A high-precision printed circuit board intelligent visual detection system

Abstract

The application discloses a kind of high-precision printed circuit board intelligent visual detection systems, it is related to the field of precision optical detection technology, including, based on surface attachment quantum dot mark layer's printed circuit board substrate, laser wavelength is used to carry out confocal scanning, generate original image dataset;Original image dataset is carried out multispectral feature analysis, capture the focus shift law of different wavelength light, generate wavelength-focus mapping relationship compensation instruction set;According to the compensation instruction set of wavelength-focus mapping relationship, real-time adjustment imaging light path, generate high-fidelity fusion image;Analyze the optical signal feature change in high-fidelity fusion image, identify abnormal area and extract microscopic defect feature, generate the detection report containing defect feature.The application realizes the molecular level combination of solder resist layer and fluorescent mark material by the uniform dispersion of quantum dot modified ink in inert atmosphere and plasma substrate activation process.

Description

Technical Field

[0001] This invention relates to the field of precision optical inspection technology, and in particular to a high-precision intelligent visual inspection system for printed circuit boards. Background Technology

[0002] The current state of printed circuit board (PCB) inspection technology shows that with the miniaturization of electronic devices and the widespread application of 5G high-frequency and high-speed technology, line width and spacing have entered the microscopic scale, driving innovation in inspection technology. Current mainstream technologies include: confocal laser scanning microscopy, which achieves subsurface tomography through point scanning and pinhole filtering; multispectral tomography, which utilizes spectral channel acquisition to reconstruct inner layer structures; and intelligent recognition algorithms, which extract and classify defect features based on convolutional neural networks. These three technological approaches together construct a modern inspection technology system to support the industry's quality control needs.

[0003] Traditional optical methods are limited by the trade-off between diffraction limit and signal-to-noise ratio, resulting in blind spots in the identification of micro-cracks and buried hole fissures. Ultrasonic testing suffers from severe signal attenuation due to material impedance mismatch, leading to the failure to effectively detect critical quality issues. Wavelength-dependent focus drift caused by broadband light sources penetrating multilayer media results in spatial mismatch of multispectral channels. Existing compensation mechanisms rely on mechanical displacement adjustment, and their response speed cannot meet the inspection requirements of high-speed production lines, creating a technical bottleneck in manufacturing scenarios with stringent real-time requirements. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a high-precision intelligent visual inspection system for printed circuit boards to solve the problems of blind spots in micro-defect identification and limitations in real-time compensation technology.

[0006] This invention provides a high-precision intelligent vision inspection system for printed circuit boards, comprising,

[0007] The substrate marking layer coating module constructs a uniform fluorescent marking layer on the surface of the printed circuit board through quantum dot modified ink coating and photocuring process, thereby generating a printed circuit board substrate with a quantum dot marking layer attached to the surface.

[0008] The image dataset generation module uses a laser wavelength for confocal scanning on a printed circuit board substrate with a quantum dot labeling layer attached to its surface to generate the original image dataset.

[0009] The focus shift compensation module performs multispectral feature analysis on the original image dataset, captures the focus shift patterns of light at different wavelengths, and generates a wavelength-focus mapping relationship compensation instruction set.

[0010] The imaging optical path adjustment module adjusts the imaging optical path in real time according to the compensation instruction set of the wavelength-focus mapping relationship to generate a high-fidelity fused image;

[0011] The defect report generation module analyzes the changes in optical signal characteristics in high-fidelity fused images, identifies abnormal areas and extracts microscopic defect features, and generates an inspection report containing the defect features.

[0012] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0013] As a preferred embodiment of the high-precision printed circuit board intelligent visual inspection system of the present invention, the steps of the quantum dot modified ink coating and photocuring process are as follows:

[0014] Quantum dot materials and photocurable solder resist inks are mixed in an inert atmosphere, and modified inks with uniformly dispersed quantum dots are obtained through mechanical stirring and cyclic grinding processes.

[0015] A surface-activated substrate is obtained by plasma surface activation treatment of the printed circuit board substrate.

[0016] As a preferred embodiment of the high-precision intelligent visual inspection system for printed circuit boards according to the present invention, the steps for constructing a uniform fluorescent marking layer on the surface of the printed circuit board to generate a printed circuit board substrate with a quantum dot marking layer attached to its surface are as follows:

[0017] Based on a surface-activated substrate, a uniformly thick wet film coating is formed by combining modified ink with uniformly dispersed quantum dots through a slot coating process, thus generating a semi-cured substrate.

[0018] A semi-cured substrate is subjected to staged light irradiation to trigger the photocrosslinking reaction of modified ink to form a fully cured substrate. The quantum dots of the fully cured substrate are monitored simultaneously to generate a printed circuit board substrate with a quantum dot marking layer attached to the surface.

