Collaborative decision method for tgv process automatic optical detection, computing device and storage medium

Abstract

The application discloses a kind of collaborative decision-making methods for TGV process automatic optical detection, computing device and storage medium. The method is for glass via TGV whole process, constructs unified data platform, accesses white sheet detection, laser induction, etching, conductive slurry filling, line detection, electrical measurement and the like data, and constructs unified index based on the unique identification of glass substrate and global coordinate system to realize cross-device, cross-station data association. Further, through CAD understanding intelligent agent, coordinate registration intelligent agent, Recipe generation intelligent agent, defect attribution intelligent agent, model iteration intelligent agent and disposal decision-making intelligent agent collaborative work, realize detection Recipe candidate generation, re-judgment task distribution, defect attribution chain construction, model retraining task generation and process disposal suggestion output. The scheme can reduce the dependence of Recipe debugging on experience, improve model iteration efficiency and process disposal consistency.

Description

Technical Field

[0001] This application relates to the field of TGV product measurement technology, and more specifically, to a collaborative decision-making method, computing device, and storage medium for automated optical inspection of TGV processes. Background Technology

[0002] TGV (Through Glass Vias) is a core interconnect structure for 2.5D / 3D packaging, high-performance computing chips, and RF devices. Its manufacturing process is long, typically involving multiple stages such as wafer glass feeding, laser induction, etching, conductive paste filling or metallization, circuit fabrication, and subsequent electrical testing. Defects generated at each stage exhibit significant spatiotemporal continuity. For example, bubbles, foreign objects, scratches, or cracks in the wafer stage may manifest as discontinuities after laser induction, as blind vias or rough hole walls after etching, and as voids or abnormal resistance after filling.

[0003] Traditional AOI equipment primarily performs single-station defect detection. AI model iterations typically rely on manual image selection, annotation, and offline training, lacking the ability to correlate defects across stations, provide feedback from downstream processes, and offer in-process handling suggestions. Even if a particular station has a high detection rate, if it cannot be linked to CAD design, coordinate errors, process parameters, and downstream electrical testing results, it remains difficult to determine whether a defect is fatal, and it is also difficult to formulate actionable recipe adjustments or in-process handling.

[0004] Therefore, a collaborative decision-making method for automated optical inspection in TGV processes is needed to upgrade AOI from a single-point inspection device to a process decision-making platform with cross-station reasoning and engineering closed-loop capabilities. Summary of the Invention

[0005] The main purpose of this application is to provide a collaborative decision-making method for automated optical inspection of TGV process. First, it establishes a unified index by connecting multi-source process data. Then, it uses a specialized intelligent agent (a program or entity that autonomously executes tasks) to complete design analysis, coordinate unification, recipe generation, defect attribution, model iteration, and disposal decision-making, ultimately achieving full-process control and closed-loop iteration.

[0006] To achieve the above objectives, the first aspect of this application proposes a collaborative decision-making method for automated optical inspection in TGV processes, comprising: Acquire automatic optical inspection data from each station in the TGV process, connect the automatic optical inspection data from each station to a unified data platform, and establish a global correlation relationship for the acquired data based on the multi-dimensional unique identifier of the glass substrate. The design file of the glass substrate is obtained, the CAD parsing agent is called to perform primitive parsing on the design file, the structured primitive set is extracted, and the structured primitive is rendered as a design prior layer and written into the unified data platform. The coordinate registration agent is invoked to map the local coordinates of different devices to the global coordinates of the glass substrate, and the coordinate confidence and coordinate error are calculated. The coordinate confidence and coordinate error are then written into the unified data platform. The design prior layer, equipment capability parameters, defect risks, and historical detection results are obtained from the unified data platform, and the Recipe generation agent is invoked to generate detection Recipe candidates and their scores. Defect detection is performed according to the candidate Recipe, and the defect data is written to the unified data platform in real time. The defect attribution agent is called to generate the defect attribution chain and the process risk score for each defect attribution chain. Based on the process risk score, the detection recipe candidates, the coordinate confidence, and the process threshold in the rule base, the disposal decision-making agent is invoked to output disposal suggestions; Based on the difficult sample generation model iterative update task in the detection process, the difficult sample includes false detection samples, missed detection samples, low confidence samples, inconsistent manual re-judgment, and failed samples in the later stage.

[0007] In one embodiment of this disclosure, the step of acquiring automated optical inspection data from each station of the TGV process, connecting the automated optical inspection data from each station to a unified data platform, and establishing a global association relationship for the acquired data based on the multi-dimensional unique identifier of the glass substrate includes: A unique identifier is generated for each glass substrate and bound to the production batch ID, sub-board ID, design version number, inspection / process parameter version number, and TGV process station identifier. The TGV process station includes blank inspection, laser induction, etching, conductive paste filling, circuit inspection, electrical testing equipment, and manufacturing execution system. All raw images, defect information, process parameters, equipment status, post-processing results, and manual review results generated at all workstations are written into a unified data platform to establish a data index. The data index includes the unique identifier of the glass substrate, production batch identifier, raw material substrate identifier, workstation identifier, equipment identifier, camera serial number, recipe version number, AI model version, design file version, inspection timestamp, global coordinates of the glass substrate, local coordinates of the equipment, defect type, defect severity level, inspection confidence level, manual review results, and post-processing results.

[0008] In one embodiment of this disclosure, the process includes obtaining a design file for a glass substrate, calling a CAD parsing agent to perform primitive parsing on the design file, extracting a set of structured primitives, and rendering the structured primitives as a design prior layer and writing it into the unified data platform. Obtain design files or custom primitive tables in any of the following design formats: Gerber, DXF, GDS, and ODB++. Identify and extract structured primitives, including holes, lines, pads, positioning points, restricted areas, tolerance zones, global positioning feature points, electrical networks, and layout layers. Record the design coordinates, dimensions, tolerances, layer, topology, and allowed detection strategies for each primitive. The structured primitives are rendered into a machine-readable design prior layer, which includes at least one of via mask, line mask, pad area mask, restricted area mask, edge distance map, tolerance band map, and expected topology map.

