A circuit board micro-fine line defect visual detection method based on edge computing

Abstract

This invention discloses a visual inspection method for micro-line defects on circuit boards based on edge computing, belonging to the field of circuit board visual inspection technology. The specific steps of the method are as follows: the edge master controller drives the composite polarization mode group to acquire multiple types of polarization images simultaneously; global and hierarchical Stokes vectors are solved, and a three-dimensional stress field is constructed by combining four-phase structured light; multi-dimensional polarization features and stress gradient field are fused and input into the neural network to detect microcracks; finally, fatigue damage is calculated and the production line closed-loop iterative optimization of acquisition and model parameters is driven. This invention adopts a composite polarization optics and hierarchical Stokes solution architecture to penetrate surface obstruction and simultaneously capture optical signals of the copper foil surface and interface, and eliminates transmission delay by combining edge computing localized processing; multi-source polarization feature fusion and stress weighted neural network improve the accuracy of microcrack identification, and rely on the production line closed-loop continuous adaptive iteration to balance detection efficiency and long-term stability.

Description

Technical Field

[0001] This invention relates to the field of circuit board visual inspection technology, specifically to a method for visual inspection of micro-line defects in circuit boards based on edge computing. Background Technology

[0002] With the continuous upgrading of high-density circuit board manufacturing processes and the increasing refinement of circuit wiring dimensions, microcracks in the buried layers of fine circuits and hidden damage at interfaces have become key hidden dangers restricting the yield of finished circuit boards. The industry's requirements for the depth, accuracy, and online real-time performance of defect detection are constantly increasing. At present, machine vision has become the mainstream method for screening circuit board defects. Conventional visual inspection relies on visible light imaging to identify surface appearance defects. Polarization imaging and structured light detection technologies are gradually being introduced into the field of electronic component quality inspection. Edge computing hardware architecture has also begun to be deployed in batches on automated production lines. The junction between the copper foil and the substrate of the circuit board is easily affected by process stress and thermal deformation, which can easily cause hidden damage. Such defects are hidden below the surface layer and cannot be directly determined by conventional two-dimensional images. It is necessary to combine optical polarization characteristics and material stress change laws to carry out in-depth detection. Related composite detection technologies have become a key research and development direction in the field of intelligent manufacturing of high-end circuit boards.

[0003] Traditional circuit board defect detection mostly uses a single optical path imaging scheme, which can only collect two-dimensional image information of the workpiece surface. It can only identify exposed surface cracks and has difficulty distinguishing hidden damage at the interface between the copper foil surface and the substrate. The missed detection rate of deeply buried microcracks remains high. Conventional detection schemes lack the ability to disassemble layered optical signals and cannot rely on optical parameters to invert the internal stress distribution of the circuit. They cannot predict potential fatigue damage from the perspective of material stress. Most detection systems upload raw image data to cloud servers for processing. Large-volume image transmission is prone to data lag and transmission delay, which cannot adapt to the online detection requirements of high-speed production lines. At the same time, traditional detection models do not have the ability to self-optimize in a closed loop on the production line. The detection parameters are fixed and cannot be adjusted automatically with changes in the production environment when facing mixed production scenarios with multiple substrates and copper foils of different thicknesses. The detection accuracy is poor and has poor versatility. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a visual inspection method for micro-circuit defects on circuit boards based on edge computing. This method relies on a composite polarization optical module to simultaneously acquire two types of polarization images. A Stokes layered decoupling algorithm is used to separate the optical information at the interface between the copper foil surface and the substrate. A three-dimensional stress gradient field is constructed by combining structured light data. Through a neural network that fuses multi-source polarization features and adaptively adjusts stress weights, the size and embedding information of microcracks are accurately obtained. A damage algorithm is then used to calculate circuit fatigue damage. The detection data is stored locally and connected to the production line system. Based on batch inspection results, optical parameters, network weights, and acquisition hardware configuration are iteratively optimized in a closed loop. The entire solution achieves integrated detection of surface and deeply embedded defects, improves the accuracy of microcrack detection, and meets the requirements of online automated quality inspection in PCB production lines.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a visual inspection method for micro-line defects on circuit boards based on edge computing, the specific steps of which are as follows: S1. Raw Image Acquisition: The edge master controller drives the coaxial integrated composite polarization optical module to acquire reference polarization reflection images and phase modulation polarization images in a time-division manner and store them in the edge storage. S2. Global Stokes parameter solution: The global composite Stokes vector is calculated from the reference polarization reflection image in step S1, and the Stokes vector at the interface between the copper foil surface and the copper foil substrate is obtained by using the Stokes layer decoupling algorithm. The result is then bound to the phase modulation polarization image in step S1 according to the workpiece number. S3. Construction of three-dimensional stress field: Based on the layered Stokes vector in step S2 and the four-phase structured light image in step S1, the optical response of the copper foil surface and interface is separated, and the three-dimensional stress gradient field of the line is generated by the depth stress mapping algorithm and transferred to step S4. S4. Microcrack Feature Detection: Extract the multidimensional polarization features of the original polarization image from step S1 and fuse them. Combine the stress gradient field from step S3 with the input polarization feature enhancement neural network to output the location, width, and depth of the microcracks, and store them locally. S5. Damage Calculation and Closed-Loop Optimization: Using the stress gradient field data from step S3 and the crack parameters from step S4, fatigue damage is calculated using the crack-coupled damage algorithm and output to the production line control system. Based on batch statistical iteration, the calibration database, neural network, and optical hardware configuration are adjusted, and the edge controller is updated to optimize the acquisition and control in the next round of step S1.

