Method for identifying bad plates in substrates

Abstract

The invention discloses a method for identifying a bad plate in a substrate, belongs to the technical field, and realizes high fusion of bad plate identification and a production process through a precise marking process and a dynamic data collaboration mechanism. A high-energy laser carbonization technology is adopted, an irreversible permanent mark is formed on the surface of a substrate, and a microscopic carbonized layer structure and a base material form remarkable optical contrast. The mark can still keep stable gray scale characteristics under strong light irradiation, and the problem of misrecognition caused by the fact that a traditional ink mark is prone to falling off and pollution is solved. Meanwhile, the closed-loop control of laser parameters ensures that the marking area is only modified within the depth of 0.1 mm of the surface layer, substrate warping or inner layer circuit damage caused by diffusion of a heat affected zone is avoided, and the reliability and safety of marking of the damaged plate are guaranteed from the physical level.

Description

Technical Field

[0001] This invention belongs to the field of substrate manufacturing technology, specifically a method for identifying defective boards in a substrate. Background Technology

[0002] Board fabrication refers to the process of manufacturing a printed circuit board (PCB) with a predetermined circuit pattern by processing copper-clad laminates through a series of processes, such as drilling, pattern transfer, etching, solder masking, and character printing. This process requires extremely high precision and consistency, directly affecting the performance and reliability of the PCB. Board defect inspection is the quality check performed on the PCB after fabrication, mainly including visual inspection, electrical testing, and functional testing. The purpose of inspection is to identify and reject defective boards, such as those with short circuits, open circuits, misaligned holes, foreign objects, and poor solder pads, ensuring that every board meets the high standards required for electronic products, guaranteeing the stable operation of electronic equipment and production efficiency.

[0003] However, existing technologies mainly rely on manual visual inspection or ink marking to identify defective boards, which has problems such as easy marking falling off and insufficient contrast leading to optical misjudgment. At the same time, the lack of real-time data links prevents production equipment from dynamically obtaining the location information of defective boards, resulting in component waste and a high rework rate. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying defective boards in a substrate in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: A method for identifying defective boards in a substrate, the method comprising the following steps:

[0006] S1: During the substrate design phase, based on signal integrity analysis, a circular badmark area with a diameter of 0.5mm is set in the non-sensitive area, and a corresponding coordinate mapping file is generated;

[0007] S2: During the substrate manufacturing process, the substrate surface is scanned in real time using an optical inspection system to identify potential defects;

[0008] S3: When a bad board is detected, the laser device is triggered to perform high-precision carbonization and blackening in the predetermined badmark area to form an irreversible mark with a contrast of ≥95%.

[0009] S4: Multispectral imaging technology is used to perform secondary verification on the blackened badmark to ensure that the mark has significant features in both the visible and infrared bands;

[0010] S5: Synchronize the coordinate data and marking status of defective boards to the cloud database and dynamically associate them with the production queue of SMT equipment;

[0011] S6: The SMT placement machine uses an integrated AI vision system to analyze substrate images in real time and automatically skips bad boards based on the shape and spectral characteristics of bad marks;

[0012] S7: After each batch of production, the energy parameters and marking and positioning algorithms for laser blackening are dynamically optimized by comparing the actual defective board distribution with the design defect model;

[0013] S8: Based on deep learning models, analyze historical defective board data, predict process steps that are prone to defects, and proactively adjust equipment parameters to reduce the defective board rate.

[0014] S9: Perform reverse analysis on the recovered bad boards, feed back the failure modes to the design end, and iteratively optimize the badmark layout rules and detection logic.

[0015] In a preferred embodiment, in step S1, a circular marking area with a diameter of 0.5 mm is set at a non-sensitive location away from high-speed signal lines and power segmentation areas. This area must be at least 2 mm away from the nearest component pad to prevent subsequent processing from affecting the functional circuit. After the layout is completed, the system automatically generates a mapping file containing the coordinates of all marking points. This file is stored in a binary encrypted format to ensure that the data is not tampered with when transmitted to the manufacturing end.

[0016] In a preferred embodiment, in step S2, after the substrate enters the production line, an optical inspection system equipped with a high-resolution line scan camera and a ring LED light source performs a full-frame scan of the substrate surface at a rate of 120 frames per second. The system employs an adaptive threshold segmentation algorithm to identify defects such as microcracks and uneven copper layers. When a defect area exceeding 0.01 mm is detected... 2 If the length of the edge burr exceeds 50μm, the system triggers an alarm mechanism and transmits the defect coordinate information to the laser processing control unit in real time in XML format.

