Photovoltaic power station unmanned aerial vehicle automatic inspection system integrated with machine vision

Abstract

The invention relates to the technical field of photovoltaic power station inspection, and discloses a photovoltaic power station unmanned aerial vehicle automatic inspection system integrated with machine vision. An image data acquisition module of the system captures a photovoltaic module surface image sequence through unmanned aerial vehicle camera equipment to form initial image data; the fault prediction module processes the initial image data and outputs a potential fault signal and a performance degradation signal; the path optimization module combines the initial image data and the potential fault signal to generate an optimal inspection track signal; the adaptive control module defines a flight constraint condition according to the initial image data and outputs a safe flight threshold value; and the unmanned aerial vehicle cooperation module comprises an automatic repair execution unit, a flight control unit and an interactive early warning platform which are respectively responsible for adjusting flight parameters and attitude parameters and processing signal output comprehensive risk indexes. Data interaction is realized among the modules, a whole cooperative work is formed, and the intelligent and automatic level of photovoltaic power station inspection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power station inspection, in particular to an unmanned aerial vehicle automatic inspection system for photovoltaic power station integrated with machine vision. BACKGROUND

[0002] With the transformation of global energy structure to clean energy, the installed capacity of photovoltaic power stations has shown explosive growth. The operation and maintenance of large-scale photovoltaic arrays has become a key link to ensure power generation efficiency. Photovoltaic modules are exposed to outdoor environments for a long time, and are easily affected by problems such as dust coverage, hot spot effect, glass breakage, junction box failure, etc. If not detected and repaired in time, it will lead to a significant decrease in power generation efficiency, and even cause safety accidents. Therefore, efficient and accurate photovoltaic module inspection technology has become a research hotspot in the industry.

[0003] Traditional photovoltaic power station inspection mainly relies on manual methods. Inspection personnel observe photovoltaic panel status by naked eye or handheld devices. This method has obvious limitations: on the one hand, large photovoltaic power stations are often distributed in complex terrains such as deserts and plateaus. Manual inspection has high labor intensity and low efficiency. The daily inspection area is usually less than 50,000 square meters, which is difficult to meet the operation and maintenance needs of large-scale power stations. On the other hand, manual judgment is easily affected by subjective experience, environmental light, etc. The identification accuracy rate of early faults such as subtle hot spots and hidden cracks is less than 60%, which easily causes missed inspection or misjudgment. In addition, manual inspection under extreme weather conditions also has high safety risks, which further restricts the timeliness of operation and maintenance.

[0004] To improve the inspection efficiency, some power stations have introduced an unmanned aerial vehicle-based inspection scheme. High-definition cameras are carried to obtain photovoltaic module images, which are then analyzed by manual post-processing. Although this method improves the inspection efficiency by 3-5 times, it still has technical bottlenecks: first, the flight path of the unmanned aerial vehicle is mostly preset fixed trajectory, which cannot be dynamically adjusted according to the actual state of the module, easily causing repeated shooting or missed inspection; second, image data relies on manual judgment, which is not only time-consuming (2-3 hours are needed for a single megawatt power station), but also lacks consistency in fault identification (the accuracy rate difference between different personnel is 20%-30%); third, there is a lack of real-time warning mechanism. The period from fault discovery to processing is as long as several days, during which the loss of power generation cannot be ignored.

[0005] In recent years, the development of machine vision and artificial intelligence technology provides a new direction for photovoltaic inspection, and related research attempts to automatically detect component failures through image recognition algorithms. However, existing systems focus on single fault type recognition, and lack comprehensive detection capability for multiple defects (such as hidden cracks, snail patterns, and ablation), and have not formed a closed-loop system of "data collection-fault prediction-path optimization-autonomous control". At the same time, the flight safety of unmanned aerial vehicles in complex environments, the coordination of multi-vehicle cooperative operation and other issues have not been effectively solved, resulting in that the system stability and robustness in practical application cannot meet the operation and maintenance requirements of large-scale power stations.

[0006] Under this background, a photovoltaic power station unmanned aerial vehicle automatic inspection system integrated with machine vision is developed to realize the full-process automation from image collection, intelligent analysis to autonomous decision-making and cooperative operation, which is of great significance for improving the operation and maintenance efficiency of photovoltaic power stations, reducing power generation loss, and ensuring system safety. SUMMARY

[0007] The purpose of the present application is to provide a photovoltaic power station unmanned aerial vehicle automatic inspection system integrated with machine vision to solve the problems raised in the above background.

[0008] To achieve the above purpose, the present application provides a photovoltaic power station unmanned aerial vehicle automatic inspection system integrated with machine vision, which comprises:

[0009] An image data acquisition module for capturing a sequence of surface images of photovoltaic components by a camera device carried by an unmanned aerial vehicle to form initial image data;

[0010] A fault prediction module for receiving and processing the initial image data and outputting potential fault signals and performance degradation signals;

[0011] A path optimization module for processing the initial image data and the potential fault signals and generating an optimal inspection trajectory signal;

[0012] An adaptive control module for defining flight constraint conditions according to the initial image data and outputting a safe flight threshold;

[0013] An unmanned aerial vehicle cooperation module including an automatic repair execution unit, a flight control unit and an interactive warning platform, the automatic repair execution unit adjusts the flight parameters of the unmanned aerial vehicle according to the optimal inspection trajectory signal, the flight control unit adjusts the attitude parameters of the unmanned aerial vehicle according to the safe flight threshold and the initial image data, and collects real-time monitoring signals, and the interactive warning platform processes the performance degradation signals and the potential fault signals and outputs a comprehensive risk index.

[0014] The image data acquisition module transmits the captured data to the fault prediction module, the fault prediction module transmits the prediction result to the path optimization module, the path optimization module feeds back the optimization signal to the adaptive control module, and the adaptive control module sends the constraint condition to the unmanned aerial vehicle cooperation module, so as to realize data interaction between the modules of the system.

[0015] Preferably, the image data acquisition module includes a thermal imaging sensor unit and a spectral analysis unit, is arranged above the hovering position of the unmanned aerial vehicle and the photovoltaic array, and collects infrared thermal image signals, the spectral analysis unit obtains spectral reflectance data of the photovoltaic module through near-infrared spectral technology and forms spectral feature signals, the image data acquisition module integrates the infrared thermal image signals and the spectral feature signals to realize multi-modal data fusion and form enhanced image data, and the enhanced image data includes prediction enhanced image data, optimization enhanced image data and regulation enhanced image data.

[0016] Preferably, the fault prediction module includes a convolutional neural network-based image classification model and a Bayesian network model, the prediction enhanced image data includes surface dirt distribution data, hot spot abnormal data, spectral reflectance data, historical maintenance records and historical efficiency attenuation data, the surface dirt distribution data, the historical maintenance records, the hot spot abnormal data, the spectral reflectance data and the historical efficiency attenuation data are input into the image classification model and output the potential fault signal, and the surface dirt distribution data, the hot spot abnormal data and the spectral reflectance data are fused through the improved Bayesian network model and output the performance degradation signal.

