was less than 0.02W. Substituting the above parameters into the beam modulation model, the polarization state switching delay of the beam is less than 4μs when the incident angle is ≤45° and less than 6μs when the incident angle is >45°. The energy output stability is significantly improved throughout the entire sequence. The resulting multi-angle, multi-polarization state modulated excitation beam has high stability, high resolution and precise phase control, and can clearly capture the microscopic details of the aperture wall in subsequent synchronous exposure acquisition.

[0041] S1.3: The modulated excitation beam is used to irradiate the inner wall of the PCB via, triggering a high-resolution line scan camera to perform synchronous exposure and acquisition, thereby obtaining the original reflected light intensity signal of the via wall that reflects the microscopic state of the via wall surface.

[0042] The input conditions include a modulated excitation beam generated by a ring-shaped LED array light source and modulated by multi-angle, multi-polarization, microsecond-level pulses, as well as the geometric curvature, material reflection characteristics, and surface state information of the inner wall of the PCB via as the irradiation object. Based on the above input, the synchronous trigger interface of the high-resolution line scan camera is invoked to time-lock the light source pulse trigger signal with the exposure command of the camera control unit, ensuring that the camera completes the exposure start operation the instant the modulated beam reaches the via wall. The line scanning speed of the camera's photosensitive array is precisely set to match the beam scanning path, avoiding inter-line displacement errors caused by scanning speed mismatch during imaging. According to the predetermined incident angle and polarization direction, the spatial incident coverage range of the beam on the inner wall of the via is adjusted, and the spectral response curve of the camera is used to match the band energy distribution of the beam to maximize the effective capture rate of the reflected light intensity signal. Instantaneous current integration and analog-to-digital conversion are performed on the charge signal collected by each row of pixels to quantize the corresponding light intensity value into a digital grayscale value and store it in the frame buffer to form a two-dimensional matrix representation of the original via wall reflected light intensity. Through the above processing method, the modulation beam illumination result of the previous step is converted into the original aperture wall reflected light intensity signal carrying the microscopic physical state information of the aperture wall surface, realizing the conversion of physical state into an analyzable digital signal.

[0043] S1.4: Perform spatiotemporal alignment processing on the original aperture wall reflected light intensity signal and the light source incident angle, polarization direction and pulse timing mark in the light source excitation instruction set to generate an aperture wall reflected light intensity sequence containing complete physical metadata.

[0044] S1.5: Perform data integrity verification and noise suppression processing based on the hole wall reflected light intensity sequence, and output a standard hole wall reflected light intensity sequence for constructing a hole wall micro-deformation response model.

[0045] Step S2: Based on the reflected light intensity sequence of the aperture wall and the known Young's modulus, Poisson's ratio, copper plating thickness parameters, and aperture wall geometric curvature radius, a micro-deformation response model of the aperture wall is constructed that maps the micrometer-level elastic displacement distribution on the aperture wall surface under light pressure to a local reflectivity modulation function. Specifically, this includes: S2.1: Obtain the aperture wall reflected light intensity sequence containing the incident angle, polarization direction and pulse timing mark of the light source, as well as the known Young's modulus, Poisson's ratio, copper plating thickness parameters and aperture wall geometric curvature radius of the known copper material, perform photon momentum transfer analysis on the aperture wall reflected light intensity sequence to calculate the equivalent illumination pressure distribution under each excitation condition, thereby generating equivalent illumination pressure distribution data characterizing the external excitation intensity.

[0046] Based on the standard aperture wall reflected light intensity sequence from the previous steps, and combined with the known Young's modulus, Poisson's ratio, copper plating thickness parameters, and aperture wall geometric curvature radius, a physical property matching relationship of the input data is established to determine the initial conditions for the analysis.

[0047] The sequence of reflected light intensity from the aperture wall is grouped according to the incident angle of the light source, the polarization direction, and the pulse timing mark, to ensure that each group of data has a unique illumination excitation identifier and can be traced back to the specific imaging physical state.

[0048] Using the principle of photon momentum transfer, the light pressure value acting on the surface of the hole wall is calculated for each set of light intensity data. The radiation pressure equation is used to convert the light intensity and reflection characteristics into pressure distribution.

[0049] For multi-angle incident conditions, a directional decomposition of the pressure vector is established, decomposing the pressure value into two components along the normal and tangential directions to characterize the local pressure coupling effect under different incident angles.

[0050] By combining polarization direction information and introducing polarization correlation correction coefficients, directional weighting is applied to the pressure distribution to ensure that the pressure calculation on the surface of polarization-sensitive materials accurately reflects the interaction between light and materials.

[0051] The corrected pressure distribution is mapped onto the three-dimensional curvature model of the borehole wall, and spatial projection is performed using the curvature radius data to generate an equivalent illumination pressure distribution matrix on the geometric surface of the borehole wall.

[0052] By analyzing photon momentum transfer and correcting incident parameters, the light intensity sequence reflected from the aperture wall in the previous step is transformed into equivalent light pressure distribution data that can be used for mechanical modeling, thereby quantifying the intensity of external excitation.

[0053] S2.2: Based on the equivalent illumination pressure distribution data and the known Young's modulus, Poisson's ratio, copper plating thickness parameters, and geometric curvature radius of the hole wall, the stress-strain tensor transformation is performed using the linear elasticity constitutive equation to derive the micron-level elastic displacement field of the hole wall surface under multidimensional illumination excitation, thereby generating micron-level elastic displacement field data characterizing the micro-deformation state of the hole wall.

[0054] Based on the equivalent light pressure distribution data and the known Young's modulus, Poisson's ratio, copper plating thickness parameters, and hole wall geometric curvature radius of the copper material, the linear elastic mechanical constitutive equation is selected as the stress-strain calculation framework, and the equivalent light pressure vector field under each excitation condition is used as the external load input.

[0055] A three-dimensional stress tensor representation is established for each local surface element corresponding to each radius of curvature. The assumption of a constant thickness copper layer is adopted and a Poisson effect correction coefficient is introduced to ensure that the coupling relationship between longitudinal and transverse strain calculations conforms to the actual material properties.

[0056] Based on the linear elastic constitutive relation in a three-dimensional coordinate system, the stress tensor is mapped to the Young's modulus matrix of the material by matrix multiplication to generate the corresponding three-dimensional strain tensor. The strain components must satisfy the volume continuity equation to avoid numerical divergence.

[0057] The deformation amplification factor is calculated by calling the copper plating layer thickness parameter and the radius of curvature parameter. The deformation amplification factor is used to proportionally correct each strain component, thereby reflecting the modulation effect of the copper plating thickness on the local displacement response.

[0058] By combining the corrected strain tensor with the initial geometry of the borehole wall, a micrometer-scale elastic displacement field data matrix is ​​generated node by node through displacement calculation.

[0059] Through the above processing method, the equivalent light pressure distribution data of the previous step is transformed into micron-level elastic displacement field data that characterizes the microscopic deformation state of the hole wall, thereby realizing a quantitative mapping between light excitation conditions and physical displacement response.

[0060] S2.3: Based on the micron-level elastic displacement field data and the geometric curvature radius of the hole wall, Fresnel reflection law and micro-surface element normal vector correction method are used to perform local surface normal distortion calculation to quantify the reflection angle offset caused by elastic displacement, thereby generating a set of local surface normal distortion parameters characterizing the change of geometric optical path.

[0061] Based on known micrometer-level elastic displacement field data and the geometric curvature radius of the hole wall, microscopic deformation information is used as input for geometric optics calculations. Local surface micro-element is selected as the analysis unit to establish a quantitative relationship between the normal vector and the external excitation displacement disturbance. The initial normal value is calculated under curvature radius constraints for the three-dimensional coordinates and displacement vector of the center point of each micro-element, obtaining the reference normal direction data under undeformed conditions. Using the reference normal direction data and the actual displacement vector as the computational objects, the normal vector correction method is called to update the normal direction of each micro-element through a vector-weighted iterative method, reflecting the surface micro-angle shift caused by light pressure. The corrected normal direction is substituted into Fresnel's law of reflection to calculate the reflection angle shift value, as shown in the following formula: Where Δθ is the reflection angle offset, and n' is the corrected normal vector. The reference normal vector is used. By combining the reflection angle offset with the incident light direction vector, the local geometric optical path change matrix is ​​derived, which serves as the final output set of local surface normal distortion parameters. Through the above vector correction and Fresnel reflection law calculation, the physical mechanical deformation results are transformed into quantitative optical parameters reflecting the changes in the light path, achieving precise quantification of the changes in the aperture wall's geometric optical path.

[0062] S2.4: Based on the local surface normal distortion parameter set and the copper plating thickness parameter, a copper layer density dielectric constant perturbation factor is introduced to perform complex amplitude reflection coefficient remapping processing to eliminate macroscopic curvature interference and extract the reflectivity fluctuation component determined by microscopic flatness, thereby generating a local reflectivity modulation function characterizing the microscopic physical state of the copper layer.

[0063] Based on the local surface normal distortion parameter set and copper thickness parameters obtained in the previous steps, the corresponding dielectric constant perturbation factor is extracted from the copper layer density database as the input for optical reflectivity correction. The local surface normal distortion parameter set is converted into an expression for the path difference between incident and reflected light, and the nominal reflectivity of each pixel position under different incident conditions is calculated using the full-vector complex amplitude reflectivity formula. A macroscopic curvature compensation model is established, and the global offset caused by curvature in the path difference expression is removed as a separation term to ensure that only the local reflectivity fluctuation component caused by microscopic flatness is retained. The dielectric constant perturbation factor is introduced to perform complex amplitude remapping on the reflectivity after curvature removal. The complex amplitude data generated during the remapping process is decomposed into real and imaginary parts. The component related to reflection intensity is extracted from the real part, and the component related to phase change is retained in the imaginary part. Independent quantization of microscopic flatness is achieved through signal separation. Based on the analytical result of this complex amplitude component, it is encapsulated into a local reflectivity modulation function to characterize the microscopic physical state of the copper layer, ensuring that the function can accurately reflect the physical consistency of the same position under different laser incident conditions. By using complex amplitude remapping processing, the geometric optical path change parameters from the previous step are transformed into local reflectivity modulation data that eliminates macroscopic curvature interference, thereby enabling accurate extraction of microscopic smoothness fluctuation components and providing highly reliable input for subsequent multiphysics residual verification.