[0017] In a preferred embodiment, in step S3, after receiving the defect coordinates, the laser device precisely moves the laser focus to the center of the preset bad mark area using a three-axis linkage positioning system. A pulsed fiber laser with a wavelength of 1064nm is used to carbonize the marking area with a peak power of 800W and a pulse width of 20ns. This process reduces the surface reflectivity of the marking area from 75% to below 5%, forming a black circular spot with a contrast of 97% with the surrounding substrate. After blackening is completed, the system uses a 20-megapixel industrial camera to perform a secondary confirmation of the mark, ensuring that the diameter error of the circular spot is less than ±5μm.

[0018] The system dynamically adjusts the laser energy density using the following formula:

[0019]

[0020] Where: Pavg is the average power (adjustable from 300-500W), t_ptp is the pulse width (10-30ns), and A_{spot}Aspot is the spot area (0.5mm diameter corresponds to 0.196mm). 2 ), where \eta_{abs}ηabs is the material absorptivity (FR-4 is 0.82). When a change in substrate material is detected, the system completes parameter reorganization within 50ms to ensure that the energy density is always maintained in the 110-120% range of the ablation threshold, thus ensuring both the marking blackness and avoiding substrate delamination.

[0021] In a preferred embodiment, in step S4, after detection in the visible light band (400-700nm), the system switches to the near-infrared band (850nm) for a transmissive scan of the bad mark. Since the absorption characteristics of the carbonized region for infrared light are significantly different from those of the normal substrate material, the system calculates the grayscale value difference ratio ΔG = (G_visible - G_IR) / G_visible × 100% between the two bands. When ΔG > 85%, the mark is deemed valid. This dual verification mechanism eliminates false judgments caused by ambient light interference, and the verification data is synchronously uploaded to the MES system to generate a digital twin record.

[0022] In a preferred embodiment, in step S5, the cloud database updates a data snapshot every 5ms. The SMT equipment subscribes to database updates in real time via the OPC UA protocol, dynamically removing substrate numbers containing valid bad marks from the production queue. When the pick-and-place machine nozzle moves to the coordinates corresponding to the bad board, the motion controller immediately executes a Z-axis lifting command to ensure that the distance between the nozzle and the bad board remains at least 0.5mm, with a response time of less than 2ms.

[0023] In a preferred embodiment, in step S6, the vision system integrated into the SMT placement machine is equipped with a TensorRT-accelerated ResNet-50 model to perform real-time semantic segmentation of the substrate image. The model distinguishes valid markings from surface contaminants with 99.2% accuracy by analyzing edge gradient changes of bad marks (e.g., detecting whether the slope of the grayscale value abrupt change at the edge of a circular spot is greater than 15% / pixel) and infrared absorption features. When a valid marking is identified, the device automatically skips the current substrate and triggers a buzzer, while simultaneously writing the skip record to a local log file.

[0024] In a preferred embodiment, in step S7, after each batch of production is completed, the system compares the actual defective board distribution data with the defect prediction model from the design phase. Principal component analysis is used to extract the top three key parameters affecting the defect rate (e.g., laser power fluctuation, ambient humidity, and copper foil roughness), and particle swarm optimization algorithm is used to dynamically adjust the laser's output energy curve. In the last 20 iterations, this mechanism has improved laser positioning accuracy from ±15μm to ±8μm and reduced the standard deviation of marking consistency by 42%.

[0025] In a preferred embodiment, step S8 involves constructing a multimodal deep learning model to fuse and analyze multidimensional dynamic data from the substrate manufacturing process with historical defective board distribution patterns, thereby enabling the prediction and proactive intervention of process defects. Specifically, this includes:

[0026] S8-1: Data Fusion Layer: Real-time acquisition of temperature gradient T(x,y,t), etching solution flow rate v(t), laser head wear degree Wn sensor data, and synchronous loading of substrate material properties dielectric constant εr, copper thickness hCu and bad board spatial distribution matrix Dbad(x,y) output by optical detection system;

[0027] S8-2: Spatiotemporal Feature Extraction: A 3D convolutional neural network is used to extract the correlation features of process parameters in the time-space dimension. The formula for establishing the mapping relationship between the etching solution flow rate fluctuation Δv(t) and the bad board accumulation region Dbad(x,y) is as follows:

[0028]

[0029] Where α(t) is the time decay factor, This represents the spatiotemporal convolution operation;

[0030] S8-3: Defect probability prediction: A dynamic risk score R(t) is generated through a gated recurrent unit (GRU) network. An early warning is triggered when R(t) > θ (threshold).

