Photovoltaic detection optimization method and system based on deep learning, and terminal

Abstract

The invention discloses a photovoltaic detection optimization method, system and terminal based on deep learning, and the method comprises the steps: inputting an operation image into a chain network model for feature extraction and normalization adjustment, and outputting an initial feature map; inputting the initial feature map into an attention training model for weighting processing to obtain a target feature map, inputting the target feature map into the attention training model for training, outputting a plurality of loss values, and performing weighting processing on the target feature map according to the loss values to obtain a training sample; and obtaining a target training sample after screening, and inputting the target training sample into the chain network model and the attention training model for optimization training. According to the method, the chain network model and the attention mechanism are introduced into the deep learning network, so that the photovoltaic module image is processed, the problems of photovoltaic panel shielding, hot spot effect, cracks, inverter faults and the like are rapidly and accurately detected, model training is performed by using the image, various fault features are learned and identified, and efficient automatic detection is realized.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic detection optimization technology, and in particular to a photovoltaic detection optimization method, system, terminal and computer-readable storage medium based on deep learning. Background Art

[0002] As a clean and renewable energy technology, photovoltaic systems play an important role in the transformation of the global energy structure. Their performance and reliability directly affect economic and environmental benefits. In recent years, with the rapid development of deep learning technology, applying it to the diagnosis of photovoltaic systems has become a new trend.

[0003] In photovoltaic systems, common faults include aging of photovoltaic modules, hot spot effects, shading, inverter failures, etc. These faults will not only reduce the power generation efficiency of the photovoltaic system, but may also cause damage to the equipment. The photovoltaic system diagnosis model based on the YOLOv8 framework has a large model scale and insufficient feature extraction capabilities, resulting in low detection efficiency and unable to cope with the continuous evolution of defects.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] The main purpose of the present invention is to provide a photovoltaic detection optimization method, system, terminal and computer-readable storage medium based on deep learning, aiming to solve the problems in the prior art of large data volume requirements, low detection accuracy and efficiency during detection of current photovoltaic systems.

[0006] To achieve the above object, the present invention provides a photovoltaic detection optimization method based on deep learning, and the photovoltaic detection optimization method based on deep learning comprises the following steps:

[0007] Constructing a chain network model, integrating the chain network model into a backbone network, acquiring multiple historical operation images of a photovoltaic detection target, inputting the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and outputting multiple initial feature maps;

[0008] Constructing an attention training model, integrating the attention training model into the backbone network, inputting all the initial feature maps into the attention training model for weighted processing, and outputting a plurality of target feature maps;

[0009] Inputting all the target feature maps into the attention training model for training, outputting multiple loss values ​​of all the target feature maps, and performing weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples;

[0010] Screening all the training samples according to a preset standard to obtain a plurality of target training samples, and inputting the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model;

[0011] Acquire multiple target running images to be detected, input all the target running images into the target chain network model, output multiple high-quality initial feature maps, input all the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

[0012] Optionally, the photovoltaic detection optimization method based on deep learning, wherein the chain network model is constructed, the chain network model is integrated into the backbone network, a plurality of historical operation images of the photovoltaic detection target are obtained, the plurality of historical operation images are input into the chain network model for feature extraction and normalization adjustment, and a plurality of initial feature maps are output, specifically including:

[0013] Constructing a chain network model and integrating the chain network model into the backbone network;

[0014] Acquire multiple historical operation images input by a user, and input all of the historical operation images into the chain network model;

[0015] Extracting multiple basic feature information of all the historical running images in the initial convolutional layer of the chain network model;

[0016] Extracting multiple complex feature information of all the historical operation images in the middle layer of the chain network model;

[0017] Extracting multiple abstract feature information of all the historical running images in the deep convolutional layer of the chain network model;

[0018] By using image preprocessing technology, all the basic feature information, all the complex feature information and all the abstract feature information are standardized and a plurality of the initial feature maps are output.

