A deep learning method for sorting and inspecting photovoltaic silicon wafers
By using deep learning methods and 6G millimeter-wave technology to automatically adjust the light source and reflector of the silicon wafer sorting machine, the problems of low efficiency and detection errors caused by manual adjustment are solved, achieving efficient and accurate silicon wafer detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JINWAN GAOJING SOLAR ENERGY TECH CO LTD
- Filing Date
- 2023-09-26
- Publication Date
- 2026-06-30
AI Technical Summary
Existing silicon wafer sorting machines require manual adjustment of the light source and reflector when switching between different sizes or types of silicon wafers, resulting in low production efficiency and easy human measurement errors, which affect the quality of testing.
Using deep learning methods, a convolutional neural network model and attention mechanism are used to identify silicon wafer images, automatically adjust the position of the light source and reflector, and combine 6G millimeter wave technology for data transmission to achieve automated adjustment.
It improves the production efficiency and inspection quality of silicon wafer sorting machines, saves labor costs, and ensures the timeliness and accuracy of data transmission.
Smart Images
Figure CN117253121B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of silicon wafer image processing technology, specifically relating to a deep learning method for sorting and inspecting photovoltaic silicon wafers. Background Technology
[0002] In the production process of solar cells, silicon rods are first cut into silicon wafers, which are then cleaned. A silicon wafer sorting machine then performs multiple inspections on the wafers, such as checking for surface contamination, thickness, size, and microcracks. In practical applications, due to the wide variety of silicon wafer types and their different sizes after cutting, existing silicon wafer sorting machines require manual adjustment of the light source and reflector when switching between different sizes or types of wafers. The specific process is as follows: the dimensions of the silicon wafer are measured manually, and then, according to standard operating instructions, the positions of the light source and reflector of the silicon wafer sorting machine that need adjustment are determined. Finally, the light source and reflector are adjusted on-site.
[0003] The existing methods have the following drawbacks:
[0004] 1) Adjusting the light source and reflector of the silicon wafer sorting machine requires a lot of manpower and time, which will result in low production efficiency.
[0005] 2) When manually measuring silicon wafer dimensions and consulting relevant standards, human factors are uncontrollable and can easily lead to errors during measurement or consultation, causing the sorting machine to be unable to effectively inspect the silicon wafers and resulting in quality problems. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a deep learning method for sorting and inspecting photovoltaic silicon wafers.
[0007] Firstly, the objective of this invention is achieved through the following technical solution:
[0008] A deep learning method for sorting and inspecting photovoltaic silicon wafers includes:
[0009] Acquire silicon wafer images;
[0010] Based on the silicon wafer image and the trained convolutional neural network model, the classification result of the silicon wafer is obtained;
[0011] Based on the classification results, the light source information and reflector information of the corresponding silicon wafer in the preset MES system are called;
[0012] Based on the classification results, the light source information, and the reflector information, a light source adjustment command and a reflector adjustment command are sent to a preset PLC.
[0013] Preferably, the process of obtaining the silicon wafer classification result based on the silicon wafer image and the trained convolutional neural network model includes:
[0014] The silicon wafer image is fed into a trained convolutional neural network model to extract image features from the silicon wafer image;
[0015] Based on the image features and the preset attention mechanism, the attention features of the silicon wafer image are calculated;
[0016] The attention features are fed into a preset classifier to obtain the classification results of the silicon wafers.
[0017] Preferably, the step of issuing light source adjustment commands and reflector adjustment commands to a preset PLC based on the classification results, the light source information, and the reflector information includes:
[0018] Based on the classification results, the light source information, and the reflector information, 6G millimeter wave technology is used to send light source adjustment commands and reflector adjustment commands to a preset PLC.
[0019] Preferably, before obtaining the classification result of the silicon wafer based on the silicon wafer image and the trained convolutional neural network model, the method further includes:
[0020] Acquire images of multiple different silicon wafer samples;
[0021] Assign label information to the key features of each silicon wafer sample image;
[0022] Deep learning models are built using deep learning frameworks;
[0023] The silicon wafer sample image and the label information are fed into the deep learning model for training to obtain a convolutional neural network model.
