Real-time AI automatic identification shouldering opening detection method for single crystal furnace

Abstract

The invention relates to a real-time AI automatic identification shouldering opening detection method for a single crystal furnace, which solves the problems of low measurement precision, complex operation, easy influence by human factors and the like in the prior art, and adopts the technical scheme that the method comprises the following steps of: obtaining an image of a shouldering stage in a crystal growth process in real time to obtain an original image data set; based on the original image data set, labels are set according to a shouldering opening and a shouldering non-opening, and a marked expanded image data set is obtained; dividing the expanded image data set into a training sample set and a test sample set; constructing an AI network model; inputting the training sample set into an AI network model, and training the model to obtain a pre-training weight; and adjusting the AI network model based on the pre-training weight. The method has the advantages that the detection accuracy is improved; the detection real-time performance can be enhanced; the detection operation process can be simplified; the single crystal production efficiency can be improved; and safe production can be realized.

Description

Technical Field

[0001] The present invention relates to the field of shoulder opening detection in a crystal pulling process, and in particular to a real-time AI automatic recognition shoulder opening detection method for a single crystal furnace. Background Art

[0002] Crystal pulling refers to melting high-purity silicon raw materials at high temperature and immersing single-crystal silicon seed crystals into the melt. With the control of temperature and the pulling of the seed crystals, the melt gradually solidifies around the seed crystals and grows into a complete single-crystal silicon rod. This method can produce high-purity, high-quality single-crystal silicon materials, which are widely used in semiconductors and solar energy.

[0003] In the crystal pulling process of semiconductor materials, the principle of shoulder release is mainly based on the control of thermal field and temperature gradient. In the single crystal furnace, by accurately adjusting the heating power and crystal pulling speed, the solidification rate of the silicon melt can be controlled, thereby achieving the gradual expansion of the crystal diameter. The main function of shoulder release is to ensure the uniformity and quality of single crystal silicon growth. The shoulder release opening is used to detect whether the shoulder release process is properly controlled, which is directly related to the quality of single crystal silicon. However, most of the existing shoulder release opening measurement methods are based on manual observation or simple mechanical measurement, and there are problems such as low measurement accuracy, complex operation, and susceptibility to human factors. Therefore, being able to accurately detect the shoulder release opening is the key to ensuring the quality of single crystal silicon. The development of an efficient, accurate, and automatic shoulder release opening detection method and its special device has important practical value. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems existing in the prior art and provide a real-time AI automatic recognition shoulder opening detection method for a single crystal furnace

[0005] The above technical objectives of the present invention are mainly solved by the following technical solutions: A real-time AI automatic identification shoulder opening detection method for a single crystal furnace, characterized in that:

[0006] The image of the shoulder release stage during the crystal growth process is obtained in real time to obtain the original image data set;

[0007] Based on the original image dataset, labels are set according to shoulder opening and shoulder non-opening to obtain an expanded image dataset with good labels;

[0008] Divide the augmented image dataset into a training sample set and a test sample set;

[0009] Build AI network models;

[0010] Input the training sample set into the AI ​​network model, train the model, and obtain pre-trained weights;

[0011] Adjust the AI ​​network model based on pre-trained weights.

[0012] As a further improvement and supplement to the above technical solution, the present invention adopts the following technical measures: a process of generating an expanded image data set: using marking software to mark the acquired original image, marking the target in the image with a prediction box, and enhancing the original image data set to increase the amount of data in the original image data set to obtain an expanded image data set, wherein the target includes shoulder opening and shoulder non-opening.

[0013] Preferably, when training the model, reasoning is required, including approximate detection: first determine the prediction confidence of each label. If the prediction confidence exceeds the set threshold, it is considered that there is a target in the corresponding label, and the approximate location of the target is obtained.

[0014] Preferably, after completing the rough detection step, a post-processing step is performed: the prediction boxes with existing targets are screened according to the non-maximum suppression algorithm, and the repeated prediction boxes corresponding to the same target are eliminated to remove redundant prediction boxes, retain the most likely prediction box, and obtain the prediction result.

