Glass via process defect ai detection system based on near-infrared images

Abstract

The application relates to the technical field of AI detection, and specifically discloses an AI detection system for glass via hole process defects based on a near-infrared image, a production line adaptation module, an infrared light source module, a calculation and detection module, and the like. The production line adaptation module acquires transmission parameters of a TGV etching production line, calculates single-substrate detection time and a detection window, and obtains a correction speed. The infrared light source module superimposes and interferes reflected signal light in a hole with reference light, and acquires an output light intensity distribution relationship. The calculation and detection module performs phase extraction and unwrapping to obtain depth information, and performs weighted fusion to obtain a unified three-dimensional morphology. The calculation and detection module screens a hole area based on the unified three-dimensional morphology, calculates hole diameter deviation and hole depth deviation, classifies glass via hole processes, and performs iterative prediction. The application quantifies hole diameter and hole depth deviation indexes, establishes A, B and C three-level classification standards, and clearly defines product quality grades, thereby solving the problems of low detection precision and poor production line adaptability of existing TGV etching process defects, and improving detection efficiency and quality control level.

Description

Technical Field

[0001] This invention relates to the field of AI detection technology, specifically to an AI detection system for defects in glass through-hole manufacturing processes based on near-infrared imaging. Background Technology

[0002] With the advent of the post-Moore's Law era, three-dimensional heterogeneous integration has become a core technological path to overcome the bottlenecks in chip performance, power consumption, and integration density. Among various high-density packaging solutions, through-glass vias (TGV) have become one of the key interconnect technologies in high-end AI chips, RF devices, power devices, and other packages due to their low dielectric loss, high flatness, excellent long-term reliability, and good compatibility with semiconductor processes.

[0003] The TGV (Through-Video) process mainly includes core steps such as laser modification, etching for via formation, conductive deposition, and electroplating for buried vias. Among these, the etching process directly determines the final quality of the vias. Hole diameter accuracy, hole depth consistency, and sidewall integrity are key indicators affecting product yield and long-term reliability. However, in actual mass production, the TGV via formation process is susceptible to fluctuations in laser energy and uneven etching solution concentration, resulting in defects such as hole diameter deviation, inconsistent hole depth, and hole collapse. If these defects are not effectively identified, they will directly lead to subsequent packaging failures, causing significant economic losses.

[0004] Most mainstream automated optical inspection (AOI) equipment uses visible light imaging, which can only identify macroscopic defects visible on the surface of glass substrates, but cannot penetrate the glass medium to effectively observe the deep morphology within the holes. Some technologies attempt to use infrared imaging methods, but these suffer from low resolution, insufficient signal-to-noise ratio, and complex interference fringe calculations, making it difficult to achieve precise measurement of micron-level hole dimensions. Furthermore, existing technologies mostly employ offline sampling inspection, resulting in long inspection cycles and an inability to achieve real-time full inspection at the production line level. This leads to defects not being detected and reported in a timely manner, resulting in a large number of defective products. Therefore, there is an urgent need for a method for detecting defects in glass through-hole processes. Summary of the Invention

[0005] The purpose of this invention is to provide an AI detection system for defects in glass through-hole manufacturing processes based on near-infrared images, thereby solving the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions: A near-infrared imaging-based AI-powered defect detection system for glass through-hole manufacturing processes includes: Production line adaptation module: acquires the transmission parameters of the TGV etching production line, including the glass substrate transmission speed v, the glass substrate length L, and the total number of TGV holes M on the substrate; Calculate the detection time of a single substrate T=M×t, where t represents the detection time of a single hole. Calculate the detection window W=v×T. Based on the detection window duration W, adjust the transmission speed v to obtain the corrected speed xv. Set the transmission speed of the TGV etching production line to xv. Infrared light source module: A Cartesian coordinate system is pre-set on the glass substrate, and three sets of DFB-LD laser emission wavelengths are pre-set. The DFB-LD lasers are activated to irradiate the glass substrate and acquire the reflected signal light from the aperture. The reflected signal light from the aperture is superimposed and interfered with a pre-set reference light to obtain the output light intensity distribution relationship. , among which, I r I represents the intensity of the reference light. s ω represents the intensity of the reflected signal light inside the aperture, ω0 represents the initial phase of the superimposed interference system, n represents the glass refractive index of the glass substrate, d(x,y) represents the equivalent optical depth of the midpoint (x,y) of the preset Cartesian coordinate system, and λ represents the incident light wavelength corresponding to the DFB-LD laser. Solution module: Extracts the phase and unwraps the superimposed interference fringes to obtain depth information. ; A unified three-dimensional topography is obtained by weighted fusion of depth information calculated by three sets of DFB-LD lasers. Calculation and detection module: Based on a unified 3D topography, the pore regions are selected, and the following operations are performed within the pore regions: The micropore profile is extracted from the orifice plane, and an equivalent circle is fitted to obtain the actual pore diameter Ds and the average depth Hs of the bottom region of the micropore is obtained. Calculate the aperture deviation δ D =|Ds-D| / D, calculate the hole depth deviation δ H =|Hs-H| / H, where D represents the nominal aperture, H represents the nominal aperture depth, and the value is based on the aperture deviation δ. D Hole depth deviation δ H The glass through-hole process is classified.