[0031] S8-4: Parameter Self-Optimization: Based on gradient backpropagation, process parameter correction ΔP is generated, and the focal depth δz of the exposure machine is adjusted. The calculation formula is as follows:

[0032]

[0033] Where η is the learning rate. This represents the spatial gradient operator.

[0034] In a preferred embodiment, in step S9, after the recovered defective board is ultrasonically cleaned, a focused ion beam (FIB) is used to perform nanoscale cross-sectional analysis on the marked area. By measuring the carbide layer thickness (typically 3-5 μm) and elemental energy spectrum, it is determined whether the laser energy penetrates the solder resist layer and damages the underlying copper foil. After this data is fed back to the design end, the layout algorithm automatically adds spacing constraints between the marked points and the inner layer traces. The latest version has optimized the minimum spacing between the marked points and adjacent vias from 0.3 mm to 0.6 mm, effectively avoiding secondary defects caused by thermal stress.

[0035] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0036] 1. This invention achieves a high degree of integration between defective board identification and the production process through a precision marking process and a dynamic data collaboration mechanism. High-energy laser carbonization technology is used to form an irreversible, permanent mark on the substrate surface, with its microscopic carbonized layer structure creating a significant optical contrast with the substrate. This mark maintains stable grayscale characteristics even under strong light irradiation, solving the problem of misidentification caused by the easy detachment and contamination of traditional ink markings. Simultaneously, closed-loop control of laser parameters ensures that the marking area is modified only within a 0.1mm depth of the surface layer, avoiding substrate warping or damage to inner layer circuitry caused by heat-affected zone diffusion, thus physically guaranteeing the reliability and safety of defective board marking.

[0037] 2. In this invention, the constructed real-time data stream system transforms defective board information into dynamic instructions driving production decisions. Through millisecond-level cloud synchronization technology, SMT equipment can accurately locate the coordinates of substrates with valid markings before placement. This feedforward control mode not only avoids the lag of traditional manual sampling but also achieves deep collaboration between material flow and information flow. Specifically, when the pick-and-place machine nozzle moves to the coordinates of the defective board, the equipment preemptively executes path avoidance based on the real-time coordinate mapping file, completely eliminating invalid placement actions and significantly improving production efficiency in high-density packaging scenarios. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] Example:

[0041] Reference Figure 1A method for identifying defective boards in a substrate, the method comprising the following steps:

[0042] S1: During the substrate design phase, based on signal integrity analysis, a circular badmark area with a diameter of 0.5mm is set in the non-sensitive area, and a corresponding coordinate mapping file is generated;

[0043] S2: During the substrate manufacturing process, the substrate surface is scanned in real time using an optical inspection system to identify potential defects;

[0044] S3: When a bad board is detected, the laser device is triggered to perform high-precision carbonization and blackening in the predetermined badmark area to form an irreversible mark with a contrast of ≥95%.

[0045] S4: Multispectral imaging technology is used to perform secondary verification on the blackened badmark to ensure that the mark has significant features in both the visible and infrared bands;

[0046] S5: Synchronize the coordinate data and marking status of defective boards to the cloud database and dynamically associate them with the production queue of SMT equipment;

[0047] S6: The SMT placement machine uses an integrated AI vision system to analyze substrate images in real time and automatically skips bad boards based on the shape and spectral characteristics of bad marks;

[0048] S7: After each batch of production, the energy parameters and marking and positioning algorithms for laser blackening are dynamically optimized by comparing the actual defective board distribution with the design defect model;

[0049] S8: Based on deep learning models, analyze historical defective board data, predict process steps that are prone to defects, and proactively adjust equipment parameters to reduce the defective board rate.

[0050] S9: Perform reverse analysis on the recovered bad boards, feed back the failure modes to the design end, and iteratively optimize the badmark layout rules and detection logic.

