An intelligent analysis system for optical communication carrier board detection with integrated Moe structure large model

Abstract

The present invention relates to the field of optoelectronic detection technology, and specifically to an intelligent analysis system for detecting optical communication carrier boards with an integrated Moe structure large-scale model. The system comprises a dynamic optical signal acquisition module, a multidimensional feature preprocessing module, a Moe structure dynamic analysis module, a defect map reconstruction module, and an intelligent diagnosis output module. The system comprises: a dynamic optical signal acquisition module for acquiring raw optical signal data from the carrier board surface; a multidimensional feature preprocessing module for outputting a multidimensional feature tensor; a dynamic Moe structure analysis module for outputting a spatial heat map; a defect map reconstruction module for generating a three-dimensional defect distribution map; and an intelligent diagnosis output module for outputting a structured diagnosis report. By constructing an intelligent analysis system that integrates multidimensional optical feature analysis with expert dynamic modeling, the system achieves high-precision three-dimensional reconstruction of optical communication carrier board defects and links them to repair decision-making, significantly improving detection accuracy and diagnostic execution efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of optoelectronic detection, and in particular to an intelligent analysis system for detecting a large-scale optical communication carrier board with an integrated MoE structure. Background Art

[0002] With the development of high-frequency and high-speed optical communication technology, optical communication substrates, as key components of core optoelectronic interconnect structures, have microscopic defects that significantly affect signal integrity, optical path consistency, and overall system reliability. During the substrate manufacturing and assembly process, typical defects such as scratches, bubbles, and delamination frequently occur. Their morphology is complex and their distribution is three-dimensional, making it difficult to achieve accurate identification and deep positioning through traditional two-dimensional imaging or surface vision inspection methods.

[0003] Current mainstream inspection methods rely heavily on manual experience-assisted judgment or simple image feature extraction. These methods struggle to identify deep structural defects and provide a traceable basis for repair decisions, severely hindering the high-quality delivery and long-term stable operation of optical communication systems. Therefore, an intelligent analysis system for optical communication carrier board inspection, integrating a large-scale MoE structure model, is urgently needed to address these issues. Summary of the Invention

[0004] Based on the above objectives, the present invention provides an intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model.

[0005] An intelligent analysis system for optical communication carrier board detection with an integrated Moe structure large model includes an optical signal dynamic acquisition module, a multi-dimensional feature preprocessing module, a Moe structure dynamic analysis module, a defect map reconstruction module, and an intelligent diagnosis output module; wherein:

[0006] Optical signal dynamic acquisition module: used to obtain the original optical signal data on the carrier surface through the confocal laser scanning device, and output a three-dimensional light field matrix including scattering intensity, phase offset, and polarization angle;

[0007] Multidimensional feature preprocessing module: used to receive the three-dimensional light field matrix, generate an enhanced light field matrix through a wavelet domain dynamic noise suppression algorithm, extract frequency domain energy distribution features and spatial gradient features, and output a multidimensional feature tensor;

[0008] Moe structure dynamic analysis module: used to input multi-dimensional feature tensors into the Moe structure large model composed of N expert sub-networks. Each expert sub-network dynamically assigns weights to scratches, bubbles, and delamination defects, and outputs defect probability distribution vectors and spatial heat maps.

[0009] Defect map reconstruction module: Based on the spatial heat map and combined with the refractive index parameters of the substrate material, the three-dimensional spatial coordinates of the defects are reconstructed through the inverse ray tracing algorithm to generate a three-dimensional defect distribution map with defect type annotations;

[0010] Intelligent diagnosis output module: used to match the preset process parameter library based on the three-dimensional defect distribution map, and output a structured diagnosis report containing defect location, type, and repair suggestions.

[0011] Optionally, the optical signal dynamic acquisition module includes a laser emission unit, a confocal scanning unit, an optical demodulation unit and a data conversion unit; wherein:

[0012] Laser emitting unit: used to emit a continuous intensity linearly polarized laser beam at a central wavelength of 532nm, with the laser beam output power controlled within the range of 10mW to 50mW;

[0013] Confocal scanning unit: used to perform point-by-point laser focus scanning on the substrate surface with an optical axial resolution better than 1μm, supported by an XYZ three-dimensional adjustable platform, and obtain reflected and scattered return signals at each scanning point;

[0014] Optical demodulation unit: used to perform interferometric measurement analysis on the optical signal returned by scanning, extract the scattering intensity, phase offset and polarization angle parameters of each point, and eliminate non-linear interference components;

[0015] Data conversion unit: used to convert the extracted three types of optical measurement parameters into a three-dimensional light field matrix with a unified structure.

[0016] Optionally, the multidimensional feature preprocessing module includes a wavelet decomposition unit, a threshold noise suppression unit, a reconstruction generation unit, a frequency domain feature extraction unit, a spatial gradient feature extraction unit, and a feature tensor assembly unit; wherein:

[0017] Wavelet decomposition unit: receives the three-dimensional light field matrix output by the optical signal dynamic acquisition module, performs a three-level two-dimensional discrete wavelet transform on each two-dimensional slice, and obtains high-frequency coefficients at three scales and low-frequency coefficients at one scale;

[0018] Threshold denoising unit: This unit is used to denoise the high-frequency coefficients at each scale using a fixed hard threshold method. The hard threshold for the first scale is set to 0.018, the hard threshold for the second scale is set to 0.011, and the hard threshold for the third scale is set to 0.006. Coefficients with amplitudes below the corresponding threshold are directly set to zero, while coefficients with amplitudes above the threshold remain unchanged.