[0007] As a further aspect of the present invention: in the production line adaptation module, adjusting the transmission speed v based on the detection window W to obtain the corrected speed xv specifically includes: If the detection window W > L, no adjustment will be made, and the current transmission speed v will be maintained; If the detection window W≤L, let the correction speed xv=v(L / W+η), where η represents the preset redundancy coefficient, and η>0.

[0008] As a further aspect of the present invention: in the calculation and detection module, based on the aperture deviation δ D Hole depth deviation δ H The classification of glass through-hole processes includes: If δD and δ H If all values ​​are less than or equal to 2%, it is recorded as Grade A, indicating a good product. If δ D and δ H If any item exceeds 5%, it is classified as Grade C, and the product is considered unqualified. The rest are marked as Grade B, indicating that the product has minor defects, and staff are advised to take action.

[0009] As a further aspect of the present invention: in the solution module, three sets of DFB-LD lasers are pre-set to emit wavelengths of λ1, λ2, and λ3, respectively, and the constraints are satisfied. , where λ min , λ max These represent the preset minimum wavelength difference and maximum wavelength difference, respectively.

[0010] As a further aspect of the present invention: the solution module, which performs weighted fusion based on the depth information calculated by the three sets of DFB-LD lasers to obtain a unified three-dimensional topography, specifically includes: The depth information calculated by the three DFB-LD lasers is weighted and fused to obtain a unified three-dimensional topography: d 3D (x, y) = w1×d1(x, y) + w2×d2(x, y) + w3×d3(x, y), where d1(x, y), d2(x, y), and d3(x, y) represent the depth information obtained by phase extraction and unwrapping of the first, second, and third groups of DFB-LD lasers, respectively, and w1, w2, and w3 represent the preset weight coefficients, and w1+w2+w3=1.

[0011] As a further aspect of the present invention: in the infrared light source module, the angle of the DFB-LD laser is adjusted to ensure that the incident light irradiates the glass substrate perpendicularly.

[0012] As a further aspect of the present invention: in the calculation and detection module, the pore region is screened based on a unified three-dimensional morphology, including: d on the glass substrate 3D The region (x, y) > T is denoted as the hole region, where T represents the preset depth threshold.

[0013] As a further aspect of the present invention: calculate the percentage of Grade A glass substrates in this batch; when the percentage of Grade A glass substrates in Y consecutive batches is less than 90%, prompt the staff to inspect the production equipment, where Y represents the preset judgment batch.

[0014] As a further aspect of the present invention, it also includes a closed-loop control module, which acquires the grading status of the glass through-hole process in this batch, where L represents a preset sampling batch, and calculates the yield rate, where the yield rate is the ratio of the number of grade A glass substrates in this batch to the total number of products. An LSTM-based AI prediction model is used, with the input dataset including the aperture deviation δ of this batch. D The mean value and hole depth deviation δ H The average value, yield, etching temperature, etching solution flow rate, etching solution concentration and processing time are used to output the ideal yield for the next batch. If the ideal yield is greater than or equal to 95%, the current process parameters are maintained and continuous monitoring is performed. If the ideal yield rate is less than 95%, then set the adjustment range for each parameter. The parameter adjustment range includes: Etching temperature adjustment range [W0, 65℃], in 0.5℃ increments; Etching solution flow rate adjustment range [W1, 50L / h], step size 1L / h; Etching solution concentration adjustment range [W2, 25%], step size 0.5%; Processing time adjustment range [W3, 60min], step size 1min, min represents minutes; Where W0, W1, W2, and W3 represent the etching temperature, etching solution flow rate, etching solution concentration, and processing time for this batch, respectively. After obtaining and adjusting the parameters, input them into the AI ​​prediction model, record the output as the actual yield rate, and repeat the iteration until the actual yield rate is greater than or equal to 95%.