[0064] S2.5: Based on the local reflectivity modulation function and the reflected light intensity sequence of the aperture wall, perform multi-physics field coupling residual verification and parameter iterative optimization to correct the nonlinear error terms in the model and ensure the causal consistency between displacement distribution and reflectivity change, thereby generating a final usable aperture wall micro-deformation response model instance.

[0065] The 3D reconstructed mesh of the aperture wall is a three-dimensional spatial discretization representation constructed during the offline calibration or initialization phase, using the sequence of reflected light intensity from the aperture wall acquired under multi-angle and multi-polarization state excitation conditions, combined with the imaging geometry model of the line scan camera. The construction process is as follows: I. Data Preparation and Camera Calibration Before inspecting PCB through-holes, the high-resolution line scan camera is first calibrated internally and externally to obtain the camera's focal length, principal point coordinates, distortion coefficient, and spatial pose relative to the ring LED array light source. Simultaneously, multi-angle reflected light intensity images are acquired using standard aperture plates of known dimensions (such as calibration samples with different aperture diameters of 0.3mm, 0.5mm, and 0.8mm) to establish a mapping relationship between image pixel coordinates and the actual three-dimensional coordinates of the aperture wall.

[0066] II. Epipolar matching and depth estimation of multi-view light intensity sequences The sequence of reflected light intensity from the aperture wall generated by S1 is grouped according to the incident angle of the light source, forming an image sequence of the aperture wall from different viewpoints. Epipolar correction and feature matching are performed on the images of the same aperture wall location at different incident angles, and the depth value of each pixel is calculated using the triangulation principle in stereo vision. Since the aperture wall is an approximately cylindrical surface, the depth information changes continuously along the periphery of the aperture. The structured light principle based on active illumination or the multi-view photometric stereo technique can be used, combined with the known reflection characteristics of copper material (such as the BRDF model), to optimize the solution and obtain the three-dimensional coordinates of the aperture wall surface corresponding to each pixel.

[0067] III. Point Cloud Generation and Mesh Reconstruction The 3D coordinates of all pixels are merged to form a point cloud of the hole wall surface. After denoising, outlier removal, and hole wall cylindrical fitting correction, the discrete point cloud is connected into a continuous triangular mesh using Poisson surface reconstruction or Delaunay triangulation. The vertex density of the mesh can be controlled according to the detection accuracy requirements (e.g., no less than 100 vertices per square millimeter), and each vertex is assigned a unique index number and corresponding 3D spatial coordinates (x, y, z). This mesh is the "hole wall 3D reconstruction mesh".

[0068] IV. Mesh Structure Description The 3D reconstruction mesh for the aperture wall is a two-dimensional manifold mesh composed of a set of vertices and a set of faces. Its topological structure approximates a cylindrical surface, arranged regularly along the aperture axis and periphery. Each vertex of the mesh corresponds to a tiny physical region on the aperture wall. Subsequent steps, such as acquiring reflected light intensity sequences and response fingerprints, will be mapped to this vertex, thus achieving a one-to-one correspondence between spatial location and physical state. The mesh resolution (number of vertices) can be set according to the actual detection accuracy and computational resources, with a typical value of approximately 500–2000 vertices per aperture.

[0069] For example, in the inspection of a 0.3mm diameter PCB via, a line scan camera acquired the reflected light intensity sequence of the via wall at four different incident angles (0°, 22.5°, 45°, and 67.5°). After epipolar matching and depth estimation, a point cloud of approximately 1200 points was generated. Poisson reconstruction yielded a 3D reconstructed mesh of the via wall containing 1024 vertices and approximately 2000 triangular facets. The vertices of this mesh are distributed in 32 layers along the via axis and 32 columns along the periphery, forming a 32×32 regular mesh structure. Each vertex has a unique index (such as (row, column) or a linear index). This mesh serves as a reference frame in S3, used to map the multi-angle light intensity sequence to each mesh vertex, generating a set of light intensity response vectors for each mesh cell.

[0070] Through the above construction method, this scheme generates a three-dimensional reconstruction mesh of the hole wall in the offline or initialization stage, providing an accurate geometric reference framework for the subsequent reverse decoupling and spatial consistency analysis of the response fingerprint.

[0071] like Figure 2 As shown, step S3 involves using the 3D reconstructed mesh of the aperture wall as a reference frame, inputting the reflected light intensity sequence of the aperture wall into the aperture wall micro-deformation response model for inverse decoupling, fitting the slope, saturation threshold, and hysteresis interval of the response curve for each mesh unit under different illumination conditions, and generating a response fingerprint characterizing the comprehensive physical state of local copper layer compactness and micro-flatness. Specifically, this includes: S3.1: Obtain the coordinate data of the three-dimensional reconstructed mesh of the aperture wall and the reflected light intensity sequence of the aperture wall containing the incident angle and polarization direction of the light source. Based on the spatial projection transformation method, map the multi-dimensional light intensity sequence to each grid cell vertex of the three-dimensional reconstructed mesh of the aperture wall to generate a set of grid cell light intensity response vectors carrying multi-angle excitation information.

[0072] Obtain the 3D reconstruction mesh coordinate data of the aperture wall from the previous step, as well as the aperture wall reflected light intensity sequence containing the incident angle and polarization direction markers of the light source, and establish a correspondence between the mesh vertex coordinates and the spatial attribute labels in the light intensity sequence.

[0073] A spatial projection matrix from the outer surface of the aperture wall to the imaging plane is constructed based on a three-dimensional geometric coordinate system. The matrix parameters are dynamically adjusted by the incident angle and polarization direction markings to adapt to the changes in the optical path under multi-angle excitation conditions.

[0074] Perform spatial projection transformation, using the projection matrix to map the reflected light intensity sequence frame by frame to the vertex positions of the 3D reconstructed mesh, and bind the light intensity value of each vertex at different incident angles.

[0075] The mapping results are stored using a multidimensional array structure, with the dimensions corresponding to the grid cell index, the incident angle sequence index, and the polarization direction sequence index, forming a light intensity response data volume with clear physical redundancy.

[0076] Angle normalization processing is performed on the multidimensional light intensity response data volume, and the normalized light intensity value is calculated.

[0077] The normalized light intensity values ​​are aggregated by grid cells, and grid cell light intensity response vectors are generated by indexing the incident angle and polarization direction to ensure that the vector indexes correspond one-to-one with the physical excitation conditions.

[0078] Through the above spatial projection and normalization processing, the three-dimensional reconstruction data of the hole wall and the reflected light intensity sequence of the previous step are transformed into a set of grid cell light intensity response vectors that carry multi-angle excitation information and are adapted to the input requirements of the micro-deformation response model, so as to achieve the high-precision data input effect of subsequent reverse decoupling.

[0079] S3.2: Call the constructed micro-deformation response model of the hole wall, take the light intensity response vector set of the grid cell as the input variable, and reconstruct the gray-scale response trajectory of each grid cell under different light pressures based on the nonlinear least squares fitting method to generate a set of grid cell response curve parameters including the initial slope factor, the maximum saturation threshold and the hysteresis width.

[0080] Using the set of light intensity response vectors of the grid cells generated in the previous steps as input conditions, the pre-constructed micro-deformation response model of the aperture wall is called. The parameter instances of the equivalent illumination pressure distribution and the local reflectivity modulation function are loaded inside the model to establish a nonlinear mapping relationship between the input light intensity and the physical state.

[0081] The light intensity response vector set of the grid cells is input into the multiphysics coupling operator of the model. The model is driven by multidimensional illumination pressure as a control variable to generate the predicted grayscale sequence of each grid cell under different excitation conditions. The actual acquired response vector is used as the target sequence for error calculation.

[0082] A nonlinear least squares fitting method is adopted, and the fitting curve of the gray-level response trajectory of each grid cell is constructed based on the criterion of minimizing the sum of squared residuals. This allows the shape of the fitting curve to simultaneously reflect the initial illumination sensitivity, gray-level saturation trend, and hysteresis characteristics of the unloading process.

[0083] During the fitting process, the cost function is set as follows: in To fit the curve under light pressure The predicted grayscale value is below. To observe grayscale values ​​in order to obtain the optimal combination of curve parameters.

[0084] The local first derivative value of the fitted curve near zero light pressure is extracted as the initial slope factor. The light pressure position of the curve when the gray value approaches saturation is determined as the maximum saturation threshold. The width of the pressure interval that the fitted curve delays falling back during the unloading process is quantified as the hysteresis width.

[0085] Through the above processing method, the light intensity response vector set of the grid cells in the previous step is transformed into a set of grid cell response curve parameters containing the initial slope factor, the maximum saturation threshold and the hysteresis width, so as to achieve a quantitative description of different physical sensitivity characteristics.

[0086] S3.3: Based on the initial slope factor in the parameter set of the grid cell response curve and the Young's modulus constraint of the copper material, the micron-level surface displacement induced by illumination is calculated using the elasticity inversion mechanism, and the displacement is converted into a local reflectivity modulation coefficient to generate a compactness mapping index characterizing the density of the copper layer grain orientation.

[0087] The initial slope factor in the parameter set of the grid cell response curve is obtained as the quantitative starting point of the gray-scale change rate under incident light conditions, and a mechanical constraint matrix is ​​constructed in combination with the preset value of Young's modulus of copper material to limit the elastic boundary of displacement calculation.

[0088] By invoking the elasticity inversion mechanism, the initial slope factor and the corresponding light pressure distribution parameters are embedded into the inverse solution of the stress-strain tensor, and the micrometer-level displacement field vector of the unit vertex under multi-directional light illumination is calculated.

[0089] The displacement field vector is geometrically projected using a local curvature correction method to ensure that the displacement is distributed along a true physical path under the constraint of the hole wall curvature radius.

[0090] By substituting the corrected micrometer-level displacement into the Fresnel reflectivity model and establishing a nonlinear mapping relationship between the displacement and reflectivity changes, the local reflectivity modulation coefficient is calculated.

[0091] The local reflectivity modulation coefficients of each grid cell are normalized to generate a compactness mapping index that can be compared under different lighting conditions, serving as a quantitative characterization of the grain orientation density.

[0092] Through the above processing method, the response curve parameter set of the previous step is transformed into a compactness mapping index, thereby realizing the physical quantification of the local copper layer grain orientation density.