[0017] Preferably, the image classification model includes a feature extraction submodule, a multi-source fusion submodule, a risk assessment submodule and a result display module, the feature extraction submodule performs edge detection algorithm processing on the hot spot abnormal data, extracts a region meeting the crack propagation feature range as a crack precursor signal, the multi-source fusion submodule uses an attention mechanism weighted convolutional neural network to process the surface dirt distribution data, the hot spot abnormal data and the spectral reflectance data and outputs the data to the risk assessment submodule, the risk assessment submodule defines a risk quantization method, calculates a risk index using a maximum value of dirt coverage, an average value of hot spot temperature and an average value of spectral reflectance, and the result display module generates a dynamic heat map of the risk index based on a color gradient and uses different color steps to represent different risk levels of the photovoltaic module positions.

[0018] Preferably, the path optimization module processes the optimization enhanced image data and the potential fault signal through a trajectory-fault nonlinear relationship model, outputs a preliminary flight signal, and processes the preliminary flight signal through a particle swarm optimization algorithm to obtain the optimal inspection trajectory signal; the optimization enhanced image data includes component tilt angle data and unmanned aerial vehicle power parameters, the trajectory-fault nonlinear relationship model is constructed through the component tilt angle data, the potential fault signal and the unmanned aerial vehicle power parameters, and the trajectory-fault nonlinear relationship model describes the relevance of flight path length and dirt coverage degree, hot spot density and unmanned aerial vehicle speed.

[0019] Preferably, the path optimization module further includes a performance matching database, a network connectivity analysis submodule and a trajectory planning submodule, the performance matching database stores historical power parameters of historical unmanned aerial vehicle flights, the network connectivity analysis submodule calculates a connectivity coefficient of the photovoltaic array according to the unmanned aerial vehicle power parameters and the flight path length, and the trajectory planning submodule obtains the optimal inspection trajectory signal through a particle swarm optimization algorithm according to the connectivity coefficient, the trajectory-fault nonlinear relationship model and the unmanned aerial vehicle power parameters.

[0020] Preferably, the adaptive control module defines a flight constraint condition according to the regulation enhanced image data, outputs a safe flight threshold, and dynamically adjusts the safe flight threshold to be not lower than a safe flight minimum value through fuzzy logic, the regulation enhanced image data includes a real-time wind speed signal and a terrain height signal, the flight constraint condition is defined by the real-time wind speed signal and the terrain height signal, the safe flight threshold is calculated based on a minimum flight height, an average wind speed extreme value and a terrain undulation degree, and the fuzzy logic dynamically adjusts an unmanned aerial vehicle hovering height and a flight speed to make the safe flight threshold not lower than the safe flight minimum value.

[0021] Preferably, the interactive early warning platform includes a three-dimensional visualization interface and a risk early warning engine, the three-dimensional visualization interface receives the enhanced image data and the real-time monitoring signal and dynamically renders a photovoltaic component fault diffusion process, the risk early warning engine processes the performance degradation signal and the potential fault signal according to an improved technique for order preference by similarity to ideal solution algorithm and outputs a comprehensive risk index, when the comprehensive risk index exceeds a preset risk threshold, the interactive early warning platform controls the automatic repair execution unit to perform emergency repair, the three-dimensional visualization interface supports arbitrary viewing angle profile analysis, real-time display of the spatial relationship between the fault area and the photovoltaic array, and simulation of component stress distribution under different flight schemes based on a finite element method.

[0022] Preferably, the risk warning engine comprises color identification of multiple levels of warning states, namely a yellow warning state, an orange warning state and a red warning state, when the efficiency degradation value in the potential failure signal falls to a first degradation threshold, the yellow warning state is started, and the flight speed of the unmanned aerial vehicle is reduced to an alert speed, when the efficiency degradation value in the potential failure signal falls to a second degradation threshold, the unmanned aerial vehicle is suspended and the automatic repair execution unit is controlled to perform emergency repair, and when the efficiency degradation value in the potential failure signal falls to a third degradation threshold, the risk warning engine sends an emergency failure signal and controls the integrated machine vision photovoltaic power station unmanned aerial vehicle automatic inspection system to stop urgently.

[0023] Preferably, the optimal inspection trajectory signal comprises inspection area coordinates, unmanned aerial vehicle type priority, flight speed range and path planning scheme, the automatic repair execution unit comprises an intelligent flight controller and an integrated temperature monitoring device, the intelligent flight controller can identify the optimal inspection trajectory signal and adjust the flight height and flight angle as needed, and the integrated temperature monitoring device receives the optimal inspection trajectory signal and adjusts the unmanned aerial vehicle cooling strategy in real time based on infrared temperature measurement technology, and the real-time monitoring signal comprises device temperature, device humidity, device voltage, flight height, component pressure and flight distance.

[0024] Compared with the prior art, the present application has the following advantages:

[0025] The integrated machine vision photovoltaic power station unmanned aerial vehicle automatic inspection system captures a sequence of surface images of the photovoltaic module by using the camera equipment carried by the unmanned aerial vehicle through the image data acquisition module, forms initial image data, and provides an original basis for subsequent failure analysis and path optimization. The failure prediction module processes the initial image data and outputs potential failure signals and performance degradation signals, which can identify possible problems of the photovoltaic module at an early stage and facilitate timely response measures.

[0026] The path optimization module generates an optimal inspection trajectory signal in combination with the initial image data and the potential failure signal, so that the inspection path of the unmanned aerial vehicle can be dynamically adjusted according to the actual state of the component, avoiding the limitations of traditional preset routes, reducing invalid flight, and making the inspection process more efficient. The adaptive control module defines flight constraint conditions and outputs safety flight thresholds according to the initial image data, which provides protection for the safe flight of the unmanned aerial vehicle and ensures that the unmanned aerial vehicle can operate stably in a complex power station environment, reducing the flight risk caused by environmental factors.

[0027] The automatic repair execution unit in the unmanned aerial vehicle cooperation module adjusts the unmanned aerial vehicle flight parameters according to the optimal inspection track signal, so that the unmanned aerial vehicle can accurately inspect according to the optimized path; the flight control unit adjusts the unmanned aerial vehicle attitude parameters according to the safe flight threshold and the initial image data, and collects real-time monitoring signals, thereby enhancing the adaptability and controllability of the unmanned aerial vehicle in the flight process; the interactive early warning platform processes the performance degradation signal and the potential fault signal and outputs a comprehensive risk index, thereby realizing comprehensive evaluation of the overall operation state of the photovoltaic power station and facilitating relevant personnel to comprehensively understand the power station.

[0028] Data interaction is realized between the modules of the system, the image data acquisition module, the fault prediction module, the path optimization module, the adaptive control module and the unmanned aerial vehicle cooperation module form an organic whole, so that the image acquisition, fault analysis, path adjustment, flight control and other links are closely coordinated, the coherence and cooperation of the entire inspection process are improved, and the inspection work of the photovoltaic power station is more intelligent and automated. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A working principle diagram of the integrated machine vision photovoltaic power station unmanned aerial vehicle automatic inspection system is provided.

[0030] Figure 2 A flowchart of multi-modal data fusion is provided.

[0031] Figure 3 A flowchart of image classification model risk assessment is provided.

