Photovoltaic module recovery value prediction method and device based on multi-dimensional feature fusion

Abstract

The invention relates to a photovoltaic module recovery value prediction method and device based on multi-dimensional feature fusion, belongs to the technical field of photovoltaic modules, and solves the problems that an existing recovery value evaluation method depends on static data and cannot dynamically consider uncertainty, so that an evaluation result is lack of perspectiveness and accuracy. The method comprises the following steps: acquiring an image of a photovoltaic module by using an infrared thermal imager and a visible light camera, and acquiring electrical performance data of the photovoltaic module; performing preliminary denoising on the photovoltaic module image through wavelet transform, and performing refined denoising on the image after preliminary denoising through a generative adversarial network; based on a multi-modal convolutional neural network model, fusing the features of the infrared thermal imaging image and the visible light image after noise reduction processing, and outputting a defect type and a severity score; dynamically adjusting a recovery value index from the aspects of internal characteristics and external uncertainty of the photovoltaic module; and constructing a multi-objective function and corresponding constraint conditions. And the recovery value of the photovoltaic module can be evaluated more accurately.

Description

Technical Field

[0001] This application relates to the field of photovoltaic module technology, and in particular to a method and apparatus for predicting the recycling value of photovoltaic modules based on multi-dimensional feature fusion. Background Technology

[0002] With the transformation of the global energy structure, photovoltaic (PV) power generation, as a clean and renewable energy source, has been widely applied globally. PV modules are the core equipment of PV power generation systems. As their service life increases, the performance of PV modules will decline to some extent, especially when exposed to the natural environment for extended periods, potentially affected by factors such as temperature, humidity, and ultraviolet radiation. Problems such as material aging, electrical performance degradation, surface contamination, and microcracks within the modules all contribute to a reduction in the overall power generation efficiency of the PV modules. Assessing the recycling value of PV modules after retirement has become a key issue for the sustainable development of the PV industry. Efficient and accurate recycling value assessment can not only promote resource reuse but also reduce environmental pollution. Therefore, how to accurately assess the recycling value of retired PV modules has become an important problem that urgently needs to be solved.

[0003] Currently, the assessment of the recycling value of photovoltaic (PV) modules mainly relies on traditional methods such as electrical performance testing, visual inspection, and environmental monitoring. Traditional methods assess power output degradation by conducting electrical performance tests on PV modules, such as IV curve testing. Furthermore, visible defects such as surface cracks, contamination, and burns are typically detected through manual visual inspection or simple image processing techniques. However, these methods have significant limitations. First, manual inspection is inefficient and prone to significant subjective errors, especially when dealing with complex cracks or minor contamination, easily missing some defects. Second, while electrical performance testing provides some information, it cannot comprehensively reflect the overall health of the module, especially in the early stages of degradation. Third, existing defect detection technologies are mostly limited to detecting surface problems, neglecting potential internal thermal degradation and microscopic damage that gradually manifest over time. Moreover, existing recycling value assessment methods rely primarily on static data, such as current electrical performance and appearance characteristics, failing to dynamically account for uncertainties such as market changes, policy subsidies, and external environmental factors, resulting in a lack of foresight and accuracy in the assessment results. Therefore, accurately assessing the recycling value of PV modules based on a comprehensive consideration of multidimensional data has become a major challenge in the current PV module recycling field. Summary of the Invention

[0004] In view of the above analysis, the embodiments of this application aim to provide a method and apparatus for predicting the recycling value of photovoltaic modules based on multi-dimensional feature fusion, so as to solve the problem that the existing recycling value assessment methods mainly rely on static data and cannot dynamically consider uncertainties such as market changes, policy subsidies, and external environmental factors, resulting in a lack of foresight and accuracy in the assessment results.

[0005] On one hand, embodiments of this application provide a method for predicting the recycling value of photovoltaic modules based on multi-dimensional feature fusion, including: acquiring photovoltaic module images using an infrared thermal imager and a visible light camera, and collecting electrical performance data of the photovoltaic module using multiple sensors, wherein the photovoltaic module images include infrared thermal imaging images and visible light images; performing preliminary denoising on the photovoltaic module images using wavelet transform, and then performing refined denoising on the preliminary denoised images using a generative adversarial network to generate a denoised photovoltaic module image; based on a multimodal convolutional neural network model, fusing features from the denoised infrared thermal imaging images and visible light images, and outputting defect type and severity scores; combining fuzzy logic algorithms, LSTM models, and model Carlo simulations to dynamically adjust the recycling value index from the perspectives of internal features and external uncertainties of the photovoltaic module; and constructing objective functions and corresponding constraints for maximizing recycling value, maximizing environmental benefits, and minimizing processing costs to assist decision-makers in selecting the optimal recycling strategy.

[0006] The beneficial effects of the above technical solution are as follows: By combining wavelet transform with generative adversarial networks, not only can noise in the image be effectively removed, but also minute features such as hot spot edges and cracks can be finely optimized, improving image quality. The recycling value index is further corrected through a single-objective optimization method, and combined with a multi-objective optimization algorithm, the optimal selection of photovoltaic module recycling strategies is achieved. This ensures a more accurate and comprehensive recycling value assessment, providing strong support for recycling decisions.

[0007] Further improvements to the above method include using an infrared thermal imager and a visible light camera to acquire key data of the photovoltaic module and using multiple sensors to collect electrical performance data of the photovoltaic module. This further includes: collecting surface temperature field data of the photovoltaic module using the infrared thermal imager to generate a thermal map to identify areas of abnormal temperature; capturing an image of the photovoltaic module's appearance using the visible light camera to capture visual fault information; and using current sensors, voltage sensors, and power sensors to measure the current, voltage, and power of the photovoltaic module to assess its electrical performance and health status.

[0008] Further improvements to the above method include preliminary denoising of the photovoltaic module image using wavelet transform, which further comprises: decomposing the photovoltaic module image into low-frequency coefficients and high-frequency coefficients using a three-level discrete wavelet transform; estimating the noise standard deviation based on the first-level horizontal high-frequency coefficients; generating a threshold for the j-th layer based on the noise standard deviation and the total number of pixels in the image during soft-threshold denoising, and then generating high-frequency coefficients after soft-threshold processing based on the threshold of the j-th layer and the high-frequency coefficients of the j-th layer; and reconstructing the low-frequency coefficients and the high-frequency coefficients after soft-threshold processing to generate the wavelet-denoised image.

