A method for detecting a seeding diameter

Abstract

This invention relates to the field of monocrystalline silicon rod production, and specifically to a method for detecting the diameter of a lead crystal. The method includes: defining a lead crystal diameter detection area in an image of the lead crystal process; extracting a set of feature points representing the outer contour of the bright portion of the lead crystal aperture within the detection area; performing circle fitting on the outer contour feature point set to obtain a fitted circle diameter in pixels, and converting the fitted circle diameter into a physical diameter, denoted as D0; standardizing the image within the detection area; inputting a pre-trained classification model, and outputting a long point similarity judgment result, denoted as A1; processing multiple consecutive long point similarity judgment results A1 to obtain the probability of a long point existing in the current lead crystal aperture, denoted as A2; and combining the physical diameter D0 with the long point existence probability A2 to calculate the compensated lead crystal diameter, thereby achieving real-time identification of abnormally protruding long points in the lead crystal aperture and dynamically compensating the diameter measurement value accordingly.

Description

Technical Field

[0001] This invention relates to the field of monocrystalline silicon rod production, and specifically to a method for detecting the diameter of a silicon lead crystal. Background Technology

[0002] Monocrystalline silicon rods are essential raw materials for the semiconductor and photovoltaic industries, and their usage is very large in the production of photovoltaic modules. Currently, the main methods for manufacturing monocrystalline silicon rods are the Czochralski method and the zone melting method. The Czochralski method grows high-quality monocrystalline silicon rods by pulling them from molten silicon. Due to its high maturity and controllable cost, it is the dominant technology for monocrystalline silicon preparation. The main process flow of the Czochralski method includes:

[0003] 1. Loading and Melting: Polycrystalline silicon is loaded into a quartz crucible and placed in a graphite heater. It is heated to the melting point of silicon in an argon atmosphere, and the temperature is precisely controlled to ensure that the silicon melt is uniform and to avoid excessive temperature fluctuations that could cause lattice defects.

[0004] 2. Seed Crystal Welding: After the melting process is complete and the temperature stabilizes, the crucible needs to be raised to maintain a suitable distance from the guide tube. A high-purity single-crystal silicon seed crystal is fixed on a rotating hammer. Once the temperature of the molten silicon in the furnace is suitable, the seed crystal is slowly brought into contact with the liquid surface. The temperature is controlled by adjusting the power, allowing the part of the seed crystal in contact with the molten silicon to melt and recrystallize, ensuring lattice matching.

[0005] 3. Necking and crystal pulling: After the furnace temperature is suitable, quickly pull up the seed crystal to reduce the crystal diameter to about 5mm. Then, continue to pull the crystal to a certain extent according to the 5mm diameter, usually between 200 and 250mm, in order to eliminate dislocation defects. During the crystal pulling process, it is necessary to keep the diameter relatively constant.

[0006] 4. Shoulder Formation: Reduce the pulling speed to gradually increase the crystal diameter to the target diameter. During the shoulder formation process, temperature step control is required, and the crucible lifting speed needs to be adjusted to maintain a relatively stable liquid outlet position, ensuring uniform crystal growth and avoiding stress-induced cracking.

[0007] 5. Constant diameter growth: The pulling speed, temperature and liquid outlet distance are adjusted by feedback such as diameter measurement to maintain the constant diameter growth of the crystal.

[0008] 6. Finishing: Increase the pulling speed and gradually reduce the diameter to form a conical tail to avoid dislocation multiplication caused by thermal stress.

