Photovoltaic precise quality inspection method and device based on neural architecture search and storage medium

Abstract

According to the photovoltaic precise quality inspection method and device based on neural architecture search and the storage medium, defect detection and classification of the battery pieces in the production process are achieved through the ADC technology and the deep learning algorithm. The technology not only can monitor the production process in real time, but also can efficiently identify and position bad products, and provides powerful support for quality control of the photovoltaic manufacturing industry. In the production process of photovoltaic cells, EL data are captured in real time through an integrated sensor and a data acquisition system, the data contain technological parameters, sensor readings, equipment states and the like, and comprehensive monitoring is provided for the production process. According to the method, an advanced deep learning algorithm is utilized, machine learning and statistical analysis technologies are combined, obtained EL data are analyzed, and defect types and causes thereof are quickly and accurately identified. Therefore, engineers and operators can clearly understand the causes of the defective products and take corrective measures accordingly.

Description

Technical Field

[0001] The present invention relates to the technical field of quality inspection for the photovoltaic industry, and in particular to a method, device and storage medium for precise photovoltaic quality inspection based on neural architecture search. Background Art

[0002] In the manufacturing process of the photovoltaic industry, precise intelligent defect detection and strict quality supervision play a key role. The quality inspection of photovoltaic cells is an indispensable part of the production process. However, existing detection methods have limitations in identifying and locating the root causes of defects, which limits their accuracy. Therefore, there is an urgent need for an advanced defect identification technology that can assist manufacturers in tracing the source of defects and effectively deal with the problem of substandard products. The application of this technology will significantly improve the performance and quality of photovoltaic products, thereby promoting the progress of the entire photovoltaic manufacturing industry.

[0003] In the manufacturing chain of the photovoltaic industry, cell quality assurance is a core link. During the production process, the instability of process parameters and fluctuations in equipment status may cause defects in cells, which will directly affect the performance and quality of the product. Traditional quality control methods usually focus on production monitoring and final product testing, but these methods are limited in quickly and accurately identifying the causes of defects, which may lead to an increase in defective rates and a decrease in production efficiency.

[0004] In response to these challenges, fully automated defect detection technology, that is, analysis technology based on automatic defect classification (ADC), has been introduced into the production process of photovoltaic cells in recent years. This technology uses advanced image processing and machine learning algorithms to monitor and analyze key parameters and image data in the production process in real time. Through this analysis, defects such as hidden cracks, concentric circles, contamination, black spots, black dots and edge contamination on the surface of the cell can be accurately identified, providing a powerful tool for subsequent quality analysis and defect cause tracking.

[0005] However, despite the many advantages brought by fully automated ADC technology, it still faces some challenges in the field of photovoltaic manufacturing. First, the amount of data generated in the production process is huge and complex, containing many parameters and image features, which requires an efficient data management and processing system. Secondly, in order to achieve real-time defect classification in a complex production environment, more accurate models and algorithms need to be developed. In addition, it is also a challenge to diagnose faults quickly and accurately and trace corrective measures, especially in large-scale production, where it is crucial to quickly locate the source of defects and take effective measures to reduce the number of defective products.

[0006] Therefore, in order to further improve the quality control capability of photovoltaic cell manufacturing, reduce the defective rate and improve production efficiency, it is necessary to conduct more in-depth research and development on fully automated ADC defect detection technology. The development of this technology will bring a higher level of quality assurance and optimization of production processes to the photovoltaic industry, and promote the technological progress and market competitiveness of the entire industry. Summary of the invention

[0007] The present invention proposes a photovoltaic precision quality inspection method, device and storage medium based on neural architecture search, which can solve at least one of the technical problems in the background technology.

[0008] To achieve the above object, the present invention adopts the following technical solutions: A photovoltaic precision quality inspection method based on neural architecture search, which continuously optimizes and iterates the inspection process through continuous data collection, model training and performance evaluation to adapt to various changes and challenges that may arise in the production process. The process mainly includes the following steps: S1: Information acquisition: the cell emits light under specific excitation conditions, the test equipment captures and records these luminescent images, and the electroluminescence (EL) images of the cell are collected through optical test equipment; Optical testing includes but is not limited to scanning electron microscopy (SEM), transmission electron microscopy (TEM), etc.; S2: Obtain the data to be analyzed and related configuration information; S3: Manually collect a part of data for training to obtain the initial model; S4: Release model; S5: Based on the published model, the image type is predicted by inference and manually sampled; S6: Summarize the results of the random inspection and obtain the accuracy score of the random inspection; S7: For models below a given accuracy threshold, re-judge the training set data; S8: Combine the re-judged training set data with the sampled data, retrain the model, obtain the optimal model of multiple algorithm models, and publish the optimal model; S9: When the accuracy score of the random inspection reaches the set threshold, the model update is suspended, and further manual review is performed according to the type of defect to verify and determine the root cause of the failure, and steps S4 to S8 are repeated; S10: Take measures and use corresponding measures according to the analyzed defect causes; Further, in step S1, the information is obtained in the following manner: step S11: S11: Carefully screen product names, production processes, equipment identification and machine numbers based on a specific time period to ensure that the selected data is representative and accurate.

