Multi-type retired photovoltaic module intelligent identification system and method

Abstract

The invention discloses an intelligent identification system and method for multiple types of retired photovoltaic modules, belongs to the technical field of photovoltaic waste management, and solves the problems that the back plate removal effect is judged only through image identification, the detection object is limited to the local processing quality of the back plate layer of the module, and the overall characteristics of the module cannot be evaluated and identified in the existing method. The method comprises the steps that an intelligent recognition end scans a photovoltaic module in real time, module scanning recognition parameters are obtained, a module recognition model based on deep learning is pre-constructed, the module recognition model recognizes and analyzes the module scanning recognition parameters, posture positioning and correction are conducted on the photovoltaic module, and a module disassembling instruction is triggered based on posture positioning and correction results. According to the method, the component scanning identification parameters including the global information of the photovoltaic component are obtained through scanning, various complex defects and different types of components are automatically learned and identified through the component identification model, and the component scanning identification parameters are identified and analyzed while the detection precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic waste management technology, specifically involving an intelligent identification system and method for multiple types of decommissioned photovoltaic modules. Background Technology

[0002] With the rapid development and large-scale application of the photovoltaic industry, early-used photovoltaic modules are gradually entering their retirement period. It is estimated that a massive number of retired photovoltaic modules will be generated globally over the next few decades. How to efficiently, environmentally, and economically recycle and dispose of these retired modules, achieving resource recycling, has become a critical issue that the industry urgently needs to address, and is of great significance for promoting green and low-carbon development and the circular economy.

[0003] In the recycling and processing of photovoltaic (PV) modules, identification and classification are the crucial first steps. Retired PV modules are diverse, primarily falling into two categories: single-glass modules and double-glass modules. These two types differ significantly in structure (such as the number of glass layers and encapsulation materials), dismantling processes, and subsequent processing routes. Furthermore, during service or after retirement, the glass covers may be intact, cracked, or even broken during transportation and storage, and the frames and junction boxes may also suffer varying degrees of damage or displacement. These differences and conditions directly impact the selection of subsequent dismantling equipment, the setting of dismantling parameters, and the flow of materials.

[0004] Chinese patent CN119919356A discloses a method for recycling photovoltaic modules and an auxiliary visual inspection method. The auxiliary visual inspection method includes: acquiring a first image of the photovoltaic module before the backsheet layer is removed; identifying the position of the busbar within the outline of the photovoltaic module based on the first image, so as to measure the thickness and grind the backsheet layer; acquiring a second image of the photovoltaic module after the backsheet layer is removed; identifying the under-grinding or over-grinding area on the side of the solar cell facing away from the glass layer based on the second image; and judging the removal effect of the backsheet layer based on the under-grinding or over-grinding area. However, the existing method only judges the removal effect of the backsheet layer through image recognition, and its detection object is limited to the local processing quality of the module's backsheet layer. It cannot evaluate and identify the overall characteristics of the module (glass condition, frame integrity, junction box position, model identification, etc.), which leads to the inability to provide comprehensive module information for subsequent disassembly. To address the above problems, we propose an intelligent identification system and method for multiple types of retired photovoltaic modules. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent identification system and method for multiple types of decommissioned photovoltaic modules. This solves the problem that existing methods only rely on image recognition to determine the backsheet removal effect, and their detection objects are limited to the local processing quality of the module's backsheet layer. They cannot evaluate and identify the overall characteristics of the module, thus failing to provide comprehensive module information for subsequent disassembly.

[0006] This invention is implemented as follows: a method for intelligent identification of multiple types of decommissioned photovoltaic modules, the method comprising:

[0007] S10, automatic feeding of photovoltaic modules, the intelligent recognition terminal scans the photovoltaic modules in real time, obtains the module scanning and recognition parameters, and uploads the module scanning and recognition parameters to the database in real time. The intelligent recognition terminal includes a lower recognition module and an upper recognition module.

[0008] S20: Pre-build a deep learning-based component recognition model, retrieve a modeling sample set from the database, use the modeling sample set to iteratively train the component recognition model, and output a converged component recognition model;

[0009] S30, taking the component scanning and recognition parameters as input, the component recognition model identifies and analyzes the component scanning and recognition parameters, and outputs the component recognition and analysis results;

[0010] S40, using the component identification and analysis results as constraints, performs attitude positioning and correction on the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results;

[0011] S50, at least one set of component disassembly modules responds to the component disassembly command, executes the automatic disassembly command for photovoltaic modules, and transports the disassembled photovoltaic modules to different processing stations based on the component identification and analysis results.

[0012] Preferably, the method for iteratively training the component recognition model using a modeling sample set includes:

[0013] Obtain the modeling sample set, perform augmentation processing on the modeling sample set, and divide the augmented modeling sample set into a training set and a test set, with the ratio of the training set to the test set being 3:1;

[0014] Load a pre-built component recognition model, pre-setting the training epochs, hyperparameters, activation function, and joint loss function of the component recognition model;

[0015] The joint loss function of the component identification model is expressed as:

[0016]

[0017]

[0018] in, Denotes the joint loss function. These represent the loss functions for convolutional neural networks, weighted bidirectional feature pyramid networks, type recognition loss functions, and super-resolution loss functions, respectively.

[0019] Load the training set, freeze the autoencoder, pre-train the convolutional neural network and the weighted bidirectional feature pyramid network, freeze the pre-trained convolutional neural network and the weighted bidirectional feature pyramid network, train the type recognition layer and the super-resolution reconstruction layer using an alternating optimization strategy, perform full-model joint fine-tuning on the component recognition model, and output a converged component recognition model.

