A detection analysis system and method for perovskite solar cells

Abstract

This invention discloses a detection and analysis system and method for perovskite solar cells, belonging to the field of perovskite solar cell detection technology. In the system, a transmission unit transports the solar cell substrate; an electroluminescence detection unit applies a preset voltage to the substrate to excite an electroluminescence signal; an image acquisition unit acquires the electroluminescence image of the substrate and transmits it to an AI processing unit; the AI ​​processing unit sequentially performs image preprocessing, feature extraction, multimodal fusion inference, and defect classification processing, generating a defect judgment result and transmitting it to a module sorting unit, while simultaneously outputting defect statistics; the module sorting unit performs sorting operations on the perovskite solar cell substrates based on the defect judgment result, transferring qualified products to a post-processing unit via the transmission unit. This invention improves detection accuracy and production efficiency through automated detection and defect identification, multimodal fusion analysis, and sorting.

Description

Technical Field

[0001] This invention relates to the field of perovskite solar cell detection technology, specifically to a detection and analysis system and method for perovskite solar cells. Background Technology

[0002] Perovskite solar cells, as an emerging solar cell technology, have attracted widespread attention due to their advantages such as high efficiency, low cost, and environmental friendliness. However, during the production process, various visual defects may occur due to the influence of processes and equipment, which can affect the performance and lifespan of the cells.

[0003] As an improvement, to promptly detect and repair these defects, the EL (Electroluminescent) tester used in perovskite solar cell visual defect inspection instruments has become an indispensable tool. The EL tester is a device that uses the principle of electroluminescence to detect internal defects in solar cells. It applies a certain voltage to the solar cell, exciting electrons inside the cell, causing the cell to emit a light signal, thus achieving visual detection of internal defects. EL defect testing requires a specialized EL tester, which should possess characteristics such as high precision, high stability, high resolution, and fast response capabilities. Simultaneously, to ensure the accuracy of test results, the testing equipment must also have multi-wavelength excitation capabilities to adapt to the detection needs of different types of perovskite solar cells. Firstly, the stability of the test environment must be ensured when conducting EL defect testing. Factors such as temperature and humidity in the test environment will affect the test results. Therefore, before testing, the test environment needs to be adequately preheated and stabilized. Then, appropriate test parameters, such as voltage, current, and exposure time, must be selected. These parameters need to be set according to the specific type and structure of the perovskite solar cell. The perovskite solar cell is placed on the test stage of an EL (electroluminescence) meter, and a certain voltage is applied. Electroluminescence (EM) is then generated inside the cell. These EEM signals are captured and recorded by the EL meter's camera. The captured EEM signals are then digitally processed and analyzed. Using specialized image processing software, the types and distribution of defects in the perovskite material can be visually observed.

[0004] However, due to the limitations of standard imaging techniques, it is impossible to distinguish whether darker areas are caused by a large trap site or many smaller traps, making it difficult to determine why they form only in certain areas. Because perovskite solar cells inherently have low repeatability, inconsistencies in defects and color depth are difficult to identify, and simply determining the source of defects is challenging. Furthermore, inconsistent on-voltage across each cell necessitates adjustments for each test, requiring analysis of glass blemishes, cell voids, shallow low-brightness areas, and internal contaminants. In summary, current perovskite solar cell testing and analysis suffers from limitations in standard imaging techniques. It struggles to distinguish the causes of dark areas and pinpoint the location and contributing factors to defect formation; poor cell consistency makes defect identification relatively difficult, hindering the accurate differentiation of different defect types. Summary of the Invention

[0005] This invention provides a detection and analysis system and method for perovskite solar cells, aiming to solve the problems in the current detection and analysis of perovskite solar cells, such as the limitations of standard imaging technology, difficulty in distinguishing the causes of dark areas, difficulty in locating the location and causes of defects, and the difficulty in accurately distinguishing different types of defects due to poor cell consistency.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a detection and analysis system for perovskite solar cells, comprising a device transmission unit, an image acquisition unit, an electroluminescence detection unit, an AI processing unit, a component sorting unit, and a post-processing unit; wherein: The equipment transmission unit is used to transmit perovskite solar cell substrates; the electroluminescence detection unit is set at the detection station of the equipment transmission unit to apply a preset voltage to the perovskite solar cell substrate in order to excite electroluminescence signals; the image acquisition unit is installed at the detection station of the electroluminescence detection unit to acquire electroluminescence images of the perovskite solar cell substrate and transmit the electroluminescence images to the AI ​​processing unit. The AI ​​processing unit sequentially performs image preprocessing, feature extraction, multimodal fusion inference, and defect classification on the electroluminescent image, generates defect judgment results, and transmits them to the component sorting unit, while outputting defect statistics. The component sorting unit performs sorting operations on the perovskite solar cell substrate based on the defect judgment results, removes unqualified products, and transmits qualified products to the post-processing unit through the equipment transmission unit.

