Quality inspection method and system for photovoltaic module busbar and solder strip

Abstract

The application discloses a quality inspection method for photovoltaic module busbars and solder strips, which comprises the following steps: obtaining images, AI detection, threshold filtering and processing detection, first, obtaining images at predetermined positions in the photovoltaic module busbar and solder strip area, processing the obtained images to obtain original images required by AI detection software, and saving the original images to a predetermined address; then, AI detection is performed, including model design, data set preparation and training, model prediction and AI return of defect information, collecting and training the data obtained by model processing, then predicting a defect dictionary, returning the defect dictionary to the system software, then performing threshold filtering, filtering the AI defect information, and saving the filtered defect image to a predetermined position; finally, based on the filtered defect image saved to the predetermined position, a detection result is output, and whether the photovoltaic module busbar and solder strip are qualified is judged. The quality inspection method and system provided by the application can effectively improve the quality inspection accuracy and production efficiency, and greatly reduce the labor cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic module equipment quality inspection, and particularly relates to a quality inspection method and system for photovoltaic module bus bars and solder strips. BACKGROUND

[0002] A photovoltaic module is a device that realizes photoelectric conversion by using a photoelectric effect. After a cell module is processed by a soldering machine, some serious defects, including tin virtual welding and solder strips that are too long, may occur in the bus bar and solder strip area, and there is a risk of degradation of the module. At present, there are two methods of manual quality inspection and computer vision software quality inspection.

[0003] For manual quality inspection, because the defect types that may occur in the bus bar and solder strip of the photovoltaic module are various, a lot of time is spent on checking a complete module, and there is a risk of missing some defects, especially for many subtle defect types, which seriously affects the production of the workshop line and the overall quality of the photovoltaic module.

[0004] For computer vision software quality inspection, a 3D point cloud image of a target solder joint is first obtained, then the 3D point cloud image of the target solder joint is processed to obtain a plurality of continuous 2D images corresponding to the target solder joint, then the plurality of continuous 2D images corresponding to the target solder joint are input into a preset three-dimensional convolutional neural network-based detection model, and finally the plurality of continuous 2D images corresponding to the target solder joint are checked by the detection model, and the checking result is that the target solder joint is qualified or unqualified. The technical problems are as follows:

[0005] 1. The current computer vision solder strip quality inspection software can only inspect the solder joint itself, and cannot detect other defect types such as fish hooks, no solder strips, and cell string spacing, which cannot meet the actual quality inspection needs of the photovoltaic industry;

[0006] 2. The current computer vision solder strip quality inspection software has high imaging requirements, so the cost of the camera and lens is a big problem, and the operation requirements are also increased, making it difficult to use;

[0007] 3. The current computer vision solder strip quality inspection software has low accuracy, and the robustness and generalization performance of the data are also low. Once the appearance of the module is adjusted by the manufacturer, the accuracy of the detection is unstable, and it is necessary to collect data and retrain, which is very cumbersome to use. SUMMARY

[0008] In a first aspect, the present application provides a quality inspection method for photovoltaic module bus bars and solder strips, which comprises the following steps in sequence:

[0009] S10, acquiring a picture, acquiring images at predetermined positions of a photovoltaic module busbar and a solder strip area, processing the acquired images to obtain an original picture required by AI detection software, and saving the original picture to a predetermined address;

[0010] S20, AI detection, comprising:

[0011] S21, model design, respectively designing a large picture cutting small picture model, a small picture defect detection model and a defect picture small target detection model;

[0012] S22, data set making and training, labeling a cell piece label corresponding to the large picture cutting small picture model to obtain coordinate information data of the large picture cutting small picture; labeling a defect type corresponding to the small picture defect detection model to obtain defect type data; labeling a welding scar corresponding to the defect picture small target detection model to obtain length information data of the welding scar on the solder strip; and making these data into training and verification data sets for AI target detection model training;

[0013] S23, model prediction, predicting defects of the photovoltaic module busbar and the solder strip required to be returned to the system based on the trained data to form a defect dictionary;

[0014] S24, AI returns defect information, returning the defect dictionary to the system software;

[0015] S30, threshold filtering, filtering AI defect information, and saving the filtered defect picture to a predetermined position;

[0016] S40, processing detection, outputting a detection result based on the filtered defect picture saved to the predetermined position, wherein the detection result includes a result image, a defect type, a defect position corresponding to the defect type and a final result of whether it is qualified.