[0015] The beneficial effects of the present invention are as follows: (1) It solves the pain points of low defect detection accuracy and poor production line adaptability in the existing TGV etching process, improves detection efficiency and quality control level, and achieves matching between the detection window and the glass substrate by dynamically adjusting the transmission speed through the production line adaptation module to avoid missed detection. It adapts to the mass production needs of substrates of different specifications, breaks the limitations of traditional offline sampling inspection, and realizes real-time full inspection of the entire process. It adopts three sets of DFB-LD laser near-infrared interferometry imaging, combined with phase calculation and multi-wavelength weighted fusion method to obtain the three-dimensional morphology inside the hole, which can effectively identify deep defects such as hole diameter deviation and hole depth unevenness. The detection effect is better than traditional visible light AOI and ordinary infrared detection, improving product yield and reliability.

[0016] (2) By quantifying the deviation index of aperture and depth, a three-level grading standard of A, B and C is established to clarify the product quality level, which makes it easier for staff to quickly handle minor defects and screen unqualified products, thereby solving the problem of separation between traditional testing and process control and helping enterprises reduce production costs. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the structure of the AI ​​detection system for glass through-hole process defects based on near-infrared images according to the present invention; Figure 2 This is a flowchart illustrating the AI ​​detection system for glass through-hole process defects based on near-infrared imaging, as per the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is an AI-based glass through-hole process defect detection system based on near-infrared imaging, comprising: Production line adaptation module: acquires the transmission parameters of the TGV etching production line, including the glass substrate transmission speed v, the glass substrate length L, and the total number of TGV holes M on the substrate; Calculate the detection time of a single substrate T=M×t, where t represents the detection time of a single hole. Calculate the detection window W=v×T. Based on the detection window duration W, adjust the transmission speed v to obtain the corrected speed xv. Set the transmission speed of the TGV etching production line to xv. Infrared light source module: A Cartesian coordinate system is pre-set on the glass substrate, and three sets of DFB-LD laser emission wavelengths are pre-set. The DFB-LD lasers are activated to irradiate the glass substrate and acquire the reflected signal light from the aperture. The reflected signal light from the aperture is superimposed and interfered with a pre-set reference light to obtain the output light intensity distribution relationship. , among which, I r I represents the intensity of the reference light. s ω represents the intensity of the reflected signal light inside the aperture, ω0 represents the initial phase of the superimposed interference system, n represents the glass refractive index of the glass substrate, d(x,y) represents the equivalent optical depth of the midpoint (x,y) of the preset Cartesian coordinate system, and λ represents the incident light wavelength corresponding to the DFB-LD laser. Solution module: Extracts the phase and unwraps the superimposed interference fringes to obtain depth information. ; A unified three-dimensional topography is obtained by weighted fusion of depth information calculated by three sets of DFB-LD lasers. Calculation and detection module: Based on a unified 3D topography, the pore regions are selected, and the following operations are performed within the pore regions: The micropore profile is extracted from the orifice plane, and an equivalent circle is fitted to obtain the actual pore diameter Ds and the average depth Hs of the bottom region of the micropore is obtained. Calculate the aperture deviation δ D =|Ds-D| / D, calculate the hole depth deviation δ H =|Hs-H| / H, where D represents the nominal aperture, H represents the nominal aperture depth, and the value is based on the aperture deviation δ. D Hole depth deviation δ H The glass through-hole process is classified.

[0021] It should be noted that, firstly, the production line adaptation module's function is to match the transmission rhythm of the TGV etching production line with the micro-via inspection cycle, avoiding incomplete inspection coverage due to excessively high production line speed. The inspection time for a single substrate is calculated as T = M × t, where t represents the inspection time for a single micro-via, i.e., the average time taken to complete the signal acquisition and depth calculation for a single micro-via. The product of M × t is the total time required to complete the inspection of all micro-vias on the entire substrate. Based on this, the inspection window W = v × T is calculated. This formula represents the physical length window reserved by the production line for the entire substrate inspection at the current transmission speed, i.e., the length range that the inspection system can cover as the substrate moves from entering the inspection area to leaving the inspection area.