[0015] Preferably, based on the classification probability of the filtered prediction box, the index corresponding to the maximum probability is taken as the classification index number of the target, thereby obtaining the category of the target, which is used to determine whether the target is shoulder-open or shoulder-unopen.

[0016] As a preference, a function is selected to fuse the convolutional layer and the BN layer in the AI ​​network model;

[0017] After training is completed, if the BN layer mean and BN layer variance in the inference phase and the mean and variance of the training set are fixed values, that is, constants, the determination function is

[0018] Where: μ is the mean of the BN layer; σ 2 is the variance of the BN layer; w is the weight of the convolution kernel; b is the bias of the convolution kernel; ∈ is a constant used to prevent the divisor from being 0; γ is the expansion coefficient of the BN layer; β is the offset coefficient of the BN layer;

[0019] When training the model, γ and β are learnable parameters;

[0020] The loss function used by the AI ​​network model is the CIoU loss function, that is

[0021] Where:

[0022] CIoU stands for Complete Intersection over Union, which means complete intersection over union ratio and CIoU loss;

[0023] IoU stands for Intersection over Union, which means intersection over union ratio and IoU loss;

[0024] Represents the center point distance loss, which is the distance between the center point of the predicted box and the real box.

[0025] αv represents the aspect ratio loss, which considers the difference in aspect ratio between the predicted box and the real box;

[0026] ρ(b,b gt ) is the Euclidean distance between the center points of the predicted box and the true box;

[0027] c is the diagonal length of the minimum enclosing box that contains the predicted box and the true box;

[0028] α is a balance parameter used to balance IoU loss and aspect ratio loss;

[0029] v is a measure of the difference in aspect ratio between the predicted box and the true box.

[0030] Preferably, when training the model, the training set data is input, the gradient error is calculated during forward propagation, the AI ​​network model parameters are updated using the gradient descent method during back propagation, the residual learning mechanism is used to establish a mapping relationship between the input image and the label data, and the loss function is used to gradually optimize to the minimum; the test set data is input, and the hyperparameters of the AI ​​network model are fine-tuned.

[0031] Preferably, when training the model, a corresponding number of training times is set. As the number of training times increases, the loss function curve of the AI ​​network model gradually converges. When the loss function curve converges and stabilizes, the AI ​​network model is trained to the optimal state and a pre-trained weight file is generated.

[0032] Preferably, after each round of training and reasoning, the performance of the AI ​​network model is detected using a test sample set, and the learning rate or corresponding hyperparameters of the AI ​​network model are adjusted according to the test results to obtain and save the test training weight file.

[0033] Preferably, the detecting step comprises:

[0034] Load the pre-trained weight file;

[0035] When reaching the shoulder release process, the shoulder release image is collected in real time;

[0036] Then the image is sent to the loaded AI network model, and the AI ​​network model preprocesses the image;

[0037] The preprocessing results are passed into the backbone network of the AI ​​network model to extract image features;

[0038] Perform reasoning, filter multiple prediction boxes of the same target, and retain the best prediction box;

[0039] Then post-process the prediction results, including label smoothing and IoU threshold adjustment;

[0040] The result is output, and then after multiple tests, it is determined whether the shoulder is open or not.

[0041] The present invention has the beneficial effects:

[0042] Helps improve detection accuracy:

[0043] The industrial camera installed in the crystal pulling furnace captures real-time images of the crystal pulling process, and uses the AI ​​network model for image recognition to achieve accurate detection of the shoulder opening.

[0044] Combined with multiple tests and calculations, the shoulder opening is obtained, which effectively reduces the error that may be caused by a single measurement and improves the overall measurement accuracy;

[0045] Helps to enhance real-time performance:

[0046] It can obtain the image data of the shoulder release process in real time, process and analyze it, so as to realize the real-time monitoring and adjustment of the shoulder release opening;

[0047] It helps to timely discover and solve problems that may occur during the shoulder release process, ensuring the stability of the production process and product quality;

[0048] Helps to simplify the operation process:

[0049] It realizes automated measurement and reduces the dependence on operators;