[0051] In step S1, a circular marking area with a diameter of 0.5 mm is set in a non-sensitive location away from high-speed signal lines and power segmentation areas. This area must be at least 2 mm away from the nearest component pad to prevent subsequent processing from affecting the functional circuit. After the layout is completed, the system automatically generates a mapping file containing the coordinates of all marking points. This file is stored in a binary encrypted format to ensure that the data is not tampered with when transmitted to the manufacturing end.

[0052] In step S2, after the substrate enters the production line, an optical inspection system equipped with a high-resolution line scan camera and a ring LED light source performs a full-frame scan of the substrate surface at a rate of 120 frames per second. The system employs an adaptive threshold segmentation algorithm to identify defects such as microcracks and copper layer inhomogeneity. When a defect area exceeding 0.01 mm is detected... 2 If the length of the edge burr exceeds 50μm, the system triggers an alarm mechanism and transmits the defect coordinate information to the laser processing control unit in real time in XML format.

[0053] In step S3, after receiving the defect coordinates, the laser device precisely moves the laser focus to the center of the preset bad mark area using a three-axis linkage positioning system. A pulsed fiber laser with a wavelength of 1064nm is used to carbonize the marked area with a peak power of 800W and a pulse width of 20ns. This process reduces the surface reflectivity of the marked area from 75% to below 5%, forming a black circular spot with a contrast of 97% with the surrounding substrate. After blackening is completed, the system uses a 20-megapixel industrial camera to perform a secondary confirmation of the mark, ensuring that the diameter error of the circular spot is less than ±5μm.

[0054] The system dynamically adjusts the laser energy density using the following formula:

[0055]

[0056] Where: Pavg is the average power (adjustable from 300-500W), t_ptp is the pulse width (10-30ns), and A_{spot}Aspot is the spot area (0.5mm diameter corresponds to 0.196mm). 2 ), where \eta_{abs}ηabs is the material absorptivity (FR-4 is 0.82). When a change in substrate material is detected, the system completes parameter reorganization within 50ms to ensure that the energy density is always maintained in the 110-120% range of the ablation threshold, thus ensuring both the marking blackness and avoiding substrate delamination.

[0057] In step S4, after detection in the visible light band (400-700nm), the system switches to the near-infrared band (850nm) for a transmissive scan of the bad mark. Since the absorption characteristics of the carbonized region for infrared light are significantly different from those of the normal substrate material, the system calculates the grayscale value difference ratio ΔG = (G_visible - G_IR) / G_visible × 100% between the two bands. When ΔG > 85%, the mark is considered valid. This dual verification mechanism eliminates false positives caused by ambient light interference, and the verification data is simultaneously uploaded to the MES system to generate a digital twin record.

[0058] In step S5, the cloud database updates a data snapshot every 5ms. The SMT equipment subscribes to database updates in real time via the OPC UA protocol, dynamically removing substrate numbers containing valid bad marks from the production queue. When the pick-and-place machine nozzle moves to the coordinates corresponding to the bad board, the motion controller immediately executes a Z-axis lifting command to ensure that the distance between the nozzle and the bad board remains at least 0.5mm, with a response time of less than 2ms.

[0059] In step S6, the vision system integrated into the SMT placement machine uses a TensorRT-accelerated ResNet-50 model to perform real-time semantic segmentation of the substrate image. The model analyzes edge gradient changes of bad marks (e.g., detecting whether the slope of the grayscale value abrupt change at the edge of a circular spot is greater than 15% / pixel) and infrared absorption features, distinguishing valid marks from surface stains with an accuracy of 99.2%. When a valid mark is identified, the device automatically skips the current substrate and triggers a buzzer, while simultaneously writing the skip record to a local log file.

[0060] In step S7, after each batch of production is completed, the system compares the actual defective board distribution data with the defect prediction model from the design phase. Principal component analysis is used to extract the top three key parameters affecting the defect rate (e.g., laser power fluctuation, ambient humidity, and copper foil roughness), and particle swarm optimization algorithm is used to dynamically adjust the laser's output energy curve. In the last 20 iterations, this mechanism has improved laser positioning accuracy from ±15μm to ±8μm and reduced the standard deviation of marking consistency by 42%.

