An AI-based method for quality inspection of photovoltaic modules

Abstract

The present invention relates to the field of artificial intelligence and photovoltaic detection technology, and particularly to a method for detecting the quality of photovoltaic modules based on artificial intelligence. The method includes: constructing a spectral-geometric model and a multi-field coupling model through multi-modal data acquisition, performing feature fusion by combining a hierarchical attention mechanism, using a generative adversarial network to extract key deep features, accurately locating the defect area and evaluating the severity. In addition, the solution simulates the operating conditions of the module through digital twin technology, and combines a time-series coupling prediction network to predict the defect expansion trend, achieving the efficiency and foresight of defect detection. The present invention can be widely applied to the field of photovoltaic power generation, significantly improving the accuracy and efficiency of module defect detection, and providing technical support for the efficient operation and maintenance of photovoltaic systems.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and photovoltaic detection, and particularly to a method for detecting the quality of photovoltaic modules based on artificial intelligence. Background Art

[0002] Photovoltaic modules are the core equipment of a photovoltaic power generation system, and their quality directly affects the power generation efficiency and system life. However, photovoltaic modules operate in a complex environment for a long time (such as high temperature, ultraviolet rays, humidity, etc.), and are prone to defects such as hidden cracks, hot spots, broken grids, and virtual soldering. These defects will lead to a decrease in power generation efficiency and even cause more serious failures.

[0003] In the prior art (Chinese published invention patent, publication number: CN114399484A, title: A Method, Device, Equipment and Medium for Detecting Defects in Photovoltaic Modules), electro-luminescence (EL) image detection and thermal imaging detection methods are generally adopted. These methods mainly rely on single-modal image features. For example, EL images are used to identify surface defects such as surface cracks and broken grids in photovoltaic modules, and thermal imaging mainly identifies hot spots. However, these technologies have the following defects: they cannot comprehensively identify internal defects (such as circuit short circuits) and dynamic defects (such as hidden crack expansion); limited by single-modal data, they lack a comprehensive characterization of complex defects, resulting in a high false negative or false positive rate; relying on manual inspection and sampling detection, the detection efficiency is difficult to meet the large-scale detection requirements of photovoltaic power stations. Summary of the Invention

[0004] In view of the above-mentioned many problems existing in the prior art, the present invention provides a method for detecting the quality of photovoltaic modules based on artificial intelligence. The present invention collects multi-modal data (including spectral, thermal, current and geometric data), constructs a spectral-geometric model and a multi-field coupling model, and uses feature mapping and deep learning for multi-modal feature fusion and defect detection. At the same time, combined with digital twin and time series prediction technologies, the operating state of photovoltaic modules is dynamically simulated, defects are accurately identified, and the defect expansion trend is predicted. The present invention significantly improves the accuracy and efficiency of photovoltaic module defect detection, and provides a scientific basis for the operation and maintenance management of modules.

[0005] A method for detecting the quality of photovoltaic modules based on artificial intelligence includes the following steps:

[0006] Collect the spectral polarization data, thermal imaging data, current imaging data and stereo depth data of the photovoltaic module, perform time synchronization and spatial calibration on all data, and generate preliminary multi-modal feature data;

[0007] Based on the preliminary multi-modal feature data, construct a spectral-geometric model and a multi-field coupling model, and generate model data respectively; perform fusion processing on the generated model data through a feature mapping algorithm, and input it into a generative adversarial network for feature abstraction to obtain key deep feature data;

[0008] Input the key deep feature data into a hierarchical attention mechanism for global and local feature fusion to generate fused defect feature data; use a region detection algorithm to locate the defect regions in the fused defect feature data, and classify and evaluate the severity of the defect regions to generate defect classification data and defect evaluation data;

[0009] Load the model data into a digital twin platform to simulate the operating conditions and generate virtual detection data; perform time series analysis on the virtual detection data to predict the expansion trend of the defect and optimize the detection system parameters.

[0010] Preferably, the spectral polarization data captures the multi-angle spectral characteristics of the surface of the photovoltaic module by adjusting the polarization angle, the thermal imaging data is collected by recording the temperature distribution characteristics of the surface of the photovoltaic module, the current imaging data is obtained by applying a small current to the photovoltaic module and synchronously recording the current distribution characteristics, and the stereo depth data generates the three-dimensional morphological characteristics of the surface of the module through multi-angle ray tracing.

[0011] Preferably, the time synchronization is based on the timestamp technology, and a unified time reference is used to label the acquisition times of the spectral polarization data, thermal imaging data, current imaging data, and stereo depth data; the spatial calibration is performed through a feature point matching algorithm to align the abnormal regions in the thermal imaging data and the current imaging data, and map the aligned results to the three-dimensional coordinate space of the stereo depth data to generate preliminary multi-modal feature data after spatial calibration.

[0012] Preferably, the feature mapping algorithm realizes the dynamic mapping of the spectral-geometry model data by calculating the correlation matrix between the spectral reflection characteristics and the geometric morphology. At the same time, by performing multi-field tensor decomposition on the thermal imaging data and the current imaging data, the correlation features of the internal stress field and the electric field of the module are extracted, and the correlation features are jointly modeled to generate the fused model data after feature mapping.

[0013] Preferably, the fused model data is input into a generative adversarial network, which includes a generator and a discriminator. The generator optimizes to generate deep abstract features through backpropagation, and the discriminator optimizes the feature data output by the generator by judging the difference between the generated features and the real features to generate key deep feature data after enhancement.

[0014] Preferably, the hierarchical attention mechanism performs weighted processing on the overall spectral distribution, thermal distribution, and current characteristics of the photovoltaic module through a global feature extraction module, and at the same time focuses on the spectral abnormal points, hot spot positions, and current abnormal points in the defect regions through a local feature extraction module to achieve the efficient fusion of global features and local features and generate fused defect feature data.