[0009] In one embodiment of this disclosure, a coordinate registration agent is invoked to map the local coordinates of different devices to the global coordinates of the glass substrate, and coordinate confidence and coordinate error are calculated. The coordinate confidence and coordinate error are then written into the unified data platform, including: Based on Mark point detection results, device calibration matrix, camera distortion parameters, platform motion error, stitching residual, and differences in coordinate formats of third-party devices, the transformation relationship from local coordinates to global coordinates is obtained through Mark point fitting: G = T(C) + W(C), where G is the global coordinate of the glass substrate, C is the local coordinate of the device, T is the rigid, similar or affine transformation, and W(C) is the compensation term for local nonlinear deformation caused by jigsaw puzzle, platform movement, glass expansion and contraction or lens distortion. Calculate the total coordinate error after superimposing all error sources, and generate a safety margin Margin=k·σ_total based on the coordinate error, where k is a margin coefficient greater than 1 and σ_total is the total error. When σ_total or Margin exceeds the preset process allowable threshold, trigger recalibration, reshooting, or manual verification. Store the global coordinates G, coordinate confidence = 1 / (1 + σ_total), safety margin, and calibration status flag into a unified data platform.

[0010] In one embodiment of this disclosure, the Recipe candidate includes at least the detection region ROI, light source channel, exposure time, camera gain, image acquisition magnification, re-judgment trigger threshold, traditional edge detection parameters, AI model version, CAD tolerance band, defect category mapping rules, result output format, and re-judgment trigger threshold; the Recipe score is determined based on at least two of the following: detection rate, false detection rate, missed detection risk, cycle time impact, equipment operation stability, the effect of historical similar Recipes, and manual confirmation cost.

[0011] In one embodiment of this disclosure, defect detection is performed according to the detected Recipe candidate, defect data is written to the unified data platform in real time, and a defect attribution agent is invoked to generate defect attribution chains and process risk scores for each defect attribution chain, including: When the detection Recipe generated by the Recipe needs to be sent to the production equipment, the following checks are performed: whether the CAD version is consistent with the design version of the current production product, whether the equipment model is compatible with the Recipe, whether the workstation number corresponds to the Recipe, whether the product model matches, whether the coordinate system version is consistent with the current equipment coordinate system, whether the AI ​​model version has been pre-installed on the equipment, and whether manual approval has been completed. Once all these conditions are met, an executable Recipe file is generated and sent to the equipment. After the Recipe is executed on the device, the detection results, manual review results, post-test / packaging results, and engineer's modification records of the Recipe are written back to the unified data platform. The original image, defect screenshot, CAD version, Recipe version, model version, coordinate system version, manual review results, and post-test results corresponding to each detection result are also recorded. Defects on the same glass substrate, such as white film defects, laser-induced anomalies, blind holes or rough hole walls after etching, voids filled by conductive paste, circuit expansion or contraction or circuit offset, and electrical test anomalies, are spatiotemporally correlated based on the unique identifier of the glass substrate and global coordinates using a unified index, forming one or more defect attribution chains; The process failure probability P(Fail|X,Z) is calculated using a conditional probability model, Bayesian network, causal graph, gradient boosting tree, graph neural network, or rule-model hybrid method, where X includes defect image features, geometric features, coordinate features, and CAD interference features, and Z includes process parameters, equipment parameters, and recipe version. The process risk score for each defect attribution chain is also output.

[0012] In one embodiment of this disclosure, based on the process risk score, the detection recipe candidates, the coordinate confidence score, and the process threshold in the rule base, a disposal decision-making agent is invoked to output disposal suggestions, including: The system employs a combination of rule gating and model scoring to output handling recommendations. The rule gating is used to constrain high-risk actions involving production stoppage, scrapping, recipe distribution, and model deployment. The model scoring is used to generate low-risk re-judgment, continued processing, sampling re-inspection, or recipe candidate ranking. The handling recommendations carry decision-making criteria, which include at least the triggering defect, defect coordinates, related CAD elements, associated workstations, risk score, historical similar cases, current recipe version, candidate recipe differences, and manual confirmation status.

[0013] In one embodiment of this disclosure, a model iteration agent is invoked to generate a model iteration update task based on difficult examples during the detection process. These difficult examples include false positives, false negatives, low-confidence samples, inconsistencies in manual review, and subsequent failure samples. Collect false detection samples, missed detection samples, low confidence samples, inconsistent cross-model judgment samples, inconsistent manual re-judgment and model judgment samples, failed back-end electrical test samples, and duplicate samples of the same type from the entire link. Deduplicate the samples based on image perceptual hash, defect location, defect category, image feature vector and CAD primitive number. The representative samples after deduplication are entered into the difficult sample pool. Generate annotation tasks, divide the training set / validation set, and after completing model training, calculate performance metrics including precision, recall, F1 score, critical defect recall, false positive rate, false negative rate, online beat impact, and cross-batch stability on the validation set, recent online samples, and critical defect samples. When the model reaches the deployment threshold, it is deployed to a specified product, device, or batch in a gray-scale manner. If the false negative rate, false positive rate, or cycle time does not meet the requirements during the gray-scale deployment phase, it is rolled back to the previous stable version. When any performance indicator fails to meet the deployment threshold, release is prohibited and supplementary sample suggestions are not generated.

[0014] A second aspect of this disclosure provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method provided in the first aspect.

[0015] A third aspect of this disclosure provides a computing device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method provided in the first aspect.