[0019] In a preferred embodiment of the high-precision printed circuit board intelligent visual inspection system of the present invention, the steps for the printed circuit board substrate based on the surface-attached quantum dot marking layer are as follows:

[0020] A solution for generating motion trajectories is proposed by vacuum adsorption and fixing the quantum dot labeled substrate on a printed circuit board substrate with a surface-attached quantum dot labeling layer, combined with machine vision positioning.

[0021] Based on the motion trajectory scheme, the laser source emits an excitation beam that matches the intrinsic response of the quantum dot, and performs point-by-point scanning to generate a fluorescence signal stream.

[0022] In a preferred embodiment of the high-precision printed circuit board intelligent vision inspection system of the present invention, the step of using laser wavelength for confocal scanning to generate an original image dataset is as follows.

[0023] The wavelength signal of the fluorescence signal stream is separated by spectral filtering, converted into digital coded information in real time, and a digital matrix associated with spatial location is generated.

[0024] The digital matrix of associated spatial locations is reconstructed by pixel mapping according to the spatial order of the trajectory, and the original image dataset is generated by verifying data integrity.

[0025] As a preferred embodiment of the high-precision printed circuit board intelligent visual inspection system of the present invention, the multispectral feature analysis of the original image dataset refers to using a multispectral separation algorithm to extract the fluorescence wavelength channel and the excitation wavelength residual channel, and performing spatial frequency domain analysis to identify the feature point coordinate set.

[0026] As a preferred embodiment of the high-precision printed circuit board intelligent vision inspection system of the present invention, the steps for capturing the focus shift pattern of light of different wavelengths and generating a wavelength-focus mapping relationship compensation instruction set are as follows:

[0027] Calculate the spatial position offset vector of the feature point in the feature point coordinate set under the fluorescence wavelength channel and the excitation wavelength channel, and generate the wavelength-focus offset characteristic curve;

[0028] Based on the wavelength-focus offset characteristic curve, the required focus compensation amount for each wavelength is calculated, and the compensation sequence is optimized using the gradient descent method to generate a compensation instruction set for the wavelength-focus mapping relationship.

[0029] As a preferred embodiment of the high-precision intelligent visual inspection system for printed circuit boards described in this invention, the steps for adjusting the imaging optical path in real time based on the compensation instruction set of the wavelength-focus mapping relationship to generate a high-fidelity fused image are as follows:

[0030] The wavelength partition identifier and compensation priority in the compensation instruction set of the wavelength-focus mapping relationship are identified by the protocol parsing engine, and an optical path control parameter package is generated.

[0031] Based on the optical path control parameter package, the graphene microstructure is driven to undergo nanoscale deformation and angle deflection, and real-time position feedback is received simultaneously to generate a real-time optical path calibration signal stream.

[0032] The original spectral image is acquired by triggering the optical path calibration signal stream, and frame accumulation noise reduction and timestamp registration are performed to generate a high-fidelity fused image.

[0033] In a preferred embodiment of the high-precision intelligent visual inspection system for printed circuit boards described in this invention, the steps for analyzing changes in optical signal features in a high-fidelity fused image are as follows:

[0034] A partitioned histogram equalization algorithm is used to extract optical features from high-fidelity fused images to enhance the image;

[0035] An edge-guided region growing algorithm is used to identify regions of abrupt brightness changes in optically enhanced images, and texture analysis and contour comparison are performed to generate anomaly mask images.

[0036] As a preferred embodiment of the high-precision printed circuit board intelligent vision inspection system of the present invention, the steps of identifying abnormal areas and extracting microscopic defect features to generate an inspection report containing defect features are as follows.

[0037] Based on the abnormal mask image, extract the geometric feature vector and optical feature vector of the brightness change region, and obtain the defect feature dataset;

[0038] Based on the defect feature dataset, a multi-rule joint decision engine is used to analyze the defect features and generate a detection report containing the defect features.

[0039] The beneficial effects of this invention are as follows: by uniformly dispersing quantum dot modified ink in an inert atmosphere and activating it with plasma substrate, the molecular-level bonding between the solder resist layer and the fluorescent labeling material is identified, which transforms the stimulated emission characteristics of quantum dots into an active defect sensing mechanism; based on the dynamic optical path compensation mechanism of wavelength-dependent focus shift law, the nanosecond-level deformation tuning and real-time alignment of multispectral channels of graphene microstructure are identified, which eliminates the dispersion distortion when broadband light sources penetrate multilayer media. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of a high-precision intelligent vision inspection system for printed circuit boards.

[0042] Figure 2 A flowchart of the quantum dot labeling layer fabrication process.

[0043] Figure 3 A flowchart for generating the wavelength focus mapping relationship compensation instruction set.

[0044] Figure 4A flowchart for generating high-fidelity fused images. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-4 This is one embodiment of the present invention, which provides a high-precision intelligent visual inspection system for printed circuit boards, comprising the following steps:

[0049] The substrate marking layer coating module constructs a uniform fluorescent marking layer on the surface of the printed circuit board through quantum dot modified ink coating and photocuring process, thereby generating a printed circuit board substrate with a quantum dot marking layer attached to the surface.