[0024] Preferably, the step of building a deep learning model using a deep learning framework includes:
[0025] Construct multiple convolution kernels to perform dot products on the input image;
[0026] An activation function is inserted into the convolution kernel to perform non-linear processing on the input image after the dot product;
[0027] Set up multiple pooling layers;
[0028] Configure multiple pooling layers with pooling functions to convert the non-linearly processed input image into a one-dimensional vector;
[0029] Configure a fully connected layer to convert the one-dimensional vector into image features;
[0030] The input image includes the silicon wafer sample image and the silicon wafer image.
[0031] Preferably, the convolution kernel has 9 kernels, and the kernel size is 7×7.
[0032] Preferably, the activation function includes the GeLU function.
[0033] Preferably, the multiple pooling layers include three Average pooling layers and three Max pooling layers.
[0034] Preferably, the pooling function includes a dim function and a sigmoid function, and the output dimension of the dim function is set to 1.
[0035] Secondly, the objective of this invention is achieved by the following technical solution:
[0036] A silicon wafer sorting machine is provided, which incorporates the aforementioned deep learning method for sorting and inspecting photovoltaic silicon wafers. The sorting machine includes an industrial camera, an industrial computer, a host computer, a MES system, a PLC, and actuators.
[0037] The industrial camera is used to acquire images of silicon wafers;
[0038] The industrial computer has a built-in convolutional neural network module and a silicon wafer sorting module. The convolutional neural network module is used to obtain the classification result of the silicon wafer based on the silicon wafer image and the trained convolutional neural network model. The silicon wafer sorting module is used to call the light source information and reflector information of the corresponding silicon wafer in the MES system based on the classification result. The silicon wafer sorting module is also used to send light source adjustment commands and reflector adjustment commands to the host computer based on the classification result, the light source information and the reflector information.
[0039] The MES system uses 6G millimeter wave technology and communicates with the industrial computer through an information queue.
[0040] The host computer uses 6G millimeter wave technology and communicates with the industrial computer through an information queue; the host computer is used to forward the light source adjustment command and the reflector adjustment command to the PLC;
[0041] The PLC is used to control the actuator to adjust the corresponding light source position, light source angle, reflector position, and reflector angle.
[0042] Compared with the prior art, the present invention has at least the following advantages:
[0043] 1) This invention combines convolutional neural network model and attention mechanism and applies them to silicon wafer sorting machine, so that it can adapt to the sorting of silicon wafers of different sizes or types. The PLC is used to adaptively adjust the position and angle of the light source, the position and angle of the reflector, thereby improving the production efficiency and sorting quality of silicon wafer sorting machine.
[0044] 2) The method of this invention is designed for 6G millimeter wave technology applications, solves the problem of long-distance data transmission, ensures the timeliness and accuracy of data output, and saves a lot of manpower costs. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0046] Figure 1 This is a flowchart of the deep learning method for sorting and inspecting photovoltaic silicon wafers in Embodiment 1 of the present invention;
[0047] Figure 2 This is a flowchart of step S20 in the deep learning method for sorting and inspecting photovoltaic silicon wafers in Embodiment 1 of the present invention;
[0048] Figure 3 This is a flowchart of steps S100-S400 in the deep learning method for sorting and inspecting photovoltaic silicon wafers in Embodiment 1 of the present invention;
[0049] Figure 4 This is a flowchart of step S300 in the deep learning method for sorting and inspecting photovoltaic silicon wafers in Embodiment 1 of the present invention;
[0050] Figure 5 This is a schematic diagram of the convolutional neural network model in Embodiment 1 of the present invention;
[0051] Figure 6 This is a schematic diagram of the silicon wafer sorting machine in Embodiment 2 of the present invention. Detailed Implementation
[0052] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0053] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0054] This invention provides a deep learning method for sorting and inspecting photovoltaic silicon wafers and a silicon wafer sorting machine to solve the problems existing in the prior art.
[0055] Example 1
[0056] like Figure 1 As shown, the deep learning method for sorting and inspecting photovoltaic silicon wafers according to an embodiment of the present invention includes:
[0057] S10: Obtain silicon wafer image.
[0058] Specifically, an industrial camera is used to photograph the silicon wafer to be inspected, thereby obtaining the corresponding silicon wafer image.
[0059] S20: Based on the silicon wafer image and the trained convolutional neural network model, the classification results of the silicon wafer are obtained.
[0060] In this embodiment, the classification results include at least the model information and size information of the silicon wafer, and the classification results correspond to the preset silicon wafer related standards (including the light source parameter setting standards and the reflector parameter setting standards).
