Detection method capable of improving quality of quartz wafer

Abstract

The invention relates to the technical field of quartz wafer quality detection, in particular to a detection method capable of improving the quality of a quartz wafer, which adopts innovative means such as a multi-modal fusion defect detection technology, a dynamic self-adaptive defect scoring system, an intelligent evaluation model based on a knowledge graph, closed-loop feedback process optimization and the like. The problems that an existing quartz wafer detection method is single in model, static in score, not considering defect positions and severity, insufficient in data utilization and the like are solved, the crossing of quartz wafer quality detection from post grading to pre-prevention is achieved, the detection precision and adaptability are effectively improved, and the reject ratio is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of quartz wafer quality detection, and in particular to a detection method capable of improving the quality of quartz wafers. BACKGROUND

[0002] As a key basic component in the field of electronic information, quartz wafers are widely used in many fields such as communication, aerospace, automotive electronics, etc. The quality of quartz wafers directly affects the performance, stability and reliability of electronic devices. With the rapid development of electronic technology, the quality requirements for quartz wafers are becoming higher and higher, and the traditional detection methods have gradually exposed many problems.

[0003] In the prior art, the defect detection method of quartz wafers mainly relies on a single detection means, such as constructing a defect region analysis model only through a convolutional neural network (CNN). This method has insufficient detection capability for small defects under complex lighting conditions and multiple defect overlapping scenes, and is prone to missed detection or misjudgment. For example, the quartz wafers used in some precision electronic devices may have very fine internal defects due to their small size, and traditional single models are difficult to accurately identify.

[0004] In addition, there is a static problem in defect quantification scoring. Existing defect scoring tables are often based on fixed indexes and cannot adapt to changes in production processes and the emergence of new defects. For example, when new materials or production processes are introduced, new defect types may be generated, but the original scoring system cannot accurately assess them. Moreover, the scoring only considers the size of the defect without considering the depth and direction of the defect, which are key dimensions that have an important impact on the strength and electrical performance of the wafer.

[0005] In the appearance quality evaluation model, the K-Nearest Neighbor algorithm (KNN) is currently relied on. This algorithm is greatly affected by the balance of historical data samples, and if there are few samples of a certain type of defect, it is easy to misjudge the "minority submits to the majority". At the same time, the evaluation coordinate system only takes the defect type as the dimension without integrating multiple data sources such as defect location and wafer size tolerance, resulting in a single evaluation dimension that cannot fully and accurately reflect the quality of the wafer.

[0006] In terms of data utilization, historical data labeling mainly relies on manual work, which is costly and has insufficient labeled samples for rare defects, resulting in weak model generalization ability. Furthermore, unsupervised learning or semi-supervised learning techniques are not fully utilized to exploit the value of unlabeled data, resulting in low data utilization. In the production process, the detection results are only used for quality grading and are not linked to production process parameters, making it impossible to reduce defects from the source.

[0007] With the development of emerging technologies such as 5G communication and artificial intelligence, higher requirements are put forward for the quality and performance of quartz wafers, and there is an urgent need for a new detection method that can improve detection accuracy, adapt to production changes, and fully utilize data value. SUMMARY

[0008] To solve the problems in the prior art, the application provides a detection method for improving the quality of quartz wafers.

[0009] The technical solution adopted by the application to solve the technical problem is: a detection method for improving the quality of quartz wafers, comprising the following steps: introducing "optical image + infrared thermal imaging + machine vision" multi-modal data acquisition to construct a multi-task learning model for defect detection; constructing a dynamic adaptive defect scoring system, quantitatively scoring defects based on three-level index expansion and real-time weight adjustment; building an intelligent evaluation model based on a knowledge graph, and performing quality evaluation through feature encoding, graph matching and risk prediction; establishing a closed-loop feedback process optimization mechanism to realize detection and production linkage and continuous learning.