[0017] According to an embodiment of the present application, in step S10, the images at the predetermined positions are three images of the photovoltaic module busbar and the solder strip area on the left, in the middle and on the right, and the predetermined address is local or cloud.

[0018] According to an embodiment of the present application, in step S21, the model structures of the large picture cutting small picture model, the small picture defect detection model and the defect picture small target detection model are the same, and all include convolution layers, up-sampling layers and pooling layers, batch normalization layers, activation functions and loss functions, wherein each convolution layer is composed of a plurality of convolution units, and the parameters of each convolution unit are obtained by a back propagation algorithm.

[0019] According to an embodiment of the present application, the loss function includes a cross-entropy loss function, a positioning loss function, a confidence loss function and a complete intersection-over-union loss function.

[0020] According to an embodiment of the present application, in step S22, the data is made into a training and verification data set, including data collection, data preprocessing and data training, wherein the data collection includes: collecting real defects of the busbar and the solder strip of the photovoltaic module on the pipeline in priority, and then, for defects less than a predetermined number, supplementing the data amount by artificially creating defects or generating defects by data simulation;

[0021] The data preprocessing includes: first, data cleaning is performed to filter out dirty data affecting the detection accuracy of the model, and then data enhancement operation is performed on the labeled defects;

[0022] The data training includes: maintaining the weight of each round during the training process, obtaining the model with the best learning effect through loss result and PR curve evaluation parameters, and ensuring that the performance can meet the pipeline beat requirements.

[0023] According to an embodiment of the present application, in step S23, the defects of the busbar and the solder strip of the photovoltaic module that need to be returned to the system are predicted based on the trained data to form a defect dictionary, specifically including:

[0024] S231, based on the storage address of the original image, an image at a predetermined position of the busbar and the solder strip of the photovoltaic module is obtained, scaling processing is performed to obtain a small image, the small image is put into the large image cutting small image model for detection to obtain a defect frame and coordinate information of each string of cell pieces, and the original image is divided into small images corresponding to each string of cell pieces based on the coordinate information;

[0025] S232, respectively calling threads to further scale the small images corresponding to each string of cell pieces, and then putting them into the small image defect detection model to detect small image defect frames and spacing area frames including string spacing, busbar spacing and creeping distance;

[0026] S233, based on the small image defect frame, the small image corresponding to each string of cell pieces is cut out, then scaling processing is performed, and the small image is transmitted to the small target defect detection model to obtain small defect data information, and then the small defect data information is inversely calculated into real millimeter values through the millimeter-pixel ratio to form the defect dictionary.

[0027] According to an embodiment of the present application, in step S232, the small image defect frame includes a virtual welding defect frame, and for virtual welding, in addition to returning to the small image defect frame, it also includes cutting out a small image through the virtual welding defect frame coordinates, then performing CV operation on the small image to obtain a mask corresponding to the tin part, calculating the area of white pixel points, and adding them to the defect dictionary.

[0028] According to an embodiment of the present application, in step S232, the small picture defect frame further comprises a welding rod too long, a welding deviation and a welding strip too short defect frame, and for the defects of the welding rod too long, the welding deviation and the welding strip too short, the small picture defect frame is returned, and the defect length and / or width corresponding to the requirement are obtained and then added to the defect dictionary.

[0029] According to an embodiment of the present application, in step S232, for the spacing area frame, a small picture is cut out through spacing area frame coordinates, and then a spacing type result is calculated through a CV image algorithm, wherein the spacing type result comprises results of string spacing, busbar spacing and creeping distance.