[0009] Further, in step S2: requesting the required data from step S1, parsing and obtaining the configuration file through real-time transmission data, specifically including the following steps S21 to S23: S21: The ADC loader service monitors new image data in real time. When new EL data is stored, it identifies and parses the EL data and configuration files. S22: Through step S21, the data of the specified time is filtered and further stored in a database with a specified table structure; S23: Through step S22, extract data information, including image path and corresponding unique identification ID information, and produce the extracted information in the form of Kafka communication to provide input for subsequent reasoning services; Further, in step S3: collecting historical EL image data, training the model and labeling the model, specifically including the following steps S31 to S33; S31: screening and labeling historical EL image data; S32: Based on step S31, mount the labeled data into the training container S33: Based on step S32, multiple deep learning network algorithms are selected, such as an improved version of the EfficientNet algorithm, an improved version of the YOLO algorithm, an improved version of the Faster-RCNN algorithm, an improved version of the ResNet algorithm, etc., and model files of each algorithm model are obtained through model training, and the scores of the models are marked; Based on the improved version of the EfficientNet algorithm, first, it has a more efficient training speed: it uses a mixed convolution and layer normalization technology to improve the training speed; second, it has better performance. In the image classification task, the improved version of EfficientNet has achieved higher accuracy while maintaining computational efficiency. The core lies in its compound scaling method, the specific formula is as follows:

[0010] in, , and is the scaling factor, is the control parameter of the scaling factor.

[0011] Introduce a dynamic adjustment mechanism to make the scaling factor , and Adaptively adjust according to different characteristics of input data.

[0012] Based on the improved version of the YOLO algorithm, it performs well in real-time classification detection while maintaining high accuracy and supports multiple frameworks; Based on the improved version of the Faster-RCNN algorithm, the region proposal network (RPN) is introduced to greatly improve the detection speed and accuracy, while optimizing the classification and regression tasks and enhancing the detection effect. The specific formula is as follows:

[0013] in, is the classification loss, is the regression loss, is the equilibrium parameter, is the predicted class probability, is the real category, and are the true and predicted bounding box parameters, respectively; Based on the improved version of the ResNet algorithm, firstly, it has higher flexibility and scalability, and uses a group convolution structure, allowing the network to use computing resources more efficiently; secondly, it has better performance and provides higher accuracy; its residual formula is as follows:

[0014] in, is the number of groups, is the conversion function, It is The weight of each group; Furthermore, in step S4: publishing the trained optimal model, specifically including the following steps S41 to S42: S41: Based on step S33, the best model is selected based on the model scores; S42: Based on step S41, the training platform sends the optimal model to the inference platform through Kafka communication; Further, in step S5: based on the published model, the image type is inferred and predicted, and a manual sampling is performed, which specifically includes the following steps S51 to S52: S51: Based on the optimal model obtained in step S42, the data obtained in step S23 are inferred to obtain corresponding classification results; S52: Based on step S51, the inference results are manually sampled to obtain the scores of each inference data of the model; Further, in step S6: the results of the random inspection are summarized to obtain the accuracy score of the random inspection, and the specific steps are as follows: steps S61 to S6; S61: Based on step S52, summarizing the results of the random inspection; S62: Based on step S61, the accuracy of the sampling results is scored; Further, in step S7: for the model with accuracy lower than the given accuracy threshold, the training set data is re-judged, specifically in the following steps S71 to S73: S71: Based on step S61, the scoring result of the random inspection is compared with the set accuracy threshold; S72: Based on step S71, for models with accuracy below a given threshold, the training set is relabeled on the training platform; S73: Based on step S71, continue monitoring for models with accuracy higher than a given threshold; Further, in step S8: the re-judged training set data is combined with the sampled data, the model is re-trained, the optimal model of multiple algorithm models is obtained, and the optimal model is released, specifically in the following steps S81 to S84: S81: Based on step S72, merging the relabeled data set with the sampled data; S82: Based on step S81, multiple algorithm model training is performed on the collected data; S83: Based on step S82, the model scores of multiple algorithms are sorted, and the model with the highest accuracy score is selected; S84: Based on step S83, publish the optimal model Further, in step S9: when the accuracy score of the random inspection reaches the set threshold, the model update is suspended, and further manual review is performed according to the type of defect to verify and determine the root cause of the failure, and steps S4 to S8 are looped, specifically steps S91 to S9 as follows: S91: comparing the accuracy scoring result of S62 with a set accuracy threshold; S92: When the accuracy score of the model is lower than the set threshold, the training data is cleaned based on step S8, and the model is trained, and steps S4 to S8 are repeated; S93: When the accuracy score of the model is not lower than the set threshold, further verify and determine the cause of the failure. On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0015] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0016] It can be seen from the above technical solutions that the photovoltaic precision quality inspection method based on neural architecture search of the present invention innovatively applies electroluminescence (EL) detection technology to the manufacturing field of photovoltaic cells. Through the integration of ADC technology and deep learning algorithms, neural architecture search and multi-algorithm fusion. Utilize the integrated sensor and data acquisition system to obtain EL image data in real time, and automatically parse and store these images and configuration files through the ADC loader service. The system can transmit data to the inference platform through the Kafka message queue, supporting large-scale real-time data processing. In the process of training and model iteration, multiple algorithms such as EfficientNet improved version, ResNet improved version, YOLO improved version, Faster-RCNN improved version and other algorithms are used to train models in parallel, and the model with the highest accuracy is selected for release for inference service. At the same time, for models with poor performance, data cleaning and re-judgment will be combined with manual sampling results, and the model will be retrained and iteratively optimized. By evaluating the accuracy of the model inference results, if the model accuracy does not reach the predetermined threshold, the new data will be automatically added to the training set and the model will be updated. This mechanism ensures that the detection model can maintain high accuracy under different production conditions. It realizes the defect detection and classification of solar cells during the production process. This technology can not only monitor the production process in real time, but also efficiently identify and locate defective products, providing strong support for quality control in the photovoltaic manufacturing industry.

[0017] First, during the production of photovoltaic cells, EL data is captured in real time through integrated sensors and data acquisition systems. These data cover process parameters, sensor readings, and equipment status, providing comprehensive monitoring of the production process.

[0018] Secondly, the acquired EL data is analyzed using advanced deep learning algorithms combined with machine learning and statistical analysis techniques. The algorithm is capable of feature extraction and defect detection, quickly and accurately identifying defect types and their causes.

[0019] Furthermore, intuitive visualization tools and report generation functions are provided, which enable engineers and operators to clearly understand the causes of defective products through charts, trend analysis and fault trees, and take corrective measures accordingly. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural schematic diagram of the present invention; Figure 2 Iterative flow chart for defect identification. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0022] refer to Figure 1 to Figure 2 This paper proposes a photovoltaic precision quality inspection system based on Neural Architecture Search (NAS). The system builds a quality inspection system with self-correction and adaptive learning capabilities through a closed-loop mechanism of model re-judgment and label re-training, as well as a combination of manual sampling and automatic reasoning results. This dynamic model optimization and data re-judgment mechanism can significantly improve the accuracy and stability of quality inspection, and is a very innovative part of the entire technical process.

[0023] The present invention uses the Kafka communication mechanism to push the model to the reasoning platform (S42) in real time, forming a full-link automated management from model training to practical application. The results of real-time reasoning are manually sampled and re-judged, and then fed back to the training system, building a closed-loop process of continuous learning and optimization, and improving the adaptability of the system in a dynamic production environment.