[0010] Specifically, the multi-modal data acquisition is: using optical image detection to detect surface cracks, scratches, notches and other visible defects; using infrared thermal imaging to identify internal stress concentration areas; and using structured light three-dimensional scanning to measure the edge perpendicularity of the wafer and other geometric tolerances.

[0011] Specifically, the three-level index expansion in the dynamic adaptive defect scoring system includes: the first-level index is the defect type, covering cracks, scratches, notches and stress spots; the second-level index is the defect geometric feature, including length, width, depth and curvature; and the third-level index is the defect position weight and direction, such as edge area, center area and electrode area weight.

[0012] Specifically, the real-time weight adjustment in the dynamic adaptive defect scoring system is to combine the failure data of the production batch, dynamically update the weight coefficients of each defect type through Bayesian optimization, and automatically increase the scoring weight of "edge notch" when the failure ratio caused by edge notch in a batch exceeds 30%.

[0013] Specifically, the quantitative scoring formula of the dynamic adaptive defect scoring system is , wherein is the defect type weight (dynamically updated), is the defect geometric feature score (predicted based on a deep learning regression model), is the position weight coefficient (edge area = 1.5, center area = 1, and electrode area = 2).

[0014] Specifically, the intelligent evaluation model architecture based on the knowledge graph comprises an entity layer, a relationship layer and an inference layer; the entity layer comprises defect types, sizes, positions, wafer models and production process parameters; the relationship layer represents the influence degree of defects on frequency stability and mechanical strength and the causal relationship between process parameters and defects; and the inference layer realizes defect correlation analysis based on a graph neural network (GNN), and when an "edge notch" is detected, "excessive cutting process pressure" is automatically associated and process adjustment suggestions are given.

[0015] Specifically, the intelligent evaluation model evaluation process based on the knowledge graph comprises: converting detected defect parameters into graph nodes for feature coding; searching for historical performance data of similar defect combinations in the knowledge graph for graph matching; and outputting a quality level (qualified / suspected / unqualified) and failure risk prediction through causal reasoning.

[0016] Specifically, the detection-production linkage mechanism of the closed-loop feedback process optimization mechanism comprises: real-time communication between the detection system and the production equipment, automatic triggering of process parameter adjustment (such as a 5% reduction in grinding pressure) when the same type of defects (such as "center area scratches") appear in three consecutive batches of a certain type of wafer, and automatic generation of a process optimization work order through association rule mining (such as the Apriori algorithm) to identify strong correlation between defects and process parameters when the "edge corner collapse" defect rate is greater than 5% and the "cutting speed is greater than 200 mm / s".

[0017] Specifically, the continuous learning system of the closed-loop feedback process optimization mechanism comprises: inputting new detection data (including unlabeled samples) into a semi-supervised learning model for online incremental training; periodically updating the defect detection and evaluation model without the need for manual re-labeling of full-quantity data; and simulating the influence of different defect combinations on frequency stability through a wafer performance twin model to verify the evaluation accuracy of the detection algorithm in reverse.

[0018] Specifically, the method improves the detection accuracy of quartz wafers by more than 20% and reduces the defect rate by 15-20% through multi-dimensional detection, dynamic adaptability, knowledge-driven decision-making and closed-loop intelligence.

[0019] The beneficial effects of the present application are: 1. Improve detection accuracy: Through multi-modal fusion defect detection technology, break through the limitations of single detection means, fuse multi-source data such as optics, infrared and three-dimensional structure, effectively identify complex defects, improve detection accuracy by more than 20%, and more accurately find various defects of quartz wafers to ensure product quality.

[0020] 2. Enhance adaptability: The dynamic adaptive defect scoring system can automatically adjust the scoring weights based on real-time production data according to changes in production processes and the appearance of new defects, adapt to different production scenarios, solve the static problem of traditional scoring systems, and make quality evaluation more scientific and reasonable.