[0006] Furthermore, in S1, the original image acquisition includes two sequentially executed image acquisition stages. The first image acquisition stage sequentially completes the acquisition of reference polarization reflection images at four fixed polarization angles: 0°, 45°, 90°, and 135°. The total time from the start of acquisition to the completion of storage of the images at the four polarization angles does not exceed 20ms. The second image acquisition stage completes the acquisition of phase-modulated polarization images corresponding to four fixed spatial phases. All acquired image files are stored in the edge storage unit in an uncompressed bitmap format.

[0007] Furthermore, in S1, during the original image acquisition, in the second image acquisition stage, phase-modulated structured light is projected onto the micro-circuit area of ​​the circuit board under test at a fixed projection angle of 30°. After each phase-modulated polarization image is acquired, it is bound with four types of identification information: corresponding workpiece number, acquisition time, polarization type, and structured light phase. The identification information is bound to the corresponding image file and stored in the edge storage unit.

[0008] Furthermore, in S2, the global Stokes parameter solution process, the global composite Stokes vector solution process only calls the reference polarization reflection image data collected in S1 corresponding to the same circuit board under test, and does not call the phase modulation polarization image data. The Stokes vector of the copper foil surface and the Stokes vector of the copper foil substrate interface are stored in two independent partitions of the edge storage unit, respectively. The two types of Stokes vector data and the phase modulation polarization image collected in S1 are associated with each other through a unified workpiece number.

[0009] Furthermore, in S3, during the construction of the three-dimensional stress field, based on the preset numerical standards that the depolarization rate of the reflected light from the copper foil surface is less than 15% and the depolarization rate of the reflected light from the copper foil substrate interface is greater than 40%, the optical response components of the copper foil surface and the optical response components of the copper foil substrate interface are distinguished and separated. The calibration database built into the edge storage unit stores the optical loss parameters of three types of substrates: FR4 substrate, aluminum substrate, and PI substrate. At the same time, it stores the optical loss parameters corresponding to three copper foil thicknesses of 18μm, 35μm, and 70μm.

[0010] Furthermore, in S3, during the construction of the three-dimensional stress field, the calculated stress data of the copper foil surface and the stress data of the copper foil substrate interface are matched and stitched together at the pixel level according to the spatial coordinate system corresponding to the thickness of the copper foil on the circuit board under test. The stitched and integrated three-dimensional stress gradient field dataset is equipped with a unique data flow identifier and is transmitted and stored in the independent cache partition corresponding to S4 in the edge storage unit.

[0011] Furthermore, in S4, the microcrack feature detection process reads all the original polarization images stored in S1 one by one, extracts the gray-scale change data of the line edge, the polarization reflection intensity change data, and the structured light stripe distortion data in the image, and performs unified pixel dimension alignment processing on the feature maps of different imaging sources and different pixel resolutions. After the alignment is completed, the multi-class feature data is subjected to pixel-level fusion and recombination processing to generate a composite polarization feature dataset of unified specifications.

[0012] Furthermore, in S4, during the microcrack feature detection, the polarization feature enhancement neural network performs inference operations based on the stress distribution coordinate data of the three-dimensional stress gradient field. According to the coordinate region where the stress value in the three-dimensional stress gradient field exceeds the preset threshold, the image feature recognition weight corresponding to the region is increased by 20% to 50%. The microcrack location coordinates, crack width values, and crack burial depth values ​​output by the neural network are uniformly organized into a standardized structured data table, and the structured data table is completely stored in the edge local storage unit.

[0013] Furthermore, in S5, during damage calculation and closed-loop optimization, standardized detection data is generated after fatigue damage calculation is completed using the crack-coupled damage algorithm. One path of the standardized detection data is permanently stored in the edge local non-volatile storage medium, and the other path is transmitted to the production line control system in real time through the industrial fieldbus communication protocol to complete data interaction and storage.

[0014] Furthermore, in S5, damage calculation and closed-loop optimization, the iteratively adjusted configuration parameters include three sets of parameters. The first set of parameters consists of optical attenuation parameters corresponding to the thickness of various substrates and copper foils in the calibration database. The second set of parameters consists of feature weight parameters corresponding to all network layers of the polarization feature enhancement neural network. The third set of parameters consists of hardware timing configuration parameters for the image acquisition stage in S1. All iteratively updated parameter sets are completely written into the edge master control program storage area. When the circuit board under test performs step S1, the edge master control directly calls the updated parameter sets to complete the image acquisition configuration.

[0015] Compared with existing technologies, this edge computing-based visual inspection method for micro-line defects on circuit boards has the following advantages: I. This invention utilizes a detection architecture combining composite polarization optics with layered Stokes algorithm to overcome the limitations of traditional single visible light imaging, which can only observe the surface morphology of circuits. By relying on a dual-layer optical signal splitting mechanism at the surface and substrate interface, it simultaneously captures the optical response information of the copper foil surface and the deeply buried interface of the circuit board. This allows it to penetrate surface material obstructions and capture internal latent defect signals. Edge computing enables localized data processing throughout the entire process of image acquisition, parameter calculation, and stress modeling, eliminating data latency and information loss issues associated with cloud transmission. Various imaging data and computation results are stored and scheduled locally, shortening the data interaction time of the entire detection chain. Layered optical data is used to reconstruct the deep stress field, characterizing potential circuit damage from the perspective of the material's stress origin. This goes beyond simply identifying crack appearance features and can locate stress-rich areas that have not yet shown visible damage, expanding the detection range of micro-defects on circuit boards.