[0009] Based on a further improvement of the above method, the generative adversarial network includes a generator and a discriminator. The process of refining the denoised image using the generative adversarial network to generate a denoised photovoltaic module image further includes: generating an optimized image based on the wavelet-denoised image using the generator; then using the discriminator with a convolutional neural network structure to determine whether the input image is a true noise-free image; obtaining the joint loss function of the generative adversarial network based on the sum of the loss of the conditional generator and the loss of the discriminator; generating a pixel-level loss function based on the denoised image output by the generator, the corresponding true noise-free image during training, and the total number of pixels in the image; performing a weighted sum of the joint loss function and the pixel-level loss function as the final training objective; and performing local contrast enhancement processing on the generator output image to generate a high-quality denoised image, wherein the high-quality denoised image includes a denoised infrared thermal imaging image and a denoised visible light image.

[0010] Further improvements to the above method, based on a multimodal convolutional neural network model, fuse the features of the denoised infrared thermal imaging image and the visible light image to output defect type and severity scores, further include: concatenating service life and material type into an embedding vector that is not an image parameter; performing a convolution operation on the denoised infrared thermal imaging image to generate an infrared feature map, and then activating the infrared feature map to generate an activated infrared feature map; performing a convolution operation on the denoised visible light image to generate a visible light feature map, and then activating the visible light feature map to generate an activated visible light feature map; and then performing a convolution operation on the activated infrared feature map and the activated visible light feature map... The visible light feature maps are stitched together to obtain image fusion features; global average pooling is performed on the image fusion features to obtain the average value of each channel, thus obtaining the image fusion feature vector; then, the image fusion feature vector and the embedding vector of the non-image parameters are stitched together using a multilayer perceptron to obtain the fused final feature vector; and the probability distribution of the defect type is calculated using a Softmax classifier based on the feature extraction and fused final feature vector, the weight matrix of the classification layer, and the bias term of the classification layer; the defect severity score is calculated based on the weight matrix of the regression layer, the feature extraction and fused final feature vector, the sigmoid function, and the bias term of the regression layer.

[0011] Further improvements to the above method, combining fuzzy logic algorithms, LSTM models, and Model Carlow simulations, dynamically adjusting the recycling value index based on the internal characteristics and external uncertainties of photovoltaic modules further include: classifying the defect levels of the photovoltaic modules into no obvious defects, minor appearance defects, moderate defects, and severe defects; quantifying the defect levels by mapping them to severity scores; using fuzzy logic algorithms to dynamically adjust the weights of each feature according to the type, service life, and defect level of the photovoltaic module, to generate dynamic weights for the type of photovoltaic module, the service life, and the defect level; and based on the module type... The initial recovery value index is calculated by weighting and summing the dynamic weights of the rating and type, the service life rating and its dynamic weight, and the defect level rating and its dynamic weight. The power prediction value is then predicted using the LSTM model, and the power degradation rate of the photovoltaic module is calculated based on the initial power and the predicted power value. A degradation correction factor is calculated based on the power degradation rate and the material degradation coefficient. Finally, a corrected recovery value index is calculated based on the initial recovery value index, the degradation correction factor, the price fluctuation correction factor, the policy subsidy correction factor, the service life index degradation coefficient, and the service life of the photovoltaic module.

[0012] Based on further improvements to the above method, the price volatility correction factor is calculated based on the recovery price of a random sample in the Monte Carlo simulation and the average of historical recovery prices; and the policy subsidy correction factor is calculated based on the policy subsidy intensity of a random sample in the Monte Carlo simulation and the average of historical policy subsidies.

[0013] Further improvements to the above method include constructing objective functions and corresponding constraints for maximizing recycling value, maximizing environmental benefits, and minimizing processing costs to assist decision-makers in selecting the optimal recycling strategy. These include: constructing an objective function for maximizing recycling value based on the modified recycling value index; constructing an objective function for maximizing environmental benefits based on the environmental benefits generated by recycling the i-th photovoltaic module; and constructing an objective function for minimizing processing costs based on the processing cost of the i-th photovoltaic module. The constraints include: resource capacity being less than or equal to the maximum processing capacity of the recycling facility; and the minimum recycling threshold being greater than or equal to the minimum operational requirements of the recycling facility.

[0014] On the other hand, embodiments of this application provide a photovoltaic module recycling value prediction device based on multi-dimensional feature fusion, comprising: an image acquisition module, used to acquire photovoltaic module images using an infrared thermal imager and a visible light camera, and to collect electrical performance data of the photovoltaic module using multiple sensors, wherein the photovoltaic module images include infrared thermal imaging images and visible light images; a denoising module, used to perform preliminary denoising on the photovoltaic module images through wavelet transform, and then to perform refined denoising on the preliminary denoised images through a generative adversarial network to generate a denoised photovoltaic module image; a defect type and scoring module, used to fuse features of the denoised infrared thermal imaging images and visible light images based on a multimodal convolutional neural network model, and output defect type and severity scores; an index dynamic adjustment module, used to dynamically adjust the recycling value index from the perspectives of internal features and external uncertainties of the photovoltaic module by combining fuzzy logic algorithms, LSTM models, and model Carlo simulations; and an objective function construction module, used to construct an objective function and corresponding constraints for maximizing recycling value, maximizing environmental benefits, and minimizing processing costs, to assist decision-makers in selecting the optimal recycling strategy.

[0015] Based on further improvements to the above-mentioned device, the noise reduction module includes a wavelet noise reduction module and a generative adversarial network (GAN) noise reduction module. The wavelet noise reduction submodule includes: a decomposition submodule, used to decompose the photovoltaic module image into low-frequency coefficients and high-frequency coefficients using a three-level discrete wavelet transform; an estimation submodule, used to estimate the noise standard deviation based on the first-layer horizontal high-frequency coefficients; a denoising submodule, used to generate a threshold for the j-th layer based on the noise standard deviation and the total number of pixels in the image during soft-threshold denoising, and then generate the soft-thresholded high-frequency coefficients based on the threshold and the j-th layer high-frequency coefficients; and a reconstruction submodule, used to reconstruct the low-frequency coefficients and the soft-thresholded high-frequency coefficients to generate the wavelet-denoised image. The GAN noise reduction module includes: a generation and judgment submodule. The generator is used to generate an optimized image based on the wavelet-denoised image, and then the discriminator uses a convolutional neural network structure to determine whether the input image is a real noise-free image; the loss function submodule is used to obtain the joint loss function of the generative adversarial network based on the sum of the loss of the conditional generator and the loss of the discriminator; the pixel-level loss function is generated based on the denoised image output by the generator, the real noise-free image corresponding during the training process, and the total number of pixels in the image; the joint loss function and the pixel-level loss function are weighted and summed as the final training target; the denoising submodule is used to perform local contrast enhancement processing on the image output by the generator to generate a high-quality denoised image in the final output, wherein the high-quality denoised image includes a denoised infrared thermal imaging image and a denoised visible light image.

