An artificial intelligence-based solar panel defect detection device

Abstract

The present invention discloses an artificial intelligence solar panel defect detection device, which relates to the technical field of solar panel defect detection, including a defect detection mechanism, a conveying mechanism is arranged on the side of the defect detection mechanism, the conveying mechanism is used to convey the solar panel to the defect detection mechanism and receive solar panels with different defect levels, and a control module is loaded on the defect detection mechanism, the control module collects a historical training data set in advance through the defect detection mechanism, and trains a machine learning model for predicting the defect level of the solar panel based on the historical training data set; the control module collects multi-angle image data of the solar panel in real time through the defect detection mechanism; the control module is completely determined by the intelligent control module, which greatly improves the efficiency of solar panel defect detection, and also improves the accuracy of solar panel defect detection, which is very suitable for large-scale detection in factories, is beneficial to reducing costs and increasing efficiency of factories, and improves the competitiveness of factories.

Description

Technical Field

[0001] The present invention relates to the technical field of solar panel defect detection, and in particular to an artificial intelligence solar panel defect detection device. Background Art

[0002] The artificial intelligence solar panel defect detection device uses machine vision technology to inspect the defects of the solar panel, such as an artificial intelligence solar panel defect detection device disclosed in the Chinese patent application with publication number CN117074422A and an automatic optical inspection device disclosed in the Chinese patent application with publication number CN108375582A;

[0003] This type of detection equipment is not highly intelligent when detecting solar panel defects, which makes it difficult to improve the efficiency and accuracy of solar panel defect detection, limiting the factory's large-scale production of solar panels. Summary of the invention

[0004] In order to overcome the above-mentioned technical problems, the purpose of the present invention is to provide an artificial intelligence solar panel defect detection device to solve the problem in the prior art that the detection equipment has a low degree of intelligence when detecting solar panel defects, which limits the efficiency and accuracy of the inspection equipment and makes it difficult for the detection equipment to form large-scale detection in the factory.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] Specifically, an artificial intelligence solar panel defect detection device is provided, including a defect detection mechanism, a plurality of conveying mechanisms are arranged on the side of the defect detection mechanism, the plurality of conveying mechanisms are respectively used to convey the solar panel to the defect detection mechanism and receive solar panels with different defect levels, a control module is loaded on the defect detection mechanism, the control module collects a historical training data set in advance through the defect detection mechanism, and trains a machine learning model for predicting the defect level of the solar panel based on the historical training data set; the control module collects multi-angle image data of the solar panel in real time through the defect detection mechanism, and predicts the corresponding defect level of the solar panel based on the multi-angle image data using the machine learning model, the multi-angle image data at least includes pictures taken at three angles of the solar panel; the defect detection mechanism conveys the solar panel to the corresponding conveying mechanism based on the corresponding defect level of the solar panel.

[0007] As a further solution of the present invention: the historical training data set includes N groups of training data, each group of training data includes feature data and label data, and the feature data is multi-angle image data;

[0008] The multi-angle image data are pictures taken at three angles of the solar panel collected by the defect detection mechanism when collecting each set of training data, and the pictures taken at the three angles are 45° pictures, 90° pictures and 135° pictures respectively;

[0009] The label data is the defect level of the solar panel when each set of training data is collected.

[0010] As a further solution of the present invention: the method for the control module to train a machine learning model for predicting the defect level of the solar panel is:

[0011] The control module uses the feature data in each set of training data as the input of the machine learning model, uses the defect level predicted by the machine learning model for the feature data as the output, uses the defect level in the label data corresponding to each set of feature data as the prediction target, and uses minimizing the sum of all prediction accuracies as the training target;

[0012] The calculation formula for the prediction accuracy is ud=(md-nd) 2 , where ud is the prediction accuracy, md is the predicted defect level corresponding to the d-th group of feature data, and nd is the defect level in the d-th group of label data. The machine learning model is trained until the sum of the prediction accuracies reaches convergence and the training is stopped.