[0016] II. This invention integrates multi-source polarization feature fusion with a stress-weighted neural network recognition mode, combining multi-dimensional image information such as grayscale changes, polarization reflection, and structured light distortion, and completes pixel alignment fusion. This enriches the data dimensions for defect recognition, improves the accuracy of identifying micro-cracks under the interference of complex circuit textures, dynamically adjusts feature recognition weights based on stress distribution, enhances the ability to identify defects in high-stress areas, improves the problem of missing detection of fine buried deep cracks, and builds a closed-loop optimization link for the production line based on measured damage data. Based on the batch product inspection results, the optical parameters of the substrate, network weights, and optical acquisition hardware configuration are updated synchronously, continuously adapting to the inspection conditions of different plates and copper foils of different specifications. The entire inspection system can automatically iterate and adapt to fluctuations in production conditions, gradually converge the inspection error during long-term operation, and balance the inspection efficiency of online inspection with long-term inspection stability.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart of a visual inspection method for micro-line defects on circuit boards based on edge computing; Figure 2This is a schematic diagram illustrating the data transmission between each step of a visual inspection method for micro-circuit defects on circuit boards based on edge computing. Figure 3 This is a schematic diagram of data transmission within the three-dimensional stress field construction process. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Reference Figure 1 One embodiment of the present invention proposes a visual inspection method for micro-line defects on circuit boards based on edge computing. It adopts a coaxial integrated composite polarization optical time-division imaging architecture and a data processing mechanism that integrates the Stokes layered decoupling algorithm and depth stress mapping. It can realize the layered optical parameter calculation of the copper foil surface and copper foil substrate interface by relying on edge local storage and edge master control hardware without relying on real-time interaction of cloud computing power. It can also complete the multi-dimensional quantitative detection of micro-crack location, width and burial depth of micro-line. At the same time, it can realize the closed-loop iterative optimization of the entire link of acquisition hardware, algorithm model and calibration parameters by relying on batch detection data.

[0022] The method described in this embodiment specifically includes: The edge master control of the coaxial integrated composite polarization optical module divides the original image acquisition into two independent time stages, acquiring the reference polarization reflection image and the phase modulation polarization image respectively. All image files are stored in the edge storage unit according to the specified storage format. The global composite Stokes vector is solved using a reference polarization reflection image alone. The Stokes vector on the surface of the copper foil and the Stokes vector at the interface of the copper foil substrate are obtained by the Stokes layer decoupling algorithm. The two types of Stokes data and the phase-modulated polarization image are established by relying on a unified workpiece number. Based on a fixed depolarization rate standard, the optical response components corresponding to the copper foil surface and substrate interface are separated. Combined with the optical loss parameters of various substrates and copper foil thicknesses that are embedded in the edge storage, a three-dimensional stress gradient field of the circuit is constructed through a depth stress mapping algorithm. The generated stress dataset is directionally transmitted and cached to an independent partition of the edge storage specifically for defect detection. All polarization original images in the edge storage are read sequentially, and the inherent feature information of the images is extracted from multiple dimensions. Pixel alignment processing is performed on the feature maps with size and resolution differences caused by different imaging conditions. After pixel-level fusion and recombination, a composite polarization feature dataset of uniform specifications is generated. The composite polarization feature dataset and the three-dimensional stress gradient field data are fed into the polarization feature enhancement neural network for inference operations. The feature recognition weights of local areas are adjusted in a targeted manner based on the stress field distribution. The network outputs standardized microcrack location, crack width, and crack burial depth data, which are then stored in a structured manner locally at the edge. Combining the stress parameters related to the three-dimensional stress gradient field and the crack geometry parameters output by the neural network, the fatigue damage quantification calculation of the circuit is completed using the crack coupling damage algorithm. The generated standardized detection data is stored and transmitted in two separate paths: one path is permanently stored in the edge non-volatile storage medium, and the other path is transmitted to the production line control system via the industrial bus. Based on the statistical data summarized from the inspection of the entire batch of circuit boards, the calibration database parameters, neural network weight parameters, and image acquisition hardware timing parameters are adjusted iteratively in three categories. All updated parameters are written to the edge master control storage space, and the updated configuration is directly called when the image acquisition step is started in subsequent workpiece inspections, such as... Figure 2 As shown.

[0023] Specifically, the technical principle of this invention lies in building a fully autonomous circuit board defect detection link at the edge. This link first relies on a coaxial composite polarization optical module to complete the time-division acquisition and storage of two types of polarization images. Then, it uses the Stokes layered decoupling algorithm to separate and peel off the optical parameters of the upper and lower layers of the copper foil. Combined with preset substrate and copper foil calibration parameters, the polarization parameters are converted into three-dimensional stress data with continuous distribution in the depth direction. Then, it integrates multi-source image features and stress distribution data into a neural network to complete the fine detection of microcracks. Finally, it uses the damage data formed by batch detection to back-optimize the configuration parameters of the entire process of early imaging, parameter calculation, and model recognition, forming a closed-loop workflow from image acquisition, layered parameter solution, stress field construction, defect recognition to system parameter self-optimization.

[0024] Optionally, the edge master control coaxial integrated composite polarization optical module completes the original image acquisition in two independent time-series stages: the first image acquisition stage switches four polarization angles (0°, 45°, 90°, and 135°) in a fixed order, and completes the acquisition of reference polarization reflection images angle by angle. The overall time from the start of acquisition of the first angle image to the writing of all four images into the edge storage is strictly controlled within 20ms. The second image acquisition stage fixes the structured light projection tilt angle at 30°, adjusts the four different spatial phases of the modulated structured light to be projected onto the micro-circuit area to be tested on the circuit board in sequence, and completes the phase modulation polarization image acquisition operation phase by phase. For each phase modulation polarization image acquired in the second stage, four unique identification information are simultaneously bound: workpiece number, acquisition time, polarization type, and structured light phase. All images generated in all stages are stored in a unified uncompressed bitmap format.