[0016] As can be seen from the above scheme, the collaborative decision-making method for automated optical inspection of TGV process provided in this disclosure integrates the inspection data of all processes in the TGV process, including white film inspection, laser induction, etching, conductive paste filling, circuit inspection, and electrical testing, as well as the management data of the manufacturing execution system and quality management system, into a unified platform. It achieves accurate data binding across equipment and workstations through unique identifiers for single glass substrates, production batch identifiers, process stage identifiers, product mark points, and a globally unified coordinate system. This ensures that the inspection data and process parameters of the same glass substrate in different processes can correspond one-to-one, and also solves the problems of inconsistent inspection coordinate systems and mismatched defect locations of different equipment, providing a complete and accurate data foundation for subsequent analysis.

[0017] By analyzing the design documents of TGV products, the normal structure, defect-sensitive areas, and key inspection points of the product are clearly identified, providing a design-based benchmark for subsequent inspection parameter generation and defect judgment. Images and inspection results collected from different workstations and equipment are uniformly mapped to a global coordinate system, and spatial alignment of cross-process inspection data is achieved through Mark point matching, ensuring that the inspection data of the same defect in different processes can be accurately correlated. Combining the current product model, process stage, type of defect to be inspected, and a successful inspection parameter library of similar products in the past, inspection recipe candidates adapted to the current scenario are automatically generated and can be directly sent to AOI equipment for execution, significantly reducing the dependence of recipe development on engineers' experience and shortening the debugging cycle of new products and new production lines. When AOI detects a defect, it retrieves the inspection data and process parameters of the corresponding glass substrate throughout the entire process. By matching the defect features with the abnormal features of each process, a complete defect attribution chain is generated, solving the problems of slow root cause location and difficulty in defining responsibility across processes. The new defect samples after attribution are organized into a labeled dataset to evaluate the missed detection and false positive rates of the existing detection model. If the model performance is not up to standard, retraining is automatically triggered. The updated detection model can be automatically sent to the AOI equipment without manual intervention to calibrate samples and train the model, greatly improving the iteration efficiency of the detection model. Combining the severity level of the defect, the attribution results, the overall quality status of the current batch, and the quality judgment rules, standardized handling suggestions are generated to avoid handling deviations caused by inconsistent manual judgment standards.

[0018] This disclosure not only solves the problems of TGV process recipe debugging relying on experience, difficulty in defect root cause location, slow model iteration, and inconsistent handling standards, but also opens up a closed loop from inspection data to process optimization, which can effectively improve the yield and production consistency of TGV process. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 A flowchart illustrating a collaborative decision-making method for automated optical inspection in TGV processes, provided in an embodiment of this disclosure; Figure 2 A schematic diagram of a collaborative decision-making system architecture for automated optical inspection in TGV processes, provided in an embodiment of this disclosure; Figure 3 This is a structural block diagram of a computing device provided in an embodiment of the present disclosure. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The TGV (Through Glass Via) process involves multiple stations, including blank wafer inspection, laser-induced drilling, etching, conductive paste filling, circuit fabrication, and electrical testing. Defect types cover a wide range, including microscopic hole shape defects, uneven filling, cracks, open / short circuits, and functional failures. Traditional AOI (Automated Optical Inspection) stores data independently for each station, making it impossible to correlate defects across stations. This makes it impossible to trace the root cause of defects or predict their impact on subsequent processes. Furthermore, debugging the inspection recipe (a set of inspection parameters including illumination, camera, algorithm thresholds, and inspection area) relies heavily on individual engineer experience, resulting in long debugging cycles and inconsistent solutions generated by different personnel. Finally, there is no unified standard for process handling, and model iterations cannot adapt to the real-time changes in defect distribution on the production line, leading to long iteration cycles.

[0023] This solution addresses the aforementioned issues by proposing a closed-loop processing method based on a unified data platform and involving multiple clearly defined intelligent agents (programs or entities that perceive the environment and autonomously execute tasks and make decisions). All agents receive input from the unified data platform, and their outputs are written back to the platform for use by other agents, forming a closed-loop iterative chain. This upgrades AOI equipment from a single defect detection tool to an integrated intelligent decision-making system encompassing detection, attribution, remediation, and iteration. By establishing bidirectional data flow and agent collaboration between Design Intent (CAD), Manufacturing Execution System (MES / equipment), AOI, and post-processing verification (electrical testing), a paradigm shift from passive detection to proactive prevention and precise remediation is achieved.

[0024] Figure 1 This is a flowchart illustrating the collaborative decision-making method for automated optical inspection in the TGV process provided in this application. (Refer to...) Figure 1 As shown, firstly, in step S102, the automatic optical inspection data of each station in the TGV process is acquired, and the automatic optical inspection data of each station is connected to a unified data platform. A global correlation relationship is established for the acquired data based on the multi-dimensional unique identifier of the glass substrate.

[0025] First, a unique identifier is generated for each glass substrate, and bound to the production batch ID, sub-board ID, design version number, inspection / process parameter version number, and TGV process station identifier. The TGV process station includes blank inspection, laser induction, etching, conductive paste filling, circuit inspection, electrical testing equipment, and manufacturing execution system.

[0026] The data content and function of each process are shown in the table below:

[0027] All raw images, defect information, process parameters, equipment status, downstream results, and manual review results generated at all workstations are written into a unified data platform to establish a data index. The data index includes the unique identifier of the glass substrate (PanelID), the production batch identifier (LotID), the raw material substrate identifier (SubstrateID), the workstation identifier (StageID), the equipment identifier (EquipmentID), the camera serial number (CameraID), the recipe version number (CameraID), the AI ​​model version (ModelVersion), the design file version (CADVersion), the inspection timestamp (InspectionTime), the global coordinates of the glass substrate (GlobalCoordinateX / Y), the local coordinates of the equipment (LocalCoordinateU / V), the defect type (DefectType), the defect severity level (DefectSeverity), the inspection confidence level (Confidence), the manual review result (ReviewResult), and the downstream result (DownstreamResult).