[0020] Obtain a test set, use the test set as input, execute the component recognition model, output the test results, and determine whether the test results meet the preset test accuracy threshold. If the test results meet the preset test accuracy threshold, output the converged component recognition model. If the test results do not meet the preset test accuracy threshold, use the Adam optimizer to optimize the hyperparameters of the component recognition model.

[0021] Preferably, when pre-constructing the deep learning-based component recognition model, a convolutional neural network (CNN) is used as the basic architecture. The CNN includes a convolutional layer, a pooling layer, an SSNBDL network, and a fully connected layer connected in sequence. The CNN serves as the initial model for the component recognition model. The initial model also includes an input layer and an output layer. An autoencoder is introduced between the input layer and the CNN to encode and label the component scanning and recognition parameters. A weighted bidirectional feature pyramid network is introduced between the CNN and the output layer. The weighted bidirectional feature pyramid network captures the features of the photovoltaic module defect area based on the CBAM attention mechanism, and the features are weighted... A type recognition layer is introduced between the bidirectional feature pyramid network and the output layer. The type recognition layer embeds the ResNet18 and CSRESNEXT50 architectures, which are used for photovoltaic module defect type recognition and photovoltaic module model recognition, respectively. A super-resolution reconstruction layer is introduced between the type recognition layer and the output layer. The super-resolution reconstruction layer virtually models the photovoltaic module based on a deep learning super-resolution algorithm. The loss function of the convolutional neural network is the cross-entropy loss function, the loss function of the weighted bidirectional feature pyramid network is the mean squared error loss function, and the loss function of the type recognition layer is the adversarial loss function.

[0022] Preferably, the component recognition model's method for identifying and analyzing component scanning recognition parameters includes:

[0023] Obtain component scanning and recognition parameters, perform noise reduction preprocessing on the component scanning and recognition parameters, and output a noise reduction parameter set;

[0024] The autoencoder compresses and encodes the noise reduction parameter set to generate a low-dimensional feature vector. A convolutional neural network is then used to extract features from the low-dimensional feature vector to extract local spatial features of the photovoltaic module. These local spatial features include glass morphology features, junction box contour features, frame texture features, and defect contour features.

[0025] A global feature map of a photovoltaic module is generated based on the local spatial features of the module, and the robustness of the local spatial features is enhanced by the SSNBDL network to output the global feature map of the module after scanning noise suppression.

[0026] Load the global feature map of the component, and use a weighted bidirectional feature pyramid network to capture the features of the defect area of ​​the photovoltaic module based on the CBAM attention mechanism. Then, fuse the local spatial features of the photovoltaic module at multiple scales and output the global feature map of the component with the CBAM attention mechanism weighted localization of the defect area.

[0027] Obtain the global feature map of the component in the defect region using the CBAM attention mechanism for weighted localization. Based on the ResNet18 architecture and the CSRESNEXT50 architecture, identify the defect type and component model in the defect region, and output the probability distribution of the defect region and the photovoltaic module model.

[0028] A high-resolution virtual component model is generated based on the probability distribution of defect areas and the photovoltaic module model. The var_threshold function is used to perform secondary detection of minor defects on the high-resolution virtual component model, and the component identification and analysis results are output. The component identification and analysis results include the component comprehensive evaluation value, defect area location, component model, glass condition, three-dimensional dimensions, junction box positioning, and frame positioning results.

[0029] When using the var_threshold function to perform secondary detection of minor defects in a high-resolution virtual component model, the formula for calculating local variance is as follows:

[0030]

[0031] in, For local variance, These are the average pixel value of the window and the average pixel value of the neighboring window, respectively.

[0032] Preferably, the method for noise reduction preprocessing of component scanning and identification parameters includes:

[0033] Load the component scanning and recognition parameters, process outliers and missing values ​​in the component scanning and recognition parameters, and obtain the component scanning and recognition parameters after processing outliers and missing values;

[0034] The component scanning and recognition parameters are processed by bilateral filtering, and the contrast of the filtered component scanning and recognition parameters is increased. The contrast-increased component scanning and recognition parameters are then filtered by a low-pass filter, and the filtered component scanning and recognition parameters are output.

[0035] When performing bilateral filtering on the component scanning and recognition parameters, the bilateral filtering output is expressed as follows:

[0036]

[0037]

[0038] in, This indicates the output result of the bilateral filter. For parameter normalization factor, For pixel coordinates and neighborhood coordinates, These are neighborhood parameters and component scanning and recognition parameters, respectively. These are the parameter space domain weights and the pixel value domain weights, respectively. This is a parameter normalization constant.

[0039] The component scanning and recognition parameters after filtering are obtained, the image grayscale threshold is determined based on the image grayscale histogram, the pixel grayscale values ​​in the component scanning and recognition parameters are extracted, and the pixel grayscale values ​​in the component scanning and recognition parameters are compared with the image grayscale threshold to generate component scanning and recognition parameters with the image grayscale threshold.

[0040] Principal component analysis is used to reduce the dimensionality of the component scanning and identification parameters, and the dimensionality-reduced component scanning and identification parameters are output.

[0041] Preferably, the method for attitude positioning and correction of photovoltaic modules includes:

[0042] Obtain the component identification and analysis results, identify the defect area location, component model, junction box location, and frame location results in the component identification and analysis results, and use the second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model. When extracting the pixel-level edges of the high-resolution virtual component model, the second-order differential method is used to calculate the magnitude and direction of the gradient, and non-maximum suppression is applied to the gradient intensity. The edge with the optimal distance is selected as the pixel-level edge from all collinear edge contours merged in the high-resolution virtual component model.