[0050] Quantum dot materials and photocurable solder resist inks are mixed in an inert atmosphere, and modified inks with uniformly dispersed quantum dots are obtained through mechanical stirring and cyclic grinding processes.

[0051] Furthermore, the quantum dot material and the photocurable solder resist ink are mixed in an inert atmosphere environment, and mechanical stirring is carried out by an impeller-type stirrer to maintain a stable inert atmosphere concentration during the stirring process. The mixture is then transferred to a circulating grinding equipment for wet grinding treatment. The grinding media is an array of zirconia beads, and the temperature inside the grinding chamber is controlled to be constant. The dispersion uniformity of the quantum dot material in the ink is identified by adjusting the filling ratio of the grinding beads and the rotor speed. Finally, the modified ink with qualified quantum dot dispersion is output.

[0052] A surface-activated substrate is obtained by plasma surface activation treatment of the printed circuit board substrate;

[0053] Furthermore, the printed circuit board substrate is placed inside a vacuum processing chamber, and process gas is introduced under specific pressure conditions. A high-frequency alternating voltage is applied to the parallel electrode plate to excite the gas to generate low-temperature plasma. The active particles in the plasma bombard the surface of the printed circuit board substrate for a specific time, causing the molecular chains on the substrate surface to break, generating free radicals and forming oxygen-containing polar groups, ultimately generating a surface-activated substrate with enhanced ink adhesion.

[0054] Based on a surface-activated substrate, a uniformly thick wet film coating is formed by combining modified ink with uniformly dispersed quantum dots through a slot coating process, thus generating a semi-cured substrate.

[0055] Furthermore, the surface-activated substrate is fixed in a temperature-controlled device, and the modified ink with quantum dot dispersion meets the standard is input into a precision slit coating device; the gap between the coating blade and the substrate is adjusted to a set size (determined based on the maximum agglomerate size of quantum dots and the surface roughness of the substrate), and the blade completes a single stroke at a constant speed to form a wet film coating; the substrate enters the pre-curing zone to receive ultraviolet irradiation for a set duration (obtained based on the ultraviolet absorption coefficient and thermal conductivity coefficient), so that the surface of the coating is initially cross-linked and the interior remains in a plastic state, and finally a semi-cured substrate is output;

[0056] A semi-cured substrate is subjected to staged light irradiation to trigger the photocrosslinking reaction of modified ink to form a fully cured substrate. The quantum dots of the fully cured substrate are monitored simultaneously to generate a printed circuit board substrate with a quantum dot marking layer attached to the surface.

[0057] Furthermore, the semi-cured substrate is subjected to the first stage of light irradiation under a nitrogen protective atmosphere, which triggers the decomposition of photoinitiator in the shallow region of the wet film coating to generate free radicals; the second stage of light irradiation is then performed to trigger the ink photocrosslinking reaction to fully cure the coating in the depth direction; simultaneously, a fiber optic spectrometer is used to monitor the fluorescence emission spectrum at the excitation wavelength in real time to confirm that the characteristic fluorescence peak wavelength of the quantum dots is within the set range (determined based on the intrinsic emission spectrum peak of the quantum dot material and the allowable deviation of process fluctuations); when the degree of curing and crosslinking reaches the required standard and the fluorescence characteristics of the quantum dots remain stable, a printed circuit board substrate with a quantum dot marking layer attached to the surface is output.

[0058] The image dataset generation module uses a laser wavelength for confocal scanning on a printed circuit board substrate with a quantum dot labeling layer attached to its surface to generate the original image dataset.

[0059] A solution for generating motion trajectories is proposed by vacuum adsorption and fixing the quantum dot labeled substrate on a printed circuit board substrate with a surface-attached quantum dot labeling layer, combined with machine vision positioning.

[0060] Furthermore, a printed circuit board substrate with a quantum dot label layer attached to its surface is placed in the central area of ​​a vacuum adsorption stage, and a vacuum generator is activated to adjust the adsorption pressure to a set value. An industrial camera takes multi-angle photos of the Fiducial marks pre-placed on the surface of the printed circuit board substrate, extracting the pixel coordinates of the center point of each Fiducial mark. The obtained pixel coordinates are then matched with the theoretical coordinates in a preset file (generated by computer-aided software based on the circuit wiring diagram before the printed circuit board is manufactured, containing the theoretical position coordinates of all Fiducial marks) at the sub-pixel level to generate a coordinate transformation matrix. Based on the coordinate transformation matrix, the boundary coordinates of the scanning area are obtained, a collision-free path is planned, and a sequence of starting point coordinates, path turning point coordinates, and scanning speed parameters are generated. Finally, a motion trajectory scheme containing spatial path and motion parameters is output.