[0030] In a second aspect, the present application further provides a quality inspection system for photovoltaic module busbars and welding strips, which adopts the quality inspection method for photovoltaic module busbars and welding strips as described above, and comprises:

[0031] a picture acquisition module, configured to acquire images at predetermined positions of the photovoltaic module busbar and welding strip area, process the acquired images to obtain original pictures required by AI detection software, and save the original pictures to a predetermined address;

[0032] an AI detection module, comprising a model design unit, a data set making and training unit, a model prediction unit and an AI returned defect information unit;

[0033] a threshold filtering module, configured to obtain defects of the photovoltaic module busbar and welding strip required to be returned to the system based on the trained data prediction, and form a defect dictionary;

[0034] a processing and detection module, configured to output a detection result based on the filtered defect pictures saved to the predetermined position, wherein the detection result comprises a result image, a defect type, a defect position corresponding to the defect type and a final result of whether it is qualified.

[0035] The present application has the following beneficial effects compared with the prior art: the quality inspection method and system for photovoltaic module busbars and welding strips provided by the present application can accurately control the quality inspection accuracy by adjusting the detection threshold in real time, the system has strong operability, the missed detection rate of the photovoltaic module busbar and welding strip area can reach <=5%, the false detection rate can reach <=1%, the production efficiency of the photovoltaic module can be effectively improved, the quality of the photovoltaic module can be ensured, and the labor cost can be greatly reduced.

[0036] These and other objects, features and advantages of the present application will become apparent with reference to the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the quality inspection method for photovoltaic module busbars and welding strips of the present application is shown.

[0038] Figure 2 A structure diagram of a quality inspection system for photovoltaic module busbars and solder strips according to the present application is shown. DETAILED DESCRIPTION

[0039] The following description is provided to enable those skilled in the art to realize the present application. The preferred embodiments in the following description are only examples to illustrate the present application. Other obvious modifications can be made by those skilled in the art. The basic principles defined in the following description can be applied to other embodiments, modifications, improvements, equivalents and other technical solutions without departing from the spirit and scope of the present application.

[0040] Those skilled in the art should understand that in the disclosure of the specification, the orientation or positional relationship indicated by the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like are based on the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore, the above terms cannot be understood as a limitation of the present application.

[0041] It can be understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of one element can be one, and in another embodiment, the number of the element can be multiple, and the term "one" cannot be understood as a limitation on the number.

[0042] Reference Figure 1 In a first aspect, a quality inspection method for photovoltaic module busbars and solder strips according to a preferred embodiment of the present application will be described in detail below, wherein the quality inspection method for photovoltaic module busbars and solder strips comprises the following steps in sequence:

[0043] S10, acquiring pictures, adjusting the camera for shooting pictures, such as a line scanning camera, using the camera software to shoot images at the predetermined positions of the photovoltaic module busbar and solder strip area, such as acquiring three pictures of the left, middle and right component busbar and solder strip area respectively, and then processing the obtained images through rotation and cropping to obtain the original image required by the AI detection software, and saving the original image to a predetermined address, wherein the predetermined address can be a local computer or a cloud server, wherein the adjustment of the camera can be adjusted through a support, and the camera is installed beside the conveyor belt of the pipeline for detection through the support without direct contact with the conveyor belt to prevent the camera from shaking, in addition, three cameras can be installed directly above the three busbar positions of the photovoltaic module to be able to shoot the complete busbar and solder strip area of the photovoltaic module, and the busbar is located in the middle of the imaging, in addition, a small spotlight can also be used to supplement the light source to avoid insufficient image brightness affecting the detection effect, the image needs to be clear and the brightness needs to be appropriate, for example, the solder strip area of the busbar and the tin need to be bright and unobstructed, and there is no case of overexposure, and the line scanning camera line frequency and other parameters also need to be controlled to prevent image compression;

[0044] S20, AI detection, including:

[0045] S21, model design, the camera shooting original image is three pictures with a width of about 20000 pixel points and a height of 1280 pixel points, corresponding to the three busbar areas of the component, since the picture is too large, three models with different parameter sizes and network depths but the same overall structure are designed, which are respectively designed as: a large picture cutting small picture model, a small picture defect detection model and a small target detection model after defect detection of the small picture, and the functions realized by them are respectively large picture cutting small picture, small picture defect detection and small target detection after defect detection of the small picture;

