CN117468085B

CN117468085B - Crystal bar growth control method and device, crystal growth furnace system and computer equipment

Info

Publication number: CN117468085B
Application number: CN202311815614.0A
Authority: CN
Inventors: 傅林坚; 刘华; 李亮
Original assignee: Zhejiang Jingsheng Mechanical and Electrical Co Ltd; Zhejiang Qiushi Semiconductor Equipment Co Ltd
Current assignee: Zhejiang Jingsheng Mechanical and Electrical Co Ltd; Zhejiang Qiushi Semiconductor Equipment Co Ltd
Priority date: 2023-12-27
Filing date: 2023-12-27
Publication date: 2024-05-28
Anticipated expiration: 2043-12-27
Also published as: CN117468085A

Abstract

The application relates to a crystal bar growth control method, a crystal growth furnace system and computer equipment, wherein the method comprises the steps of acquiring a real-time image in the crystal growth furnace; inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result; if the detection result is abnormal, retrying Wen Lajie based on the target area image; the growth detection model comprises a self-adaptive contrast enhancement module, a main network, a neck network and a detection network which are connected in sequence; the self-adaptive contrast enhancement module enhances the real-time image to obtain an enhanced image; the trunk network extracts crystal characteristics based on the enhanced image; the backbone network also comprises an attention fusion module for enhancing crystal characteristics; the neck network fuses the crystal characteristics to obtain fused crystal characteristics; the detection network determines a target area image based on the fused crystal characteristics and a corresponding crystal bar growth detection result. By adopting the method, the accuracy and the safety of crystal bar growth judgment can be improved.

Description

Crystal bar growth control method and device, crystal growth furnace system and computer equipment

Technical Field

The application relates to the technical field of crystal bar growth, in particular to a crystal bar growth control method and device, a crystal growth furnace system and computer equipment.

Background

Sapphire is a crystalline material formed by melting and resolidifying aluminum and oxygen to form an alumina compound under specific temperature conditions. The sapphire in the seeding stage can comprise three steps of seed crystal lowering, liquid contact and temperature test drawknot.

In the process of temperature test drawknot of sapphire, need the manual work to add the window that gold-plated glass formed through electric welding glass, whether the temperature test drawknot step in the observation stove normally goes on, because the step duration is long, the manual work observes the crystal bar growth process, and on the one hand is difficult to observe the crystal bar growth phenomenon, leads to the yields of sapphire lower, on the other hand, under high temperature, strong bright environment, long-term observation easily causes the harm to the human eye, has the influence to operating personnel's personal safety.

Therefore, the problems of low accuracy of crystal bar growth judgment and low safety of operators in the temperature test and pulling process still exist in the prior art.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a method and apparatus for controlling ingot growth, a crystal growth furnace system, and a computer device that can improve the accuracy of ingot growth judgment and the safety of operators.

In a first aspect, the present embodiment provides a method for controlling growth of a crystal bar, the method comprising:

Acquiring a real-time image in the crystal growth furnace;

inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result;

If the detection result is abnormal growth of the crystal bar, re-performing Wen Lajie on the basis of the target area image;

the growth detection model comprises a self-adaptive contrast enhancement module, a main network, a neck network and a detection network which are connected in sequence;

the self-adaptive contrast enhancement module is used for enhancing the real-time image to obtain an enhanced image;

The backbone network is used for extracting crystal characteristics based on the enhanced image; the backbone network further comprises an attention fusion module for enhancing the crystal characteristics;

The neck network is used for fusing the crystal characteristics to obtain fused crystal characteristics;

The detection network is used for determining the target area image and a crystal bar growth detection result corresponding to the target area image based on the fused crystal characteristics.

In some of these embodiments, the detection network is further configured to:

determining the target area image based on the fused crystal characteristics;

determining the crystal thickness in the target area image based on the target area image;

if the crystal thickness in the target area image is within a preset thickness interval, the crystal bar growth detection result is that the crystal bar grows normally;

If the crystal thickness in the target area image is outside the preset thickness interval, the crystal bar growth detection result is abnormal crystal bar growth.

In some of these embodiments, the re-performing the temperature-trial drawknot based on the target region image includes:

Adjusting the power of the crystal growth furnace based on the crystal thickness in the target area image;

And after the temperature in the crystal growth furnace is stable, acquiring a real-time image in the crystal growth furnace again, and inputting the real-time image into the growth detection model.

In some embodiments, the inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image, and after the crystal growth detection result, further includes:

And if the detection result is that the crystal bar grows normally, storing the real-time image, and performing seed crystal purification and pulling operation.