[0016] Compared with the prior art, this application can achieve at least one of the following beneficial effects: 1. This application addresses the problem that existing recycling value assessment methods cannot comprehensively consider different data sources by employing multi-dimensional feature fusion technology. By combining multiple information sources such as infrared thermal imaging images, visible light images, and electrical performance data, this application utilizes a multimodal convolutional neural network model for defect detection and recycling value assessment of photovoltaic modules. This method can accurately extract features from images from different sources and comprehensively consider the health status of the modules, providing more comprehensive and accurate data input for recycling value assessment.

[0017] 2. This application employs innovative denoising techniques in image preprocessing. By combining wavelet transform with generative adversarial networks, not only can noise in the image be effectively removed, but also minute features such as hot spot edges and cracks can be meticulously optimized, improving image quality. This processing method significantly enhances image detail and accuracy, providing high-quality input data for subsequent defect detection and recovery value assessment.

[0018] 3. This application innovatively introduces non-image parameters such as service life and material type, and combines this information with image features through weighted fusion and dynamic feature adjustment to further quantify the impact of defects on photovoltaic module performance. Using fuzzy logic algorithms, the model can dynamically adjust the weights of each feature based on module type, service life, and defect level, ensuring more scientific and reasonable evaluation results.

[0019] 4. This application innovatively combines LSTM models and Monte Carlo simulations in the recycling value assessment process, enabling the quantification of the impact of future power decay rates and external uncertainties on recycling value. The recycling value index is further refined through a single-objective optimization method, and combined with a multi-objective optimization algorithm, the optimal selection of photovoltaic module recycling strategies is achieved. This ensures a more accurate and comprehensive recycling value assessment, providing strong support for recycling decisions.

[0020] In this application, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this application will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing this application. The objectives and other advantages of this application can be realized and obtained from the specific points highlighted in the description and accompanying drawings. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of a photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to an embodiment of this application; Figure 2 A thermal map is generated based on temperature field data of the photovoltaic module surface collected by an infrared thermal imager according to embodiments of this application; Figure 3 The image shows the appearance of a photovoltaic module captured by a visible light camera according to an embodiment of this application. Figure 4 This is an IV curve diagram of a photovoltaic module according to an embodiment of this application; Figure 5 The image is a wavelet-denoised image generated by wavelet reconstruction according to an embodiment of this application; Figure 6 A structural diagram of an LSTM model according to an embodiment of this application; and Figure 7 This is a block diagram of a photovoltaic module recycling value prediction device based on multi-dimensional feature fusion according to an embodiment of this application. Detailed Implementation

[0022] The preferred embodiments of this application are described in detail below with reference to the accompanying drawings, which constitute a part of this application and are used together with the embodiments of this application to illustrate the principles of this application, but are not intended to limit the scope of this application.

[0023] refer to Figure 1 One specific embodiment of this application discloses a method for predicting the recycling value of photovoltaic modules based on multi-dimensional feature fusion, which includes the following steps.

[0024] In step S101, images of the photovoltaic module are acquired using an infrared thermal imager and a visible light camera, and electrical performance data of the photovoltaic module are collected using multiple sensors. The photovoltaic module images include both infrared thermal images and visible light images. Relevant data for the photovoltaic module are collected by simultaneously acquiring key physical parameters and status information of the photovoltaic module through multi-source heterogeneous devices, constructing a multi-dimensional dataset to provide high-precision, highly consistent input data for subsequent image processing, defect analysis, and recycling strategy modeling.

[0025] The acquisition of key data from photovoltaic (PV) modules using infrared thermal imagers and visible light cameras, along with the collection of electrical performance data from multiple sensors, further includes: collecting surface temperature field data of PV modules using infrared thermal imagers to generate thermal maps for identifying areas of temperature anomalies; capturing images of the PV modules' appearance using visible light cameras to obtain visual characteristic data for detecting cracks, surface contamination, and other visible faults; and measuring the current, voltage, and power of PV modules using current, voltage, and power sensors to assess their electrical performance and health status. IV curve testing is used to measure the module's current, voltage, and power output, collecting electrical performance data and assessing its power degradation and health status.

[0026] In step S102, the photovoltaic module image is initially denoised by wavelet transform, and then the image after initial denoising is further refined by generative adversarial network to generate a denoised photovoltaic module image.

[0027] (1) Two-dimensional discrete wavelet decomposition (DWT). The photovoltaic module image is decomposed into low-frequency coefficients and high-frequency coefficients through a three-level discrete wavelet transform.

[0028] For the input image Perform a three-level discrete wavelet transform to decompose it into low-frequency approximation coefficients. and high frequency detail coefficient : ; in, Indicates the height and width of the image; Indicates the number of wavelet decomposition levels; Indicates the first Low-frequency approximation coefficients of the layer; Indicates the first High-frequency detail factor of the layer; This represents the third-order wavelet transform function. H j This represents the high-frequency detail coefficients in the horizontal direction of the j-th layer. It reflects the variations and detail information of the signal in the horizontal direction. For example, in image processing, H... j It can capture horizontal edges and textures in an image. V j V represents the high-frequency detail coefficients in the vertical direction of the j-th layer. It reflects the signal's variations and detail information in the vertical direction. In image processing, V... j It can capture vertical edges and textures. D j This represents the high-frequency detail coefficients in the diagonal direction of the j-th layer. It reflects the variations and detail information of the signal in the diagonal direction. In image processing, D... j It can capture diagonal edges and textures.

[0029] (2) Noise intensity estimation. The noise standard deviation is estimated based on the high-frequency coefficients in the horizontal direction of the first layer. The noise in the image follows a zero-mean Gaussian distribution. Using the first layer of horizontal detail coefficients To estimate the noise standard deviation : ; in: This represents the first-level level detail coefficient matrix.

[0030] (3) Soft Thresholding. In the soft thresholding process, the threshold of the j-th layer is generated based on the noise standard deviation and the total number of pixels in the image. Then, the high-frequency coefficients after soft thresholding are generated based on the threshold of the j-th layer and the high-frequency coefficients of the j-th layer.