[0019] Reconstruction generation unit: used to perform inverse wavelet transform on the high-frequency coefficients after threshold noise suppression and the original low-frequency coefficients to generate an enhanced light field matrix;

[0020] Frequency domain feature extraction unit: used to perform three-dimensional fast Fourier transform on the enhanced light field matrix, calculate the energy density of the three wave number bands of 0–1 / 3 Nyquist, 1 / 3–2 / 3 Nyquist, and 2 / 3–Nyquist, and form the frequency domain energy distribution feature vector;

[0021] Spatial gradient feature extraction unit: used to calculate the first-order differential along the x, y, and z directions in the spatial domain, summarize the three-dimensional gradient amplitude array, and perform maximum pooling on the array to compress it into a fixed-length gradient feature vector;

[0022] Feature tensor assembly unit: used to splice the frequency domain energy distribution feature vector and the spatial gradient feature vector in index order, and output a multi-dimensional feature tensor in a unified format.

[0023] Optionally, the Moe structure dynamic analysis module includes a feature distribution unit, an expert sub-network group, a dynamic weighted fusion unit, a defect probability decoding unit, and a spatial thermal mapping unit; wherein:

[0024] Feature distribution unit: used to receive the multidimensional feature tensor output by the multidimensional feature preprocessing module, calculate the correlation weight coefficients of N expert sub-networks through a fully connected gating network, and broadcast the complete feature tensor to all expert sub-networks simultaneously in the same clock cycle;

[0025] Expert subnetwork group unit: including the first expert subnetwork for scratch defects, the second expert subnetwork for bubble defects, and the third expert subnetwork for delamination defects, which are used to output the confidence score vector and spatial attention map of the corresponding defect type;

[0026] Dynamic weighted fusion unit: used to perform weighted summation on the confidence score vectors and spatial attention maps output by each expert sub-network based on the correlation weight coefficient generated by the feature distribution unit, to obtain the global defect probability intermediate result and the fused spatial attention matrix;

[0027] Defect probability decoding unit: used to normalize the intermediate results of global defect probability to form a one-dimensional defect probability distribution vector, in which each component corresponds to the probability of occurrence of three types of defects: scratches, bubbles, and delamination;

[0028] Spatial heat mapping unit: used to perform channel-level superposition of the fused spatial attention matrix and the enhanced light field matrix to generate a spatial heat map that corresponds to the carrier size one by one. The heat value range is between 0 and 1.

[0029] Optionally, the feature distribution unit includes:

[0030] Input normalization subunit: used to flatten the multi-dimensional feature tensor into a one-dimensional vector, calculate the mean and standard deviation of the vector, and then complete the linear normalization of the vector;

[0031] Gated activation subunit: used to input the normalized feature vector into the fully connected gating network and calculate the activation vector through the weight matrix and bias vector. The formula is: ,in, is the weight matrix; is the normalized eigenvector; is the bias vector; is the activation vector;

[0032] Probability mapping subunit: used to apply Soft-Max mapping to the activation vector to generate the correlation weight coefficient corresponding to each expert sub-network. The formula is: ,in, For the The weight coefficient of each expert sub-network; is the activation vector elements; is the activation vector elements.

[0033] Optionally, the dynamic weighted fusion unit includes:

[0034] Score fusion subunit: It is used to receive the correlation weight coefficient output by the feature distribution unit and the confidence score vector output by each expert subnetwork, and perform weighted superposition on the scores corresponding to each expert subnetwork according to the weight coefficient to obtain the intermediate result of the global defect probability; the formula is: ,in, is the global defect probability intermediate result vector; For the The confidence score vector of the expert sub-network; is the number of expert sub-networks;

[0035] Thermal fusion subunit: It is used to perform pixel-by-pixel weighted summation of the spatial attention maps output by each expert subnetwork according to the corresponding weight coefficient to obtain the fused spatial attention matrix, which is used to reflect the global defect spatial distribution trend; the formula is: ,in, To fuse the spatial attention matrix at position The value of For the The spatial attention map of the expert sub-network is at position The value of .

[0036] Optionally, the defect map reconstruction module includes a thermal threshold segmentation unit, a refractive index correction unit, a reverse ray tracing unit, a coordinate fusion unit, a defect annotation unit and a map output unit; wherein:

[0037] Thermal threshold segmentation unit: used to receive the spatial thermal map generated by the dynamic weighted fusion unit, divide the substrate surface into defect candidate areas and background areas based on the thermal value threshold, and record the thermal peak position for each candidate area;

[0038] Refractive index correction unit: used to read the pre-stored refractive index parameters of the carrier material and perform optical path refraction correction on the thermal peak position of the candidate area;

[0039] Reverse ray tracing unit: It is used to perform reverse ray tracing along the incident light direction with pixel steps, starting from the incident coordinates of the refraction-corrected surface. It searches for a path with increasing thermal value in the 3D voxel space until the thermal gradient drops to zero, thereby determining the center position of the defective voxel.

[0040] Coordinate fusion unit: used to spatially cluster the voxel center positions output by the inverse ray tracing unit, calculate the geometric centroid of each cluster, and generate a list of three-dimensional spatial coordinates of defects;

[0041] Defect annotation unit: calls the defect probability distribution vector output by the defect probability decoding unit, matches the defect type with the highest probability with the corresponding three-dimensional coordinates, and forms a defect entity record with type annotation;

[0042] Atlas output unit: used to map defect entity records to a unified three-dimensional model coordinate system according to three-dimensional coordinates, and generate a three-dimensional defect distribution map with defect type annotations.