In some of these embodiments, the adaptive contrast enhancement module comprises a plurality of convolution layers, a fully-connected layer, and a contrast enhancement unit connected in sequence,

The plurality of convolution layers are divided into a first convolution group and a second convolution group; the convolution layers sequentially convolve the real-time image, and when the convolution layers in the second convolution group convolve, the convolution results of the convolution layers of the matched first convolution group are spliced to obtain a fusion feature map;

the full-connection layer is used for determining preset parameters based on the fusion feature map;

The contrast enhancement unit is used for enhancing the contrast of the real-time image based on the preset parameters and a contrast enhancement algorithm.

In some embodiments, the attention fusion module includes a channel reorganization downsampling unit, a fusion unit, and an attention weighted fusion unit connected in sequence, wherein:

The channel reorganization downsampling unit is used for carrying out channel reorganization downsampling on the crystal characteristics extracted by the backbone network to obtain reorganized crystal characteristics;

The fusion unit is used for carrying out multiple convolutions on the recombined crystal characteristics and fusing characteristic graphs of the multiple convolutions to obtain convolution crystal characteristics;

The attention weighted fusion unit is used for weighting the fused crystal characteristics to obtain weighted crystal characteristics, and fusing the fused crystal characteristics with the weighted crystal characteristics to obtain reinforced crystal characteristics.

In some of these embodiments, the inputting the real-time image into the pre-trained growth detection model is preceded by:

acquiring an acquired training image set, wherein the training image set comprises a plurality of crystal bar growth abnormal images marked with a target area image by a Labelme tool and a plurality of crystal bar growth normal images;

dividing the training image set based on a preset proportion to obtain a training set, a verification set and a test set; the training set and the verification set comprise a plurality of abnormal crystal bar growth images, and the test set comprises a plurality of normal crystal bar growth images and abnormal crystal bar growth images;

and building a neural learning network, and training the neural learning network based on the training set, the verification set and the test set to obtain the growth detection model.

In a second aspect, the present embodiment provides a crystal growth control apparatus, the apparatus comprising:

The acquisition module is used for acquiring a real-time image in the crystal growth furnace;

The detection module is used for inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result;

the processing module is used for retrying Wen Lajie based on the target area image if the detection result is that the crystal bar growth is abnormal;

In a third aspect, the embodiment provides a crystal growth furnace system, which comprises a crystal growth furnace, an area array camera and an industrial personal computer, wherein an observation window is arranged on a furnace cover of the crystal growth furnace, the area array camera is arranged around the observation window through a camera fixing rod and is used for acquiring real-time images in the crystal growth furnace and transmitting the real-time images to the industrial personal computer; the industrial personal computer is used for realizing the steps of the method.

In a fourth aspect, the present embodiment provides a computer device comprising a memory storing a computer program and a processor implementing the steps of the method as described above when the processor executes the computer program.

According to the crystal bar growth control method, the crystal growth control device, the crystal growth furnace system and the computer equipment, real-time images in the crystal growth furnace are obtained; inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result; if the detection result is abnormal growth of the crystal bar, re-performing Wen Lajie on the basis of the target area image; the automatic detection of the growth of the crystal bar can be realized, the condition that operators need to observe through naked eyes is avoided, and the safety of the operators is improved. The growth detection model comprises a self-adaptive contrast enhancement module, a main network, a neck network and a detection network which are connected in sequence; the self-adaptive contrast enhancement module is used for enhancing the real-time image to obtain an enhanced image; the main network is used for extracting crystal characteristics based on the enhanced image; the main network further comprises an attention fusion module for enhancing the crystal characteristics, the crystal characteristics of the real-time image are enhanced through the modularized self-adaptive contrast enhancement module under the condition that the real-time image is only required to be input according to an original image format, and the attention fusion module is combined, so that the accuracy of the main network in extracting the crystal characteristics is higher, and the effect of improving the accuracy of crystal bar growth judgment is achieved.

Drawings

FIG. 1 is a diagram of an environment for an application of a method for controlling growth of an ingot in one embodiment;

FIG. 2 is a flow chart illustrating a method of controlling growth of a boule in one embodiment;

FIG. 3 is a flow chart illustrating a method of controlling growth of a boule in another embodiment;

FIG. 4 is a schematic diagram of a crystal feature in a real-time image according to one embodiment;

FIG. 5 is a block diagram of the ACEA-Yolov model in one embodiment;

FIG. 6 is a schematic diagram of an adaptive contrast enhancement module in one embodiment;

FIG. 7 is a graph of contrast before and after adaptive contrast enhancement of real-time images in one embodiment;

FIG. 8 is a block diagram of the structure of an attention fusion module in one embodiment;

FIG. 9 is a block diagram illustrating an apparatus for controlling growth of an ingot in one embodiment;

FIG. 10 is a block diagram of a crystal growth furnace system in one embodiment;

FIG. 11 is an internal block diagram of a computer device in one embodiment.

Detailed Description

The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.