[0031] For the high frequency coefficients of each layer Apply soft thresholding: ; Then apply the soft threshold function: ; in Indicates the first Detail factor in a certain direction of a layer; This represents the coefficient after soft thresholding. Indicates the first The threshold of the layer; This represents the total number of pixels in the image.

[0032] The advantage of soft thresholding is that it can remove noise while maintaining the smoothness of the image, making it suitable for scenarios that preserve edge information.

[0033] (4) Inverse wavelet transform. The low-frequency coefficients and the high-frequency coefficients after soft thresholding are reconstructed to generate the wavelet-denoised image. The processed low-frequency coefficients... and the updated high frequency coefficient Refactor: ; in: This represents the image after wavelet denoising. This represents the third-order inverse wavelet transform function.

[0034] Specifically, wavelet transform is used for initial image denoising, preserving the main structure of the image while removing noise. Then, a generative adversarial network (GAN) is used to further optimize the details of the denoised image, enhancing the visibility of important features. The combined use of these two methods effectively improves the quality of photovoltaic module images, providing high-quality image input for subsequent fault diagnosis and recycling value assessment.

[0035] Based on wavelet preprocessing, generative adversarial networks are further used to refine the image denoising, especially to enhance weak structures such as hot spot boundaries and cracks.

[0036] (1) GAN Model Structure. The generator produces an optimized image based on the wavelet-denoised image, and then a discriminator using a convolutional neural network structure determines whether the input image is a true noise-free image. Generator The U-Net architecture is used, and the input is the wavelet-denoised image. The output is the optimized image. Discriminator A convolutional neural network structure is used to determine whether the input image is a "real, noise-free image".

[0037] (2) Training Objective Function. The joint loss function of the generative adversarial network is obtained by summing the losses of the conditional generator and the discriminator. The following joint loss function is minimized: ; in, This represents the corresponding real, noise-free image in the training set; This represents a noisy image after wavelet preprocessing. This represents the denoised image output by the generator; It is the probability that the discriminator outputs that the image is a real image.

[0038] (3) Pixel-level reconstruction loss. A pixel-level loss function is generated based on the denoised image output by the generator, the corresponding real noise-free image during training, and the total number of pixels in the image.

[0039] To accelerate convergence and improve image detail quality, pixel-level loss is incorporated: ; The final training objective is a weighted sum of the joint loss function and the pixel-level loss function. The final training objective is in the form of a weighted combination: ; in Represents the balancing weights (usually set to...). ); Indicates resistance to loss; This indicates pixel-level error loss.

[0040] (4) Post-processing. Local contrast enhancement processing is performed on the generator output image to generate a high-quality denoised image as the final output. The high-quality denoised image includes the denoised infrared thermal image and the denoised visible light image.

[0041] To further improve image contrast, the output image of the GAN can be modified. Perform local contrast enhancement processing: ; Among them, CLAHE represents restricted contrast adaptive histogram equalization; It is the final output high-quality denoised image.

[0042] In step S103, based on the multimodal convolutional neural network model, the features of the denoised infrared thermal imaging image and the visible light image are fused to output the defect type and severity score.

[0043] By combining features from infrared thermal imaging and visible light images through a dual-stream network and employing an attention mechanism for dynamic feature fusion, the model can handle multimodal input data. The model uses Softmax and Sigmoid regression to classify defect types and predict severity scores.

[0044] (1) Data input.

[0045] 1) Input image definition Infrared thermal imaging images The image was obtained using an infrared thermal imager and is approximately [size missing]. (Height and width) and One channel.

[0046] Visible light images The image was taken using a conventional visible light camera and is [size missing]. and One channel.

[0047] 2) Non-image parameter embedding Service life This indicates the service life of the photovoltaic modules. To standardize input, the service life should be... It will be normalized to ensure that it is Within the range: ; in: It represents the maximum service life of all photovoltaic modules in the dataset. It is the normalized service life, ranging from between.

[0048] Material type This is the One-Hot code for the material type of the photovoltaic module, monocrystalline silicon: Polycrystalline silicon: ,film: .

[0049] Embedded vector Service life and material type Concatenation yields an embedding vector that is not an image parameter. : ; in, It is the dimension of the service life and material type embedding vector. It is an embedding vector containing service life and material type, used as a non-image feature input model.

[0050] (2) Dual-stream feature extraction.

[0051] 1) Infrared Image Feature Extraction: A convolution operation is performed on the denoised infrared thermal imaging image to generate an infrared feature map. Then, the infrared feature map is activated to generate an activated infrared feature map. Convolution operation: For the input infrared image To perform convolution operations, use a convolution kernel. The output feature map is obtained. : ; in, It is the input infrared image. It is a convolution kernel with a size of . It is the feature map obtained after convolution, representing the first... Position at The value of each channel. It is a bias term. As the name of the overall output tensor. Indicates spatial location in the feature map First, Second A single scalar output value for each channel.

[0052] Activation function (ReLU): Activates the feature map after convolution. Perform ReLU activation to introduce a nonlinear factor: ; in, It is a feature map after ReLU activation.

[0053] 2) Visible Light Image Feature Extraction: A convolution operation is performed on the denoised visible light image to generate a visible light feature map. Then, the visible light feature map is activated to generate an activated visible light feature map. Convolution operation: For the input visible light image To perform convolution operations, use a convolution kernel. The output feature map is obtained. : ; Activation function (ReLU): convolutional feature maps Perform ReLU activation: ; This represents the number of channels in the input image, i.e., the feature dimension contained in each pixel; This represents the number of channels in the output feature map, corresponding to the number of convolutional kernels. Each convolutional kernel independently extracts a spatial-channel joint feature. in, These are visible light image features after ReLU activation.

[0054] (3) Feature fusion.

[0055] 1) Concatenation and Global Average Pooling Feature stitching: The activated infrared feature map and the activated visible light feature map are stitched together to obtain image fusion features.

[0056] Features extracted from infrared and visible light image streams and By splicing, the fusion features are obtained. : ; Global average pooling: for the fused feature map Perform global average pooling to obtain the average value for each channel, and output the fused feature vector. : ; 2) Fusion with non-image parameter embedding vectors Multilayer perceptron fusion: The image fusion feature vector and the embedding vector of non-image parameters are concatenated by a multilayer perceptron to obtain the final fused feature vector.

[0057] Features after global average pooling embedding vectors of non-image parameters The features are concatenated to obtain the final feature vector, which is then input into a multilayer perceptron for further processing. ; It is a feedforward neural network structure, typically consisting of several fully connected layers, activation functions, and optional normalization layers. It represents a multilayer perceptron, which contains multiple linear layers and activation functions.