[0050] The device has a simple structure and is easy to operate, which reduces the frequency of operators entering the production workshop and the frequency of operating near the crystal pulling furnace, and reduces the difficulty of operation and labor intensity;

[0051] Helps improve production efficiency:

[0052] Automated measurement reduces human intervention and improves production efficiency;

[0053] Real-time monitoring and adjustment help optimize the crystal pulling process, reduce raw material waste and crystal rod scrap rate, and improve product qualification rate and production efficiency;

[0054] Conducive to achieving safe production:

[0055] Reduce the direct contact between operators and dangerous environments such as high temperature and high pressure, thus reducing safety risks;

[0056] Accurate detection of the shoulder opening helps prevent material waste caused by an overly large shoulder diameter or ingot scrap or degradation caused by a too small shoulder diameter, further improving production safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 The invention relates to a flow chart of a detection method.

[0058] Figure 2 It is a schematic image of a crystal rod without shouldering involved in the present invention.

[0059] Figure 3 It is a schematic diagram of the crystal rod shouldering involved in the present invention. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0061] Example: Figure 1-Figure 3 As shown, in Figure 3 In the figure, the boxed part is the shoulder of the crystal rod.

[0062] A real-time AI automatic recognition and shoulder opening detection method for a single crystal furnace:

[0063] The image of the shoulder release stage during the crystal growth process is obtained in real time by an industrial camera to obtain the original image data set;

[0064] Based on the original image dataset, labels are set according to shoulder opening and shoulder non-opening to obtain an expanded image dataset with good labels;

[0065] Divide the augmented image dataset into a training sample set and a test sample set;

[0066] Build an AI network model, that is, build an improved YOLOv5 network model;

[0067] Input the training sample set into the AI ​​network model, train the model, and obtain pre-trained weights;

[0068] Adjust the AI ​​network model based on pre-trained weights.

[0069] In practical applications, the process of generating an extended image data set is as follows: the acquired original image is marked using marking software, the target in the image is marked with a prediction box, and the original image data set is enhanced to increase the amount of data in the original image data set to obtain an extended image data set, wherein the targets include shoulder opening and shoulder non-opening.

[0070] In practical applications, when training the model, reasoning is required, including approximate detection: first determine the prediction confidence of each label. If the prediction confidence exceeds the set threshold, it is considered that there is a target in the corresponding label, and the approximate location of the target is obtained.

[0071] In practical applications, after completing the rough detection step, the post-processing step is performed: the prediction boxes with targets are screened according to the non-maximum suppression algorithm, and the repeated prediction boxes corresponding to the same target are eliminated to remove redundant prediction boxes, retain the most likely prediction box, and obtain the prediction result.

[0072] In practical applications, according to the classification probability of the filtered prediction box, the index corresponding to the maximum probability is taken as the classification index number of the target, so as to obtain the category of the target, that is, to determine whether the target is shoulder-open or shoulder-unopen.

[0073] In practical applications, in order to improve the measurement accuracy, a function is selected to fuse the convolutional layer and the BN layer in the AI ​​network model;

[0074] After training is completed, if the BN layer mean and BN layer variance in the inference phase and the mean and variance of the training set are fixed values, that is, constants, the determination function is

[0075] Where: μ is the mean of the BN layer; σ 2 is the variance of the BN layer; w is the weight of the convolution kernel; b is the bias of the convolution kernel; ∈ is a constant used to prevent the divisor from being 0; γ is the expansion coefficient of the BN layer; β is the offset coefficient of the BN layer;

[0076] When training the model, γ and β are learnable parameters;

[0077] The loss function used by the AI ​​network model is the CIoU loss function, that is

[0078] Where:

[0079] CIoU stands for Complete Intersection over Union, which means complete intersection over union ratio and CIoU loss;

[0080] IoU stands for Intersection over Union, which means intersection over union ratio and IoU loss;

[0081] Represents the center point distance loss, which is the distance between the center point of the predicted box and the real box.