[0022] If the window does not match the substrate length, the transmission speed v is adjusted based on the detection window duration W to obtain the correction speed xv. The transmission speed of the TGV etching production line is set to xv to ensure that the detection system can completely cover all the micropores of the entire substrate.

[0023] The infrared light source module first pre-sets a planar rectangular coordinate system on the glass substrate to provide a spatial reference for subsequent topography reconstruction. Simultaneously, it pre-sets the emission wavelengths of three DFB-LD lasers, utilizing multi-wavelength interference to avoid phase ambiguity issues in single-wavelength measurements. Then, the DFB-LD lasers are activated to illuminate the glass substrate, acquiring the reflected signal light from within the aperture. This reflected signal light is then superimposed and interfered with a pre-set reference light. The variation of the interference light intensity with the depth of the measured point is described by a formula. The interference light intensity consists of the interference terms of the intensity of the reference light and the signal light. The phase of the interference term is determined by the system's initial phase ω0 and the optical path difference term 4πnd / λ, which is related to the depth d. In the optical path difference term, 4πnd corresponds to the phase change of the light traveling back and forth within the glass. By analyzing the phase distribution of the interference light intensity, the depth information of the measured point can be deduced.

[0024] In addition, it should be noted that the near-infrared detection module can be replaced by non-destructive detection technologies such as white light interferometry (WLI), optical coherence tomography (OCT), or confocal measurement. All of these alternative technologies can achieve non-destructive detection of glass through holes and have the same detection accuracy and effect as the original near-infrared detection module.

[0025] The solution module is responsible for processing the interference fringes, converting light intensity information into three-dimensional depth data. First, it extracts and unwraps the phase of the superimposed interference fringes, establishing a correspondence between the phase change of the interference light intensity and the optical path difference. Using a scaling factor of 4πnd / λ, the phase difference is directly converted into geometric depth d(x, y), achieving the solution from phase information to physical depth. Then, the depth information calculated from three sets of DFB-LD lasers is weighted and fused to obtain a unified three-dimensional topography.

[0026] The calculation and detection module filters out the hole regions based on a unified 3D topography, identifying the spatial location and extent of micropores through topographic features and eliminating interference from non-hole regions. Then, the following operations are performed within the hole region: the micropore contour is extracted from the hole opening plane, an equivalent circle is fitted to obtain the actual hole diameter, and the average depth of the bottom region of the micropore is extracted. The equivalent circle fitting transforms the irregular micropore contour into a standard circle for easy comparison with the nominal hole diameter, while the average bottom depth reflects the actual etching depth of the micropore. Next, the hole diameter deviation and hole depth deviation are calculated. Finally, the glass through-hole process is graded based on the hole diameter and depth deviations, classifying the process into good, slightly defective, and unacceptable levels according to the deviation threshold. This clarifies the defect level.

[0027] In another preferred embodiment of the present invention, adjusting the transmission speed v based on the detection window W to obtain the corrected speed xv specifically includes: If the detection window W > L, no adjustment will be made, and the current transmission speed v will be maintained; If the detection window W≤L, let the correction speed xv=v(L / W+η), where η represents the preset redundancy coefficient, and η>0.

[0028] It is worth noting that when the detection window W≤L, it means that at the current speed v, the physical length required to complete the full board detection is insufficient to cover the substrate length L. Choosing to accelerate, by increasing the substrate moving speed, allows the entire substrate to pass through the detection area in a shorter time, enabling the detection system to complete the full hole scan more efficiently. This allows the production line cycle to be more compact and improves the overall production capacity without increasing the detection time. In addition, the redundancy coefficient is increased by speed on top of the minimum multiple. The purpose is to reserve a safety margin to cope with sudden delays in the production line or momentary blockages in the detection system, and to avoid efficiency losses due to insufficient speed.

[0029] The traditional approach is to slow down and expand the inspection window to prioritize the integrity of the inspection; this approach accelerates and compresses the inspection time to prioritize production line efficiency. Through preset multipliers and redundancy, the inspection system can still complete full-hole inspection at higher speeds.