[0019] Optionally, the photovoltaic detection optimization method based on deep learning, wherein the attention training model is constructed, the attention training model is integrated into the backbone network, all the initial feature maps are input into the attention training model for weighted processing, and multiple target feature maps are output, specifically including:

[0020] Constructing an attention training model and integrating the attention training model into the backbone network;

[0021] Input all the initial feature maps into the attention training model, and calculate the importance weight value of each feature pixel in each of the initial feature maps;

[0022] According to each of the initial feature maps and the corresponding importance weight values, generating a weighted feature map corresponding to each of the initial feature maps;

[0023] Each of the weighted feature maps is fused with the corresponding initial feature map, and a target feature map corresponding to each of the initial feature maps is output.

[0024] Optionally, the photovoltaic detection optimization method based on deep learning, wherein all the target feature maps are input into the attention training model for training, multiple loss values ​​of all the target feature maps are output, and all the target feature maps are weighted according to all the loss values ​​to obtain multiple training samples, specifically includes:

[0025] Inputting all the target feature maps into the attention training model for training, generating multiple prediction boxes corresponding to all the target feature maps, and obtaining multiple real boxes in the running image corresponding to all the target feature maps;

[0026] Generate a loss value for each target feature map according to all predicted frames and all corresponding real frames of each target feature map, where the loss value is used to measure the degree of overlap between each feature map and the corresponding real frame;

[0027] Evaluate all the target feature maps according to all the loss values ​​to obtain an evaluation result;

[0028] According to the evaluation result and each loss value, each target feature map is weighted to obtain a training sample corresponding to each target feature map.

[0029] Optionally, the photovoltaic detection optimization method based on deep learning, wherein, according to the evaluation result and each of the loss values, weighted processing is performed on each of the target feature maps to obtain a training sample corresponding to each of the target feature maps, specifically including:

[0030] Determine the quality weight value corresponding to each loss value according to the evaluation result and preset weight mapping information;

[0031] Multiply each of the quality weight values ​​by the corresponding loss value to obtain a target weight value corresponding to each of the target feature maps;

[0032] Assign each of the target weight values ​​to the corresponding target feature map to obtain a training sample corresponding to each of the target feature maps.

[0033] Optionally, in the photovoltaic detection optimization method based on deep learning, the preset standards include: clarity standard, illumination condition standard, annotation accuracy standard, data distribution standard and noise level;

[0034] The method of screening all the training samples according to the preset standard to obtain a plurality of target training samples, and inputting the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model, specifically includes:

[0035] Determining, according to the target weight value of each training sample, whether each training sample meets the clarity standard, the lighting condition standard, the labeling accuracy standard, the data distribution standard or the noise level;

[0036] If the clarity standard, the lighting condition standard, the annotation accuracy standard, the data distribution standard or the noise level are met, all the corresponding training samples are respectively input into the chain network model and the attention training model, and the chain network model and the attention training model are optimized to obtain the target chain network model and the target attention training model.

[0037] Optionally, the photovoltaic detection optimization method based on deep learning, wherein the step of acquiring multiple target operation images to be detected, inputting all the target operation images into the target chain network model, outputting multiple high-quality initial feature maps, inputting all the high-quality initial feature maps into the target attention training model, and outputting multiple high-quality feature maps, specifically includes:

[0038] Acquire multiple target running images to be detected, input all the target running images into the target chain network model, and output multiple high-quality initial feature maps;

[0039] All of the high-quality initial feature maps are input into the target attention training model, and a plurality of the high-quality feature maps are output.

[0040] In addition, to achieve the above-mentioned purpose, the present invention also provides a photovoltaic detection optimization system based on deep learning, wherein the photovoltaic detection optimization system based on deep learning includes:

[0041] An image preprocessing module is used to construct a chain network model, integrate the chain network model into the backbone network, obtain multiple historical operation images of photovoltaic detection targets, input the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and output multiple initial feature maps;

[0042] A feature map weighting module is used to construct an attention training model, integrate the attention training model into the backbone network, input all the initial feature maps into the attention training model for weighted processing, and obtain multiple target feature maps;

[0043] A loss weighting module, used for inputting all the target feature maps into the attention training model for training, outputting multiple loss values ​​of all the target feature maps, and performing weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples;

[0044] A model training module, used to screen all the training samples according to a preset standard to obtain a plurality of target training samples, and input the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model;

[0045] The image detection module is used to obtain multiple target running images to be detected, input all the target running images into the target chain network model, output multiple high-quality initial feature maps, input all the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

[0046] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a deep learning-based photovoltaic detection optimization program stored in the memory and executable on the processor, wherein the deep learning-based photovoltaic detection optimization program, when executed by the processor, implements the steps of the deep learning-based photovoltaic detection optimization method as described above.