[0061] Specifically, by feeding silicon wafer images into a trained convolutional neural network model, the classification results of silicon wafers are obtained, thereby reducing the uncontrollability of manual measurement and improving the efficiency of silicon wafer detection.
[0062] Among them, such as Figure 2 As shown, step S20 includes:
[0063] S21: Input the silicon wafer image into the trained convolutional neural network model to extract the image features of the silicon wafer image;
[0064] S22: Calculate the attention features of the silicon wafer image based on image features and a preset attention mechanism;
[0065] S23: Input the attention features into the preset classifier to obtain the classification result of the silicon wafer.
[0066] In this embodiment, in order to achieve backpropagation of data, an existing soft attention mechanism is adopted. The soft attention mechanism includes one of the following: self-attention mechanism, multi-head attention mechanism, separable attention mechanism, global average pooling attention mechanism, and nonlinear attention mechanism. The classifier can be a fully connected layer built into the convolutional neural network model, or it can be an additional classifier.
[0067] It should be noted that if backpropagation is required to train the convolutional neural network model, the classifier is the fully connected layer built into the convolutional neural network model.
[0068] Specifically, the silicon wafer image is placed into a convolutional neural network model to obtain the corresponding image features. Then, combined with a soft attention mechanism, the image features are backpropagated to the fully connected layer built into the convolutional neural network model for training to obtain attention weights. This allows the convolutional neural network model to focus on processing the most similar related features to the silicon wafer image and use the most similar related features as attention features. Finally, the classifier obtains the classification result of the silicon wafer based on the attention features.
[0069] Among them, such as Figure 3 As shown, before step S20, the method of the present invention further includes:
[0070] S100: Acquires images of multiple different silicon wafer samples;
[0071] S200: Assign label information to the key features of each silicon wafer sample image;
[0072] S300: Uses a deep learning framework to build deep learning models;
[0073] S400: Input the silicon wafer sample image and label information into the deep learning model for training to obtain the convolutional neural network model.
[0074] In this embodiment, multiple different silicon wafer sample images should contain multiple different silicon wafer size information, shape information, and surface feature information (such as cracks, reflectivity, and pits). All of the above information serves as key features of the silicon wafer sample images. The label information includes at least size data, shape identification, and silicon wafer model. The key features and label information of the silicon wafer sample images enable the deep learning model to perform self-learning and achieve automatic classification or sorting of silicon wafers. The deep learning framework uses Python.
[0075] In other embodiments, in order to enhance the diversity and reliability of training samples, additional data augmentation steps can be set, such as expanding the dataset for training deep learning models by rotating, flipping, scaling and translating silicon wafer sample images, so as to improve the robustness and generalization ability of the models.
[0076] Specifically, before inputting a silicon wafer image in the method of the present invention, a convolutional neural network model needs to be trained and generated. This model can be pre-stored in an existing computer device after being trained with multiple silicon wafer sample images, and then retrieved and used when step S20 is executed.
[0077] It should be noted that during the training of a deep learning model, the cross-entropy loss function can be used to measure the model's classification accuracy, and an optimization algorithm (gradient descent) can be used to update the model parameters.
[0078] Among them, such as Figure 4 and Figure 5 As shown, step S300 includes:
[0079] S301: Construct multiple convolution kernels to perform dot products on the input image;
[0080] S302: Insert an activation function into the convolution kernel to perform non-linear processing on the input image after the dot product;
[0081] S303: Set multiple pooling layers;
[0082] S304: Configure multiple pooling layers with pooling functions to convert the non-linearly processed input image into a one-dimensional vector;
[0083] S305: Set up a fully connected layer to convert a one-dimensional vector into image features;
[0084] S306: Input images include silicon wafer sample images and silicon wafer images.
[0085] Preferably, there are 9 convolution kernels, and the size of each convolution kernel is 7×7;
[0086] Preferably, the activation function is the GeLU function;
[0087] Preferably, the multiple pooling layers consist of 3 Average pooling layers and 3 Maxpooling layers;
[0088] Preferably, the pooling function consists of a dim function and a sigmoid function. The dim function has an output dimension of 1 and is used to convert the non-linearly processed input image into a one-dimensional vector. The sigmoid function is used to merge the input images that have passed through the Averagepooling layer and the Maxpooling layer.