[0032] Figure 4 A flowchart of path optimization module submodule cooperation is provided. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0034] Please refer to Figure 1 The present application provides an integrated machine vision photovoltaic power station unmanned aerial vehicle automatic inspection system, which realizes intelligent inspection and fault management of photovoltaic modules through multi-module cooperation. The system comprises an image data acquisition module, a fault prediction module, a path optimization module, an adaptive control module and an unmanned aerial vehicle cooperation module.

[0035] The image data acquisition module is mounted on the unmanned aerial vehicle and captures a sequence of surface images of the photovoltaic module through a camera device to form initial image data.

[0036] The fault prediction module receives the initial image data, processes through a convolutional neural network and a Bayesian network model, and outputs potential fault signals and performance degradation signals.

[0037] The path optimization module combines the initial image data and the potential fault signals, and generates an optimal inspection trajectory signal using a particle swarm optimization algorithm.

[0038] The adaptive control module defines flight constraints based on the initial image data and outputs dynamically adjusted safe flight thresholds.

[0039] The UAV coordination module includes an automatic repair execution unit, a flight control unit, and an interactive warning platform. The automatic repair execution unit adjusts flight parameters based on the optimal inspection trajectory signal. The flight control unit regulates the UAV attitude in combination with the safe flight threshold and real-time monitoring signals. The interactive warning platform integrates performance degradation signals and potential fault signals, outputs three-dimensional visual risk indicators, and triggers emergency responses.

[0040] Each module realizes closed-loop control through data flow interaction. The image data acquisition module transmits data to the fault prediction module. The fault prediction module outputs to the path optimization module. The path optimization module feeds back signals to the adaptive control module. Finally, the UAV coordination module executes dynamic regulation and control.

[0041] Example 1: Referring to Figure 2 , the image data acquisition module serves as the front-end sensing unit of the system and is composed of a thermal imaging sensor unit and a spectral analysis unit. It is deployed at the bottom of the UAV body and the hovering position to ensure complete detection coverage of the photovoltaic array. The thermal imaging sensor unit uses a high-resolution infrared camera with a working wavelength of 814 microns, which can capture temperature distribution differences on the surface of photovoltaic modules. This unit collects infrared images at a rate of 5 frames per second during flight, generating infrared thermal image signals containing temperature gradient data for each region of the module. The spectral analysis unit is equipped with a near-infrared spectrometer with a wavelength range of 900-1700 nanometers. It obtains the spectral characteristics of photovoltaic modules through reflectance spectroscopy. This unit scans the target module under the hovering state of the UAV and records reflectivity data at different wavelengths to form spectral characteristic signals.

[0042] The data of the thermal imaging sensor unit and the spectral analysis unit are integrated by a multimodal fusion algorithm. After the infrared thermal image signal and the spectral feature signal are aligned in time and space, a weighted fusion strategy is used to generate enhanced image data. During the fusion process, the infrared data is used to identify hot spots and temperature anomaly areas, and the spectral data is used to analyze the dirt and material degradation characteristics on the surface of the components. The enhanced image data is divided into three categories: the prediction enhanced image data contains hot spot distribution, dirt coverage, and spectral reflectivity information, which is used as input for the fault prediction module; the optimization enhanced image data integrates the component tilt angle, unmanned aerial vehicle power parameters, and environmental lighting conditions, which serves the path optimization module; the regulation enhanced image data includes real-time wind speed, terrain height, and environmental temperature and humidity, which is used by the adaptive control module.

[0043] The automatic repair execution unit, as the core component of the unmanned aerial vehicle cooperation module, is composed of an intelligent flight controller and an integrated temperature monitoring device. The intelligent flight controller receives the optimal inspection trajectory signal generated by the path optimization module, which contains the three-dimensional coordinates of the inspection area, the priority of the unmanned aerial vehicle type, the flight speed range, and the path planning scheme. The controller dynamically adjusts the flight height and pitch angle of the unmanned aerial vehicle according to the coordinate information to ensure that the camera equipment is always aimed at the target component. In complex terrain areas, the controller automatically raises the flight height in combination with the terrain height data to avoid collision risks. The flight speed range is dynamically adjusted according to the dirt coverage of the photovoltaic array, and the flight speed is reduced in areas with severe dirt to improve detection accuracy.

[0044] The integrated temperature monitoring device monitors the temperature changes of the key components of the unmanned aerial vehicle in real time based on infrared temperature measurement technology, including the motor, battery, and electronic speed controller. The device is equipped with multiple temperature sensors with a sampling frequency of 10 Hz, and the data is transmitted to the flight control unit through the CAN bus. When the detected temperature exceeds the preset threshold, the device automatically adjusts the heat dissipation strategy of the unmanned aerial vehicle, such as reducing the motor load or starting the auxiliary cooling fan. In high-temperature environments, the device optimizes the flight path in combination with the optimal inspection trajectory signal to preferentially detect shadow areas or photovoltaic components with lower temperatures, reducing the exposure time of the unmanned aerial vehicle in high-temperature environments.

[0045] Real-time monitoring signals are collected by the flight control unit, including device temperature, humidity, voltage, flight height, component pressure, and flight distance parameters. Device temperature and humidity data are obtained in real time by onboard sensors to assess the operating status of the unmanned aerial vehicle; voltage data monitor the remaining battery capacity and trigger a low-power warning to automatically return; flight height and component pressure data are measured by barometers and pressure sensors to determine the relative position relationship between the unmanned aerial vehicle and the photovoltaic components; flight distance data record the cumulative inspection mileage of the unmanned aerial vehicle to assist in maintenance cycle prediction. These signals are transmitted to the interactive warning platform through the wireless communication module to form a closed-loop monitoring network.

[0046] The flight control unit of the UAV coordination module adjusts the flight attitude according to the safety flight threshold output by the adaptive control module. The safety flight threshold includes the maximum allowed wind speed, the minimum flight height, and the terrain fluctuation tolerance. When the real-time wind speed exceeds the threshold, the flight control unit automatically reduces the speed of the UAV and switches to the wind-resistant mode; when the terrain height changes dramatically, the unit dynamically adjusts the flight height to avoid collision. The flight attitude parameters are adjusted in real time through the PID controller to ensure the stability of the camera equipment and reduce motion blur during image acquisition.

[0047] The interactive early warning platform receives enhanced image data and real-time monitoring signals and dynamically displays the health status of the photovoltaic module through a three-dimensional visualization interface. The platform uses OpenGL rendering technology to support multi-angle viewing of the hot spot, dirt, and crack distribution on the surface of the module. The risk assessment engine generates a comprehensive risk index based on performance degradation signals and potential failure signals, and triggers a multi-level early warning mechanism when the index exceeds the preset threshold. In the yellow warning state, the platform prompts the operator to pay special attention to the specific area; in the orange warning state, the automatic repair execution unit starts the cleaning or local repair program; in the red warning state, the system immediately stops the UAV flight and disconnects the power supply of the photovoltaic array to prevent fault propagation.

[0048] The data flow of the entire system realizes closed-loop control. The image data acquisition module transmits the enhanced image data to the fault prediction module, the fault prediction result is transmitted to the path optimization module, the optimized trajectory signal is fed back to the adaptive control module, and finally the UAV coordination module executes the inspection task. The data interaction between modules is realized through high-speed bus and wireless communication, ensuring the real-time and reliability of the system.