[0047] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a photovoltaic detection optimization program based on deep learning, and when the photovoltaic detection optimization program based on deep learning is executed by a processor, the steps of the photovoltaic detection optimization method based on deep learning as described above are implemented.

[0048] In the present invention, a chain network model is constructed, the chain network model is integrated into the backbone network, a plurality of historical operation images of photovoltaic detection targets are obtained, the plurality of historical operation images are input into the chain network model for feature extraction and normalization adjustment, and a plurality of initial feature maps are output; an attention training model is constructed, the attention training model is integrated into the backbone network, all the initial feature maps are input into the attention training model for weighted processing, and a plurality of target feature maps are output; all the target feature maps are input into the attention training model for training, a plurality of loss values ​​of all the target feature maps are output, and the loss values ​​of all the target feature maps are output according to the weighted processing of all the target feature maps. The loss value is used to weight all the target feature maps to obtain multiple training samples; all the training samples are screened according to preset standards to obtain multiple target training samples, and the multiple target training samples are respectively input into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model; multiple target operation images to be detected are obtained, all the target operation images are input into the target chain network model, multiple high-quality initial feature maps are output, all the high-quality initial feature maps are input into the target attention training model, and multiple high-quality feature maps are output. The present invention introduces the chain network model and the attention mechanism into the deep learning network to process the photovoltaic module image, quickly and accurately detect common problems such as photovoltaic panel shading, hot spot effect, cracks and inverter failure, use images for model training, learn and identify various fault characteristics, and realize efficient automatic detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0050] Figure 2 It is a dust and bird droppings detection diagram of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0051] Figure 3 It is a bypass diode short-circuit diagram of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0052] Figure 4 It is a hot spot map of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0053] Figure 5 It is a long crack diagram of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0054] Figure 6 It is a YOLO segmentation map of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0055] Figure 7 It is a UNET segmentation diagram of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0056] Figure 8 It is a model iteration and convergence curve diagram of a preferred embodiment of the photovoltaic detection optimization method based on deep learning of the present invention;

[0057] Fig. 9 It is a schematic diagram of the principle of a preferred embodiment of a photovoltaic detection optimization system based on deep learning of the present invention;

[0058] Fig.10 Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] The photovoltaic detection optimization method based on deep learning described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the photovoltaic detection optimization method based on deep learning includes the following steps:

[0061] Step S10, constructing a chain network model, integrating the chain network model into the backbone network, acquiring multiple historical operation images of the photovoltaic detection target, inputting the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and outputting multiple initial feature maps.

[0062] Photovoltaic panels are exposed to sunlight for a long time, and are often interfered by reflections during detection, which affects the detection accuracy of the model. For this reason, a chain network model is added to the backbone network of the YOLOv8 algorithm (for example, the CPA-Enhancer chain thinking network, a model that can improve the object detection performance under unknown multiple degradation conditions). Through feature extraction and normalization adjustment, the model's perception of lighting changes is improved, thereby enhancing the detection accuracy under strong light and reflective conditions.

[0063] Specifically, a chain network model is constructed and integrated into the backbone network; multiple historical operation images input by the user are obtained, and all the historical operation images are input into the chain network model; multiple basic feature information of all the historical operation images are extracted in the initial convolution layer of the chain network model; multiple complex feature information of all the historical operation images are extracted in the middle layer of the chain network model; multiple abstract feature information of all the historical operation images are extracted in the deep convolution layer of the chain network model; all the basic feature information, all the complex feature information and all the abstract feature information are standardized through image preprocessing technology, and multiple initial feature maps are output.