[0093] For example, for a PCB sample with an aperture of 0.3 mm, a radius of curvature of 0.15 mm, and a copper plating thickness of 15 μm, the initial slope factor of the grayscale response acquired when the incident angle of the input light source is 45° is 0.08, and the Young's modulus of the copper material is set to 110 GPa. Elasticity inversion calculations yielded a unit vertex displacement of 0.35 μm under this incident condition, which stabilized at 0.33 μm after curvature correction. Substituting this displacement into the Fresnel reflectivity mapping formula, with a static reflectivity baseline R0 = 0.52 and a material optical sensitivity coefficient k = 0.04 / μm, the modulated reflectivity R = 0.5332 was obtained. Normalizing the reflectivity modulation value of each unit according to the global maximum and minimum reflectivity range yielded a compactness mapping index value of 0.64. This index is directly comparable under different lighting conditions and grid positions. During verification, it can significantly improve the ability to identify loosely arranged grain regions, and the output is a complete compactness mapping matrix, which can be applied to subsequent texture feature sensitive scale segmentation. S3.4: Combining the loop hysteresis width and maximum saturation threshold in the parameter set of the grid cell response curve, analyze the multiple scattering effect of submicron-level pores on the optical path, calculate the smoothness deviation value reflecting the degree of micro-undulation of the coating surface, and generate a smoothness mapping index characterizing the micro-geometric state of the copper layer surface.

[0094] The input conditions are the loop hysteresis width and maximum saturation threshold data in the parameter set of the grid cell response curve, as well as the hole wall material parameters and the known uniform light source characteristics of the optical measurement environment.

[0095] Optical energy differential calculations are performed based on the loop hysteresis width and the maximum saturation threshold to obtain the nonlinear deviation curve parameters between energy input and light intensity output during loading and unloading.

[0096] Using the above deviation curve parameters as the incident conditions for multiple scattering, the optical path deviation caused by submicron-level pores is calculated, and the scattering path length distribution function is updated.

[0097] By introducing a light intensity attenuation coefficient and a phase delay factor into the scattering path length distribution function, and using the energy conservation constraint of multiple scattering to solve for the percentage of light intensity loss caused by surface undulations, the flatness deviation value is calculated using the following formula: Where ΔP is the flatness deviation value, and I is the current scattered light intensity. The reference light intensity is unfluctuated, and N is the number of sampling points.

[0098] The flatness deviation value is mapped to the two-dimensional surface geometric state parameter domain to generate a flatness mapping index that characterizes the micro-geometric state of the copper layer surface.

[0099] By constructing a flatness mapping index, the scattering effect is quantified into structured geometric information that can directly participate in the response fingerprint packaging, thereby achieving accurate characterization of the micro-undulations of the copper layer.

[0100] S3.5: Integrate the compactness mapping index and the flatness mapping index, and use a multi-dimensional feature encoding strategy to encapsulate the above two physical state indices into structured data objects, generating a response fingerprint that uniquely identifies the physical characteristics of each grid cell, which serves as the direct input basis for subsequent division of the physical sensitive scale domain.

[0101] Using the grid cell compactness mapping index and the flatness mapping index as input variables, a physical state fusion matrix is ​​constructed by calling a multi-dimensional feature encoding method module. The two physical state indices are normalized in the numerical domain to eliminate dimensional differences. The normalized compactness index vector and the flatness index vector are concatenated at the feature axis level, and principal component analysis is used to determine the contribution rate of each feature component in the overall physical state description. The two indices are then fused and encoded. Based on the fusion encoding results, a unique identifier generation rule is constructed, and each row of data in the fusion matrix is ​​hashed to generate a fixed-length structured data object identifier. This identifier, along with the fusion matrix values, is encapsulated into a response fingerprint data object, ensuring that the physical characteristics of each grid cell are associated with a unique fingerprint. By outputting the fusion encoding of the compactness and flatness indices as structured response fingerprint data, direct input data preparation for subsequent physical sensitive scale domain partitioning is achieved.

[0102] like Figure 3 As shown, step S4 involves dividing the response fingerprint into three physically sensitive scale domains—a high-steepness domain, a medium-slope domain, and a significant hysteresis domain—based on the steepness, monotonicity, and repeatability stability characteristics, corresponding to nanoscale grain orientation differences, submicron-level porosity distribution, and micron-level coating undulations, respectively. Specifically, this includes: S4.1: Perform first-order differential operation on the gray-scale response difference sequence contained in the response fingerprint to obtain a set of response curve slopes that characterize the rate of change of local reflectivity with illumination angle, and use this set of response curve slopes as the basic data input for quantifying the micro-deformation sensitivity of the copper layer surface.

[0103] A fixed-step differential operator is configured to perform differential operations on the grayscale response difference sequence contained in the response fingerprint data object generated in the previous steps, forming a first-order differential input matrix for calculating the reflectivity change rate. Angle sequence index mapping is introduced during the differential operation to ensure that each grayscale difference value is precisely associated with its corresponding incident illumination angle, thus eliminating slope errors caused by angular misalignment. Based on the paired grayscale difference values ​​and illumination angle sets, the numerical slope calculation module is called to perform linear regression fitting within the continuous illumination angle interval of each response fingerprint, obtaining initial values ​​of the local grayscale change rate and storing them as unit-level intermediate results. A secondary precision correction is performed on these intermediate results, employing piecewise regression and anomaly residual removal mechanisms to remove high-amplitude discrete points caused by local noise or acquisition anomalies, thereby stabilizing the slope estimation. The corrected slope value set is encapsulated into a structured slope index object, serving as the basic data input for further quantifying the sensitivity of the copper layer surface to microscopic deformation, realizing the conversion of local physical states into mathematical indices.

[0104] By employing slope difference operations, angle index pairing, linear fitting, and residual elimination processing, the grayscale difference data of the response fingerprint from the previous step is transformed into a set of response curve slopes that characterize the rate of change of local reflectivity with the illumination angle, thereby achieving a quantitative characterization of the sensitivity to microscopic deformation of the copper layer.

[0105] For example, in PCB copper via wall detection, the grayscale response difference sequence in the response fingerprint is sorted according to the incident angle sequence, with a difference step size set to 1°, and each sequence contains at least 20 sampling points. After difference calculation, a grayscale change matrix is ​​formed, where each row corresponds to the grayscale increase or decrease value of a grid cell under adjacent illumination angles. For each grid cell, a linear regression fitting model is used to calculate the local slope value. After outlier removal processing based on the threshold 3σ rule, a stable slope set is obtained for sensitivity quantification. When the input angle range is from 0° to 30°, the slope value sensitive to micro-deformation is usually higher than 0.8 grayscale / degree. This value is judged as a high-steepness domain feature in the subsequent scale domain division, thereby effectively supporting the identification of nanoscale grain orientation differences. The detection system's ability to identify subtle non-uniformities is significantly improved in this embodiment.

[0106] S4.2: Based on the set of slopes of the response curves, perform extreme point detection and linearity fitting methods to generate a steepness index that characterizes the degree of drastic change of the response fingerprint within a specific illumination range and a monotonicity index that characterizes the consistency of the change trend. The steepness index and the monotonicity index are used as key criteria to distinguish the differences in orientation of nanoscale grains.

[0107] The slope set of the response curve output from the preceding sub-step S4.1 is input as data, serving as the basis matrix for subsequent quantitative analysis of slope change trends. An extreme point detection method is invoked to perform differential sign change analysis on the slope sequence corresponding to each grid cell in the slope set, detecting all local maxima and minima, and establishing an extreme point location index table according to the illumination interval index. Based on the extreme point location index table, a least-squares straight-line fitting operation is performed on each response slope sub-interval using a linearity fitting method, calculating the mean square value of the fitting residuals and recording the average growth rate of the slope. The extreme value amplitude of each sub-interval is combined with the mean square value of the fitting residuals to perform a comprehensive evaluation operation, obtaining a steepness index characterizing the drastic change of the slope curve within a specific illumination interval. The sign sequence of the absolute values ​​of the slopes in each sub-interval is used as the monotonicity judgment input, and the monotonicity index is calculated using trend consistency judgment logic. The steepness index and the monotonicity index are packaged and encapsulated into a structured feature object, serving as a key criterion for distinguishing differences in nanoscale grain orientation.

[0108] By using extreme point detection and linearity fitting methods, the slope set of the response curve from the previous step is transformed into a quantitative index that characterizes the steepness and monotonicity of the curve's rate of change and trend consistency, thereby achieving accurate identification of sensitive regions with differences in nanoscale grain orientation.

[0109] S4.3: The temporal variance of the response fingerprint of the same grid cell is calculated by using the sequence of reflected light intensity from the hole wall acquired in multiple consecutive frames, so as to extract the repeatability stability index that characterizes the fluctuation amplitude of the copper layer density under different excitation conditions, and the repeatability stability index and the steepness index are used together as the characteristic basis for identifying the distribution of submicron porosity.

[0110] Temporal data extraction is performed on the response fingerprint of the same grid cell. The sequence of reflected light intensity from the aperture wall acquired in multiple consecutive frames is arranged in chronological order to form a grayscale response time vector for that grid cell.

[0111] Perform frame-by-frame difference calculation on the above grayscale response time vector, and take the difference between the grayscale response of each frame and the previous frame as the instantaneous fluctuation amplitude to obtain the time-series difference sequence.

[0112] Input the time series difference sequence into the variance calculation module, and use the following formula to solve for the fluctuation variance: Where m is the sequence length. Let be the instantaneous difference value of the i-th frame. σ is the mean of the instantaneous differences, and σ is the variance of the fluctuation.

[0113] The variance of the fluctuations is normalized to generate a repeatability stability index for the fluctuation amplitude of the copper layer density under different excitation conditions.

[0114] The above-mentioned repeatability stability index is correlated with the steepness index output in the preceding step S4.2. A multidimensional feature cross-validation strategy is used to evaluate the common pattern distribution of the two to identify response fingerprint samples that conform to the submicron pore distribution characteristics.

[0115] By using the time-series variance calculation and index fusion processing method, the slope characteristics of the response curve in the previous step are transformed into quantifiable stability data that characterizes the periodic fluctuation amplitude of the copper layer's density, thereby enabling automatic identification of sensitive areas for submicron-level porosity distribution.