[0025] Specifically, the edge controller first issues timing control commands, and the device hardware prioritizes entering the first image acquisition stage. The polarization devices inside the optical module automatically switch according to a preset angle sequence. After each polarization angle switch, the camera synchronously completes exposure and imaging. After imaging, the image data is transmitted and written to the edge storage unit in real time. The device hardware has undergone factory timing calibration, which can strictly control the total time limit for four polarization imaging and data storage, avoiding the impact of single-batch acquisition timeouts on production line efficiency. After all the reference images in the first stage are stored, the edge controller automatically issues a light path switching command, and the device enters the second image acquisition stage. The mechanical structure of the projection light path is locked and remains stationary, always maintaining a fixed projection tilt angle of 30°. Only the internal phase parameters of the structured light are changed to achieve the switching of four spatial phases. After each phase switch is completed, polarization image acquisition is initiated. After a single phase-modulated image is generated, the edge controller captures four types of information in real time: the workpiece code recorded on the production line, the precise acquisition time generated by the system clock, the polarization type of the current imaging, and the structured light phase code corresponding to this imaging. These four identification information items are bound and indexed with the corresponding image files. The uncompressed bitmap storage method can completely retain the original pixel information of the image, prevent the loss of pixel details caused by image compression, and prevent data deviations in subsequent Stokes parameter solutions.

[0026] For example, in a mass industrial online inspection scenario for PCB circuit boards, after receiving the acquisition command from the edge controller, the device prioritizes the acquisition of the reference polarization image. The entire four-angle image capture and storage process is successfully completed within the specified time. The device then automatically switches to the structured light projection imaging mode, maintaining a stable projection tilt angle, and sequentially alternating between the four spatial phases to capture the corresponding images. At the moment each phase image is generated, the system automatically captures the current circuit board's unique workpiece code, the acquisition timestamp accurate to milliseconds, the imaging mode classification mark, and the phase number. The above information and image files are bound and archived one by one, and all acquired images are stored in the corresponding folder of the edge storage in uncompressed bitmap format.

[0027] Optionally, the step of using a reference polarization reflection image to solve the global composite Stokes vector separately, and obtaining the copper foil surface Stokes vector and the copper foil substrate interface Stokes vector by using a Stokes layer decoupling algorithm, includes: the entire calculation process of the global composite Stokes vector only retrieves the reference polarization reflection image data acquired in the first stage of the current workpiece under test, and does not call the phase modulation polarization image related data generated in the second stage; based on the mathematical calculation formula corresponding to the Stokes layer decoupling algorithm, the copper foil surface Stokes vector and the copper foil substrate interface Stokes vector are calculated from the global composite Stokes vector; the two types of Stokes vector data obtained by decomposition are stored in two independent partitions within the edge storage that do not interfere with each other, and an association index between the two types of vector data and the phase modulation polarization image is established based on the unified number of the workpiece under test.

[0028] Specifically, the mathematical expression for the Stokes hierarchical decoupling algorithm is: ; in: Represents the global composite Stokes vector. Represents the Stokes vector of the copper foil surface. Stokes vector representing the copper foil-substrate interface. Represents the surface dynamic light attenuation correction coefficient. Represents the interface transmission attenuation correction factor. This represents the thickness of the copper foil on the circuit board being tested. Represents the substrate material coefficient. Representative Reliance Numerical and Numericalized compensation items for process deviations in sheet metal; , , , , This is the calibration constant for surface loss. , This is the calibration constant for interface transmission loss. , , The constants for fitting process deviations are all pre-calibrated offline using standard samples of copper foil of different thicknesses and different substrates and stored in the calibration database.

[0029] In actual operation, the system reads the raw pixel grayscale data of four reference polarization images of the same workpiece, first calculates the overall value of the global composite Stokes vector, and then, based on the actual copper foil thickness and substrate material of the circuit board under test, matches the corresponding attenuation correction coefficient and process deviation compensation parameters from the edge calibration database, and substitutes them into the calculation formula to complete the layered splitting of the global vector. After the splitting is completed, the Stokes vector of the copper foil surface layer is uniformly written to the first dedicated partition of the edge storage, and the Stokes vector of the copper foil substrate interface is stored in the second independent partition. When it is necessary to retrieve the phase modulation image for subsequent calculations, only the workpiece number needs to be entered to quickly index and match the corresponding layered Stokes data.

[0030] For example, when performing Stokes parameter calculations on conventional mass-produced circuit boards, the algorithm only calls the reference polarization image data acquired in the early stage to perform calculations. Based on the actual material of the circuit board and the thickness of the copper foil processing, the corresponding correction coefficients and process compensation parameters are retrieved from the local calibration library, and then used to split the data into two layers of Stokes vector data using the layered calculation formula. The two types of data are stored separately in different storage partitions of the edge storage. The data is linked with the phase modulation image by relying on the workpiece number of the circuit board. When subsequent processes retrieve data, the data pairing can be quickly completed by using the number.

[0031] Optionally, the step of separating the optical response components corresponding to the copper foil surface and substrate interface based on a fixed depolarization rate standard, and constructing a three-dimensional stress gradient field of the circuit using a depth stress mapping algorithm, combined with the optical loss parameters of multi-specification substrates and copper foil thicknesses stored in the edge storage, includes: distinguishing the two layers of optical data according to a preset judgment rule, where the depolarization rate of the reflected light corresponding to the copper foil surface is less than 15%, and the depolarization rate of the reflected light corresponding to the copper foil substrate interface is greater than 40%, and splitting the two types of optical response components based on this numerical boundary; reading the parameter content that has been pre-fixed and saved in the edge storage calibration database, with the parameters covering FR The optical loss parameters of three commonly used substrates—substrate, aluminum substrate, and PI substrate—and the optical loss parameters of three mainstream copper foil thicknesses—18μm, 35μm, and 70μm—are calculated pixel by pixel using the depth stress mapping formula. The calculated surface stress data and interface stress data are then pixel-level spliced ​​and integrated based on the spatial coordinate system matching the copper foil thickness of the circuit board. The integrated three-dimensional stress gradient field dataset is configured with a unique transfer identifier and directionally transmitted and saved to an independent cache partition in edge storage for use in the defect detection step.