[0046] S22, making data set and training, corresponding to the large picture cutting small picture model, labeling the battery piece label, including battery piece, spot pressing trace and jumper, to obtain the coordinate information data of the large picture cutting small picture; corresponding to the small picture defect detection model, labeling the defect type, wherein the defect type includes tin virtual welding, solder strip being too long, solder strip being too short, soldering deviation, no solder strip, fishhook, solder strip separating from the silver paste point, battery piece dirt, busbar foreign matter, string spacing, bus connection distance and creepage distance, to obtain the defect type data; corresponding to the small target detection model of the defect picture, labeling the welding scar, to obtain the length information data of the welding scar on the solder strip; then the above data (coordinate information data of the large picture cutting small picture, defect type data and length information data of the welding scar) are made into training and verification data set for AI target detection model training;

[0047] S23, model prediction, based on the trained data to predict the defects of the photovoltaic module busbar and solder strip returned to the system to form a defect dictionary, wherein in the prediction, the float32 floating point parameters in the trained model are converted to half-precision (half) data type to reduce the GPU memory consumption and improve the prediction speed of the AI target detection model, and the picture data input into the model is transposed from (H, W, C) to (C, H, W), then needs to be upsampled to (N, C, H, W), and converted to FP16 half-precision and transmitted to the half-precision AI target detection model for prediction; the prediction result exists in a case of multiple prediction boxes for a target, therefore, DIOU-NMS is used to select the prediction box closest to the real position of the target, and DIOU is more capable of ensuring the center point of the prediction box and approaching the center point of the real target than IOU;

[0048] Specifically comprising:

[0049] S231, based on the storage address of the original image, such as a local computer or a cloud server, the image at the predetermined position of the photovoltaic module busbar and solder strip area is obtained, three threads are called to scale the corresponding three large images to 1280*1280 small images, and then the small images are put into the large image cutting small image model for detection to obtain the defect frame and the coordinate information of each string of cell pieces, based on which the original image is divided into small images corresponding to each string of cell pieces, that is, the long image obtained is cut into small images of each string of cell pieces by the first AI, and the coordinate information of this step is used for cutting the image by the first AI;

[0050] S232, the corresponding small images of each string of cell pieces are scaled to 1080*1080 in size respectively by calling threads, and then put into the small image defect detection model to detect the small image defect frame and the spacing area frame including the string spacing, busbar spacing and creepage distance, which is equivalent to transmitting the small image cut out by the first AI to the second AI;

[0051] S233, based on the small image defect frame, the corresponding small image of each string of cell pieces is cut out, then scaled, the small image is further scaled to 320*320, and then transmitted to the small target defect detection model to obtain the small defect data information, and then the small defect data information is calculated back to the real millimeter value through the millimeter-pixel ratio to form the defect dictionary, for example, the small target detection model for virtual welding scab can obtain the length of the small welding scab, which is accumulated to form the small welding scab defect in the defect dictionary, and the virtual welding detection standard can be set according to different factories, such as whether the virtual welding is bare or has a welding scab below 1mm;

[0052] S24, the AI returns the defect information, and returns the defect dictionary to the system software, wherein the defect information in the returned defect dictionary includes the specific coordinates of the small defect frame, the row and column coordinates of the photovoltaic module corresponding to the defect, the length (width, area), and the defect confidence;

[0053] S30, threshold filtering, filtering the AI defect information through the threshold filtering software, setting the confidence threshold of each defect category detection item of the AI, and returning the required defect information to the system software according to the confidence threshold rule by the AI software, saving the corresponding detection result to the local computer or the cloud server, and filtering out the AI defect information with low defect severity through the threshold filtering software by setting the rules of length, width or area, and saving the defect map left after the rule filtering to the local computer or the cloud server, wherein the high and low of the defect severity can be adjusted and set by the relevant subject party, such as the factory or the use end;

[0054] S40, processing detection, based on the filtered defect map saved to the predetermined position, through the docking of the system software and the result software interface, the system software transmits the camera original map and the detection result to the result display software, and outputs the detection result, wherein the detection result includes the result image, the defect type (or the defect name), the defect position corresponding to the defect type (or the defect name), and the final result of whether it is qualified, in addition, if there is an unqualified case, i.e. NG case, an alarm signal can be formed through the PLC control system to remind manual processing.