[0116] For example, in a PCB copper plating process monitoring scenario, the continuous acquisition frame count for a single via mesh cell is set to 20 frames, with a frame interval of 0.5ms and a grayscale response value range of 0 to 255. Subtracting the grayscale response value of the previous frame from the grayscale response value of each frame forms a temporal difference sequence of length 19. The mean μ of the difference sequence is calculated to be 2.8, and the variance σ is calculated using the formula, yielding a value of 1.45. After normalization, the repeatability stability index is found to be 0.62 within the range of 0 to 1. This index, along with a steepness index of 0.75, is input into a submicron porosity feature recognition model. The model determines that the mesh cell simultaneously exhibits high steepness and moderate stability fluctuation, classifying it as belonging to the submicron porosity sensitive region. This determination can significantly improve porosity detection accuracy and reduce false positives in industrial experiments, achieving stable quality monitoring results.

[0117] S4.4: Based on the pulse energy threshold required for the grayscale response in the response fingerprint to reach saturation and the signal fallback lag interval during the unloading process, calculate the hysteresis significance parameter that characterizes the optical hysteresis effect caused by macroscopic fluctuations in the coating, and use the hysteresis significance parameter as the core quantification factor for defining micron-level coating fluctuations.

[0118] Pulse energy mapping processing is performed on the response fingerprint dataset, which includes the pulse energy threshold required for the grayscale response to reach saturation and the signal fallback hysteresis interval during the unloading process. The pulse energy corresponding to the maximum grayscale value of each grid cell is located as the optical saturation point and the value is recorded as the energy threshold input.

[0119] The pulse energy is calculated by multiplying the illumination power and the incident time based on the energy threshold input.

[0120] The time difference between the grayscale fallback time series of the unloading process and the corresponding point of the energy threshold is calculated, and the delay time required for the signal to fall back to the initial grayscale value is extracted as the hysteresis interval length parameter.

[0121] The hysteresis significance parameter is calculated by performing normalization based on the hysteresis interval length parameter and the energy threshold.

[0122] Multigrid statistical analysis was performed on the hysteresis significance parameter. The stability of the optical hysteresis effect caused by macroscopic coating fluctuations was measured by the mean and standard deviation. The obtained hysteresis significance parameter was used as the core quantitative factor to define micron-level coating fluctuations.

[0123] Through the above processing method, the response fingerprint data of the previous step is transformed into a hysteresis significance parameter that characterizes the hysteresis properties of macroscopic coating undulations, thereby achieving accurate identification of micron-level coating undulation sensitive areas.

[0124] For example, in a PCB through-hole inspection scenario, the pulse energy corresponding to the maximum grayscale value of the acquired grid cell grayscale response curve is 1.5 mJ. This value is used as the energy threshold input. The incident light illumination power is 150 mW, and the pulse duration is 10 ms. The pulse energy E is calculated to be 1.5 mJ using the formula. The grayscale fallback delay time ΔT during unloading is measured to be 0.25 ms. Substituting the above value into the hysteresis significance parameter calculation formula, H is obtained as 0.166. Statistical analysis of the hysteresis significance parameters of multiple grid cells shows that the mean is stable at 0.16, and the standard deviation is 0.01, indicating that the hysteresis effect caused by macroscopic undulations in the coating is consistent in most grid cells. The final output hysteresis significance parameter is used to directly identify micron-level coating undulation regions during subsequent scale domain division, thereby significantly improving the division accuracy of the macroscopic undulation-sensitive scale domain and optimizing the multi-scale recognition performance of texture analysis.

[0125] S4.5: Combining the steepness index, the monotonicity index, the repeatability stability index, and the hysteresis significance parameter, perform a multidimensional threshold clustering segmentation operation to automatically generate three physically sensitive scale domain boundaries that divide the hole wall texture space into a high steepness domain, a medium slope domain, and a hysteresis significance domain, and use the three physically sensitive scale domains as the activation basis for the subsequent generation of gating enable signals.

[0126] The multidimensional physical state quantification data, including steepness, monotonicity, repeatability, and hysteresis significance parameters, is processed as input loading and used as the feature vector set for clustering segmentation methods. A four-dimensional Euclidean distance metric space is established based on this feature vector set. Selective normalization eliminates the dimensional differences between different indicators, ensuring that the contribution weight of each indicator during clustering is controllable. Multidimensional threshold vectorization mapping is performed on the normalized indicator set, generating an initial label matrix for dividing the physically sensitive scale domains by referring to a preset threshold group corresponding to the physical meaning. Based on this initial label matrix, a multidimensional threshold clustering operator is invoked, using a threshold discrimination hyperplane to divide each data point into corresponding high-steepness, medium-slope, and hysteresis significance domains according to the indicator combination relationship, forming sensitive scale domain boundary data with spatial segmentation consistency. Boundary smoothing and topological consistency correction operations are performed on the divided sensitive scale domain boundary data to ensure that there are no isolated or overlapping abnormal regions between scale domains. By using multidimensional threshold clustering segmentation and boundary correction processing, the response fingerprint quantification index of the previous step is transformed into spatial boundary information of high steepness domain, medium slope domain and hysteresis significant domain, realizing the automatic division of the physical scale domain and providing activation basis for subsequent generation of gating enable signal.

[0127] Step S5: The trend consistency of the response fingerprints of each physically sensitive scale domain acquired in three consecutive frames is determined and compared with a signal-to-noise ratio threshold to generate a gating enable signal characterizing whether each scale domain meets the activation condition. The gating enable signal is used to filter the response fingerprints, and Gabor direction response, local contrast gradient histogram, and Fourier frequency domain energy spectrum extraction are independently performed only on the scale domains that meet the activation condition, generating a multi-scale texture feature channel output through nonlinear gating. Specifically, this includes: S5.1: Perform time-series sliding window traversal processing on the high-steepness domain response fingerprint, medium-slope domain response fingerprint, and hysteresis-significant domain response fingerprint acquired in three consecutive frames to obtain the response fingerprint sequence data of each physical sensitive scale domain in the time dimension.

[0128] S5.2: Based on the response fingerprint sequence data, perform a trend consistency determination method and a signal-to-noise ratio threshold comparison operation to generate a binary gating enable signal characterizing whether each physical sensitive scale domain meets the activation condition.

[0129] Based on the response fingerprint sequence data of each physical sensitive scale domain obtained by S5.1, the trend consistency determination method is called to perform a direction offset sign check operation on each scale domain sequence in the time dimension to determine whether the grayscale response change sign of three consecutive frames is consistent.

[0130] Based on the above judgment results, the absolute value calculation method of inter-frame amplitude difference is used to extract the change amplitude vector of consecutive frames, and the sequence samples with the same amplitude change direction but fluctuation amplitude below the effective detection limit are removed accordingly.

[0131] The signal-to-noise ratio (SNR) is calculated for the remaining sequence samples. The power spectrum estimation method is used to obtain the ratio of signal power to noise power, and the ratio is compared with the preset SNR threshold to form a preliminary judgment signal that meets or does not meet the SNR requirements.

[0132] Execute logic and combination operations on the trend consistency determination result and the signal-to-noise ratio determination result. Only when both are in a valid state is the state of the scale domain marked as active; otherwise, it is marked as disabled.

[0133] To form standardized control quantities that can be called upon in subsequent steps, the above-mentioned activatable / disabled states are mapped to binary gating enable signals, where the activation state corresponds to the value 1 and the disabled state corresponds to the value 0. This binary encoding enables precise control of feature channels in each scale domain.

[0134] By using a chain-like combination of trend consistency judgment and signal-to-noise ratio comparison, the time series data from the previous step is transformed into a binary gating enable signal that characterizes whether each physical sensitive scale domain meets the activation condition, thereby achieving precise start-stop control for subsequent texture feature extraction.

[0135] S5.3: Use the binarized gating enable signal to perform logical AND operation filtering on the high steepness domain response fingerprint, medium slope domain response fingerprint and hysteresis significant domain response fingerprint to output a subset of response fingerprints to be processed that contains only valid physical states.

[0136] Using response fingerprints from three consecutive frames acquired in the high-steepness, medium-slope, and hysteresis-signal domains as input, the binarized gating enable signal matrix generated in the previous sub-step is invoked to perform channel filtering on the response fingerprint sets of the three types of physically sensitive scale domains. An element-wise logical AND operation is performed between the binarized gating enable signal and the corresponding response fingerprint set to ensure that each grid cell retains its original grayscale response data only when the activation condition is met in its corresponding scale domain. The result set after the logical AND operation is indexed and rearranged to remove null or invalid data nodes caused by gating closure, forming a compact subset of response fingerprints with accurate mapping relationships. A structured encapsulation operation is performed on the rearranged subset, binding the grayscale response curve parameter groups and their physically sensitive scale domain labels of each grid cell into a unified data object, facilitating batch calls by the subsequent multi-domain texture feature extraction module. Through the above filtering and structured encapsulation processing, the results of the previous step are transformed into a subset of response fingerprints containing only valid physical states, achieving filtering of noisy data and preservation of physical relevance.

[0137] For example, in a PCB copper plating inspection scenario, the response fingerprint array for the high-steepness domain has a length of 100 grid cells, the response fingerprint array for the medium-slope domain has a length of 120 grid cells, and the response fingerprint array for the hysteresis-signal domain has a length of 90 grid cells. The binarized gating enable signal matrix has a dimension of 3×120, where the first row corresponds to the gating signal for the high-steepness domain, the second row corresponds to the medium-slope domain, and the third row corresponds to the hysteresis-signal domain. A logical AND operation is performed between the gating signal matrix and the response fingerprint set. For example, if the gating signal value of the 25th grid cell in the high-steepness domain is 0, the grayscale response curve at that position is invalidated and deleted during index rearrangement. After index rearrangement, the number of effective grid cells in the high-steepness domain is 68, in the medium-slope domain it is 95, and in the hysteresis-signal domain it is 72. The initial slope, maximum saturation threshold, hysteresis width, and other parameter groups of the grayscale response curve of each effective grid cell are assembled into a data object with a scale domain label, and a physical state identifier is added to the data structure. Verification showed that the selected subset significantly improved the effectiveness and physical interpretability of feature extraction during subsequent Gabor directional response filtering, ensuring the stability and accuracy of the multi-scale texture feature channel construction process.

[0138] S5.4: For the subset of response fingerprints to be processed, Gabor directional response filtering, local contrast gradient histogram statistics, and Fourier frequency domain energy spectrum decomposition operations are performed in parallel to generate the original multi-domain texture feature vectors corresponding to nanoscale grain orientation differences, submicron-level pore distribution, and micron-level coating undulations.