[0013] As a further solution of the present invention: the defect detection mechanism includes a mechanism body, the bottom surface of the mechanism body is fixedly connected to a driving disk, the driving disk is rotatably connected to a rotating disk, a conveyor belt is installed on the inner side of the mechanism body, moving rails are arranged on both sides of the top surface of the mechanism body, and a sliding multi-angle synchronous acquisition mechanism is arranged on the moving rails.

[0014] As a further solution of the present invention: the multi-angle synchronous acquisition mechanism includes an arc track, both ends of the arc track are fixedly connected with ball nuts, and the side of the arc track is embedded with an image acquisition mechanism.

[0015] As a further solution of the present invention: the image acquisition mechanism includes a track slider, the inner side of which is penetrated by an arc groove matching the arc track, the bottom surface of the track slider is fixedly connected to the image collector, and the center position of the top surface of the track slider is fixedly connected to the drive box.

[0016] As a further solution of the present invention: the control module pre-processes the multi-angle image data, including the following steps:

[0017] S1: Normalize and grayscale the pictures taken at three angles of the solar panel to form the same width and height and corresponding grayscale values;

[0018] S2: Segment the captured image into m X×Y feature regions, and calculate the quantity parameters of the grayscale values ​​in the feature regions. The calculation formula of the quantity parameters is:

[0019] ;

[0020] Among them, M is the quantity parameter, i is the i-th gray value in the feature area;

[0021] S3: The quantity parameters are used as the input layer of the convolutional neural network;

[0022] S4: The convolution layer of the convolutional neural network includes several convolution kernels. Each convolution kernel slides on the captured image to calculate the weighted sum of the feature area and generate a feature map. The calculation formula of the weighted sum is:

[0023] ;

[0024] Among them, Z is the output of the convolution layer, Mi is the number parameter of the i-th feature area in the input captured image, Wi is the convolution kernel of the i-th feature area, and b is the bias term;

[0025] S5: Introduce nonlinear activation function and perform pooling operation;

[0026] S6: Expand the feature map into a one-dimensional vector and input it into the fully connected layer.

[0027] As a further solution of the present invention: the step S6 further comprises the following steps:

[0028] S6.1: Expand the multi-dimensional feature maps output by the convolutional layer and the pooling layer into a one-dimensional vector;

[0029] S6.2: The fully connected layer performs a linear transformation on the input one-dimensional vector;

[0030] R=SZ+c;

[0031] Among them, R is the result vector of linear transformation, Z is the output of the convolutional layer, S is the weight matrix, and c is the bias vector;

[0032] S6.3: Apply an activation function to the result vector of the linear transformation:

[0033] a = activation(R);

[0034] Where a is the output vector after applying the activation function;

[0035] S6.4: Use the Softmax activation function to convert the output vector into a probability distribution:

[0036] ;

[0037] Among them, Pj is the normalized probability, n is the number of elements in the output vector a, aj is the jth element in the output vector a, and Softmax outputs the probability distribution of each category, selecting the category with the highest probability as the prediction result.

[0038] As a further solution of the present invention: the defect level includes no defect, crack defect, black spot defect, and scratch defect;

[0039] Among them, crack defects, black spot defects and scratch defects include three levels: high, medium and low.

[0040] As a further solution of the present invention: several groups of the conveying mechanisms correspond to no defects, crack defects, black spot defects and scratch defects respectively.

[0041] Beneficial effects of the present invention:

[0042] 1. In the present invention, after the training of the machine learning model for predicting the defect level of the solar panel is completed, it can be used in the defect detection mechanism. The specific process is that one of the conveying mechanisms for conveying the undetected solar panel to the defect detection mechanism conveys the solar panel to the top surface of the defect detection mechanism. When the solar panel enters the top surface of the defect detection mechanism, the control module collects multi-angle image data of the solar panel in real time through the defect detection mechanism, and uses the machine learning model based on the multi-angle image data to predict the defect level corresponding to the solar panel. The defect detection mechanism conveys the solar panel to the corresponding conveying mechanism based on the defect level corresponding to the solar panel. In the process of solar panel defect detection, it is completely determined by the intelligent control module, which greatly improves the efficiency of solar panel defect detection, and also improves the accuracy of solar panel defect detection. It is very suitable for large-scale detection in factories, which is beneficial to reducing costs and increasing efficiency of factories and improving the competitiveness of factories.

