Photovoltaic power equipment fault detection method and system based on unmanned aerial vehicle

Abstract

The invention provides a photovoltaic power equipment fault detection method and system based on an unmanned aerial vehicle, and belongs to the technical field of photovoltaic operation and maintenance and image intelligent recognition. The method comprises the following steps: carrying out aerial photography on a photovoltaic power module through an unmanned aerial vehicle to obtain an infrared image; the collected image is preprocessed to enhance fault features, and labeling processing is carried out; carrying out model training by adopting an improved Faster R-CNN algorithm in combination with transfer learning; deploying the trained model to an edge calculation unit of the unmanned aerial vehicle; finally, the real-time detection of the fault of the photovoltaic module is realized. According to the method, the hot spot defect in the photovoltaic equipment can be effectively identified, the multi-fault target detection precision is improved, the method has high environmental adaptability and deployment flexibility, and powerful technical support can be provided for intelligent operation and maintenance of a photovoltaic power system and an unmanned aerial vehicle edge computing platform.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of photovoltaic operation and maintenance and intelligent image recognition, and particularly relates to a photovoltaic power equipment fault detection method and system based on a UAV, and is particularly suitable for large-area inspection and intelligent diagnosis of outdoor photovoltaic modules. BACKGROUND

[0002] Under the background of global energy transformation, developing a clean energy system that is coordinated with environmental sustainability has become a key task. As a typical renewable energy, photovoltaic power generation has attracted much attention due to its clean and sustainable advantages. With the widespread deployment of photovoltaic power generation systems, its operation and fault detection problems have become increasingly prominent. Since photovoltaic modules are exposed to the outdoor environment for a long time, they face multiple challenges such as extreme weather and dust accumulation, which can easily cause various faults. During operation, hot spot effects often occur due to environmental obstruction and local damage, which can affect the overall power generation efficiency and even cause safety hazards. Traditional manual detection is low in efficiency and high in missed detection rate, while existing image recognition-based hot spot detection methods have low recognition rates in complex backgrounds, especially for small and low-contrast targets, and there is still much room for improvement in detection effect.

[0003] Current mainstream target detection algorithms have certain limitations when processing small target detection tasks. For example, single-stage detectors are a series of algorithms that have an advantage in detection speed but perform poorly in detection accuracy. In contrast, the two-stage detection framework Faster R-CNN has a clear advantage in detection accuracy due to its end-to-end training method, and is more suitable for the requirements of hot spot recognition accuracy in photovoltaic operation and maintenance. However, its performance still has the following key problems: the hot spot target in aerial images usually only occupies tens of pixels, and the continuous downsampling operation of the deep convolutional network causes severe loss of small target feature information; secondly, the reflection and shadow interference of the metal frame on the surface of the photovoltaic module are highly similar to the visual features of the real hot spot, increasing the risk of false detection; in addition, the traditional Intersection over Union (IOU) loss function is not accurate enough in regressing the bounding box of irregular hot spot regions. At the same time, to adapt to the processing capability of the UAV edge computing platform, the feature extraction network determines the real-time processing performance.

[0004] Although existing deep learning methods have achieved certain results, most of them are applied to offline server detection, and therefore a comprehensive UAV detection process for acquisition, training, and deployment needs to be developed. Therefore, there is a need in the art for a detection method and system that solves the above problems.

[0005] Therefore, there is a need in the art for a detection method and system that solves the above problems. SUMMARY

[0006] The application provides a photovoltaic power equipment fault detection method and system based on a UAV, which comprises the following steps: taking a photograph of a photovoltaic power module by using a UAV to obtain an infrared image; pre-processing the collected image to enhance the fault features and performing labeling processing; training a model by using an improved Faster R-CNN algorithm combined with transfer learning; deploying the trained model to an edge computing unit of the UAV; and finally realizing real-time detection of photovoltaic module faults. The method can effectively identify hot spot defects in photovoltaic equipment, improve the multi-fault target detection accuracy, has strong environmental adaptability and deployment flexibility, and can provide strong technical support for intelligent operation and maintenance of a photovoltaic power system and a UAV edge computing platform.

[0007] A photovoltaic power equipment fault detection method based on a UAV, characterized in that the method comprises the following steps:

[0008] S1. Taking a photograph of a photovoltaic power module by using a UAV to obtain infrared image data.