[0009] Currently, the industry commonly uses computer vision to measure changes in aperture pixel values ​​to determine diameter changes during crystal pulling, and controls the crystal lifting speed based on target settings to maintain the crystal diameter control within the set requirements. Current single-crystal furnace vision systems primarily rely on measuring aperture pixel values ​​to determine the crystal diameter. However, temperature fluctuations during crystal pulling significantly impact diameter control. Specifically, when the temperature fluctuates downwards, the crystal pulling aperture does not expand uniformly; instead, the inner aperture first bulges outwards. Typically, about 5 seconds after the temperature fluctuation, a growth point phenomenon appears in the crystal pulling aperture, meaning outward-growing crystal points appear on the inner aperture, lasting for 10-15 seconds. During this process, the diameter value obtained using existing measurement methods shows no significant change. Only after the inner crystal point bulges outwards to a certain extent will the outer aperture suddenly expand. At this point, to bring the abnormally increased diameter back to the target value, the system must significantly increase the crystal pulling speed to achieve a gradual diameter correction, resulting in a significant lag in diameter control. Summary of the Invention

[0010] To address the problems of existing technologies, this invention provides a method for detecting the diameter of a crystal driver.

[0011] The objective of this invention can be achieved through the following technical solution: A method for detecting the diameter of a crystal driver includes:

[0012] S1: In the image of the crystal pulling process, a crystal diameter detection area is set; the crystal pulling process is the process of drawing according to the target diameter after the diameter reduction is completed in the crystal pulling stage.

[0013] S2: Extract the outer contour feature point set of the bright part of the pilot aperture within the detection area; use a feature extraction algorithm to perform circle fitting on the outer contour feature point set to obtain the fitted circle diameter in pixels; based on the preset conversion coefficient between pixel diameter and physical diameter, convert the fitted circle diameter into a physical diameter in millimeters, denoted as D0.

[0014] S3: Standardize the image within the detection area; input the standardized image into a pre-trained classification model, and output the long point similarity determination result through the classification model, denoted as A1;

[0015] S4: The mean filter is used to process the similarity determination results A1 of multiple consecutive long points to obtain the probability that there are long points in the current pilot aperture, which is denoted as A2.

[0016] S5: Based on the preset upper limit of diameter offset, combined with the physical diameter D0 and the probability of long point existence A2, the compensated crystal diameter is calculated.

[0017] In a further improvement, in S2, the feature extraction algorithm is the least squares method; the conversion coefficient between the pixel diameter and the physical diameter is the ratio of the physical value of the actual pilot diameter to the pixel diameter value of the fitted circle corresponding to the actual pilot diameter.

[0018] Further improvements in S3 include the following specific steps in the standardization process: scaling and cropping the image within the detection area by a specified scale, using the center point of the lead-in aperture as a reference; performing grayscale stretching on the cropped image to ensure consistent image size and brightness; the pre-trained classification model is a convolutional neural network classification model, which, during the pre-training phase, uses labeled images with two categories: those with long points and those without; the output rule for the long point similarity judgment result A1 is: if the classification model determines that the image does not have long points, then A1 = 0; if the classification model determines that the image has long points, then A1 is the confidence level corresponding to the judgment result, and the value range of A1 is 0.5-1.

[0019] In a further improvement, in S4, the specific operation of the mean filtering includes: continuously collecting multiple similarity judgment results A1 of the long points; summing all the collected A1 and taking the arithmetic mean as the probability of the long point existence A2; the mean filtering is used to smooth the random fluctuations of the A1 data, and the value range of A2 is 0-1.

[0020] In a further improvement, in S5, the preset upper limit of diameter offset is denoted as D2, and the value of D2 is 5%-10% of the target diameter of the lead crystal, and D2 is determined according to the on-site single crystal silicon pulling process standard; the compensated lead crystal diameter is denoted as D1, and D1 is calculated by the formulas D1=D0 and D2×A2.

[0021] Further improvements include S6: transmitting the compensated lead diameter D1 to the PLC control system; the PLC control system compares D1 with the preset lead target diameter to stabilize the lead diameter within the target range.

[0022] In a further improvement, the PLC control system adopts a PID control strategy, which drives the crystal lifting mechanism to adjust the crystal pulling speed according to the changing trend of D1 under different operating conditions. The specific process is as follows:

[0023] When the temperature of the thermal field is stable or rising: the long point detection results A1=0, A2=0, D1=D0, and the PLC controls the pulling speed to remain stable, so that the lead diameter is consistent with the target diameter;

[0024] When the thermal field temperature fluctuates downwards: long spots appear in the crystal pulling aperture, and the detection result A1 of the long spots increases accordingly, which in turn leads to an increase in A2, ultimately causing D1 to show a sudden increase trend; after the PLC control system detects the upward trend of D1, it rapidly increases the crystal pulling speed.