[0043] 2. In the present invention, by setting up a multi-angle synchronous acquisition mechanism, the drive motor is turned on, and the output end of the drive motor drives the gear through the drive box. The rotating gear cooperates with the arc track to realize the free sliding of the track slider on the side of the arc track. Since it is necessary to collect pictures of the solar panel at three angles, the pictures at the three angles are 45° pictures, 90° pictures and 135° pictures respectively. Therefore, when working, the track slider slides to the 45°, 90° and 135° positions on the side of the arc track in turn, and the image collector collects a picture of the solar panel at the 45°, 90° and 135° positions respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below in conjunction with the accompanying drawings:

[0045] Figure 1It is a schematic diagram of the structure of the artificial intelligence solar panel defect detection device of the present invention;

[0046] Figure 2 is a top view of the artificial intelligence solar panel defect detection device of the present invention;

[0047] Figure 3 It is a structural schematic diagram of a defect detection mechanism in an artificial intelligence solar panel defect detection device of the present invention;

[0048] Figure 4 It is a structural schematic diagram of a multi-angle synchronous acquisition mechanism in an artificial intelligence solar panel defect detection device of the present invention;

[0049] Figure 5 It is a structural schematic diagram of the image acquisition mechanism in the artificial intelligence solar panel defect detection device of the present invention;

[0050] Figure 6 It is a flow chart of a method for pre-processing multi-angle image data by a control module in an artificial intelligence solar panel defect detection device of the present invention.

[0051] Reference numerals:

[0052] 1. Defect detection mechanism; 11. Mechanism body; 12. Driving disk; 13. Rotating disk; 14. Conveyor belt; 15. Moving track; 16. Multi-angle synchronous acquisition mechanism; 161. Arc track; 162. Ball nut; 163. Image acquisition mechanism; 1631. Track slider; 1632. Arc groove; 1633. Image collector; 1634. Driving box; 1635. Driving motor; 2. Conveying mechanism. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Example 1

[0055] like Figure 1 - Figure 5As shown, the present invention discloses an artificial intelligence solar panel defect detection device, including a defect detection mechanism 1, on the side of which are arranged a plurality of groups of conveying mechanisms 2, the plurality of groups of conveying mechanisms 2 are respectively used to convey the solar panels to the defect detection mechanism 1 and to receive solar panels with different defect levels, a control module is loaded on the defect detection mechanism 1, the control module collects a historical training data set in advance through the defect detection mechanism 1, and trains a machine learning model for predicting the defect level of the solar panel based on the historical training data set; the control module collects multi-angle image data of the solar panel in real time through the defect detection mechanism 1, and predicts the corresponding defect level of the solar panel using the machine learning model based on the multi-angle image data, the multi-angle image data at least including pictures taken at three angles of the solar panel; the defect detection mechanism 1 conveys the solar panel to the corresponding conveying mechanism 2 based on the corresponding defect level of the solar panel.

[0056] It should be noted that a microprocessor can be installed in the defect detection mechanism 1, and the control module is loaded on the microprocessor on the defect detection mechanism 1. Figure 1 As shown, several groups of conveying mechanisms 2 are evenly distributed on the side of the defect detection mechanism 1, one of which is used to convey the undetected solar panel to the defect detection mechanism 1. When the solar panel enters the top surface of the defect detection mechanism 1, the defect detection mechanism 1 can perform defect detection on the solar panel on its top surface. It should be emphasized that the solar panel is continuously moving on the top surface of the defect detection mechanism 1, and the defect detection mechanism 1 can perform defect detection on the continuously moving solar panel. When the defect detection mechanism 1 determines the defect level of the solar panel, it conveys the solar panel to another corresponding conveying mechanism 2. The conveying mechanism 2 is used to receive the detected solar panel. The length and specifications of the conveying mechanism 2 are adaptively selected by those skilled in the art according to the specific model and scenario of the solar panel. The specific process of the defect detection mechanism 1 on the solar panel is as follows:

[0057] First, the technical personnel in this field collects a historical training data set in advance through the defect detection mechanism 1. The historical training data set includes N groups of training data, where N is a positive integer. Each group of training data includes feature data and label data. The feature data is multi-angle image data. The amount of historical training data set collected is adaptively selected by the technical personnel in this field according to the specific data training situation.