[0007] In some implementations, the AI ​​processing unit includes a dynamic reference correction module, which is used to construct a dynamic reference light intensity curve. The calculation formula for the dynamic reference light intensity curve is as follows: ; in, The set of electroluminescence intensities of all pixels within the sliding window. The average grayscale value of the pixels within the sliding window. This is the statistical function for the 95th percentile. To calculate only the robust local standard deviation of the sample between the median and the maximum, This is the compensation coefficient.

[0008] Furthermore, the dynamic baseline correction module is also used to calculate the defect severity index, and the formula for calculating the defect severity index is as follows: ; in, The measured light intensity per pixel. α is the noise floor constant, and α is the sensitivity gain coefficient.

[0009] In some implementations, the AI ​​processing unit includes a dual-branch fusion convolutional neural network module, which includes an image feature extraction branch, a process parameter fusion branch, and a fusion and classification head.

[0010] Furthermore, the input to the image feature extraction branch is a two-channel tensor containing the original electroluminescent grayscale image and the defect severity index image. The backbone network of the image feature extraction branch is an improved ResNet-50 network. Each stage of the improved ResNet-50 network embeds a CBAM attention module. The output of the image feature extraction branch is the image feature vector obtained by global average pooling.

[0011] Furthermore, the process parameter fusion branch is a multilayer perceptron structure. The input of the process parameter fusion branch is a process parameter vector containing coating speed, annealing temperature, ambient humidity and open circuit voltage. The multilayer perceptron includes two fully connected layers and the activation function is ReLU. The output of the process parameter fusion branch is a process feature vector.

[0012] Furthermore, the fusion and classification head includes a feature stitching unit, a Dropout layer, and a fully connected output layer. The feature stitching unit is used to stitch together image feature vectors and process feature vectors, and the fully connected output layer outputs the defect category probability distribution through the Softmax function.

[0013] In some implementations, the AI ​​processing unit uses the Focal Loss function for model training. The formula for calculating the Focal Loss function is as follows: ; in, To predict the probability of the target class for the model. To focus parameters, This is the balance coefficient.

[0014] In some implementations, the AI ​​processing unit further includes a defect evolution prediction module and a federated learning collaboration module; the defect evolution prediction module is used to collect electroluminescence image sequences of perovskite solar cell substrates at different aging stages and predict the defect expansion trend through a time series model; the federated learning collaboration module is used to aggregate local model parameter updates from multiple production sites and generate a globally optimal model.

[0015] This invention also provides a method for detecting and analyzing perovskite solar cells, which is based on the above-mentioned detection and analysis system for perovskite solar cells, and includes the following steps: The equipment transmission unit carries and transports the perovskite solar cell substrate, delivering it to the corresponding detection station of the electroluminescence detection unit. The electroluminescence detection unit applies a preset voltage to the perovskite solar cell substrate, causing the perovskite solar cell substrate to emit an electroluminescence signal; The image acquisition unit acquires electroluminescence images of the perovskite solar cell substrate and transmits the electroluminescence images to the AI ​​processing unit in the form of electrical signals; The AI ​​processing unit sequentially performs image preprocessing, feature extraction, multimodal fusion inference, and defect classification on the electroluminescent image, generates defect judgment results, transmits the defect judgment results to the component sorting unit, and outputs defect statistics. The component sorting unit performs sorting operations on perovskite solar cell substrates based on the defect judgment results, rejects unqualified products, and transfers qualified products to the post-processing unit through the equipment transfer unit.