[0025] When the long spot phenomenon subsides: the long spot detection result A1 gradually decreases to 0, A2 also gradually decreases to 0, and D1 shows a downward trend; after the PLC control system detects the downward trend of D1, it gradually reduces the crystal pulling speed.

[0026] Compared with existing technologies, the beneficial effects of the crystal diameter detection method of the present invention are as follows:

[0027] By visually monitoring the long point phenomenon of the seeding aperture in real time, combined with a dynamic diameter compensation mechanism, the long point signal can be captured in advance and the crystal lifting mechanism can be driven to adjust the pulling speed when the thermal field temperature drops and the outer diameter of the aperture has not changed. This effectively avoids the abnormal sudden increase in the seeding diameter caused by the lag response of traditional methods.

[0028] When the thermal field temperature is stable and there is no long spot phenomenon, the long spot detection value is always 0. At this time, the actual diameter of the crystal puller is completely consistent with the diameter measured by the aperture, and no additional compensation is required. This ensures the accuracy of diameter detection under stable working conditions and avoids interference from invalid adjustments to the crystal puller process.

[0029] When the crystal pulling speed is increased and the long point phenomenon subsides, the long point detection value gradually drops back to 0. The PLC system can gradually reduce the pulling speed according to the downward trend of the final measured value to slow down the diameter shrinkage rate and avoid the problem of crystal breakage caused by rapid diameter shrinkage from the source.

[0030] Through the integrated design of real-time visual monitoring and PLC closed-loop control, the system enables instant identification of long point anomalies, dynamic compensation of diameter, and precise control of pulling speed. It covers all operating conditions, including stable temperature, temperature drop, and long point disappearance, without the need for manual intervention or machine shutdown for calibration, thus significantly improving the continuity and stability of the crystal pulling process. Attached Figure Description

[0031] Figure 1 A schematic diagram illustrating the long-point phenomenon in crystal germination.

[0032] Figure 2 Schematic diagram of a crystal growth point-free phenomenon.

[0033] Figure 3 A schematic diagram illustrating the actual control effect of traditional detection methods on crystal diameter.

[0034] Figure 4 This is a schematic diagram illustrating the actual control effect of the crystal diameter in this detection method.

[0035] Figure 5 The attached figure shows the comparison data of the crystal diameter in specific embodiment 1.

[0036] Figure 6 The attached figure shows the comparison data of the crystal diameter in specific embodiment 2.

[0037] Figure 7 The attached figure shows the comparison data of the crystal puller diameter in specific embodiment 3.

[0038] Figure 8 The accompanying figure shows the comparison data of the crystal diameter in specific embodiment 4.

[0039] Figure 9 The attached figure shows the comparison data of the crystal diameter in specific embodiment 5. Detailed Implementation

[0040] The following describes the embodiments and appendices. Figures 1-9 The technical solution of the present invention will be further described below.

[0041] Example 1

[0042] A method for detecting the diameter of a seed crystal, comprising:

[0043] S1: In the image of the crystal pulling process, a crystal diameter detection area is set; the crystal pulling process is the process of drawing according to the target diameter after the diameter reduction is completed in the crystal pulling stage.

[0044] S2: Extract the outer contour feature point set of the bright part of the pilot aperture within the detection area; use a feature extraction algorithm to perform circle fitting on the outer contour feature point set to obtain the fitted circle diameter in pixels; based on the preset conversion coefficient between pixel diameter and physical diameter, convert the fitted circle diameter into a physical diameter in millimeters, denoted as D0.

[0045] S3: Standardize the image within the detection area; input the standardized image into a pre-trained classification model, and output the long point similarity determination result through the classification model, denoted as A1;

[0046] S4: The mean filter is used to process the similarity determination results A1 of multiple consecutive long points to obtain the probability that there are long points in the current pilot aperture, which is denoted as A2.