[0021] 3. Intelligent decision making: The intelligent evaluation model based on the knowledge graph can realize the causal reasoning of defect-performance-process. It can not only accurately evaluate the quality level, but also trace the defect source to provide scientific suggestions for production process adjustment, support preventive quality control, and avoid the generation of a large number of defective products.

[0022] 4. Optimization of production process: The closed-loop feedback process optimization mechanism realizes real-time linkage and continuous learning of detection and production, reduces defects from the source, builds a "detection-analysis-optimization" full-link closed loop, and makes the production process from passive detection to active quality improvement, effectively reduces the quartz wafer defect rate by 15%-20%, and improves production efficiency and enterprise economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0023] The present application will be further described below in conjunction with the drawings and examples.

[0024] Fig. 1 The architecture diagram of the multi-modal fusion defect detection method provided by the present application; Fig. 2 The architecture diagram of the intelligent evaluation model based on the knowledge graph provided by the present application; Fig. 3 The process flow diagram of the closed-loop feedback process optimization mechanism provided by the present application. DETAILED DESCRIPTION

[0025] In order to make the technical means, creative features, purposes and effects achieved by the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0026] As shown in Figs. 1-3 The detection method for improving the quality of quartz wafers described in the present application comprises: Multi-modal fusion defect detection technology, introducing "optical image + infrared thermal imaging + machine vision" multi-modal data acquisition method. Optical image can directly detect visible defects such as cracks, scratches, and notches on the wafer surface; infrared thermal imaging can deeply identify stress concentration areas inside the wafer, which often indicate potential crack risks; structured light three-dimensional scanning can accurately measure the edge perpendicularity and other geometric tolerances of the wafer. By constructing a multi-task learning model, different modal data features are fused using the Transformer architecture to realize cross-modal feature fusion and improve the recognition rate of complex defects. At the same time, based on the attention mechanism, dynamic ROI (region of interest) positioning is realized to preferentially detect defects in positions that have the greatest impact on performance, such as within 3mm of the wafer edge.

[0027] Dynamic adaptive defect scoring system, build dynamic scoring table, adopt three-level index expansion. The first-level index covers multiple defect types such as cracks, scratches, notches, and stress spots; the second-level index is refined into defect geometric features including length, width, depth, and curvature; the third-level index introduces defect position weight (different weights for edge area, center area, and electrode area) and orientation (such as double weight for edge crack orientation). Combined with production batch failure data, the weight coefficients of each defect type are dynamically updated through Bayesian optimization to realize real-time weight adjustment. The quantitative scoring formula is , wherein is the defect type weight (dynamically updated), is the defect geometric feature score (predicted based on a deep learning regression model), is the position weight coefficient (edge area = 1.5, center area = 1, electrode area = 2), which accurately quantifies the defects.

[0028] Intelligent evaluation model based on knowledge graph, build knowledge graph model containing entity layer, relationship layer and reasoning layer. The entity layer contains defect type, size, position, wafer model, production process parameters and other information; the relationship layer represents the influence degree of defects on frequency stability and mechanical strength, as well as the causal relationship between process parameters and defects; the reasoning layer realizes defect correlation analysis based on graph neural network (GNN). The evaluation process is as follows: first, the detected defect parameters are converted into graph nodes for feature encoding; then the historical performance data of similar defect combinations are retrieved in the knowledge graph for graph matching; finally, the quality level (qualified / suspected / unqualified) and failure risk prediction are output through causal reasoning, providing a more scientific basis for quality evaluation.

[0029] Closed-loop feedback process optimization, establish detection-production linkage mechanism, detection system and production equipment (cutting machine, grinding machine, etc.) real-time communication. When the same type of defects (such as "center area scratch") appear in 3 consecutive batches of a certain type of wafer, automatically trigger process parameter adjustment, such as reducing grinding pressure by 5%. Use association rule mining (such as Apriori algorithm) to identify strong association between defects and process parameters, when "edge corner collapse" defect rate > 5%, associated with "cutting speed > 200 mm / s", automatically generate process optimization work order. The continuous learning system inputs new detection data (including unlabeled samples) into the semi-supervised learning model for online incremental training, regularly updates the defect detection and evaluation model, without the need for manual re-labeling of full-quantity data, and simulates the impact of different defect combinations on frequency stability through wafer performance twin model, to verify the evaluation accuracy of the detection algorithm.