The crystal bar growth control method provided by the embodiment of the application can be applied to an application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network. The data storage system may store data that the server 104 needs to process. The data storage system may be integrated on the server 104 or may be located on a cloud or other network server. The real-time image may be collected by storing the real-time image in the server 104 and forwarding the real-time image to the terminal 102, where the terminal 102 obtains the real-time image in the crystal growth furnace; inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result; and if the detection result is abnormal crystal bar growth, carrying out temperature test drawknot again based on the target area image. The terminal 102 may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, internet of things devices, and portable wearable devices. The portable wearable device may be a smart watch, smart bracelet, headset, or the like. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers.

In one embodiment, as shown in fig. 2, a method for controlling growth of a seed crystal is provided, and the method is applied to the terminal 102 in fig. 1, and is illustrated as an example, and includes the following steps:

and step S100, acquiring a real-time image in the crystal growth furnace.

In this embodiment, the crystal growth furnace may be an apparatus for producing a sapphire single crystal material. The real-time image in the crystal growth furnace can be acquired by image acquisition at the top of the crystal growth furnace.

And step 200, inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result.

The growth detection model can be obtained by training a neural learning network consisting of a self-adaptive contrast enhancement module, a main network, a neck network and a detection network which are connected in sequence; the self-adaptive contrast enhancement module is used for enhancing the real-time image to obtain an enhanced image; the backbone network is used for extracting crystal characteristics based on the enhanced image; the backbone network further comprises an attention fusion module for enhancing the crystal characteristics; the neck network is used for fusing the crystal characteristics to obtain fused crystal characteristics; the detection network is used for determining the target area image and a crystal bar growth detection result corresponding to the target area image based on the fused crystal characteristics.

The data set adopted by the training neural network can be based on a labeled training image set, a plurality of images in the training image set are labeled with target area images, and the target area images can be target images with abnormal crystal rod growth.

The adaptive contrast enhancement module may enhance the real-time image by a contrast enhancement algorithm. The adaptive contrast enhancement module may be configured to contrast enhance the real-time image by determining a gain parameter based on the real-time image and based on a contrast enhancement algorithm and the gain parameter. Illustratively, before the real-time image is enhanced, the adaptive contrast enhancement module performs convolution processing based on the real-time image to obtain a feature map of the real-time image, predicts a gain parameter of a contrast enhancement algorithm based on the feature map of the real-time image, substitutes the gain parameter into the contrast enhancement algorithm, and performs contrast enhancement processing on the real-time image, thereby realizing adaptive contrast enhancement of the real-time image.

The main network is used for extracting crystal characteristics of the enhanced image. The crystal may be a crystal formed around the ingot during the growth of the ingot, and the crystal feature may be a feature of the enhanced image that includes the crystal. And fusing the crystal characteristics by the neck network to obtain fused crystal characteristics.

The attention fusion module is used for enhancing the crystal characteristics, and the attention fusion module is used for combining the crystal characteristics extracted by the main network of the growth detection model at each stage to carry out interactive fusion on the crystal characteristics at the same stage, so that richer and more complete crystal characteristics are obtained. It can be understood that in the process of growth detection, the enhanced image itself will be processed into feature images with different scales, and in the process of downsampling, local features of the smaller region will be lost in the smaller region, and through the attention fusion module, richer and complete crystal feature information can be obtained based on feature images with different scales, and the recognition effect on small targets is improved.

And step S300, if the detection result is abnormal growth of the crystal bar, carrying out temperature test drawknot again based on the target area image.

If the detection result is that the crystal bar grows abnormally, the temperature test drawknot step is wrong, and the yield is possibly reduced, and then the temperature test drawknot is required to be retried after the parameters of the crystal growth furnace are adjusted.

According to the crystal bar growth control method provided by the embodiment, a real-time image in a crystal growth furnace is obtained; inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result; if the detection result is abnormal growth of the crystal bar, re-performing Wen Lajie on the basis of the target area image; the automatic detection of the growth of the crystal bar can be realized, the condition that operators need to observe through naked eyes is avoided, and the safety of the operators is improved. The growth detection model comprises a self-adaptive contrast enhancement module, a main network, a neck network and a detection network which are connected in sequence; the self-adaptive contrast enhancement module is used for enhancing the real-time image to obtain an enhanced image; the main network is used for extracting crystal characteristics based on the enhanced image; the main network further comprises an attention fusion module for enhancing the crystal characteristics, the crystal characteristics of the real-time image are enhanced through the modularized self-adaptive contrast enhancement module under the condition that the real-time image is only required to be input according to an original image format, and the attention fusion module is combined, so that the accuracy of the main network in extracting the crystal characteristics is higher, and the effect of improving the accuracy of crystal bar growth judgment is achieved.

In some of these embodiments, the detection network is further configured to:

determining the target area image based on the fused crystal characteristics;

The target area image is determined based on the fused crystal characteristics, and the actual position of the crystal is predicted according to the fused crystal characteristics, so that the target area image in the real-time image is obtained. Further, a plurality of prediction target frames are obtained based on the fused crystal characteristics, each prediction target frame corresponds to one area image in the real-time image, and the final target area image is determined by screening the prediction target frames.