[0055] In one embodiment, in step S21, the model structures of the large-to-small image model, the small image defect detection model and the defect map small target detection model are the same, and all include convolution layers, up-sampling layers and pooling layers, batch normalization layers, activation functions and loss functions;

[0056] Each of the convolution layers is composed of a plurality of convolution units, and the parameters of each convolution unit are obtained by the back propagation algorithm;

[0057] The activation function is introduced mainly to increase the nonlinearity of the model, such as the introduction of the RELU activation function, which has small calculation amount, stable nonlinearity, and can effectively alleviate the problems of gradient disappearance and overfitting;

[0058] Wherein for the batch normalization layer, mainly considering that in the training process of the neural network, as the depth deepens, the input value distribution will be offset, and will be close to the upper and lower ends of the value interval, such as the Sigmoid function, which will cause the gradient of the low-level neural network to disappear during back propagation, the batch normalization layer (Batch Normalization, referred to as BN layer) can forcibly pull the distribution of the input value of each neural network layer to the standard normal distribution with a mean of 0 and a variance of 1 through certain normalization means, thereby avoiding problems such as gradient disappearance and gradient explosion, and can accelerate model convergence and improve training efficiency.

[0059] Further preferably, the upsampling layer and the pooling layer are also referred to as the upsampling layer and the subsampling layer, wherein the upsampling layer is used to restore the output layer size to ensure that it can be stacked with the input layer in the channel dimension, i.e. the routing operation, and the subsampling layer is used to reduce the size of the input data while preserving the original features of the input data, such as using a convolution with a stride of 2, and adaptively adjusting the parameters of the convolution, which is more flexible than the fixed pooling layer, which can effectively ensure the accuracy and quality of quality inspection.

[0060] Further preferably, the loss function includes a cross-entropy loss function, a positioning loss function, a confidence loss function, and a complete intersection-over-union loss function.

[0061] Wherein for the cross-entropy loss function, since the classification result uses a logistic function to obtain a probability, when combined with the cross-entropy function, especially when the model effect is poor, the learning speed can be improved.

[0062] Wherein the positioning loss function uses a mean squared error loss function (MSELoss);

[0063] Wherein the confidence loss function uses IOULoss, wherein IOU is a standard for measuring the accuracy of detecting corresponding objects in a specific data set, and the specific calculation method is to divide the overlapping part of the two regions by the set part of the two regions, and compare the result with the set threshold value;

[0064] Wherein the complete intersection-over-union loss function, i.e. CIOULoss, is used in training, so that the center point of the prediction frame can be close to the center point of the real frame, and the length-width ratio of the prediction frame and the real frame is also considered;

[0065] After obtaining these loss values, the gradient calculation and back propagation are used to adjust the weights and biases to optimize the overall performance of the neural network.

[0066] It is also worth mentioning that in step S22, the making of the data into training and verification data sets includes data collection, data preprocessing and data training, wherein the data collection includes: collecting real defects of the busbars and the welding strips of the photovoltaic components on the production line in priority, and then supplementing the data quantity by artificially creating defects for the defects whose number is less than the predetermined number, i.e. the defects of the category with less number, and for some defects that are difficult to create, the defects can also be generated by data simulation, such as data simulation by PS or generative adversarial network.

[0067] The data preprocessing includes: firstly, data cleaning is performed to filter out dirty data affecting the detection accuracy of the model, and then data augmentation operation is performed on the labeled defects to enhance the diversity of the data, ensure the generalization of the AI, and improve the learning efficiency of the model, wherein the data augmentation includes changing the brightness, contrast, rotation angle, offset distance, size, etc. of the image, and Mix-up (mixing two random samples in proportion, and the classification result is distributed in proportion) and Mosaic (separately flipping, scaling, etc. multiple pictures, and placing them as one picture, combining the picture and the target box) are also used.