[0015] Preferably, the area detection algorithm performs multi-level feature extraction on the fused defect feature data. First, it roughly determines the defective area through an edge detection algorithm, and then optimizes the fine-grained boundary of the defective area through an improved area extraction algorithm. Finally, it outputs the precise position coordinates and area boundary information of the defective area.

[0016] Preferably, the timing analysis includes performing time series decomposition on the virtual detection data, extracting the dynamic change patterns of the defective area, including the hot spot expansion rate, the resistance change rate, and the spectral reflectance fluctuation, and extrapolating the defect expansion trend based on the timing coupling prediction network to generate defect trend prediction data.

[0017] Preferably, optimizing the detection system parameters includes adjusting the acquisition frequency of the spectral imaging device, dynamically adjusting the temperature sensitivity threshold of the thermal imaging device, optimizing the sampling interval of the current imaging system, and updating the learning rate of the generative adversarial network, so as to meet the defect detection requirements of photovoltaic modules under different operating environments.

[0018] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0019] Through the multi-modal data acquisition and fusion technology means, the present invention realizes the comprehensive characterization of the surface and internal defects of photovoltaic modules;

[0020] Through the hierarchical attention mechanism technology means, the present invention realizes the efficient fusion of global and local features, improving the accuracy of defect recognition;

[0021] Through the timing coupling prediction network technology means, the present invention realizes the accurate prediction of the defect expansion trend, providing forward-looking support for the preventive maintenance of photovoltaic modules;

[0022] Through the dynamic parameter optimization technology means, the present invention realizes the adaptive adjustment of the detection system to complex operating environments, improving the detection efficiency and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic flow chart of the method of the present invention;

[0024] Figure 2 is a flow chart of feature fusion and generative adversarial network in the present invention;

[0025] Figure 3 is a flow chart of defect detection and evaluation in the present invention;

[0026] Figure 4 is a flow chart of virtual working condition simulation and timing analysis in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0027] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0028] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0030] As Figure 1 shown, an artificial intelligence-based photovoltaic module quality inspection method includes the following steps:

[0031] Collect spectral polarization data, thermal imaging data, current imaging data, and stereo depth data of the photovoltaic module, perform time synchronization and spatial calibration on all data, and generate preliminary multimodal feature data;

[0032] Preferably, the spectral polarization data captures the multi-angle spectral characteristics of the surface of the photovoltaic module by adjusting the polarization angle, the thermal imaging data is collected by recording the temperature distribution characteristics of the surface of the photovoltaic module, the current imaging data is obtained by applying a small current to the photovoltaic module and synchronously recording the current distribution characteristics, and the stereo depth data generates the three-dimensional morphological characteristics of the surface of the module through multi-angle ray tracing.

[0033] Principle and Application of Spectral Polarization Data Acquisition

[0034] Principle: Spectral polarization imaging technology utilizes the polarization characteristics of light to capture the reflection spectra of different angles on the surface of the photovoltaic module by changing the polarization angle of the incident light. The interaction between polarized light and the surface structure and optical properties of the material can reveal subtle differences in surface cracks, scratches, and coating defects. By adjusting the polarization angle (such as 0°, 45°, 90°, etc.) and the incident direction, the full-angle spectral characteristics of the surface can be obtained, forming a set of polarization spectral data.

[0035] Implementation: In the present invention, an automated polarization angle regulator is adopted to control the angle change of the polarizer, and a spectral camera is used to record the reflected spectral intensity at each polarization angle. The generated data can be expressed by the following formula:

[0036] where, represents a certain polarization angle and wavelength at which the spectral intensity is the reflectivity; is the polarization degree function of the light source. Through the captured spectral polarization data, the hidden cracks and scratch areas on the surface of the component can be accurately identified, especially the subtle optical property differences that are difficult to detect by the naked eye.

[0037] Thermal imaging technology records the temperature distribution on the surface of the photovoltaic module in a non-contact manner, and uses the local temperature rise phenomenon caused by internal hidden cracks, hot spots or material aging during the operation of the module to generate a thermal image of the module. The infrared sensor captures the infrared energy on the surface of the material according to the principle of thermal radiation and converts it into temperature distribution data. In the present invention, the thermal imaging camera collects infrared images in the operating state at a fixed frequency. By calculating the radiance of each pixel point and converting it into temperature :

[0038]

[0039] where, is the emissivity of the material, is the Stefan-Boltzmann constant, is the infrared radiance. Through the thermal imaging data, the abnormal heating areas caused by hidden cracks, hot spots or poor soldering can be quickly identified, and basic data for defect classification can be provided.

[0040] Current imaging records the distribution characteristics of the current on the surface of the photovoltaic module by applying a small current to the module, and is used to capture local resistance anomalies caused by hidden cracks or poor soldering. The distribution of the current is synchronously collected by a current sensor and a thermal infrared imaging device, and the integrity of the current path can be analyzed. In the present invention, a highly sensitive current sensor and a thermal infrared imaging device are used to work together to collect current imaging data under the condition of injecting a small current, and the resistance distribution is calculated by the following formula:

[0041]

[0042] where, is the local resistance at the pixel position , is the voltage at this position, is the injected current. The generated current imaging data can accurately locate the areas with local resistance anomalies. When combined with thermal imaging data for joint analysis, it can effectively improve the reliability of detecting hidden cracks and soldering defects.