[0058] The multi-angle image data are pictures taken at three angles of the solar panel collected by the defect detection mechanism 1 when collecting each set of training data. The pictures taken at the three angles are 45° pictures, 90° pictures and 135° pictures respectively. It should be noted that the 45° pictures, 90° pictures and 135° pictures refer to pictures taken at 45° on one side of the solar panel, pictures taken directly above the solar panel and pictures taken at 45° on the other side of the solar panel respectively. Those skilled in the art can also design pictures taken at any angle and collect any number of pictures of the same solar panel. These pictures are all used as training data for the machine learning model.

[0059] The label data is the defect level of the solar panel when each set of training data is collected. It should be noted that the defect level of the solar panel is preset by technicians in this field based on the actual defects of the solar panel.

[0060] After the technicians in this field have collected the historical training data set, they can train the designed machine learning model. The method for the control module to train the machine learning model for predicting the defect level of the solar panel is as follows:

[0061] The control module uses the feature data in each set of training data as the input of the machine learning model, and the defect level predicted by the machine learning model for the feature data as the output. The defect level in the label data corresponding to each set of feature data is used as the prediction target, and the training target is to minimize the sum of all prediction accuracies.

[0062] The calculation formula for prediction accuracy is ud= (md-nd) 2 , where ud is the prediction accuracy, md is the predicted defect level corresponding to the d-th group of feature data, and nd is the defect level in the d-th group of label data. The machine learning model is trained until the sum of the prediction accuracies converges. It should be noted that the convergence standard is selected by those skilled in the art according to the specific model training situation.

[0063] When the machine learning model for predicting the defect level of the solar panel is trained, it can be used on the defect detection mechanism 1. The specific process is that one of the conveying mechanisms 2 for conveying the uninspected solar panel to the defect detection mechanism 1 conveys the solar panel to the top surface of the defect detection mechanism 1. When the solar panel enters the top surface of the defect detection mechanism 1, the control module collects multi-angle image data of the solar panel in real time through the defect detection mechanism 1, and uses the machine learning model based on the multi-angle image data to predict the corresponding defect level of the solar panel. The defect detection mechanism 1 conveys the solar panel to the corresponding conveying mechanism 2 based on the corresponding defect level of the solar panel. In the process of solar panel defect detection, it is completely determined by the intelligent control module, which greatly improves the efficiency of solar panel defect detection, and also improves the accuracy of solar panel defect detection, which is very suitable for large-scale detection in factories, and is beneficial to reducing costs and increasing efficiency of factories, and improving the competitiveness of factories.