[0139] The subset of response fingerprints to be processed, after logical AND filtering, is loaded into the texture feature extraction parallel processing module, establishing a multi-threaded computation queue indexed by the physically sensitive scale domain label. For each computation queue, the Gabor directional response filtering operator is called, setting the filter center frequency and directional angular resolution based on the scale domain range obtained in the previous steps, and performing a two-dimensional convolution operation to extract the high-frequency texture energy distribution matrix consistent with the grain orientation direction. For the same computation queue, the local contrast gradient histogram computation operator is called, setting the gradient amplitude binning interval and brightness normalization window, calculating the gradient amplitude distribution of each pixel's neighborhood, and constructing a normalized histogram to generate statistical features characterizing the local contrast changes of the hole wall texture. For the same computation queue, Fourier frequency domain energy spectrum decomposition is performed. First, a fast Fourier transform is performed on the response fingerprint grayscale matrix to obtain the frequency domain complex amplitude spectrum, then the energy integral value in the radial frequency interval is calculated to form a frequency domain feature vector reflecting the spatial periodicity of the hole wall texture. By employing a feature domain concatenation method, the Gabor directional response matrix is ​​compressed into a directional energy feature vector. This vector is then sequentially concatenated with the local contrast gradient histogram feature vector and the Fourier frequency domain energy spectrum feature vector, according to scale domain labels, to generate the original multi-domain texture feature vectors corresponding to nanoscale grain orientation differences, submicron-level porosity distribution, and micron-level coating undulations. Through the aforementioned parallel filtering, statistical, and frequency domain decomposition processing methods, the effective physical state response fingerprint from the previous step is transformed into multi-domain texture feature data, enabling independent capture and refined expression of multi-scale physically sensitive features.

[0140] For example, in PCB copper plating quality inspection, the high-slope domain response fingerprint matrix is ​​input into a Gabor filter module, with the filter center frequency set to 8Hz, directional angular resolution to 15 degrees, convolution window size to 7×7 pixels, and output directional energy feature vector length to 8. The medium-slope domain response fingerprint matrix is ​​input into a local contrast gradient histogram module, with gradient magnitude binning interval set to 0.1 grayscale unit and histogram length to 16. The output contrast features reflect the brightness variation pattern caused by submicron pores. The hysteresis-significant domain response fingerprint matrix is ​​input into a Fourier frequency domain energy spectrum decomposition module, performing a 256×256-point FFT on the matrix, dividing the radial frequency into 5 intervals, calculating the energy integral of each interval, and outputting a 5-dimensional frequency domain feature vector. The above 8-dimensional directional energy, 16-dimensional local contrast, and 5-dimensional frequency domain energy vectors are concatenated according to scale domain labels to obtain a 29-dimensional original multi-domain texture feature vector. In this example, the high-frequency energy extracted by Gabor significantly enhances the peak value when characterizing the difference in grain arrangement direction, the local contrast feature significantly improves the gradient amplitude concentration when identifying the porosity distribution, the frequency domain energy spectrum shows a low-frequency energy peak in the detection of macroscopic fluctuations in the coating, and the final output feature set shows higher physical consistency and defect localization accuracy in subsequent graph convolution aggregation.

[0141] S5.5: Perform element-wise multiplication nonlinear gating fusion processing on the original multi-domain texture feature vector and the binarized gating enable signal to generate a multi-scale texture feature channel with spatial physical consistency and noise interference removal.

[0142] The original multi-domain texture feature vectors from the high-steepness, medium-slope, and hysteresis-signal domains are fused with their corresponding binarized gating enable signals according to an element-wise multiplication rule. The inputs are the multi-domain texture feature vector matrix generated in S5.4 and the gating signal matrix generated in S5.2. For the feature vectors of each scale domain, an element-wise multiplication operation matrix is ​​first defined, and each component of the feature vector is multiplied with the corresponding binarized gating signal to form a preliminary fusion matrix. Based on a nonlinear gating strategy, a response consistency confidence correction weight coefficient is introduced into the element-wise multiplication result. The correction coefficient and the multiplication result are then multiplied again element-wise to suppress local noise interference and enhance the feature components in physically consistent regions. For the corrected fusion results of all scale domains, normalization is performed to ensure that the amplitude distribution of each fusion channel is within a uniform scale range, avoiding weight shifts in the subsequent feature aggregation process. Using feature domain correlation analysis, the cross-correlation coefficients between each fusion channel are calculated, highly correlated redundant feature components are removed, and effective feature sets with independent physical meaning are retained. Through the above chain processing, the original feature vector generated by S5.4 and the gated signal are transformed into a multi-scale texture feature channel after nonlinear gated fusion, achieving the expected effect of spatial physical consistency and noise interference removal.

[0143] For example, in the real-time detection process of a PCB copper plating production line, the original multi-domain texture feature vector in the high-steepness domain is set as a vector of length 64, and the binarized gate signal is a Boolean sequence of length 64, in which the proportion of effective activation bits is 0.75. After element-wise multiplication, a preliminary fused vector is obtained, with 48 non-zero components. A consistency confidence correction weight coefficient is introduced, which is derived from the trend consistency detection results of three consecutive frames, with a value range of 0.65 to 0.95, and noise suppression is achieved through element-wise multiplication. In the normalization process, the amplitude value of the fused vector is mapped to the [0,1] interval. Using a cross-correlation coefficient threshold set to 0.85, redundancy removal is performed on the fused multi-scale texture feature channels, and finally 47, 52, and 45 effective feature components are retained in the high-steepness domain, medium-slope domain, and hysteresis significant domain, respectively. This processing result effectively improves the recognition of nanoscale grain differences and submicron pore distribution when matching spatial node data in subsequent graph convolutional networks, and significantly enhances the physical interpretability and anti-interference ability of global texture embedding vectors.

[0144] Step S6: The multi-scale texture feature channels are fed into a graph convolutional network. Using the hole wall mesh nodes as graph vertices and response fingerprint similarity as edge weights, a cross-node texture propagation aggregation operation is performed to generate a hole wall texture embedding vector with spatial physical consistency. Specifically, this includes: The Lightweight Graph Convolutional Network is a lightweight graph neural network model specifically designed for evaluating the uniformity of copper plating on PCB via walls. Its core function is to spatially propagate and aggregate multi-scale texture feature channels across the via wall mesh topology, generating a low-dimensional embedding vector that reflects the overall texture distribution pattern of the via walls. This network is designed to balance low computational overhead with high real-time performance, making it suitable for deployment in industrial edge computing devices.

[0145] I. Network Input and Output Input: Multi-scale texture feature channels output by S5 nonlinear gating, each grid node corresponds to a feature vector (dimension D, e.g., 12-dimensional). Topological graph structure matrix generated by S6.1, where nodes are vertices of the hole wall grid, and edge weights are the similarity of response fingerprints between adjacent nodes (value range [0,1]).

[0146] Output: Hole wall texture embedding vector (low-dimensional dense vector, dimension d) D (e.g., 32-dimensional), representing the texture distribution pattern of the entire pore wall region.

[0147] II. Network Structure The lightweight graph convolutional network consists of the following modules connected in series: Feature initialization layer: Loads the multi-scale texture feature vector of each mesh node as the initial node state vector. The dimension is D.

[0148] Graph Convolutional Layers (Lightweight Design): Employ simplified first-order graph convolution (similar to GCN) or graph attention mechanisms, adjusting the contributions of neighboring nodes through edge weights. The calculation formula for each graph convolution layer is: Where N(i) is the set of neighboring nodes of node i. For edge weights, Let i be the degree of node i. Let be the trainable weight matrix of the l-th layer, and σ be a non-linear activation function (such as ReLU). To reduce computational cost, this network uses only two layers of graph convolutions, and the hidden dimension of each layer is kept below D / 2.

[0149] Global pooling layer: Performs global average pooling (or global max pooling) on ​​the final hidden states of all nodes to obtain a fixed-length global feature vector. Where N is the total number of nodes, and L=2.

[0150] Dimensionality reduction output layer: The global feature vector is mapped to a low-dimensional dense vector through a fully connected layer (output dimension d), and L2 normalization is performed to obtain the final hole wall texture embedding vector.

[0151] III. Network Construction and Training Process The network is pre-built based on historical data during the offline phase, and the construction steps are as follows: Step 1: Dataset Preparation Collect multi-scale texture feature channel data and corresponding via wall quality labels (such as uniformity level and defect type) for several batches of PCB copper plating samples. For each sample, calculate the topological graph structure matrix according to the method in S6.1 to construct a graph structure sample.

[0152] Step 2: Network Initialization Set the input feature dimension D (e.g., 12), the number of graph convolutional layers L=2, the first hidden layer has a dimension of D, the second hidden layer has a dimension of D / 2 (e.g., 6), and the output embedding dimension d=32. The weight matrix is ​​initialized using Xavier, and the bias is initialized to 0.

[0153] Step 3: Training Optimization A supervised learning approach was adopted, using pore wall quality labels as the supervision signal. The loss function was either cross-entropy (for classification tasks) or mean squared error (for regression tasks). The optimizer was Adam, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. During training, the loss was monitored on the validation set, and an early stopping strategy was employed to prevent overfitting.

[0154] Step 4: Lightweight Optimization To accommodate the limited computing power of industrial controllers, the trained model is quantized and compressed: the weight parameters are converted from 32-bit floating-point numbers to 8-bit integers, and the scaling factors in layer normalization are merged. The optimized model has fewer than 100KB of parameters and a single forward inference time of less than 10ms (for a graph with 1000 nodes).

[0155] For example, the 3D reconstructed mesh of the hole wall contains 1024 nodes, and the multi-scale texture feature vector of each node has a dimension of 12. The topological graph structure matrix (a 1024×1024 sparse matrix with non-zero edge weights of approximately 2.5%) is calculated via S6.1. A lightweight graph convolutional network aggregates the features of its neighboring nodes for each node (higher edge weights contribute more). After two layers of graph convolution, the hidden state dimension of each node is reduced to 6. Global pooling yields a 1024×6 vector, which is then mapped to a 32-dimensional embedding vector through a fully connected layer. This embedding vector is input into the nonlinear regression mapping model in S7 to generate dynamic tuning factors.

[0156] Through the aforementioned lightweight graph convolutional network, this invention achieves efficient propagation and aggregation of multi-scale texture features in the hole wall spatial topology, significantly improving the physical consistency and computational real-time performance of copper plating uniformity evaluation.