[0061] Based on the motion trajectory scheme, the laser source emits an excitation beam that matches the intrinsic response of the quantum dot, and performs point-by-point scanning to generate a fluorescence signal stream;

[0062] Furthermore, based on the starting point coordinates, path turning point sequence, and scanning speed parameters in the motion trajectory scheme, an excitation beam is controlled to emit a wavelength that matches the peak value of the intrinsic absorption spectrum of the quantum dot. The three-dimensional displacement stage moves the laser focusing position point by point according to the path turning point sequence, and stays at each position for a set time (determined comprehensively based on the fluorescence lifetime of the quantum dot and the signal-to-noise ratio requirements of the detection unit) to excite the characteristic fluorescence of the quantum dot. The fluorescence signal is detected by a photomultiplier tube, and after background noise is separated by a narrow-band filter, it is input into a current-to-voltage conversion circuit. The converted analog voltage signal is acquired by an ADC analog-to-digital converter at a set sampling rate, and a spatial coordinate label from a position encoder is added simultaneously to generate a fluorescence signal stream containing spatiotemporal dimensions.

[0063] It should be noted that the intrinsic response of quantum dots refers to the unique optical behavior exhibited by quantum dot materials under laser excitation at a specific wavelength. Its physical essence stems from the quantum confinement effect and band structure characteristics. When the excitation beam wavelength precisely matches the peak value of the quantum dot's intrinsic absorption spectrum, valence band electrons in the quantum dot are excited to transition to the conduction band, forming electron-hole pairs. The excited states of these electron-hole pairs release energy through radiative recombination, producing characteristic fluorescence with a precise wavelength. The core characteristics manifest in three dimensions: size dependence ensures a positive correlation between fluorescence wavelength and quantum dot diameter; surface state modulation ensures material stability through molecular ligand modification; and the defect response mechanism induces changes in fluorescence intensity in microscopic defect regions, forming an active defect sensing signal source. This defect response mechanism transforms quantum dots into high-precision optical probes. Its value lies in identifying the strict matching between excitation / emission wavelengths and optical devices, improving the signal-to-noise ratio through micro-region signal variations, and exhibiting excellent performance in terms of process compatibility, providing a physical basis for breaking through the limits of traditional optical detection.

[0064] The wavelength signal of the fluorescence signal stream is separated by spectral filtering, converted into digital coded information in real time, and a digital matrix associated with spatial location is generated.

[0065] Furthermore, the fluorescence signal stream is input into a multi-stage spectral filtering component. The first stage uses a long-pass filter to filter out residual signals at the excitation wavelength, the second stage uses a band-pass filter to extract the characteristic fluorescence bands of the quantum dots, and the third stage uses a short-pass filter to suppress long-wavelength stray light. After filtering, the fluorescence signal is converted into a pulsed current signal by a photomultiplier tube, and then converted into a voltage signal by a transimpedance amplifier. The voltage signal is input into a high-speed ADC analog-to-digital converter for real-time digitization, and the spatial coordinate data output by the position encoder is acquired synchronously. The digitized voltage value is bound to the corresponding coordinate and encapsulated into a timestamp-coordinate-grayscale value structure data packet, and finally a digital matrix associated with the spatial location is generated.

[0066] The digital matrix of associated spatial locations is reconstructed by pixel mapping according to the spatial order of the trajectory, and the original image dataset is generated by verifying data integrity.

[0067] Furthermore, the digital matrix associated with spatial location is loaded into the data processing unit. The sequence of path turning points in the motion trajectory scheme is read as the index order, and the data packets in the digital matrix are rearranged according to the traversal rules of ascending X coordinate and ascending Y coordinate. For the gaps in the path turning points in the motion trajectory scheme, the bilinear interpolation algorithm is used to fill the gray value gaps. After the digital matrix is ​​reassembled, data integrity verification is performed: the sum of gray values ​​in each row and column is obtained and compared with the preset theoretical total value of the corresponding area in the motion trajectory scheme (calculated based on the material reflectance coefficient and quantum dot concentration distribution in the substrate design parameters), and the continuity of the data packet timestamp is verified. After the verification is passed, a two-dimensional digital matrix is ​​output and encapsulated into HDF5 format data packets, finally generating the original image dataset.

[0068] It should be noted that pixel mapping and reconstruction is the core processing flow that transforms discretely acquired optical signals into spatially ordered images. It establishes a precise mapping between the laser scanning path and pixels through spatial location reconstruction—based on the path turning point sequence of the motion trajectory scheme, the digital matrix data is reconstructed according to the ascending XY coordinate rule; a bilinear interpolation algorithm is used to intelligently fill the gaps between path turning points; rigorous data integrity verification is performed; and finally, a standardized dataset is output through two-dimensional spatial topology reconstruction, HDF5 format encapsulation, and metadata annotation. This overcomes the limitations of traditional image reconstruction: adaptive path mapping solves the problem of locating irregular scanning trajectories, intelligent gap compensation maintains the geometric fidelity of microstructure edges, and physical constraint verification ensures data reliability, providing a high-precision spatial topology data foundation for defect identification.

[0069] The focus shift compensation module performs multispectral feature analysis on the original image dataset, captures the focus shift patterns of light at different wavelengths, and generates a wavelength-focus mapping relationship compensation instruction set.