[0030] Example one: multi-modal fusion defect detection technology implementation In a large quartz wafer production workshop, a new batch of quartz wafers begins to enter the detection link. The optical image acquisition equipment in the workshop uses a high-resolution industrial camera equipped with a customized telecentric lens, which can effectively avoid defect misjudgment caused by perspective distortion. At the same time, the light source system uses a multi-angle adjustable LED light source. For quartz wafers with different surface characteristics, the best contrast surface image can be obtained by adjusting the light source angle and intensity, and micron-level surface cracks, scratches and other defects can be clearly captured.

[0031] The infrared thermal imager has high-sensitivity infrared detectors that can detect very small temperature differences. When scanning quartz wafers, its temperature resolution can reach 0.05°C, and the frame rate is 30 frames per second, ensuring that no internal stress concentration area that may indicate potential crack risk is missed during rapid detection. The thermal imager is also equipped with an intelligent temperature compensation algorithm that can automatically adjust detection parameters according to environmental temperature changes to ensure the accuracy of detection results.

[0032] The structured light three-dimensional scanning device projects a series of carefully designed Gray code structured light patterns onto the surface of the quartz wafer and uses a high-speed camera to collect reflected light information. Through complex triangulation principles and high-precision algorithms, the edge perpendicularity and other geometric tolerances of the wafer are accurately calculated, with a measurement accuracy of ±0.01mm.

[0033] After the data collection of the three modalities is completed, it is transmitted synchronously to the multi-task learning model. The Transformer architecture in the model has multiple attention mechanism modules that can deeply fuse the features of different modalities of data. For example, when fusing the surface defect features in the optical image and the internal stress features in the infrared thermal image, the Transformer architecture automatically assigns weights according to the importance of the features, strengthens those features that play a key role in defect identification, and effectively improves the recognition rate of complex defects such as micro-cracks hidden under the surface. At the same time, based on the attention mechanism, the model learns from a large amount of historical data that the 3mm range of the wafer edge is an area with high incidence of defects and has a greater impact on performance, so it automatically focuses on this area and prioritizes detecting defects in these locations. After actual production verification, compared with the traditional single image detection method, the detection accuracy of this multi-modal fusion detection technology has been improved from 75% to more than 95%.

[0034] Example Two: Implementation of Dynamic Adaptive Defect Scoring System During the production process of a batch of quartz wafers, the detection equipment successively found multiple defects. First, there was a crack with a length of 2.5mm and a width of 0.12mm located in the edge area of the wafer; there was also a scratch with a length of 3.2mm and a depth of 0.06mm located in the center area of the wafer. The staff accurately entered this defect information into the detection system.

[0035] The system starts scoring according to the dynamic adaptive defect scoring system. For the crack, the first-level index is determined as the "crack" category according to its type; the second-level index is determined by the measured length, width, and other geometric characteristics; and the position weight of the third-level index is determined as 1.5 because it is located in the edge area. At this time, the system queries the historical failure data of the production line in the recent period and finds that the product failure caused by edge cracks accounts for 40% in the last week, exceeding the preset threshold of 30%. Therefore, the system automatically increases the weight coefficient of the edge crack defect type from the initial 1.0 to 1.3 through the Bayesian optimization algorithm.

[0036] For the scratch, the first-level index is also determined as the "scratch" category, and the second-level index is determined according to the geometric characteristics, and the position weight is 1 because it is in the center area. Next, the system calculates according to the quantitative scoring formula . Among them, Through the deep learning regression model, it is predicted that the model has been trained with a large amount of labeled data and can accurately predict the impact score of the defect on the wafer quality according to the geometric characteristics of the defect.