[0009] S2. Pre-processing the collected infrared image to enhance the fault features and labeling the image.

[0010] S3. Training a hot spot fault recognition model by using an improved Faster R-CNN algorithm based on the pre-processed image data and transfer learning.

[0011] S4. Deploying the trained hot spot fault recognition model to an edge computing unit of the UAV.

[0012] S5. Real-time fault detection of the photovoltaic power module by using a UAV system loaded with the edge computing unit, identifying hot spot defects, and returning to a ground station.

[0013] The image pre-processing in step S1 comprises denoising, contrast enhancement, and the like, and is fused with visible light image data features to improve the image quality and feature extraction effect.

[0014] In the feature labeling step in step S2, the image data collected in S1 is labeled by using an artificial labeling method or a semi-automatic labeling algorithm, and the hot spot types of the photovoltaic module are mainly classified into the following categories: large-area hot spot, single-point small hot spot, abnormally low temperature, diode short circuit, and mixed hot spot defect, which are used to construct a training data set.

[0015] The candidate frame size optimization and main network improvement in the Faster R-CNN algorithm in step S3 are as follows: in view of the diversity of the size of the hot spot target of the photovoltaic module, the K Means clustering algorithm is used to analyze the width-height ratio and area distribution of the labeled frame in the training set, the IOU of the anchor frame and the real boundary frame is calculated as the clustering distance measurement, and finally the anchor frame size configuration and number suitable for small target hot spot detection are generated.

[0016] The main network of the Faster R-CNN algorithm in step S3 adopts Resnet50 to replace the original VGG16 network, and the deeper network structure and residual connection mechanism of Resnet50 are used to effectively extract multi-scale features of different types of fault hot spots, such as point hot spots, large-area hot spots, and associated low-temperature area hot spots, and especially enhance the feature expression ability of small abnormal areas.

[0017] The training process adopts a transfer learning strategy, and parameters are fine-tuned based on a pre-trained model to speed up the convergence of the model and improve the detection accuracy.

[0018] Step S4 includes that the edge computing unit includes an embedded processor deployed on the unmanned aerial vehicle, and has model inference and data storage capabilities.

[0019] Step S5 includes that the edge computing unit also integrates model dynamic loading, temperature compensation algorithm and data caching functions, and the detection result is transmitted in a structured format to the ground monitoring system in real time through a wireless module.

[0020] The present application realizes accurate detection of the hot spot of the photovoltaic module under complex lighting and background interference conditions by combining deep learning and edge computing, has the advantages of high accuracy, fast response and strong adaptability, and is suitable for unmanned operation and maintenance scenes of modern intelligent power stations. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the method of the present application.

[0022] Figure 2 is a schematic diagram of the improved Faster R-CNN structure. DETAILED DESCRIPTION

[0023] In order for those skilled in the art to better understand the technical solutions of the present application, the photovoltaic power equipment fault detection method and system based on an unmanned aerial vehicle provided by the present application are described in detail below with reference to the embodiments. The following embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0024] In one embodiment, as Figure 1As shown, a power inspection robot detection and deployment method, the method a kind of unmanned aerial vehicle-based photovoltaic power equipment fault detection method and system includes the following steps.

[0025] Step S1: using the unmanned aerial vehicle system built, using the infrared sensor and visible light camera carried out aerial photography on photovoltaic power module, obtains infrared and image data;Wherein the image preprocessing includes denoising, contrast enhancement and other operations, while the infrared image and visible light image data are fused to improve image quality and feature extraction effect.

[0026] Step S2: feature labeling step through the image data collected in S1, manually labeled or using semi-automatic labeling algorithm, the hot spot type of photovoltaic module is mainly divided into: large area hot spot, single point small hot spot, abnormal low temperature, diode short circuit and mixed hot spot defect and the like, for building diversified training data set.

[0027] Step S3: as shown in Figure 2 The candidate frame size optimization and main network improvement in Faster R-CNN algorithm, specifically: for the diversity of photovoltaic module hot spot target size, the K Means clustering algorithm is used to analyze the width-height ratio and area distribution of the labeled frame in the training set. Take the IOU of anchor frame and real boundary frame as the clustering distance measure, randomly select K initial anchor frames, calculate the IOU of each real boundary frame and all initial anchor frames, and assign it to the anchor frame category with the maximum IOU. Then, recalculate the mean of the width-height ratio and area of all boundary frames in each category, and update the anchor frame size. After several iterations, the anchor frame size converges. In the photovoltaic thermal imaging data set with high proportion of small target hot spots, the K value is determined through multiple experiments, and finally the anchor frame size configuration suitable for small target hot spot detection is generated, covering different width-height ratio and area range, effectively improving the recall rate and detection accuracy of small target hot spots.