[0032] Specifically, the mathematical expression for the depth stress mapping algorithm is: ; in: Represents the depth of the copper foil Stress value at the location, For the stress transformation coefficient of deeply correlated materials, Surface stress weighting coefficient Let be the mapping function from surface Stokes parameters to surface stress. This is the stress gradient distribution function calculated based on the constitutive model of copper foil material and the surface stress boundary conditions; ; ; ; ; ; , For depth-thickness correlation calibration coefficients, This represents the optical loss coefficient of the copper foil surface layer. The optical loss coefficient at the substrate interface. , The constants for fitting the surface polarization response are: , The constants for fitting the interface polarization response are defined by various coefficients that have been pre-calibrated offline using standard samples and then stored in the calibration database.

[0033] After separating the two layers of optical response components, the system identifies the substrate type and actual copper foil thickness of the circuit board under test. It retrieves matching optical loss parameters from the solidified calibration parameters to determine the values ​​of various coefficients in the formula. Then, it substitutes each pixel along the depth direction of the copper foil into the formula to calculate the stress data at the corresponding depth. After all single-point stress calculations are completed, the system aligns the pixel coordinates of the discrete stress data from the surface and interface layers with the spatial coordinate standard corresponding to the current board specifications, completing the full-frame stress data stitching. The complete 3D stress dataset generates a unique data flow code, which is automatically stored in a pre-defined independent cache area in the edge storage according to the preset data flow direction, reserving data resources for subsequent neural network defect detection.

[0034] For example, in the online inspection process, the system automatically calculates the depolarization rate index of the reflected light at two points on the circuit board, and successfully distinguishes the optical response data of the copper foil surface and the substrate interface by comparing it with the predetermined threshold. Combining the actual material selection of the circuit board and the copper foil thickness, the system retrieves the built-in loss parameters in the calibration library, and solves the depth stress pixel by pixel based on the depth stress mapping formula. All single-point stress data are spliced ​​and summarized according to the unified pixel coordinates to generate a complete three-dimensional stress gradient field file with a unique transfer number. The system automatically pushes the stress data file to the defect detection dedicated cache partition for storage and backup.

[0035] Optionally, the step of sequentially reading all original polarization images in the edge storage, extracting inherent image feature information from multiple dimensions, performing pixel alignment processing on feature maps with size and resolution differences caused by different imaging conditions, and generating a composite polarization feature dataset of uniform specifications through pixel-level fusion and recombination includes: retrieving the reference polarization original image and phase modulation polarization original image stored in the edge storage one by one according to the workpiece number, and loading all image files to be processed completely; extracting three types of feature information for each loaded image, including line edge grayscale change data, polarization reflection intensity change data, and structured light stripe distortion data; addressing the problem of inconsistent feature map resolution and inconsistent image size caused by different imaging devices and shooting modes, performing pixel-dimensional alignment operation of the full feature map based on the selected reference map size as the standard, and after the alignment work is completed, performing multi-type feature data recombination and fusion at the single pixel granularity to generate a composite polarization feature dataset of uniform overall specifications.

[0036] Specifically, the system loads all corresponding original polarization images in batches according to the workpiece index order. In the feature extraction stage, the gray-scale abrupt change feature focuses on capturing the abrupt change points of the gray-scale values ​​of pixels at the edge of the copper foil of the micro-circuit. The polarization reflection feature counts the gray-scale fluctuation range of the same pixel position under various polarization imaging conditions. The structured light stripe distortion feature tracks the degree of stripe deformation offset formed by structured light projection. Since the reference polarization image and the phase modulation image have objective differences in imaging focal length and sampling specifications, the extracted feature maps will have problems with image size and pixel resolution mismatch. The algorithm selects the pixel size of the reference polarization image as a unified standard, and completes the pixel alignment of the remaining feature maps through edge padding, scale scaling and other methods. After all features are mapped to the same pixel grid, the three types of feature data are superimposed and summarized at each pixel position to eliminate the data misalignment problem caused by heterogeneous images, and finally form a composite polarization feature dataset with completely standardized format and size.

[0037] For example, all the original polarization images corresponding to a single circuit board under test are sequentially retrieved and loaded by the system. The algorithm extracts the grayscale changes of the line edge, polarization reflection fluctuations, and structured light stripe deformation related feature information from all images in different dimensions. In view of the objective differences in resolution and image size between different images, the algorithm performs full feature map alignment and calibration according to the pixel specifications of the reference image. On the calibrated pixel grid, multi-class feature fusion and recombination are realized, and finally a set of composite polarization feature dataset with uniform specifications and regular data is output for subsequent defect detection calculations.

[0038] Optionally, the step of feeding the composite polarization feature dataset and the three-dimensional stress gradient field data into the polarization feature enhancement neural network for inference calculation, and adjusting the feature recognition weights of local areas in a targeted manner based on the stress field distribution, and outputting standardized microcrack location, crack width, and crack depth data by the network includes: importing the preprocessed composite polarization feature dataset and the three-dimensional stress gradient field data with accompanying spatial coordinate information into the polarization feature enhancement neural network, and starting the model's forward inference calculation; traversing the stress values ​​corresponding to all coordinate points within the three-dimensional stress gradient field, filtering out coordinate regions where the stress values ​​exceed the system's preset judgment threshold, and increasing the image feature recognition weights corresponding to these regions by 20% to 50%; after the neural network completes all inference calculations, summarizing all detected defect information, and organizing the three types of parameters—microcrack coordinate location, crack width, and crack depth—into a standardized structured data table, which is then fully stored in the edge local storage unit.