[0049] In the specific implementation process, the calibration of the thermal imaging sensor unit and the spectral analysis unit is crucial. The thermal imaging sensor needs to be calibrated regularly by a blackbody radiation source to ensure that the temperature measurement accuracy is within ±2℃; the spectral analysis unit needs to be calibrated using a standard reflector to eliminate environmental light interference. The intelligent flight controller of the automatic repair execution unit needs to be pre-loaded with three-dimensional map data of the photovoltaic power station to match the actual coordinates and image acquisition area. The threshold setting of the temperature monitoring device needs to consider the model and heat dissipation performance of the UAV to avoid false triggering or delayed response.

[0050] The collection frequency and transmission delay of real-time monitoring signals directly affect the response speed of the system. The flight control unit uses a real-time operating system to ensure that the sampling interval of sensor data does not exceed 100 milliseconds; the wireless communication module uses 5G or a dedicated frequency band to reduce data transmission delay. The risk assessment algorithm of the interactive early warning platform needs to be updated regularly, combined with historical fault data to optimize the early warning threshold and reduce false positives and false negatives.

[0051] The system deployment needs to consider the scale and environmental conditions of the photovoltaic power station. For large power stations, multiple UAVs can be used for cooperative inspection, and distributed computing can be used to optimize the path planning. For high wind speed or high altitude areas, a UAV model with stronger wind resistance performance needs to be selected, and the parameters of the safety flight threshold need to be adjusted. The maintenance work of the system includes regular inspection of sensor accuracy, updating of three-dimensional map data, and optimization of fault prediction models to ensure the stability of long-term operation.

[0052] Embodiment 2: see Figure 3 The fault prediction module is the core analysis unit of the system, and adopts a convolutional neural network image classification model and a Bayesian network model to work together. The module receives prediction-enhanced image data from the image data acquisition module, including photovoltaic module surface dirt distribution data, hot spot anomaly data, spectral reflectivity data, and historical maintenance records and historical efficiency decay data. After preprocessing, these data are input into different submodules for feature extraction and fusion analysis.

[0053] The convolutional neural network image classification model is composed of a feature extraction submodule, a multi-source fusion submodule, a risk assessment submodule, and a result display module. The feature extraction submodule first processes the hot spot anomaly data, and uses an edge detection algorithm to identify the temperature anomaly area on the surface of the module. The algorithm analyzes the trend of the temperature gradient and extracts the area profile that meets the crack propagation characteristics, outputting a crack precursor signal. This signal contains information about the length, width, and direction of the crack, which is used for subsequent risk assessment. The multi-source fusion submodule uses a convolutional neural network structure with attention mechanism weighting to process surface dirt distribution data, hot spot anomaly data, and spectral reflectivity data simultaneously. The network structure is designed with multiple branches for input, with each branch corresponding to a data type. The contribution of different data sources is dynamically adjusted through attention weights. The dirt distribution data reflects the degree of contamination on the surface of the module, the hot spot data indicates the risk of local overheating, and the spectral data represents the material aging state. The fusion of the three outputs forms a comprehensive feature vector.

[0054] The risk assessment submodule receives the comprehensive feature vector output by the multi-source fusion submodule and defines three core indicators: the maximum dirt coverage, the average hot spot temperature, and the average spectral reflectivity. The dirt coverage is calculated by image segmentation technology to determine the proportion of the area covered by the module surface; the average hot spot temperature is calculated by counting the temperature deviation from the normal range in the abnormal area; and the average spectral reflectivity is calculated by analyzing the reflection intensity change in a specific waveband. After normalization, the three indicators are weighted to obtain the risk indicator value of each photovoltaic module. The result display module maps the risk indicator value to the color gradient space to generate a dynamic heat map. The heat map uses a gradient color scale from green to red to represent the risk level, with green representing normal state, yellow indicating attention, orange indicating warning, and red indicating serious failure. The heat map supports interactive viewing, allowing users to click on any module to view detailed risk parameters.

[0055] The Bayesian network model handles the relationship between surface contamination distribution data, hot spot anomaly data, and spectral reflectance data as a probabilistic inference tool. The network structure includes three types of nodes: observation nodes correspond to input data, hidden nodes represent state variables that cannot be directly measured, and output nodes generate performance degradation signals. The conditional probability distribution of the observation nodes is trained based on historical data to establish the relationship between the degree of contamination and the probability of hot spot occurrence, the relationship between spectral changes and material aging, etc. During model operation, input data updates the confidence of each node through a probability propagation algorithm, and finally outputs the performance degradation signal. This signal contains key parameters such as component efficiency decay rate prediction and remaining useful life estimation, which are used for long-term health management of the system.

[0056] Historical efficiency decay data and image classification model output potential fault signals together constitute a complete fault prediction result. Historical efficiency decay data comes from long-term records of power station monitoring systems, reflecting the power output decline trend of each component. These data are compared with current detection results in time sequence to identify abnormal patterns of accelerated decay. Potential fault signals integrate crack precursors, severe contamination areas, and hot spot distribution information, marking the locations of components that need priority attention. The correlation analysis of the two types of data in the time dimension can distinguish between temporary performance fluctuations and persistent fault development.

[0057] During the implementation of the system, the training of the convolutional neural network model uses a transfer learning method. The pre-trained model is based on a large-scale photovoltaic component image dataset and is adapted to the environmental characteristics of a specific power station through fine-tuning. The training data includes samples collected under different seasons and weather conditions, ensuring the model's generalization ability. The structure design of the Bayesian network involves domain experts to determine the causal relationships between key variables. Network parameters are learned from historical case data and updated regularly to adapt to changes in component aging patterns.

[0058] The data processing flow adopts a distributed computing architecture. Raw image data is preprocessed on edge computing nodes, including noise removal, geometric correction, and feature extraction operations. The preprocessed feature data is uploaded to the cloud server for model inference, reducing network transmission load. The inference results are cached in the local database for real-time calling by the path optimization module and interactive warning platform. The system sets up a data quality check mechanism to automatically filter out fuzzy images or abnormal sensor readings, avoiding the impact of false inputs on analysis results.

[0059] The output interface design of the fault prediction module takes into account the needs of downstream modules. Potential fault signals adopt a structured data format, including fields such as component number, fault type, severity, and location coordinates, facilitating the path optimization module to plan detection routes. Performance degradation signals contain time-series prediction data, encapsulated in JSON format, supporting visualization on the interactive early warning platform. Module operation status monitoring information is provided through a separate service interface, including operational data such as model version, computation time, and resource usage.

[0060] In practical deployments, the module's inference performance is optimized through various techniques. Convolutional neural network models are converted to a dedicated inference engine format, utilizing GPUs for accelerated computation. Bayesian network inference employs approximation algorithms to achieve a balance between accuracy and efficiency. A batch processing mechanism for input data reduces the overhead of model calls, prioritizing computational resources for operations with high real-time requirements. The system includes a model update mechanism that triggers a model retraining process when newly acquired data deviates from the existing model by more than a threshold.

[0061] The module's exception handling mechanism covers both data anomalies and system anomalies. Data anomalies include sensor failures, communication interruptions, etc., in which the system automatically switches to degraded mode and uses the most recent valid data for inference. System anomalies, such as hardware failures or software crashes, are handled through heartbeat detection and a watchdog timer for rapid recovery. All anomalies are logged in the system log for post-event analysis and improvement.