[0043] Load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculations on the pixel-level edges, and use graph-based weights to filter out interference points in the high-resolution virtual component model to extract the virtual central axis of the high-resolution virtual component model.

[0044] Extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and transform the perpendicular coordinates of the pixel-level edges and virtual central axis into photovoltaic module pose information through affine transformation.

[0045] The photovoltaic module pose information is obtained, and the photovoltaic module pose information is corrected based on the gray-scale equalization algorithm. The corrected photovoltaic module pose information is then output.

[0046] On the other hand, the present invention also provides a multi-type decommissioned photovoltaic module intelligent identification system, the multi-type decommissioned photovoltaic module intelligent identification system comprising:

[0047] The intelligent identification terminal automatically feeds photovoltaic modules, scans the photovoltaic modules in real time, obtains the module scanning and identification parameters, and uploads the module scanning and identification parameters to the database in real time.

[0048] The identification and analysis module takes the component scanning and identification parameters as input, and the component identification model identifies and analyzes the component scanning and identification parameters, and outputs the component identification and analysis results.

[0049] The attitude positioning module uses the component identification and analysis results as constraints to perform attitude positioning and correction of the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results;

[0050] The module disassembly module, at least one set of module disassembly modules, responds to the module disassembly command, executes the automatic disassembly command of photovoltaic module, and transports the disassembled photovoltaic module to different processing stations based on the module identification and analysis results.

[0051] Preferably, the intelligent recognition terminal includes a lower recognition module and an upper recognition module. The lower recognition module is used to scan the component size, identify single-glass and double-glass components, and identify the glass shape. The upper recognition module is used to scan the component size, identify single-glass and double-glass components, and locate the junction box and frame disassembly position.

[0052] Preferably, the attitude positioning module includes:

[0053] The edge determination unit is used to obtain the component identification and analysis results, identify the location of defect areas, component models, junction box positioning and border positioning results in the component identification and analysis results, and use a second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model;

[0054] The centerline determination unit is used to load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculations on the pixel-level edges, and use graph-based weights to filter out interference points of the high-resolution virtual component model to extract the virtual centerline of the high-resolution virtual component model.

[0055] The pose extraction unit is used to extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and transform the perpendicular coordinates of the pixel-level edges and virtual central axis into photovoltaic module pose information through affine transformation.

[0056] The pose correction unit is used to acquire the pose information of the photovoltaic module, correct the pose information of the photovoltaic module based on the gray-scale equalization algorithm, and output the corrected pose information of the photovoltaic module.

[0057] Compared with the prior art, the embodiments of this application have the following main advantages:

[0058] In this embodiment of the invention, component scanning and identification parameters containing global information of photovoltaic modules are obtained through scanning. A component identification model automatically learns and identifies various complex defects and different types of components, improving detection accuracy while simultaneously analyzing the component scanning and identification parameters. This enables a comprehensive analysis of the overall state of the component, providing complete and comprehensive component information for disassembly.

[0059] This invention provides a component recognition model and its training method. During model training, an alternating optimization strategy is employed to train the type recognition layer and the super-resolution reconstruction layer. This balances the model across different task types, preventing any single task from dominating the training process. This allows the model to achieve good performance across multiple tasks, including defect type recognition, component model recognition, and super-resolution reconstruction. The component recognition model is based on a convolutional neural network (CNN) architecture. CNNs automatically learn local spatial features of component images, such as glass morphology, junction box contours, border textures, and defect contours, providing rich feature information for subsequent recognition and analysis. Simultaneously, the robustness of local spatial features is enhanced through an SSNBDL network, ensuring stability even under noise and illumination variations, thus improving the model's anti-interference capability. The type recognition layer embeds ResNet18 and CSResNeXt50 architectures for defect type recognition and component model recognition, respectively, enabling joint learning across multiple tasks and improving the model's overall performance. The super-resolution reconstruction layer generates high-resolution virtual component models, providing a clearer image foundation for subsequent minor defect detection and evaluation.

[0060] This invention provides a method for noise reduction preprocessing of component scanning and recognition parameters. Noise reduction preprocessing of component scanning and recognition parameters can effectively improve the quality and reliability of component scanning and recognition parameters. During the preprocessing, from handling outliers and missing values ​​to bilateral filtering, contrast enhancement, threshold segmentation and principal component analysis, each step provides a better data foundation for subsequent feature extraction and recognition. Moreover, the data after dimensionality reduction is more compact, reducing the amount of computation, improving the efficiency of subsequent processing steps, and saving computing resources. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the implementation process of the intelligent identification method for multiple types of decommissioned photovoltaic modules provided by the present invention.

[0062] Figure 2 The diagram illustrates the process of iteratively training a component identification model using a modeling sample set.

[0063] Figure 3The diagram illustrates the implementation process of the component identification model for identifying and analyzing component scanning parameters.

[0064] Figure 4 A schematic diagram of the implementation process of the method for noise reduction preprocessing of component scanning and identification parameters is shown.

[0065] Figure 5 A schematic diagram of the process for attitude positioning and correction of photovoltaic modules is shown.

[0066] Figure 6 A schematic diagram of the structure of a smart identification system for multiple types of decommissioned photovoltaic modules is shown.