[0064] Example 2

[0065] like Figure 1-Figure 5 As shown, the defect detection mechanism 1 includes a mechanism body 11, the bottom surface of the mechanism body 11 is fixedly connected with a driving disk 12, and the driving disk 12 is rotatably connected with a rotating disk 13. It should be noted that the rotating disk 13 can be directly fixed on the ground, and the center position of the top surface of the rotating disk 13 is connected to the driving disk 12 through a bearing. A motor can be installed inside the rotating disk 13, and the output shaft of the motor is meshed with the rotating shaft of the driving disk 12 through a gear. When in use, turn on the motor, and the output shaft of the motor will directly drive the driving disk 12 to rotate through the gear, and the rotating driving disk 12 can drive the mechanism body 11 to rotate. Since several conveying mechanisms 2 are equidistantly arranged in a circular arc shape on the side of the defect detection mechanism 1, when the control module predicts the defect level corresponding to the solar panel, the defect detection mechanism 1 needs to rotate to the corresponding conveying mechanism 2, and then convey the solar panel on the top surface of the defect detection mechanism 1 to the conveying mechanism 2 to complete the defect detection of the solar panel. As well as transportation, a conveyor belt 14 is installed on the inner side of the mechanism body 11, and moving tracks 15 are arranged on both sides of the top surface of the mechanism body 11, and a sliding multi-angle synchronous collection mechanism 16 is arranged on the moving track 15. It should be noted that a rotating shaft is installed at both ends of the inner side of the conveyor belt 14, and a motor is installed at a position close to the rotating shaft inside the mechanism body 11. The output shaft of the motor is directly connected to the rotating shaft through a coupling. When in use, turn on the motor, and the power of the motor will be transmitted to the rotating shaft through the output shaft, and the rotating rotating shaft will drive the conveyor belt 14 to rotate. When the solar panel is on the top surface of the conveyor belt 14, the rotating conveyor belt 14 will drive the solar panel, thereby realizing the function of transporting the solar panel. In addition, it should be noted that the top surface of the conveyor belt in the conveying mechanism 2 should be flush with the top surface of the conveyor belt 14, so as to ensure that the solar panel will not vibrate when moving between the conveyor belt 14 and the conveying mechanism 2, thereby improving the protection effect of the solar panel.

[0066] like Figure 3 - Figure 5 As shown, the multi-angle synchronous acquisition mechanism 16 includes an arc track 161, and ball nuts 162 are fixedly connected to both ends of the arc track 161. The side of the arc track 161 is embedded with an image acquisition mechanism 163. It should be noted that a lead screw is installed inside the movable track 15 through a rotating shaft. The lead screw and the ball nuts 162 at both ends of the arc track 161 match each other. A motor is installed near the end of the lead screw inside the movable track 15. The output shaft of the motor is directly connected to the lead screw through a coupling. When in use, the motor is turned on. The output shaft of the motor transmits the motor to the screw through the coupling. The rotating screw cooperates with the ball nut 162 to realize the free movement of the arc track 161 along the track direction on the top surface of the conveyor belt 14 and control the moving speed of the arc track 161. When the image acquisition mechanism 163 needs to collect multi-angle image data of the solar panel, the moving speed of the arc track 161 is consistent with the rotation speed of the conveyor belt 14, thereby ensuring the relative stillness of the image acquisition mechanism 163 and the solar panel, thereby improving the acquisition quality of multi-angle image data.

[0067] like Figure 4 and Figure 5 As shown, the image acquisition mechanism 163 includes a track slider 1631, the inner side of which is penetrated by an arc groove 1632 matching the arc track 161, the bottom surface of the track slider 1631 is fixedly connected with an image collector 1633, and the center position of the top surface of the track slider 1631 is fixedly connected with a drive box 1634. It should be noted that the top surface of the drive box 1634 is connected with a drive motor 1635 by bolts. The drive box 1634 is a reduction box, including an input end and an output end. The rotating shaft of the output end is connected with a gear, and the gear passes through the track slider 1631 and meshes with the side of the arc track 161. The output shaft of the drive motor 1635 is directly connected to the input end of the drive box 1634. When in use, the drive motor 1635 is turned on and the drive motor 1635 is turned on. The output end drives the gear through the drive box 1634, and the rotating gear cooperates with the arc track 161, so that the track slider 1631 can slide freely on the side of the arc track 161. Since it is necessary to collect pictures of the solar panel at three angles, the pictures at the three angles are 45° pictures, 90° pictures and 135° pictures respectively. Therefore, when working, the track slider 1631 slides to the 45°, 90° and 135° positions on the side of the arc track 161 in turn, and the image collector 1633 collects pictures of the solar panel at the 45°, 90° and 135° positions respectively. It should be noted that the image collector 1633 can be a high-definition camera, and the specific model and specifications are adaptively selected by technical personnel in this field according to the pictures taken.