[0070] Based on the original image dataset, a multispectral separation algorithm was used to extract the fluorescence wavelength channel and the excitation wavelength residual channel, and spatial frequency domain analysis was performed to identify the coordinate set of feature points.

[0071] More specifically, the original image dataset is input into the multispectral separation algorithm processing unit, and the blind source separation algorithm is applied to decompose the mixed spectral signal to extract the quantum dot feature fluorescence wavelength channel and the excitation wavelength residual channel; two-dimensional fast Fourier transform is performed on the separated wavelength channel images to generate frequency domain maps, and spatial frequency domain features are extracted through bandpass frequency filters; energy accumulation regions in the frequency domain map are identified as feature candidate points, the centroid coordinates of the candidate points are obtained and mapped back to the spatial domain position, generating a set of coordinate pairs containing the feature point coordinates of the fluorescence wavelength channel and the feature point coordinates of the excitation wavelength channel, and finally outputting the feature point coordinate set;

[0072] It should be noted that the multispectral separation algorithm is the core processing step for achieving accurate optical feature extraction. Through a blind source separation computational framework, the mixed optical signals in the original image are decomposed into independent spectral components. Signal demixing is achieved by utilizing the intrinsic wavelength difference between the quantum dot characteristic fluorescence and the excitation residual light. The multispectral separation algorithm simulates the spectral separation mechanism of human eye cone cells but overcomes biological limitations, achieving nanometer-level wavelength resolution through independent component analysis. Its core innovative advantages include adaptive demixing capability, frequency domain feature enhancement, and spatial positioning accuracy. It successfully solves the spectral aliasing problem caused by broadband light sources penetrating multilayer media, ensuring signal reliability in complex noise environments, and improving the positioning accuracy of micron-level feature points to the nanometer level. This provides a precise spatial reference for high-density PCB inspection and improves the inner layer defect recognition rate to an industry-leading level.

[0073] Calculate the spatial position offset vector of the feature point in the feature point coordinate set under the fluorescence wavelength channel and the excitation wavelength channel, and generate the wavelength-focus offset characteristic curve;

[0074] More specifically, based on the feature points of the fluorescence wavelength channel and the excitation wavelength channel in the feature point coordinate set, a correspondence is established according to the spatial Euclidean distance matching rule. When the straight-line distance between two points does not exceed the set maximum tolerance distance (set comprehensively based on the maximum feature point position deviation allowed by the substrate manufacturing process and the spatial resolution limit of the detection system), it is recorded as a valid coordinate pair. For each valid coordinate pair, a coordinate difference operation is performed: extract the X coordinate of the excitation channel feature point and subtract the X coordinate of the fluorescence channel feature point to obtain the horizontal offset, and extract the Y coordinate of the excitation channel feature point and subtract the Y coordinate of the fluorescence channel feature point to obtain the vertical offset. All offset data are grouped by wavelength range, and the arithmetic mean of the horizontal and vertical offsets within the same group is calculated. The variation law of wavelength value and offset mean is established through data fitting, and the wavelength-focus offset characteristic curve is finally output in JSON file format, including wavelength-horizontal offset mapping table, wavelength-vertical offset mapping table and statistical analysis confidence interval parameters.

[0075] Spatial position offset and characteristic curve generation formula:

[0076] ;

[0077] in, Represents the average offset vector of the groups; Indicates the first The number of effective coordinate logs of the group; This represents a point-to-point index; Indicates the first fluorescent channel feature points coordinate; Indicates the first For the excitation channel feature points coordinate; Indicates the first fluorescent channel feature points coordinate; Indicates the first For the excitation channel feature points coordinate; Indicates the activation channel; This indicates a fluorescence channel.

[0078] Based on the wavelength-focus offset characteristic curve, the required focus compensation amount for each wavelength is calculated, and the compensation sequence is optimized using the gradient descent method to generate a compensation instruction set for the wavelength-focus mapping relationship.

[0079] More specifically, based on the wavelength-horizontal offset mapping table and the wavelength-vertical offset mapping table in the wavelength-focus offset characteristic curve, the horizontal and vertical focus compensation amounts corresponding to each wavelength node are extracted in ascending order of wavelength; the compensation amount sequence is input into an iterative optimization algorithm, and the compensation amount values ​​are adjusted through continuous iteration to gradually reduce the compensation error function value, generating an optimized focus compensation sequence for each wavelength; the compensation sequence is encapsulated into a binary protocol, and finally, a compensation instruction set for the wavelength-focus mapping relationship is output;

[0080] ;

[0081] in, Indicates the set of compensation instructions; Indicates the first Compensation amount for each wavelength node; Indicates the first Compensation amount for each wavelength node; Indicates wavelength spacing; Indicates the fidelity weight; Indicates the number of wavelength points; Indicates the first Estimated initial compensation values ​​for each wavelength node;

[0082] The imaging optical path adjustment module adjusts the imaging optical path in real time according to the compensation instruction set of the wavelength-focus mapping relationship to generate a high-fidelity fused image;

[0083] The wavelength partition identifier and compensation priority in the compensation instruction set of the wavelength-focus mapping relationship are identified by the protocol parsing engine, and an optical path control parameter package is generated.