[0067] Figure 7 A schematic diagram of the installation of the intelligent identification terminal in an embodiment of the present invention is shown. Detailed Implementation

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0069] Existing methods rely solely on image recognition to assess backsheet removal effectiveness, limiting detection to localized processing quality of the backsheet layer and failing to evaluate the overall characteristics of the module. This results in a lack of comprehensive module information for subsequent disassembly. To address this issue, we propose an intelligent identification system and method for multiple types of decommissioned photovoltaic (PV) modules. The method involves an intelligent identification terminal 100 that scans the PV module in real-time, acquiring scanning and identification parameters, and uploading these parameters to a database. A deep learning-based module identification model is pre-built, analyzing the scanning and identification parameters, outputting the results, performing attitude positioning and correction on the PV module, and triggering a disassembly command based on the results. The disassembled PV module is then transported to different processing stations based on the identification and analysis results. In this embodiment, scanning acquires scanning and identification parameters containing global information about the PV module. The module identification model automatically learns and identifies various complex defects and different module types, improving detection accuracy while simultaneously analyzing the scanning and identification parameters. This allows for a comprehensive analysis of the overall module status, providing complete and comprehensive module information for disassembly.

[0070] This invention provides a method for intelligent identification of multiple types of decommissioned photovoltaic modules. Figure 1 This diagram illustrates the implementation flow of a method for intelligent identification of multiple types of decommissioned photovoltaic modules. Specifically, this method includes:

[0071] S10, automatic feeding of photovoltaic modules, intelligent identification terminal 100 scans photovoltaic modules in real time, obtains module scanning identification parameters, and uploads the module scanning identification parameters to the database in real time. The intelligent identification terminal 100 includes a lower identification module and an upper identification module.

[0072] It should be noted that the component scanning parameters include, but are not limited to, three-dimensional dimension data (length, width, height), component type identification, glass shape data, junction box positioning data, frame positioning data, and defect contour data.

[0073] S20: Pre-build a deep learning-based component recognition model, retrieve a modeling sample set from the database, use the modeling sample set to iteratively train the component recognition model, and output a converged component recognition model;

[0074] S30, taking the component scanning and recognition parameters as input, the component recognition model identifies and analyzes the component scanning and recognition parameters, and outputs the component recognition and analysis results;

[0075] S40, using the component identification and analysis results as constraints, performs attitude positioning and correction on the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results;

[0076] S50, at least one set of component disassembly modules responds to the component disassembly command, executes the automatic disassembly command for photovoltaic modules, and transports the disassembled photovoltaic modules to different processing stations based on the component identification and analysis results.

[0077] In this embodiment of the invention, component scanning and identification parameters containing global information of photovoltaic modules are obtained through scanning. A component identification model automatically learns and identifies various complex defects and different types of components, improving detection accuracy while simultaneously analyzing the component scanning and identification parameters. This enables a comprehensive analysis of the overall state of the component, providing complete and comprehensive component information for disassembly.

[0078] This invention provides a method for iteratively training a component identification model using a modeling sample set. Figure 2 This diagram illustrates the implementation process of a method for iteratively training a component recognition model using a modeling sample set. The method specifically includes:

[0079] S101, Obtain the modeling sample set, perform enhancement processing on the modeling sample set, including but not limited to rotation, brightness adjustment, and noise adjustment, and divide the enhanced modeling sample set into a training set and a test set, wherein the ratio of the training set to the test set is 3:1;

[0080] S102, Load the pre-built component recognition model, and preset the training rounds, hyperparameters, activation function and joint loss function of the component recognition model;

[0081] The joint loss function of the component identification model is expressed as:

[0082]

[0083]

[0084] in, Denotes the joint loss function. These represent the loss functions for convolutional neural networks, weighted bidirectional feature pyramid networks, type recognition, and super-resolution, respectively. These are the loss weight coefficients for the convolutional neural network, the weighted bidirectional feature pyramid network, the type recognition loss, and the super-resolution loss, respectively. In this embodiment, the loss weight coefficients for the convolutional neural network, the weighted bidirectional feature pyramid network, the type recognition loss, and the super-resolution loss can be 0.2, 0.3, 0.2, and 0.3, respectively. In this embodiment, by presetting the training epochs, hyperparameters, activation functions, and joint loss functions, clear guidance and optimization directions are provided for model training. Reasonable hyperparameter settings can accelerate the model's convergence speed and improve training efficiency; the selection of activation functions can enhance the model's nonlinear fitting ability; the joint loss function comprehensively considers the losses of multiple tasks, enabling the model to optimize multiple objectives simultaneously and improve the model's overall performance.

[0085] S103: Load the training set, freeze the autoencoder, pre-train the convolutional neural network and the weighted bidirectional feature pyramid network, freeze the pre-trained convolutional neural network and the weighted bidirectional feature pyramid network, train the type recognition layer and the super-resolution reconstruction layer using an alternating optimization strategy, perform full-model joint fine-tuning on the component recognition model, and output a converged component recognition model. In this embodiment, a staged training strategy is adopted. First, the convolutional neural network and the weighted bidirectional feature pyramid network are pre-trained, which can quickly extract low-level features and defect region features of components, providing more accurate feature representations for subsequent type recognition and super-resolution reconstruction. Freezing the pre-trained network parts and then training the type recognition layer and the super-resolution reconstruction layer can avoid mutual interference between networks and improve the stability and efficiency of training.

[0086] S104, Obtain the test set, use the test set as input, execute the component recognition model, and output the test results;

[0087] S105, determine whether the test result meets the preset test accuracy threshold. It should be noted that the test accuracy threshold can be 0.85-0.91.

[0088] S106, If the test results meet the preset test accuracy threshold, output the converged component recognition model.

[0089] If the test results do not meet the preset test accuracy threshold, the Adam optimizer is used to optimize the hyperparameters of the component recognition model, and the process returns to S103 to continue iterative training of the model.