The crystal thickness of the crystal bar is determined based on the target area image, the crystal position of the crystal bar can be determined according to the target area image, and the crystal thickness of the crystal bar can be determined by calculating the actual thickness of the crystal bar crystal according to the pixel diameter of the crystal bar in proportion.

And if the crystal thickness in the target area image is within the preset thickness interval, the crystal bar growth detection result is that the crystal bar grows normally. It can be understood that the preset thickness may not be a fixed value, but may also be an error allowable range of the preset thickness, and when the crystal thickness in the target area image is within the error allowable range of the preset thickness, the crystal thickness may be considered to be within the preset thickness interval. When the thickness of the crystal is within the preset thickness interval, the current temperature and the like in the crystal growth furnace are in ideal states, and the crystal rod grows normally.

If the crystal thickness in the target area image is outside the preset thickness interval, the crystal growth is slowed down or accelerated, and the crystal growth detection result is abnormal crystal growth.

According to the crystal bar growth control method, whether the crystal bar grows normally or not is determined through the thickness of crystals, and judgment of the growth condition of the crystal bar can be achieved.

The adjusting the power of the crystal growth furnace based on the crystal thickness in the target area image may be determining to increase or decrease the power of the crystal growth furnace according to the relation between the crystal thickness and the preset thickness. It can be appreciated that when the thickness of the crystal is greater than the preset thickness, which means that the solidification speed is greater than the ideal speed, the power is increased, so that the temperature in the furnace is increased; when the thickness of the crystal is smaller than the preset thickness, which means that the solidification speed is lower than the ideal speed, the power is required to be reduced, so that the temperature in the furnace is reduced.

And after the temperature in the crystal growth furnace is stable, re-acquiring a real-time image in the crystal growth furnace, inputting the real-time image into the growth detection model, wherein the real-time temperature measurement can be performed on the temperature in the furnace, judging whether the temperature tends to be stable or not based on a temperature measurement result, or calculating the time of temperature stabilization based on the adjusted power, re-acquiring the real-time image in the crystal growth furnace after the temperature stabilization time, and inputting the real-time image into the growth detection model.

According to the crystal bar growth control method, the power of the crystal growth furnace is adjusted based on the crystal thickness, so that the adjustment can be performed under the condition that the temperature in the crystal growth furnace is unsuitable, the most suitable seeding temperature is found, and the effect of improving the yield of the sapphire single crystal is achieved.

According to the crystal bar growth control method, by storing the real-time images under the condition that the crystal bar grows normally, real-time recording of the crystal bar growth process can be achieved, historical image support is provided for later judging whether the crystal bar grows or not, and judging accuracy of crystal bar growth is improved.

The convolution layer is used for convolving the real-time image to obtain a plurality of feature images. In this embodiment, each convolution layer may be formed using the same size and number of convolution kernels. In a specific embodiment, each convolution layer may use 32 convolution kernels with a size of 3×3, a step size of 1, and the number of zero-pixel points for the periphery of the feature map is 1. It can be understood that by setting the convolution kernels with the same size and number to form each convolution layer, the feature image size output by each convolution is ensured to be consistent with the output feature image size, and the loss of the edge information of the features in the image is avoided.

The plurality of convolution layers are divided into a first convolution group and a second convolution group, the number of the convolution layers in the first convolution group and the number of the convolution layers in the second convolution group can be the same, in this embodiment, the first convolution group and the second convolution group adopt a symmetrical cascade structure, when the convolution layers in the second convolution group carry out convolution, the convolution results of the convolution layers of the matched first convolution group are spliced to obtain a fusion feature map, so that the feature map of the shallow layer and the feature map of the deep layer symmetrical with the feature map of the shallow layer can be spliced in the channel direction, shallow layer information and deep layer features are effectively fused, and the neural network can learn the global features and local features of the image better. Furthermore, due to the reinforcement of the characteristics, the cascade structure can accelerate the convergence of the neural network to a certain extent.

The full connection layer is used for determining preset parameters based on the fusion feature map, and the preset parameters can be used for being substituted into a contrast enhancement algorithm to enhance the real-time image. In a specific embodiment, the full connection layer includes m×n neurons, where M and N are respectively the length and width of the image, and each neuron is used to predict a parameter of the contrast enhancement algorithm, that is, a preset parameter corresponding to each position in the image, so as to implement targeted adjustment of the pixel level.