[0064] Among them, multiple historical operation images input by the user can be taken by drones, and high-definition images can be used for model training, so as to realize defect detection of photovoltaic inspection systems, and use these historical operation images to train the model. First, low-level feature extraction is performed, and the historical operation images are input into the initial convolution layer to extract basic features such as edges and textures of the historical operation images. For example, small cracks and textures on the surface of photovoltaic modules, or Figure 2 As shown in the figure, the features of dust and bird droppings on the photovoltaic panel are detected; then intermediate feature extraction is performed to extract more complex features of the historical running image in the middle layer of the chain network model, such as the shape and structural information of the photovoltaic panel, and the overall shape and installation angle of the photovoltaic module are identified, such as Figure 3 As shown in the figure, the features of bypass diode short circuit are detected; finally, advanced feature extraction is performed to extract more abstract features in the deep convolutional layer of the chain network model, such as the comprehensive features of different types of faults, identifying obstructions and hot spot effects on photovoltaic panels (under certain conditions, the shaded solar cell components in a series branch will be treated as loads to consume the energy generated by other illuminated solar cell components. The shaded solar cell components will heat up at this time, which is the hot spot effect), as shown in the figure. Figure 4 As shown, the hot spot effect on the photovoltaic panel is detected, such as Figure 5 As shown, detecting long cracks on a photovoltaic panel.

[0065] Furthermore, through image preprocessing technology, the acquired basic feature information, complex feature information and abstract feature information are standardized (i.e., adaptive adjustment to lighting changes). First, each feature information is adaptively adjusted to strong light. Under strong light conditions, the contrast and brightness of the historical operation image are adjusted to make each feature information more obvious. For example, the photovoltaic panel image under direct sunlight is enhanced to reduce the reflection effect, so that each historical operation image is more convenient for model detection; then, each feature information is adaptively adjusted to shadow. Under shadow conditions, the brightness of the historical operation image is adjusted and shadow compensation is increased to enhance the image details of each feature information in the historical operation image, thereby improving the visibility of details of the photovoltaic components under partial shadow coverage; finally, the images under different lighting conditions are standardized so that the model can adapt to various lighting conditions, that is, the operation images under different lighting environments can be uniformly processed, so that the feature information of the operation image can also remain consistent under different lighting environments, thereby improving the accuracy of photovoltaic detection.

[0066] After completing the multi-scale feature extraction of the historical running images and the adaptive adjustment of the illumination changes, the initial feature map is output for subsequent model training, thereby improving the model's sensitivity and adaptability to illumination changes.

[0067] Step S20, construct an attention training model, integrate the attention training model into the backbone network, input all the initial feature maps into the attention training model for weighted processing, and output multiple target feature maps.

[0068] An attention training model is added to the backbone network of the YOLOv8 algorithm (for example, an EMA attention mechanism, i.e., an Exponential Moving Average attention mechanism) and the small target features are weighted by the attention training model to improve the detection accuracy of small target defects.

[0069] Specifically, an attention training model is constructed and integrated into the backbone network; all the initial feature maps are input into the attention training model, and the importance weight value of each feature pixel in each initial feature map is calculated; according to each initial feature map and the corresponding importance weight value, a weighted feature map corresponding to each initial feature map is generated; each weighted feature map is fused with the corresponding initial feature map, and a target feature map corresponding to each initial feature map is output.

[0070] Among them, after the initial feature map is input into the attention training model, each feature pixel in each initial feature map is extracted, and the importance weight of each feature pixel is calculated. Then, according to the importance weight and the corresponding feature pixel, a weighted feature map corresponding to each initial feature map can be generated. Then, each weighted feature map is fused with the corresponding initial feature map, and the target feature map corresponding to each initial feature map is output for subsequent model training, thereby improving the weight of the small target features in the initial feature map and improving the detection accuracy of the model for small target defects.

[0071] Step S30: input all the target feature maps into the attention training model for training, output multiple loss values ​​of all the target feature maps, perform weighted processing on all the target feature maps according to all the loss values, and obtain multiple training samples.

[0072] In the process of training the model, the WIOU loss function (Wise-IoU loss function, an improved IoU loss function used to solve the problem of small target prediction box regression) can be used to weight the loss of the training process, so as to balance the impact of high-quality and low-quality samples.

[0073] Specifically, all the target feature maps are input into the attention training model for training, multiple prediction boxes corresponding to all the target feature maps are generated, and multiple real boxes in the running image corresponding to all the target feature maps are obtained; according to all the prediction boxes of each target feature map and all the corresponding real boxes, a loss value of each target feature map is generated, and the loss value is used to measure the degree of overlap between each feature map and the corresponding real box; according to all the loss values, all the target feature maps are evaluated to obtain evaluation results; according to the evaluation results and each loss value, each target feature map is weighted to obtain a training sample corresponding to each target feature map.