[0016] Compared with the prior art, the detection and analysis system and method for perovskite solar cells of the present invention have the following advantages: This invention discloses a detection and analysis system for perovskite solar cells. By establishing a detection system encompassing perovskite solar cell substrate transmission, electroluminescence signal excitation, electroluminescence image acquisition, intelligent image analysis, defect determination, automatic sorting, and post-processing of qualified products, it improves upon the existing decentralized and highly manual operation mode of detection, achieving automated and precise operation for perovskite solar cell defect detection. The electroluminescence detection unit can stably apply a preset voltage at a fixed detection station to excite the electroluminescence signal, working in conjunction with the image acquisition unit to complete accurate image acquisition and transmission, ensuring the stability and reliability of the detection data and avoiding detection deviations caused by manual placement and parameter adjustments. AI processing is also included. The unit can independently complete image preprocessing, feature extraction, multimodal fusion inference, and defect classification, outputting relatively accurate defect judgment results. It can replace manual visual inspection, improve the subjectivity and fatigue error of manual judgment, and enhance the consistency and accuracy of defect judgment. The component sorting unit automatically performs non-conforming product rejection and qualified product conveying operations based on the defect judgment results, realizing synchronous linkage between detection and sorting, improving the efficiency of detection and sorting. The synchronous output of defect statistics can provide direct data support for production process optimization, improve the limitations of existing EL detection imaging, fuzzy defect discrimination, and cumbersome detection process, and make defect detection of perovskite solar cells more efficient and accurate. Attached Figure Description

[0017] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a schematic diagram of the overall architecture of a detection and analysis system for a perovskite solar cell according to the present invention.

[0019] The system includes: 1. Equipment transmission unit; 2. Image acquisition unit; 3. Electroluminescence detection unit; 4. AI processing unit; 5. Component sorting unit; and 6. Post-processing unit. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0023] It should be noted that the apparatus and methods disclosed in the embodiments herein can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments herein. In this regard, each block in a flowchart or block diagram may represent a module, program, or part of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system to perform the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0024] In addition, the functional modules in the various embodiments of this article can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0025] like Figure 1 As shown, the present invention discloses a detection and analysis system for perovskite solar cells, comprising a device transmission unit 1, an image acquisition unit 2, an electroluminescence detection unit 3, an AI processing unit 4, a component sorting unit 5, and a post-processing unit 6; wherein: The equipment transmission unit 1 is used to transmit the perovskite solar cell substrate; the electroluminescence detection unit 3 is set at the detection station of the equipment transmission unit 1 and is used to apply a preset voltage to the perovskite solar cell substrate in order to excite the electroluminescence signal; the image acquisition unit 2 is installed at the detection station of the electroluminescence detection unit 3 and is used to acquire the electroluminescence image of the perovskite solar cell substrate and transmit the electroluminescence image to the AI ​​processing unit 4. AI processing unit 4 is used to sequentially perform image preprocessing, feature extraction, multimodal fusion inference and defect classification processing on electroluminescent images, generate defect judgment results and transmit them to component sorting unit 5, and output defect statistics. Component sorting unit 5 is used to perform sorting operations on perovskite solar cell substrates according to defect judgment results, remove unqualified products and transmit qualified products to post-processing unit 6 through equipment transmission unit 1.

[0026] This invention effectively eliminates false positives caused by uneven film thickness by introducing a dynamic benchmark correction algorithm, significantly improving the detection rate of minute defects such as microcracks and shallow traps. By combining image features with process parameters, it not only determines the presence of defects but also assists in analyzing their causes. An optimized CNN architecture supports online inference of hundreds of images per second, meeting the cycle time requirements of high-speed production lines. This invention can replace manual visual inspection, eliminating fatigue and subjective errors and ensuring the consistency of quality inspection standards. It supports online incremental learning and federated updates, adapting to the emergence of new materials and unknown defect types. Furthermore, this invention possesses defect evolution analysis capabilities, providing data support for product reliability assessment and lifespan assurance.

[0027] In some operating conditions, the perovskite solar cell substrate enters the inspection station via equipment transfer unit 1. Electroluminescence detection unit 3 has an electroluminescence device that applies a specific voltage. Image acquisition unit 2 uses a high-resolution camera to acquire EL images. AI processing unit 4 performs dynamic benchmark correction, feature extraction, and multimodal fusion inference; based on the judgment results, it controls the sorting mechanism of component sorting unit 5, rejecting defective products and allowing qualified products to flow into the next process; simultaneously, it provides feedback to the process end on defect statistics and optimization suggestions.

[0028] In this invention, to address the pseudo-defect problem caused by uneven thickness of perovskite thin films, a dynamic benchmark correction algorithm based on local statistics is introduced to construct a defect feature map decoupled from thickness. For any pixel in the EL image, a sliding window is constructed centered on that pixel. A dynamic benchmark light intensity curve is constructed, and the theoretical optimal luminous intensity of the local region is estimated using the gray-level distribution of pixels within the window. ; in, For sliding windows The set of EL intensities for all pixels within the range. The average grayscale value of the pixels within the window is used as a proxy variable for the local film thickness. This is a high percentile statistical function, taking the value at the 95th percentile of the set, representing the upper limit of normal luminescence at this thickness. To ensure robust local standard deviation, only the sample standard deviation between the median and the maximum value is calculated to exclude the impact of defective points on the benchmark. This is the compensation coefficient, with a value ranging from 2.0. Version 3.0 is used to cover normal statistical fluctuations.