[0090] In this embodiment, the pre-constructed deep learning-based component recognition model uses a convolutional neural network (CNN) as its basic architecture. The CNN comprises sequentially connected convolutional layers, pooling layers, an SSNBDL network, and a fully connected layer. The CNN serves as the initial model for component recognition, further including an input layer and an output layer. An autoencoder is introduced between the input layer and the CNN to encode and label the component scanning and recognition parameters. A weighted bidirectional feature pyramid network is introduced between the CNN and the output layer. This network captures the defect region features of the photovoltaic module based on the CBAM attention mechanism. A type recognition layer is introduced between the weighted bidirectional feature pyramid network and the output layer. The type recognition layer embeds the ResNet18 and CSRESNEXT50 architectures, which are used for photovoltaic module defect type recognition and photovoltaic module model recognition, respectively. A super-resolution reconstruction layer is introduced between the type recognition layer and the output layer. The super-resolution reconstruction layer virtually models the photovoltaic module based on a deep learning super-resolution algorithm. The loss function of the convolutional neural network is the cross-entropy loss function, the loss function of the weighted bidirectional feature pyramid network is the mean squared error loss function, and the loss function of the type recognition layer is the adversarial loss function.

[0091] This invention provides a component recognition model and its training method. During model training, an alternating optimization strategy is employed to train the type recognition layer and the super-resolution reconstruction layer. This balances the model across different task types, preventing any single task from dominating the training process. This allows the model to achieve good performance across multiple tasks, including defect type recognition, component model recognition, and super-resolution reconstruction. The component recognition model is based on a convolutional neural network (CNN) architecture. CNNs automatically learn local spatial features of component images, such as glass morphology, junction box contours, border textures, and defect contours, providing rich feature information for subsequent recognition and analysis. Simultaneously, the robustness of local spatial features is enhanced through an SSNBDL network, ensuring stability even under noise and illumination variations, thus improving the model's anti-interference capability. The type recognition layer embeds ResNet18 and CSResNeXt50 architectures for defect type recognition and component model recognition, respectively, enabling joint learning across multiple tasks and improving the model's overall performance. The super-resolution reconstruction layer generates high-resolution virtual component models, providing a clearer image foundation for subsequent minor defect detection and evaluation.

[0092] This invention provides a method for identifying and analyzing component scanning and identification parameters using a component identification model. Figure 3 This diagram illustrates the implementation flow of the component identification model's method for identifying and analyzing component scanning parameters. Specifically, the method includes:

[0093] S201: Obtain component scanning and recognition parameters, perform noise reduction preprocessing on the component scanning and recognition parameters, and output a noise reduction parameter set. It should be noted that the component scanning process may be affected by environmental noise, equipment errors, and other factors, resulting in noise in the scan data. Noise reduction preprocessing can effectively remove this noise and improve the quality and reliability of the data.

[0094] In step S202, the autoencoder compresses and encodes the noise reduction parameter set to generate a low-dimensional feature vector. A convolutional neural network is then used to extract features from this low-dimensional feature vector, extracting local spatial features of the photovoltaic module. These local spatial features include glass morphology features, junction box contour features, frame texture features, and defect contour features. In this embodiment, the autoencoder can compress high-dimensional scanning parameters into low-dimensional feature vectors, effectively removing redundant information, retaining key features, and improving data compactness and processability. The convolutional neural network can automatically learn and extract local spatial features of the photovoltaic module, such as glass morphology, junction box contour, frame texture, and defect contours. These features form the basis for subsequent identification and analysis, providing rich information for the model.

[0095] S203, Generate a global feature map of the photovoltaic module based on the local spatial features of the photovoltaic module, and enhance the robustness of the local spatial features through the SSNBDL network, and output the global feature map of the module after scanning noise suppression. In this embodiment, the SSNBDL network can further enhance the robustness of the local spatial features, so that the features remain stable when facing interference such as noise and light changes, and improve the anti-interference ability of the model.

[0096] S204, Load the global feature map of the component, the weighted bidirectional feature pyramid network captures the features of the defect area of ​​the photovoltaic module based on the CBAM attention mechanism, and fuses the local spatial features of the photovoltaic module at multiple scales, and outputs the global feature map of the component with the CBAM attention mechanism weighted localization of the defect area;

[0097] S205: Obtain the global feature map of the component in the defect area using the CBAM attention mechanism for weighted localization. Based on the ResNet18 architecture and the CSResNext50 architecture, identify the defect type and component model in the defect area, and output the probability distribution of the defect area and the photovoltaic module model.

[0098] S206. A high-resolution virtual component model is generated based on the probability distribution of defect areas and the photovoltaic module model. The var_threshold function is used to perform secondary detection of minor defects on the high-resolution virtual component model and output the component identification and analysis results. The component identification and analysis results include the component comprehensive evaluation value, defect area location, component model, glass condition, three-dimensional dimensions, junction box positioning and frame positioning results.

[0099] In this embodiment of the invention, the use of the var_threshold function for secondary detection of minor defects can further improve the accuracy of defect detection and uncover some minor defects that may be overlooked in the initial detection. The output component identification and analysis results include the component's comprehensive evaluation value, defect area location, component model, glass condition, three-dimensional dimensions, junction box positioning, and frame positioning results. These results can comprehensively reflect the component's condition and characteristics, providing comprehensive information support for subsequent processing and decision-making.

[0100] When using the var_threshold function to perform secondary detection of minor defects in a high-resolution virtual component model, the formula for calculating local variance is as follows:

[0101]

[0102] in, For local variance, These are the average pixel value of the window and the average pixel value of the neighboring window, respectively.