[0068] Example 3

[0069] like Figure 6 As shown, the control module pre-processes the multi-angle image data, including the following steps:

[0070] S1: Normalize and grayscale the pictures taken at three angles of the solar panel to form the same width and height and corresponding grayscale values;

[0071] It should be noted that the pictures taken at three angles are first converted into grayscale pictures, and then the pixel values ​​of the grayscale pictures are normalized to the range of 0 to 1. Although the pictures taken at three angles are taken of the same solar panel, the pixel distribution of the grayscale pictures obtained at the end will also change due to the different shooting angles. Therefore, the grayscale values ​​of the pictures taken at the three angles are different.

[0072] S2: Segment the captured image into m X×Y feature regions, and calculate the quantity parameters of the grayscale values ​​in the feature regions. The calculation formula of the quantity parameters is:

[0073] ;

[0074] Among them, M is the quantity parameter, i is the i-th grayscale value in the feature area. It should be noted that the number of m is determined by X×Y and the height and width of the captured image. For example, if the height and width of the captured image are 300×400 respectively, then X and Y can be 3 and 4, so that the captured image can be divided into 100 3×4 feature areas. The control module needs to calculate the number parameters of the grayscale values ​​in each feature area;

[0075] S3: The quantity parameters are used as the input layer of the convolutional neural network;

[0076] It should be noted that the quantity parameter is used as the input parameter of the input layer of the convolutional neural network;

[0077] S4: The convolution layer of the convolutional neural network includes several convolution kernels. Each convolution kernel slides on the captured image to calculate the weighted sum of the feature area and generate a feature map. The calculation formula of the weighted sum is:

[0078] ;

[0079] Wherein, Z is the output of the convolution layer, Mi is the number parameter of the i-th feature region in the input captured image, and Wi is the convolution kernel of the i-th feature region. It should be noted that since the captured image is divided into m X×Y feature regions, the size of the convolution kernel is set to be the same as the feature region, that is, the size of the feature region and the convolution kernel are both X×Y. The specific value of X×Y is adaptively selected by those skilled in the art according to the specific resolution of the captured image. If there are m feature regions in the captured image, there will be m number parameters M, namely M1, M2...Mm. b is the bias term, which is an additional learnable parameter. It is added to the weighted sum after each convolution operation to introduce more flexibility and expressiveness. The role of the bias term in the neural network is similar to the constant term in a linear equation. It helps the model produce non-zero output when the input is zero, thereby improving the performance of the model.

[0080] S5: Introduce a nonlinear activation function and perform a pooling operation. It should be noted that the type of the nonlinear activation function is adaptively selected by those skilled in the art according to the convolutional neural network.

[0081] S6: Expand the feature map into a one-dimensional vector and input it into the fully connected layer; specifically, it includes the following steps:

[0082] S6.1: Expand the multi-dimensional feature maps output by the convolutional layer and the pooling layer into a one-dimensional vector;

[0083] S6.2: The fully connected layer performs a linear transformation on the input one-dimensional vector;

[0084] R=SZ+c;

[0085] Among them, R is the result vector of linear transformation, Z is the output of convolution layer, S is weight matrix, c is bias vector. It should be noted that weight matrix S is obtained through training process. At the beginning of training, weight matrix is ​​usually randomly initialized, and elements in weight matrix are randomly sampled from uniform distribution or normal distribution. c is bias vector, which is also obtained through training process. Before model training, bias vector is randomly initialized, prediction result and true label are compared, loss value is calculated, gradient of loss function with respect to bias vector is calculated, and optimization algorithm is used to update bias vector according to gradient. After multiple iterations of training, bias vector and weight matrix are gradually adjusted together, so that prediction result of model becomes more and more accurate, and finally better performance is achieved.