[0028] The unified data platform does not require all equipment to be from the same manufacturer. For third-party equipment, its inspection results can be accessed via CSV, JSON, XML, database, OPC UA, SECS / GEM, or API, and the coordinate units, coordinate origin, mark recognition method, image path, defect type, and confidence level must be recorded. For third-party equipment that cannot provide full data, its data confidence level is individually marked by the coordinate registration agent. By establishing cross-system and cross-workstation data binding indexes, any data can be quickly located to the corresponding physical object, solving the problem of cross-workstation data not being correlated.

[0029] To address the disconnect between design data and testing logic, in step S104, the design file of the glass substrate is obtained, the CAD parsing agent is invoked to perform primitive parsing on the design file, a set of structured primitives is extracted, and the structured primitives are rendered as a design prior layer and written into the unified data platform.

[0030] The CAD understanding agent's processing logic involves performing primitive-level semantic parsing on unstructured design documents, extracting all structural features and detection requirements, and outputting a machine-readable structured design primitive library. This avoids biases caused by human interpretation of design documents and directly provides design basis for subsequent detection logic, such as clarifying the rule that "defects within the tolerance zone will not trigger an alarm, while defects within the restricted zone will be directly judged as serious defects."

[0031] In one embodiment of this disclosure, a design file or a custom primitive table in any design format (Gerber, DXF, GDS, ODB++) is acquired. Structured primitives are identified and extracted. These structured primitives include primitive types such as Via (through-hole / blind via), Pad (pad), Line (line), RDL (rerouting layer), Mark (reference point), KeepOut (restricted area, i.e., zero-tolerance defect zone), ToleranceBand (tolerance band, i.e., the allowable deviation range for defects), Net (electrical network group), and Layer (process layer, such as glass front and back, metal layer). The design coordinates, dimensions, tolerances, layer, topology, and allowed detection strategies for each primitive are recorded. The structured primitives are rendered into a machine-readable design prior layer. This design prior layer includes at least one of the following: via mask, line mask, pad area mask, restricted area mask, edge distance map (calculating the distance between defects and lines / vias to determine if they affect electrical performance), tolerance band map, and expected topology map (ideal line / via shape on the design side, used to compare with the actual detection map to find defects). For example, Via objects are rendered as hole masks, Line objects as line masks, ToleranceBand objects as tolerance band maps, and KeepOut objects as restricted area masks. In this way, the Recipe-generated agent can accurately determine which areas should be prioritized for detection, which areas are only allowed weak alarms, and which areas are prohibited from automatic handling.

[0032] To address the issues of incompatibility of coordinates among multiple devices and the inability to locate defects across processes in the TGV process, in step S106, a coordinate registration agent is invoked to map the local coordinates of different devices to the global coordinates of the glass substrate, and the coordinate confidence level and coordinate error are calculated. The coordinate confidence level and coordinate error are then written into the unified data platform.

[0033] Specifically, based on the Mark point detection results, equipment calibration matrix, camera distortion parameters, platform motion error, stitching residual, and differences in coordinate formats of third-party equipment, the transformation relationship from local coordinates to global coordinates is obtained through Mark point fitting: G=T(C)+W(C), where G is the global coordinates of the glass substrate, C is the local coordinates of the equipment, T is the rigid / similar / affine transformation, which is responsible for the basic translation, rotation, scaling and other global coordinate transformations. The rigid transformation has no deformation, the similar transformation allows uniform scaling, and the affine transformation allows shearing, adapting to the distortion differences of different cameras. W is the local nonlinear deformation compensation term, which compensates for the non-uniform deformation of the glass substrate caused by thermal expansion and contraction, chemical etching, and mechanical extrusion during the manufacturing process, as well as the nonlinear deviations such as small vibrations and stitching errors of the equipment platform.

[0034] Calculate the total coordinate error after superimposing all error sources, and generate a safety margin based on the coordinate error: Margin=k σ_total, where k is a margin coefficient greater than 1, and σ_total is the total error. When σ_total or Margin exceeds the preset process allowable threshold, recalibration, re-shooting, or manual verification is triggered, and the global coordinates G, coordinate confidence = 1 / (1 + σ_total), error margin Margin, and calibration status flag are stored in the unified data platform.

[0035] If the number of Mark points is insufficient, the Mark residual exceeds the threshold, the splicing residual exceeds the threshold, or the third-party device lacks a coordinate origin definition, the coordinate confidence will decrease. The decision-making agent will not be allowed to automatically issue high-risk recipes, but will only be allowed to generate review or manual confirmation tasks.

[0036] In step S108, the design prior layer, equipment capability parameters, defect risks and historical detection results are obtained from the unified data platform, and the Recipe generation agent is called to generate detection Recipe candidates and their scores.

[0037] In one embodiment of this disclosure, the Recipe generating agent integrates the design prior layer output by the CAD understanding agent, the hardware capabilities of the current device (camera resolution, depth of field range, motion speed, light source type), the defect risks of the corresponding process stage (such as micro-cracks easily appearing in the laser station, and rough hole walls easily appearing in the etching station), and historical detection results (historical detection rate and false detection rate of the same product and the same equipment), and outputs detection Recipe candidates.

[0038] The generated detection recipe candidates include at least the detection area ROI (only for key areas such as hole periphery and dense line areas to improve cycle time), light source channel (side light / back light / ring light, etc., to adapt to the imaging needs of different defects), exposure time, camera gain, image acquisition magnification, re-judgment trigger threshold (automatic manual re-judgment when AI confidence is below the threshold), traditional algorithm parameters (edge ​​detection, morphological operation, etc.), AI model version, CAD tolerance band, defect category mapping rules, result output format, and re-judgment trigger threshold.