[0039] Specifically, the entire inference and computation of the neural network is completed using local edge computing power, eliminating the need to transmit data to the cloud for processing. Pixel areas with excessive stress values ​​represent stress concentration within the circuit, where the probability of microcracks is much higher than in ordinary areas. By locally increasing the feature recognition weights, the neural network's focus on subtle defects in these areas can be enhanced, effectively reducing the probability of missing micro-buried cracks. After the model computation is complete, the system automatically organizes all crack detection results, standardizes the table field format, and uniformly records key information such as workpiece number, crack plane coordinates, crack transverse width, and crack depth. The generated structured data table is directly stored in the local edge storage space.

[0040] For example, after the composite polarization feature dataset and the three-dimensional stress field data are all imported into the neural network, the algorithm traverses all point values ​​in the stress field, filters out multiple image areas where the stress level exceeds the preset threshold, adjusts the feature recognition weight of the corresponding area according to the rules, and the model completes the full-image defect recognition based on the adjusted weight parameters. It summarizes the location, width, and burial depth information of all detected cracks and organizes them into a standardized table, and saves the table file to the edge local storage area.

[0041] Optionally, the step of combining the stress parameters related to the three-dimensional stress gradient field and the crack geometric parameters output by the neural network, and using a crack coupling damage algorithm to complete the quantitative calculation of circuit fatigue damage, generates standardized detection data which is stored and transmitted separately in two paths. This includes: extracting the characteristic stress values ​​of the copper foil surface and the characteristic stress values ​​of the substrate interface from the three-dimensional stress gradient field, and simultaneously retrieving the crack surface extension length and crack depth data output by the neural network; substituting all the above parameters into the crack coupling damage calculation formula to quantitatively calculate the total fatigue damage of the circuit, and generating a standardized detection data package based on the calculation results; the standardized detection data is processed through two transmission links. The first link writes the detection data to the edge local non-volatile storage medium for permanent archiving, and the second link uses the industrial fieldbus communication protocol to transmit the data to the production line control system in real time to complete data interaction, such as... Figure 3 As shown.

[0042] Specifically, the mathematical expression for the crack-coupled damage algorithm is: ; in: This represents cumulative fatigue damage to the circuit. The number of stress cycles that have been withstood. To determine the surface characteristic stress Crack surface extension length Crack depth And fatigue life calculated from material SN curves and fracture mechanics parameters; ; ; ; ; These are the calibration constants for damage fitting. The average stress on the surface of the entire image. The average stress of the entire image interface is determined by the calibration constants, which are stored in the calibration database after being calibrated through fatigue tests on multiple batches of standard plates with defects.

[0043] The system matches and selects corresponding damage correction coefficients based on the substrate type and actual crack propagation morphology of the circuit board under test. Stress parameters and crack size parameters are then input into the calculation formula to solve for fatigue damage values. These damage values ​​are combined with comprehensive crack detection information to form a standardized detection message. After message generation, a dual-path distribution strategy is executed simultaneously. One copy of the data is written to a non-volatile storage device that is not lost even when power is off for long-term storage. The other copy is uploaded in real-time to the production line's supporting control PLC system via a field industrial bus, allowing the production line to adjust pre-processing steps such as material loading and etching based on the defect data.

[0044] For example, the system extracts the stress data of the surface and interface features inside the stress field, and simultaneously retrieves the measured parameters of crack length, width and depth that have been stored. It selects an appropriate damage correction coefficient based on the circuit board material selection and crack growth morphology, and substitutes it into the coupling calculation formula to complete the fatigue damage calculation of a single circuit. A complete detection data package is generated. One copy of the data package is permanently stored in the edge local storage hardware, and the other copy is transmitted to the production line main control system in real time via the industrial bus to complete the data synchronization.

[0045] Optionally, based on the statistical data summarized from the inspection of the entire batch of circuit boards, the calibration database parameters, neural network weight parameters, and image acquisition hardware timing parameters are adjusted iteratively in three categories. The updated parameters are written to the edge controller's storage space. These include: the configuration parameters that need iterative optimization and adjustment are divided into three independent parameter sets. The first parameter set consists of optical attenuation parameters corresponding to various substrate materials and copper foil thicknesses within the calibration database; the second parameter set consists of feature weight parameters corresponding to all network layers of the polarization feature enhancement neural network; and the third parameter set consists of hardware timing configuration parameters corresponding to the original image acquisition process. The crack detection error, fatigue damage distribution, and imaging acquisition anomaly data of all circuit boards in the same batch are statistically analyzed, and the above three parameter sets are corrected and iterated based on the batch statistical results. The updated set of parameters is uniformly written to the program storage partition of the edge controller. When a new batch of circuit boards performs the S1 image acquisition step, the edge controller directly retrieves the updated parameter set to complete the hardware and algorithm initialization configuration.

[0046] Specifically, after all the inspection work on a single batch of circuit boards is completed, the system automatically summarizes the inspection log for the entire batch, compares the deviation range between the crack size calculated by the algorithm and the actual physical size, and the error distribution law of the Stokes parameter solution. Based on the error distribution, it reversely corrects the optical attenuation parameters corresponding to the substrate and copper foil in the calibration database. According to the statistical results of batch defect false detection and missed detection samples, it fine-tunes the weights of each level of the neural network to optimize the model's defect recognition capability. Combining the exposure time and polarization device switching delay data during the batch imaging process, it corrects the hardware switching sequence and exposure start-stop sequence of the optical module. All optimized and updated parameters are uniformly written into the fixed storage area of ​​the edge master controller. When the image acquisition of a new workpiece is started in the next round, the edge master controller prioritizes loading the latest parameter configuration and directly drives the optical hardware and algorithm modules to carry out a new round of inspection according to the optimized standard.