[0068] The data training includes: maintaining the weight of each round in the training process, obtaining the model with the best learning effect by evaluating the parameters through the loss result and the PR curve (P represents precision (precision), and R represents recall (recall), which represents the relationship between precision and recall. Generally, the recall is set as the horizontal coordinate, and the precision is set as the vertical coordinate. The area surrounded by the PR curve is AP, and the average value of all class AP is Map), and ensuring that the performance can meet the beat requirements of the production line to ensure that the production and processing efficiency of the production line is not reduced.

[0069] Further preferably, in step S232, the small defect frame includes a virtual welding defect frame, and for virtual welding, in addition to returning to the small defect frame, a small frame is also cut out through the virtual welding defect frame coordinates, and then the small frame is subjected to CV operation such as binarization to obtain a mask corresponding to the tin plating part, the area of the white pixel points is calculated, and is added to the defect dictionary.

[0070] Further preferably, in step S232, the small defect frame also includes a welding rod too long, welding offset and welding strip too short defect frame, and for the defects of the welding rod too long, welding offset and welding strip too short, the small defect frame is returned, and the defect length and / or width corresponding to the requirements (such as welding rod too long, welding offset and welding strip too short) are obtained, and then added to the defect dictionary.

[0071] Further preferably, in step S232, a small picture is cut out by the pitch area frame coordinates, i.e. the coordinates obtained by the aforementioned second AI prediction, for the pitch area frame, and then a pitch type result is calculated by a CV image algorithm, wherein the pitch type result includes the results of string pitch, busbar pitch and creeping distance, and the CV image algorithm includes binarization, opening operation, closing operation, contour extraction and the like.

[0072] Therefore, the quality inspection method for the busbar and solder strip of the photovoltaic module provided by the application can accurately control the quality inspection accuracy by adjusting the detection threshold in real time, the system has strong operability, and the miss detection rate of the quality inspection of the busbar and solder strip of the photovoltaic module can reach <= 5%, the false detection rate can reach <= 1%, the production efficiency of the photovoltaic module can be effectively improved, the quality of the photovoltaic module can be ensured, and the labor cost can be greatly reduced.

[0073] In a second aspect, in combination with Figure 2 Based on the same principle, the application further provides a quality inspection system for a busbar and solder strip of a photovoltaic module, which adopts the quality inspection method for the busbar and solder strip of the photovoltaic module as described above, and comprises:

[0074] a picture acquisition module, configured to acquire images at predetermined positions of a busbar and solder strip region of a photovoltaic module, process the acquired images to obtain original pictures required by AI detection software, and save the original pictures to a predetermined address;

[0075] an AI detection module, comprising a model design unit, a data set making and training unit, a model prediction unit and an AI defect information returning unit, wherein the model design unit comprises a large picture cutting small picture model, a small picture defect detection model and a defect picture small target detection model designed respectively, the data set making and training unit is configured to label cell piece labels corresponding to the large picture cutting small picture model to obtain coordinate information data of large picture cutting small picture, label defect types corresponding to the small picture defect detection model to obtain defect type data, and label welding scars corresponding to the defect picture small target detection model to obtain length information data of welding scars on the solder strip, and make these data into training and verification data sets for AI target detection model training, the model prediction unit is configured to predict defects of the busbar and solder strip of the photovoltaic module required to be returned to the system based on the trained data to form a defect dictionary, and the AI defect information returning unit is configured to return the defect dictionary to the system software;

[0076] a threshold filtering module, configured to predict defects of the busbar and solder strip of the photovoltaic module required to be returned to the system based on the trained data to form a defect dictionary;

[0077] The processing detection module is configured to output a detection result based on the filtered defect map saved at the predetermined position, wherein the detection result comprises a result image, a defect type, a defect position corresponding to the defect type, and a final result of whether the product is qualified.