[0043] Optionally, the reverse ray tracing unit includes:

[0044] Ray initialization subunit: used to take the incident coordinates of the refraction-corrected surface as the starting position vector and read the unit direction vector of the incident light , and set the pixel-level step size scalar ;

[0045] Step iterator unit: It is used to perform reverse stepping along the direction vector in the three-dimensional voxel space according to a fixed step size to update the current position. The update formula is: ,in, For the Position information after iterations; For the Position information at the iteration;

[0046] Gradient determination subunit: used to obtain the thermal value of the current position after each step , while calculating the local thermal gradient ,when The stop sign is triggered when

[0047] Center positioning subunit: used to record the maximum position vector in the thermal value sequence , and output it as the defect voxel center coordinates.

[0048] Optionally, the coordinate fusion unit includes:

[0049] Initial clustering subunit: used to perform single-round density clustering on the voxel center coordinate set output by the reverse ray tracing subunit according to the Euclidean distance threshold, and calculate the distance between any two points. and Euclidean distance between ,when Less than or equal to the preset distance threshold , then these two points are classified into the same initial cluster;

[0050] Cluster merging subunit: used to compare all initial clusters pairwise. If the minimum distance between clusters is not greater than , then the corresponding clusters are merged until the minimum distance between any clusters is greater than ;

[0051] Centroid calculation subunit: used to calculate the centroid after merging Within a cluster The geometric center of mass is obtained from the coordinates of the voxel center. The formula is: ,in, For the The geometric centroid coordinates of the clusters; For the The number of voxel center coordinates within a cluster; The first Voxel center coordinates;

[0052] List output subunit: used to generate the list according to the centroid calculation subunit A list of defect three-dimensional spatial coordinates is constructed sequentially, and a cluster identification number is assigned to each record.

[0053] Optionally, the intelligent diagnosis output module includes a defect attribute extraction unit, a parameter matching unit, a repair suggestion generation unit, and a report structured output unit; wherein:

[0054] Defect attribute extraction unit: This unit receives the 3D defect distribution map output by the defect map reconstruction module, extracts the spatial coordinates, type identification, and distribution scale information of each defect, and converts them into structured defect attribute tuples. Each attribute tuple includes the corresponding location coordinates, defect type label, volume estimation value, and local thermal significance level.

[0055] Parameter matching unit: Used to search for the best matching item in the preset process parameter library based on the type label and volume estimate in the defect attribute tuple. The process parameter library is indexed by defect type + size range and stores the corresponding treatment process code, process feasibility level and recommended treatment path. The matching item outputs the recommended treatment process entry corresponding to the defect.

[0056] Repair suggestion generation unit: This unit combines the processing process items returned by the parameter matching unit and formulates repair suggestions for the corresponding defects based on processing equipment resources, repair priority rules, and defect spatial distribution density. The suggestions include treatment methods, recommended tool models, recommended time windows, and precautions.

[0057] Report structured output unit: used to classify, encode and format the spatial coordinates, defect types, matching treatment processes and repair suggestions of all defects, and generate a structured diagnostic report according to a unified report template.

[0058] Beneficial effects of the present invention:

[0059] The present invention constructs an intelligent analysis system integrating a large Moe structure model, introduces confocal laser scanning, multi-dimensional feature preprocessing, expert network dynamic fusion, defect map reconstruction and diagnostic output modules, and achieves high-precision recognition and three-dimensional visual positioning of surface and sub-surface defects of optical communication carrier boards; the system can automatically classify and judge different types of defects, output probability distribution and spatial heat map, and significantly improve the accuracy of detection results and spatial resolution capabilities.

[0060] The present invention combines the defect map with the material refractive index parameters, uses the inverse ray tracing algorithm to accurately reconstruct the three-dimensional coordinates of the defect, and links it with the preset process parameter library to automatically generate a structured diagnostic report containing the defect location, type, and treatment suggestions, forming an intelligent closed loop from detection to repair suggestion output, which greatly improves the fault handling efficiency and process response accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0062] Figure 1 Schematic diagram of an intelligent analysis system for optical communication carrier board detection according to an embodiment of the present invention;

[0063] Figure 2 Schematic diagram of the Moe structure dynamic analysis module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0065] like Figure 1-Figure 2 As shown, an integrated Moe structure large model optical communication carrier board detection intelligent analysis system includes an optical signal dynamic acquisition module, a multi-dimensional feature preprocessing module, a Moe structure dynamic analysis module, a defect map reconstruction module, and an intelligent diagnosis output module; wherein:

[0066] Optical signal dynamic acquisition module: used to obtain the original optical signal data on the carrier surface through the confocal laser scanning device, and output a three-dimensional light field matrix including scattering intensity, phase offset, and polarization angle;

[0067] Multidimensional feature preprocessing module: used to receive the three-dimensional light field matrix, generate the enhanced light field matrix through the wavelet domain dynamic noise suppression algorithm, extract the frequency domain energy distribution characteristics and spatial gradient characteristics, and output the multidimensional feature tensor;

[0068] Moe structure dynamic analysis module: used to input multi-dimensional feature tensors into the Moe structure large model composed of N expert sub-networks. Each expert sub-network dynamically assigns weights to scratches, bubbles, and delamination defects, and outputs defect probability distribution vectors and spatial heat maps.

[0069] Defect map reconstruction module: Based on the spatial heat map and combined with the refractive index parameters of the substrate material, the three-dimensional spatial coordinates of the defects are reconstructed through the inverse ray tracing algorithm to generate a three-dimensional defect distribution map with defect type annotations;

[0070] Intelligent diagnosis output module: used to match the preset process parameter library based on the three-dimensional defect distribution map, and output a structured diagnosis report containing defect location, type, and repair suggestions.