[0103] This invention provides a method for noise reduction preprocessing of component scanning and identification parameters. Figure 4 This diagram illustrates the implementation flow of a method for noise reduction preprocessing of component scanning and identification parameters. The method specifically includes:

[0104] S2011, Load component scanning and recognition parameters, process outliers and missing values ​​of component scanning and recognition parameters, and obtain component scanning and recognition parameters after outlier and missing value processing;

[0105] S2012, perform bilateral filtering on the component scanning recognition parameters, and increase the contrast of the component scanning recognition parameters after bilateral filtering. Then, filter the component scanning recognition parameters after increasing the contrast through a low-pass filter and output the filtered component scanning recognition parameters.

[0106] When performing bilateral filtering on the component scanning and recognition parameters, the bilateral filtering output is expressed as follows:

[0107]

[0108]

[0109] in, This indicates the output result of the bilateral filter. For parameter normalization factor, For pixel coordinates and neighborhood coordinates, These are neighborhood parameters and component scanning and recognition parameters, respectively. These are the parameter space domain weights and the pixel value domain weights, respectively. This is a parameter normalization constant.

[0110] S2013, obtain the filtered component scanning recognition parameters, determine the image grayscale threshold based on the image grayscale histogram, extract the pixel grayscale values ​​in the component scanning recognition parameters, compare the pixel grayscale values ​​in the component scanning recognition parameters with the image grayscale threshold, and generate component scanning recognition parameters with the image grayscale threshold;

[0111] S2014 uses principal component analysis to reduce the dimensionality of component scanning and identification parameters, and outputs the dimensionality-reduced component scanning and identification parameters.

[0112] This invention provides a method for noise reduction preprocessing of component scanning and recognition parameters. Noise reduction preprocessing of component scanning and recognition parameters can effectively improve the quality and reliability of component scanning and recognition parameters. During the preprocessing, from handling outliers and missing values ​​to bilateral filtering, contrast enhancement, threshold segmentation and principal component analysis, each step provides a better data foundation for subsequent feature extraction and recognition. Moreover, the data after dimensionality reduction is more compact, reducing the amount of computation, improving the efficiency of subsequent processing steps, and saving computing resources.

[0113] This invention provides a method for attitude positioning and correction of photovoltaic modules. Figure 5 This diagram illustrates the implementation flow of a method for attitude positioning and correction of photovoltaic modules. The method specifically includes:

[0114] S301, Obtain the component identification and analysis results, identify the defect area location, component model, junction box location and frame location results in the component identification and analysis results, and use the second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model. When extracting the pixel-level edges of the high-resolution virtual component model, the second-order differential method is used to calculate the magnitude and direction of the gradient, and the gradient intensity is suppressed by non-maximum. Among all collinear edge contours of the high-resolution virtual component model, the edge with the optimal distance is selected as the pixel-level edge.

[0115] S302, load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculation on the pixel-level edges, and use graph-based weights to filter out interference points of the high-resolution virtual component model and extract the virtual central axis of the high-resolution virtual component model.

[0116] S303, extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and transform the perpendicular coordinates of the pixel-level edges and virtual central axis into photovoltaic module pose information through affine transformation.

[0117] S304: Obtain the pose information of the photovoltaic module, correct the pose information of the photovoltaic module based on the gray-scale equalization algorithm, and output the corrected pose information of the photovoltaic module.

[0118] In this embodiment of the invention, a method for attitude positioning and correction of photovoltaic modules is provided. When positioning and correcting the photovoltaic modules, the second-order differential algorithm can accurately calculate the magnitude and direction of the gradient. Through non-maximum suppression, redundant information in edge detection can be effectively removed, and clear and accurate pixel-level edges can be extracted. Through edge fitting calculation, discrete pixel-level edges can be smoothed to generate continuous edge curves, which facilitates subsequent geometric analysis and calculation. Finally, the perpendicular coordinates are converted into pose information through affine transformation, which can convert the two-dimensional information in the image into actual three-dimensional pose information, providing accurate guidance for subsequent mechanical operation and processing.

[0119] On the other hand, embodiments of the present invention also provide an intelligent identification system for multiple types of decommissioned photovoltaic modules. Figure 6 A schematic diagram of a multi-type decommissioned photovoltaic module intelligent identification system is shown. The system specifically includes:

[0120] The intelligent identification terminal 100 automatically feeds photovoltaic modules and scans the photovoltaic modules in real time to obtain module scanning and identification parameters, and uploads the module scanning and identification parameters to the database in real time.

[0121] It should be noted that, as Figure 7 The diagram shows the installation schematic of the intelligent identification terminal 100 in an embodiment of the present invention. The intelligent identification terminal 100 includes a lower identification module and an upper identification module. The lower identification module is used to scan the component size, identify single-glass and double-glass components, and identify the glass shape. The upper identification module is used to scan the component size, identify single-glass and double-glass components, and locate the junction box and frame disassembly position.

[0122] The identification and analysis module 200 takes the component scanning and identification parameters as input, the component identification model identifies and analyzes the component scanning and identification parameters, and outputs the component identification and analysis results.

[0123] The attitude positioning module 300 uses the component identification and analysis results as constraints to perform attitude positioning and correction on the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results;

[0124] The component disassembly module 400, at least one set of component disassembly modules 400, responds to the component disassembly command, executes the photovoltaic module automatic disassembly command, and delivers the disassembled photovoltaic module to different processing stations based on the component identification and analysis results.