[0157] S6.1: Obtain multi-scale texture feature channels and 3D reconstructed mesh data of hole walls output by nonlinear gating, calculate the correlation strength between adjacent mesh nodes based on the response fingerprint similarity measurement method, and generate a topological graph structure matrix with hole wall mesh nodes as vertices and response fingerprint similarity as edge weights.

[0158] Multi-scale texture feature channel data processed by nonlinear gated fusion and 3D reconstructed mesh coordinate data of hole walls are used as input conditions to perform precise correlation calculations on the physical state features of each mesh node. For each pair of adjacent mesh nodes, the response fingerprint similarity measurement method is invoked. By calculating the difference between the compactness mapping index and the flatness mapping index at each scale domain, an initial similarity value between nodes is obtained. This initial similarity value is introduced into the normalization module to eliminate the weight imbalance caused by the difference in feature dimensions at different scales, and obtain a similarity coefficient with consistent comparability. A threshold filtering strategy is used to remove node pairs with similarity coefficients lower than a preset physical correlation threshold, ensuring that the topology only retains adjacency relationships with physical texture associations. For the retained node pairs, edge weight assignment is performed according to the similarity coefficient to form an adjacency matrix with nodes as vertices and edge weights representing the strength of response fingerprint associations. The adjacency matrix is ​​loaded into a topology graph structure matrix through index mapping, providing physically consistent network structure constraints for subsequent graph convolution operations. Through the above processing method, the multi-scale texture feature channels output by nonlinear gating are transformed into a hole wall topology graph structure matrix with response fingerprint similarity as weight, thereby realizing the physical correlation encoding of multi-scale texture features between spatial grid nodes.

[0159] For example, in a PCB through-hole inspection scenario, the three texture features after nonlinear gated fusion—high-steepness domain, medium-slope domain, and hysteresis-significant domain—are 128-dimensional, 64-dimensional, and 32-dimensional vectors, respectively, and the 3D reconstructed mesh contains 1024 nodes. The similarity measurement method uses a weighted combination of Euclidean distance and cosine similarity, with a weight ratio of 0.6:0.4. The formula for calculating the difference in texture feature vectors between any two adjacent nodes i and j is: in, Let be the texture feature vector of node i at scale index s. Let be the texture feature vector of node j at scale index s. This represents the L2 norm of a vector. By filtering out low-relevance node pairs with a threshold of 0.75, the edge weights are concentrated in regions of high physical relevance. This calculation results in a 1024×1024 topological graph structure matrix, where non-zero edge weights account for only about 2.5% of the total elements, significantly improving the computational efficiency of subsequent graph convolutions and enhancing the effectiveness of physical feature propagation.

[0160] S6.2: Perform feature mapping initialization processing on each vertex in the topology graph structure matrix, load the multi-scale texture feature channels of the corresponding mesh unit as the initial node state vector, and generate an original node feature set containing information on nanoscale grain orientation, submicron-level pore distribution and micron-level coating undulation.

[0161] S6.3: Based on the original node feature set and the topology graph structure matrix, a lightweight graph convolution operator is used to perform a weighted aggregation operation on neighborhood information. The contribution of neighboring nodes to the central node is dynamically adjusted through edge weights to generate a hidden state representation of intermediate-layer nodes that incorporates local physical context information.

[0162] Based on the original node feature set and the topology graph structure matrix, a lightweight graph convolution operator kernel weight matrix is ​​initialized, and a neighbor node contribution adjustment coefficient is set according to the edge weight values. The neighbor node index list of the target node is extracted from the topology graph matrix, and the state vector of the corresponding neighbor node is obtained by calling the feature retrieval interface. This state vector is then multiplied by the edge weight values ​​using a multidimensional element-wise multiplication to generate a weighted neighborhood feature set. A summation and aggregation operation is performed on the weighted neighborhood feature set to obtain a neighborhood aggregation vector that centrally reflects the local physical context. This vector is then linearly combined with the target node's own feature vector according to a set ratio to obtain a preliminary fusion vector. A nonlinear activation function from the graph convolution operator is used to perform element-wise transformation on the preliminary fusion vector to enhance the expressive power of node features at different aggregation scales and suppress irrelevant physical noise features. Normalization is performed on the nonlinearly transformed fusion vector to ensure that the hidden state representations of all nodes have balanced scale features within the same feature space. Through the chain-like derivation process described above, the original node feature set from the previous step is transformed into an intermediate layer node hidden state representation that incorporates local physical context information, thereby achieving smooth propagation of cross-node texture information in lightweight graph convolutional networks.

[0163] For example, a local region containing 1000 grid nodes is selected as the topological graph substructure in the 3D reconstructed mesh of the hole wall. The edge weights are output by the response fingerprint similarity processing method, with values ​​ranging from 0.1 to 0.95. The node feature dimension is 12-dimensional multi-scale texture channel values. The lightweight graph convolution operator kernel weight matrix is ​​initialized to a shape of 12×8, using the ReLU activation function and batch normalization strategy. In the neighborhood information weighted aggregation, the target node index is 500, and its neighbor nodes are extracted to be 6, corresponding to edge weight values ​​of 0.85, 0.78, 0.43, 0.62, 0.55, and 0.91, respectively. The neighbor node feature vectors and edge weights are multiplied in multiple dimensions and then summed to obtain the aggregation vector [3.21, 4.56, 2.18, 1.95, 5.02, 0.88, 4.11, 1.72]. The aggregated vector and the target node's own features are linearly combined at a ratio of 0.7:0.3 and then input into the ReLU function. All output values ​​are greater than 0, demonstrating its significant enhancement effect in capturing local texture features in high-steep and medium-slope domains. After batch normalization, the hidden state values ​​of each node are distributed within the range of [-1.0, 1.0], verifying the feature scale balance. The final output of the intermediate layer node hidden state representations preserves high-frequency texture details in subsequent multi-level nonlinear transformations and significantly improves spatial physical consistency during cross-node texture propagation.

[0164] S6.4: Perform multi-level nonlinear transformation and residual connection processing on the hidden state representation of the intermediate layer nodes to eliminate the gradient diffusion effect in the deep network propagation process and retain high-frequency texture details, generating a deep graph embedding feature tensor with cross-scale spatial smoothness characteristics.

[0165] Multi-level nonlinear transformations are performed on the hidden state representations of intermediate-layer nodes that incorporate local physical context information to enhance the saliency of texture features at different spatial scales. Based on a pre-defined multi-layer perceptual sub-unit structure, the hidden states of intermediate-layer nodes are input into corresponding nonlinear activation functions to perform mapping transformations, forming primary-scale enhanced features. Multi-scale convolutional kernels are invoked and combined with node topological relationships to perform domain-specific convolution operations to extract structural patterns in different spatial frequency bands, aligning the convolutional output with the primary-scale enhanced features in the feature dimension and preserving the corresponding spatial location encoding. The convolutional output and the primary-scale enhanced features are then fused element-wise with weights, and the weight coefficients are automatically adjusted through a residual connection strategy to improve global smoothness without destroying the original high-frequency texture details. A gradient flow channel preservation method is adopted, and batch normalization constraints are introduced at each layer's residual connection points to prevent gradient vanishing during deep propagation and to enhance the feature's ability to be sensitive to both nanoscale and microscale textures simultaneously. The fused multi-layer feature sequence is subjected to feature resampling to ensure that the spatial reference frame of each scale domain remains consistent in the deep graph embedding feature tensor. Through this multi-level nonlinear transformation and residual connection processing, the hidden state of the intermediate layer node in the previous step is transformed into a deep graph embedding feature tensor with cross-scale spatial smoothness, so as to achieve the synchronous preservation of spatial physical consistency and multi-scale texture information.

[0166] S6.5: Perform global pooling and dimensionality reduction mapping operations based on the deep graph embedding feature tensor to extract low-dimensional dense vectors that characterize the texture distribution pattern of the entire hole wall region, and generate a hole wall texture embedding vector with spatial physical consistency as the final output result.

[0167] Based on the deep graph embedding feature tensor obtained through multi-level nonlinear transformation and residual connection processing, the global pooling operator is called to perform traversal accumulation operation on the embedding vectors of all grid nodes, so as to uniformly count the spatial physical context information between nodes in the global scope and eliminate the differences in local coordinate systems.

[0168] A standardized parameter matrix is ​​introduced into the global cumulative result. Based on the relationship between the number of nodes and the edge weights in the topology matrix, the mean and variance of the features are normalized and adjusted to eliminate the distribution offset of the feature components in different scale domains.

[0169] Principal component analysis transformation matrix is ​​used to decompose the normalized global feature set into feature covariance matrix, and the eigenvectors corresponding to the first few eigenvalues ​​are obtained. The vectors are then mapped to a low-dimensional dense space through projection matrix, retaining the feature components that contribute the most to the distribution law of pore wall texture.

[0170] For the projection transformation process, the principal components are selected based on the condition that the cumulative contribution rate exceeds a preset threshold, and the low-dimensional mapping results are calculated.

[0171] Low-dimensional dense vectors are normalized using the L2 norm to ensure their magnitude consistency across spatial scales and to improve the numerical stability of subsequent nonlinear mapping models.

[0172] By using global pooling and dimensionality reduction mapping, the deep graph embedding feature tensor from the previous step is transformed into a spatially and physically consistent hole wall texture embedding vector, thus achieving a complete representation of the multi-scale texture distribution pattern of the hole wall region in a single low-dimensional space.

[0173] Step S7: Based on the nonlinear mapping relationship between the hole wall texture embedding vector and the copper plating process parameters, calculate the dynamic optimization factor used to compensate for the delay in image feature and process control response, and generate process correction instructions for fluctuations in the copper plating process. Specifically, this includes: S7.1: Obtain the hole wall texture embedding vector with spatial physical consistency generated in the previous steps and the historical sequence of copper plating process parameters acquired in real time. Use the time alignment method to perform timestamp matching on the two to eliminate asynchronous errors in the data acquisition and transmission process, and generate a texture-process association dataset containing spatiotemporal synchronization information.

[0174] S7.2: Based on the texture-process association dataset, call the pre-trained nonlinear regression mapping model to perform projection transformation processing from high-dimensional feature space to process parameter space, so as to quantify the potential influence weight of the micro-texture state of the hole wall on the macro-copper plating quality, and generate an initial static compensation coefficient set that characterizes the current process deviation trend.