[0030] In another preferred embodiment of the present invention, the calculation and detection module is based on the aperture deviation δ D Hole depth deviation δ H The classification of glass through-hole processes includes: If δ D and δ H If all values ​​are less than or equal to 2%, it is recorded as Grade A, indicating a good product. If δ D and δ H If any item exceeds 5%, it is classified as Grade C, and the product is considered unqualified. The rest are marked as Grade B, indicating that the product has minor defects, and staff are advised to take action.

[0031] Understandably, a quality evaluation standard for glass through-holes has been established. Grade A is qualified and can be directly transferred to the next process; Grade C is seriously unqualified and needs to be scrapped or reworked; Grade B is slightly defective and is used to warn of process drift on the production line, assisting staff in timely troubleshooting and adjustment, thus ensuring product quality stability.

[0032] In another preferred embodiment of the present invention, the emission wavelengths of the three DFB-LD lasers are preset to λ1, λ2, and λ3, respectively, and the constraints are satisfied. , where λ min , λ max These represent the preset minimum wavelength difference and maximum wavelength difference, respectively.

[0033] It is important to note that the complementarity of multi-wavelength interferometry can be used to solve the problems of phase ambiguity and depth calculation accuracy in single-wavelength measurements. By constraining the wavelength difference between two adjacent laser sets to between a preset minimum and maximum value, we can prevent the interference fringes of different wavelengths from being too close due to a small wavelength difference, making it difficult to distinguish phase information and thus preventing ambiguity during phase calculation. On the other hand, we can prevent the contrast of the interference signal from decreasing and the light intensity from being too attenuated due to a large wavelength difference, ensuring that the measurement signals at different wavelengths have a sufficient signal-to-noise ratio.

[0034] This wavelength constraint design enables the three sets of lasers to form a gradient multi-wavelength measurement system. The interference fringe periods of different wavelengths differ, and the phase problem of single-wavelength measurement can be effectively eliminated through combined calculation, expanding the measurable depth range. At the same time, the wavelength difference after constraint allows the measurement results of each wavelength to complement each other during phase calculation, improving the resolution and anti-interference ability of three-dimensional topography reconstruction, and providing a data basis for subsequent calculation of the deviation of glass through-hole diameter and depth.

[0035] In another preferred embodiment of the present invention, the weighted fusion of depth information calculated by three sets of DFB-LD lasers to obtain a unified three-dimensional topography specifically includes: The depth information calculated by the three DFB-LD lasers is weighted and fused to obtain a unified three-dimensional topography: d 3D (x, y) = w1×d1(x, y) + w2×d2(x, y) + w3×d3(x, y), where d1(x, y), d2(x, y), and d3(x, y) represent the depth information obtained by phase extraction and unwrapping of the first, second, and third groups of DFB-LD lasers, respectively, and w1, w2, and w3 represent the preset weight coefficients, and w1+w2+w3=1.

[0036] It should be noted that single sets of data are susceptible to phase noise and the measurement blind zone of the wavelength itself, and it is difficult to guarantee the stability of the entire area when used alone. By performing linear weighted fusion with preset weight coefficients, the contribution of different wavelengths can be allocated according to their performance in specific measurement scenarios. For example, wavelengths with stronger anti-interference capabilities can be given higher weights, the influence of noisier data can be weakened, and the weight sum can be kept to 1 to maintain the consistency of the physical meaning and dimensions of the data.

[0037] This fusion method not only preserves the effective information of each set of measurements, but also suppresses the error fluctuation of individual data through weight allocation. It effectively reduces the measurement deviation caused by factors such as phase solution ambiguity and environmental vibration, improves the overall resolution and reliability of the three-dimensional morphology, and provides basic data for subsequent micropore contour extraction and pore diameter and depth deviation calculation.

[0038] In another preferred embodiment of the present invention, the angle of the DFB-LD laser is adjusted to ensure that the incident light irradiates the glass substrate perpendicularly.

[0039] It is understandable that by adjusting the installation angle of the DFB-LD laser, the incident light is ensured to illuminate the surface of the glass substrate in a vertical direction, avoiding optical path difference shift caused by oblique incidence, ensuring that the interference phase of the reflected signal light and the reference light in the hole corresponds to the actual depth of the measured point, and improving the accuracy of depth calculation.

[0040] In another preferred embodiment of the present invention, the pore region is selected based on a unified three-dimensional morphology, including: d on the glass substrate 3D The region (x, y) > T is denoted as the hole region, where T represents the preset depth threshold.