[0061] In step S8, a multimodal deep learning model is constructed to integrate and analyze multidimensional dynamic data from the substrate manufacturing process with historical defective board distribution patterns, enabling the prediction and proactive intervention of process defects. Specifically, this includes:

[0062] S8-1: Data Fusion Layer: Real-time acquisition of temperature gradient T(x,y,t), etching solution flow rate v(t), laser head wear degree Wn sensor data, and synchronous loading of substrate material properties dielectric constant εr, copper thickness hCu and bad board spatial distribution matrix Dbad(x,y) output by optical detection system;

[0063] S8-2: Spatiotemporal Feature Extraction: A 3D convolutional neural network is used to extract the correlation features of process parameters in the time-space dimension. The formula for establishing the mapping relationship between the etching solution flow rate fluctuation Δv(t) and the bad board accumulation region Dbad(x,y) is as follows:

[0064]

[0065] Where α(t) is the time decay factor, This represents the spatiotemporal convolution operation;

[0066] S8-3: Defect probability prediction: A dynamic risk score R(t) is generated through a gated recurrent unit (GRU) network. An early warning is triggered when R(t) > θ (threshold).

[0067] S8-4: Parameter Self-Optimization: Based on gradient backpropagation, process parameter correction ΔP is generated, and the focal depth δz of the exposure machine is adjusted. The calculation formula is as follows:

[0068]

[0069] Where η is the learning rate. This represents the spatial gradient operator.

[0070] In step S9, after ultrasonic cleaning, the recovered defective boards undergo nanoscale cross-sectional analysis of the marked areas using focused ion beam (FIB). By measuring the carbide layer thickness (typically 3-5 μm) and elemental energy spectrum, it is determined whether the laser energy penetrates the solder resist layer and damages the underlying copper foil. This data is fed back to the design team, and the layout algorithm automatically adds spacing constraints between the marked points and inner layer traces. The latest version has optimized the minimum spacing between the marked points and adjacent vias from 0.3 mm to 0.6 mm, effectively avoiding secondary defects caused by thermal stress.

[0071] As can be seen from the above, this invention achieves a high degree of integration between defective board identification and the production process through a precision marking process and a dynamic data collaboration mechanism. High-energy laser carbonization technology is used to form irreversible, permanent marks on the substrate surface, with a significant optical contrast between the microscopic carbonized layer structure and the substrate. This marking maintains stable grayscale characteristics even under strong light irradiation, solving the problems of misidentification caused by the easy detachment and contamination of traditional ink markings. Simultaneously, closed-loop control of laser parameters ensures that the marking area is modified only within a 0.1mm depth on the surface, avoiding substrate warping or damage to inner layer circuitry caused by heat-affected zone diffusion, thus physically guaranteeing the reliability and safety of defective board marking.

[0072] In this invention, a real-time data stream system is constructed to transform defective board information into dynamic instructions that drive production decisions. Through millisecond-level cloud synchronization technology, SMT equipment can accurately locate the coordinates of substrates with valid markings before placement. This feedforward control mode not only avoids the lag of traditional manual sampling but also achieves deep collaboration between material flow and information flow. Specifically, when the pick-and-place machine nozzle moves to the coordinates of the defective board, the equipment preemptively executes path avoidance based on the real-time coordinate mapping file, completely eliminating invalid placement actions and significantly improving production efficiency in high-density packaging scenarios.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0074] The foregoing description enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying defective boards in a substrate, characterized in that: The method includes the following steps: S1: During the substrate design phase, based on signal integrity analysis, a circular badmark area with a diameter of 0.5mm is set in the non-sensitive area, and a corresponding coordinate mapping file is generated; S2: During the substrate manufacturing process, the substrate surface is scanned in real time using an optical inspection system to identify potential defects; S3: When a bad board is detected, the laser device is triggered to perform high-precision carbonization and blackening in the predetermined badmark area to form an irreversible mark with a contrast of ≥95%. S4: Multispectral imaging technology is used to perform secondary verification on the blackened badmark to ensure that the mark has significant features in both the visible and infrared bands; S5: Synchronize the coordinate data and marking status of defective boards to the cloud database and dynamically associate them with the production queue of SMT equipment; S6: The SMT placement machine uses an integrated AI vision system to analyze substrate images in real time and automatically skips bad boards based on the shape and spectral characteristics of bad marks; S7: After each batch of production, the energy parameters and marking and positioning algorithms for laser blackening are dynamically optimized by comparing the actual defective board distribution with the design defect model; S8: Based on deep learning models, analyze historical defective board data, predict process steps that are prone to defects, and proactively adjust equipment parameters to reduce the defective board rate. S9: Perform reverse analysis on the recovered bad boards, feed back the failure modes to the design end, and iteratively optimize the badmark layout rules and detection logic.

2. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S1, a circular marking area with a diameter of 0.5 mm is set in a non-sensitive location away from high-speed signal lines and power segmentation areas; this area must be more than 2 mm away from the nearest component pad to prevent subsequent processing from affecting the functional circuit.

3. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S2, after the substrate enters the production line, an optical inspection system equipped with a high-resolution line scan camera and a ring LED light source performs a full-frame scan of the substrate surface at a speed of 120 frames per second.

4. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S3, after receiving the defect coordinates, the laser device precisely moves the laser focus to the center of the preset bad mark area through a three-axis linkage positioning system; a pulsed fiber laser with a wavelength of 1064nm is used to carbonize the marked area with a peak power of 800W and a pulse width of 20ns; this process reduces the surface reflectivity of the marked area from 75% to below 5%, forming a black circular spot with a contrast of 97% with the surrounding substrate.

5. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S4, after detection in the visible light band, the system switches to the near-infrared band to perform a transmissive scan of the bad mark. Since the absorption characteristics of the carbonized region to infrared light are significantly different from those of the normal substrate material, the system calculates the gray value difference ratio ΔG = / G_visible × 100% under the two bands. When ΔG > 85%, the mark is determined to be valid.

6. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S5, the cloud database updates a data snapshot every 5ms; the SMT equipment subscribes to database updates in real time via the OPC UA protocol and dynamically removes substrate numbers with valid bad marks from the production queue; when the pick-and-place machine nozzle moves to the coordinates corresponding to the bad board, the motion controller immediately executes the Z-axis lifting command to ensure that the distance between the nozzle and the bad board is maintained at a safe distance of more than 0.5mm, and the response time of this action is less than 2ms.

7. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S6, the vision system integrated into the SMT placement machine is equipped with a ResNet-50 model accelerated by TensorRT to perform real-time semantic segmentation on the substrate image. The model distinguishes between valid marks and surface stains with an accuracy of 99.2% by analyzing the edge gradient changes and infrared absorption features of bad marks. When a valid mark is identified, the device automatically skips the current substrate and triggers a buzzer prompt, while writing the skip record to the local log file.

8. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S7, after each batch of production is completed, the system compares the actual defective board distribution data with the defect prediction model in the design stage; principal component analysis is used to extract the top 3 key parameters affecting the defective board rate, and particle swarm optimization algorithm is used to dynamically adjust the output energy curve of the laser.

9. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S8, a multimodal deep learning model is constructed to integrate and analyze multidimensional dynamic data from the substrate manufacturing process with historical defective board distribution patterns, thereby enabling the prediction and proactive intervention of process defects. Specifically, this includes: S8-1: Data Fusion Layer: Real-time acquisition of temperature gradient T(x,y,t), etching solution flow rate v(t), laser head wear degree Wn sensor data, and synchronous loading of substrate material properties dielectric constant εr, copper thickness hCu and bad board spatial distribution matrix Dbad(x,y) output by optical detection system; S8-2: Spatiotemporal Feature Extraction: A 3D convolutional neural network is used to extract the correlation features of process parameters in the time-space dimension. The formula for establishing the mapping relationship between the etching solution flow rate fluctuation Δv(t) and the bad board accumulation region Dbad(x,y) is as follows: Where α(t) is the time decay factor, This represents the spatiotemporal convolution operation; S8-3: Defect probability prediction: A dynamic risk score R(t) is generated through a gated recurrent unit network. An early warning is triggered when R(t) > θ. S8-4: Parameter Self-Optimization: Based on gradient backpropagation, process parameter correction ΔP is generated, and the focal depth δz of the exposure machine is adjusted. The calculation formula is as follows: Where η is the learning rate. This represents the spatial gradient operator.

10. The method for identifying defective boards in a substrate as described in claim 1, characterized in that: In step S9, after the recovered defective board is ultrasonically cleaned, a focused ion beam is used to perform nanoscale cross-sectional analysis on the marked area. By measuring the thickness of the carbide layer and the elemental energy spectrum, it is determined whether the laser energy penetrates the solder resist layer and damages the underlying copper foil. After these data are fed back to the design end, the layout algorithm automatically adds spacing constraints between the marked points and the inner layer traces. The latest version has optimized the minimum spacing between the marked points and adjacent vias from 0.3mm to 0.6mm, effectively avoiding secondary defects caused by thermal stress.

Images