[0074] Among them, Figure 6 and Figure 7 As shown in the figure, when calculating the loss, the IOU value (Intersection over Union, a metric commonly used in computer vision) of each predicted box and the corresponding real box of all target feature maps is calculated to measure the degree of overlap between the predicted box and the corresponding real box, and then the basic loss (i.e., loss value) of each sample (target feature map) is calculated based on the IOU value. Combined with the pre-set evaluation criteria, each sample is weighted to obtain the training sample corresponding to each sample.

[0075] Furthermore, based on the evaluation results and preset weight mapping information, the quality weight value corresponding to each loss value is determined; each quality weight value is multiplied by the corresponding loss value to obtain the target weight value corresponding to each target feature map; each target weight value is assigned to the corresponding target feature map to obtain a training sample corresponding to each target feature map.

[0076] Among them, when weighting the samples, it is necessary to calculate the basic loss of each sample and the quality weight value set according to the evaluation standard to obtain the final target weight value of each sample, that is, multiply the basic loss of each sample by the quality weight value set according to the evaluation standard to obtain the target weight value corresponding to each sample (that is, the loss value corresponding to each sample), such as Figure 8 As shown in the figure, according to different training cycles, loss values ​​of different training types can be obtained, including bounding box loss, classification loss and distribution focus loss, etc., and training samples of corresponding samples can be generated according to the target weight value, which is used to optimize the training of the chain network model and the attention training model, thereby improving the detection robustness of the two models under various lighting conditions. All target weight values ​​can also be added to obtain the total loss value of all samples (that is, all target feature maps), and the model parameters can also be updated according to the total loss value.

[0077] Step S40, screening all the training samples according to preset standards to obtain multiple target training samples, and inputting the multiple target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model.

[0078] Among them, not all training samples obtained need to be used for model optimization training. They can be screened according to preset standards to improve the robustness of model training. The preset standards include: clarity standard, lighting condition standard, labeling accuracy standard, data distribution standard and noise level.

[0079] Specifically, according to the target weight value of each of the training samples, determine whether each of the training samples meets the clarity standard, the lighting condition standard, the annotation accuracy standard, the data distribution standard or the noise level; if it meets the clarity standard, the lighting condition standard, the annotation accuracy standard, the data distribution standard or the noise level, then all the corresponding training samples are respectively input into the chain network model and the attention training model, and the chain network model and the attention training model are optimized and trained to obtain the target chain network model and the target attention training model.

[0080] Furthermore, the training samples may vary according to the weather conditions and environment in different regions. The parameters of the model can be adjusted according to different environmental information to ensure the detection accuracy of the model, so that the model can continue to learn and optimize after deployment, reduce the impact of extreme environmental changes on the model, and gradually improve the performance of the model. It can also be combined with edge computing devices to realize the processing and preliminary detection of real-time data on site, and upload the preliminary detection results and raw data to the cloud for more complex analysis and long-term data storage. This can not only reduce the computing pressure on the cloud, but also achieve efficient data processing and fault response. When the computing resources of the edge device are limited, the structure of the YOLOv8 model can be optimized, and a lightweight detection model can be designed. The model pruning, quantization and other technologies can be used to reduce model parameters and calculation amount, while maintaining a high detection accuracy to ensure efficient operation on low-power devices.

[0081] Among them, by analyzing the target weight value corresponding to each training sample, the quality level of the corresponding training sample can be determined. For example, images with high clarity (for example, images without blur and moderate contrast) are assigned higher target weight values, images with uniform illumination (for example, images without overexposure or excessive shadows) are assigned higher target weight values, and samples with accurate annotations (for example, the annotation box is highly consistent with the actual target) are assigned higher target weight values. Samples that are common or rare but important in the data set will be assigned higher target weight values, while samples with low noise are assigned higher weights, and samples with high noise have lower weights. Low-quality training samples are screened out, thereby improving the efficiency of model training and the accuracy of photovoltaic detection.

[0082] Furthermore, when training samples are insufficient, data augmentation technology and generative adversarial networks (GAN) can be used to generate more high-quality training data. Through data augmentation (such as rotation, cropping, and noise addition), the diversity of training data can be increased and the generalization ability of the model can be improved. GAN can be used to generate images of simulated fault scenarios to further enrich the training data set, thereby improving the robustness of the optimized trained model.