[0125] In this embodiment, the attitude positioning module 300 includes:

[0126] The edge determination unit 310 is used to obtain the component identification and analysis results, identify the location of the defect area, component model, junction box positioning and border positioning results in the component identification and analysis results, and use a second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model;

[0127] The centerline determination unit 320 is used to load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculations on the pixel-level edges, and use graph-based weights to filter out interference points of the high-resolution virtual component model to extract the virtual centerline of the high-resolution virtual component model.

[0128] The pose extraction unit 330 is used to extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and affine transform the perpendicular coordinates of the pixel-level edges and virtual central axis into the pose information of the photovoltaic component.

[0129] The pose correction unit 340 is used to acquire the pose information of the photovoltaic module, correct the pose information of the photovoltaic module based on the gray-scale equalization algorithm, and output the corrected pose information of the photovoltaic module.

[0130] It should be noted that the steps of the intelligent identification system for multiple types of decommissioned photovoltaic modules correspond to those of the intelligent identification method for multiple types of decommissioned photovoltaic modules described above, and will not be repeated here.

[0131] In summary, this invention provides an intelligent identification system and method for multiple types of decommissioned photovoltaic modules. In the embodiments of this invention, module scanning and identification parameters containing global information of the photovoltaic module are obtained through scanning. The module identification model automatically learns and identifies various complex defects and different types of modules, improving detection accuracy while simultaneously analyzing the module scanning and identification parameters. This enables a comprehensive analysis of the overall condition of the module, providing complete and comprehensive module information for disassembly.

[0132] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0133] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of multiple types of decommissioned photovoltaic modules, characterized in that, The method includes: S10, automatic feeding of photovoltaic modules, the intelligent recognition terminal scans the photovoltaic modules in real time, obtains the module scanning and recognition parameters, and uploads the module scanning and recognition parameters to the database in real time. The intelligent recognition terminal includes a lower recognition module and an upper recognition module. S20: Pre-build a deep learning-based component recognition model, retrieve a modeling sample set from the database, use the modeling sample set to iteratively train the component recognition model, and output a converged component recognition model; S30, taking the component scanning and recognition parameters as input, the component recognition model identifies and analyzes the component scanning and recognition parameters, and outputs the component recognition and analysis results; S40, using the component identification and analysis results as constraints, performs attitude positioning and correction on the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results; S50, at least one set of component disassembly modules responds to the component disassembly command, executes the automatic disassembly command for photovoltaic modules, and transports the disassembled photovoltaic modules to different processing stations based on the component identification and analysis results.

2. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 1, characterized in that: The method of iteratively training the component recognition model using a modeling sample set includes: Obtain the modeling sample set, perform augmentation processing on the modeling sample set, and divide the augmented modeling sample set into a training set and a test set, with the ratio of the training set to the test set being 3:1; Load a pre-built component recognition model, pre-setting the training epochs, hyperparameters, activation function, and joint loss function of the component recognition model; Load the training set, freeze the autoencoder, pre-train the convolutional neural network and the weighted bidirectional feature pyramid network, freeze the pre-trained convolutional neural network and the weighted bidirectional feature pyramid network, train the type recognition layer and the super-resolution reconstruction layer using an alternating optimization strategy, perform full-model joint fine-tuning on the component recognition model, and output a converged component recognition model. Obtain a test set, use the test set as input, execute the component recognition model, output the test results, and determine whether the test results meet the preset test accuracy threshold. If the test results meet the preset test accuracy threshold, output the converged component recognition model. If the test results do not meet the preset test accuracy threshold, use the Adam optimizer to optimize the hyperparameters of the component recognition model.

3. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 2, characterized in that: The pre-constructed deep learning-based component recognition model uses a convolutional neural network (CNN) as its basic architecture. The CNN comprises sequentially connected convolutional layers, pooling layers, an SSNBDL network, and a fully connected layer. The CNN serves as the initial model for component recognition, including an input layer and an output layer. An autoencoder is introduced between the input layer and the CNN to encode and label component scanning and recognition parameters. A weighted bidirectional feature pyramid network (BNPB) is introduced between the CNN and the output layer. The BNPB captures photovoltaic module defect region features based on the CBAM attention mechanism. A type recognition layer is introduced between the BNPB and the output layer, embedding a ResNet18 and CSRESNEXT50 architecture. The ResNet18 and CSRESNEXT50 architectures are used for photovoltaic module defect type recognition and photovoltaic module model recognition, respectively. A super-resolution reconstruction layer is introduced between the type recognition layer and the output layer. The super-resolution reconstruction layer virtually models the photovoltaic module based on a deep learning super-resolution algorithm. The loss function of the CNN is the cross-entropy loss function, the loss function of the BNPB is the mean squared error loss function, and the loss function of the type recognition layer is the adversarial loss function.

4. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 3, characterized in that: The component recognition model includes a method for identifying and analyzing component scanning and recognition parameters, comprising: Obtain component scanning and recognition parameters, perform noise reduction preprocessing on the component scanning and recognition parameters, and output a noise reduction parameter set; The autoencoder compresses and encodes the noise reduction parameter set to generate a low-dimensional feature vector. A convolutional neural network is then used to extract features from the low-dimensional feature vector to extract local spatial features of the photovoltaic module. These local spatial features include glass morphology features, junction box contour features, frame texture features, and defect contour features. A global feature map of a photovoltaic module is generated based on the local spatial features of the module, and the robustness of the local spatial features is enhanced by the SSNBDL network to output the global feature map of the module after scanning noise suppression. Load the global feature map of the component, and use a weighted bidirectional feature pyramid network to capture the features of the defect area of ​​the photovoltaic module based on the CBAM attention mechanism. Then, fuse the local spatial features of the photovoltaic module at multiple scales and output the global feature map of the component with the CBAM attention mechanism weighted localization of the defect area. Obtain the global feature map of the component in the defect region using the CBAM attention mechanism for weighted localization. Based on the ResNet18 architecture and the CSRESNEXT50 architecture, identify the defect type and component model in the defect region, and output the probability distribution of the defect region and the photovoltaic module model. A high-resolution virtual component model is generated based on the probability distribution of defect areas and the photovoltaic module model. The var_threshold function is used to perform secondary detection of minor defects on the high-resolution virtual component model, and the component identification and analysis results are output. The component identification and analysis results include the component comprehensive evaluation value, defect area location, component model, glass condition, three-dimensional dimensions, junction box positioning, and frame positioning results.

5. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 4, characterized in that: The method for noise reduction preprocessing of component scanning and identification parameters includes: Load the component scanning and recognition parameters, process outliers and missing values ​​in the component scanning and recognition parameters, and obtain the component scanning and recognition parameters after processing outliers and missing values; The component scanning and recognition parameters are processed by bilateral filtering, and the contrast of the bilaterally filtered component scanning and recognition parameters is increased. The contrast-increased component scanning and recognition parameters are then filtered by a low-pass filter, and the filtered component scanning and recognition parameters are output.

6. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 5, characterized in that: The method for noise reduction preprocessing of component scanning and identification parameters further includes: The component scanning and recognition parameters after filtering are obtained, the image grayscale threshold is determined based on the image grayscale histogram, the pixel grayscale values ​​in the component scanning and recognition parameters are extracted, and the pixel grayscale values ​​in the component scanning and recognition parameters are compared with the image grayscale threshold to generate component scanning and recognition parameters with the image grayscale threshold. Principal component analysis is used to reduce the dimensionality of the component scanning and identification parameters, and the dimensionality-reduced component scanning and identification parameters are output.

7. The intelligent identification method for multiple types of decommissioned photovoltaic modules as described in claim 6, characterized in that: The method for attitude positioning and correction of photovoltaic modules includes: Obtain the component identification and analysis results, identify the defect area location, component model, junction box location, and frame location results in the component identification and analysis results, and use the second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model. When extracting the pixel-level edges of the high-resolution virtual component model, the second-order differential method is used to calculate the magnitude and direction of the gradient, and non-maximum suppression is applied to the gradient intensity. The edge with the optimal distance is selected as the pixel-level edge from all collinear edge contours merged in the high-resolution virtual component model. Load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculations on the pixel-level edges, and use graph-based weights to filter out interference points in the high-resolution virtual component model to extract the virtual central axis of the high-resolution virtual component model. Extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and transform the perpendicular coordinates of the pixel-level edges and virtual central axis into photovoltaic module pose information through affine transformation. The photovoltaic module pose information is obtained, and the photovoltaic module pose information is corrected based on the gray-scale equalization algorithm. The corrected photovoltaic module pose information is then output.

8. A multi-type decommissioned photovoltaic module intelligent identification system, used to implement the multi-type decommissioned photovoltaic module intelligent identification method as described in any one of claims 1-7, characterized in that: The intelligent identification system for multiple types of decommissioned photovoltaic modules includes: The intelligent identification terminal automatically feeds photovoltaic modules, scans the photovoltaic modules in real time, obtains the module scanning and identification parameters, and uploads the module scanning and identification parameters to the database in real time. The identification and analysis module takes the component scanning and identification parameters as input, and the component identification model identifies and analyzes the component scanning and identification parameters, and outputs the component identification and analysis results. The attitude positioning module uses the component identification and analysis results as constraints to perform attitude positioning and correction of the photovoltaic module, and triggers the component disassembly command based on the attitude positioning and correction results; The module disassembly module, at least one set of module disassembly modules, responds to the module disassembly command, executes the automatic disassembly command of photovoltaic module, and transports the disassembled photovoltaic module to different processing stations based on the module identification and analysis results.

9. The intelligent identification system for multiple types of decommissioned photovoltaic modules as described in claim 8, characterized in that: The intelligent recognition terminal includes a lower recognition module and an upper recognition module. The lower recognition module is used to scan the component size, identify single-glass and double-glass components, and identify the glass shape. The upper recognition module is used to scan the component size, identify single-glass and double-glass components, and locate the junction box and frame disassembly position.

10. The intelligent identification system for multiple types of decommissioned photovoltaic modules as described in claim 9, characterized in that: The attitude localization module includes: The edge determination unit is used to obtain the component identification and analysis results, identify the location of defect areas, component models, junction box positioning and border positioning results in the component identification and analysis results, and use a second-order differential algorithm to extract the pixel-level edges of the high-resolution virtual component model; The centerline determination unit is used to load the pixel-level edges of the high-resolution virtual component model, perform edge fitting calculations on the pixel-level edges, and use graph-based weights to filter out interference points of the high-resolution virtual component model to extract the virtual centerline of the high-resolution virtual component model. The pose extraction unit is used to extract the pixel-level edges and virtual central axis of the high-resolution virtual component model, determine the perpendicular coordinates of the pixel-level edges and virtual central axis, and transform the perpendicular coordinates of the pixel-level edges and virtual central axis into photovoltaic module pose information through affine transformation. The pose correction unit is used to acquire the pose information of the photovoltaic module, correct the pose information of the photovoltaic module based on the gray-scale equalization algorithm, and output the corrected pose information of the photovoltaic module.

Images