[0058] (4) Output layer design In photovoltaic module defect detection, a Softmax classifier is used to determine which defect type each image or sample belongs to. Since we have multiple defect types, the Softmax classifier calculates the probability distribution for each defect type and ultimately outputs the defect type with the highest probability.

[0059] The Softmax function is commonly used for multi-class classification tasks. It transforms a vector (the score or raw output for each class) into a probability distribution. The probability value for each class is between 0 and 1, and the sum of the probabilities for all classes is 1. The Softmax classifier calculates the probability distribution of defect types based on the final feature vector after feature extraction and fusion, the weight matrix of the classification layer, and the bias term of the classification layer. The formula is as follows: ; in, It is the probability distribution of defect types, representing the predicted probability of each category. It is the weight matrix of the classification layer, representing the influence of each input feature on each category. It is a feature vector derived from feature extraction and fusion. It is a bias term for the classification layer.

[0060] The function transforms the result of a linear combination into a probability distribution, and the specific calculation formula is as follows: ; in, It is the result of a linear combination of the inputs, that is... . It represents the number of categories. It calculates the index for each category, ensuring that all output probabilities are positive, and that larger scores correspond to higher probabilities.

[0061] The Sigmoid function is commonly used in binary classification tasks. It maps a real number to the range of 0 to 1, which can be interpreted as the probability of a certain class. The defect severity score is calculated based on the weight matrix of the regression layer, the final feature vector after feature extraction and fusion, the Sigmoid function, and the bias term of the regression layer. The calculation formula is as follows: ; in, This is a defect severity score, indicating the degree of severity of the defect. The value ranges from 0 to 1, where 0 indicates a minor defect and 1 indicates a severe defect. It is the weight matrix of the regression layer, representing the influence of input features on severity scores. These are feature vectors extracted from a two-stream network. It is the bias term of the regression layer.

[0062] It is the Sigmoid function, and its calculation formula is: ; in, It is a linear combination of the inputs.

[0063] x is the linear input term of the regression model, which serves as the input to the Sigmoid function.

[0064] In step S104, the recycling value index is dynamically adjusted from the perspectives of internal characteristics and external uncertainties of photovoltaic modules by combining fuzzy logic algorithm, LSTM model and model Carlo simulation.

[0065] By combining fuzzy logic algorithms, LSTM models, and Monte Carlo simulations, the recovery value index is dynamically adjusted from both the internal characteristics of the components and external uncertainties. Ultimately, by correcting the recovery value index (RVI), the scientific rigor and practicality of the assessment are improved.

[0066] (1) Definition and quantification of defect levels.

[0067] Photovoltaic modules are classified into four defect levels: no obvious defects, minor appearance defects, moderate defects, and severe defects. Specifically: Grade A: No obvious defects, performance close to that of a brand new module. Grade B: Minor appearance defects, not affecting function. Grade C: Moderate defects, may affect efficiency or lifespan. Grade D: Severe defects, requiring downgrading or scrapping.

[0068] (2) Quantification method of defect level.

[0069] Defect levels are quantified by mapping defect levels to severity scores. For example, defect levels are mapped to severity scores S∈[0,1]: Grade A: S=0.1 (minor defect) Grade B: S=0.3 (moderate defect) Grade C: S=0.6 (serious defect) Grade D: S=0.9 (serious defect).

[0070] (3) Dynamic weight adjustment.

[0071] The core of dynamic weight adjustment is to use fuzzy logic algorithms to dynamically adjust the weights of each feature based on the component type, service life, and defect level.

[0072] Fuzzy logic algorithms are reasoning methods used to handle uncertainty and fuzzy information. Based on linguistic variables and membership functions, the algorithm converts quantitative inputs into fuzzy sets and performs reasoning operations through a pre-defined fuzzy rule base. The membership degree of each input variable represents its strength of belonging in different fuzzy intervals, and the fuzzy rules output corresponding fuzzy conclusions based on the combination relationships between variables. Finally, a defuzzification process transforms the fuzzy output into specific numerical results, achieving automatic adjustment of dynamic weights for multiple features.

[0073] This algorithm enables flexible modeling of the influence of photovoltaic module features, exhibiting good adaptability and generalization capabilities. Even under conditions of complex defect information, incomplete or highly volatile state data, the fuzzy logic mechanism still provides stable and reliable weighted output results, significantly enhancing the robustness of recycling value assessment. It allows the weights of each key feature to dynamically change according to the actual state of the module, avoiding errors that may arise from static models and improving the personalization and accuracy of the overall assessment.

[0074] 1) Input variables Component Type

[0075] Service life Normalized to: ; in, It represents the maximum service life of photovoltaic modules in the dataset.

[0076] Defect Level Quantified as a severity score .

[0077] 2) Membership function design: Define a Sigmoid membership function for each input variable to characterize the influence of these variables on the weights.

[0078] 3) Fuzzy rule base: A fuzzy rule base is constructed based on component type, service life, and defect level.

[0079] 4) Weight Calculation: Using a fuzzy logic algorithm, the weights of each feature are dynamically adjusted based on the type, service life, and defect level of the photovoltaic module to generate dynamic weights for the type of photovoltaic module, service life, and defect level.

[0080] The weights of each variable are calculated using a fuzzy rule base and membership functions: ; , , These are the dynamic weights for component type, service life, and defect level, respectively.

[0081] , , : The initial weight corresponding to each rule.

[0082] (4) Preliminary RVI calculation: The preliminary recovery value index is calculated by weighted summation based on component type score and dynamic weight of type, service life score and dynamic weight of service life, and defect level score and dynamic weight of defect level.

[0083] Based on dynamic weight adjustment, the preliminary recovery value index calculation formula is as follows: ; in, These represent the dynamic weights for component type, service life, and defect level, respectively. This indicates the component type score. To represent the exponential decay function of service life, use the exponential decay function: ; The defect level rating is defined as follows: ; in It is a defect severity score, with a range of . It is the attenuation coefficient, which determines the impact of service life on recycling value.

[0084] (5) Power decay prediction. LSTM models are used to capture long-term dependencies in time series. When building an LSTM model, we need to define the input layer, the LSTM layer, and the fully connected layer.

[0085] refer to Figure 6The core of the LSTM model lies in its gating mechanism, particularly the input gate, forget gate, and output gate. Through these gating mechanisms, LSTM can effectively learn long-term dependencies in time series data.

[0086] Input gate ( ): Controls the impact of the current input on the model's memory units.

[0087] Forgotten Gate ( ): Determine how much information from the previous time step to forget.