[0175] A texture-process association dataset containing spatiotemporal synchronization information is obtained, and a pre-trained nonlinear regression mapping model is used as the core module for high-dimensional input processing. The hole wall texture embedding vector is normalized and the covariance matrix is ​​estimated in the feature space to determine the stability of the input distribution.

[0176] Based on the normalized embedded vector and historical data of copper plating process parameters, a high-dimensional feature space and process parameter space mapping weight solution is performed. A nonlinear regression structure in the form of a kernel function is adopted, and a high-dimensional projection relationship matrix is ​​constructed according to the kernel type selected by the model.

[0177] By utilizing the optimization mechanism within the model, the sum of squared residuals between the predicted process parameter deviations and the actual parameter deviations is minimized through iterative gradient descent calculations, ensuring the convergence and interpretability of the mapping parameters.

[0178] During the projection transformation, correlation coefficient analysis is performed on the kernel function output to quantify the influence of a single embedded vector component on the copper plating process quality indicators to form an influence weight set, and low significance weights are removed to reduce computational redundancy.

[0179] The remaining weight set is grouped and coded according to the process quality index category to generate an initial static compensation coefficient set, which provides basic data for subsequent dynamic phase compensation with lag time constant.

[0180] Through the above processing method, the hole wall texture embedding vector of the previous step is transformed into an initial static compensation coefficient that characterizes the current process deviation trend, realizing a quantifiable mapping from image micro-features to macro-copper plating quality indicators.

[0181] S7.3: Introduce the inherent response lag time constant of the system into the initial static compensation coefficient set, and perform dynamic phase compensation operation using the phase lead correction method to predict and offset the control delay effect generated from the completion of image analysis to the effective action of the actuator, thereby generating the original value of the dynamic tuning factor after time domain correction.

[0182] The system's inherent response lag time constant is introduced as the input benchmark for the time-domain correction factor in the initial static compensation coefficient set. The value of the lag time constant is comprehensively determined based on the mechanical inertia of the actuator, the chemical reaction rate of the copper plating bath, and the transmission delay of the control signal, and then loaded into the tuning calculation module. A discrete time-domain phase response model is used to establish the correspondence between frequency components and control delay, and the initial static compensation coefficient set is expanded into amplitude and phase spectra in the frequency domain. For the phase spectrum components, the phase lead correction method is used to calculate the phase lead required to offset the delay. The calculated phase lead is superimposed on the original phase spectrum while keeping the amplitude spectrum unchanged, forming a phase-corrected frequency-domain compensation coefficient set. An inverse Fourier transform is performed on the phase-corrected frequency-domain compensation coefficient set to restore it to its time-domain form, thereby generating the original value of the time-domain corrected dynamic tuning factor. Through the above processing method, the initial static compensation coefficient set of the previous step is transformed into the original value of the dynamic tuning factor with delay offset capability, realizing the prediction and offsetting of the control delay effect during the period from the completion of image analysis to the effective action of the actuator.

[0183] S7.4: Based on the original value of the dynamic tuning factor and the current electrochemical stability constraints of the copper plating bath, an adaptive amplitude limiting filter strategy is adopted to perform safety boundary truncation processing to prevent severe fluctuations in process parameters caused by overcompensation, and to generate a safety dynamic tuning factor that meets the physical limits of the actuator.

[0184] Based on the original value of the dynamic tuning factor generated by phase lead correction and the current electrochemical stability constraint data of the copper plating bath, the electrochemical state assessment module is invoked to quantify the conductivity, ion concentration, temperature, and pH parameters of the bath, generating a set of process safety boundary coefficients that can serve as the initial reference for amplitude limiting filtering. A difference analysis is performed between the original value of the dynamic tuning factor and the set of safety boundary coefficients to extract tuning components exceeding the safety range. An adaptive amplitude limiting filter is used to attenuate these abnormal components within the boundary, and the filter amplitude range is adjusted in real time according to the bath stability through an internal weight adjustment mechanism. The safety tuning component output by the amplitude limiting filter is superimposed and synthesized with the original tuning factor to form a preliminary boundary-truncation set of tuning factor candidate values. A secondary truncation operation based on the physical limits of the actuator is performed on each factor in the candidate value set to calculate the final safety dynamic tuning factor that meets the safety requirements of the actuator. The safety dynamic tuning factor is marked as a closed-loop control executable state to achieve safety compensation control for parameter fluctuations in the copper plating process. By using adaptive amplitude limiting filtering and physical limit truncation processing, the original value of the dynamic tuning factor in the previous step is transformed into a safe dynamic tuning factor that meets the electrochemical stability of the bath solution and the physical safety boundary of the actuator, thereby achieving the technical effect of preventing overcompensation from causing drastic fluctuations in process parameters.

[0185] For example, in the copper plating process, the conductivity of the bath solution is 58.3 mS / cm, and the Cu²⁺ content is... + Under conditions of ion concentration of 0.85 mol / L, temperature of 45.2 ℃, and pH of 1.1, the initial value of the dynamic tuning factor obtained after phase lead correction was 2.75. By inputting conductivity, ion concentration, temperature, and pH into the electrochemical state assessment module, the process safety boundary coefficient was calculated to be 2.40. An adaptive limiting filter was used to limit the tuning factor amplitude range to ±(safety boundary coefficient × 0.05), i.e., ±0.12. The limiting output value was calculated using the following formula: Where V is the original dynamic tuning factor value and B is the process safety boundary coefficient, the calculated value after limiting is 2.40. The difference between the limited value and the original value is attenuated and corrected, resulting in an effective value of 2.42 in the candidate value set. Then, a second truncation is performed based on the physical limit of the actuator (set to 2.45), ultimately outputting a safety dynamic tuning factor of 2.42. After this processing, in actual execution, the copper plating current density adjustment remains within the predetermined safety range, the chemical flow rate compensation value changes smoothly, the uniformity of the copper plating layer on the hole wall is significantly improved, and the system operates stably.

[0186] S7.5: Map the safety dynamic tuning factor to a specific current density adjustment increment and liquid flow rate compensation pulse width, encapsulate it into a standardized control data frame using the industrial fieldbus protocol, generate an adaptive process correction command for fluctuations in the copper plating process, and output it to the lower-level control system.

[0187] Step S8: If the confidence level of the dynamic tuning factor is detected to be lower than the preset stability threshold, then the issuance of the process correction command is paused and the model self-update process is triggered; otherwise, the process correction command is sent to the copper plating control system to perform closed-loop feedback optimization. Specifically, this includes: S8.1: Obtain the dynamic tuning factor and preset stability threshold generated in the previous main step, perform statistical variance calculation and historical trend consistency test on the dynamic tuning factor, and generate a confidence quantification index that characterizes the reliability of the current control command.

[0188] S8.2: Based on the confidence quantification index and the preset stability threshold, a logical comparison is performed. If the confidence quantification index is lower than the preset stability threshold, an instruction pause signal is generated and the model is marked as in failure. Otherwise, an instruction execution permission signal is generated.

[0189] S8.3: In response to the instruction pause signal, call the online learning algorithm to perform abnormal sample reweighting on the historical training dataset of the hole wall micro-deformation response model, and generate a model self-updating parameter set containing the latest process fluctuation characteristics.

[0190] Obtain the historical training dataset of the hole wall micro-deformation response model triggered by the instruction pause signal, and filter out sample records with abnormal texture features caused by fluctuations in the copper plating process.

[0191] Multidimensional feature anomaly measurement was performed on the selected abnormal samples. The deviation of each sample in the feature space was quantified by the joint analysis method of Mahalanobis distance and local outlier factor, and an anomaly score matrix was generated.

[0192] A dynamic reweighting factor is set based on the anomaly score matrix, and the contribution coefficient of abnormal samples in model training is adjusted using a weighted update formula.

[0193] Normalization is performed on the reweighted factor matrix to ensure that the sum of all sample weights remains consistent with the original dataset and that the original weight ratios of normal samples are preserved.

[0194] The parameter update operation of the online learning algorithm is performed using the reweighted training dataset. The incremental gradient descent mechanism is called to adjust the key weight vector of the response model in chronological order, generating a model self-updating parameter set containing the latest copper plating process fluctuation characteristics.

[0195] By using the above-mentioned abnormal sample reweighting and online learning processing methods, the historical training data from the previous step is transformed into model update parameters that reflect the new operating conditions, thereby improving the adaptability of the response model under abnormal fluctuation conditions.

[0196] For example, in a copper plating production line, when the confidence level of the dynamic tuning factor drops to 0.35, a model self-update process is triggered, acquiring the micro-deformation response features of the hole walls and corresponding process parameters from the most recent 200 batches as a historical training dataset. For the 15 batches of anomalous samples identified through extreme value analysis, anomaly scores are calculated, ranging from 2.5 to 4.8. An anomaly penalty coefficient α = 0.2 is set, and a reweighting coefficient is calculated, resulting in anomalous sample weights ranging from 0.51 to 0.69. After normalization, the total weights of all samples are kept at 200, and incremental gradient descent online learning is performed with a learning rate of 0.005. After 50 iterations, the model self-updated parameter set is obtained. The updated model can stably identify texture feature changes in the subsequent 10 batches and adapt to the condition where the copper plating current density fluctuation range expands to 15 units, significantly improving defect identification accuracy under fluctuating conditions.

[0197] S8.4: The internal weight matrix of the micro-deformation response model of the pore wall is iteratively corrected using the self-updating parameter set of the model to generate an updated micro-deformation response model of the pore wall with the ability to adapt to new working conditions, so as to complete the system self-healing.

[0198] A matrix dimension consistency check is performed on the self-updating parameter set of the model to ensure that the updated weight values ​​in the new parameter set match the row and column structure of the weight matrix inside the micro-deformation response model of the pore wall.

[0199] Based on the validated model self-updating parameter set, the update increment of each weight element is calculated through a gradient iterative adjustment mechanism.

[0200] The calculated weight update increment is added to the existing weight matrix to form a candidate set of weight matrices after iterative correction.

[0201] Regularization constraints are applied to the candidate set of weight matrices, using the L2 norm to limit the excessive growth of weight values.

[0202] The regularized weight matrix is ​​subjected to performance verification. By loading it into the simulation environment and inputting a test dataset containing anomalous samples, the model's response accuracy and stability under new operating conditions are tested.

[0203] Replace the internal weight matrix of the original pore wall micro-deformation response model with the weight matrix that has passed performance verification to generate an updated pore wall micro-deformation response model instance with the ability to adapt to new working conditions.