[0071] The optical signal dynamic acquisition module includes a laser emission unit, a confocal scanning unit, an optical demodulation unit and a data conversion unit; wherein:

[0072] Laser emission unit: used to emit a continuous intensity linearly polarized laser beam at a central wavelength of 532nm. The laser beam output power is controlled within the range of 10mW to 50mW to ensure sufficient irradiation penetration into the surface structure of the substrate.

[0073] Confocal scanning unit: used to perform point-by-point laser focus scanning on the substrate surface with an optical axial resolution better than 1μm, supported by an XYZ three-dimensional adjustable platform, and obtain reflected and scattered return signals at each scanning point;

[0074] Optical demodulation unit: used to perform interferometric measurement analysis on the optical signal returned by scanning, extract the scattering intensity, phase offset and polarization angle parameters of each point, and eliminate non-linear interference components;

[0075] Data conversion unit: used to convert the three types of extracted optical measurement parameters into a structurally unified three-dimensional light field matrix form; the above unit organically integrates the functional units of laser emission, spatial scanning, optical demodulation and data matrix conversion. It can obtain the three types of optical response characteristics of the carrier surface with high precision at micron-level resolution, forming a three-dimensional light field matrix in a unified format, providing a structured, high signal-to-noise ratio input basis for subsequent defect identification and modeling analysis.

[0076] The multidimensional feature preprocessing module includes a wavelet decomposition unit, a threshold noise suppression unit, a reconstruction generation unit, a frequency domain feature extraction unit, a spatial gradient feature extraction unit, and a feature tensor assembly unit; wherein:

[0077] Wavelet decomposition unit: receives the three-dimensional light field matrix output by the optical signal dynamic acquisition module, performs a three-level two-dimensional discrete wavelet transform on each two-dimensional slice, and obtains high-frequency coefficients at three scales and low-frequency coefficients at one scale;

[0078] Threshold denoising unit: This unit is used to denoise the high-frequency coefficients at each scale using a fixed hard threshold method. The hard threshold for the first scale is set to 0.018, the hard threshold for the second scale is set to 0.011, and the hard threshold for the third scale is set to 0.006. Coefficients with amplitudes below the corresponding threshold are directly set to zero, while coefficients with amplitudes above the threshold remain unchanged.

[0079] Reconstruction generation unit: used to perform inverse wavelet transform on the high-frequency coefficients after threshold noise suppression and the original low-frequency coefficients to generate an enhanced light field matrix with improved signal-to-noise ratio;

[0080] Frequency domain feature extraction unit: used to perform three-dimensional fast Fourier transform on the enhanced light field matrix, calculate the energy density of the three wave number bands of 0–1 / 3 Nyquist, 1 / 3–2 / 3 Nyquist, and 2 / 3–Nyquist, and form the frequency domain energy distribution feature vector;

[0081] Spatial gradient feature extraction unit: used to calculate the first-order differential along the x, y, and z directions in the spatial domain, summarize the three-dimensional gradient amplitude array, and perform maximum pooling on the array to compress it into a fixed-length gradient feature vector;

[0082] Feature tensor assembly unit: used to splice the frequency domain energy distribution feature vector and the spatial gradient feature vector in index order, and output a multidimensional feature tensor in a unified format for call by the Moe structural dynamic analysis module; the above unit sets a hard threshold with a fixed value for the three-layer high-frequency coefficients and performs hierarchical wavelet noise suppression. This multidimensional feature preprocessing module removes random noise and periodic stripe noise while maintaining the defect edge details, thereby improving the signal-to-noise ratio of the enhanced light field matrix by approximately 9dB, thereby improving the accuracy of subsequent defect identification.

[0083] The Moe structural dynamic analysis module includes a feature distribution unit, an expert sub-network group, a dynamic weighted fusion unit, a defect probability decoding unit, and a spatial thermal mapping unit; among which:

[0084] Feature distribution unit: used to receive the multidimensional feature tensor output by the multidimensional feature preprocessing module, calculate the correlation weight coefficients of N expert sub-networks through a fully connected gating network, and broadcast the complete feature tensor to all expert sub-networks simultaneously in the same clock cycle;

[0085] Expert subnetwork group unit: This includes a first expert subnetwork for scratch defects, a second expert subnetwork for bubble defects, and a third expert subnetwork for delamination defects. Each expert subnetwork adopts a convolutional attention structure and completes end-to-end training on its own dedicated defect dataset to output a confidence score vector and spatial attention map for the corresponding defect type;

[0086] Dynamic weighted fusion unit: used to perform weighted summation on the confidence score vectors and spatial attention maps output by each expert sub-network based on the correlation weight coefficient generated by the feature distribution unit, to obtain the global defect probability intermediate result and the fused spatial attention matrix;

[0087] Defect probability decoding unit: used to normalize the intermediate results of global defect probability to form a one-dimensional defect probability distribution vector, in which each component corresponds to the probability of occurrence of three types of defects: scratches, bubbles, and delamination;

[0088] Spatial thermal mapping unit: used to perform channel-level superposition of the fused spatial attention matrix and the enhanced light field matrix to generate a spatial heat map that corresponds to the substrate size. The thermal value range is between 0 and 1, which is used to reflect the spatial significance of the defect. The above unit uses a multi-expert collaboration mechanism driven by gated weights. The Moe structural dynamic analysis module can simultaneously realize defect probability quantification and spatial significance positioning in a single forward reasoning process, which not only improves the defect recognition accuracy, but also provides visual heat map results, enhancing the spatial accuracy of subsequent defect map reconstruction.