[0086] S6.3: Apply an activation function to the result vector of the linear transformation:

[0087] a = activation(R);

[0088] Where a is the output vector after applying the activation function;

[0089] S6.4: Use the Softmax activation function to convert the output vector into a probability distribution:

[0090] ;

[0091] Among them, Pj is the normalized probability, n is the number of elements in the output vector a, aj is the jth element in the output vector a, and Softmax outputs the probability distribution of each category. The category with the highest probability is selected as the prediction result to classify the pictures taken by the solar panel. For example, the output vector a=[a1, a2, ..., an], first calculate the index of each element in the input vector , and then calculate the sum of all exponentials , and finally divide each exponent by the sum to get the normalized probability Pj.

[0092] The defect levels include no defect, crack defect, black spot defect and scratch defect, among which crack defect, black spot defect and scratch defect all include three levels: high, medium and low. When the solar panel detected by the defect detection agency 1 is defect-free, it means that the solar panel is of good quality and can proceed to the next process directly. When the solar panel detected by the defect detection agency 1 is a low-level crack defect, black spot defect or scratch defect, a special EL test is required to further determine whether its internal defects meet the production requirements. When the solar panel detected by the defect detection agency 1 is a moderate or high-level crack defect, black spot defect or scratch defect, it is necessary to conduct a key manual inspection on the solar panel and give a recycling instruction for the solar panel.

[0093] Example 3

[0094] like Figure 1 and Figure 2 As shown, several groups of conveying mechanisms 2 correspond to no defects, crack defects, black spot defects and scratch defects respectively. It should be noted that when the defect detection mechanism 1 performs defect detection on the solar panel, Figure 2For example, the leftmost one is conveying mechanism No. 1, and then conveying mechanisms No. 2 to No. 6 are marked clockwise, where conveying mechanism No. 1 is used to convey uninspected solar panels to defect detection mechanism 1 for defect detection, and conveying mechanism No. 4 is used to receive non-defective solar panels. If the solar panel detected by defect detection mechanism 1 is non-defective, the defect detection mechanism 1 directly conveys the solar panel to conveying mechanism No. 4. If the solar panel detected by defect detection mechanism 1 is a low-level crack defect, black spot defect, or scratch defect, a special EL test is required, and the solar panel on the defect detection mechanism 1 can be conveyed to conveying mechanism No. 3 or conveying mechanism No. 5. Conveying mechanism No. 3 or conveying mechanism No. 5 is used to convey the solar panel to the EL detection equipment for EL detection. If the solar panel detected by the defect detection mechanism 1 is a medium-level or high-level crack defect, black spot defect, or scratch defect, it is necessary to focus on manual inspection of the solar panel, specifically, the defect detection mechanism 1 conveys it to conveying mechanism No. 2 or conveying mechanism No. 6, and the solar panel is manually inspected. The inspection method is adaptively selected by technical personnel in this field according to the specific situation on site.