[0037] For example, for this crack, according to its length, width, and other characteristics, the model predicts 0.8; for the scratch, the model predicts 0.6. Finally, the score of the crack is 1.3 x 0.8 x 1.5 = 1.56, and the score of the scratch is 1 x 0.6 x 1 = 0.6. Through this dynamic scoring method, compared with the traditional static scoring, it can more accurately reflect the impact of different defects on the wafer quality under the current production conditions, and provide a more reliable basis for subsequent quality evaluation and decision-making.

[0038] Example Three: Implementation of Intelligent Evaluation Model Based on Knowledge Graph In the detection of a batch of quartz wafers, an "edge notch" defect was found. The detection system quickly converts the parameters of this defect, such as the diameter of the edge notch being 1.2 mm and being located at a distance of 2 mm from the end point of the wafer edge, into graph nodes and completes feature encoding.

[0039] Then, the system searches for historical performance data of similar defect combinations in the pre-constructed large knowledge graph. This knowledge graph is constructed through the collection and analysis of a large amount of quartz wafer production data, detection data, and performance test data over many years, and stores rich data on the relationship between different defect combinations and wafer performance. Through an efficient graph search algorithm, the system quickly finds historical cases similar to the current "edge notch" defect combination.

[0040] Next, the causal reasoning is performed using a graph neural network (GNN). The GNN model is trained deeply and can learn the complex associations between various entities and relationships in the knowledge graph. In this case, the model automatically associates the cause of "excessive cutting process pressure" based on the causal relationship between process parameters and defects stored in the knowledge graph. Because in the historical data, a large number of similar defect combinations are related to cutting pressure abnormalities. The model not only associates the cause, but also gives specific process adjustment suggestions according to relevant experience and algorithms, such as reducing the cutting pressure from the current 5MPa to 4.5MPa.

[0041] Finally, the system outputs the quality level of the batch of wafers as "suspicious" according to the performance prediction model in the knowledge graph, and gives a high failure risk prediction. Compared with the traditional KNN algorithm, the accuracy and comprehensiveness of the intelligent evaluation model are greatly improved, which can provide more valuable reference for production decision-making.

[0042] Example Four: Closed-loop feedback process optimization-detection-production linkage mechanism implementation In a quartz wafer production workshop, the detection system is connected in real time with cutting machines, grinding machines and other production equipment through industrial Ethernet. Recently, a type of wafer has been detected with center area scratch defects in three consecutive production batches. The detection system immediately transmits this information to the control system of the grinding machine.

[0043] The control system of the grinding machine is built-in with an intelligent decision-making module, which analyzes according to pre-set rules. When receiving the information of continuous occurrence of center area scratch defects, the decision-making module judges that it may be caused by excessive grinding pressure leading to wafer surface damage. Therefore, it automatically reduces the grinding pressure from the current 8MPa to 7.6MPa, and records the time, adjustment amplitude and detection situation of the subsequent wafers.

[0044] At the same time, the data analyst of the workshop uses the association rule mining algorithm (such as Apriori algorithm) to deeply analyze the production data of the past few months. When the "edge collapse angle" defect rate exceeds 5%, the algorithm discovers through mining a large number of production parameters and defect data that there is a strong correlation between it and "cutting speed > 200mm / s". Based on this analysis result, the system automatically generates a process optimization work order, suggesting to adjust the cutting speed from the current 220mm / s to 180mm / s, and sends the work order to the cutting machine operator. After receiving the work order, the operator adjusts the cutting speed according to the suggestion and observes the defect situation in the subsequent production. Through this detection-production linkage mechanism, the rapid response from detection to production process adjustment is realized, effectively reducing the generation of defects from the source.