[0084] More specifically, the compensation instruction set for the wavelength-focus mapping relationship is input into the protocol parsing engine. First, the four bytes of the protocol header are parsed to obtain the wavelength partition identifier; then, the priority flag field is read; the compensation data body is decoded in wavelength partition order to extract the axial compensation amount and radial compensation amount corresponding to each wavelength node; the compensation parameters are converted into voltage control instructions and finally encapsulated into a structured optical path control parameter package.

[0085] Based on the optical path control parameter package, the graphene microstructure is driven to undergo nanoscale deformation and angle deflection, and real-time position feedback is received simultaneously to generate a real-time optical path calibration signal stream.

[0086] More specifically, the graphene nanounits are deflected at an angle via an interdigital electrode array based on radial voltage commands; real-time position feedback signals from the piezoelectric micro-displacement device are received simultaneously; a proportional-integral-derivative closed-loop control algorithm is used to obtain the deformation depth error and angle deflection error; the real-time deformation depth value, deflection angle value, and calibration status identifier are encapsulated into a data stream, and finally, a real-time optical path calibration signal stream is generated.

[0087] It should be noted that the radial voltage command is derived from the axial displacement parameters in the wavelength-focus mapping compensation command set, which are obtained by analyzing the voltage-displacement conversion function.

[0088] It should be noted that graphene microstructures are micro-nano scale optical elements composed of a single layer of carbon atoms. The honeycomb lattice structure enables real-time control of the optical path through voltage control: when a radial voltage is applied to the interdigitated electrodes, it induces rearrangement of the π-bond electron cloud, producing a high-precision deformation response (in-plane strain dynamically adjusts the bond length of carbon atoms, and out-of-plane bending changes the hybrid orbital angle); the asymmetric electrode design establishes a voltage gradient to induce non-uniform stress in the lattice, achieving high-speed deflection control over a large angle range; forming a closed loop with piezoelectric micro-displacement devices, optical monitoring tracks the deformation depth and deflection angle in real time; graphene microstructures have wide spectral control capabilities, enabling dynamic adjustment of refractive index, plasmonic characteristic tuning, and precise optical path compensation, breaking through the limitations of slow response speed and short lifespan of traditional optical elements, achieving high-precision axial chromatic aberration compensation in multispectral imaging, and meeting the testing needs of precision electronic manufacturing.

[0089] The original spectral image is acquired by triggering the optical path calibration signal stream, and frame accumulation noise reduction and timestamp registration are performed to generate a high-fidelity fused image.

[0090] More specifically, the unprocessed optical data acquired by the multi-channel synchronous acquisition device contains core features of optical information from three independent channels: ultraviolet channel data, captured by a back-illuminated sCMOS detector, mainly includes quantum dot fluorescence features and microcrack scattering signals; visible light channel data, acquired by a global shutter CMOS detector, records the substrate surface morphology and solder joint reflection characteristics; and near-infrared channel data, acquired by an InGaAs array detector, penetrates multiple layers to reveal inner layer structure information. Each channel data undergoes pixel-level frame accumulation calculation, and spatiotemporal synchronization is achieved through timestamp registration. An adaptive weighting algorithm is used to dynamically adjust the number of accumulated frames based on the signal-to-noise ratio: reducing the number of frames in high signal-to-noise ratio regions to improve efficiency, and increasing the number of frames in low signal-to-noise ratio regions to enhance details. The registered 3D dataset constitutes the original spectral image cube, with the spatial dimension recording the physical coordinates of the substrate, the spectral dimension carrying wavelength feature information, and the temporal dimension preserving dynamic process data. In subsequent processing, the wavelength-focus mapping relationship compensation instruction set is called to perform multi-channel feature point spatial alignment, eliminating optically generated dispersion distortion. When merging spectral channels using a weighted pyramid fusion algorithm, the fusion weights are adaptively adjusted based on the fluorescence intensity distribution of quantum dots: the ultraviolet channel enhances microcrack features (weight increase), the visible light channel preserves surface texture (baseline weight), and the near-infrared channel suppresses background noise (weight decrease). After fusion, adaptive edge sharpening compensates for optical diffraction effects, and the image is finally converted to the sRGB color space to generate a high-fidelity image that conforms to the visual characteristics of the human eye.

[0091] The defect report generation module analyzes the changes in optical signal characteristics in high-fidelity fused images, identifies abnormal areas and extracts microscopic defect features, and generates an inspection report containing the defect features.