[0083] Step S50, obtain multiple target operation images to be detected, input all of the target operation images into the target chain network model, output multiple high-quality initial feature maps, input all of the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

[0084] Specifically, a plurality of target operation images to be detected are obtained, all of the target operation images are input into the target chain network model, and a plurality of high-quality initial feature maps are output; all of the high-quality initial feature maps are input into the target attention training model, and a plurality of the high-quality feature maps are output.

[0085] Among them, the target running image is input into the target chain network model, and multiple high-quality initial feature maps are output. Then, the multiple high-quality initial feature maps are input into the target attention training model, and multiple high-quality feature maps are output. According to different types of images, high-quality feature maps corresponding to the type can be output, thereby obtaining more accurate detection results under various lighting conditions.

[0086] Furthermore, when extracting relevant features of the image (including basic feature information, complex feature information and abstract feature information), these relevant features can also be combined with other sensor data (such as temperature, humidity and radiation intensity, etc.) to build a multimodal data fusion photovoltaic detection system. By analyzing data from different data sources and displaying the operation status of the photovoltaic system through charts or heat maps, it helps operation and maintenance personnel to understand the detection status in real time, so as to improve the accuracy and robustness of photovoltaic system fault detection. For example, combined with thermal imaging data, hot spots and electrical faults can be accurately detected.

[0087] Furthermore, when conducting photovoltaic inspections, blockchain technology is introduced to record and track the operating data and fault information of the photovoltaic system. The tamper-proof nature of blockchain ensures the authenticity and integrity of the data, improves the transparency and credibility of the inspection system, and provides reliable data support for subsequent operation and maintenance and analysis. When fault information is detected, historical data and predictive models are combined to provide specific maintenance recommendations and optimization strategies to predict possible future failures, making it easier for operation and maintenance personnel to arrange maintenance in advance.

[0088] The present invention introduces a chain network model and an attention mechanism into a deep learning network to process images of photovoltaic modules, quickly and accurately detect common problems such as photovoltaic panel shading, hot spot effect, cracks and inverter failure, uses images for model training, learns and identifies various fault characteristics, and realizes efficient automatic detection. In addition, a drone is used as an image acquisition device. The drone is equipped with a visible light camera and an infrared camera to take aerial photos of the photovoltaic power station. The obtained high-definition images are used as the input of the detection system. The improved system learns and identifies images to finally complete the detection of defects in the photovoltaic power station.

[0089] Furthermore, if Fig. 9 As shown, based on the above photovoltaic detection optimization method based on deep learning, the present invention also provides a photovoltaic detection optimization system based on deep learning, wherein the photovoltaic detection optimization system based on deep learning includes:

[0090] An image preprocessing module 51 is used to construct a chain network model, integrate the chain network model into the backbone network, obtain multiple historical operation images of photovoltaic detection targets, input the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and output multiple initial feature maps;

[0091] A feature map weighting module 52 is used to construct an attention training model, integrate the attention training model into the backbone network, input all the initial feature maps into the attention training model for weighted processing, and obtain multiple target feature maps;

[0092] A loss weighting module 53 is used to input all the target feature maps into the attention training model for training, output multiple loss values ​​of all the target feature maps, and perform weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples;

[0093] A model training module 54 is used to screen all the training samples according to a preset standard to obtain a plurality of target training samples, and input the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model;

[0094] The image detection module 55 is used to obtain multiple target operation images to be detected, input all the target operation images into the target chain network model, output multiple high-quality initial feature maps, input all the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

[0095] Furthermore, if Fig.10 As shown, based on the above-mentioned photovoltaic detection optimization method and system based on deep learning, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig.10 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0096] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a photovoltaic detection optimization program 40 based on deep learning is stored on the memory 20, and the photovoltaic detection optimization program 40 based on deep learning can be executed by the processor 10, thereby realizing the photovoltaic detection optimization method based on deep learning in the present application.

[0097] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the deep learning-based photovoltaic detection optimization method.

[0098] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0099] In one embodiment, when the processor 10 executes the deep learning-based photovoltaic detection optimization program 40 in the memory 20 , the steps of the deep learning-based photovoltaic detection optimization method described above are implemented.