[0041] It is worth noting that areas exceeding a preset threshold depth are marked as hole regions. By utilizing the difference in depth between the micropores and the surrounding flat substrate, non-hole substrate surface areas are excluded, avoiding interference from irrelevant topography. The depth threshold-based screening method is logically simple and can meet the real-time requirements of high-speed production line inspection while ensuring identification accuracy.

[0042] Meanwhile, by setting a reasonable threshold, the identification needs of micropores of different specifications can be taken into account. This avoids missing shallow pores due to excessively high thresholds, and also prevents surface scratches and minor defects from being misjudged as pore areas due to excessively low thresholds. This ensures the accuracy and stability of pore area positioning and provides a reliable area boundary for subsequent pore diameter fitting and pore depth calculation.

[0043] In a preferred embodiment, the percentage of Grade A glass substrates in this batch is calculated. When the percentage of Grade A glass substrates in Y consecutive batches is less than 90%, the staff is prompted to inspect the production equipment. Y represents the preset batch for judgment.

[0044] It is worth noting that the percentage of Grade A glass substrates in this batch reflects the yield rate of the current batch. When the percentage of Grade A in a continuously preset batch falls below 90%, an equipment maintenance prompt is triggered to avoid misjudgments caused by fluctuations in a single batch. By identifying process drift trends across multiple batches, early warnings are issued for equipment wear and parameter deviations, preventing the continuous production of defective products. This ensures the stability of product quality and transforms passive rework into proactive prevention, reducing production costs and scrap rates.

[0045] In a preferred embodiment, the method further includes a closed-loop control module, which acquires the grading information of the glass through-hole process in this batch, where L represents a preset sampling batch, and calculates the yield rate, which is the ratio of the number of grade A glass substrates in this batch to the total number of products. An LSTM-based AI prediction model is used, with the input dataset including the aperture deviation δ of this batch. D The mean value and hole depth deviation δ H The average value, yield, etching temperature, etching solution flow rate, etching solution concentration and processing time are used to output the ideal yield for the next batch. If the ideal yield is greater than or equal to 95%, the current process parameters are maintained and continuous monitoring is performed. If the ideal yield rate is less than 95%, then set the adjustment range for each parameter. The parameter adjustment range includes: Etching temperature adjustment range [W0, 65℃], in 0.5℃ increments; Etching solution flow rate adjustment range [W1, 50L / h], step size 1L / h; Etching solution concentration adjustment range [W2, 25%], step size 0.5%; Processing time adjustment range [W3, 60min], step size 1min, min represents minutes; Where W0, W1, W2, and W3 represent the etching temperature, etching solution flow rate, etching solution concentration, and processing time for this batch, respectively. After obtaining and adjusting the parameters, input them into the AI ​​prediction model, record the output as the actual yield rate, and repeat the iteration until the actual yield rate is greater than or equal to 95%.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An AI-based glass through-hole process defect detection system based on near-infrared imaging, characterized in that, Includes the following steps: Production line adaptation module: acquires the transmission parameters of the TGV etching production line, including the glass substrate transmission speed v, the glass substrate length L, and the total number of TGV holes M on the substrate; Calculate the detection time of a single substrate T=M×t, where t represents the detection time of a single hole. Calculate the detection window W=v×T. Based on the detection window duration W, adjust the transmission speed v to obtain the corrected speed xv. Set the transmission speed of the TGV etching production line to xv. Infrared light source module: A Cartesian coordinate system is pre-set on the glass substrate, and three sets of DFB-LD laser emission wavelengths are pre-set. The DFB-LD lasers are activated to irradiate the glass substrate and acquire the reflected signal light from the aperture. The reflected signal light from the aperture is superimposed and interfered with a pre-set reference light to obtain the output light intensity distribution relationship. , among which, I r I represents the intensity of the reference light. s ω represents the intensity of the reflected signal light inside the aperture, ω0 represents the initial phase of the superimposed interference system, n represents the glass refractive index of the glass substrate, d(x,y) represents the equivalent optical depth of the midpoint (x,y) of the preset Cartesian coordinate system, and λ represents the incident light wavelength corresponding to the DFB-LD laser. Solution module: Extracts the phase and unwraps the superimposed interference fringes to obtain depth information. ; A unified three-dimensional topography is obtained by weighted fusion of depth information calculated by three sets of DFB-LD lasers. Calculation and detection module: Based on a unified 3D topography, the pore regions are selected, and the following operations are performed within the pore regions: The micropore profile is extracted from the orifice plane, and an equivalent circle is fitted to obtain the actual pore diameter Ds and the average depth Hs of the bottom region of the micropore is obtained. Calculate the aperture deviation δ D =|Ds-D| / D, calculate the hole depth deviation δ H =|Hs-H| / H, where D represents the nominal aperture, H represents the nominal aperture depth, and the value is based on the aperture deviation δ. D Hole depth deviation δ H The glass through-hole process is classified.

2. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, In the aforementioned production line adaptation module, the adjustment of the transmission speed v based on the detection window W to obtain the corrected speed xv specifically includes: If the detection window W > L, no adjustment will be made, and the current transmission speed v will be maintained; If the detection window W≤L, let the correction speed xv=v(L / W+η), where η represents the preset redundancy coefficient, and η>0.

3. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, In the aforementioned calculation and detection module, based on the aperture deviation δ D Hole depth deviation δ H The classification of glass through-hole processes includes: If δ D and δ H If all values ​​are less than or equal to 2%, it is recorded as Grade A, indicating a good product. If δ D and δ H If any item exceeds 5%, it is classified as Grade C, and the product is considered unqualified. The rest are marked as Grade B, indicating that the product has minor defects, and staff are advised to take action.

4. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, In the aforementioned solution module, three sets of DFB-LD lasers are pre-set to emit wavelengths of λ1, λ2, and λ3, respectively, and the constraints are satisfied. , where λ min , λ max These represent the preset minimum wavelength difference and maximum wavelength difference, respectively.

5. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, The solution module, which performs weighted fusion of depth information calculated by three sets of DFB-LD lasers to obtain a unified three-dimensional topography, specifically includes: The depth information calculated by the three DFB-LD lasers is weighted and fused to obtain a unified three-dimensional topography: d 3D (x, y) = w1×d1(x, y) + w2×d2(x, y) + w3×d3(x, y), where d1(x, y), d2(x, y), and d3(x, y) represent the depth information obtained by phase extraction and unwrapping of the first, second, and third groups of DFB-LD lasers, respectively, and w1, w2, and w3 represent the preset weight coefficients, and w1+w2+w3=1.

6. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, In the infrared light source module, the angle of the DFB-LD laser is adjusted to ensure that the incident light irradiates the glass substrate perpendicularly.

7. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, In the aforementioned calculation and detection module, the pore regions selected based on a unified three-dimensional morphology include: d on the glass substrate 3D The region (x, y) > T is denoted as the hole region, where T represents the preset depth threshold.

8. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 3, characterized in that, Calculate the percentage of Grade A glass substrates in this batch. When the percentage of Grade A glass substrates in Y consecutive batches is less than 90%, prompt the staff to inspect the production equipment. Y represents the preset judgment batch.

9. The AI ​​detection system for glass through-hole process defects based on near-infrared imaging according to claim 1, characterized in that, It also includes a closed-loop control module, which acquires the grading status of the glass through-hole process in this batch, where L represents the preset sampling batch, and calculates the yield rate, which is the ratio of the number of grade A glass substrates in this batch to the total number of products. An LSTM-based AI prediction model is used, with the input dataset including the aperture deviation δ of this batch. D The mean value and hole depth deviation δ H The average value, yield, etching temperature, etching solution flow rate, etching solution concentration and processing time are used to output the ideal yield for the next batch. If the ideal yield is greater than or equal to 95%, the current process parameters are maintained and continuous monitoring is performed. If the ideal yield rate is less than 95%, then set the adjustment range for each parameter. The parameter adjustment range includes: Etching temperature adjustment range [W0, 65℃], in 0.5℃ increments; Etching solution flow rate adjustment range [W1, 50L / h], step size 1L / h; Etching solution concentration adjustment range [W2, 25%], step size 0.5%; Processing time adjustment range [W3, 60min], step size 1min, min represents minutes; Where W0, W1, W2, and W3 represent the etching temperature, etching solution flow rate, etching solution concentration, and processing time for this batch, respectively. After obtaining and adjusting the parameters, input them into the AI ​​prediction model, record the output as the actual yield rate, and repeat the iteration until the actual yield rate is greater than or equal to 95%.

Images