[0089] The feature distribution unit includes:

[0090] Input normalization subunit: used to flatten the multidimensional feature tensor into a one-dimensional vector, calculate the mean and standard deviation of the vector, and then complete the linear normalization of the vector; the normalization formula is: ,in, is the one-dimensional eigenvector after flattening; is the mean of a one-dimensional vector; is the standard deviation of a one-dimensional vector; is the normalized eigenvector;

[0091] Gated activation subunit: used to input the normalized feature vector into the fully connected gating network and calculate the activation vector through the weight matrix and bias vector. The formula is: ,in, is the weight matrix; is the normalized eigenvector; is the bias vector; is the activation vector;

[0092] Probability mapping subunit: used to apply Soft-Max mapping to the activation vector to generate the correlation weight coefficient corresponding to each expert sub-network. The formula is: ,in, For the The weight coefficient of each expert sub-network; is the activation vector elements; is the activation vector elements; through the step-by-step processing of normalization, gated activation and probability mapping, the feature distribution unit can output numerically stable and interpretable weight coefficients within a single inference cycle, ensuring that each expert sub-network obtains accurate dynamic weights based on the relevance of the current sample, thereby improving the discrimination accuracy of defect probability fusion.

[0093] The dynamic weighted fusion unit includes:

[0094] Score fusion subunit: It is used to receive the correlation weight coefficient output by the feature distribution unit and the confidence score vector output by each expert subnetwork, and perform weighted superposition on the scores corresponding to each expert subnetwork according to the weight coefficient to obtain the intermediate result of the global defect probability; the formula is: ,in, is the global defect probability intermediate result vector; For the The confidence score vector of the expert sub-network; is the number of expert sub-networks;

[0095] Thermal fusion subunit: It is used to perform pixel-by-pixel weighted summation of the spatial attention maps output by each expert subnetwork according to the corresponding weight coefficient to obtain the fused spatial attention matrix, which is used to reflect the global defect spatial distribution trend; the formula is: ,in, To fuse the spatial attention matrix at position The value of For the The spatial attention map of the expert sub-network is at position The above subunits perform weighted calculations through the score fusion subunit and the thermal fusion subunit. The dynamic weighted fusion unit can accurately superimpose various defect judgments and spatial response results based on the feature correlation of each expert subnetwork, and generate intermediate outputs that are both discriminative and have spatial distribution clarity, providing a consistent and high-quality input basis for defect probability decoding and thermal map generation.

[0096] The defect map reconstruction module includes a thermal threshold segmentation unit, a refractive index correction unit, a reverse ray tracing unit, a coordinate fusion unit, a defect annotation unit, and a map output unit; wherein:

[0097] Thermal threshold segmentation unit: used to receive the spatial thermal map generated by the dynamic weighted fusion unit, divide the substrate surface into defect candidate areas and background areas based on the thermal value threshold, and record the thermal peak position for each candidate area;

[0098] Refractive index correction unit: used to read the pre-stored refractive index parameters of the carrier material, perform optical path refraction correction on the thermal peak position of the candidate area, and output the surface incident coordinate set adjusted by the material refraction effect;

[0099] Reverse ray tracing unit: It is used to perform reverse ray tracing along the incident light direction with pixel steps, starting from the incident coordinates of the refraction-corrected surface. It searches for a path with increasing thermal value in the 3D voxel space until the thermal gradient drops to zero, thereby determining the center position of the defective voxel.

[0100] Coordinate fusion unit: used to spatially cluster the voxel center positions output by the inverse ray tracing unit, calculate the geometric centroid of each cluster, and generate a list of three-dimensional spatial coordinates of defects;

[0101] Defect annotation unit: calls the defect probability distribution vector output by the defect probability decoding unit, matches the defect type with the highest probability with the corresponding three-dimensional coordinates, and forms a defect entity record with type annotation;

[0102] Atlas output unit: used to map defect entity records into a unified 3D model coordinate system according to 3D coordinates, generate a 3D defect distribution map with defect type annotations, and pass this map to the intelligent diagnosis output module; through a hierarchical process of threshold segmentation, refractive index correction, reverse ray tracing and type annotation, the defect map reconstruction module can accurately locate the 3D coordinates of defects inside the carrier and attach defect type information, significantly improving the resolution and classification accuracy of defect space reconstruction, and providing a reliable 3D visualization basis for subsequent structured diagnosis reports.

[0103] The reverse ray tracing unit includes:

[0104] Ray initialization subunit: used to take the incident coordinates of the refraction-corrected surface as the starting position vector and read the unit direction vector of the incident light , and set the pixel-level step size scalar ;

[0105] Step iterator unit: It is used to perform reverse stepping along the direction vector in the three-dimensional voxel space according to a fixed step size to update the current position. The update formula is: ,in, For the Position information after iterations; For the Position information at the iteration;

[0106] Gradient determination subunit: used to obtain the thermal value of the current position after each step , while calculating the local thermal gradient , whose expression is: ,in,- For the Thermal gradient during segment stepping; The thermal values ​​of the corresponding positions are The stop sign is triggered when

[0107] Center positioning subunit: used to record the maximum position vector in the thermal value sequence , and output it as the coordinates of the defect voxel center; through the hierarchical process of ray initialization, step-by-step tracking, gradient monitoring and extreme value positioning, the reverse ray tracing subunit can determine the three-dimensional coordinates of the defect with voxel-level accuracy, avoid positioning errors caused by material refraction, and significantly improve the accuracy and stability of defect space reconstruction.