[0043] Stereo depth data generates three-dimensional morphological features through ray tracing technology, and uses a multi-angle light field to capture the surface structure information of the component. The light field sensor calculates the depth information of the component surface by recording the light distribution at different incident angles. The present invention adopts a multi-view light field imaging system. Cameras with a fixed angular interval collect the light distribution on the component surface from different perspectives, and use the triangulation method to calculate the depth:

[0044]

[0045] where is the depth, is the camera focal length, is the baseline distance, is the parallax. The stereo depth data can be used to identify macroscopic geometric defects on the component surface, such as warping, depression, and crack areas, providing a geometric basis for constructing a spectral-geometric model.

[0046] Example: In a certain actual application scenario, after long-term operation, the photovoltaic module shows power attenuation, and there are no obvious abnormalities visible to the naked eye on the surface. Through the above data acquisition technology, the following detections are carried out:

[0047] Use spectral polarization data to identify hidden crack areas on the surface, and find that the abnormal spectral reflectivity points are associated with geometric characteristics; locate the hot spot area through thermal imaging data to verify the local temperature rise at the hidden crack points; combine the current imaging data to further confirm the resistance abnormal area, which is consistent with the thermal imaging data; find small surface warping through stereo depth data to further verify that the crack affects the mechanical stability of the module.

[0048] Finally, based on the above data, multi-modal feature data is generated to complete the defect assessment of the photovoltaic module, providing support for subsequent optimized operation and maintenance decisions.

[0049] Preferably, the time synchronization is based on the timestamp technology, and a unified time reference is used to mark the acquisition time of the spectral polarization data, thermal imaging data, current imaging data, and stereo depth data; the spatial calibration is performed through a feature point matching algorithm to align the abnormal areas in the thermal imaging data and the current imaging data in terms of position, and map the aligned result to the three-dimensional coordinate space of the stereo depth data to generate the preliminary multi-modal feature data after spatial calibration.

[0050] Time synchronization aligns the acquisition times of multi-modal data to a unified time reference to ensure the temporal consistency of different-modal data, thereby avoiding timing errors caused by acquisition delays or hardware clock drifts. The present invention assigns high-precision time stamps to the data acquired each time through the time stamp technology, enabling the multi-modal data to accurately correspond to the component states at the same time point.

[0051] A unified high-precision time server is adopted, and the internal clocks of each sensor (including spectral cameras, thermal imagers, current sensors, and light field sensors) are aligned with the time reference by using NTP (Network Time Protocol). During the data acquisition process, the acquisition time of each frame of data is recorded using time stamps. For example, the acquisition time of spectral data is marked as , the thermal imaging data is marked as , and by comparing the time difference , it is judged whether the synchronization condition is met. Through time synchronization, it is ensured that the spectral, thermal, current, and stereo depth data reflect the true state of the photovoltaic module at the same time point, providing accurate temporal correlation for subsequent fusion analysis.

[0052] Spatial calibration aligns the spatial coordinate systems of different-modal data to a unified three-dimensional coordinate system to ensure the spatial consistency of the data. The present invention adopts a feature point matching algorithm. First, the positions of abnormal regions in the thermal imaging data and current imaging data are matched, and then the alignment result is mapped to the three-dimensional coordinate space of the stereo depth data to generate spatially consistent multi-modal feature data.

[0053] Through spatial calibration, the data of different modalities are aligned in space, and the spatial correlations between hot spots, resistance anomalies, and three-dimensional geometric features of the photovoltaic module can be accurately identified, laying a spatial foundation for multi-modal feature fusion.

[0054] As Figure 2 shown, based on the preliminary multi-modal feature data, a spectral-geometry model and a multi-field coupling model are constructed to generate model data respectively; the generated model data is fused through a feature mapping algorithm and input into a generative adversarial network for feature abstraction to obtain key deep feature data;

[0055] The spectral-geometry model combines spectral polarization data and stereo depth data to comprehensively characterize surface cracks, coating defects, and geometric morphology anomalies of the photovoltaic module. The multi-field coupling model combines thermal imaging data and current imaging data to simulate the interaction between the internal stress field and the electric field of the module, thereby reflecting the internal characteristics of the module under operating conditions.

[0056] The temperature distribution in the thermal imaging data is combined with the resistance distribution in the current imaging data using tensor decomposition technology to generate a multi-field tensor. The finite element analysis method is used to calculate the interaction characteristics between the stress field and the electric field to generate multi-field coupling model data.

[0057] The spectral-geometric model can comprehensively characterize the optical properties and geometric morphological characteristics of the surface of the photovoltaic module, while the multi-field coupling model reveals the thermal-electric coupling relationship inside the module, providing comprehensive data support for defect detection.

[0058] The feature mapping algorithm fuses the spectral-geometric model data and the multi-field coupling model data by constructing a cross-mapping matrix between multi-modal features, thereby generating a unified multi-modal feature representation.

[0059] The generative adversarial network (GAN) consists of a generator and a discriminator. The generator optimizes to generate abstract feature data, and the discriminator is used to judge the quality of the generated features, thereby realizing the deep feature abstraction of the multi-modal fusion data.

[0060] The generative adversarial network performs deep abstraction on the multi-modal data to obtain key deep feature data, which can accurately reflect the defect characteristics of the photovoltaic module and significantly improve the accuracy of subsequent detection and evaluation.

[0061] Preferably, the feature mapping algorithm realizes the dynamic mapping of the spectral-geometric model data by calculating the correlation matrix between the spectral reflection characteristics and the geometric morphology. At the same time, through the multi-field tensor decomposition of the thermal imaging data and the current imaging data, the correlation features between the stress field and the electric field inside the module are extracted, and the correlation features are jointly modeled to generate the fused model data after feature mapping.