[0088] Output gate ( ): Determines the output at the current moment.

[0089] The LSTM calculation process is based on the following formula: Input Gate: ; Forgotten Gate: ; Output gate: ; Cell status update: ; Hidden status update: ; in: It is the input data at the current moment. It is the hidden state from the previous moment. It is the weight matrix of each gate. These are the bias terms for each gate. It represents the current unit state and controls the transmission of information.

[0090] Ultimately, LSTM outputs the future. Annual power forecast .

[0091] Based on the prediction results of the LSTM model, the power degradation rate of the photovoltaic module is calculated. The LSTM model is used to predict the power output, and then the power degradation rate of the photovoltaic module is calculated based on the initial power and the predicted power output. The calculation formula is as follows: ; in: It is the initial power, that is, the power of the component at the initial moment. Is the future number The predicted power for the year.

[0092] The attenuation correction factor is calculated based on the power attenuation rate and material attenuation coefficient of the photovoltaic module. Attenuation Correction Factor Calculated using the following formula: ; in, It is the material attenuation coefficient, which represents the attenuation characteristics of different photovoltaic module materials.

[0093] (6) Quantification of uncertainty To account for uncertainties such as market volatility, changes in policy subsidies, and component lifespan, we use Monte Carlo simulations to quantify these uncertainties.

[0094] 1) Setting random variables Market price fluctuations: Future photovoltaic module recycling prices will follow a normal distribution. .

[0095] Policy subsidy changes: Subsidy levels follow a uniform distribution .

[0096] Component lifespan correction: service life Uncertainty (such as) ).

[0097] 2) Simulation process generate Sub-random scenarios (such as) ).

[0098] For each scenario, calculate the corrected RVI: ; RVI sim : This represents the statistical expected recovery value obtained by performing multiple simulations of the residual value during the component's lifespan using the Monte Carlo simulation method.

[0099] 3) Correction factor calculation The price volatility correction factor is calculated based on the recovery price of a random sample from a Monte Carlo simulation and the average of historical recovery prices. Price volatility correction factor: ; in, It is the recovery price of a random sample in a Monte Carlo simulation. It is the average of historical recycling prices.

[0100] The policy subsidy adjustment factor is calculated based on the policy subsidy intensity of a random sample in a Monte Carlo simulation and the average historical policy subsidy. Policy subsidy adjustment factor: ; in, This refers to the level of policy subsidies in a random sample from a Monte Carlo simulation. It is the average of historical policy subsidies.

[0101] (7) Optimized recovery value index formula The final RVI correction formula is as follows: The corrected recovery value index is calculated based on the initial recovery value index, degradation correction factor, price fluctuation correction factor, policy subsidy correction factor, service life index degradation coefficient, and the service life of photovoltaic modules.

[0102] Based on the predicted power degradation of photovoltaic modules and the statistical characteristics of Monte Carlo simulations, the final correction formula for the recycling value index is constructed as follows: ; RVI cor The revised value of the recovery value index is based on preliminary forecasts (such as logical models or expert rules) and adjusted by comprehensively considering the following factors: photovoltaic power degradation factor. Service life index Policy subsidy adjustment factor Market price volatility factor . This represents the preliminary recovery value index, a preliminary value calculated based on a fuzzy logic algorithm. This represents the power attenuation correction factor. This represents the service life index decay coefficient. This represents the price volatility correction factor. This indicates the policy subsidy adjustment factor. This indicates the service life of the photovoltaic module.

[0103] Through these steps of correction and optimization, the revised recycling value index was finally obtained. This value integrates multiple factors such as component defect type, severity score, power decay rate, market fluctuations, and policy subsidies, and can provide an accurate basis for recycling decisions.

[0104] In step S105, an objective function and corresponding constraints are constructed to maximize recycling value, maximize environmental benefits, and minimize treatment costs, in order to assist decision-makers in selecting the optimal recycling strategy.

[0105] (1) Construction of the objective function.

[0106] Objective 1: Maximize economic benefits. Recycling value is a key indicator in the photovoltaic module recycling process. An objective function for maximizing recycling value is constructed based on a modified recycling value index. The economic benefit formula is: ; It is the modified recycling value index, which can be calculated based on the aforementioned model, x i Let represent the recycling decision variable for the i-th photovoltaic module.

[0107] Objective 2: Maximize environmental benefits. Based on the environmental benefits generated by recycling the i-th photovoltaic module, construct an objective function to maximize environmental benefits. The formula for calculating environmental benefits is: ; in, It is the first The environmental benefits of recycling photovoltaic modules, such as carbon emission reductions or resource savings.

[0108] The formula for calculating carbon emission reduction is: ; The formula for calculating resource recovery volume is: ; in, It is the first Each photovoltaic module reduces carbon emissions through recycling. It is the carbon emissions corresponding to the new materials required to produce this component. It refers to the carbon emissions generated from producing the same components using recycled materials. No. The resource recycling value of each component. It is the first Material recycling rate of each component. It is the unit value of the material (e.g., yuan / kg).

[0109] Objective 3: Minimize processing cost. Construct an objective function to minimize processing cost based on the processing cost of the i-th photovoltaic module. The formula for calculating processing cost is: ; Indicates the first The processing cost of each component (RMB / component) includes transportation, dismantling and reprocessing costs.

[0110] (1) Constraints include: resource capacity less than or equal to the maximum processing capacity of the recycling facility; minimum recycling threshold greater than or equal to the minimum operational requirements of the recycling facility. To ensure the feasibility of the recycling strategy, the model needs to meet the following constraints: Resource capacity limit: ; This is the maximum processing capacity of the recycling facility.

[0111] Minimum recycling threshold: ; These are the minimum operating requirements for recycling facilities.

[0112] (3) Implementation steps of NSGA-II algorithm 1) Initialize the population Encoding method: Uses binary encoding, with each gene bit indicating whether a component is recycled.

[0113] Population size: Set the population size .

[0114] Initialization: Randomly generate solutions that satisfy the constraints, ensuring that: ; 2) Non-dominated ranking and crowding calculation Non-dominated sorting The populations are ordered according to Pareto dominance. The Pareto front is divided into multiple levels, and crowding is calculated. The crowding distance for each individual is calculated using the following formula: ; Represents an individual The degree of congestion.

[0115] and Describe the objective function In individuals The target values ​​of the two neighbors on the right and left sides.

[0116] and Describe the objective function The maximum and minimum values ​​in the entire population.

[0117] 3) Genetic manipulation Based on roulette or tournament selection, solutions with high non-dominant levels and low crowding are prioritized.