[0047] S5: Based on the preset upper limit of diameter offset, combined with the physical diameter D0 and the probability of long point existence A2, the compensated crystal diameter is calculated.

[0048] In S2, the feature extraction algorithm is the least squares method; the conversion factor between the pixel diameter and the physical diameter is the ratio of the physical value of the actual pilot diameter to the value of the fitted circle pixel diameter corresponding to the actual pilot diameter.

[0049] In S3, the specific process of the standardization process includes: scaling and cropping the image within the detection area by a specified scale using the center point of the lead-in aperture as a reference; performing grayscale stretching on the cropped image to make the image size and brightness reach a consistent standard; the pre-trained classification model is a convolutional neural network classification model, which is trained under supervision in the pre-training stage using labeled images with two categories: those with long points and those without; the output rule of the long point similarity judgment result A1 is: if the classification model determines that the image does not have long points, then A1=0; if the classification model determines that the image has long points, then A1 is the confidence level corresponding to the judgment result, and the value range of A1 is 0.5-1.

[0050] In S4, the specific operation of the mean filtering includes: continuously collecting multiple long point similarity judgment results A1; summing all collected A1 and taking the arithmetic mean as the long point existence probability A2; the mean filtering is used to smooth the random fluctuation of A1 data, and the value range of A2 is 0-1.

[0051] In S5, the preset upper limit of diameter offset is denoted as D2, and the value of D2 is 5%-10% of the target diameter of the crystal puller. D2 is determined according to the on-site single crystal silicon pulling process standard. The compensated crystal puller diameter is denoted as D1, and D1 is calculated by the formulas D1=D0 and D2×A2.

[0052] It also includes S6: transmitting the compensated lead diameter D1 to the PLC control system; the PLC control system compares D1 with the preset lead target diameter to stabilize the lead diameter within the target range.

[0053] The PLC control system adopts a PID control strategy, which drives the crystal lifting mechanism to adjust the crystal pulling speed according to the changing trend of D1 under different working conditions. The specific process is as follows:

[0054] When the temperature of the thermal field is stable or rising: the long point detection results A1=0, A2=0, D1=D0, and the PLC controls the pulling speed to remain stable, so that the lead diameter is consistent with the target diameter;

[0055] When the thermal field temperature fluctuates downwards: long spots appear in the crystal pulling aperture, and the detection result A1 of the long spots increases accordingly, which in turn leads to an increase in A2, ultimately causing D1 to show a sudden increase trend; after the PLC control system detects the upward trend of D1, it rapidly increases the crystal pulling speed.

[0056] When the long spot phenomenon subsides: the long spot detection result A1 gradually decreases to 0, A2 also gradually decreases to 0, and D1 shows a downward trend; after the PLC control system detects the downward trend of D1, it gradually reduces the crystal pulling speed.

[0057] Specific Implementation Example 1 (corresponding to) Figure 5 (Comparison of diameter prediction 2 and control 1)

[0058] This embodiment focuses on the pulling scenario of single-crystal silicon with a target diameter of 5mm, verifying the diameter control accuracy of the method of the present invention (prediction 2) compared with the traditional method (control 1). The basic experimental parameters are as follows:

[0059] Target diameter for crystal growth: 5mm;

[0060] Diameter offset upper limit D2: Take 5% of the target diameter, i.e., 0.25mm;

[0061] Pixel and physical diameter conversion factor: determined to be 0.1mm / px through calibration (a physical diameter of 5mm corresponds to a fitted circle pixel diameter of 50px).

[0062] Pre-trained classification model: A CNN model is used to train on labeled images with / without long points;

[0063] Mean filtering window: Collect the A1 value of 5 consecutive frames of images and take the average value as A2.

[0064] Experimental procedure and data recording:

[0065] 1. S1-S2 (Basic Diameter Detection): The central area of ​​the lead-in aperture is set as the detection area. The outer contour is fitted by the least squares method to obtain D0 stable at 5.0±0.1mm (corresponding to a pixel diameter of 50±1px).