[0029] In this invention, the Defect Severity Index is calculated, and a normalized defect enhancement channel is generated. As an additional input to the neural network: ; in, The measured light intensity is for each pixel. The noise basis constant (e.g., 1) e 5) Prevent the denominator from being zero. This is the sensitivity gain coefficient, with a value of 5.0. Version 10.0 amplifies weak defect signals. It converts absolute light intensity differences into relative deviations, effectively eliminating background interference caused by film thickness variations and allowing the network to focus on identifying abnormal attenuation regions.

[0030] The dual-branch fusion convolutional neural network, employing a Dual-Branch Fusion CNN architecture, fuses image visual features with production process parameters to achieve accurate multimodal classification. The dual-branch fusion convolutional neural network module includes an image feature extraction branch, a process parameter fusion branch, and a fusion and classification head.

[0031] In image feature extraction, the dual-channel tensor contains the original EL grayscale image and the defect severity index image. The backbone network uses an improved ResNet-50. The ResNet-50 contains four stages, each composed of stacked Bottleneck residual blocks. A Convolutional Block Attention Module (CBAM) is embedded at the end of each stage to adaptively weight the features along the channel and spatial dimensions, enhancing the feature responses of key defect areas such as microcracks and holes, and suppressing background noise. After global average pooling (GAP), a 512-dimensional image feature vector is obtained.

[0032] In the process parameter fusion branch, the process parameter vector includes coating speed, annealing temperature, ambient humidity, and open-circuit voltage. The structure employs a multilayer perceptron, containing two fully connected layers with ReLU activation function. The 32-dimensional process feature vector implicitly represents the potential constraints of process conditions on defect formation.

[0033] The fusion and classification head includes feature concatenation, a Dropout layer, and a fully connected layer; the Dropout layer has a dropout rate of 0.5 to prevent overfitting. The fully connected layer maps to the number of classes and outputs a probability distribution via the Softmax function.

[0034] ; in, This is the defect category prediction probability vector output by the model after passing through a fully connected layer and normalization processing, used to characterize the prediction probability of various defects corresponding to the perovskite solar cell substrate. The weight matrix of the fully connected output layer is used to perform a linear transformation on the fused feature vector and extract the mapping relationship between features and defect categories. It is a multimodal fusion feature vector obtained by concatenating image feature vectors and process feature vectors, integrating image visual features and production process parameter features; This is the bias vector for the fully connected output layer, used to adjust the output of the linear transformation and improve the model fitting accuracy. This is used to predict the probability that the sample to be tested is a hole defect. The probability value for predicting the test sample as a scratch defect category for the model; This is used to predict the probability value of the sample being tested being classified as a dirt defect; The system predicts the probability value of the sample being classified as normal and defect-free. In addition to holes, scratches, dirt, and the normal category, the system can also be designed to identify the predicted probability values ​​for other perovskite solar cell defect categories as needed in actual operation.

[0035] This invention addresses the problem of scarce defect samples (extreme imbalance between positive and negative samples) in industrial scenarios by designing the following training strategy.

[0036] The loss function is Focal Loss, which reduces the weight of easily classified samples and focuses on difficult-to-classify samples (minor defects): ; in, Predict the probability of the target class for the model. The focusing parameter is set to 2.0 to suppress gradients in simple samples. The balancing coefficient is set to 0.25 to adjust the contribution ratio of positive and negative samples.

[0037] As one example, it is shown in Table 1 below; Table 1 Training Data

[0038] During the development phase, multiple baseline models (ResNet-50, EfficientNet-B4, ViT) are trained in parallel, and the overall training success rate is defined for optimization. ; Given the serious consequences of missing a defect (classifying it as normal); [the following is set...] =0.5, = =0.25.

[0039] Confusion matrix, precision / recall for each class, ROC curves, and AUC values.