[0062] Collaboration with other modules is achieved through message queues. The fault prediction module subscribes to data published by the image data acquisition module and simultaneously pushes analysis results to the path optimization module and the interactive early warning platform. The message format adopts an industry-standard protocol to ensure compatibility with devices from different manufacturers. Important messages are equipped with an acknowledgment mechanism; data that fails to be delivered is placed in a retransmission queue to ensure the reliable transmission of critical information.

[0063] The system's scalability is reflected in the modular design of its models and functions. New detection algorithms can be integrated into the existing framework as plug-ins without affecting core processes. Horizontal scaling of hardware resources supports the deployment needs of larger-scale power plants. The remote management interface allows operations and maintenance personnel to adjust model parameters and business rules to adapt to changes in power plant operation strategies.

[0064] The maintenance plan comprises two parts: periodic calibration and preventative maintenance. Calibration primarily focuses on the accuracy of input data, such as temperature calibration of thermal imaging sensors and optical calibration of spectrometers. Preventative maintenance addresses the health status of computing nodes, including storage space monitoring, thermal system checks, and backup data verification. Maintenance records are digitally managed, creating a device lifecycle archive.

[0065] Example 3: See Figure 4The path optimization module processes the optimized enhanced image data from the image data acquisition module and the potential fault signals output by the fault prediction module through a trajectory-fault nonlinear relationship model. The optimized enhanced image data contains key information such as the inclination angle of the photovoltaic module, the power parameters of the unmanned aerial vehicle, and the environmental lighting conditions, which together with the potential fault signals constitute the input parameters for path planning. The inclination angle of the module is obtained through the inertial measurement unit carried by the unmanned aerial vehicle, reflecting the actual attitude of the photovoltaic panel installation; the power parameters of the unmanned aerial vehicle include real-time state data such as battery remaining capacity, motor speed, and propeller efficiency; the environmental lighting conditions are collected by the light intensity sensor, affecting the quality of image acquisition.

[0066] The trajectory-fault nonlinear relationship model establishes a mathematical correlation between the flight path length and the degree of dirt coverage, the density of hot spots, and the speed of the unmanned aerial vehicle. The model uses a multivariate nonlinear regression method to construct, and its core expression is:

[0067] L = a · D + g · T + b · v β δ ∈

[0068] where L represents the path length, D is the degree of dirt coverage, T is the density of hot spots, and v represents the flight speed of the unmanned aerial vehicle. a and g are weight coefficients, reflecting the influence degree of different factors on the path length; b, d, and e are exponential parameters, describing the nonlinear effects of each variable. The weight coefficients and exponential parameters are obtained by training historical flight data and are updated regularly to adapt to the inspection requirements under different seasons and environmental conditions. The degree of dirt coverage D comes from the analysis results of the image classification model, representing the proportion of the component surface covered in percentage form; the hot spot density T is calculated as the number of abnormally high temperature regions per unit area; the flight speed v is dynamically adjusted according to the model and power state of the unmanned aerial vehicle.

[0069] The preliminary flight signals output by the model include path length, recommended speed, and detection priority. These signals are input into the particle swarm optimization algorithm for further processing. When the particle swarm optimization algorithm is initialized, historical power parameters are loaded from the performance matching database as references, including optimal flight speed, typical path length, and energy consumption distribution under different weather conditions. Each particle represents a possible flight scheme, and the optimal solution is found by iteratively updating the position and speed. The fitness function of the particle considers three objectives: path coverage, energy efficiency, and fault detection urgency. Path coverage is calculated as the ratio of detected areas to total areas; energy efficiency evaluates power consumption per unit distance; fault detection urgency is classified according to the severity of potential fault signals.

[0070] ​​The network connectivity analysis submodule evaluates the influence of the physical layout of the photovoltaic array on the flight path. This submodule calculates the connectivity coefficient by the UAV power parameters and the length of the flight path, quantifying the accessibility between different areas. The connectivity coefficient takes into account the following factors: the array spacing determines the turning radius limit of the UAV, the component height difference affects the dynamic adjustment of the flight height, and the obstacle distribution constrains the safe flight channel. The calculation process uses the graph theory method to model the photovoltaic array as a node network, and the weight of the connection between nodes reflects the difficulty of flight. The weight assignment rules include: the weight of flat areas is lower, the weight of inclined areas increases moderately, and the weight of areas near obstacles increases significantly. The ideal traffic route between nodes is generated by the shortest path algorithm, and the final inspection sequence is determined in combination with the particle swarm optimization result.

[0071] The trajectory planning submodule integrates the connectivity coefficient, the trajectory-fault nonlinear relationship model, and the real-time power parameters to generate the optimal inspection trajectory signal. The specific content of the signal includes: the three-dimensional coordinate sequence of the inspection area, which defines the position of the UAV's passing point; the speed adjustment strategy, which sets different flight speeds for different areas; and the detection mode parameters, which control the focal length and shooting frequency of the camera equipment. The coordinate sequence is smoothed by B-spline curve to eliminate sharp turns and sudden changes in speed; the speed adjustment strategy reduces the flight speed to 70% of the standard value in heavily soiled areas and restores to 100% in normal areas; and the detection mode parameters are automatically matched according to the component type, with high-resolution mode for monocrystalline silicon components and multi-spectral scanning for thin-film components.

[0072] The real-time updating mechanism of the UAV power parameters ensures the adaptability of the path planning. During flight, the battery management system continuously monitors the remaining power, triggering the path shortening strategy when the power falls below the threshold; the motor temperature sensor data is used to dynamically adjust the maximum speed limit to prevent overheating damage; and the environmental wind speed measurement results are used to real-time correct the flight stability parameters, automatically increasing the path redundancy under strong wind conditions. These parameters are transmitted to the trajectory planning submodule through the bus system, triggering online recalculation of the trajectory. The recalculation process retains the state of the completed detection areas and only adjusts the path direction of the un-inspected part to avoid repeated detection.

[0073] The design of the performance matching database supports the effective use of historical data. The database uses a time series storage structure to record the complete parameter set of each inspection, including date and time, weather conditions, equipment status, and path effects. When searching for data, the database prioritizes matching historical records similar to the current environmental conditions and extracts relevant parameters as the initial values for the optimization algorithm. The database establishes an indexing mechanism to accelerate the query response for massive historical data, with index keys including temperature intervals, wind speed levels, and component types. The data cleaning process periodically removes outdated or low-quality records to maintain the reference value of the database.

[0074] The fault-tolerant processing mechanism of the system is designed for possible abnormal situations. When the sensor data is lost, the module automatically switches to a conservative mode and continues to run with preset safety parameters; when the path planning is timed out, the system falls back to a simplified grid scanning mode to ensure basic coverage; when the communication is interrupted, the UAV executes the last received valid trajectory for inspection, while caching local data for uploading after recovery. All abnormal events are recorded in the flight log, including the time of occurrence, duration, and handling measures, etc. for subsequent analysis and system improvement.