[0062] The spectral reflection characteristics reflect the optical changes of the surface coating, hidden cracks or scratches of the photovoltaic module, while the geometric morphology reveals the macroscopic characteristics such as surface deformation or warping. By calculating the correlation matrix between the spectral reflection characteristics and the geometric morphology, the coupling relationship between the spectral and geometric features can be dynamically mapped, thereby revealing the comprehensive performance of hidden cracks or surface defects. The specific implementation steps include:

[0063] Generate a spectral reflection matrix for the spectral polarization data , where is the two-dimensional coordinate, is the wavelength;

[0064] Use the stereo depth data to calculate the three-dimensional morphological coordinates of the module ;

[0065] Construct a correlation matrix between the spectral reflection characteristics and the geometric morphology :

[0066]

[0067] Among them, is the surface feature function of geometric points, reflecting the local changes in the surface morphology.

[0068] Map the correlation matrix to the spectral-geometric model data , and complete the dynamic mapping. Through the dynamic mapping, the spectral characteristics and geometric morphology are expressed uniformly, providing a comprehensive basis for analyzing the multimodal characteristics caused by surface cracks or warping of photovoltaic modules.

[0069] The thermal imaging data reflects the temperature distribution characteristics of the component surface, and the current imaging data reflects the internal resistance distribution characteristics. The multi-field tensor decomposition technology can extract the coupling relationship between these two modalities, revealing the interaction characteristics of heat and electric field, such as the combined effect of hot spots and current anomalies caused by local cracks. The specific implementation steps include:

[0070] Use the thermal imaging data and the current imaging data to construct the initial tensor :

[0071]

[0072] Perform principal component decomposition on the tensor to extract the principal eigenmatrix :

[0073]

[0074] Among them, and are the left and right singular value matrices of the tensor respectively, is the singular value diagonal matrix;

[0075] The principal eigenmatrix represents the heat-electric field interaction characteristics, and convert it into multi-field coupling model data . The multi-field tensor decomposition can accurately extract the common characteristics of hot spots and resistance anomalies, providing key data support for analyzing internal cracks or poor soldering.

[0076] Through the joint modeling technology of the feature mapping algorithm, the spectral-geometric model data and the multi-field coupling model data are comprehensively expressed to generate the fusion model data, which uniformly characterizes the surface and internal defect characteristics of the component.

[0077] Use the spectral-geometric model data and the multi-field coupling model data to construct the feature mapping matrix :

[0078]

[0079] Among them, represents a similarity calculation function, which is used to measure the correlation between two model data;

[0080] Generate fused model data according to the weights of the feature mapping matrix :

[0081]

[0082] The fused model data comprehensively integrates the surface features and internal characteristics of the photovoltaic module, and can uniformly analyze surface hidden cracks, coating defects and internal thermal-electric interaction characteristics, providing optimized input for subsequent deep feature abstraction.

[0083] In an embodiment, during the operation detection of a certain photovoltaic module, it is found that there are optical reflection anomalies on the surface, and at the same time, hot spots and resistance anomalies appear in local areas. The following operations are completed through the feature mapping algorithm:

[0084] Construct spectral-geometric model data , and analyze the surface hidden crack and warping areas;

[0085] Extract the thermal-electric interaction characteristics of the thermal imaging data and current imaging data through multi-field tensor decomposition to generate multi-field coupling model data ;

[0086] Jointly model the above two types of model data using the feature mapping matrix to generate fused model data ;

[0087] Analyze the fused model data, identify the key defect areas of the photovoltaic module, verify the interaction between surface hidden cracks and internal hot spots and resistance anomalies, and finally provide suggestions for repair solutions.

[0088] Preferably, the fused model data is input into a generative adversarial network, which includes a generator and a discriminator. The generator optimizes to generate deep abstract features through backpropagation, and the discriminator optimizes the feature data output by the generator by judging the difference between the generated features and the real features to generate enhanced key deep feature data.

[0089] The generative adversarial network (GAN) is a dual-network structure, where the generator is responsible for generating feature representations close to the real features, and the discriminator is responsible for distinguishing between the generated features and the real features, thereby improving the quality of the generated features through adversarial training.

[0090] In the present invention, the input of the GAN is the data of the fusion model, and the goal is to abstract key deep features from it to enhance the characterization ability of photovoltaic module defects. Through the collaborative training of the generator and the discriminator, the generator optimizes the accuracy of the generated features, and the discriminator optimizes the evaluation criteria for the generated features.

[0091] Generator design: Input the data of the fusion model , extract multi-level features through a convolutional neural network (CNN). Use the backpropagation algorithm to optimize the generated features , with the goal of making the generated features as close as possible to the real features . The objective function of the generated features is defined as:

[0092]

[0093] where is the loss function of the generator, is the quadratic norm, which is used to measure the difference between the generated features and the real features.

[0094] Discriminator design: The input is the generated features and the real features , and the discriminant network is used to evaluate the similarity between the two. The goal of the discriminator is to maximize the difference between the generated features and the real features, and the optimization objective function is:

[0095]

[0096] where is the loss function of the discriminator, represents the output of the discriminant network.

[0097] The generator and the discriminator are continuously improved through adversarial optimization: the generator aims to deceive the discriminator and generate data closer to the real features; the discriminator aims to distinguish between the real features and the generated features and strengthen the judgment criteria for the generated features. Finally, through multiple rounds of iteration, the quality of the generated features reaches the ideal standard, and key deep feature data is obtained .

[0098] Adopt a dynamic learning rate adjustment strategy to optimize the learning efficiency of the generator and the discriminator according to the gradient changes during the training process; introduce the Batch Normalization technology to stabilize the training process and prevent the mode collapse phenomenon.