[0204] Through the above gradient iteration, regularization constraints and performance verification processing, the model self-updating parameter set of the previous step is transformed into the final weight matrix that meets the adaptability requirements of the new working conditions, thereby realizing the self-healing of the hole wall micro-deformation response model.

[0205] For example, in the online inspection of a PCB copper plating production line, the historical training dataset contains 5000 samples of reflected light intensity sequences from hole walls, with an abnormal sample ratio of 20%. The learning rate η is set to 0.005, and the regularization coefficient λ is 0.01. During matrix dimension verification, the model's self-updating parameter set and weight matrix are both 256×256, meeting the update conditions. During iterative correction, gradient calculation is performed on the loss function L = mean squared error, with the gradient range of each weight element being [-0.02, 0.015], corresponding to an update increment range of [-0.0001, 0.000075]. After matrix addition, the maximum element value of the candidate weight matrix is ​​0.45, and the minimum element value is -0.37. After regularization, the weight values ​​converge significantly, with the maximum value decreasing to 0.42 and the minimum value increasing to -0.34. During the performance verification phase, a new weight matrix was loaded into the response model. Using abnormal samples as test inputs, the fitting error of the model's response curve under the new operating conditions was reduced to half that of the original model, and the stability index was improved to above the set fitness threshold. The final updated hole wall micro-deformation response model showed significantly improved automatic adjustment capabilities in subsequent copper plating process control, effectively offsetting uniformity deviations caused by various process fluctuation factors.

[0206] S8.5: In response to the instruction execution permission signal, the dynamic tuning factor is mapped to a specific current density adjustment amount and a chemical flow rate compensation value, an adaptive process correction instruction for fluctuations in the copper plating process is generated and sent to the copper plating control system for closed-loop feedback optimization.

[0207] In this invention, the copper plating control system refers to an automated control platform for performing the chemical copper plating process, which includes at least a bath concentration detector, a temperature sensor, an automatic addition device, and a programmable logic controller.

[0208] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0209] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0210] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A computer vision-based real-time evaluation method for the uniformity of copper plating on PCB via walls, specifically including: S1: Collect data of the inner wall of the PCB through-hole to generate a sequence of reflected light intensity from the hole wall; S2: Based on the reflected light intensity sequence of the hole wall and the known Young's modulus, Poisson's ratio, copper thickness parameters of copper material and geometric curvature radius of the hole wall, a micro-deformation response model of the hole wall is constructed that maps the micron-level elastic displacement distribution of the hole wall surface under light pressure to a local reflectivity modulation function. S3: Using the three-dimensional reconstructed mesh of the aperture wall as a reference frame, input the reflected light intensity sequence of the aperture wall into the aperture wall micro-deformation response model to generate a response fingerprint; S4: Based on the steepness, monotonicity and repeatability of the response fingerprint, three physical sensitive scale domains are divided, corresponding to the high steepness domain, medium slope domain and significant hysteresis domain, respectively, which correspond to the differences in nanoscale grain orientation, submicron-level pore distribution and micron-level coating undulations. S5: Perform trend consistency determination and signal-to-noise ratio threshold comparison on the response fingerprints of each physical sensitive scale domain acquired in three consecutive frames, generate a gating enable signal that characterizes whether each scale domain meets the activation condition, and generate a multi-scale texture feature channel based on the gating enable signal. S6: Feed the multi-scale texture feature channels into the graph convolutional network, use the hole wall mesh nodes as the graph structure vertices and the response fingerprint similarity as the edge weights, perform cross-node texture propagation aggregation, and generate hole wall texture embedding vectors. S7: Based on the nonlinear mapping relationship between the hole wall texture embedding vector and the copper plating process parameters, calculate the dynamic optimization factor used to compensate for the delay in image features and process control response, and generate process correction instructions for fluctuations in the copper plating process.

2. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, Following S7, the following also includes: S8: If the confidence level of the dynamic tuning factor is detected to be lower than the preset stability threshold, the issuance of process correction instructions will be paused and the model self-update process will be triggered; otherwise, the process correction instructions will be sent to the copper plating control system to perform closed-loop feedback optimization.

3. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, The process of collecting data from the inner wall of PCB vias to generate a sequence of reflected light intensity from the via walls involves using a ring-shaped LED array light source with controllable polarization angle and microsecond-level pulse modulation capability to excite the inner wall of the PCB vias in a multi-angle, multi-polarization state sequence. Simultaneously, a high-resolution line scan camera is triggered to collect and record the sequence of reflected light intensity from the via walls, which includes the incident angle of the light source, polarization direction, and pulse timing markers.

4. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, The step of inputting the aperture wall reflected light intensity sequence into the aperture wall micro-deformation response model to generate a response fingerprint specifically involves inputting the aperture wall reflected light intensity sequence into the aperture wall micro-deformation response model for inverse decoupling, fitting the slope, saturation threshold, and hysteresis interval of the response curve of each grid unit under different illumination conditions, and generating a response fingerprint that characterizes the comprehensive physical state of local copper layer compactness and micro-flatness.

5. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, The method of generating multi-scale texture feature channels based on gating enable signals specifically involves using gating enable signals to filter response fingerprints, and independently performing Gabor direction response, local contrast gradient histogram, and Fourier frequency domain energy spectrum extraction only on scale domains that meet the activation conditions, thereby generating multi-scale texture feature channels with nonlinear gating output.

6. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, The three-dimensional reconstruction mesh of the aperture wall is generated by performing multi-view geometric reconstruction and meshing processing based on the aperture wall reflected light intensity sequence and the internal and external parameters of the line scan camera.

7. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, The graph convolutional network includes a feature initialization layer, a graph convolutional layer, a global pooling layer, and a dimensionality reduction output layer.

8. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, S3 specifically includes: The coordinate data of the three-dimensional reconstructed mesh of the aperture wall and the reflected light intensity sequence of the aperture wall containing the incident angle and polarization direction of the light source are obtained. Based on the spatial projection transformation method, the multi-dimensional light intensity sequence is mapped to the vertices of each mesh unit of the three-dimensional reconstructed mesh of the aperture wall, and a set of mesh unit light intensity response vectors carrying multi-angle excitation information is generated. The micro-deformation response model of the hole wall is called, and the light intensity response vector set of the grid cell is used as the input variable. The gray-scale response trajectory of each grid cell under different light pressure is reconstructed to generate a set of grid cell response curve parameters including the initial slope factor, the maximum saturation threshold and the hysteresis width. Based on the initial slope factor in the parameter set of the grid cell response curve and the Young's modulus constraint of copper material, the micron-scale surface displacement induced by illumination is calculated, and the displacement is converted into a local reflectivity modulation coefficient to generate a compactness mapping index characterizing the density of copper layer grain orientation. By combining the loop hysteresis width and maximum saturation threshold in the parameter set of the grid cell response curve, the multiple scattering effect of submicron pores on the optical path is analyzed, the smoothness deviation value reflecting the degree of micro-undulation of the coating surface is calculated, and a smoothness mapping index characterizing the micro-geometric state of the copper layer surface is generated. By integrating the compactness mapping index and the flatness mapping index, a multi-dimensional feature encoding strategy is adopted to encapsulate the two physical state indices into structured data objects, generating a response fingerprint that uniquely identifies the physical characteristics of each grid cell, which serves as the direct input basis for subsequent division of the physical sensitive scale domain.

9. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, S4 specifically includes: The gray-scale response difference sequence contained in the response fingerprint is processed by first-order differential operation to obtain the set of response curve slopes that characterize the rate of change of local reflectivity with illumination angle, and the set of response curve slopes is used as the basic data input for quantifying the micro-deformation sensitivity of the copper layer surface. An extreme point detection and linearity fitting method is performed based on the slope set of the response curve to generate a steepness index that characterizes the drastic change of the response fingerprint within a specific illumination range and a monotonicity index that characterizes the consistency of the change trend. The steepness index and the monotonicity index are used as key criteria to distinguish the orientation differences of nanoscale grains. The temporal variance of the response fingerprint of the same grid cell is calculated by using the sequence of reflected light intensity from the hole wall acquired in multiple consecutive frames, so as to extract the repeatability stability index that characterizes the fluctuation amplitude of the copper layer compactness under different excitation conditions. The repeatability stability index and the steepness index are used together as the characteristic basis for identifying the distribution of submicron pores. Based on the pulse energy threshold required for the grayscale response in the response fingerprint to reach saturation and the signal fallback lag interval during the unloading process, the hysteresis significance parameter, which characterizes the optical hysteresis effect caused by macroscopic fluctuations in the coating, is calculated, and the hysteresis significance parameter is used as the core quantitative factor for defining micron-level coating fluctuations. By combining steepness, monotonicity, repeatability, and hysteresis significance parameters, multidimensional threshold clustering segmentation is performed to generate three physically sensitive scale domain boundaries that divide the hole wall texture space into a high-steepness domain, a medium-slope domain, and a hysteresis significance domain. These three physically sensitive scale domains are then used as the activation basis for the subsequent generation of gating enable signals.

10. The real-time evaluation method for PCB hole wall copper uniformity based on computer vision according to claim 1, characterized in that, S5 specifically includes: A time-series sliding window traversal process is performed on the response fingerprints of the high-steepness domain, the medium-slope domain, and the hysteresis-significant domain acquired in three consecutive frames to obtain the response fingerprint sequence data of each physical sensitive scale domain in the time dimension. Based on the response fingerprint sequence data, a trend consistency determination method and a signal-to-noise ratio threshold comparison operation are performed to generate a binary gating enable signal characterizing whether each physical sensitive scale domain meets the activation condition. The binarized gating enable signal is used to perform logical AND operation filtering on the response fingerprints of high steepness domain, medium slope domain, and hysteresis significant domain to output a subset of response fingerprints to be processed that contains only valid physical states. Gabor directional response filtering, local contrast gradient histogram statistics, and Fourier frequency domain energy spectrum decomposition are performed in parallel for the subset of response fingerprints to be processed, in order to generate original multi-domain texture feature vectors corresponding to nanoscale grain orientation differences, submicron-level pore distribution, and micron-level coating undulations. The original multi-domain texture feature vector and the binarized gated enable signal are subjected to element-wise multiplication nonlinear gating fusion processing to generate a multi-scale texture feature channel with spatial physical consistency and noise interference removal.

Images