[0075] The cooperation with other modules is realized through standardized interface protocols. The path optimization module receives potential fault signals from the fault prediction module, and the signal format contains structured fields such as component number, fault type, and location coordinates. The optimal inspection trajectory signal output to the adaptive control module uses waypoint sequence encoding, and each waypoint contains attributes such as latitude, longitude, height, and dwell time. Data exchange with the UAV cooperation module is through a high-speed bus, with transmission delay controlled at the millisecond level. Interface protocols define necessary verification mechanisms such as CRC verification and timeout retransmission to ensure reliable delivery of critical instructions.

[0076] The module's computing resource management adopts a dynamic allocation strategy. In the path planning phase, the algorithm preferentially uses GPU accelerated computing to shorten the optimization time; in the flight monitoring phase, CPU resources are focused on processing real-time sensor data; memory allocation sets a soft upper limit to avoid a single task exhausting resources affecting the overall response of the system. Resource usage status is displayed in real time through a monitoring service, including indicators such as occupancy, temperature, and power consumption, and abnormal situations trigger alarm notifications to operation and maintenance personnel.

[0077] In embodiment 4, the adaptive control module processes and enhances image data to dynamically define flight constraint conditions for the UAV and output safe flight thresholds. The enhanced image data includes real-time wind speed signals, terrain height signals, and environmental temperature and humidity data, which are collected and fused by multiple sensors to form the basis for flight decision-making. Real-time wind speed signals are measured by an ultrasonic anemometer carried by the UAV, with a sampling frequency of 20 Hz. The data is filtered by a sliding average filter to eliminate transient fluctuations. Terrain height signals are derived from laser radar scanning results, which construct a digital elevation model of the photovoltaic array area. Environmental temperature and humidity data are obtained by an onboard weather station, which is used to assess the impact of air density on flight performance.

[0078] The calculation of the safe flight threshold is based on three core parameters: minimum flight height, maximum average wind speed, and terrain undulation. The minimum flight height is determined based on the installation height of the photovoltaic components to avoid collisions between the UAV and the components. The maximum average wind speed is calculated by taking the maximum wind speed over the past 30 seconds to reflect the stability of the current wind field. The terrain undulation quantifies the rate of elevation change in the inspection area by analyzing the height difference between adjacent locations through gridding. These three parameters are dynamically adjusted through a fuzzy logic system to form flight constraints that adapt to different environmental conditions. The rule base of the fuzzy logic contains 72 empirical rules covering various wind speed, height, and terrain combinations for safe operation guidelines. For example, when the wind speed increases, the system automatically raises the minimum flight height; when the terrain undulation is severe, the upper limit of the UAV's speed is correspondingly reduced.

[0079] The following table shows an example of safe flight threshold adjustment for a photovoltaic power station under different environmental conditions:

[0080]

[0081] The actual operation case of a UAV in a coastal photovoltaic power station in Jiangsu Province demonstrates the working mechanism of this module. The power station is located in an open area with high average wind speed throughout the year. During a certain inspection period, real-time monitoring showed that the wind speed gradually increased from 4.2 m / s to 7.8 m / s, with a southeast direction. The adaptive control module detected that the wind speed exceeded the current threshold of 6.5 m / s and immediately started the dynamic adjustment program: first, the minimum flight height was raised from 1.5 meters to 2.2 meters to avoid the risk of height fluctuations caused by sudden wind; then the cruising speed was reduced from 4.0 m / s to 3.2 m / s to increase flight stability; at the same time, the terrain undulation threshold was tightened to 8% to limit the UAV's maneuvering range near high-voltage transmission lines. These adjustments take effect in real time through the flight control unit, and the UAV's attitude control system correspondingly increases the motor power output to compensate for the loss of thrust caused by the increase in wind speed.

[0082] The terrain adaptation mechanism performed well in the application of a mountainous photovoltaic power station in Yunnan Province. The power station is built on the mountain, with a maximum installation height difference of 35 meters. Laser radar scanning showed that there was a 4.2-meter elevation jump within a 10-meter range in a certain inspection area. The adaptive control module identified that the terrain undulation in this area reached 28%, exceeding the preset threshold, triggering the following response: the flight height was automatically adjusted to 3.5 meters above the components, leaving sufficient safety margin; the flight path was re-planned to bypass the steep slope area and change to a step-by-step gradual detection; the focal length of the camera equipment was dynamically adjusted to compensate for the image scale difference caused by height changes. Throughout the process, the UAV maintained stable image acquisition quality and did not experience image blurring or data loss caused by terrain.

[0083] The processing of environmental temperature and humidity data is demonstrated in the case of a power plant in a humid region of Guangdong Province. During summer afternoons, the ambient temperature reaches 38°C and the relative humidity is 85%, causing a decrease in air density of approximately 12%. The module detects a decrease in power system efficiency and takes three measures: the battery cooling system power is increased by 20% to prevent a sudden decrease in capacity caused by high temperatures; the motor temperature monitoring interval is shortened from 10 seconds to 5 seconds to provide early warning of overheating risks; and flight tasks are segmented, with a landing every 15 minutes for 5 minutes of cooling under a sunshade. These measures enable the UAV to complete 85% of the scheduled inspection tasks in extreme environments, with the remaining areas being inspected after the temperature drops.

[0084] The redundancy design of the system is verified by a sudden situation at a high-altitude power plant in Sichuan Province. At an altitude of 3100 meters, a sudden failure of the barometric pressure sensor caused abnormal height data. The module automatically switches to the backup plan: using laser range finder data to replace the barometric altimeter, and fusing inertial navigation data through Kalman filtering; at the same time, the safe flight threshold is switched to a conservative mode, reducing the upper limit of flight speed by 15%. This seamless switching ensures the safe operation of the UAV during sensor failure, without triggering an emergency landing.

[0085] The interaction with the UAV coordination module uses a priority scheduling mechanism. When the safety threshold issued by the adaptive control module conflicts with the trajectory planning of the path optimization module, the system processes according to the following priority: terrain safety limit > wind speed limit > device temperature limit > original path planning. For example, during a flight, the UAV approaches a high-temperature component area, and the device temperature is close to the critical value. At this time, the module prioritizes the cooling strategy and suspends the original low-speed detection task until the temperature returns to normal.

[0086] The self-learning function of the module is reflected in the accumulation of data during long-term operation. The system records the correspondence between each environmental condition and the adjustment parameters, and identifies typical scenarios through cluster analysis. For example, the frequent sea wind attack mode at a power plant is marked as a specific type, and when similar meteorological features are detected, the historical optimal parameters are directly called to reduce real-time calculation delay. The learning results are updated every quarter to form a flight strategy library specific to the power plant.

[0087] Maintenance personnel can adjust the basic parameters through the configuration interface to adapt to power plant modifications or equipment upgrades. For example, when a lightning rod is added to a certain area, the administrator labels this obstacle in the digital elevation model, and the module automatically increases the terrain relief weight of this area by 30%. Configuration changes require double verification to prevent misoperation affecting flight safety.

[0088] Example 5: The interactive early warning platform serves as the integrated decision-making and visualization terminal of the system, integrating the two core functions of three-dimensional visualization interface and risk early warning engine. The platform receives performance degradation signals and potential failure signals from the fault prediction module, as well as real-time monitoring signals from the unmanned aerial vehicle coordination module, and generates a panoramic view of the health status of photovoltaic modules through multi-dimensional analysis. The three-dimensional visualization interface is developed based on WebGL technology, using a hierarchical rendering strategy that prioritizes loading component models within the current window, supporting smooth display of detailed features of over ten thousand photovoltaic modules on ordinary workstations. The interface operation logic is designed as an exploration mode, allowing users to view fault features in any location through functions such as free perspective rotation, cross-section cutting, and regional focusing.