[0039] For each candidate recipe, a comprehensive score is calculated from multiple dimensions, including detection rate, false positive rate, missed detection risk, impact on production cycle time, equipment operational stability, the effectiveness of historically similar recipes, and manual review cost. The recipe with the highest score is selected as the default candidate, and multiple candidates can be generated for subsequent decision-making. For example, for each recipe candidate r, the system calculates a recipe score Score(r). The score can be composed of detection capability, false positive cost, missed detection risk, cycle time impact, historical stability, and manual verification cost. Score(r) = a1 Recall_key(r)-a2 FalseAlarm(r)-a3 CycleTime(r) + a4 Stability(r)-a5 ManualCost(r) Where Recall_key(r) represents the critical defect recall rate, FalseAlarm(r) represents the false alarm rate or the number of false alarms per unit area, CycleTime(r) represents the single-piece detection time, Stability(r) represents cross-batch stability, ManualCost(r) represents the manual verification cost, and a1 to a5 are weights.

[0040] In step S110, defect detection is performed according to the detected Recipe candidate, defect data is written to the unified data platform in real time, and the defect attribution agent is called to generate defect attribution chains and process risk scores for each defect attribution chain.

[0041] In one embodiment of this disclosure, when the detection Recipe output by the Recipe-generated intelligent agent needs to be sent to the production equipment, the system verifies whether the CAD version is consistent with the design version of the current production product, whether the equipment model is compatible with the Recipe, whether the workstation number corresponds to the Recipe, whether the product model matches, whether the coordinate system version is consistent with the current equipment coordinate system, whether the AI ​​model version has been pre-installed on the equipment, and whether manual approval has been completed. Once all conditions are met, an executable Recipe file is generated and sent to the equipment to avoid batch detection accidents caused by incorrect Recipe downloads.

[0042] After the detection recipe is executed on the device, the detection results, manual review results, post-production electrical testing / packaging results, and engineer's modification records of the recipe are written back to the unified data platform. The platform also records the original image, defect screenshot, CAD version, recipe version, model version, coordinate system version, manual review results, and post-production results corresponding to each detection result, supporting end-to-end traceability of any result. The written-back data can be used to update the recipe scoring library: if the false positive rate / false negative rate of a recipe exceeds the standard after execution, its score is automatically lowered, and its priority is reduced when generating subsequent candidates; to update the AI ​​model sample pool: manually confirmed defect samples and false negative samples of post-production failures are automatically added to the difficult example sample library for subsequent model iterations; and to update the defect attribution agent: if there is a deviation between the post-production failure data of a certain type of defect and the prediction results of the attribution model, the model weights are automatically adjusted to improve the accuracy of subsequent attributions.

[0043] Defects on the same glass substrate, including unfinished glass defects, laser-induced anomalies, blind vias or rough via walls after etching, voids filled with conductive paste, circuit expansion or misalignment, and electrical test anomalies, are spatiotemporally correlated based on the unique identifier and global coordinates of the glass substrate using a unified index, forming one or more defect attribution chains. For example, "scratches on the unfinished glass surface → microcracks appear at the scratches after laser induction → cracks propagate along the via walls after etching → voids appear at the cracks during paste filling → circuit misalignment after circuit fabrication → open circuit failure in electrical testing," clearly identifying the origin and propagation path of the defects.

[0044] The process failure probability, P(Fail|X,Z), is calculated using conditional probability models, Bayesian networks, causal graphs, gradient boosting trees, graph neural networks, or a rule-model hybrid approach. Here, X represents defect features (image features, geometric features, coordinate features, CAD interference features), and Z represents process / equipment parameters (laser power, etching temperature, slurry viscosity, equipment vibration, etc.) and the Recipe version. A process risk score is output for each defect attribution chain, providing a basis for subsequent handling decisions. For example, if a white wafer bubble highly overlaps with a subsequent blind via within the same CAD via risk area, and similar patterns in historical data correspond to a high failure probability, the failure risk score for that chain is increased.

[0045] In step S112, based on the process risk score, the detection recipe candidate, the coordinate confidence score, and the process threshold in the rule base, the disposal decision-making agent is invoked to output disposal suggestions.

[0046] This disclosure employs a hybrid logic combining rule gating and model scoring. Rule gating acts as a hard constraint, requiring pre-defined rules to be met before triggering high-risk operations such as production stoppages, scrapping, recipe distribution, and model deployment. For example, a production stoppage rule is triggered only if the detection rate of the same type of defect in the same batch at the same location exceeds 5% and the probability of subsequent failure exceeds 80%, thus preventing unnecessary production losses due to AI misjudgments. Model scoring is used for low-risk decisions, automatically generating ranking suggestions based on defect risk scores, subsequent failure probabilities, and historical similar cases for scenarios such as continued processing, manual review, sampling re-inspection, and recipe candidate ranking.

[0047] For low-risk defects, it is recommended to continue processing or conduct random sampling for re-inspection; for medium-risk defects, trigger a review camera for re-evaluation; for high-risk defects, it is recommended to stop the production line, rework, scrap, or prohibit the issuance of further recipes. All handling recommendations are accompanied by complete decision-making basis, including at least the triggering defect type, defect coordinates, associated CAD elements, associated workstation, process risk score, historical similar cases, current recipe version, differences in candidate adjusted recipes, and manual confirmation status.

[0048] In step S114, the model is iteratively updated based on the difficult sample generation task in the detection process. The difficult sample includes false detection samples, missed detection samples, low confidence samples, inconsistent manual re-judgment samples, and failed samples in the later stage.

[0049] It can collect false detection samples, missed detection samples, low confidence samples, inconsistent cross-model judgment samples, inconsistent manual re-judgment and model judgment samples, failed back-end electrical test samples, and duplicate samples of the same type. The samples are deduplicated based on image perceptual hash, defect location, defect category, image feature vector and CAD primitive number. The representative samples after deduplication are entered into the difficult sample pool.

[0050] Generate annotation tasks, divide the training set / validation set, lock the validation set throughout the process for model performance evaluation, and calculate performance metrics including precision, recall, F1 score (a comprehensive indicator for evaluating the quality of the classification model), critical defect recall, false positive rate, false negative rate, online beat impact, and cross-batch stability on the validation set, recent online samples, and critical defect samples.