In a specific embodiment, the input of the contrast enhancement algorithm may be a preset parameter q (i, j) output by the previous neural network, where i and j are used to represent the position of the pixel to be processed in the real-time image, and it is understood that the preset parameter may be a positive number or a negative number, and, as an example, when the pixel located in (i, j) in the image is a high-frequency information value, the preset parameter may be a positive value, and when it is low-frequency information, the value may be a negative value. Assuming that the gray value of the (i, j) th pixel point in the real-time image is x (i, j), the gray mean value m (i, j) and the gray variance of each pixel point can be sequentially determined by the following formula：

Wherein x (i, j), q (i, j), m (i, j) andRespectively the gray value, the preset parameter, the gray mean value and the gray variance of the (i, j) th pixel point. By/>The gray standard deviation/>, of the (i, j) th pixel point can be obtained。

By presetting parameters q (i, j) and m (i, j) andSubstituting the formula to calculate the final gray value:

Wherein f (i, j) is the gray value of the finally determined (i, j) th pixel point, and d is a constant coefficient.

According to the crystal bar growth control method, the convolution results of the matched convolution layers of the first convolution group are spliced when the convolution layers of the second convolution group are convolved to obtain the fusion feature map, so that the shallow feature map and the deep feature map symmetrical to the shallow feature map can be spliced in the channel direction, shallow information and deep features are effectively fused, and the neural network can learn global features and local features of images better; and predicting preset parameters corresponding to each position in the real-time image through a plurality of neurons of the full-connection layer, so as to realize targeted adjustment of the pixel point level.

The channel reorganization downsampling unit may divide the crystal feature extracted by the backbone network into a plurality of sub-feature graphs through downsampling, and stack the plurality of sub-feature graphs on the channel to obtain the reorganized crystal feature after channel reorganization.

The fusion unit can be composed of a plurality of convolution layers which are connected in sequence, feature graphs of different layers obtained by the plurality of convolution layers are mutually fused, fusion of deep features and shallow features is achieved, and fusion crystal features are obtained.

The attention weighted fusion unit is used for weighting the fusion crystal characteristics and fusing the weighted crystal characteristics with the fusion crystal characteristics so as to realize the enhancement of the crystal characteristics.

In a specific embodiment, the attention weighted fusion unit may be composed of an adaptive average pooling layer, a convolution layer, an activation layer, a convolution layer, and an activation layer. The self-adaptive average pooling layer adopts AdaptiveAvgPool d functions for carrying out average pooling on global features, the first activation layer adopts ReLU functions for carrying out linear rectification on the features input by the upper layer, and the second activation layer adopts Sigmoid functions for carrying out nonlinear activation processing on the features input by the upper layer.

According to the crystal bar growth control method, the attention fusion module is formed by the channel recombination downsampling unit, the fusion unit and the attention weighting fusion unit which are sequentially connected, so that the crystal characteristics extracted at each stage in the model can be better combined, the crystal characteristic diagram is weighted through an attention mechanism, the weight of the crystal characteristic can be improved, the weight of the amorphous crystal characteristic is reduced, the characteristic diagram with more obvious crystal characteristics is obtained, and the effect of improving the accuracy of crystal growth judgment is achieved.

In a specific embodiment, the training image set may be an image of the sapphire seeding amorphous ingot growth state and the growth state containing the ingot acquired using an area array camera at the temperature test drawknot stage. In the temperature test drawknot process, the crystal bar is in a rotating state, and images belonging to the same seeding part are spliced into a complete state diagram by adjusting the shooting interval of a camera and the rotating speed of a seed rod, and 1000 to 2000 or even more crystal state diagrams are acquired and used as images of a training image set. The division ratio of the training image set may be 8:1:1.

According to the crystal bar growth control method, the built neural learning network is trained by adopting the crystal bar growth abnormal image, and the test set comprising the crystal bar growth normal image and the crystal bar growth abnormal image is adopted for testing, so that a more accurate growth detection model can be obtained through training, and the effect of improving the crystal bar growth judgment accuracy is achieved.

In order to more clearly illustrate the technical solution of the present application, as shown in fig. 3, this embodiment also provides a detailed embodiment. In this embodiment, a method for controlling growth of a seed rod is provided, including:

Image collection and preprocessing: and acquiring images of the sapphire seeding normal state and the crystal thickness which do not meet the preset thickness in the temperature test and drawknot stage by using an area array camera. In the temperature test drawknot process, the seed rod is in a rotating state, images belonging to the same seeding part are spliced into a complete state diagram by adjusting the shooting interval of a camera and the rotating speed of the seed rod, the number of the state diagrams can be 1000 to 2000, and the number of the state diagrams is not limited in the embodiment. Dividing the data set into a training set, a verification set and a test set, wherein the dividing ratio can be 8:1:1, wherein the training set and the verification set are abnormal crystal bar growth images, namely images with the crystal thickness not meeting the preset thickness, the training set and the verification set are not repeated, the test set is a normal crystal bar growth image and an abnormal crystal bar growth image, and the images can be obtained by marking software marks through labelme and can be obtained by manually marking defect positions. The training set is used for training a model, the verification set is used for adjusting model parameters, and the adjustment method can be transfer learning.