[0066] 2. Operating Condition 1: Stable temperature in the thermal field (0-100s)

[0067] There is no long point phenomenon. The classification model outputs A1=0, and after mean filtering, A2=0.

[0068] The compensation diameter D1 = D0 and 0.25 × 0 = 5.0 ± 0.1 mm;

[0069] The PLC-PID control stabilized the pulling speed between 4.5 mm / min and 5.5 mm / min, predicting a diameter fluctuation of ±0.1 mm for the two groups; the control group (traditional method) showed no abnormalities, with a diameter fluctuation of ±0.1 mm as well.

[0070] 3. Operating Condition 2: The temperature of the thermal field fluctuates downward (101-120s).

[0071] A long point appears at 101s, and the classification model outputs A1=0.6 (101s), 0.8 (103s), and 0.9 (105s). After mean filtering, A2 increases from 0 to 0.78 (105s).

[0072] The compensation diameter D1 = 5.0 and 0.25 × 0.78 ≈ 5.195 mm;

[0073] The PLC detected an increase in D1, and at 105s, the drawing speed was increased from 4.5mm / min~5.5 mm / min to 8.5mm / min~9.5mm / min. At this time, the control group did not detect the long point by the traditional method (only the outer aperture was measured), and D0 still showed 5.0mm, so the drawing speed was not adjusted.

[0074] 4. Operating Condition 3: Long-term decline (121-140s)

[0075] The long point at 121s begins to fade, with A1 decreasing to 0.5 (121s), 0.2 (125s), and 0 (130s), and A2 decreasing to 0;

[0076] When D1 dropped to 5.0mm, the PLC gradually reduced the drawing speed to 2.5mm / min~3.5mm / min; in control group 1, the expansion of the outer aperture (D0 increased to 5.3mm) was only detected at 125s, and the drawing speed was then increased to 8.5mm / min~9.5mm / min, with a lag fluctuation in diameter.

[0077] Comparison results (corresponding) Figure 5 )

[0078] Predicted for Group 2: The diameter fluctuation range throughout the entire process is 5.0±0.2mm, with no obvious lag;

[0079] Control group 1: Diameter fluctuation range 5.0±0.3mm, with a control lag of 10-15s after temperature fluctuation, and the maximum diameter deviation reached 0.3mm.

[0080] Conclusion: When the thermal field temperature began to fluctuate downwards at 101s, our method identified the long point (A2=0.78), calculated the compensation diameter (D1≈5.195mm), and adjusted the drawing speed within 105s, with a lag time of only 4s. The traditional method, failing to detect the long point, only detected the outer aperture expansion (D0=5.3mm) at 125s, with a lag time of 24s, representing an 83% reduction in lag time. The overall diameter fluctuation range was 5.0±0.2mm, with no deviation exceeding the tolerance; the traditional method's fluctuation range was 5.0±0.3mm, with a maximum deviation of 0.3mm (6% exceeding the target diameter), resulting in a 33% improvement in accuracy.

[0081] Specific Implementation Example 2 (corresponding) Figure 6 (Comparison of diameter prediction 1 and control 5)

[0082] This embodiment addresses scenarios with high process requirements (diameter fluctuation ≤ ±0.15mm). The target diameter for chip extraction remains 5mm, and D2 is adjusted to 8% of the target diameter to highlight the sensitivity of prediction group 1. The basic experimental parameters are as follows:

[0083] Target diameter for crystal growth: 5mm;

[0084] Diameter offset upper limit D2: Take 8% of the target diameter, i.e., 0.4mm;

[0085] Pixel and physical diameter conversion factor: 0.1mm / px (same as Example 1);

[0086] Mean filtering window: 3 frames (more sensitive to long point changes);

[0087] Comparison Group 5: Typical cases of low pixel aperture measurement accuracy in traditional methods (fitting only 10 feature points of the outer aperture).

[0088] Experimental process and core comparative data

[0089] Temperature fluctuation trigger (at 80s): The temperature of the thermal field drops slightly at 80s, and a slight long point appears at 81s. The classification model outputs A1=0.55 (at 81s), and after mean filtering, A2=0.55 (there is no other data within 3 frames, so A1 is taken directly).