[0051] When the model reaches the deployment threshold, it is deployed to designated products, devices, or batches in a gray-scale manner. If the false negative rate, false positive rate, or cycle time does not meet the requirements during the gray-scale deployment phase, it is rolled back to the previous stable version. If any performance indicator fails to meet the deployment threshold, release and generation of supplementary sample suggestions are prohibited. For example, if there are insufficient false negative samples for cracks, more than 100 microcrack samples of different morphologies need to be added.

[0052] This disclosure also provides a collaborative decision-making system for automated optical inspection in TGV processes. Figure 2 This is a schematic diagram of a collaborative decision-making system architecture for automated optical inspection in the TGV process. (Refer to...) Figure 2 As shown, the system architecture can be divided into a data access layer, a multi-agent collaboration layer, and an application output layer. The data access layer receives data from blank wafer inspection, laser induction, etching, conductive paste filling, circuit inspection, electrical testing equipment, and the manufacturing execution system. The CAD understanding agent, coordinate registration agent, recipe generation agent, defect attribution agent, model iteration agent, and handling decision agent collaborate to make decisions, outputting inspection recipes, review tasks, process handling, training tasks, version reports, and yield warnings. The input data, core processing logic, and output results of each agent are shown in the table below:

[0053] As shown in the table above, in the specific implementation, the CAD understanding agent is responsible for converting CAD drawings into machine-computable design priors; the coordinate registration agent is responsible for mapping the local coordinates of each workstation to the global glass coordinates; the recipe generation agent is responsible for outputting the detection window, optical parameters, threshold strategy, AI model version, and re-judgment trigger rules; the defect attribution agent is responsible for calculating the cross-workstation defect chain and failure probability; the model iteration agent is responsible for updating the difficult example sample pool and model version; and the disposal decision agent is responsible for converting risk scores and rule gating into engineering disposal suggestions.

[0054] The aforementioned intelligent agents do not operate independently, but rather achieve data interoperability through a unified data platform, forming a closed loop of continuous optimization. The forward process loop consists of data access → design analysis → coordinate alignment → recipe generation → defect detection → defect attribution → handling decision → model iteration → model update to improve detection performance, completing one round of full-process processing. The reverse feedback loop includes: the optimized model output by the model iteration agent is synchronously updated to the algorithm library of each AOI device, improving the accuracy of subsequent detection; the root cause information output by the defect attribution agent is fed back to the recipe generation agent, optimizing the recipe's detection area and parameters (for example, if defects caused by frequent Mark point offsets occur, the recipe will automatically add a Mark point recognition and verification step); the recipe adjustment requirements output by the handling decision agent trigger the recipe generation agent to regenerate candidate parameters adapted to the current process state. The manual optimization loop involves: all agent outputs supporting manual review and adjustment; the results of manual corrections are written back to the historical effect library, rule library, and defect knowledge graph, continuously improving the accuracy of the agent's judgment and adapting to the customized needs of different customers and products.

[0055] It should be noted that the intelligent agent in this embodiment refers to a software functional unit with clearly defined input, processing logic, output, and permission boundaries, and is not limited to a certain large language model.

[0056] The following describes the collaborative processing framework of this disclosure embodiment, using the TGV etching station AOI as an example: A batch of glass substrates with PanelID ETCH-20260707-001~100 entered the etching AOI station. Through a unified index, it was associated with the CAD version of this batch as V2.1, the corresponding equipment as ETCH-AOI-01, and the coordinate system version as V3.0. The CAD Understanding Agent parses CAD files (V2.1), extracts the expected hole positions, lines, restricted areas, and tolerance zones after etching, and renders them as a design prior layer. The coordinate registration agent calculates the global mapping relationship through Mark points. The total error is 2μm and the safety margin is 6μm, which meets the process threshold (10μm) and the coordinate mapping is effective. The Recipe generator combines the high-risk defects of "rough hole walls and side etching of lines" in the etching station, the equipment's resolution of 2μm, and historical inspection data to generate 3 Recipe candidates. The candidate with the highest score of 92 points was selected. The parameters are: ROI of 50μm+ around the hole and line area, ring side lighting, exposure of 20ms, gain of 15, 5x image acquisition, re-judgment threshold of 0.8, and AI model V4.2. The defect attribution agent correlates the defect data of the preceding white film and the laser station, and combines them with the current process parameters to predict that the root cause of the hole wall roughness defect is "the amplification of microcracks after etching due to excessive laser power" with a probability of 70% and a risk score of 75. After the test was performed, rough hole wall defects were detected in all three glass pieces at the same coordinate position, with a confidence level of 0.85. The probability of electrical test failure for the same type of defect in the subsequent section was 60%. The intelligent agent for handling decisions outputs the following suggestions: First, sample and re-examine these three pieces of glass. If defects are confirmed, adjust the power parameters of the laser station and optimize the exposure parameters of the perforated area of ​​Recipe to improve the detection rate of microcracks. The review and detection results are written back to the data platform: the Recipe score is updated, three defect samples are added to the difficult sample pool, and the weights of the defect attribution agent are adjusted to complete the closed loop of this iteration. Records the entire process, including version, lineage, and approval records, supporting subsequent traceability.

[0057] In summary, the embodiments disclosed herein integrate all inspection data from the entire TGV process, including white film inspection, laser induction, etching, conductive paste filling, circuit inspection, and electrical testing, as well as management data from the manufacturing execution system and quality management system, into a unified platform. Furthermore, through unique identifiers for each glass substrate, production batch identifiers, process stage identifiers, product mark points, and a globally unified coordinate system, precise data binding across devices and workstations is achieved. This ensures that the inspection data and process parameters of the same glass substrate correspond one-to-one across different processes, and also solves the problems of inconsistent inspection coordinate systems and mismatched defect locations across different devices, providing a complete and accurate data foundation for subsequent analysis.