[0118] Single-point intersection: Set the intersection point Then swap the parent individuals and Some genes: ; ; Uniform crossover: During the crossover process, each gene locus is randomly selected from the parent generation. or parent This is how two offspring are generated.

[0119] To avoid premature convergence, certain gene loci are randomly flipped with a small probability (e.g., 5%). ; 4) Constraint handling.

[0120] Hard constraint handling: If Then randomly add unselected components until the condition is met. .if If the limit is exceeded, then components that exceed the limit will be randomly deleted.

[0121] Soft constraint handling: Add a penalty term to the objective function: ; in, It is a penalty coefficient, which penalizes solutions that violate the constraints.

[0122] 5) Iteration termination condition Maximum number of iterations: Sets the maximum number of iterations. .

[0123] Convergence criterion: If the solution of the Pareto front does not change significantly during continuous iteration, terminate the iteration early.

[0124] refer to Figure 7 One specific embodiment of this application discloses a photovoltaic module recycling value prediction device based on multi-dimensional feature fusion, comprising: an image acquisition module 701, used to acquire photovoltaic module images using an infrared thermal imager and a visible light camera, and to collect electrical performance data of the photovoltaic module using multiple sensors, wherein the photovoltaic module images include infrared thermal imaging images and visible light images; a noise reduction module 702, used to perform preliminary noise reduction on the photovoltaic module images through wavelet transform, and then to perform refined noise reduction on the preliminary noise-reduced images through a generative adversarial network to generate a noise-reduced photovoltaic module image; a defect type and scoring module 703, used to fuse features of the noise-reduced infrared thermal imaging images and visible light images based on a multimodal convolutional neural network model, and output defect type and severity scores; an index dynamic adjustment module 704, used to dynamically adjust the recycling value index from the perspectives of internal features and external uncertainties of the photovoltaic module by combining fuzzy logic algorithms, LSTM models and model Carlo simulations; and an objective function construction module 705, used to construct an objective function and corresponding constraints for maximizing recycling value, maximizing environmental benefits and minimizing processing costs, to assist decision-makers in selecting the optimal recycling strategy.

[0125] The noise reduction module 702 includes a wavelet noise reduction module and a generative adversarial network noise reduction module.

[0126] The wavelet denoising submodule includes: a decomposition submodule, used to decompose the photovoltaic module image into low-frequency coefficients and high-frequency coefficients through a three-level discrete wavelet transform; an estimation submodule, used to estimate the noise standard deviation based on the first-level horizontal high-frequency coefficients; a denoising submodule, used to generate the threshold of the j-th layer based on the noise standard deviation and the total number of pixels in the image during the soft-threshold denoising process, and then generate the high-frequency coefficients after soft-threshold processing based on the threshold of the j-th layer and the high-frequency coefficients of the j-th layer; and a reconstruction submodule, used to reconstruct the low-frequency coefficients and the high-frequency coefficients after soft-threshold processing to generate the wavelet-denoised image.

[0127] The generative adversarial network (GAN) denoising module includes: a generation and judgment submodule, which generates an optimized image based on the wavelet-denoised image using a generator, and then uses a discriminator with a convolutional neural network structure to determine whether the input image is a true noise-free image; a loss function submodule, which obtains the joint loss function of the GAN based on the sum of the loss of the conditional generator and the loss of the discriminator; generates a pixel-level loss function based on the denoised image output by the generator, the corresponding true noise-free image during training, and the total number of pixels in the image; and performs a weighted summation of the joint loss function and the pixel-level loss function as the final training objective; and a denoising submodule, which performs local contrast enhancement processing on the generator output image to generate a high-quality denoised image, wherein the high-quality denoised image includes a denoised infrared thermal imaging image and a denoised visible light image.

[0128] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0129] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the recycling value of photovoltaic modules based on multi-dimensional feature fusion, characterized in that, include: The photovoltaic module images are acquired using an infrared thermal imager and a visible light camera, and the electrical performance data of the photovoltaic module are collected using multiple sensors, wherein the photovoltaic module images include infrared thermal images and visible light images; The photovoltaic module image is initially denoised by wavelet transform, and then refined by generative adversarial network to generate a denoised photovoltaic module image. Based on a multimodal convolutional neural network model, the features of the denoised infrared thermal imaging image and the visible light image are fused to output the defect type and severity score; By combining fuzzy logic algorithms, LSTM models, and Model Carlo simulations, the recycling value index is dynamically adjusted based on the internal characteristics and external uncertainties of photovoltaic modules. We construct objective functions and corresponding constraints for maximizing recycling value, maximizing environmental benefits, and minimizing treatment costs to assist decision-makers in selecting the optimal recycling strategy.

2. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, The use of infrared thermal imagers and visible light cameras to acquire key data of photovoltaic modules, and the use of multiple sensors to collect electrical performance data of the photovoltaic modules, further include: The infrared thermal imager collects surface temperature field data of the photovoltaic module and generates a thermal map to identify areas of abnormal temperature. The visible light camera captures an image of the photovoltaic module's appearance, capturing visual fault information; and Current, voltage, and power sensors are used to measure the current, voltage, and power of the photovoltaic module to assess its electrical performance and health status.

3. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, Preliminary denoising of the photovoltaic module image is performed using wavelet transform, and further includes: The photovoltaic module image is decomposed into low-frequency coefficients and high-frequency coefficients by a three-level discrete wavelet transform. The noise standard deviation is estimated based on the high-frequency coefficients in the horizontal direction of the first layer. In the soft thresholding denoising process, the threshold of the j-th layer is generated based on the noise standard deviation and the total number of pixels in the image. Then, the high-frequency coefficients after soft thresholding are generated according to the threshold of the j-th layer and the high-frequency coefficients of the j-th layer, where j is 1, 2 or 3. The low-frequency coefficients and the high-frequency coefficients after soft thresholding are reconstructed to generate a wavelet-denoised image.

4. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 3, characterized in that, The generative adversarial network includes a generator and a discriminator, wherein the process of refining the initially denoised image using the generative adversarial network to generate a denoised photovoltaic module image further includes: The generator generates an optimized image based on the wavelet-denoised image, and then the discriminator uses a convolutional neural network structure to determine whether the input image is a real noise-free image. The joint loss function of the generative adversarial network is obtained by summing the loss of the condition generator and the loss of the discriminator. A pixel-level loss function is generated based on the denoised image output by the generator, the corresponding real noise-free image during the training process, and the total number of pixels in the image. The joint loss function and the pixel-level loss function are weighted and summed as the final training objective; and The generator output image is subjected to local contrast enhancement processing to generate a high-quality denoised image as the final output, wherein the high-quality denoised image includes a denoised infrared thermal imaging image and a denoised visible light image.

5. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 3, characterized in that, Based on a multimodal convolutional neural network model, by fusing features from denoised infrared thermal imaging and visible light images, the output defect type and severity score further include: The service life and material type are concatenated into an embedding vector that is not an image parameter; The denoised infrared thermal imaging image is convolved to generate an infrared feature map, and then the infrared feature map is activated to generate an activated infrared feature map; the denoised visible light image is convolved to generate a visible light feature map, and then the visible light feature map is activated to generate an activated visible light feature map. The activated infrared feature map and the activated visible light feature map are concatenated to obtain image fusion features; global average pooling is performed on the image fusion features to obtain the average value of each channel, thus obtaining the image fusion feature vector; then, the image fusion feature vector and the embedding vector of the non-image parameters are concatenated using a multilayer perceptron to obtain the final fused feature vector; and The probability distribution of defect types is calculated using the Softmax classifier based on the final feature vector after feature extraction and fusion, the weight matrix of the classification layer, and the bias term of the classification layer. The defect severity score is calculated based on the weight matrix of the regression layer, the final feature vector after feature extraction and fusion, the sigmoid function, and the bias term of the regression layer.

6. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 3, characterized in that, Combining fuzzy logic algorithms, LSTM models, and Model Carlow simulations, the recycling value index is dynamically adjusted based on the internal characteristics and external uncertainties of photovoltaic modules, further including: The defect levels of the photovoltaic modules are classified into no obvious defects, minor appearance defects, moderate defects, and severe defects. Defect levels are quantified by mapping the defect levels to severity scores. Using a fuzzy logic algorithm, the weights of each feature are dynamically adjusted based on the type, service life, and defect level of the photovoltaic module to generate dynamic weights for the type of photovoltaic module, the service life, and the defect level. The preliminary recovery value index is calculated by weighted summation based on component type score and dynamic weight of type, service life score and dynamic weight of service life, and defect level score and dynamic weight of defect level. The LSTM model is used to predict the power forecast value, and then the power degradation rate of the photovoltaic module is calculated based on the initial power and the power forecast value of the photovoltaic module. The attenuation correction factor is calculated based on the power attenuation rate and material attenuation coefficient of the photovoltaic module. The corrected recycling value index is calculated based on the initial recycling value index, the degradation correction factor, the price fluctuation correction factor, the policy subsidy correction factor, the service life index degradation coefficient, and the service life of the photovoltaic module.

7. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 6, characterized in that, The price volatility correction factor is calculated based on the recovery price of a random sample in a Monte Carlo simulation and the average of historical recovery prices; and The policy subsidy correction factor is calculated based on the policy subsidy intensity of a random sample in a Monte Carlo simulation and the average value of historical policy subsidies.

8. The photovoltaic module recycling value prediction method based on multi-dimensional feature fusion according to claim 6, characterized in that, Constructing objective functions and corresponding constraints to maximize recycling value, maximize environmental benefits, and minimize treatment costs to assist decision-makers in selecting the optimal recycling strategy further includes: Construct an objective function to maximize recycling value based on the modified recycling value index; Construct an objective function that maximizes the environmental benefits based on the environmental benefits generated by recycling the i-th photovoltaic module; and Construct an objective function that minimizes the processing cost based on the processing cost of the i-th photovoltaic module; The constraints include: the resource capacity is less than or equal to the maximum processing capacity of the recycling facility; and the minimum recycling threshold is greater than or equal to the minimum operating requirements of the recycling facility.

9. A photovoltaic module recycling value prediction device based on multi-dimensional feature fusion, characterized in that, include: An image acquisition module is used to acquire images of photovoltaic modules using an infrared thermal imager and a visible light camera, and to collect electrical performance data of the photovoltaic modules using multiple sensors, wherein the photovoltaic module images include infrared thermal images and visible light images; The noise reduction module is used to perform preliminary noise reduction on the photovoltaic module image through wavelet transform, and then to perform fine noise reduction on the image after preliminary noise reduction through generative adversarial network to generate a noise-reduced photovoltaic module image. The defect type and scoring module is used to fuse features of denoised infrared thermal imaging images and visible light images based on a multimodal convolutional neural network model, and output defect type and severity score; The index dynamic adjustment module is used to dynamically adjust the recovery value index from the perspectives of internal characteristics and external uncertainties of photovoltaic modules by combining fuzzy logic algorithms, LSTM models and Model Carlow simulations. The objective function construction module is used to construct objective functions and corresponding constraints that maximize recycling value, maximize environmental benefits, and minimize treatment costs, in order to assist decision-makers in selecting the optimal recycling strategy.

10. The photovoltaic module recycling value prediction device based on multi-dimensional feature fusion according to claim 9, characterized in that, The noise reduction module includes a wavelet noise reduction module and a generative adversarial network noise reduction module, wherein the wavelet noise reduction submodule includes: The decomposition submodule is used to decompose the photovoltaic module image into low-frequency coefficients and high-frequency coefficients through a three-level discrete wavelet transform. The estimation submodule is used to estimate the noise standard deviation based on the first layer of horizontal high-frequency coefficients; The noise reduction submodule is used to generate the threshold of the j-th layer based on the noise standard deviation and the total number of pixels in the image during the soft thresholding process, and then generate the high-frequency coefficients after soft thresholding based on the threshold of the j-th layer and the high-frequency coefficients of the j-th layer, where j is 1, 2 or 3. The reconstruction submodule is used to reconstruct the low-frequency coefficients and the high-frequency coefficients after soft thresholding to generate a wavelet-denoised image; The generative adversarial network noise reduction module includes: The generation and judgment submodule is used to generate an optimized image based on the wavelet-denoised image through the generator, and then use a discriminator to determine whether the input image is a real noise-free image using a convolutional neural network structure. The loss function submodule is used to obtain the joint loss function of the generative adversarial network based on the sum of the loss of the conditional generator and the loss of the discriminator; to generate a pixel-level loss function based on the denoised image output by the generator, the corresponding real noise-free image during training, and the total number of pixels in the image; and to perform a weighted summation of the joint loss function and the pixel-level loss function as the final training target. The denoising submodule is used to perform local contrast enhancement processing on the generator output image to generate a high-quality denoised image as the final output, wherein the high-quality denoised image includes a denoised infrared thermal imaging image and a denoised visible light image.

Images