[0078] In a third aspect, based on the same principle, the embodiments of the present application further provide a storage medium for quality inspection of a photovoltaic module busbar and a solder strip, wherein the storage medium has a computer program stored thereon, and the computer program, when executed by a computer or a processor, can implement any of the foregoing quality inspection methods for the photovoltaic module busbar and the solder strip.

[0079] In the embodiments of the intelligent terminal and the computer readable storage medium provided by the present application, all the technical features of any of the foregoing quality inspection method embodiments can be included, and the description and explanation content is basically the same as that of the foregoing method embodiments, which will not be repeated here.

[0080] The embodiments of the present application further provide a computer program product, wherein the computer program product comprises computer program code, and when the computer program code is run on a computer, the computer can execute the method in various possible embodiments as described above.

[0081] The embodiments of the present application further provide a chip comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the device installed with the chip can execute the method in various possible embodiments as described above.

[0082] It can be understood that the above scenarios are only examples and do not constitute a limitation on the application scenarios of the technical solutions provided by the embodiments of the present application. The technical solutions provided by the embodiments of the present application can also be applied to other scenarios. For example, those skilled in the art can know that, as the system architecture evolves and new business scenarios appear, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0083] The above sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0084] The steps in the method embodiments of the present application can be adjusted, combined and deleted in sequence according to actual needs.

[0085] The units in the device embodiments of the present application can be combined, divided and deleted according to actual needs.

[0086] In the present application, for the same or similar term concept, technical scheme and / or application scene description, generally only the first time is described in detail, and the repeated description is not repeated in order to be brief. When understanding the technical scheme of the present application, the same or similar term concept, technical scheme and / or application scene description which is not described in detail can be referred to the related description before.

[0087] In the present application, the description of each embodiment has its own focus, and the part not described or recorded in a certain embodiment can be referred to the related description of other embodiments.

[0088] The technical features of the technical scheme of the present application can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the range recorded in the present application.

[0089] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical scheme of the present application can be embodied in the form of software product, which is stored in the above-mentioned storage medium (such as ROM / RAM, magnetic disc, optical disc), including a plurality of instructions to make a terminal device (which can be a mobile phone, iPad, computer, server, controlled terminal, or network device, etc.) execute the method of each embodiment of the present application.

[0090] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, storage disk, magnetic tape), optical media (such as DVD), or semiconductor media (such as solid state disk (SSD)), etc. Those skilled in the art should understand that the embodiments of the present application described above and shown in the drawings are only examples and do not limit the present application. The advantages of the present application have been fully and effectively realized. The functions and structural principles of the present application have been shown and described in the embodiments, and the embodiments of the present application can be modified or changed in any way without departing from the principles.

Claims

1. A quality inspection method for busbars and solder strips of photovoltaic modules, characterized in that, The steps are as follows: S10, acquire images, acquire images of the photovoltaic module busbar and solder strip area at predetermined locations, process the acquired images to obtain the original image required by the AI ​​detection software, and save the original image to a predetermined address; S20, AI detection, includes: S21, Model Design: Design a large image to small image model, a small image defect detection model, and a defect image small target detection model respectively; S22, Create and train a dataset. Label the battery cells according to the large image to small image model to obtain the coordinate information data of the large image to small image; label the defect types according to the small image defect detection model to obtain the defect type data; label the weld scars according to the defect image small target detection model to obtain the length information data of the weld scars on the weld strip; and create training and validation datasets from these data for training the AI ​​target detection model. S23, Model Prediction: Based on the trained data, the defects of the photovoltaic module busbars and solder strips that need to be returned to the system are predicted, forming a defect dictionary, specifically including: S231, based on the storage address of the original image, obtain the image at the predetermined position of the photovoltaic module busbar and solder strip area, perform scaling processing to obtain a small image, put the small image into the large image to cut small image model for detection, obtain the defect box and the coordinate information of each string of battery cells, and divide the original image into small images corresponding to each string of battery cells based on the coordinate information. S232, respectively call the thread to further scale the small image of each corresponding battery cell, and then put it into the small image defect detection model to detect the small image defect box and the spacing region box including the string spacing, the conductor spacing and the creepage spacing. S233, based on the small image defect box, extract the small image corresponding to each battery cell, then perform scaling processing, and send it to the defect image small target detection model to obtain the defect data information of the small defects. Then, through the millimeter-pixel ratio, the defect data information of the small defects is back-calculated into the real millimeter value to form the defect dictionary. S24, AI returns defect information and the defect dictionary to the system software; S30, threshold filtering, filters AI defect information and saves the filtered defect image to a predetermined location; S40, Process the detection, and output the detection result based on the filtered defect image saved to a predetermined location. The detection result includes the result image, defect type, defect location corresponding to the defect type, and final result of whether it is qualified.

2. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 1, characterized in that, In step S10, the image at the predetermined location is three images of the photovoltaic module busbars and solder strip areas (left, center, and right), and the predetermined address is either local or cloud-based.

3. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 2, characterized in that, In step S21, the large image to small image model, the small image defect detection model, and the defect image small target detection model have the same model structure, all including convolutional layers, upsampling layers and pooling layers, batch normalization layers, activation functions and loss functions. Each convolutional layer consists of several convolutional units, and the parameters of each convolutional unit are optimized through the backpropagation algorithm.

4. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 3, characterized in that, The loss functions include cross-entropy loss function, localization loss function, confidence loss function, and complete intersection-union ratio loss function.

5. The quality inspection method for busbars and solder strips of photovoltaic modules as described in any one of claims 1 to 4, characterized in that, In step S22, the process of creating training and validation datasets from these data includes data collection, data preprocessing, and data training. Data collection includes: prioritizing the collection of real defects in the busbars and solder strips of photovoltaic modules on the production line; and secondly, for defects with fewer than a predetermined number of defects, supplementing the data volume by artificially creating defects or generating defects through data simulation. The data preprocessing includes: first, data cleaning to filter out dirty data that affects the accuracy of model detection; and then data augmentation operations to be performed on the labeled defects. The data training process includes: maintaining the weights in each round of training, evaluating the parameters through loss results and PR curves to obtain the model with the best learning performance, and ensuring that its performance meets the pipeline cycle time requirements.

6. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 5, characterized in that, In step S232, the small image defect box includes a cold solder joint defect box. For cold solder joints, in addition to returning the small image defect box, the process also includes extracting the small image using the coordinates of the cold solder joint defect box, performing a CV operation on the small image to obtain the mask corresponding to the tin-melted part, calculating the area of ​​the white pixels, and adding it to the defect dictionary.

7. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 6, characterized in that, In step S232, the small image defect box also includes defect boxes for excessively long welding rods, weld deviation, and excessively short weld strips. For defects such as excessively long welding rods, weld deviation, and excessively short weld strips, the small image defect box is returned, and the corresponding required defect length and / or width is obtained, and then added to the defect dictionary.

8. The quality inspection method for busbars and solder strips of photovoltaic modules as described in claim 7, characterized in that, In step S232, for the spacing region box, a small image is extracted by the coordinates of the spacing region box, and then the spacing class result is calculated by the CV image algorithm, wherein the spacing class result includes the results of string spacing, conductor distance and creepage distance.

9. A quality inspection system for busbars and solder strips of photovoltaic modules, employing the quality inspection method for busbars and solder strips of photovoltaic modules as described in any one of claims 1 to 8, characterized in that, include: The image acquisition module is used to acquire images of predetermined locations in the busbar and solder strip areas of photovoltaic modules, process the acquired images to obtain the original image required by the AI ​​detection software, and save the original image to a predetermined address. The AI ​​detection module includes a model design unit, a dataset creation and training unit, a model prediction unit, and an AI-returned defect information unit. The threshold filtering module is used to predict the defects of photovoltaic module busbars and solder strips that need to be returned to the system based on the trained data, and form a defect dictionary; The processing and detection module is used to output detection results based on the filtered defect image saved to a predetermined location. The detection results include the result image, the defect type, the defect location corresponding to the defect type, and the final result of whether it is qualified.

Images