[0058] By analyzing the design documents of TGV products, the normal structure, defect-sensitive areas, and key inspection points of the product are clearly identified, providing a design-based benchmark for subsequent inspection parameter generation and defect judgment. Images and inspection results collected from different workstations and equipment are uniformly mapped to a global coordinate system, and spatial alignment of cross-process inspection data is achieved through Mark point matching, ensuring that the inspection data of the same defect in different processes can be accurately correlated. Combining the current product model, process stage, type of defect to be inspected, and a successful inspection parameter library of similar products in the past, inspection recipe candidates adapted to the current scenario are automatically generated and can be directly sent to AOI equipment for execution, significantly reducing the dependence of recipe development on engineers' experience and shortening the debugging cycle of new products and new production lines. When AOI detects a defect, it retrieves the inspection data and process parameters of the corresponding glass substrate throughout the entire process. By matching the defect features with the abnormal features of each process, a complete defect attribution chain is generated, solving the problems of slow root cause location and difficulty in defining responsibility across processes. The new defect samples after attribution are organized into a labeled dataset to evaluate the missed detection and false positive rates of the existing detection model. If the model performance is not up to standard, retraining is automatically triggered. The updated detection model can be automatically sent to the AOI equipment without manual intervention to calibrate samples and train the model, greatly improving the iteration efficiency of the detection model. Combining the severity level of the defect, the attribution results, the overall quality status of the current batch, and the quality judgment rules, standardized handling suggestions are generated to avoid handling deviations caused by inconsistent manual judgment standards.

[0059] This disclosure not only solves the problems of TGV process recipe debugging relying on experience, difficulty in defect root cause location, slow model iteration, and inconsistent handling standards, but also opens up a closed loop from inspection data to process optimization, which can effectively improve the yield and production consistency of TGV process.

[0060] This disclosure also provides a computing device, such as... Figure 3 As shown, the computing device includes one or more processors 301 and a memory 302. Figure 3 Take processor 301 as an example.

[0061] The computing device may also include an input device 303 and an output device 304.

[0062] The processor 301, memory 302, input device 303, and output device 304 can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0063] Processor 301 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips. The general-purpose processor can be a microprocessor or any conventional processor.

[0064] The memory 302, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 301 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 302, thereby implementing the methods in the above-described method embodiments.

[0065] Memory 302 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the use of the processing device operated by the server. Furthermore, memory 302 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 302 may optionally include memory remotely located relative to processor 301, and these remote memories can be connected to a network connection device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0066] One or more modules are stored in memory 302 and, when executed by one or more processors 301, perform the methods shown in the above embodiments.

[0067] Those skilled in the art will understand that all or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes as described in the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory (FM), hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0068] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0069] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0070] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A collaborative decision-making method for automated optical inspection in TGV processes, characterized in that, include: Acquire automatic optical inspection data from each station in the TGV process, connect the automatic optical inspection data from each station to a unified data platform, and establish a global correlation relationship for the acquired data based on the multi-dimensional unique identifier of the glass substrate. The design file of the glass substrate is obtained, the CAD parsing agent is called to perform primitive parsing on the design file, the structured primitive set is extracted, and the structured primitive is rendered as a design prior layer and written into the unified data platform. The coordinate registration agent is invoked to map the local coordinates of different devices to the global coordinates of the glass substrate, and the coordinate confidence and coordinate error are calculated. The coordinate confidence and coordinate error are then written into the unified data platform. The design prior layer, equipment capability parameters, defect risks, and historical detection results are obtained from the unified data platform, and the Recipe generation agent is invoked to generate detection Recipe candidates and their scores. Defect detection is performed according to the candidate Recipe, and the defect data is written to the unified data platform in real time. The defect attribution agent is called to generate the defect attribution chain and the process risk score for each defect attribution chain. Based on the process risk score, the detection recipe candidates, the coordinate confidence, and the process threshold in the rule base, the disposal decision-making agent is invoked to output disposal suggestions; as well as The model iterative agent is invoked to generate model iterative update tasks based on difficult examples in the detection process. The difficult examples include false detection samples, missed detection samples, low confidence samples, inconsistent manual re-judgment samples, and failed samples in the later stage.

2. The collaborative decision-making method for automated optical inspection in TGV process according to claim 1, characterized in that, The process involves acquiring automated optical inspection data from each station in the TGV process, connecting this data to a unified data platform, and establishing a global correlation between the acquired data based on the multi-dimensional unique identifier of the glass substrate. This includes: A unique identifier is generated for each glass substrate and bound to the production batch ID, sub-board ID, design version number, inspection / process parameter version number, and TGV process station identifier. The TGV process station includes blank inspection, laser induction, etching, conductive paste filling, circuit inspection, electrical testing equipment, and manufacturing execution system. All raw images, defect information, process parameters, equipment status, post-processing results, and manual review results generated at all workstations are written into a unified data platform to establish a data index. The data index includes the unique identifier of the glass substrate, production batch identifier, raw material substrate identifier, workstation identifier, equipment identifier, camera serial number, recipe version number, AI model version, design file version, inspection timestamp, global coordinates of the glass substrate, local coordinates of the equipment, defect type, defect severity level, inspection confidence level, manual review results, and post-processing results.

3. The collaborative decision-making method for automated optical inspection in TGV processes according to claim 1, characterized in that, The process of obtaining the design file for the glass substrate, calling a CAD parsing agent to perform primitive parsing on the design file, extracting a structured primitive set, and rendering the structured primitives as a design prior layer and writing it into the unified data platform includes: Obtain design files or custom primitive tables in any of the following design formats: Gerber, DXF, GDS, and ODB++. Identify and extract structured primitives, which include holes, lines, pads, positioning points, restricted areas, tolerance zones, global positioning feature points, electrical networks, and layout layers. Record the design coordinates, dimensions, tolerances, layer, topology, and allowed detection strategies for each primitive. The structured primitives are rendered into a machine-readable design prior layer, which includes at least one of via mask, line mask, pad area mask, restricted area mask, edge distance map, tolerance band map, and expected topology map.