[0082] αv represents the aspect ratio loss, which considers the difference in aspect ratio between the predicted box and the real box;

[0083] ρ(b,b gt ) is the Euclidean distance between the center points of the predicted box and the true box;

[0084] c is the diagonal length of the minimum enclosing box that contains the predicted box and the true box;

[0085] α is a balance parameter used to balance IoU loss and aspect ratio loss;

[0086] ν is a measure of the difference in aspect ratio between the predicted box and the true box.

[0087] In actual applications, when training the model, the training set data is input, the gradient error is calculated during forward propagation, the gradient descent method is used to update the AI ​​network model parameters during back propagation, the residual learning mechanism is used to establish the mapping relationship between the input image and the label data, and the loss function is used to gradually optimize to the minimum; the test set data is input to fine-tune the hyperparameters of the AI ​​network model.

[0088] In actual applications, when training the model, the corresponding number of training times is set. As the number of training times increases, the loss function curve of the AI ​​network model gradually converges. When the loss function curve converges and stabilizes, the AI ​​network model is trained to the optimal state and a pre-trained weight file is generated.

[0089] In actual applications, after each round of training and reasoning, the test sample set is used to detect the performance of the AI ​​network model, and the learning rate or corresponding hyperparameters of the AI ​​network model are adjusted according to the test results to obtain and save the test training weight file.

[0090] In practical applications, the detection step includes:

[0091] Load the pre-trained weight file;

[0092] When reaching the shoulder release process, the shoulder release image is collected in real time;

[0093] Then the image is sent to the loaded AI network model, and the AI ​​network model preprocesses the image;

[0094] The preprocessing results are passed into the backbone network of the AI ​​network model to extract image features;

[0095] Perform reasoning, filter multiple prediction boxes of the same target, and retain the best prediction box;

[0096] Then post-process the prediction results, including label smoothing and IoU threshold adjustment;

[0097] The result is output, and then after multiple tests, it is determined whether the shoulder is open or not.

[0098] Another technical subject involved in this technical solution: a device for measuring the diameter of a crystal rod in a crystal pulling furnace using a real-time AI automatic recognition and shoulder opening detection method for a single crystal furnace, comprising:

[0099] The image acquisition module is used to obtain a set of images to be tested; the first insulator image set includes an insulator body image and corresponding defect labels,

[0100] An image preprocessing module, used to adjust the size of the image;

[0101] The model training module inputs the training sample set into the improved YOLOv5 network model for training to obtain pre-training weights, and adjusts and optimizes the improved YOLOv5 network model according to the pre-training weights.

[0102] The model testing module uses the test set to evaluate the model performance after each round of training and inference, and adjusts the learning rate or other hyperparameters based on the evaluation results.

[0103] Model saving module, save the pre-training weight file.

[0104] Model loading module, loads pre-trained weight files, and performs inference on images;

[0105] Model output module, outputs the opening during the shoulder release process;

[0106] A controller module, used to receive the processing results output by the model and control the shooting parameters of the image acquisition device (such as an industrial camera);

[0107] A display module, used for connecting with the controller module and for displaying the measurement result of the seeding diameter;

[0108] The memory module is used for connecting with the controller module and storing the measurement results and other related data.

[0109] The present application provides a method, device and equipment for real-time AI automatic identification of shoulder opening detection of a single crystal furnace, the method comprising: obtaining a target image captured by an industrial camera; using the original image data set for annotation and training; and then using an AI model to identify the target image and output the result to improve the detection accuracy. The solution of the present application can detect shoulder opening more accurately.

[0110] Another technical subject involved in the technical solution: a readable medium having a computer program stored thereon, which, when executed by a processor, implements a real-time AI automatic identification and shoulder opening detection method for a single crystal furnace.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. In the above embodiments, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time AI automatic identification shoulder opening detection method for a single crystal furnace, characterized in that: The image of the shoulder release stage during the crystal growth process is obtained in real time to obtain the original image data set; Based on the original image dataset, labels are set according to shoulder opening and shoulder non-opening to obtain an expanded image dataset with good labels; Divide the augmented image dataset into a training sample set and a test sample set; Build AI network models; Input the training sample set into the AI ​​network model, train the model, and obtain pre-trained weights; Adjust the AI ​​network model based on pre-trained weights.

2. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 1 is characterized in that The process of generating an extended image data set is as follows: the original image is marked by using marking software, the target in the image is marked with a prediction box, and the original image data set is enhanced to increase the amount of data in the original image data set to obtain an extended image data set, wherein the targets include shoulder opening and shoulder non-opening.

3. The real-time AI automatic identification shoulder opening detection method of a single crystal furnace according to claim 1 or 2, characterized in that When training the model, reasoning is required, including approximate detection: first determine the prediction confidence of each label. If the prediction confidence exceeds the set threshold, it is considered that there is a target in the corresponding label, and the approximate location of the target is obtained.

4. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 3 is characterized in that After completing the rough detection step, the post-processing step is performed: the prediction boxes with targets are screened according to the non-maximum suppression algorithm, and the repeated prediction boxes corresponding to the same target are eliminated to remove redundant prediction boxes, retain the most likely prediction boxes, and obtain the prediction results.

5. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 4 is characterized in that According to the classification probability of the filtered prediction box, the index corresponding to the maximum probability is taken as the classification index number of the target, so as to obtain the category of the target, which is used to determine whether the target is shoulder-open or shoulder-unopen.

6. The real-time AI automatic identification shoulder opening detection method of a single crystal furnace according to claim 3 is characterized in that: Select a function to fuse the convolutional layer and the BN layer in the AI ​​network model; After training is completed, if the BN layer mean and BN layer variance in the inference phase and the mean and variance of the training set are fixed values, that is, constants, the determination function is Where: μ is the mean of the BN layer; σ 2 is the variance of the BN layer; w is the weight of the convolution kernel; b is the bias of the convolution kernel; ∈ is a constant used to prevent the divisor from being 0; γ is the expansion coefficient of the BN layer; β is the offset coefficient of the BN layer; When training the model, γ and β are learnable parameters; The loss function used by the AI ​​network model is the CIoU loss function, that is Where: CIoU stands for Complete Intersection over Union, which means complete intersection over union ratio and CIoU loss; IoU stands for Intersection over Union, which means intersection over union ratio and IoU loss; Represents the center point distance loss, which is the distance between the center point of the predicted box and the real box. αv represents the aspect ratio loss, which takes into account the difference in aspect ratio between the predicted box and the real box. ρ(b,b gt ) is the Euclidean distance between the center points of the predicted box and the true box; c is the diagonal length of the minimum enclosing box that contains the predicted box and the true box; α is a balance parameter used to balance IoU loss and aspect ratio loss; v is a measure of the difference in aspect ratio between the predicted box and the true box.

7. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 3 is characterized in that When training the model, input the training set data, calculate the gradient error during forward propagation, use the gradient descent method to update the AI ​​network model parameters during back propagation, use the residual learning mechanism to establish the mapping relationship between the input image and the label data, and use the loss function to gradually optimize to the minimum; input the test set data and fine-tune the hyperparameters of the AI ​​network model.

8. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 7 is characterized in that When training the model, set the corresponding number of training times. As the number of training times increases, the loss function curve of the AI ​​network model gradually converges. When the loss function curve converges and stabilizes, the AI ​​network model is trained to the optimal state and a pre-trained weight file is generated.

9. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 8 is characterized in that After each round of training and reasoning, the test sample set is used to detect the performance of the AI ​​network model, and the learning rate or corresponding hyperparameters of the AI ​​network model are adjusted according to the test results to obtain and save the test training weight file.

10. The real-time AI automatic identification shoulder opening detection method of the single crystal furnace according to claim 9 is characterized in that The detection step comprises: Load the pre-trained weight file; When reaching the shoulder release process, the shoulder release image is collected in real time; Then the image is sent to the loaded AI network model, and the AI ​​network model preprocesses the image; The preprocessing results are passed into the backbone network of the AI ​​network model to extract image features; Perform reasoning, filter multiple prediction boxes of the same target, and retain the best prediction box; Then post-process the prediction results, including label smoothing and IoU threshold adjustment; The result is output, and then after multiple tests, it is determined whether the shoulder is open or not.

Images