[0092] A partitioned histogram equalization algorithm is used to extract optical features from high-fidelity fused images to enhance the image;

[0093] Furthermore, the high-fidelity fusion image input processing unit divides the image into fixed-size blocks; performs histogram statistics on the grayscale values ​​of pixels within each block and obtains the cumulative grayscale distribution function of the block; establishes a grayscale mapping relationship based on the cumulative distribution function, mapping the original grayscale values ​​to the target grayscale range (set by analyzing the cumulative distribution function of quantum dot fluorescence intensity); implements contrast suppression for quantum dot aggregation regions during the process; reconstructs the pixel values ​​after processing all blocks, performs bilinear interpolation smoothing on the block boundaries, and finally outputs an optically enhanced image.

[0094] An edge-guided region growing algorithm is used to identify regions of abrupt brightness changes in optically enhanced images, and texture analysis and contour comparison are performed to generate anomaly mask images.

[0095] Furthermore, the optical feature enhancement image input processing unit uses an edge detection algorithm to extract the image gradient magnitude map and uses the gradient maxima as seed points for region growth. When the difference between the pixel grayscale value and the grayscale value of the seed point is less than a set percentage and the texture uniformity index is met, region growth is performed to generate a set of connected regions. Morphological contour polygons are extracted for each connected region and matched with the standard component contour polygons in the substrate file. Polygon regions whose matching distance exceeds the tolerance range (set jointly based on the position tolerance band width of the standard component contour in the substrate file and the extraction accuracy of the feature point coordinate set) are marked as abnormal regions. Finally, an abnormality mask image is output.

[0096] It should be noted that the anomaly mask image is a high-precision binary feature-marked image generated through edge-guided region growth and intelligent contour comparison. The anomaly mask image locates seed points based on gradient information from optically enhanced images, and dynamically expands connected regions using a dual criterion of gray-level difference threshold (dynamically generated by analyzing the gray-level statistical features of the local neighborhood of the component under test) and texture uniformity to form a complete boundary for the anomaly area. A morphological contour extraction algorithm transforms the connected regions into polygonal structures, and intelligent distance matching is performed with the contours of standard components in a standard file. Finally, areas exceeding the tolerance range are marked as anomalies, and the anomaly mask image is output. The anomaly mask image possesses sub-pixel-level edge localization capabilities, fully preserves the true topological structure of micro-defects, intuitively distinguishes normal and abnormal areas using black-and-white binary values, supports micrometer-level defect recognition in spatial accuracy, avoids over-segmentation and misjudgment in geometric fidelity, and adapts to process fluctuations through dynamic tolerance control. This improves the accuracy of identifying defects such as cold solder joints and micro-cracks in precision electronic manufacturing inspection.

[0097] Based on the abnormal mask image, extract the geometric feature vector and optical feature vector of the brightness change region, and obtain the defect feature dataset;

[0098] The abnormal mask image is input into the feature extraction unit. First, a connected component labeling algorithm is executed to identify the independent contours of each region with abrupt changes in brightness. For each independent connected component, geometric feature vectors are extracted, including equivalent diameter, aspect ratio, and contour roundness. Simultaneously, optical feature vectors are extracted from the corresponding high-fidelity fused image, including maximum light intensity gradient value, average fluorescence intensity, and texture uniformity. All features are organized into structured data records according to connected component identifiers. Finally, the data is packaged into a defect feature dataset containing feature vectors of all abnormal regions.

[0099] Based on the defect feature dataset, a multi-rule joint decision engine is used to analyze the defect features and generate a detection report containing the defect features.

[0100] The defect feature dataset is input into the multi-rule joint decision engine. First, it calls the preset defect classification parameter table (by statistically analyzing the geometric and optical feature distribution range of historical defect samples from the production line and combining destructive physical verification experiments to determine the quantization interval boundary values ​​of each defect type). Quantization judgment is performed on the geometric feature vector of each data record; optical feature vectors are analyzed simultaneously; location coordinate mapping is performed on successfully matched defects; when a single defect triggers a multi-rule conflict, feature priority arbitration is performed; a detection report containing three core data types is generated: a defect location heatmap, a type proportion distribution statistical chart, and a size level classification table; finally, a formatted detection report containing defect features is output.