4. The collaborative decision-making method for automated optical inspection in TGV process according to claim 1, characterized in that, The process of calling the coordinate registration agent to map the local coordinates of different devices to the global coordinates of the glass substrate, calculating the coordinate confidence and coordinate error, and writing the coordinate confidence and coordinate error into the unified data platform includes: Based on Mark point detection results, device calibration matrix, camera distortion parameters, platform motion error, stitching residual, and differences in coordinate formats of third-party devices, the transformation relationship from local coordinates to global coordinates is obtained through Mark point fitting: G = T(C) + W(C), where G is the global coordinate of the glass substrate, C is the local coordinate of the device, T is the rigid, similar or affine transformation, and W(C) is the compensation term for local nonlinear deformation caused by jigsaw puzzle, platform movement, glass expansion and contraction or lens distortion. Calculate the total coordinate error after superimposing all error sources, and generate a safety margin Margin=k·σ_total based on the coordinate error, where k is a margin coefficient greater than 1 and σ_total is the total error. When σ_total or Margin exceeds the preset process allowable threshold, trigger recalibration, reshooting, or manual verification. Global coordinates, coordinate confidence, safety margin, and calibration status flags are stored in a unified data platform.

5. The collaborative decision-making method for automated optical inspection in TGV processes according to claim 1, characterized in that, The candidate recipe includes at least the detection region ROI, light source channel, exposure time, camera gain, image acquisition magnification, re-judgment trigger threshold, traditional edge detection parameters, AI model version, CAD tolerance band, defect category mapping rules, result output format, and re-judgment trigger threshold; the recipe score is determined based on at least two of the following: detection rate, false detection rate, missed detection risk, cycle time impact, equipment operation stability, the effect of similar historical recipes, and manual confirmation cost.

6. The collaborative decision-making method for automated optical inspection in TGV processes according to claim 5, characterized in that, The process of performing defect detection according to the candidate Recipe, writing defect data to the unified data platform in real time, and calling the defect attribution agent to generate defect attribution chains and process risk scores for each defect attribution chain includes: When the detection Recipe generated by the Recipe needs to be sent to the production equipment, the following checks are performed: whether the CAD version is consistent with the design version of the current production product, whether the equipment model is compatible with the Recipe, whether the workstation number corresponds to the Recipe, whether the product model matches, whether the coordinate system version is consistent with the current equipment coordinate system, whether the AI ​​model version has been pre-installed on the equipment, and whether manual approval has been completed. Once all these conditions are met, an executable Recipe file is generated and sent to the equipment. After the Recipe is executed on the device, the detection results, manual review results, post-test / packaging results, and engineer's modification records of the Recipe are written back to the unified data platform. The original image, defect screenshot, CAD version, Recipe version, model version, coordinate system version, manual review results, and post-test results corresponding to each detection result are also recorded. Defects on the same glass substrate, such as white film defects, laser-induced anomalies, blind holes or rough hole walls after etching, voids filled by conductive paste, circuit expansion or contraction or circuit offset, and electrical test anomalies, are spatiotemporally correlated based on the unique identifier of the glass substrate and global coordinates using a unified index, forming one or more defect attribution chains; The process failure probability P(Fail|X,Z) is calculated using a conditional probability model, Bayesian network, causal graph, gradient boosting tree, graph neural network, or rule-model hybrid method, where X includes defect image features, geometric features, coordinate features, and CAD interference features, and Z includes process parameters, equipment parameters, and recipe version. The process risk score for each defect attribution chain is also output.

7. The collaborative decision-making method for automated optical inspection in TGV processes according to claim 1, characterized in that, The step of invoking a decision-making agent to output disposal suggestions based on the process risk score, the detection recipe candidates, the coordinate confidence score, and the process thresholds in the rule base includes: The system employs a combination of rule gating and model scoring to output handling recommendations. The rule gating is used to constrain high-risk actions involving production stoppage, scrapping, recipe distribution, and model deployment. The model scoring is used to generate low-risk re-judgment, continued processing, sampling re-inspection, or recipe candidate ranking. The handling recommendations carry decision-making criteria, which include at least the triggering defect, defect coordinates, related CAD elements, associated workstations, risk score, historical similar cases, current recipe version, candidate recipe differences, and manual confirmation status.

8. The collaborative decision-making method for automated optical inspection in TGV process according to claim 1, characterized in that, The invoking model iterative agent generates model iterative update tasks based on difficult examples during the detection process. These difficult examples include false positives, false negatives, low-confidence samples, inconsistencies in manual review, and failed samples in the later stages. Collect false detection samples, missed detection samples, low confidence samples, inconsistent cross-model judgment samples, inconsistent manual re-judgment and model judgment samples, failed back-end electrical test samples, and duplicate samples of the same type from the entire link. Deduplicate the samples based on image perceptual hash, defect location, defect category, image feature vector and CAD primitive number. The representative samples after deduplication are entered into the difficult sample pool. Generate annotation tasks, divide the training set / validation set, and after completing model training, calculate performance metrics including precision, recall, F1 score, critical defect recall, false positive rate, false negative rate, online beat impact, and cross-batch stability on the validation set, recent online samples, and critical defect samples. When the model reaches the deployment threshold, it is deployed to a specified product, device, or batch in a gray-scale manner. If the false negative rate, false positive rate, or cycle time does not meet the requirements during the gray-scale deployment phase, it is rolled back to the previous stable version. When any performance indicator fails to meet the deployment threshold, release is prohibited and supplementary sample suggestions are not generated.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to execute the collaborative decision-making method for automated optical inspection of the TGV process as described in any one of claims 1 to 8.

10. A computing device, characterized in that, The computing device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the collaborative decision-making method for automated optical inspection of TGV process according to any one of claims 1 to 8.

Images