[0047] For example, after the batch inspection of the entire batch of circuit boards is completed, the system summarizes the defect detection deviation data and imaging timing anomaly records of the entire batch, and corrects and optimizes the optical parameters of the calibration library, the neural network layer weights, and the image acquisition hardware timing in batches. All updated parameters are uniformly stored in the edge master control storage area. When the next batch of circuit boards is put into online inspection and the image acquisition work is started, the edge master control automatically calls the updated set of parameters to complete the preliminary configuration.

[0048] Optionally, the method further includes: statistically analyzing the concentrated locations of micro-circuit defects on circuit boards and the distribution range of fatigue damage values ​​in batches; combining the substrate material type and copper foil processing thickness information corresponding to the high-incidence defect areas to reverse-calculate and supplement the missing optical parameter entries of niche specification boards in the calibration database; and synchronously recording the newly added board parameters into the edge master control configuration file to continuously expand the equipment's adaptability and testing capabilities for different non-standard specification circuit boards.

[0049] Specifically, after completing the inspection task for each production batch, the system automatically summarizes the coordinates of crack occurrence areas and fatigue damage statistics for all workpieces in that batch. It also identifies the board specifications corresponding to high-incidence defects. If high-frequency defects are concentrated on niche substrates or copper foil products with unconventional thicknesses not covered by existing calibration parameters, it indicates that the original calibration database parameters are missing. The system uses the Stokes data and stress field data obtained from the actual measurements of this batch to reverse-engineer the optical loss parameters of the corresponding niche board materials, adds the new parameters to the edge calibration database, and synchronously updates the new parameters to the edge master control configuration area. When subsequent non-standard circuit boards of the same specifications are tested online, the equipment can directly retrieve the supplementary parameters to complete the layered calculation and stress field construction.

[0050] For example, during the inspection of a batch of non-standard customized circuit boards, interface cracks were frequently missed. The system summarized the defect data of the entire batch and determined that the missed detection problem was concentrated in specific niche substrates and copper foil products with special thicknesses. Based on the full amount of measured data of this batch, the corresponding optical loss parameters of the board were calculated and added to the calibration database. The internal configuration parameters of the edge controller were updated simultaneously. When the same type of customized circuit boards enter the production line for inspection in the future, the equipment will automatically call the newly added parameters to carry out the full process calculation, thereby improving the problem of missed detection of defects in similar products.

[0051] Optionally, after the original image acquisition is completed and before the global Stokes parameter solution, the process further includes: the edge controller automatically performs a compliance self-check of all acquired images, verifying the image file storage format, pixel bit depth parameters, and the completeness of the four binding identification information; abnormal bitmap files with image damage or missing identification information are isolated and archived separately, and only valid image data with all indicators qualified are selected to enter the subsequent Stokes parameter solution process.

[0052] Specifically, after all images are written to edge storage, the device automatically initiates a preset file self-check program. First, it verifies whether the images meet the uncompressed bitmap storage format standard and whether the pixel storage bit depth is consistent with the preset imaging parameters. Then, it checks each of the four binding identifiers—workpiece number, acquisition time, polarization type, and structured light phase—to ensure they are complete and without missing information. For abnormal images with corrupted files or missing key identifiers, the system creates a separate abnormal folder for isolated storage and retains an abnormal log. Abnormal images are no longer included in subsequent Stokes vector calculations to prevent distortion in layered parameter solutions caused by corrupted images, ensuring a stable and reliable data benchmark for subsequent parameter calculations.

[0053] For example, after all images of a single circuit board have been acquired and stored, the device automatically starts the image self-inspection process, checks the format parameters and binding identification information of each bitmap file one by one, filters out abnormal images that are damaged or missing identifications and stores them separately, and the remaining valid images that have passed all checks are successfully transferred to the next Stokes parameter solving process.

[0054] Optionally, the method further includes: edge storage packages and archives the original image files of the whole month, the parameter iteration logs of the whole batch, and the defect detection summary report according to the natural month cycle; when production line maintenance personnel need to trace back the detection data, they can realize the full-link data traceability by using the unique workpiece number and fully retrieve the entire process record from original image acquisition, parameter calculation, defect calculation to parameter iteration optimization.

[0055] Specifically, edge storage uses a calendar month as the archiving cycle, creating an independent archive directory each month. Within each directory, all data is categorized by data type, including original polarization bitmaps, layered Stokes vector data, 3D stress field files, structured crack detection tables, and parameter iteration change logs for each round. Archived files are compressed to save storage space. After the production line maintenance team inputs the target workpiece's unique number, the system retrieves the corresponding end-to-end data level by level according to the number index, completely reproducing the entire process record of the workpiece from image capture and storage to parameter iteration optimization. This facilitates subsequent product defect tracing and process problem review.

[0056] For example, when production line maintenance personnel need to review the entire process data of a single defective circuit board inspection, they can enter the corresponding workpiece number into the maintenance terminal. The system will then retrieve the original acquired image of the workpiece, Stokes layered solution data, three-dimensional stress field file, crack detection report, and the equipment parameter record of the current period from the monthly archive data package, thus completing the data traceability review of the entire process.