[0099] Through the generative adversarial network, it is possible to abstract deep multi-modal joint features from the fused model data. These features have higher semantic expression capabilities and can accurately depict the complex patterns of surface and internal defects of photovoltaic modules. The deep feature data optimized by the generator removes the possible noise and redundant information in the multi-modal data and strengthens the expression of the key characteristics of the defects. The evaluation criteria optimized by the discriminator improve the authenticity and reliability of the generated features and provide higher-quality input data for subsequent defect detection and trend prediction.

[0100] Example, in a quality inspection task of a certain photovoltaic module, through the collected fused model data M_fusion , perform the following steps:

[0101] Input into the generator, extract multi-level features through the convolutional network, and initially generate the feature representation M_generated ;

[0102] Input and the real features into the discriminator together. The discriminator evaluates the quality of the generated features and returns the loss value ;

[0103] Optimize the parameters of the generator and the discriminator through the backpropagation algorithm to make the generated features gradually approach the real features;

[0104] After multiple rounds of training, generate the final key deep feature data , which accurately characterizes the surface hidden cracks, coating defects, and internal thermal-electric field interaction characteristics of the module;

[0105] Utilize the data to successfully identify the correlation between a hidden crack on the surface of the module and a local hot spot inside, providing a direct basis for subsequent maintenance and optimization.

[0106] As Figure 3 shown, input the key deep feature data into the hierarchical attention mechanism for global and local feature fusion to generate fused defect feature data; use the region detection algorithm to locate the defect regions of the fused defect feature data, and classify and evaluate the severity of the defect regions to generate defect classification data and defect evaluation data;

[0107] The Hierarchical Attention Mechanism (HAM) is a feature fusion technology based on deep learning. The global attention module focuses on the overall feature pattern of the photovoltaic module for detecting macroscopic anomalies such as large-scale coating wear or deformation; the local attention module focuses on the detailed features of the local anomaly area to capture tiny defects such as hidden cracks and hot spots. The two are fused through a hierarchical weighting method to achieve complementary expression of global and local information.

[0108] Through the hierarchical attention mechanism, global features and local features can be efficiently fused, making the defect feature expression more comprehensive and accurate, and significantly improving the accuracy of subsequent defect area localization and classification.

[0109] The region detection algorithm divides the fused defect feature data through a Convolutional Neural Network (CNN) to identify the regions where defects may exist, determines the specific positions of the defects through the bounding box localization technology, and at the same time uses the classification module to evaluate the category and severity of each defect area.

[0110] The region detection algorithm can not only accurately locate the positions of the defect areas, but also comprehensively evaluate the defect types and severities, providing a reliable basis for the fault diagnosis of photovoltaic modules.

[0111] Preferably, the hierarchical attention mechanism performs weighted processing on the overall spectral distribution, thermal distribution, and current characteristics of the photovoltaic module through the global feature extraction module, and at the same time focuses on the spectral anomaly points, hot spot positions, and current anomaly points in the defect area through the local feature extraction module to achieve efficient fusion of global and local features and generate fused defect feature data.

[0112] The Hierarchical Attention Mechanism (HAM) is a deep learning method that combines global and local features. The global feature extraction module is responsible for capturing the overall characteristic distribution patterns of the photovoltaic module, such as spectral reflection uniformity, temperature gradient, etc.; the local feature extraction module focuses on the anomaly points in specific regions, such as spectral anomalies caused by hidden cracks, high-temperature regions of hot spots, current anomaly points, etc. Through the hierarchical structure, the global and local features are fused, and HAM can generate more accurate defect feature data.

[0113] The global feature extraction module uses the Self-Attention Mechanism to perform weighted processing on the spectral, thermal, and current data of the overall photovoltaic module to capture the distribution patterns of the overall characteristics of the module.

[0114] Input key deep feature data , and generate query vectors through linear transformation , key vector , value vector :

[0115]

[0116]

[0117]

[0118] in, , , is the trainable weight matrix.

[0119] Calculate the attention weights:

[0120]

[0121] in, is the dimension of the key vector.

[0122] Weighted global features :

[0123]

[0124] The local feature extraction module uses sliding window technology to focus on the defective area of ​​the component and extract high-dimensional features of spectral anomalies, hot spot locations, and current anomalies. Fixed window size Divide into several sub-areas , calculate the local weighted features for each sub-region:

[0125]

[0126] in, is the local weight matrix, It is a bias term. It focuses on the abnormal point features in the defect area and enhances the feature expression through high weight.

[0127] Generate fused defect feature data by weighted fusion of global and local features , which comprehensively expresses the overall characteristics and local abnormalities of the component. Perform weighted summation of global features and local features:

[0128]

[0129] in, , dynamically adjust the weights to adapt to different detection requirements.

[0130] Through the hierarchical attention mechanism, global features can capture the overall optical, electrical, and thermodynamic property distributions of components, identifying macroscopic abnormal regions; local features can precisely capture the detailed characteristics of defect regions, such as the spectral changes of hidden cracks, the high-temperature characteristics of hot spots, and the local interruptions of current paths. The finally generated fused defect feature data has higher expressive power, providing optimized input for subsequent defect localization, classification, and severity assessment.

[0131] Example, in a photovoltaic module detection task, the following steps are completed through the hierarchical attention mechanism:

[0132] Input the spectral, thermal imaging, and current imaging data of the module into the global feature extraction module to obtain the global spectral distribution, temperature gradient, and current path characteristics. The results show that there is a large range of deviation in the spectral reflectance in the lower left corner of the module, and at the same time, the temperature in this area is slightly higher than the normal value.

[0133] Perform local feature extraction on the lower left corner area of the module, focusing on the spectral abnormal points of hidden cracks and the high-temperature positions of hot spots, and extract multiple groups of local high-weight features.