[0090] Predicting that D1 = 5.0 and 0.4 × 0.55 = 5.22 mm for group 1, the PLC will increase the pulling speed from 4.5 mm / min ~ 5.5 mm / min to 8.5 mm / min ~ 9.5 mm / min in 82 seconds;

[0091] Compared with the control group 5, which had fewer feature points and lower fitting accuracy, D0 still showed 5.0mm within 81-90s, and the pulling speed remained unchanged.

[0092] At the peak value at 90s: A1 rises to 0.9, A2=0.9, D1=5.0 and 0.4×0.9=5.36mm, and the PLC pulling speed increases to 8.5mm / min~9.5 mm / min; in the control group 5, D0 is only detected to rise to 5.2mm at this time, and the pulling speed begins to increase (lagging by 8s).

[0093] Final fluctuation comparison:

[0094] Predicted Group 1: Diameter fluctuation 5.0±0.15mm, lag time ≤2s;

[0095] Control group 5: Diameter fluctuation 5.0±0.28mm, hysteresis time ≥8s (corresponding to...) Figure 6 (Predicting early response in 1, and comparing the delayed curve characteristics in 5).

[0096] Conclusion: The thermal field cools slightly after 80 seconds, and a faint long point (A1=0.55) appears after 81 seconds. This method identifies the long point within 1 frame and outputs A2=0.55 within 3 frames, with a compensation diameter D1=5.22mm. The pulling speed adjustment is completed in 82 seconds, and the entire process from the appearance of the long point to the adjustment takes only 1 second, demonstrating sensitivity far exceeding traditional methods. The diameter fluctuation throughout the process is 5.0±0.15mm, fully meeting high-standard manufacturing processes.

[0097] Specific Implementation Example 3 (corresponding to) Figure 7 (Comparison of diameter prediction 2 and control 4)

[0098] This embodiment simulates a scenario of repeated temperature fluctuations in a thermal field to verify the anti-interference capability of the proposed method (prediction 2). The four control groups are cases in the traditional method that have not undergone filtering and are susceptible to random noise. The basic experimental parameters are as follows:

[0099] Target diameter for crystal growth: 5mm;

[0100] Diameter offset upper limit D2: Take 10% of the target diameter, i.e., 0.5mm;

[0101] Mean filtering window: 5 frames (2 groups of pre-judgment); 4 groups of comparison without filtering, diameter is calculated directly based on single aperture pixel value;

[0102] Experimental period: 200s, during which the temperature fluctuated downward twice at 100s and 140s (the duration of the longer periods was 15s and 10s respectively).

[0103] Batch comparison data (corresponding) Figure 7 (See Table 1 below)

[0104]

[0105] Conclusion: During the two temperature fluctuations, the A2 value of this method transitioned smoothly (from 0, 0.85, 0, then 0, 0.7, 0), and the compensation diameter D1 did not experience sudden increases or decreases. The total fluctuation was 5.0 ± 0.18 mm, with no instances of deviation due to fluctuations. The traditional method (control 4, without filtering) was affected by furnace dust and noise, resulting in an additional ± 0.05 mm fluctuation, leading to a total deviation of 5.0 ± 0.25 mm. Furthermore, a misjudgment occurred during the second fluctuation (treating noise as a long point), causing a deviation of 0.1 mm / min in the casting speed adjustment.

[0106] Specific Implementation Example 4 (corresponding) Figure 8 (Comparison of diameter prediction 2 and control 3)

[0107] This embodiment focuses on temperature stability testing during the later stages of die pulling (at a pulling length of 200mm). The three control groups represent typical cases in traditional methods that rely solely on the outer aperture pixel values, ignoring the early signal at the long point. The basic experimental parameters are as follows:

[0108] Target diameter for crystal growth: 5mm;

[0109] Diameter offset limit D2: 0.3mm (6% of the target diameter);

[0110] Drawing length: from 180mm to 230mm (covering the critical 200mm node in the later stage of lead generation);

[0111] Temperature fluctuation: A downward temperature fluctuation occurs at 190s, with the longest point lasting 12s.