[0045] Example Five: Closed-loop feedback process optimization-continuous learning system implementation In the daily production process, the quartz wafer detection system continuously generates new detection data, including both defect data that has been manually labeled and a large number of unlabeled samples. These new data are transmitted to the semi-supervised learning model in real time.

[0046] The semi-supervised learning model uses an advanced self-training algorithm. It first trains an initial model using a small amount of labeled data. Then, the model makes predictions on a large amount of unlabeled data, selects samples with high prediction confidence, and adds them as pseudo-labeled data to the training set to further train the model. For example, a rare defect with a special shape is newly detected. Although there is not enough manually labeled data, the model can automatically generate pseudo-labels for it through the self-training algorithm, using the labeled information of similar defects around it and the characteristics of this defect, thereby learning the characteristics of the new defect.

[0047] Every certain period of time, the system automatically updates the defect detection and evaluation model. The updating process does not require manual re-labeling of the full amount of data, but uses newly accumulated labeled data and new knowledge learned by the semi-supervised learning model to fine-tune the parameters of the model. At the same time, in order to verify the accuracy of the detection algorithm, the workshop establishes a wafer performance twin model. Based on the principle of finite element analysis, this model can simulate the impact of different defect combinations on the frequency stability of the wafer. By inputting the detected defects into the twin model, the frequency change output by the model can be observed and compared with the test results on the high-precision frequency test equipment. If a deviation is found between the model prediction results and the actual test results, the related parameters in the detection algorithm will be adjusted in the opposite direction, and the detection method will be continuously optimized to ensure the accuracy and reliability of the detection.

[0048] Example Six: Comprehensive Implementation Effect In a certain quartz wafer production enterprise, the detection method of the present application has been comprehensively applied for half a year. During this period, the production efficiency and product quality of the enterprise have been significantly improved.

[0049] In terms of detection accuracy, the multi-dimensional detection capability breaks through the limitations of single image detection. In the past, when relying solely on optical image detection, some defects hidden inside or under complex lighting conditions were often missed, and the detection accuracy was only about 70%. After integrating optical, infrared, and three-dimensional structure data, through multi-modal fusion defect detection technology, wafer information can be obtained from multiple angles, and various complex defects can be effectively identified. The detection accuracy is now stable at more than 90%, greatly improving the product quality control capability.

[0050] The dynamic adaptability aspect automatically adjusts the scoring weight and process parameters based on real-time production data, effectively solving the adaptability problem of traditional methods to new defects and process changes. After introducing a new raw material supplier, a new type of micro stress spot defect appeared, and the traditional detection method could not accurately evaluate its impact on wafer quality. However, the dynamic adaptive defect scoring system of the present application quickly adjusts the scoring weight of this new type of defect by collecting real-time production data and using Bayesian optimization algorithm, making the quality evaluation more accurate. At the same time, the closed-loop feedback process optimization mechanism adjusts the production process parameters in a timely manner according to the detection results, avoiding the production of a large number of defective products due to the inadaptability of the process to new materials.

[0051] Knowledge-driven decision-making realizes causal reasoning of defect-performance-process through knowledge graph, which not only accurately evaluates the quality level, but also traces the source of defects, supporting preventive quality control. In the past, when facing product quality problems, enterprises often only conduct post-analysis, making it difficult to prevent in advance during the production process. Now, the intelligent evaluation model based on knowledge graph can quickly analyze the possible causes and give process adjustment suggestions at the moment of detecting defect combinations, avoiding the production of a large number of defective products and saving production costs.

[0052] The "detection-analysis-optimization" full-link closed loop constructed by closed-loop intelligence changes from passive detection to active quality improvement. Through continuous learning system to constantly optimize the detection model and production process, the enterprise's quartz wafer defective rate is significantly reduced from 15% to 5%-10%. The improvement of product quality greatly enhances the competitiveness of the enterprise in the market, receives more orders from high-end customers, and significantly improves the production efficiency.