In this embodiment, as shown in fig. 4, in the growth process of the ingot, the crystal existing around the seed crystal may be the distance from the small black point formed in the liquid contact step to the seed crystal, if the thickness of the crystal meets the required thickness of 3mm, the crystal appears, the ingot grows, if the crystal is too thick or too thin, the temperature is too high or too low, and the ingot grows abnormally. The thickness of the crystal can be obtained by post-treatment measurement and calculation, or can be determined based on human judgment input.

And building an ACEA-Yolov model, training and storing the model. As shown in fig. 5, the ACEA-Yolov model includes a backbone network, a neck network and a detection network connected in sequence, before the backbone network, the model further includes an input layer, and the input layer performs operations such as Mosaic data enhancement, adaptive anchor frame calculation, adaptive image scaling and the like on a real-time image, and then inputs the operations to the backbone network. The input layer also comprises a self-adaptive contrast enhancement module which is used for enhancing the real-time image to obtain an enhanced image; the backbone network is used for extracting crystal characteristics based on the enhanced image; the backbone network further comprises an attention fusion module for enhancing the crystal characteristics; the neck network comprises a bidirectional fusion network, all layers of the bidirectional fusion network are connected with each other, and all layers of the bidirectional fusion network keep each layer of characteristics and are used for fusing the crystal characteristics to obtain fused crystal characteristics; the detection network is used for determining the target area image and a crystal bar growth detection result corresponding to the target area image based on the fused crystal characteristics. The self-adaptive contrast enhancement module is introduced into the network input layer to process the input crystal bar image, so that the crystal characteristics are more obvious, network identification is facilitated, and the problem that the crystal characteristics are not obvious and are difficult to distinguish, so that the phenomenon of omission is easy to occur is solved. The ACEA-Yolov model can be trained by adopting a multi-scale training strategy to adapt to melting defects with different sizes, and in the training process, a loss function can be realized by adopting Focal EIoU double losses so as to improve the detection precision of the melting defects. Through verification of the test set, if the detection accuracy reaches 0.98, the defect detection model can be stored.

As shown in fig. 6, the adaptive contrast enhancement module of this embodiment includes six convolution layers, a full connection layer and a contrast enhancement unit connected in sequence, where the convolution layers are divided into a first convolution group and a second convolution group; the convolution layers sequentially convolve the real-time image, and when the convolution layers in the second convolution group convolve, the convolution results of the convolution layers of the matched first convolution group are spliced to obtain a fusion feature map; the full-connection layer is used for determining preset parameters based on the fusion feature map; the contrast enhancement unit is used for enhancing the contrast of the real-time image based on the preset parameters and a contrast enhancement algorithm. Besides the final output layer using the Tanh activation function, the rest layers use the leak ReLU as the activation function, so that the nonlinear processing of the feature map can be realized.

Furthermore, the convolution kernels with the size of 3 multiplied by 3 are adopted by all convolution layers, the step length S is 1, the number of zero pixel points are supplemented to the periphery of the feature map is 1, the feature map size outputted by convolution each time is ensured to be consistent with the input feature map size through the arrangement, and the loss of edge information of features in the image is prevented. The multiple convolution layers adopt a symmetrical cascade structure, the shallow characteristic images and the deep characteristic images symmetrical with the shallow characteristic images are spliced in the channel direction, in this way, the embodiment can effectively fuse the shallow information and the deep characteristics, so that the neural network can learn the global characteristics and the local characteristics of the image better, and the cascade structure can accelerate the convergence of the neural network to a certain extent. The full connection layer is provided with m×n neurons, wherein M, N is the length and width of the image, respectively. Each neuron is used for predicting a preset parameter of the adaptive contrast enhancement algorithm, namely a gain parameter corresponding to each position of the image. The adaptive contrast enhancement module provided by the embodiment is based on an adaptive contrast enhancement algorithm, predicts parameters of the adaptive contrast enhancement algorithm through a convolutional neural network, and eliminates a blurring effect caused by the adoption of a uniform global standard deviation when processing an input image, so that the adaptive contrast enhancement effect can be achieved.

The input of the contrast enhancement unit may be a preset parameter q (i, j) output by the previous neural network, where i and j are used to represent the positions of the pixels to be processed in the real-time image, and it may be understood that the preset parameter may be a positive number or a negative number, and, for example, when the pixel located in (i, j) in the image is a high-frequency information value, the preset parameter may be a positive value, and corresponds to performing high-frequency processing; when it is low frequency information, the value may be negative, corresponding to low frequency processing. Assuming that the gray value of the (i, j) th pixel point in the real-time image is x (i, j), the gray mean value m (i, j) and the gray variance of each pixel point can be sequentially determined by the following formula：

wherein f (i, j) is the gray value of the finally determined (i, j) th pixel point, and d is a constant coefficient. Fig. 7 shows a comparison of the real-time image before and after adaptive contrast enhancement.