[0095] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An artificial intelligence solar panel defect detection device, characterized in that: include: A defect detection mechanism (1) is provided with a conveying mechanism (2) on its side, and the conveying mechanism (2) is used to convey the solar panels to the defect detection mechanism (1) and receive solar panels with different defect levels; A control module, which is loaded on the defect detection mechanism (1), and the control module collects a historical training data set in advance through the defect detection mechanism (1), and trains a machine learning model for predicting the defect level of the solar panel based on the historical training data set; The control module collects multi-angle image data of the solar panel in real time through a defect detection mechanism (1), and predicts the defect level corresponding to the solar panel using a machine learning model based on the multi-angle image data, wherein the multi-angle image data includes pictures taken at least three angles of the solar panel; The defect detection mechanism (1) transports the solar panel to the corresponding transport mechanism (2) based on the defect level corresponding to the solar panel; The historical training data set includes N groups of training data, each group of training data includes feature data and label data, and the feature data is multi-angle image data; The multi-angle image data are pictures taken at three angles of the solar panel collected by the defect detection mechanism (1) when collecting each set of training data, and the pictures taken at the three angles are pictures taken at 45°, pictures taken at 90°, and pictures taken at 135°; The control module pre-processes the multi-angle image data, including the following steps: S1: Normalize and grayscale the pictures taken at three angles of the solar panel to form the same width and height and corresponding grayscale values; S2: Segment the captured image into m X×Y feature regions, and calculate the quantity parameters of the grayscale values ​​in the feature regions. The calculation formula of the quantity parameters is: ； Among them, M is the quantity parameter, i is the i-th gray value in the feature area; S3: The quantity parameters are used as the input layer of the convolutional neural network; S4: The convolution layer of the convolutional neural network includes multiple convolution kernels. Each convolution kernel slides on the captured image to calculate the weighted sum of the feature area and generate a feature map. The calculation formula of the weighted sum is: ； Among them, Z is the output of the convolution layer, Mi is the number parameter of the i-th feature area in the input captured image, Wi is the convolution kernel of the i-th feature area, and b is the bias term; S5: Introduce nonlinear activation function and perform pooling operation; S6: Expand the feature map into a one-dimensional vector and input it into the fully connected layer; The step S6 further comprises the following steps: S6.1: Expand the multi-dimensional feature maps output by the convolutional layer and the pooling layer into a one-dimensional vector; S6.2: The fully connected layer performs a linear transformation on the input one-dimensional vector; R=SZ+c; Among them, R is the result vector of linear transformation, Z is the output of the convolutional layer, S is the weight matrix, and c is the bias vector; S6.3: Apply an activation function to the result vector of the linear transformation: a = activation(R); Where a is the output vector after applying the activation function; S6.4: Use the Softmax activation function to convert the output vector into a probability distribution: ； Among them, Pj is the normalized probability, e is the base of the natural logarithm, n is the number of elements in the output vector a, aj is the jth element in the output vector a, and Softmax outputs the probability distribution of each category, selecting the category with the highest probability as the prediction result; The label data is the defect level of the solar panel when collecting each set of training data; The method for training a machine learning model for predicting defect levels of solar panels is: The control module uses the feature data in each set of training data as the input of the machine learning model, uses the defect level predicted by the machine learning model for the feature data as the output, uses the defect level in the label data corresponding to each set of feature data as the prediction target, and uses minimizing the sum of all prediction accuracies as the training target; The calculation formula for the prediction accuracy is ud=(md-nd) 2 , where ud is the prediction accuracy, md is the predicted defect level corresponding to the d-th group of feature data, and nd is the defect level in the d-th group of label data. The machine learning model is trained until the sum of the prediction accuracies reaches convergence and the training is stopped.

2. The artificial intelligence solar panel defect detection device according to claim 1 is characterized in that: The defect detection mechanism (1) comprises a mechanism body (11), the bottom surface of the mechanism body (11) is fixedly connected to a driving disk (12), the driving disk (12) is rotatably connected to a rotating disk (13), a conveyor belt (14) is installed on the inner side of the mechanism body (11), and moving rails (15) are arranged on both sides of the top surface of the mechanism body (11), and a sliding multi-angle synchronous acquisition mechanism (16) is arranged on the moving rails (15).

3. The artificial intelligence solar panel defect detection device according to claim 2 is characterized in that: The multi-angle synchronous acquisition mechanism (16) comprises an arc track (161), both ends of the arc track (161) are fixedly connected with ball nuts (162), and the side surface of the arc track (161) is embedded with an image acquisition mechanism (163).

4. The artificial intelligence solar panel defect detection device according to claim 3 is characterized in that: The image acquisition mechanism (163) comprises a track slider (1631), the inner side of the track slider (1631) being penetrated by an arc-shaped groove (1632) matching the arc-shaped track (161), the bottom surface of the track slider (1631) being fixedly connected to an image acquisition device (1633), and the center position of the top surface of the track slider (1631) being fixedly connected to a drive box (1634).

5. The artificial intelligence solar panel defect detection device according to claim 1 is characterized in that: The defect levels include no defect, crack defect, black spot defect, and scratch defect; Among them, crack defects, black spot defects and scratch defects include three levels: high, medium and low.

6. The artificial intelligence solar panel defect detection device according to claim 5, characterized in that: The conveying mechanism (2) comprises a plurality of groups, corresponding to no defect, crack defect, black spot defect and scratch defect respectively.

Images