[0028] In the training process, first, freeze the front several layers of Resnet50 main network, only train the subsequent layers and RPN region suggestion network, Faster R-CNN detection head, to utilize the low-level visual feature extraction capability of the pre-trained model; then gradually unfreeze part of the layer parameters for fine-tuning in the later training, balance the model generalization ability and specific task adaptability by setting different learning rates, to speed up the model convergence speed and improve the detection accuracy.

[0029] Step S4: The network model trained by Matlab or Pytroch software platform can integrate the infrared sensor, visible light sensor, deep learning algorithm model and image preprocessing algorithm through the Simulink platform to build a UAV photovoltaic inspection system. Among them, the infrared sensor collects thermal imaging data in real time through the GigE interface, integrates the temperature correction sub-module to compensate the influence of environmental temperature and emissivity on temperature measurement accuracy; the visible light sensor collects RGB images through the CSI-2 interface, and is matched with lens distortion correction and ROI extraction function. The image preprocessing link is built in Simulink to construct a cascade pipeline, including infrared image non-uniformity correction, histogram equalization, visible light image distortion compensation and component ROI segmentation, and the thermal imaging and visible light image sub-pixel level registration fusion is realized through ORB feature matching. In terms of deep learning algorithm model, the optimized Faster R-CNN model is deployed in ELF format, the model volume is compressed, and single-frame infrared image real-time inference is realized on Jetson embedded processor to meet the real-time requirement of UAV inspection.

[0030] Step S5: The UAV fault detection result is stored in JSON format, and the specific fields include hot spot type, center point coordinates, fault module labeled image. The detection result is transmitted to the ground monitoring system in real time through the wireless module, the transmission protocol uses UDP, and the transmission delay is controlled within 200ms. The ground monitoring system is developed based on LabView, which can display the UAV inspection trajectory and hot spot distribution map in real time, and supports historical data query, trend analysis and other functions. In addition, the system has intelligent alarm function, which pushes SMS notification to operation and maintenance personnel, and marks the fault point on the map.

[0031] The application can be widely applied to fault diagnosis, state monitoring and intelligent operation and maintenance in photovoltaic power generation scene, and provides efficient and intelligent technical support for green energy system.

Claims

1. A method and system for detecting faults in photovoltaic power equipment based on unmanned aerial vehicles, characterized by, The method comprises the following steps: S1. Using a UAV to take aerial photos of a photovoltaic power module to obtain infrared image data; S2. Preprocessing the collected infrared image to enhance the fault features and label the image; S3. Based on the preprocessed image data, using an improved Faster R-CNN algorithm for transfer learning training to construct a hot spot fault recognition model; S4. Deploying the trained hot spot fault recognition model to the edge computing unit of the UAV; S5. Using the UAV system carrying the edge computing unit to perform real-time fault detection on the photovoltaic power module, identify the hot spot defects, and return to the ground station.

2. The method of claim 1, wherein: The image preprocessing includes denoising, contrast enhancement, and the like to improve the image quality and feature extraction effect.

3. The method of claim 1, wherein, S2 comprises: The image data collected by S1 is labeled as manual labeling or semi-automatic labeling to construct a training data set.

4. The method of claim 1, wherein, S3 comprises: The size of the candidate box in the Faster R-CNN algorithm is optimized by the K Means clustering algorithm to improve the detection accuracy of small target hot spots; the main network uses Resnet50 to improve the detection accuracy of different types of fault hot spots.

5. The method of claim 1, wherein, S3 comprises: The training process uses a transfer learning strategy to fine-tune the parameters based on a pre-trained model to speed up the model convergence and improve the detection accuracy.

6. The method of claim 1, wherein, S4 comprises: The edge computing unit includes an embedded processor deployed on the UAV, which has model inference and data storage capabilities.

7. The method of claim 1, wherein, S5 comprises: The fault detection results include hot spot type, area, and temperature threshold determination results, and can be transmitted in real time to the ground monitoring system through wireless means.

Images