[0089] The risk early warning engine uses an improved approximation ideal solution sorting algorithm to process input signals and establish a multi-level risk assessment system. The algorithm first normalizes performance degradation signals and potential failure signals to eliminate dimensional differences between different indicators. Performance degradation signals include time series data such as efficiency decay rate, output power deviation, and material aging coefficient; potential failure signals include spatial distribution parameters such as hot spot level, dirt coverage area, and crack length. The normalized data is calculated by weighted distance to determine the deviation of each component from the ideal state, outputting a comprehensive risk index ranging from 0 to 1. The index value is divided into three warning intervals: 0 to 0.3 is normal, displayed as green; 0.3 to 0.7 is yellow warning, needs attention; 0.7 to 1 is red warning, must be handled immediately.

[0090] The early warning state triggers a differentiated response mechanism. When the component enters the yellow warning state, the system automatically reduces the flight speed of the unmanned aerial vehicle in that area to 60% of the standard value, prolongs the image acquisition time to improve detection accuracy. At the same time, the platform interface pops up a semi-transparent prompt box, labeling the efficiency change curve and historical maintenance records of the component in the past three months. In the orange warning state, the unmanned aerial vehicle hovers 2 meters above the faulty component and starts the high-definition review mode, capturing details of the fault from multiple angles. The platform locks the array region where the component is located, highlighting it with a pulsating flashing effect, and automatically retrieves the comparison data from the last three inspections. The red warning state triggers the full-system emergency protocol: the unmanned aerial vehicle immediately interrupts the current task and transmits the coordinates of the faulty component to the ground maintenance terminal; the photovoltaic array controller disconnects the circuit connection of the corresponding string; the platform interface switches to the emergency treatment view, providing standardized emergency operation guidelines.

[0091] The rendering pipeline of the three-dimensional visualization interface is specially optimized for photovoltaic inspection needs. The basic model uses lightweight modeling techniques, with the number of triangular facets of a single component controlled to within 500, while retaining key structural features. The material system realizes dynamic reflection effects, with real-time updates to the component surface lighting simulation according to the solar azimuth angle. The fault diffusion process uses a particle system to represent, with the hot spot area emitting an orange particle stream and the crack path generating a red trajectory line. Users can replay the fault development history through the timeline control and observe abnormal change patterns during a specific period. The cross-section analysis tool supports cutting the component model with any plane, allowing users to view the microscopic state of internal connectors and cell pieces, with the cutting plane automatically generating contour line labels for key dimensions.

[0092] The decision logic of the risk warning engine introduces a time decay factor to distinguish the handling priorities of new and old faults. Newly discovered fault features have a higher initial weight, which increases exponentially over time. Historical faults that persist in subsequent inspections have their risk index calculation increased by a recurrence penalty coefficient. This mechanism avoids overreacting to long-term, minor defects while ensuring that major risks are not missed. The engine runs a self-checking program in the background, periodically verifying the rationality of algorithm weight settings and prompting model updates when input data distribution changes significantly.

[0093] The linkage between the platform and the automatic repair execution unit is achieved through an instruction coding system. Yellow warning status sends speed adjustment instructions, including target coordinates and correction speed values; orange warning status sends re-inspection instructions, specifying shooting angles and resolution parameters; red warning status transmits emergency handling packages, containing complete parameters such as shutdown coordinates, safety distances, and isolation ranges. Instruction transmission uses a redundancy check protocol, with critical instructions requiring the receiver to return an acknowledgment signal, and a retransmission through a backup channel if there is no response within a timeout period. The execution results are fed back to the platform in real time, updating the status markers on the three-dimensional model.

[0094] User interaction design follows aviation control standards, using a double-confirmation operation process. Important instructions such as manually overriding warning states or forcibly interrupting tasks require two clicks on different locations of the confirmation button to take effect. The operation log records the complete interaction process, including timestamps, operation content, and execution results, supporting filtering queries by date, operation type, or component number. The permission management system divides access into four levels, with ordinary inspectors only able to view data, senior engineers having parameter adjustment permissions, and emergency operations requiring authorization from the power station manager.

[0095] In the actual deployment of a large photovoltaic power station in Qinghai Province, the platform successfully identified the early failure of a string inverter. Through comparative analysis, the system found that the 32 components under the jurisdiction of a certain inverter showed synchronous efficiency decay. Although the risk indicators of individual components did not reach the warning threshold, the cluster anomaly pattern triggered a special inspection recommendation. Maintenance personnel confirmed that the local overheating was caused by the failure of the inverter's cooling fan based on the platform's marked location. Similarly, in a fish-light complementary power station in Shandong Province, the platform's three-dimensional profile function helped locate the insulation layer damage at the underwater cable connection, which would have been easily overlooked in traditional two-dimensional detection.

[0096] The system's continuous learning function is reflected in the gradual improvement of warning accuracy. The platform records the correspondence between each warning and actual failure, and automatically adjusts the sensitivity of related parameters when a specific type of false alarm pattern is found. For example, false hot spots caused by morning dew in a certain power station, after training with data from the rainy season, the system learned to make more accurate judgments by combining humidity data and temperature change rate. The learning results are pushed to all deployed sites through quarterly software updates, forming a collective sharing of experience.

[0097] The maintenance terminal extension function supports on-site operation collaboration. Maintenance personnel wear augmented reality glasses that can access the three-dimensional model and historical data of the current component superimposed in the real field of view. The lightweight application on the glasses keeps data synchronized with the platform, displaying fault location guidance and standard maintenance procedures. Maintenance results are uploaded by taking photos with mobile terminals and automatically archived in the component's lifecycle file.

[0098] This implementation builds a three-dimensional monitoring system for photovoltaic power station operation and maintenance by deeply integrating three-dimensional visualization and intelligent early warning decision-making. The interactive interface translates complex device states into intuitive spatial relationship expressions, and the risk engine establishes a multi-parameter coupled evaluation model. Practical cases have verified the practical value of the system in complex environments. The platform design focuses on balancing human-machine collaboration, providing both automated decision support and retaining key judgment rights for professionals. The progressive learning mechanism and distributed deployment architecture enable the system to adapt to the individual needs of photovoltaic power stations of different sizes and continuously optimize operation and maintenance efficiency in the long run.

[0099] It should be noted that in this paper, relationship terms such as first and second are used only 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 the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or device.

[0100] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An integrated machine vision unmanned aerial vehicle automatic patrol inspection system for photovoltaic power stations, characterized in that, The application relates to a photovoltaic module inspection system based on unmanned aerial vehicle (UAV) cooperation, which comprises the following modules: an image data acquisition module, a fault prediction module, a path optimization module, an adaptive control module and a UAV cooperation module. The image data acquisition module is used for capturing a surface image sequence of a photovoltaic module by a camera device carried by an unmanned aerial vehicle, thereby forming initial image data. The fault prediction module receives and processes the initial image data, and outputs potential fault signals and performance degradation signals. The path optimization module processes the initial image data and the potential fault signals, and generates an optimal inspection trajectory signal. The adaptive control module defines flight constraint conditions according to the initial image data, and outputs a safe flight threshold. The UAV cooperation module comprises an automatic repair execution unit, a flight control unit and an interactive early warning platform. The image data acquisition module transmits the captured data to the fault prediction module, the fault prediction module transmits the prediction results to the path optimization module, the path optimization module feeds back the optimization signal to the adaptive control module, and the adaptive control module sends the constraint conditions to the UAV cooperation module, so as to realize data interaction among the modules of the system.