[0134] Perform weighted fusion of the global features and local features to generate fused defect feature data. Analysis finds that the spectral changes in the hidden crack area are highly correlated with the hot spot positions, which are the key defect areas.

[0135] Through subsequent area detection and evaluation confirmation, the hidden crack in the lower left corner of the photovoltaic module has caused local overheating, and it is recommended to replace the module or strengthen monitoring.

[0136] Preferably, the area detection algorithm performs multi-level feature extraction on the fused defect feature data. First, it roughly determines the areas with defects through the edge detection algorithm, and then uses the improved area extraction algorithm to perform fine-grained boundary optimization on the defect areas, and finally outputs the precise position coordinates and area boundary information of the defect areas.

[0137] Multi-level feature extraction is a region detection method that refines features layer by layer. First, it quickly identifies the areas that may have defects in the photovoltaic module through the edge detection algorithm, roughly determining the positions and ranges of these areas; then it uses the improved area extraction algorithm to perform refined analysis on these areas, optimizing the boundary information and outputting the precise defect area coordinates. This method combines the advantages of rough positioning and fine optimization, and can greatly improve the detection accuracy while ensuring the detection efficiency.

[0138] The edge detection algorithm identifies the edges of the areas with abnormal spectral, thermal, or current characteristics in the photovoltaic module by calculating the gray level change or gradient change of the image. Input the fused defect feature data into the edge detection module and calculate the gradient intensity of each pixel point :

[0139]

[0140] Among them, is the pixel intensity value, and are the gradients in the horizontal and vertical directions.

[0141] Set a threshold , and mark the pixels with gradient intensity greater than as edge points to generate a preliminary defect area mask . Quickly determine the area where defects may exist in the photovoltaic module, providing input for subsequent refined processing.

[0142] The improved region extraction algorithm refines the boundary information of the rough region by optimizing the defect region boundary to ensure the accuracy of region positioning. Use the region growing method to optimize the preliminary mask , and dynamically adjust the region boundary according to the spectral reflectance, temperature distribution, and current path characteristics within the region;

[0143] Calculate the boundary point set , and the optimized boundary is updated by the following formula:

[0144]

[0145] Among them, is the feature distribution within the region, is the feature distribution outside the region.

[0146] Map the optimized boundary information back to the original image, and output the accurate defect region coordinates and boundary information . The improved region extraction algorithm can refine the boundary of the rough region to ensure the accuracy of defect region positioning, and is particularly suitable for the detection of complex defects in photovoltaic modules.

[0147] Through the multi-level feature extraction method, the edge detection algorithm can quickly identify the location and scope of the defect region, saving computational costs for subsequent refinement processing; the improved region extraction algorithm accurately optimizes the boundary of the rough region, effectively improving the positioning accuracy of the defect region. The finally output defect region coordinates and boundary information provide reliable data support for defect classification and severity assessment.

[0148] Example, in the component detection task of a certain photovoltaic power station, identify possible hidden cracks and hot spots in the components through the region detection algorithm, and perform the following operations:

[0149] Input the fused defect feature data into the edge detection module to calculate the gradient intensity The result shows that there is a high-gradient area with obvious edges in the middle of the component, and this area is marked as the preliminary defect area.

[0150] The preliminary mask map The input area extraction module optimizes the boundary, eliminates some redundant boundary points, and at the same time refines the boundary range of the real defect. Finally, an accurate boundary of the defect area is generated. 。

[0151] The coordinates and boundary information of the output defect area show that there are significant anomalies in the spectral reflectivity, temperature distribution, and current path in this area, confirming the coexistence of hidden cracks and hot spots, which has a great impact on the power output of the component.

[0152] As Figure 4 shown, the model data is loaded into the digital twin platform to simulate the operating conditions and generate virtual detection data; perform time series analysis on the virtual detection data, predict the expansion trend of the defect, and optimize the detection system parameters.

[0153] Digital twin technology reproduces the operating state and behavior of physical objects in a virtual environment through the dynamic mapping of virtual models and actual physical systems. In the present invention, the spectral-geometric model data and multi-field coupling model data are loaded into the digital twin platform to construct a virtual model of the photovoltaic module, and the actual operating conditions are simulated based on the input environmental and operating parameters (such as temperature, humidity, light intensity, etc.). The digital twin platform can reflect the operating characteristics of the photovoltaic module under different operating conditions in real time, provide high-precision defect evolution data, and lay a foundation for subsequent analysis.

[0154] Time series analysis is a method of mining change laws based on time series data. By analyzing the dynamic changes of the defect area in the virtual detection data over time, the expansion trend of the defect is predicted. In the present invention, the virtual detection data is processed through a time series coupling prediction network (TCP-Net) to obtain the future development path and expansion rate of the defect. Time series analysis can accurately predict the expansion trend of the defect, provide forward-looking data support, and help formulate preventive maintenance strategies.

[0155] Based on the prediction results of the defect expansion trend, adjust the key parameters of the detection system (such as the resolution of the acquisition device, the hyperparameters of the feature extraction algorithm, etc.), and optimize the system performance to adapt to different operating environments and defect characteristics. The optimized detection system can more efficiently adapt to complex operating environments and give early warnings before defects occur, reducing the risk of component failure.

[0156] Preferably, the time series analysis includes performing time series decomposition on the virtual detection data, extracting the dynamic change patterns of the defect area, including the hot spot expansion rate, the resistance change rate, and the spectral reflectivity fluctuation, and extrapolating the defect expansion trend based on the time series coupling prediction network to generate defect trend prediction data.