[0100] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a photovoltaic detection optimization program based on deep learning, and when the photovoltaic detection optimization program based on deep learning is executed by a processor, the steps of the photovoltaic detection optimization method based on deep learning as described above are implemented.

[0101] In summary, the present invention provides a photovoltaic detection optimization method based on deep learning and related equipment, the method comprising: constructing a chain network model, integrating the chain network model into a backbone network, acquiring multiple historical operation images of photovoltaic detection targets, inputting multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and outputting multiple initial feature maps; constructing an attention training model, integrating the attention training model into the backbone network, inputting all the initial feature maps into the attention training model for weighted processing, and outputting multiple target feature maps; inputting all the target feature maps into the attention training model for training, and outputting all the target feature maps. A plurality of loss values ​​of a target feature map, weighting all the target feature maps according to all the loss values ​​to obtain a plurality of training samples; screening all the training samples according to a preset standard to obtain a plurality of target training samples, and inputting the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model; obtaining a plurality of target running images to be detected, inputting all the target running images into the target chain network model, outputting a plurality of high-quality initial feature maps, inputting all the high-quality initial feature maps into the target attention training model, and outputting a plurality of high-quality feature maps. The present invention introduces a chain network model and an attention mechanism into a deep learning network to process images of photovoltaic modules, quickly and accurately detect common problems such as photovoltaic panel shading, hot spot effect, cracks and inverter failure, uses images for model training, learns and identifies various fault characteristics, and realizes efficient automatic detection. In addition, a drone is used as an image acquisition device. The drone is equipped with a visible light camera and an infrared camera to take aerial photos of the photovoltaic power station. The obtained high-definition images are used as the input of the detection system. The improved system learns and identifies images to finally complete the detection of defects in the photovoltaic power station.

[0102] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal 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 terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0103] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0104] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A photovoltaic detection optimization method based on deep learning, characterized in that: The photovoltaic detection optimization method based on deep learning includes: Constructing a chain network model, integrating the chain network model into a backbone network, acquiring multiple historical operation images of a photovoltaic detection target, inputting the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and outputting multiple initial feature maps; Constructing an attention training model, integrating the attention training model into the backbone network, inputting all the initial feature maps into the attention training model for weighted processing, and outputting a plurality of target feature maps; Inputting all the target feature maps into the attention training model for training, outputting multiple loss values ​​of all the target feature maps, and performing weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples; Screening all the training samples according to a preset standard to obtain a plurality of target training samples, and inputting the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model; Acquire multiple target running images to be detected, input all the target running images into the target chain network model, output multiple high-quality initial feature maps, input all the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

2. The photovoltaic detection optimization method based on deep learning according to claim 1, characterized in that: The chain network model is constructed, the chain network model is integrated into the backbone network, a plurality of historical operation images of the photovoltaic detection target are obtained, the plurality of historical operation images are input into the chain network model for feature extraction and normalization adjustment, and a plurality of initial feature maps are output, specifically including: Constructing a chain network model and integrating the chain network model into the backbone network; Acquire multiple historical operation images input by a user, and input all of the historical operation images into the chain network model; Extracting multiple basic feature information of all the historical running images in the initial convolutional layer of the chain network model; Extracting multiple complex feature information of all the historical operation images in the middle layer of the chain network model; Extracting multiple abstract feature information of all the historical running images in the deep convolutional layer of the chain network model; By using image preprocessing technology, all the basic feature information, all the complex feature information and all the abstract feature information are standardized and a plurality of the initial feature maps are output.

3. The photovoltaic detection optimization method based on deep learning according to claim 1, characterized in that: The step of constructing an attention training model, integrating the attention training model into the backbone network, inputting all the initial feature maps into the attention training model for weighted processing, and outputting a plurality of target feature maps specifically includes: Constructing an attention training model and integrating the attention training model into the backbone network; Input all the initial feature maps into the attention training model, and calculate the importance weight value of each feature pixel in each initial feature map; According to each of the initial feature maps and the corresponding importance weight values, generating a weighted feature map corresponding to each of the initial feature maps; Each of the weighted feature maps is fused with the corresponding initial feature map, and a target feature map corresponding to each of the initial feature maps is output.