2. The integrated machine vision based unmanned aerial vehicle automatic inspection system for photovoltaic power station of claim 1, wherein, The image data acquisition module comprises a thermal imaging sensor unit and a spectrum analysis unit. 3.The integrated machine vision photovoltaic power station unmanned automatic inspection system of claim 2, characterized in that, The image data acquisition module is arranged at a hovering position of the unmanned aerial vehicle and above a photovoltaic array, and collects infrared thermal image signals. The spectrum analysis unit obtains spectral reflectance data of the photovoltaic module by near-infrared spectrum technology, and forms spectral feature signals. The image data acquisition module integrates the infrared thermal image signals and the spectral feature signals to realize multi-modal data fusion, and forms enhanced image data. The fault prediction module comprises a convolutional neural network-based image classification model and a Bayesian network model. The prediction enhanced image data comprises surface dirt distribution data, hot spot abnormal data, spectral reflectance data, historical maintenance records and historical efficiency attenuation data. The surface dirt distribution data, the historical maintenance records, the hot spot abnormal data, the spectral reflectance data and the historical efficiency attenuation data are input into the image classification model and output the potential fault signals. The surface dirt distribution data, the hot spot abnormal data and the spectral reflectance data are fused by an improved Bayesian network model and output the performance degradation signals.

4. The integrated machine vision based unmanned aerial vehicle automatic inspection system for photovoltaic power station of claim 3, wherein, The image classification model comprises a feature extraction submodule, a multi-source fusion submodule, a risk assessment submodule and a result display module, the feature extraction submodule performs edge detection algorithm processing on the hot spot abnormal data, extracts a region meeting a crack propagation characteristic range as a crack precursor signal, the multi-source fusion submodule processes the surface dirt distribution data, the hot spot abnormal data and the spectral reflectance data using an attention mechanism weighted convolutional neural network and outputs the data to the risk assessment submodule, the risk assessment submodule defines a risk quantification method, calculates a risk indicator using a maximum dirt coverage degree, a hot spot average temperature value and a spectral reflectance average value, and the result display module generates a dynamic heat map of the risk indicator based on a color gradient and uses different color scales to represent different risk levels of photovoltaic component positions.

5. The integrated machine vision based UAV automatic inspection system for photovoltaic power plant as claimed in claim 2, wherein, The path optimization module processes the optimization enhanced image data and the potential fault signal through a trajectory-fault nonlinear relationship model, outputs a preliminary flight signal, and processes the preliminary flight signal using a particle swarm optimization algorithm to obtain the optimal inspection trajectory signal; the optimization enhanced image data includes component tilt angle data and unmanned aerial vehicle power parameters, the trajectory-fault nonlinear relationship model is constructed through the component tilt angle data, the potential fault signal and the unmanned aerial vehicle power parameters, and the trajectory-fault nonlinear relationship model describes the relevance of flight path length and dirt coverage degree, hot spot density and unmanned aerial vehicle speed. 6.The integrated machine vision based unmanned aerial vehicle automatic inspection system for photovoltaic power station of claim 5, wherein, The path optimization module further comprises a performance matching database, a network connectivity analysis submodule and a trajectory planning submodule, the performance matching database stores historical power parameters of historical unmanned aerial vehicle flights, the network connectivity analysis submodule calculates the connectivity coefficient of the photovoltaic array according to the unmanned aerial vehicle power parameters and the flight path length, and the trajectory planning submodule obtains the optimal inspection trajectory signal through a particle swarm optimization algorithm according to the connectivity coefficient, the trajectory-fault nonlinear relationship model and the unmanned aerial vehicle power parameters. 7.The integrated machine vision based unmanned aerial vehicle automatic inspection system for photovoltaic power station of claim 2, wherein, The adaptive control module defines a flight constraint condition according to the regulation enhanced image data, outputs a safe flight threshold, and dynamically adjusts the safe flight threshold to be not lower than a safe flight minimum value through fuzzy logic, the regulation enhanced image data includes a real-time wind speed signal and a terrain height signal, the flight constraint condition is defined by the real-time wind speed signal and the terrain height signal, the safe flight threshold is calculated based on a minimum flight height, an average wind speed extreme value and a terrain undulation degree, and the fuzzy logic dynamically adjusts the unmanned aerial vehicle hovering height and flight speed to make the safe flight threshold not lower than the safe flight minimum value.

8. The integrated machine vision based UAV automatic inspection system for photovoltaic power plant as claimed in claim 1, wherein, The interactive early warning platform comprises a three-dimensional visualization interface and a risk early warning engine, the three-dimensional visualization interface receives the enhanced image data and the real-time monitoring signal and dynamically renders a photovoltaic module fault diffusion process, the risk early warning engine processes the performance degradation signal and the potential fault signal according to an improved technique for order preference by similarity to ideal solution algorithm and outputs a comprehensive risk index, when the comprehensive risk index exceeds a preset risk threshold, the interactive early warning platform controls the automatic repair execution unit to perform emergency repair, the three-dimensional visualization interface supports arbitrary visual angle profile analysis, real-time display of the spatial relationship between the fault area and the photovoltaic array, and simulation of component stress distribution under different flight schemes based on the finite element method. 9.The integrated machine vision based unmanned aerial vehicle automatic inspection system for photovoltaic power station of claim 8, wherein, The risk early warning engine comprises color identification of multiple early warning states, namely a yellow early warning state, an orange early warning state and a red early warning state, when the efficiency attenuation value in the potential fault signal decreases to a first attenuation threshold, the yellow early warning state is started, and the unmanned aerial vehicle flight speed is reduced to an alert speed, when the efficiency attenuation value in the potential fault signal decreases to a second attenuation threshold, the unmanned aerial vehicle flight is paused and the automatic repair execution unit is controlled to perform emergency repair, when the efficiency attenuation value in the potential fault signal decreases to a third attenuation threshold, the risk early warning engine sends an emergency fault signal and controls the integrated machine vision photovoltaic power station unmanned aerial vehicle automatic inspection system to stop.

10. The integrated machine vision based UAV automatic inspection system for photovoltaic power plant as claimed in claim 1, wherein, The optimal inspection trajectory signal comprises inspection area coordinates, unmanned aerial vehicle type priority, flight speed range and path planning scheme, the automatic repair execution unit comprises an intelligent flight controller and an integrated temperature monitoring device, the intelligent flight controller can identify the optimal inspection trajectory signal and adjust the flight height and flight angle as required, the integrated temperature monitoring device receives the optimal inspection trajectory signal and adjusts the unmanned aerial vehicle cooling strategy in real time based on infrared temperature measurement technology, and the real-time monitoring signal comprises device temperature, device humidity, device voltage, flight height, component pressure and flight distance.

Images