[0089] Specifically, during the model training phase, multiple silicon wafer sample images are first fed into a convolutional kernel. The convolutional kernel performs a dot product on the multiple silicon wafer sample images to extract image features. Then, the GeLU function is used to perform nonlinear processing on the image features of the silicon wafer sample images. After that, three Average pooling layers and three Max pooling layers, along with their corresponding pooling functions, convert the nonlinearly processed image features into a one-dimensional vector and output it to a fully connected layer. The fully connected layer sorts each silicon wafer sample image. During the silicon wafer sorting phase, newly acquired silicon wafer images can also be used as training objects for the model. The program logic for silicon wafer sorting is similar to the above method and will not be repeated here.
[0090] S30: Based on the classification results, call the light source information and reflector information of the corresponding silicon wafer in the preset MES system.
[0091] In this embodiment, the MES system is an existing intelligent production management system with functions such as production information aggregation, production environment monitoring, material information aggregation, and production formula parameter storage.
[0092] In this embodiment, the light source information includes light source position information and light source angle information; the reflector information includes reflector position information and reflector angle information.
[0093] Specifically, based on the size, shape, and surface feature information within the classification results, the system can automatically retrieve the corresponding silicon wafer light source and reflector information by calling the preset production formula parameters. This replaces the manual process of consulting relevant standards and determining the corresponding silicon wafer light source and reflector information, thereby effectively improving production efficiency and enabling effective inspection of silicon wafers.
[0094] S40: Based on the classification results, light source information, and reflector information, send light source adjustment commands and reflector adjustment commands to the preset PLC.
[0095] In this embodiment, the industrial computer acquires the production formula parameters of the MES system (i.e., light source position information, light source angle information, reflector position information, and reflector angle information), and then sends light source adjustment commands and reflector adjustment commands to the PLC.
[0096] Specifically, the PLC automatically adjusts the position and angle of the light source, the position and angle of the reflector according to the light source adjustment command and the reflector adjustment command, without the need for manual intervention to adjust the angle and position, thereby further improving production efficiency.
[0097] Preferably, the data transmission in steps S30 and S40 adopts the form of message queue and combines 6G millimeter wave technology as the main communication technology.
[0098] Compared to existing technologies, the use of message queues combined with 6G millimeter wave technology enables real-time uploading of image data and various parameters, ensuring the timeliness and accuracy of the data. Furthermore, it solves the problem of packet loss in long-distance data transmission, saving a significant amount of manpower and reducing production costs.
[0099] The implementation principle of this invention is as follows:
[0100] First, an industrial camera is used to photograph the silicon wafer to obtain an image. Then, the image is fed into a convolutional neural network model to obtain image features. Based on the image features and a preset attention mechanism, the attention features of the silicon wafer image are calculated. The attention features are then sent to a preset classifier to determine the classification result of the silicon wafer. Based on the classification result and 6G millimeter-wave technology, the corresponding light source information and reflector information of the silicon wafer in the MES system are called using a message queue. Finally, the called light source information and reflector information are converted into light source adjustment instructions and reflector adjustment instructions, respectively, and sent to the PLC, thereby effectively improving production efficiency and realizing effective inspection of silicon wafers.
[0101] Example 2
[0102] like Figure 6 As shown, based on Embodiment 1, this embodiment of the invention provides a silicon wafer sorting machine, which incorporates the aforementioned deep learning method for sorting and inspecting photovoltaic silicon wafers. The sorting machine includes an industrial camera, an industrial computer, a host computer, a MES system, a PLC, and an actuator.
[0103] Industrial cameras are used to acquire images of silicon wafers;
[0104] The industrial computer has a built-in convolutional neural network module and a silicon wafer sorting module. The convolutional neural network module is used to obtain the classification results of silicon wafers based on silicon wafer images and trained convolutional neural network models. The silicon wafer sorting module is used to call the light source information and reflector information of the corresponding silicon wafer in the MES system based on the classification results. The silicon wafer sorting module is also used to send light source adjustment commands and reflector adjustment commands to the host computer based on the classification results, light source information and reflector information.
[0105] The MES system uses 6G millimeter wave technology and communicates with industrial computers through information queues.
[0106] The host computer uses 6G millimeter-wave technology and communicates with the industrial computer through information queues; the host computer is used to forward light source adjustment instructions and reflector adjustment instructions to the PLC.