[0101] In summary, this invention achieves molecular-level bonding between the solder resist layer and the fluorescent labeling material through the uniform dispersion of quantum dot modified ink in an inert atmosphere and plasma substrate activation process, thereby transforming the stimulated emission characteristics of quantum dots into an active defect sensing mechanism; and achieves nanosecond-level deformation tuning and real-time alignment of multispectral channels of graphene microstructures based on a dynamic optical path compensation mechanism with wavelength-dependent focus offset law, thereby eliminating chromatic distortion when broadband light sources penetrate multilayer media.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A high-precision intelligent visual inspection system for printed circuit boards, characterized in that: include, The substrate marking layer coating module constructs a uniform fluorescent marking layer on the surface of the printed circuit board through quantum dot modified ink coating and photocuring processes, generating a printed circuit board substrate with a quantum dot marking layer attached to its surface. The steps are as follows. Quantum dot materials and photocurable solder resist inks are mixed in an inert atmosphere, and modified inks with uniformly dispersed quantum dots are obtained through mechanical stirring and cyclic grinding processes. A surface-activated substrate is obtained by plasma surface activation treatment of the printed circuit board substrate; Based on a surface-activated substrate, a uniformly thick wet film coating is formed by combining modified ink with uniformly dispersed quantum dots through a slot coating process, thus generating a semi-cured substrate. A semi-cured substrate is subjected to staged light irradiation to trigger the photocrosslinking reaction of modified ink to form a fully cured substrate. The quantum dots of the fully cured substrate are monitored simultaneously to generate a printed circuit board substrate with a quantum dot marking layer attached to the surface. The image dataset generation module uses a laser wavelength for confocal scanning on a printed circuit board substrate with a quantum dot labeling layer attached to its surface to generate the original image dataset. The focus shift compensation module performs multispectral feature analysis on the original image dataset to capture the focus shift patterns of light at different wavelengths, generating a wavelength-focus mapping compensation instruction set. The steps are as follows. Calculate the spatial position offset vector of the feature point in the feature point coordinate set under the fluorescence wavelength channel and the excitation wavelength channel, and generate the wavelength-focus offset characteristic curve; Based on the wavelength-focus offset characteristic curve, the required focus compensation amount for each wavelength is calculated, and the compensation sequence is optimized using the gradient descent method to generate a compensation instruction set for the wavelength-focus mapping relationship. ； in, Indicates the set of compensation instructions; Indicates the first Compensation amount for each wavelength node; Indicates the first Compensation amount for each wavelength node; Indicates wavelength spacing; Indicates the fidelity weight; Indicates the number of wavelength points; Indicates the first Estimated initial compensation values ​​for each wavelength node; The imaging optical path adjustment module adjusts the imaging optical path in real time according to the compensation instruction set of the wavelength-focus mapping relationship to generate a high-fidelity fused image; The defect report generation module analyzes the changes in optical signal characteristics in high-fidelity fused images, identifies abnormal areas and extracts microscopic defect features, and generates an inspection report containing the defect features.

2. The high-precision printed circuit board intelligent vision inspection system as described in claim 1, characterized in that: The steps for the printed circuit board substrate based on the surface-attached quantum dot labeling layer are as follows. A solution for generating motion trajectories is proposed by vacuum adsorption and fixing the quantum dot labeled substrate on a printed circuit board substrate with a surface-attached quantum dot labeling layer, combined with machine vision positioning. Based on the motion trajectory scheme, the laser source emits an excitation beam that matches the intrinsic response of the quantum dot, and performs point-by-point scanning to generate a fluorescence signal stream.

3. The high-precision printed circuit board intelligent vision inspection system as described in claim 2, characterized in that: The process of using laser wavelength for confocal scanning to generate the original image dataset involves the following steps. The wavelength signal of the fluorescence signal stream is separated by spectral filtering, converted into digital coded information in real time, and a digital matrix associated with spatial location is generated. The digital matrix of associated spatial locations is reconstructed by pixel mapping according to the spatial order of the trajectory, and the original image dataset is generated by verifying data integrity.

4. The high-precision printed circuit board intelligent vision inspection system as described in claim 3, characterized in that: The multispectral feature analysis of the original image dataset refers to using a multispectral separation algorithm to extract the fluorescence wavelength channel and the excitation wavelength residual channel, and then performing spatial frequency domain analysis to identify the feature point coordinate set.

5. The high-precision printed circuit board intelligent vision inspection system as described in claim 1, characterized in that: The steps for adjusting the imaging optical path in real time based on the compensation instruction set according to the wavelength-focus mapping relationship to generate a high-fidelity fused image are as follows: The wavelength partition identifier and compensation priority in the compensation instruction set of the wavelength-focus mapping relationship are identified by the protocol parsing engine, and an optical path control parameter package is generated. Based on the optical path control parameter package, the graphene microstructure is driven to undergo nanoscale deformation and angle deflection, and real-time position feedback is received simultaneously to generate a real-time optical path calibration signal stream. The original spectral image is acquired by triggering the optical path calibration signal stream, and frame accumulation noise reduction and timestamp registration are performed to generate a high-fidelity fused image.

6. The high-precision printed circuit board intelligent vision inspection system as described in claim 5, characterized in that: The steps for analyzing changes in optical signal features in high-fidelity fused images are as follows. A partitioned histogram equalization algorithm is used to extract optical features from high-fidelity fused images to enhance the image; An edge-guided region growing algorithm is used to identify regions of abrupt brightness changes in optically enhanced images, and texture analysis and contour comparison are performed to generate anomaly mask images.

7. The high-precision printed circuit board intelligent vision inspection system as described in claim 6, characterized in that: The steps for identifying abnormal regions, extracting microscopic defect features, and generating a detection report containing these defect features are as follows: Based on the abnormal mask image, extract the geometric feature vector and optical feature vector of the brightness change region, and obtain the defect feature dataset; Based on the defect feature dataset, a multi-rule joint decision engine is used to analyze the defect features and generate a detection report containing the defect features.

Images