[0108] The coordinate fusion unit includes:

[0109] Initial clustering subunit: used to perform single-round density clustering on the voxel center coordinate set output by the reverse ray tracing subunit according to the Euclidean distance threshold, and calculate the distance between any two points. and Euclidean distance between ,when Less than or equal to the preset distance threshold When , the two points are classified into the same initial cluster; the threshold Pick ; The calculation formula of Euclidean distance is ,in, for point with dot The Euclidean distance of are the three-dimensional coordinate components of the corresponding points respectively;

[0110] Cluster merging subunit: used to compare all initial clusters pairwise. If the minimum distance between clusters is not greater than , then the corresponding clusters are merged until the minimum distance between any clusters is greater than ;

[0111] Centroid calculation subunit: used to calculate the centroid after merging Within a cluster The geometric center of mass is obtained from the coordinates of the voxel center. The formula is: ,in, For the The geometric centroid coordinates of the clusters; For the The number of voxel center coordinates within a cluster; The first Voxel center coordinates;

[0112] List output subunit: used to generate the list according to the centroid calculation subunit A list of three-dimensional spatial coordinates of defects is constructed sequentially, and a cluster identification number is assigned to each record, which is then output to the defect annotation sub-unit. The above sub-units perform hierarchical processing through Euclidean distance threshold clustering, inter-cluster merging, and precise centroid calculation. The coordinate fusion unit can merge discrete voxel centers into physically consistent defect entities and generate a unique and reusable three-dimensional coordinate list, significantly improving the spatial consistency and data manageability of defect positioning.

[0113] The intelligent diagnosis output module includes a defect attribute extraction unit, a parameter matching unit, a repair suggestion generation unit, and a report structured output unit; among which:

[0114] Defect attribute extraction unit: This unit receives the 3D defect distribution map output by the defect map reconstruction module, extracts the spatial coordinates, type identification, and distribution scale information of each defect, and converts them into structured defect attribute tuples. Each attribute tuple includes the corresponding location coordinates, defect type label, volume estimation value, and local thermal significance level.

[0115] Parameter matching unit: Used to search for the best matching item in the preset process parameter library based on the type label and volume estimate in the defect attribute tuple. The process parameter library uses defect type + size range as index and stores the corresponding treatment process code, process feasibility level and recommended treatment path. The matching item outputs the recommended treatment process entry corresponding to the defect.

[0116] Table 1 Example of process parameter library

[0117] serial number Defect Type Volume range (mm³) Process Code Recommended treatment process Feasibility level Processing tool model Recommended repair time window Remarks D001 scratches 0–0.5 SC-P1 Micro polishing + transparent protective layer coating A-level (high) MP-03 Micro-Polishing Unit Within 6 hours Minor scratches on the surface, which will not affect the optical function after repair D002 scratches 0.5–2.0 SC-P2 Laser melting reconstruction + coating restoration Level B (Medium) LM-20 laser module Within 12 hours Risk of light scattering interference in moderately scratched areas D003 bubble 0–1.0 BP-F1 Hot pressing guided exhaust + transparent resin backfill A-level (high) TP-01 Heating Head Within 8 hours For local microbubbles, hot pressing repair is recommended first D004 Delamination 1.0–5.0 DL-S2 Ultrasonic curing agent injection + repackaging Level B (Medium) US-CUR Ultrasonic Needle within 24 hours Delamination structure may affect the mechanical strength of the carrier D005 Delamination >5.0 DL-S3 Laser peeling + structural reconstruction C-level (low) LS-50 laser head Within 48 hours The structure is severely damaged and the repair cost is high. It is recommended to scrap it for evaluation.

[0118] In Table 1 above, the number represents the unique identification of each parameter record; the defect type represents the one-to-one correspondence with the type field output by the defect annotation subunit; the volume range represents the defect volume range adapted to the current process; the process code is used for concise coding called within the system and is bound to the process control program; the recommended processing process represents the specific defect repair method, which must be a standardized process path; the feasibility level represents the processing priority given by the system based on experience and resource assessment; the processing tool model represents the main hardware equipment model required to execute the process; the recommended repair time window represents the recommended optimal processing time from the detection of the defect; the remarks represent supplementary explanations of the risks and priorities after processing.

[0119] Repair suggestion generation unit: This unit combines the processing process items returned by the parameter matching unit and formulates repair suggestions for the corresponding defects based on processing equipment resources, repair priority rules, and defect spatial distribution density. The suggestions include treatment methods, recommended tool models, recommended time windows, and precautions.

[0120] Report structured output unit: used to classify, encode and format the spatial coordinates, defect types, matching treatment processes and repair suggestions of all defects, and generate a structured diagnostic report according to a unified report template. Each defect entry in the report contains five fields: number-location-type-repair process-treatment suggestion, and provides a report export interface; the above unit uses structured attribute extraction, standardized parameter matching and rule-driven suggestion generation process. The intelligent diagnostic output module can not only accurately retrieve process parameters that match the defect characteristics, but also automatically push the optimal repair plan based on the defect location and type, realizing standardized and intelligent output of diagnostic results, and effectively supporting rapid closed-loop defect processing and maintenance decision-making.