[0107] The PLC is used to control the actuator to adjust the corresponding light source position, light source angle, reflector position, and reflector angle.
[0108] Compared to existing technologies, this invention combines convolutional neural network models and attention mechanisms in a silicon wafer sorting machine, enabling the machine to adapt to silicon wafers of different sizes and achieve automatic adjustment of the light source position, light source angle, reflector position, and reflector angle to complete the effective detection of silicon wafers.
[0109] Meanwhile, the sorting machine of this invention also uses 6G millimeter wave technology, which can achieve a transmission rate 50 times that of 5G and reduce latency to one-tenth of that of 5G, thus better meeting the needs of actual production.
[0110] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
Claims
1. A deep learning method for photovoltaic silicon wafer sorting inspection, characterized in that, include: Acquire silicon wafer images; Based on the silicon wafer image and the trained convolutional neural network model, the classification result of the silicon wafer is obtained; Based on the classification results, the light source information and reflector information of the corresponding silicon wafer in the preset MES system are called; Based on the classification results, the light source information, and the reflector information, a light source adjustment command and a reflector adjustment command are sent to a preset PLC. The classification results of the silicon wafers obtained based on the silicon wafer image and the trained convolutional neural network model include: The silicon wafer image is fed into a trained convolutional neural network model to extract image features from the silicon wafer image; Based on the image features and the preset attention mechanism, the attention features of the silicon wafer image are calculated; The attention features are fed into a preset classifier to obtain the classification result of the silicon wafer; The step of issuing light source adjustment commands and reflector adjustment commands to a preset PLC based on the classification results, the light source information, and the reflector information includes: Based on the classification results, the light source information, and the reflector information, 6G millimeter wave technology is used to send light source adjustment commands and reflector adjustment commands to a preset PLC.
2. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 1, wherein, Before obtaining the classification result of the silicon wafer based on the silicon wafer image and the trained convolutional neural network model, the method further includes: Acquire images of multiple different silicon wafer samples; Assign label information to the key features of each silicon wafer sample image; Deep learning models are built using deep learning frameworks; The silicon wafer sample image and the label information are fed into the deep learning model for training to obtain a convolutional neural network model.
3. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 2, wherein, The method of building a deep learning model using a deep learning framework includes: Construct multiple convolution kernels to perform dot products on the input image; An activation function is inserted into the convolution kernel to perform non-linear processing on the input image after the dot product; Set up multiple pooling layers; Configure multiple pooling layers with pooling functions to convert the non-linearly processed input image into a one-dimensional vector; Configure a fully connected layer to convert the one-dimensional vector into image features; The input image includes the silicon wafer sample image and the silicon wafer image.
4. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 3, characterized in that, The convolution kernel has 9 kernels, and the kernel size is 7×7.
5. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 3, wherein, The activation function includes the GeLU function.
6. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 3, wherein, The multiple pooling layers include three Average pooling layers and three Max pooling layers.
7. The deep learning method for photovoltaic silicon wafer sorting inspection according to claim 3, wherein, The pooling functions include the dim function and the sigmoid function, and the output dimension of the dim function is set to 1.
8. A silicon wafer sorting machine, characterized in that, The sorting machine incorporates a deep learning method for sorting and inspecting photovoltaic silicon wafers as described in any one of claims 1-7, and the sorting machine includes an industrial camera, an industrial computer, a host computer, a MES system, a PLC, and an actuator. The industrial camera is used to acquire images of silicon wafers; The industrial computer has a built-in convolutional neural network module and a silicon wafer sorting module. The convolutional neural network module is used to obtain the classification result of the silicon wafer based on the silicon wafer image and the trained convolutional neural network model. The silicon wafer sorting module is used to call the light source information and reflector information of the corresponding silicon wafer in the MES system based on the classification result. The silicon wafer sorting module is also used to send light source adjustment commands and reflector adjustment commands to the host computer based on the classification result, the light source information and the reflector information. The MES system uses 6G millimeter wave technology and communicates with the industrial computer through an information queue. The host computer uses 6G millimeter wave technology and communicates with the industrial computer through an information queue; the host computer is used to forward the light source adjustment command and the reflector adjustment command to the PLC; The PLC is used to control the actuator to adjust the corresponding light source position, light source angle, reflector position, and reflector angle.