[0157] Temporal analysis aims to reveal the evolution pattern of defect area characteristics over time by mining the dynamic changes in virtual detection data. The present invention uses the time series decomposition method to decompose the multi-modal data of the defect area (such as hot spot temperature, resistance change, and spectral reflectance fluctuation) into three parts: long-term trend, periodic fluctuation, and random fluctuation, and extracts the dynamic change patterns of each feature. Subsequently, a Temporal Coupling Prediction Network (TCP-Net) is used to model the extracted temporal features to predict the defect expansion trend.

[0158] Time series decomposition decomposes the time series data into a long-term trend , periodic fluctuation and random fluctuation :

[0159]

[0160] This decomposition can help identify the long-term expansion trend and short-term fluctuation characteristics of defects.

[0161] Sample the features of each defect area in the virtual detection data to generate a time series feature matrix:

[0162]

[0163] where , , respectively represent the changes in spectral reflectance, hot spot temperature, and resistance value over time; use the additive model for decomposition to extract the long-term trend and periodic fluctuation , and filter the random fluctuation through a denoising method .

[0164] The hot spot expansion rate is calculated by analyzing the change in the hot spot area over time :

[0165]

[0166] where, is the change amount of the hot spot area, is the time interval.

[0167] The resistance change rate is calculated by analyzing the change in the resistance value :

[0168] where, is the change in resistance. The spectral reflectance fluctuates, and the standard deviation of the spectral reflectance in the defective area is calculated , and its fluctuation intensity is quantified.

[0169] Using the Temporal Coupling Prediction Network (TCP-Net), combined with multi-modal temporal features, predict the expansion trend of future defective areas. TCP-Net extracts the spatial correlation of temporal features through a convolutional network and models the temporal correlation through a Recurrent Neural Network (RNN).

[0170] Input into TCP-Net, and calculate the eigenvalue at the future time point :

[0171]

[0172] where is the predicted future feature matrix. The prediction results include the expansion range of the hot spot area, the resistance change trend, and the spectral reflectance fluctuation intensity.

[0173] Through temporal analysis, the dynamic change patterns of the defective area can be accurately extracted, such as the hot spot expansion rate, the resistance change rate, and the spectral reflectance fluctuation, so as to identify the development trend and potential risks of the defect. Combined with the Temporal Coupling Prediction Network, it can accurately predict the future evolution of the defect and provide forward-looking maintenance suggestions.

[0174] Example, in the component detection of a certain photovoltaic power station, the following operations are realized through temporal analysis and prediction:

[0175] Collect virtual detection data , perform time series decomposition on the spectral reflectance, hot spot temperature, and resistance value of the defective area, and extract the long-term trend and periodic fluctuation. The results show that the temperature in the hot spot area shows an upward trend, the resistance value gradually increases, and the spectral reflectance fluctuation intensifies.

[0176] Calculate the hot spot expansion rate , the resistance change rate , and the standard deviation of the spectral reflectance is .

[0177] Use TCP-Net to predict that the area of the hot spot area will increase by 50%, the resistance value will increase by 20%, and the spectral reflectance fluctuation amplitude will increase by 10% within the next 4 hours.

[0178] According to the prediction results, it is recommended to cool down the hot spot area and further inspect the relevant components to prevent the further expansion of the hidden crack.

[0179] Preferably, the optimized detection system parameters include adjusting the acquisition frequency of the spectral imaging device, dynamically adjusting the temperature sensitivity threshold of the thermal imaging device, optimizing the sampling interval of the current imaging system, and updating the learning rate of the generative adversarial network, so as to meet the requirements of photovoltaic module defect detection in different operating environments.

[0180] The acquisition frequency of the spectral imaging device determines the timeliness and fineness of data acquisition. In dynamic working conditions (such as when photovoltaic modules are under strong light or rapid shadow changes), a higher acquisition frequency can capture the rapidly changing spectral reflection characteristics, while in stable working conditions, the frequency can be appropriately reduced to reduce data redundancy and processing burden.

[0181] Dynamically adjust the acquisition frequency according to the operating environment of the photovoltaic module :

[0182]

[0183] Among them, is the threshold of the light change rate. Dynamically adjusting the acquisition frequency can balance the timeliness and efficiency of data acquisition, ensure capturing the spectral changes of photovoltaic modules in a dynamic environment, and at the same time reduce data storage and processing costs.

[0184] The temperature sensitivity threshold of the thermal imaging device determines the device's detection ability to temperature difference changes. For weak temperature rises caused by small hot spots or hidden cracks, higher sensitivity is required; while in an environment with large temperature fluctuations, to avoid false alarms, the sensitivity can be appropriately reduced.

[0185] Dynamically adjust the temperature sensitivity threshold :

[0186]

[0187] Among them, is the threshold of the environmental temperature change rate. By dynamically adjusting the sensitivity, tiny hot spots can be captured more accurately, while avoiding false alarms caused by environmental temperature fluctuations, and improving the accuracy of thermal imaging detection.

[0188] The sampling interval of the current imaging system determines the time resolution of the current distribution characteristics. In the case of rapid current changes caused by hidden cracks or poor soldering, a shorter sampling interval is required to capture transient signals; while under steady-state conditions, a longer sampling interval can reduce redundant data.

[0189] Optimize the sampling interval :

[0190]

[0191] Among them, is the threshold value of the current change rate. The optimization of the sampling interval ensures capturing key features under dynamic current change conditions while improving the efficiency and accuracy of data acquisition.

[0192] The learning rate of the Generative Adversarial Network (GAN) directly affects the optimization speed and convergence of the feature abstraction process. A higher learning rate can accelerate the training speed of the network but may cause model oscillation; a lower learning rate helps the model converge stably but with a slower training speed. The present invention dynamically updates the learning rate according to the training error.