4. The photovoltaic detection optimization method based on deep learning according to claim 1, characterized in that: The step of inputting all the target feature maps into the attention training model for training, outputting multiple loss values ​​of all the target feature maps, and performing weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples specifically includes: Inputting all the target feature maps into the attention training model for training, generating multiple prediction boxes corresponding to all the target feature maps, and obtaining multiple real boxes in the running image corresponding to all the target feature maps; Generate a loss value for each target feature map according to all predicted frames and all corresponding real frames of each target feature map, where the loss value is used to measure the degree of overlap between each feature map and the corresponding real frame; Evaluate all the target feature maps according to all the loss values ​​to obtain an evaluation result; According to the evaluation result and each loss value, each target feature map is weighted to obtain a training sample corresponding to each target feature map.

5. The photovoltaic detection optimization method based on deep learning according to claim 4 is characterized in that: The step of performing weighted processing on each of the target feature maps according to the evaluation result and each of the loss values ​​to obtain a training sample corresponding to each of the target feature maps specifically includes: Determine the quality weight value corresponding to each loss value according to the evaluation result and preset weight mapping information; Multiply each of the quality weight values ​​by the corresponding loss value to obtain a target weight value corresponding to each of the target feature maps; Assign each of the target weight values ​​to the corresponding target feature map to obtain a training sample corresponding to each of the target feature maps.

6. The photovoltaic detection optimization method based on deep learning according to claim 5, characterized in that: The preset standards include: clarity standard, lighting condition standard, annotation accuracy standard, data distribution standard and noise level; The method of screening all the training samples according to the preset standard to obtain a plurality of target training samples, and inputting the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model, specifically includes: Determining, according to the target weight value of each training sample, whether each training sample meets the clarity standard, the lighting condition standard, the labeling accuracy standard, the data distribution standard or the noise level; If the clarity standard, the lighting condition standard, the annotation accuracy standard, the data distribution standard or the noise level are met, all the corresponding training samples are respectively input into the chain network model and the attention training model, and the chain network model and the attention training model are optimized to obtain the target chain network model and the target attention training model.

7. The photovoltaic detection optimization method based on deep learning according to claim 1, characterized in that: The step of acquiring a plurality of target operation images to be detected, inputting all the target operation images into the target chain network model, outputting a plurality of high-quality initial feature maps, inputting all the high-quality initial feature maps into the target attention training model, and outputting a plurality of high-quality feature maps specifically includes: Acquire multiple target running images to be detected, input all the target running images into the target chain network model, and output multiple high-quality initial feature maps; All of the high-quality initial feature maps are input into the target attention training model, and a plurality of the high-quality feature maps are output.

8. A photovoltaic detection optimization system based on deep learning, characterized in that: The photovoltaic detection optimization system based on deep learning includes: An image preprocessing module is used to construct a chain network model, integrate the chain network model into the backbone network, obtain multiple historical operation images of photovoltaic detection targets, input the multiple historical operation images into the chain network model for feature extraction and normalization adjustment, and output multiple initial feature maps; A feature map weighting module is used to construct an attention training model, integrate the attention training model into the backbone network, input all the initial feature maps into the attention training model for weighted processing, and obtain multiple target feature maps; A loss weighting module, used for inputting all the target feature maps into the attention training model for training, outputting multiple loss values ​​of all the target feature maps, and performing weighted processing on all the target feature maps according to all the loss values ​​to obtain multiple training samples; A model training module, used to screen all the training samples according to a preset standard to obtain a plurality of target training samples, and input the plurality of target training samples into the chain network model and the attention training model for optimization training to obtain a target chain network model and a target attention training model; The image detection module is used to obtain multiple target running images to be detected, input all the target running images into the target chain network model, output multiple high-quality initial feature maps, input all the high-quality initial feature maps into the target attention training model, and output multiple high-quality feature maps.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a deep learning-based photovoltaic detection optimization program stored in the memory and executable on the processor. When the deep learning-based photovoltaic detection optimization program is executed by the processor, the steps of the deep learning-based photovoltaic detection optimization method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a photovoltaic detection optimization program based on deep learning, and when the photovoltaic detection optimization program based on deep learning is executed by a processor, the steps of the photovoltaic detection optimization method based on deep learning are implemented as described in any one of claims 1 to 7.

Images