[0112] Key experimental data

[0113] 190s temperature fluctuation:

[0114] Two sets of predictions were made: at 191s, A1=0.6, A2=0.6 (average of 5 frames), D1=5.0 and 0.3×0.6=5.18mm were detected, and the PLC increased the pulling speed to 8.5mm / min~9.5 mm / min at 192s;

[0115] Compared with the three groups: at 195s (4s after the long point appeared), no change in D0 was detected (the outer aperture did not expand). At 202s, the outer aperture suddenly expanded, and D0 rose to 5.3mm, which allowed the pulling speed to be increased from 4.5mm / min~5.5 mm / min to 8.5mm / min~9.5mm / min (10s lag).

[0116] 202s long point fades:

[0117] Predicted for two groups: A1 decreases to 0.2, A2=0.2, D1=5.06mm, and the pulling speed decreases to 2.5mm / min~3.5 mm / min;

[0118] Compared to the control group 3, it took 207 seconds to detect D0 dropping back to 5.1 mm, and the pulling speed dropped to 2.5 mm / min~3.5 mm / min (with a lag of 5 seconds).

[0119] Comparison results (corresponding) Figure 8 )

[0120] Predicted for two groups: the diameter of the 200mm key node is 5.02mm, with a fluctuation of ±0.18mm throughout.

[0121] Compared with the three control groups, the diameter of the critical node with a diameter of 200mm was 5.25mm, with a fluctuation of ±0.3mm, and there was one brief out-of-tolerance (5.3mm > target diameter and 0.2mm).

[0122] Conclusion: With a temperature fluctuation of 190s, this method adjusts the pulling speed at 192s, achieving a critical node diameter of 5.02mm at 200mm (with a deviation of only 0.02mm); the traditional method adjusts the pulling speed at 202s, achieving a node diameter of 5.25mm at 200mm (with a deviation of 0.25mm), resulting in a 92% improvement in critical node accuracy. During the later stages of crystal pulling (180-230mm), the diameter fluctuation of this method remained consistently ≤±0.18mm, while the traditional method exhibited fluctuations ≥±0.3mm, with one brief instance of exceeding the tolerance (5.3mm), requiring a shutdown for calibration. This method eliminates the need for shutdown, resulting in superior process continuity.

[0123] Specific Implementation Example 5 (corresponding to) Figure 9 (Comparison of diameter prediction 2 and control 2)

[0124] This embodiment is a batch production simulation experiment, in which two single-crystal silicon rods are pulled simultaneously (one using the pre-judgment method 2, and the other using the control method 2) to verify the batch applicability of this method. The control group 2 is a case of a standardized process in the traditional method (but without long spot detection). The basic experimental parameters are as follows:

[0125] Target diameter for crystal deriving: 5mm (set uniformly for both rods);

[0126] Diameter offset upper limit D2: 0.35mm (7% of the target diameter, a common value for mass production);

[0127] Experiment duration: 300s, during which a large-scale temperature fluctuation occurred at 200s (both rods were affected simultaneously, with the longer fluctuation lasting 18s).

[0128] Batch comparison data (corresponding) Figure 9 (See Table 2 below)

[0129]

[0130] Conclusion: The diameter deviation of rod 1 controlled by this method is ≤0.02mm throughout the entire process, and the final diameter is 5.01mm after 300s; the consistency deviation of rod 2 controlled by the traditional method is ≥0.1mm, and the final diameter is 5.12-5.15mm, improving batch consistency by 80%. This method does not require individual parameter calibration for each rod (D2=0.35mm is universal).

[0131] In summary, all five embodiments validated the core logic of this method, which uses CNN to identify long points (capturing temperature fluctuation signals in advance), mean filtering to smooth interference, diameter compensation to quantize the impact, and PLC to adjust the pulling speed in advance. This completely breaks through the limitations of traditional methods that only detect the outer aperture and passively correct. The lag time of traditional methods is generally 8-24s, while the lag time of this method is ≤4s, with an average lag reduction of more than 80%.