Through the self-adaptive contrast enhancement module, noise can be restrained while the contrast of the crystal bar image is enhanced, so that the area with unobvious crystal characteristics is improved to a certain extent, and the model is better identified and extracted.

As shown in fig. 8, the attention fusion module adopts a strategy of fusion, re-weighting and re-fusion, and comprises a channel recombination downsampling unit, a fusion unit and an attention weighted fusion unit which are sequentially connected, wherein the channel recombination downsampling unit is used for downsampling the feature images to output crystal feature images with richer channel information, the fusion unit realizes fusion of the crystal feature information through multiple convolutions, and the attention weighted fusion unit weights the crystal feature images through an attention mechanism so as to improve the crystal feature weight, reduce the amorphous feature weight and obtain the feature images with more obvious crystal features by referring to the previous information. It can be understood that, because the crystal structure characteristic region includes small black points formed by the liquid contact process, the small black points are small targets, and the recognition effect on the small target image can be improved by means of fusion and weighting and fusion.

And integrating the model into an industrial personal computer, and connecting a camera and a PLC. The camera can be an area array camera arranged near the furnace cover of the crystal growth furnace, and the PLC can be a logic controller for driving and controlling the motor of the seed rod.

Collecting an image, inputting the image, and analyzing the result of the model: and acquiring an in-furnace image, a step name and a drawing time of the sapphire temperature test drawknot process in real time, and carrying out self-adaptive contrast enhancement on the real-time image to acquire an enhanced image with more obvious crystal knot characteristics. It will be appreciated that each step name corresponds to a corresponding model, and the temperature test drawknot step may correspond to a growth detection model in this embodiment, with the mapping time corresponding to the initial time of image acquisition. And detecting each image in the image group by using an ACEA-Yolov model to obtain a target frame possibly containing crystals. Further, the model may detect the real-time image, that is, the real-time image obtained after detection is divided into a plurality of candidate frames to detect, and generate a detection result reaching a confidence threshold and a IoU threshold, where the confidence threshold and the IoU threshold may be set based on actual requirements, and the portion exceeding the threshold may be screened by judging through a non-maximum suppression algorithm.

Judging whether the crystal bar grows: and the model detects the crystals of the target frame possibly containing the crystals, and judges whether the crystal bar growth condition exists in the image based on the crystal detection result. Specifically, whether crystals appear in each target frame or not can be judged based on the target frames and comparing with the crystal characteristics, if the similarity exceeds 0.5, crystals exist in the temperature test and pulling stage, the crystal thickness reaches the preset thickness to indicate the growth condition of crystal bars, and if the crystal thickness is less than the preset thickness, the temperature is regulated through the crystal thickness.

If the temperature change is not stable, the information feedback industrial personal computer automatically adjusts the power, judges the temperature change stabilization time and retries the temperature drawknot.

If the seed crystal exists, purifying the seed crystal, inserting the seed crystal into the liquid surface for 15mm, carrying out drawknot, and ending the detection. The liquid level can be adjusted by controlling a PLC driving motor through an industrial personal computer.

The temperature test and drawknot step in the embodiment can be two hours long, and the area array camera can be arranged around an observation window of a furnace cover of the crystal growth furnace through a camera fixing rod and used for collecting real-time images in the crystal growth furnace. The image acquisition may be to take a plurality of images every 20 minutes, detect, observe the seeding portion of the crystal through a window and rotate the images.

Furthermore, the training set image is the imaging of the same seed crystal part in the same time period, and in order to ensure the integrity of the seeding part in the training image, the imaging acquisition times of the seeding part are generally determined according to the working distance of the camera, the rotating speed of the seed crystal rod and the clear imaging field area. If the clear view area is large, the acquisition times are less; the clear field of view is small, and several more acquisitions are required.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides a crystal bar growth control device for realizing the crystal bar growth control method. The implementation of the solution provided by the apparatus is similar to that described in the above method, so specific limitations in one or more embodiments of the ingot growth control apparatus provided below may be found in the limitations of the ingot growth control method described above, and are not repeated here.

In one embodiment, as shown in fig. 9, there is provided a crystal rod growth control apparatus comprising: an acquisition module 100, a detection module 200, and a processing module 300, wherein:

An acquisition module 100, configured to acquire a real-time image in the crystal growth furnace;

The detection module 200 is used for inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image and a crystal bar growth detection result;

the processing module 300 is configured to re-perform the test Wen Lajie based on the target area image if the detection result is that the crystal ingot growth is abnormal;

In some of these embodiments, the detection network is further configured to:

determining the target area image based on the fused crystal characteristics;

In some of these embodiments, the processing module 300 is further configured to:

In some of these embodiments, the boule growth control apparatus further comprises a training module for:

The various modules in the ingot growth control apparatus described above may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, as shown in fig. 10, a crystal growth furnace system is provided, which comprises a crystal growth furnace, an area array camera and an industrial personal computer, wherein a furnace cover 1 of the crystal growth furnace is provided with an observation window 3, the area array camera is arranged around the observation window through a camera fixing rod 4 and is used for acquiring real-time images in the crystal growth furnace and transmitting the real-time images to the industrial personal computer; the industrial personal computer is used for realizing the crystal bar growth control method of any one of the embodiments:

Acquiring a real-time image in the crystal growth furnace;

Further, the observation window 3 is provided around the seed rod opening 2, and the seed rod opening 2 is located at the center of the furnace cover 1. The camera fixing rod 4 further comprises a detachable rod 5 and a camera clamp 6 which are connected, and the area array camera is fixed towards the observation window 3 after being clamped by the camera clamp 6 for image acquisition.

In one embodiment, a computer device is provided, which may be a terminal, and the internal structure thereof may be as shown in fig. 11. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement a method of controlling the growth of a seed crystal. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, can also be keys, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

It will be appreciated by those skilled in the art that the structure shown in FIG. 11 is merely a block diagram of some of the structures associated with the present inventive arrangements and is not limiting of the computer device to which the present inventive arrangements may be applied, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, where the memory stores a computer program, and the processor implements the ingot growth control method of any of the above embodiments when executing the computer program:

Acquiring a real-time image in the crystal growth furnace;

In one embodiment, a computer readable storage medium is provided having a computer program stored thereon, which when executed by a processor, implements the ingot growth control method of any of the above embodiments:

Acquiring a real-time image in the crystal growth furnace;

The user information (including but not limited to user equipment information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the present application are information and data authorized by the user or sufficiently authorized by each party.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magneto-resistive Memory (Magnetoresistive Random AccessMemory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (PHASE CHANGE Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in various forms such as static Random access memory (Static Random Access Memory, SRAM) or Dynamic Random access memory (Dynamic Random AccessMemory, DRAM), and the like. The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processor referred to in the embodiments provided in the present application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, or the like, but is not limited thereto.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The foregoing examples illustrate only a few embodiments of the application and are described in detail herein without thereby limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of the application should be assessed as that of the appended claims.

Claims

1. The crystal bar growth control method is characterized by comprising the following steps of:

Acquiring a real-time image in the crystal growth furnace;

The self-adaptive contrast enhancement module is used for enhancing the real-time image to obtain an enhanced image; the self-adaptive contrast enhancement module comprises a plurality of convolution layers, a full connection layer and a contrast enhancement unit which are sequentially connected, wherein the convolution layers are divided into a first convolution group and a second convolution group; the convolution layers sequentially convolve the real-time image, and when the convolution layers in the second convolution group convolve, the convolution results of the convolution layers of the matched first convolution group are spliced to obtain a fusion feature map; the full-connection layer is used for determining preset parameters based on the fusion feature map; the contrast enhancement unit is used for enhancing the contrast of the real-time image based on the preset parameters and a contrast enhancement algorithm; the input of the contrast enhancement algorithm is preset parameters q (i, j), i and j are used for representing the positions of pixel points to be processed in the real-time image, and the gray average value m (i, j) and the gray variance of each pixel point are sequentially determined ：/>，; Wherein x (i, j), q (i, j), m (i, j) and/>Respectively obtaining gray values, preset parameters, gray average values and gray variance of the (i, j) th pixel point of the real-time image; by gray variance/>Obtaining the gray standard deviation/>, of the (i, j) th pixel point; By presetting parameters q (i, j), m (i, j) andAnd calculating to obtain the finally determined gray value: /(I)Wherein f (i, j) is the gray value of the finally determined (i, j) th pixel point, and d is a constant coefficient;

2. The ingot growth control method of claim 1, wherein the detection network is further configured to:

determining the target area image based on the fused crystal characteristics;

3. The method according to claim 2, wherein the re-performing the temperature-trial drawknot based on the target area image comprises:

4. The method according to claim 1, wherein the inputting the real-time image into a pre-trained growth detection model to obtain a target area image in the real-time image, and the step of after the crystal growth detection result further comprises:

5. The ingot growth control method of claim 1, wherein the attention fusion module comprises a channel reorganization downsampling unit, a fusion unit, and an attention weighted fusion unit connected in sequence, wherein:

6. The ingot growth control method of claim 1, wherein the inputting the real-time image into a pre-trained growth detection model is preceded by:

7. A device for controlling growth of a seed rod, the device comprising:

8. The crystal growth furnace system is characterized by comprising a crystal growth furnace, an area array camera and an industrial personal computer, wherein an observation window is arranged on a furnace cover of the crystal growth furnace, the area array camera is arranged around the observation window through a camera fixing rod and is used for collecting real-time images in the crystal growth furnace and sending the real-time images to the industrial personal computer; the industrial personal computer is configured to implement the steps of the method of any one of claims 1 to 6.

9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 6 when the computer program is executed.