[0040] Furthermore, in some operating conditions, the system of this invention not only outputs defect categories but also uses the feature weights of process branches to inversely deduce the highly sensitive process parameters that lead to such defects. For example, pore defects are mainly related to low annealing temperatures, providing real-time process adjustment suggestions for the production line. During the R&D phase, the defect density distribution of batteries with different formulations is quantitatively compared to quickly assess the stability of new process windows. In addition, this invention acquires EL image sequences of the same component at different aging stages, such as initial, after light decay, and after thermal decay; it introduces LSTM or Transformer timing modules to learn the mapping relationship between defect area expansion rate and optical, thermal, and electrical stress conditions, predicting future defect development trends. A defect feature and performance degradation correlation model is established to estimate the remaining useful life (RUL) of the component, achieving a leap from acceptance judgment to reliability classification.

[0041] Furthermore, this invention utilizes federated learning for multi-plant collaborative evolution, deploying local nodes at each production base to fine-tune the model using local data. Federated learning aggregates the model parameter updates from each node (without uploading original image data), generating a globally optimal model and distributing it. This can improve the problem of data silos within a single plant, continuously accumulate knowledge of rare defect samples, and enhance confidentiality.

[0042] This invention also provides a method for detecting and analyzing perovskite solar cells, comprising the following steps: The equipment transmission unit 1 carries and transmits the perovskite solar cell substrate, and delivers the perovskite solar cell substrate to the detection station corresponding to the electroluminescence detection unit 3. The electroluminescence detection unit 3 applies a preset voltage to the perovskite solar cell substrate, causing the perovskite solar cell substrate to emit an electroluminescence signal. Image acquisition unit 2 acquires electroluminescence images of the perovskite solar cell substrate and transmits the electroluminescence images to AI processing unit 4 in the form of electrical signals; AI processing unit 4 sequentially performs image preprocessing, feature extraction, multimodal fusion reasoning, and defect classification processing on the electroluminescent image, generates defect judgment results, transmits the defect judgment results to component sorting unit 5, and outputs defect statistics data. The component sorting unit 5 performs sorting operations on the perovskite solar cell substrates based on the defect judgment results, rejects unqualified products, and transfers qualified products to the post-processing unit 6 through the equipment transfer unit 1.

[0043] In some embodiments, the present invention uses a perovskite solar cell substrate to transmit image information through a device transmission unit 1 to a high-resolution camera in an electroluminescence detection device. The image information is then compared and processed by an AI processing unit 4, during which defect labeling, analysis, and other processes are performed. Screening feedback is then provided, and unqualified components are sorted in a component sorting unit 5. Good components continue to be processed in the next process 6.

[0044] Through experimental verification, the system of the present invention was compared with the existing system. The results are shown in Table 2 below. As can be seen from Table 2, the false alarm rate of the present invention is reduced by about 44%, the process parameters can be automatically optimized, the hole defects can be reduced, the process tolerance is improved, the efficiency is improved by about 22%, and the risk of missed detection is reduced to a certain extent.

[0045] Table 2 Comparative Experimental Data

[0046] In summary, this invention provides a detection and analysis system for perovskite solar cells. It employs a dynamic benchmark correction module based on local high quantile statistics and robust standard deviation, along with a method for calculating a defect severity index. Simultaneously, it inputs EL images with enhanced channels and process parameter vectors, fusing features through an attention mechanism in a neural network structure. The system utilizes a Focal Loss-based loss function design and hyperparameter combinations tailored to perovskite defect characteristics. It maps defect classification results back to specific process parameter adjustment suggestions using control logic. Finally, it employs a method for predicting defect propagation trends and cell lifetime based on time-series EL image sequences. This invention achieves accurate and efficient detection of perovskite solar cells through transmission, EL excitation, high-resolution imaging, AI processing, and automatic sorting.

[0047] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Anyone skilled in the art can readily implement the present invention according to the description and above. Any modifications, alterations, or equivalent variations made using the technical content disclosed above are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A detection and analysis system for perovskite solar cells, characterized in that, It includes a device transmission unit (1), an image acquisition unit (2), an electroluminescence detection unit (3), an AI processing unit (4), a component sorting unit (5), and a post-processing unit (6); wherein: The equipment transmission unit (1) is used to transmit the perovskite solar cell substrate; the electroluminescence detection unit (3) is set at the detection station of the equipment transmission unit (1) and is used to apply a preset voltage to the perovskite solar cell substrate in order to excite the electroluminescence signal; the image acquisition unit (2) is installed at the detection station of the electroluminescence detection unit (3) and is used to acquire the electroluminescence image of the perovskite solar cell substrate and transmit the electroluminescence image to the AI ​​processing unit (4). The AI ​​processing unit (4) is used to perform image preprocessing, feature extraction, multimodal fusion reasoning and defect classification processing on the electroluminescent image in sequence, generate defect judgment results and transmit them to the component sorting unit (5), and output defect statistics at the same time; the component sorting unit (5) is used to perform sorting operation on the perovskite solar cell substrate according to the defect judgment results, remove unqualified products, and transmit qualified products to the post-processing process unit (6) through the equipment transmission unit (1).