[0193] Based on the learning rate adjustment strategy (Learning Rate Scheduler, LRS), the learning rate is dynamically updated :

[0194] where is the initial learning rate, is the learning rate decay coefficient, is the number of training steps. The dynamic learning rate adjustment can balance the training speed and stability, ensuring that the Generative Adversarial Network efficiently extracts key deep features and improving the robustness of defect detection.

[0195] By optimizing the key parameters of the detection system, the present invention can flexibly adapt to different operating environments, improve the accuracy of data acquisition and feature extraction, and enhance the detection efficiency and precision. Dynamically adjusting the spectral imaging frequency, thermal imaging sensitivity, current sampling interval, and GAN learning rate enables the system to respond in real time to changes in the operating state of components, optimize the defect detection performance, and reduce the false alarm rate and missed alarm rate.

[0196] In an embodiment, in the dynamic condition detection task of a certain photovoltaic module, the system performs the following optimization operations:

[0197] When the light intensity fluctuates greatly, the system increases the acquisition frequency from 30 frames per second to 50 frames per second to ensure capturing details of spectral changes; under stable light conditions, the frequency is reduced to 20 frames per second to reduce redundant data.

[0198] For a weak hot spot appearing in a certain component, the system optimizes the temperature sensitivity threshold from 0.1 °C to 0.05 °C and successfully captures the early features of hot spot expansion.

[0199] Under the condition of rapid resistance change in the crack area, the system shortens the sampling interval from 1 second to 0.5 second to accurately record the current fluctuation characteristics.

[0200] In the feature abstraction process, the initial learning rate is 0.01 and gradually decays to 0.001 as the number of training steps increases, ensuring stable convergence of the model while training efficiently. The generated key deep feature data significantly improves the accuracy of defect detection.

[0201] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0202] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A photovoltaic module quality detection method based on artificial intelligence, characterized in that: The following steps are involved: Collect spectral polarization data, thermal imaging data, current imaging data and stereo depth data of photovoltaic modules, perform time synchronization and spatial calibration on all data, and generate preliminary multi-modal feature data; Based on the preliminary multimodal feature data, a spectral-geometric model characterizing the surface optical characteristics and geometric morphological characteristics of the photovoltaic module and a multi-field coupling model characterizing the internal thermal-electric coupling relationship of the photovoltaic module are constructed to generate model data respectively; The generated model data is fused through the feature mapping algorithm and input into the generative adversarial network for feature abstraction to obtain key deep feature data; The feature mapping algorithm realizes dynamic mapping of spectral-geometric model data by calculating the correlation matrix between spectral reflectance characteristics and geometric morphology, and extracts the correlation features between the internal stress field and the electric field of the component by performing multi-field tensor decomposition on the thermal imaging data and the current imaging data, generates multi-field coupling model data, and jointly models the correlation features to generate fusion model data after feature mapping; Inputting the key deep feature data into a hierarchical attention mechanism to perform global and local feature fusion to generate fused defect feature data; The defective area of ​​the fused defect feature data is located by using a regional detection algorithm, and the defective area is classified and the severity is evaluated to generate defect classification data and defect evaluation data, wherein the hierarchical attention mechanism performs weighted processing on the overall spectral distribution, thermal distribution and current characteristics of the photovoltaic module through a global feature extraction module, and at the same time focuses on the spectral anomaly points, hot spot positions and current anomaly points of the defective area through a local feature extraction module, so as to achieve the fusion of global features and local features and generate fused defect feature data; the regional detection algorithm performs multi-level feature extraction on the fused defect feature data, firstly roughly determines the defective area through an edge detection algorithm, and then performs fine-grained boundary optimization on the defective area through an improved regional extraction algorithm, and finally outputs the precise position coordinates and regional boundary information of the defective area; The model data is loaded into the digital twin platform, the operating conditions are simulated, and virtual detection data is generated; a time series analysis is performed on the virtual detection data to predict the expansion trend of defects and optimize the detection system parameters.

2. The method according to claim 1, characterized in that The spectral polarization data is collected by adjusting the polarization angle to capture the multi-angle spectral characteristics of the surface of the photovoltaic module, the thermal imaging data is collected by recording the temperature distribution characteristics of the surface of the photovoltaic module, the current imaging data is obtained by applying a small current to the photovoltaic module and synchronously recording the current distribution characteristics, and the stereo depth data is generated by multi-angle ray tracing to generate three-dimensional morphological characteristics of the component surface.

3. The method according to claim 2, characterized in that The time synchronization is based on the time stamp technology, and a unified time reference is used to mark the acquisition time of the spectral polarization data, thermal imaging data, current imaging data and stereo depth data; The spatial calibration aligns the thermal imaging data with the abnormal areas in the current imaging data through a feature point matching algorithm, and maps the aligned result to the three-dimensional coordinate space of the stereo depth data to generate preliminary multimodal feature data after spatial calibration.

4. The method according to claim 1, characterized in that: The fusion model data is input into a generative adversarial network, which includes a generator and a discriminator. The generator generates deep abstract features through back-propagation optimization, and the discriminator optimizes the feature data output by the generator by judging the difference between the generated features and the real features, thereby generating enhanced key deep feature data.

5. The method according to claim 1, characterized in that: The timing analysis includes time series decomposition of virtual detection data, extraction of dynamic change patterns of defect areas, including hot spot expansion rate, resistance change rate and spectral reflectivity fluctuations, and extrapolation of defect expansion trends based on a timing coupled prediction network to generate defect trend prediction data.

6. The method according to claim 1, characterized in that The optimization of detection system parameters includes adjusting the acquisition frequency of the spectral imaging device, dynamically adjusting the temperature sensitivity threshold of the thermal imaging device, optimizing the sampling interval of the current imaging system, and updating the learning rate of the generative adversarial network, so as to adapt to the photovoltaic module defect detection needs under different operating environments.

Images