[0057] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A visual inspection method for micro-line defects on circuit boards based on edge computing, characterized in that, The specific steps of this method are as follows: S1. Raw Image Acquisition: The edge master controller drives the coaxial integrated composite polarization optical module to acquire reference polarization reflection images and phase modulation polarization images in a time-division manner and store them in the edge storage. S2. Global Stokes parameter solution: The global composite Stokes vector is calculated from the reference polarization reflection image in step S1, and the Stokes vector at the interface between the copper foil surface and the copper foil substrate is obtained by using the Stokes layer decoupling algorithm. The result is then bound to the phase modulation polarization image in step S1 according to the workpiece number. S3. Construction of three-dimensional stress field: Based on the layered Stokes vector in step S2 and the four-phase structured light image in step S1, the optical response of the copper foil surface and interface is separated, and the three-dimensional stress gradient field of the line is generated by the depth stress mapping algorithm and transferred to step S4. S4. Microcrack Feature Detection: Extract the multidimensional polarization features of the original polarization image from step S1 and fuse them. Combine the stress gradient field from step S3 with the input polarization feature enhancement neural network to output the location, width, and depth of the microcracks, and store them locally. S5. Damage Calculation and Closed-Loop Optimization: Using the stress gradient field data from step S3 and the crack parameters from step S4, fatigue damage is calculated using the crack-coupled damage algorithm and output to the production line control system. Based on batch statistical iteration, the calibration database, neural network, and optical hardware configuration are adjusted, and the edge controller is updated to optimize the acquisition and control in the next round of step S1.

2. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, S1, the original image acquisition, includes two sequentially executed image acquisition stages. The first image acquisition stage sequentially completes the acquisition of reference polarization reflection images at four fixed polarization angles: 0°, 45°, 90°, and 135°. The total time from the start of acquisition to the completion of storage of the images at the four polarization angles does not exceed 20ms. The second image acquisition stage completes the acquisition of phase-modulated polarization images corresponding to four fixed spatial phases. All acquired image files are stored in the edge storage unit in an uncompressed bitmap format.

3. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 2, characterized in that, In S1, during the original image acquisition, in the second image acquisition stage, phase-modulated structured light is projected onto the micro-circuit area of ​​the circuit board under test at a fixed projection angle of 30°. After each phase-modulated polarization image is acquired, it is bound with four types of identification information: corresponding workpiece number, acquisition time, polarization type, and structured light phase. The identification information and the corresponding image file are bound and stored in the edge storage unit.

4. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In S2, the global Stokes parameter solution process, the global composite Stokes vector solution process only calls the reference polarization reflection image data collected in S1 corresponding to the same circuit board under test, and does not call the phase modulation polarization image data. The Stokes vector of the copper foil surface and the Stokes vector of the copper foil substrate interface are stored in two independent partitions of the edge storage unit, respectively. The two types of Stokes vector data and the phase modulation polarization image collected in S1 are associated with each other through a unified workpiece number.

5. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In the S3 three-dimensional stress field construction, based on the preset numerical standards that the depolarization rate of the reflected light from the copper foil surface is less than 15% and the depolarization rate of the reflected light from the copper foil substrate interface is greater than 40%, the optical response components of the copper foil surface and the optical response components of the copper foil substrate interface are distinguished and separated. The calibration database built into the edge storage unit stores the optical loss parameters of three types of substrates: FR4 substrate, aluminum substrate, and PI substrate. At the same time, it stores the optical loss parameters corresponding to three copper foil thicknesses of 18μm, 35μm, and 70μm.

6. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In S3, during the construction of the three-dimensional stress field, the calculated stress data of the copper foil surface and the stress data of the copper foil substrate interface are matched and stitched together at the pixel level according to the spatial coordinate system corresponding to the thickness of the copper foil on the circuit board under test. The stitched and integrated three-dimensional stress gradient field dataset is equipped with a unique data flow identifier and is transmitted and stored in the independent cache partition corresponding to S4 in the edge storage unit.

7. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In S4, during microcrack feature detection, all polarization original images stored in S1 are read one by one, and grayscale change data of line edges, polarization reflection intensity change data, and structured light stripe distortion data are extracted from the images respectively. A unified pixel dimension alignment process is performed on feature maps with different imaging sources and different pixel resolutions. The aligned multi-class feature data is then subjected to pixel-level fusion and recombination processing to generate a composite polarization feature dataset of a unified specification.

8. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In step S4, during the microcrack feature detection, the polarization feature enhancement neural network performs inference operations based on the stress distribution coordinate data of the three-dimensional stress gradient field. According to the coordinate region where the stress value in the three-dimensional stress gradient field exceeds the preset threshold, the image feature recognition weight corresponding to the region is increased by 20% to 50%. The microcrack location coordinates, crack width values, and crack burial depth values ​​output by the neural network are uniformly organized into a standardized structured data table, and the structured data table is completely stored in the edge local storage unit.

9. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In S5, during damage calculation and closed-loop optimization, standardized detection data is generated after fatigue damage calculation is completed using the crack-coupled damage algorithm. One path of the standardized detection data is permanently stored in the edge local non-volatile storage medium, and the other path is transmitted to the production line control system in real time through the industrial fieldbus communication protocol to complete data interaction and storage.

10. The method for visual inspection of micro-circuit defects on circuit boards based on edge computing according to claim 1, characterized in that, In S5, during damage assessment and closed-loop optimization, the iteratively adjusted configuration parameters include three sets of parameters. The first set of parameters consists of optical attenuation parameters corresponding to various substrates and copper foil thicknesses within the calibration database. The second set of parameters consists of feature weight parameters corresponding to all network layers of the polarization feature enhancement neural network. The third set of parameters consists of hardware timing configuration parameters for the image acquisition stage in S1. All iteratively updated parameter sets are completely written into the edge master control program storage area. When the circuit board under test executes step S1, the edge master control directly calls the updated parameter sets to complete the image acquisition configuration.

Images