[0132] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A seed diameter detection method characterized by comprising: include: S1: In the image of the crystal pulling process, a crystal diameter detection area is set; the crystal pulling process is the process of drawing according to the target diameter after the diameter reduction is completed in the crystal pulling stage. S2: Extract the set of feature points on the outer contour of the bright part of the pilot aperture within the detection area; A feature extraction algorithm is used to fit the outer contour feature point set to a circle to obtain the diameter of the fitted circle in pixels; based on a preset conversion factor between pixel diameter and physical diameter, the fitted circle diameter is converted into a physical diameter in millimeters, denoted as D0. S3: Standardize the image within the detection area; The standardized image is input into a pre-trained classification model, and the classification model outputs a long-point similarity determination result, denoted as A1. S4: The mean filter is used to process the similarity determination results A1 of multiple consecutive long points to obtain the probability that there are long points in the current pilot aperture, which is denoted as A2. S5: Based on the preset upper limit of diameter offset, combined with the physical diameter D0 and the probability of long point existence A2, the compensated crystal diameter is calculated.

2. The method for detecting the diameter of a crystal puller according to claim 1, characterized in that, In S2, the feature extraction algorithm is the least squares method; the conversion factor between the pixel diameter and the physical diameter is the ratio of the physical value of the actual pilot diameter to the value of the fitted circle pixel diameter corresponding to the actual pilot diameter.

3. The method for detecting the diameter of a crystal driver according to claim 1, characterized in that, In S3, the specific process of the standardization process includes: scaling and cropping the image within the detection area by a specified scale using the center point of the lead-in aperture as a reference; performing grayscale stretching on the cropped image to make the image size and brightness reach a consistent standard; the pre-trained classification model is a convolutional neural network classification model, which is trained in the pre-training stage using labeled images with two categories: "existing long points" and "not existing long points"; the output rule of the long point similarity judgment result A1 is: if the classification model determines that the image "does not exist," then A1=0; if the classification model determines that the image "exists," then A1 is the confidence level corresponding to the judgment result, and the value range of A1 is 0.5-1.

4. The method for detecting the diameter of a crystal puller according to claim 1, characterized in that, In S4, the specific operation of the mean filtering includes: continuously collecting multiple long point similarity judgment results A1; summing all collected A1 and taking the arithmetic mean as the long point existence probability A2; the mean filtering is used to smooth the random fluctuation of A1 data, and the value range of A2 is 0-1.

5. The method for detecting the diameter of a crystal driver according to claim 1, characterized in that, In S5, the preset upper limit of diameter offset is denoted as D2, and the value of D2 is 5%-10% of the target diameter of the crystal puller. D2 is determined according to the on-site single crystal silicon pulling process standard. The compensated crystal puller diameter is denoted as D1, and D1 is calculated by the formula D1=D0+D2×A2.

6. The method for detecting the diameter of a crystal driver according to claim 1, characterized in that, It also includes S6: transmitting the compensated lead diameter D1 to the PLC control system; the PLC control system compares D1 with the preset lead target diameter to stabilize the lead diameter within the target range.

7. The method for detecting the diameter of a crystal driver according to claim 6, characterized in that, The PLC control system adopts a PID control strategy, which drives the crystal lifting mechanism to adjust the crystal pulling speed according to the changing trend of D1 under different working conditions. The specific process is as follows: When the temperature of the thermal field is stable or rising: the long point detection results A1=0, A2=0, D1=D0, and the PLC controls the pulling speed to remain stable, so that the lead diameter is consistent with the target diameter; When the temperature of the thermal field fluctuates downward: long spots appear in the crystal-leading aperture, the detection result A1 of the long spots increases accordingly, which in turn leads to an increase in A2, and finally causes D1 to show a sudden increase trend. After the PLC control system detects the upward trend of D1, it quickly increases the crystal pulling speed. When the long spot phenomenon subsides: the long spot detection result A1 gradually decreases to 0, A2 also gradually decreases to 0, and D1 shows a downward trend; after the PLC control system detects the downward trend of D1, it gradually reduces the crystal pulling speed.

Images