[0024] Specifically, the following steps S1 to S10 are included: Information acquisition: the cell emits light under specific excitation conditions, and the test equipment captures and records these luminous images. The cell is subjected to electroluminescence (EL) image acquisition through the optical test equipment, as shown in step S11: S11: Carefully screen product names, production processes, equipment identification and machine numbers based on a specific time period to ensure that the selected data is representative and accurate; According to step S1, requesting data, obtaining the data to be analyzed and related configuration information, specifically includes the following steps S21 to S23: S21: The ADC loader service monitors new image data in real time. When new EL data is stored, it identifies and parses the EL data and configuration files. S22: Through step S21, the data of the specified time is filtered and further stored in a database with a specified table structure; S23: Through step S22, extract data information, including image path and corresponding unique identification ID information, and produce the extracted information in the form of Kafka communication to provide input for subsequent reasoning services; According to the analysis data and configuration information obtained in step S2, a part of data is manually collected for training to obtain an initial model, which specifically includes the following steps: S31: screening and labeling historical EL image data; S32: Based on step S31, mount the labeled data into the training container S33: Based on step S32, obtain the model files of each algorithm model through model training, and mark the scores of the models; Based on the initial models obtained in step S3, the best model is selected and released. The specific steps are as follows: S41: Based on step S33, obtain the scores of multiple models and select the best model; S42: Based on step S41, the training platform sends the optimal model to the inference platform through Kafka communication; Based on the model published in step S4, the image type is inferred and predicted, and manual sampling is performed. The specific steps are as follows: S51: Based on the optimal model obtained in step S42, the data obtained in step S23 are inferred to obtain corresponding classification results; S52: Based on step S51, the inference results are manually sampled to obtain the scores of each inference data of the model; Based on step S5, the results of the random inspection are summarized to obtain the accuracy score of the random inspection. The specific steps are as follows: S61: Based on step S52, summarizing the results of the random inspection; S62: Based on step S61, the accuracy of the sampling results is scored; Based on the algorithm model scores obtained in step S6, for models below a given accuracy threshold, the training set data is re-judged. The specific steps are as follows: S71: Based on step S61, the scoring result of the random inspection is compared with the set accuracy threshold, and the system dynamically adjusts the accuracy threshold according to the actual reasoning and random inspection results; S72: Based on step S71, for models with accuracy below the given threshold, the training set is relabeled on the training platform to remove noise data or erroneous data from the training set, reduce data pollution, and improve the stability and accuracy of the model. The re-judgment mechanism in this step can effectively avoid the performance degradation of the model due to noise data or labeling errors; S73: Based on step S71, continue monitoring for models with accuracy higher than a given threshold; Based on the label data obtained in step S72, the re-judged training set data is combined with the sampled data, the model is retrained, the optimal model of multiple algorithm models is obtained, and the optimal model is released. The specific steps are as follows: S81: Based on step S72, the relabeled data set is merged with the sampled data. This process can dynamically improve the quality of the training set and ensure that the model performs more reliably in production. This re-judgment mechanism enables the model to have stronger adaptability and self-correction capabilities. S82: Based on step S81, multiple algorithm model training is performed on the collected data; S83: Based on step S82, the model scores of multiple algorithms are sorted, and the model with the highest accuracy score is selected, and the accuracy of the final model is improved through continuous iteration and optimization; S84: Based on step S83, publish the optimal model Based on step S6, when the accuracy score of the random inspection reaches the set threshold, the model update is suspended, and further manual review is performed according to the type of defect to verify and determine the root cause of the failure, and steps S4 to S8 are repeated; S10: Take measures. According to the analyzed defect causes, use corresponding measures. The specific methods are as follows: S101: Introducing advanced optical inspection equipment and technology, using advanced optical inspection equipment, such as scanning electron microscope (SEM) and transmission electron microscope (TEM), to improve the detection precision and accuracy of electroluminescence (EL) data. These high-precision equipment can capture defects and minor anomalies of the cell more carefully, thereby improving the reliability of detection; S102: Strengthen environmental management at the production site to ensure cleanliness and tidiness of the production environment and reduce the impact of pollution sources. Through strict environmental management measures, reduce the interference of pollutants such as dust and particles on the production process and ensure the stability of the production environment and the quality of the battery cells.

[0025] S103: Optimize material selection and processing technology, select high-quality materials and optimize processing technology to improve the manufacturing quality and stability of battery cells. By continuously improving material formulas and processing technologies, ensure that the produced battery cells have higher performance and longer service life; S104: Implement a strict quality management system and establish a comprehensive quality management system, including testing, evaluation and feedback mechanisms. Timely discover problems in the production process and take prompt measures to solve them. Through systematic quality management, ensure that the quality of each link meets the standards.

[0026] S105: Train and improve employee skills. Ensure that employees have professional skills in operating specifications and equipment maintenance through regular training and technical improvement. Only when operators are proficient in production processes and equipment operations can human errors be effectively reduced and production efficiency and product quality be improved.

[0027] Through the above steps S11 to S105, the accuracy of detailed screening based on product name, production process, equipment identification and machine number can be output. The higher the accuracy obtained, the higher the accuracy of the algorithm model and the higher its use value.

[0028] Photovoltaic manufacturers need to sample and test 200,000 cells produced every day. Each quality inspector can review about 1,000 cells per hour, which means that manufacturers need nearly 25 quality inspectors to work every day. After the introduction of the present invention, 33 images can be processed per second by using an NVIDIA A10 GPU card. In the case of concurrent reasoning, the workload of the current 25 quality inspectors can be completely replaced. Compared with traditional manual analysis, the present invention not only greatly improves the speed, but also has a wider range of analyzed data. The results show that the method has a significant effect in identifying and classifying the main defect types that affect the quality of cells, which greatly improves the work efficiency of photovoltaic engineers and provides effective guidance for production yield. In addition, due to long-term fatigue work, manual re-judgment may have more misjudgment risks, while the reasoning of the algorithm model will not be affected by such factors. Therefore, the present invention not only has a significant improvement in efficiency, but also provides reliable guarantees in terms of accuracy and stability, ensuring higher quality control and process improvement directions.