2. The detection and analysis system for perovskite solar cells according to claim 1, characterized in that, The AI ​​processing unit (4) includes a dynamic reference correction module, which is used to construct a dynamic reference light intensity curve. The calculation formula for the dynamic reference light intensity curve is as follows: ； in, The set of electroluminescence intensities of all pixels within the sliding window. The average grayscale value of the pixels within the sliding window. This is the statistical function for the 95th percentile. To calculate only the robust local standard deviation of the sample between the median and the maximum, This is the compensation coefficient.

3. The detection and analysis system for perovskite solar cells according to claim 2, characterized in that, The dynamic benchmark correction module is also used to calculate the defect severity index, and the formula for calculating the defect severity index is as follows: ； in, The measured light intensity per pixel. α is the noise floor constant, and α is the sensitivity gain coefficient.

4. The detection and analysis system for perovskite solar cells according to claim 1, characterized in that, The AI ​​processing unit (4) includes a dual-branch fusion convolutional neural network module, which includes an image feature extraction branch, a process parameter fusion branch, and a fusion and classification head.

5. The detection and analysis system for perovskite solar cells according to claim 4, characterized in that, The input to the image feature extraction branch is a two-channel tensor containing the original electroluminescent grayscale image and the defect severity index image. The backbone network of the image feature extraction branch is an improved ResNet-50 network. Each stage of the improved ResNet-50 network has a CBAM attention module embedded at its end. The output of the image feature extraction branch is an image feature vector obtained by global average pooling.

6. The detection and analysis system for perovskite solar cells according to claim 4, characterized in that, The process parameter fusion branch is a multilayer perceptron structure. The input of the process parameter fusion branch is a process parameter vector including coating speed, annealing temperature, ambient humidity and open circuit voltage. The multilayer perceptron includes two fully connected layers and the activation function is ReLU. The output of the process parameter fusion branch is a process feature vector.

7. The detection and analysis system for perovskite solar cells according to claim 4, characterized in that, The fusion and classification head includes a feature stitching unit, a Dropout layer, and a fully connected output layer. The feature stitching unit is used to stitch together image feature vectors and process feature vectors. The fully connected output layer outputs a defect category probability distribution through a Softmax function.

8. The detection and analysis system for perovskite solar cells according to claim 1, characterized in that, The AI ​​processing unit (4) uses the Focal Loss function for model training. The calculation formula of the Focal Loss function is as follows: ； in, To predict the probability of the target class for the model, To focus parameters, This is the balance coefficient.

9. The detection and analysis system for perovskite solar cells according to claim 1, characterized in that, The AI ​​processing unit (4) also includes a defect evolution prediction module and a federated learning collaboration module; the defect evolution prediction module is used to collect electroluminescence image sequences of perovskite solar cell substrates at different aging stages and predict the defect expansion trend through a time series model; the federated learning collaboration module is used to aggregate local model parameter updates from multiple production bases and generate a globally optimal model.

10. A method for detecting and analyzing perovskite solar cells, based on the detection and analysis system for perovskite solar cells according to any one of claims 1-9, characterized in that, Includes the following steps: The equipment transmission unit (1) carries and transmits the perovskite solar cell substrate, and delivers the perovskite solar cell substrate to the detection station corresponding to the electroluminescence detection unit (3); The electroluminescence detection unit (3) applies a preset voltage to the perovskite solar cell substrate, causing the perovskite solar cell substrate to emit an electroluminescence signal; The image acquisition unit (2) acquires electroluminescent images of the perovskite solar cell substrate and transmits the electroluminescent images to the AI ​​processing unit (4) in the form of electrical signals. The AI ​​processing unit (4) sequentially performs image preprocessing, feature extraction, multimodal fusion reasoning and defect classification processing on the electroluminescent image, generates defect judgment results, transmits the defect judgment results to the component sorting unit (5), and outputs defect statistics data. The component sorting unit (5) performs sorting operations on the perovskite solar cell substrates based on the defect judgment results, removes unqualified products, and transfers qualified products to the post-processing unit (6) through the equipment transfer unit (1).

Images