[0121] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0122] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An integrated Moe structure large model optical communication carrier board detection intelligent analysis system, characterized by: It includes optical signal dynamic acquisition module, multi-dimensional feature preprocessing module, Moe structure dynamic analysis module, defect map reconstruction module and intelligent diagnosis output module; among which: Optical signal dynamic acquisition module: used to obtain the original optical signal data on the carrier surface through the confocal laser scanning device, and output a three-dimensional light field matrix including scattering intensity, phase offset, and polarization angle; Multidimensional feature preprocessing module: used to receive the three-dimensional light field matrix, generate an enhanced light field matrix through a wavelet domain dynamic noise suppression algorithm, extract frequency domain energy distribution features and spatial gradient features, and output a multidimensional feature tensor; Moe structure dynamic analysis module: used to input multi-dimensional feature tensors into the Moe structure large model composed of N expert sub-networks. Each expert sub-network dynamically assigns weights to scratches, bubbles, and delamination defects, and outputs defect probability distribution vectors and spatial heat maps. The Moe structure dynamic analysis module includes a feature distribution unit, an expert sub-network group, a dynamic weighted fusion unit, a defect probability decoding unit, and a spatial thermal mapping unit; wherein: Feature distribution unit: used to receive the multidimensional feature tensor output by the multidimensional feature preprocessing module, calculate the correlation weight coefficients of N expert sub-networks through a fully connected gating network, and broadcast the complete feature tensor to all expert sub-networks simultaneously in the same clock cycle; Expert subnetwork group unit: including the first expert subnetwork for scratch defects, the second expert subnetwork for bubble defects, and the third expert subnetwork for delamination defects, which are used to output the confidence score vector and spatial attention map of the corresponding defect type; Dynamic weighted fusion unit: used to perform weighted summation on the confidence score vectors and spatial attention maps output by each expert sub-network based on the correlation weight coefficient generated by the feature distribution unit, to obtain the global defect probability intermediate result and the fused spatial attention matrix; Defect probability decoding unit: used to normalize the intermediate results of global defect probability to form a one-dimensional defect probability distribution vector, in which each component corresponds to the probability of occurrence of three types of defects: scratches, bubbles, and delamination; Spatial heat mapping unit: used to perform channel-level superposition of the fused spatial attention matrix and the enhanced light field matrix to generate a spatial heat map that corresponds to the carrier size. The heat value range is between 0 and 1. Defect map reconstruction module: Based on the spatial heat map and combined with the refractive index parameters of the substrate material, the three-dimensional spatial coordinates of the defects are reconstructed through the inverse ray tracing algorithm to generate a three-dimensional defect distribution map with defect type annotations; Intelligent diagnosis output module: used to match the preset process parameter library based on the three-dimensional defect distribution map, and output a structured diagnosis report containing defect location, type, and repair suggestions.

2. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 1 is characterized in that: The optical signal dynamic acquisition module includes a laser emission unit, a confocal scanning unit, an optical demodulation unit and a data conversion unit; wherein: Laser emitting unit: used to emit a continuous intensity linearly polarized laser beam at a central wavelength of 532nm, with the laser beam output power controlled within the range of 10mW to 50mW; Confocal scanning unit: used to perform point-by-point laser focus scanning on the substrate surface with an optical axial resolution better than 1μm, supported by an XYZ three-dimensional adjustable platform, and obtain reflected and scattered return signals at each scanning point; Optical demodulation unit: used to perform interferometric measurement analysis on the optical signal returned by scanning, extract the scattering intensity, phase offset and polarization angle parameters of each point, and eliminate non-linear interference components; Data conversion unit: used to convert the extracted three types of optical measurement parameters into a three-dimensional light field matrix with a unified structure.

3. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 2 is characterized in that: The multidimensional feature preprocessing module includes a wavelet decomposition unit, a threshold noise suppression unit, a reconstruction generation unit, a frequency domain feature extraction unit, a spatial gradient feature extraction unit and a feature tensor assembly unit; wherein: Wavelet decomposition unit: receives the three-dimensional light field matrix output by the optical signal dynamic acquisition module, performs a three-level two-dimensional discrete wavelet transform on each two-dimensional slice, and obtains high-frequency coefficients at three scales and low-frequency coefficients at one scale; Threshold denoising unit: This unit is used to denoise the high-frequency coefficients at each scale using a fixed hard threshold method. The hard threshold for the first scale is set to 0.018, the hard threshold for the second scale is set to 0.011, and the hard threshold for the third scale is set to 0.

006. Coefficients with amplitudes below the corresponding threshold are directly set to zero, while coefficients with amplitudes above the threshold remain unchanged. Reconstruction generation unit: used to perform inverse wavelet transform on the high-frequency coefficients after threshold noise suppression and the original low-frequency coefficients to generate an enhanced light field matrix; Frequency domain feature extraction unit: used to perform three-dimensional fast Fourier transform on the enhanced light field matrix, calculate the energy density of the three wave number bands of 0–1 / 3 Nyquist, 1 / 3–2 / 3 Nyquist, and 2 / 3–Nyquist, and form the frequency domain energy distribution feature vector; Spatial gradient feature extraction unit: used to calculate the first-order differential along the x, y, and z directions in the spatial domain, summarize the three-dimensional gradient amplitude array, and perform maximum pooling on the array to compress it into a fixed-length gradient feature vector; Feature tensor assembly unit: used to splice the frequency domain energy distribution feature vector and the spatial gradient feature vector in index order, and output a multi-dimensional feature tensor in a unified format.

4. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 1 is characterized in that: The feature distribution unit includes: Input normalization subunit: used to flatten the multi-dimensional feature tensor into a one-dimensional vector, calculate the mean and standard deviation of the vector, and then complete the linear normalization of the vector; Gated activation subunit: used to input the normalized feature vector into the fully connected gating network and calculate the activation vector through the weight matrix and bias vector. The formula is: ,in, is the weight matrix; is the normalized eigenvector; is the bias vector; is the activation vector; Probability mapping subunit: used to apply Soft-Max mapping to the activation vector to generate the correlation weight coefficient corresponding to each expert sub-network. The formula is: ,in, For the The weight coefficient of each expert sub-network; is the activation vector elements; is the activation vector elements.

5. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 4 is characterized in that: The dynamic weighted fusion unit includes: Score fusion subunit: It is used to receive the correlation weight coefficient output by the feature distribution unit and the confidence score vector output by each expert subnetwork, and perform weighted superposition on the scores corresponding to each expert subnetwork according to the weight coefficient to obtain the intermediate result of the global defect probability; the formula is: ,in, is the global defect probability intermediate result vector; For the The confidence score vector of the expert sub-network; is the number of expert sub-networks; Thermal fusion subunit: It is used to perform pixel-by-pixel weighted summation of the spatial attention maps output by each expert subnetwork according to the corresponding weight coefficient to obtain the fused spatial attention matrix, which is used to reflect the global defect spatial distribution trend; the formula is: ,in, To fuse the spatial attention matrix at position The value of For the The spatial attention map of the expert sub-network is at position The value of .

6. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 1 is characterized in that: The defect map reconstruction module includes a thermal threshold segmentation unit, a refractive index correction unit, a reverse ray tracing unit, a coordinate fusion unit, a defect annotation unit and a map output unit; wherein: Thermal threshold segmentation unit: used to receive the spatial thermal map generated by the spatial thermal mapping unit, divide the substrate surface into defect candidate areas and background areas based on the thermal value threshold, and record the thermal peak position for each candidate area; Refractive index correction unit: used to read the pre-stored refractive index parameters of the carrier material and perform optical path refraction correction on the thermal peak position of the candidate area; Reverse ray tracing unit: It is used to perform reverse ray tracing along the incident light direction with pixel steps, starting from the incident coordinates of the refraction-corrected surface. It searches for a path with increasing thermal value in the 3D voxel space until the thermal gradient drops to zero, thereby determining the center position of the defective voxel. Coordinate fusion unit: used to spatially cluster the voxel center positions output by the inverse ray tracing unit, calculate the geometric centroid of each cluster, and generate a list of three-dimensional spatial coordinates of defects; Defect annotation unit: calls the defect probability distribution vector output by the defect probability decoding unit, matches the defect type with the highest probability with the corresponding three-dimensional coordinates, and forms a defect entity record with type annotation; Atlas output unit: used to map defect entity records to a unified three-dimensional model coordinate system according to three-dimensional coordinates, and generate a three-dimensional defect distribution map with defect type annotations.

7. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 6 is characterized in that: The reverse ray tracing unit includes: Ray initialization subunit: used to take the incident coordinates of the refraction-corrected surface as the starting position vector and read the unit direction vector of the incident light , and set the pixel-level step size scalar ; Step iterator unit: It is used to perform reverse stepping along the direction vector in the three-dimensional voxel space according to a fixed step size to update the current position. The update formula is: ,in, For the Position information after iterations; For the Position information at the iteration; Gradient determination subunit: used to obtain the thermal value of the current position after each step , while calculating the local thermal gradient ,when The stop sign is triggered when Center positioning subunit: used to record the maximum position vector in the thermal value sequence , and output it as the defect voxel center coordinates.

8. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 7 is characterized in that: The coordinate fusion unit includes: Initial clustering subunit: used to perform single-round density clustering on the voxel center coordinate set output by the reverse ray tracing subunit according to the Euclidean distance threshold, and calculate the distance between any two points. and Euclidean distance between ,when Less than or equal to the preset distance threshold , then these two points are classified into the same initial cluster; Cluster merging subunit: used to compare all initial clusters pairwise. If the minimum distance between clusters is not greater than , then the corresponding clusters are merged until the minimum distance between any clusters is greater than ; Centroid calculation subunit: used to calculate the centroid after merging Within a cluster The geometric center of mass is obtained from the coordinates of the voxel center. The formula is: ,in, For the The geometric centroid coordinates of the clusters; For the The number of voxel center coordinates within a cluster; The first Voxel center coordinates; List output subunit: used to generate the list according to the centroid calculation subunit A list of defect three-dimensional spatial coordinates is constructed sequentially, and a cluster identification number is assigned to each record.

9. The intelligent analysis system for detecting and analyzing optical communication carrier boards with an integrated Moe structure large model according to claim 1 is characterized in that: The intelligent diagnosis output module includes a defect attribute extraction unit, a parameter matching unit, a repair suggestion generation unit and a report structured output unit; wherein: Defect attribute extraction unit: This unit receives the 3D defect distribution map output by the defect map reconstruction module, extracts the spatial coordinates, type identification, and distribution scale information of each defect, and converts them into structured defect attribute tuples. Each attribute tuple includes the corresponding location coordinates, defect type label, volume estimation value, and local thermal significance level. Parameter matching unit: Used to search for the best matching item in the preset process parameter library based on the type label and volume estimate in the defect attribute tuple. The process parameter library is indexed by defect type + size range and stores the corresponding treatment process code, process feasibility level and recommended treatment path. The matching item outputs the recommended treatment process entry corresponding to the defect. Repair suggestion generation unit: This unit combines the processing process items returned by the parameter matching unit and formulates repair suggestions for the corresponding defects based on processing equipment resources, repair priority rules, and defect spatial distribution density. The suggestions include treatment methods, recommended tool models, recommended time windows, and precautions. Report structured output unit: used to classify, encode and format the spatial coordinates, defect types, matching treatment processes and repair suggestions of all defects, and generate a structured diagnostic report according to a unified report template.

Images