[0029] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0030] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0031] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any one of the photovoltaic precision quality inspection methods based on neural architecture search in the above embodiments.

[0032] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above methods.

[0033] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0034] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0035] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic precision quality inspection method based on neural architecture search, characterized in that: The following steps are included: S1: Information acquisition, allowing the cell to emit light under set excitation conditions, the test equipment captures and records these luminous images, and the electroluminescence (EL) image of the cell is collected through the optical test equipment; S2: Obtain the data to be analyzed and related configuration information; S3: Manually collect a part of data for training to obtain the initial model; S4: Release model; S5: Based on the published model, the image type is predicted by inference and manually sampled; S6: Summarize the results of the random inspection and obtain the accuracy score of the random inspection; S7: For models below a given accuracy threshold, re-judge the training set data; S8: Combine the re-judged training set data with the sampled data, retrain the model, obtain the optimal model of multiple algorithm models, and publish the optimal model; S9: When the accuracy score of the random inspection reaches the set threshold, the model update is suspended, and further manual review is performed according to the type of defect to verify and determine the root cause of the failure, and steps S4 to S8 are repeated; S10: Take measures and use corresponding measures based on the analyzed causes of defects.

2. The photovoltaic precision quality inspection method based on neural architecture search according to claim 1 is characterized by: The method of obtaining information in step S1 is as follows: step S11: S11: Based on the set time period, carefully screen the product name, production process, equipment identification and machine number to ensure that the selected data is representative and accurate.

3. The photovoltaic precision quality inspection method based on neural architecture search according to claim 2 is characterized by: Step S2 requests the required data from step S1, parses and obtains the configuration file through real-time data transmission, and specifically includes the following steps S21 to S23: S21: The ADC loader service monitors new image data in real time. When new EL data is stored, it identifies and parses the EL data and configuration files. S22: Through step S21, the data of the specified time is filtered and further stored in a database with a specified table structure; S23: Through step S22, extract data information, including image path and corresponding unique identification ID information, and produce the extracted information through Kafka communication to provide input for subsequent reasoning services.

4. The photovoltaic precision quality inspection method based on neural architecture search according to claim 3 is characterized by: In step S3, historical EL image data is collected, a model is trained and a model is labeled, which specifically includes the following steps S31 to S33; S31: screening and labeling historical EL image data; S32: Based on step S31, mount the labeled data into the training container S33: Based on step S32, multiple deep learning network algorithms are selected, and model files of each algorithm model are obtained through model training, and the scores of the models are marked.

5. The photovoltaic precision quality inspection method based on neural architecture search according to claim 4 is characterized by: In step S4, the trained optimal model is published, which specifically includes the following steps S41 to S42: S41: Based on step S33, the best model is selected based on the model scores; S42: Based on step S41, the training platform sends the optimal model to the inference platform through Kafka communication.

6. The photovoltaic precision quality inspection method based on neural architecture search according to claim 5 is characterized by: In step S5: based on the published model, the image type is inferred and predicted, and a manual sampling is performed, which specifically includes the following steps S51 to S52: S51: Based on the optimal model obtained in step S42, the data obtained in step S23 are inferred to obtain corresponding classification results; S52: Based on step S51, the inference results are manually sampled to obtain the scores of each inference data of the model.

7. The photovoltaic precision quality inspection method based on neural architecture search according to claim 6 is characterized by: In step S6: the results of the random inspection are summarized to obtain the accuracy score of the random inspection. The specific steps are as follows: steps S61 to S6; S61: Based on step S52, summarizing the results of the random inspection; S62: Based on step S61, the accuracy of the sampling results is scored.

8. The photovoltaic precision quality inspection method based on neural architecture search according to claim 6 is characterized by: In step S9: when the accuracy score of the random inspection reaches the set threshold, the model update is suspended, and further manual review is performed according to the type of defect to verify and determine the root cause of the failure, and steps S4 to S8 are looped, specifically steps S91 to S9 as follows: S91: comparing the accuracy scoring result of S62 with a set accuracy threshold; S92: When the accuracy score of the model is lower than the set threshold, the training data is cleaned based on step S8, and the model is trained, and steps S4 to S8 are repeated; S93: When the accuracy score of the model is not lower than the set threshold, further verify and determine the cause of the failure.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 8.

Images