[0053] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting an improvement in the quality of a quartz wafer, characterized by, The method comprises the following steps: A multi-modal data acquisition of "optical image + infrared thermal imaging + machine vision" is introduced to construct a multi-task learning model for defect detection; A dynamic adaptive defect scoring system is constructed to quantitatively score defects based on three-level index expansion and real-time weight adjustment; An intelligent evaluation model based on a knowledge graph is built to evaluate quality through feature coding, graph matching and risk prediction; A closed-loop feedback process optimization mechanism is established to realize detection and production linkage and continuous learning.

2. The method of claim 1, wherein: The multi-modal data acquisition specifically comprises: Surface cracks, scratches, and notches are detected by optical images; Internal stress concentration areas are identified by infrared thermal imaging; The edge perpendicularity geometric tolerance of the wafer is measured by structured light three-dimensional scanning.

3. The method of claim 1, wherein the method is performed by a wafer quality enhancement system. The three-level index expansion in the dynamic adaptive defect scoring system comprises: The first-level index is the defect type, covering cracks, scratches, notches, and stress spots; The second-level index is the defect geometric feature, including length, width, depth, and curvature; The third-level index is the defect position weight and direction, such as different weights of edge area, center area, and electrode area; And the edge crack direction is doubled in weight.

4. The method of claim 1, wherein: The real-time weight adjustment in the dynamic adaptive defect scoring system is to combine the failure data of the production batch, dynamically update the weight coefficients of each defect type through Bayesian optimization, and automatically increase the scoring weight of "edge notch" when the failure rate of a batch caused by edge notch exceeds 30%.

5. The method of claim 1, wherein: The quantitative scoring formula of the dynamic adaptive defect scoring system is wherein is a defect type weight, is a defect geometric feature score, is a position weight coefficient.

6. The method of claim 1, wherein: The architecture of the intelligent evaluation model based on the knowledge graph comprises an entity layer, a relationship layer, and an inference layer; The entity layer includes defect type, size, position, wafer model, and production process parameters; The relationship layer represents the influence of defects on frequency stability and mechanical strength, as well as the causal relationship between process parameters and defects; The inference layer realizes defect correlation analysis based on graph neural networks, and when "edge notch" is detected, it automatically associates "excessive pressure in the cutting process" and gives process adjustment suggestions.

7. The method of claim 1, wherein: The evaluation process of the intelligent evaluation model based on the knowledge graph comprises: The detected defect parameters are converted into graph nodes for feature coding; Historical performance data of similar defect combinations are retrieved in the knowledge graph for graph matching; The quality level and failure risk prediction are output through causal reasoning, and the quality level is divided into qualified, suspicious, and unqualified.

8. The method of claim 1, wherein: The detection-production linkage mechanism of the closed-loop feedback process optimization mechanism comprises: The detection system communicates with the production equipment in real time, and when the same type of defect appears in three consecutive batches of a certain type of wafer, the process parameter adjustment is automatically triggered, such as reducing the grinding pressure by 5%; When the "edge collapse angle" defect rate is greater than 5%, the process optimization work order is automatically generated by associating "cutting speed > 200 mm / s".

9. The method of claim 1, wherein: The continuous learning system of the closed-loop feedback process optimization mechanism comprises: Unlabeled sample-containing new detection data are input into a semi-supervised learning model for online incremental training; The defect detection and evaluation model is updated regularly without the need for manual re-labeling of full-quantity data; The influence of different defect combinations on frequency stability is simulated through a wafer performance twin model to verify the evaluation accuracy of the detection algorithm in reverse.

10. The method of claim 1-9, wherein the method is used to improve the quality of a quartz wafer. The method improves the quartz wafer detection accuracy by more than 20% and reduces the defective rate by 15%-20% through multi-dimensional detection, dynamic adaptability, knowledge-driven decision-making, and closed-loop intelligence.

Images