A Distributed New Energy Monitoring Method and System Based on Machine Vision

Abstract

This invention discloses a distributed new energy monitoring method and system based on machine vision. It acquires real-time image data through an image acquisition device, combines anomaly detection and attitude detection algorithms to locate equipment anomalies and assess energy conversion efficiency, and integrates light and shadow distribution analysis results to identify potential anomaly areas. For anomaly areas, operational data is extracted from edge computing units, and a multi-source data fusion algorithm is used to integrate image analysis and operational parameters to generate a global status assessment index, which triggers early warning and maintenance processes. This invention significantly improves the operational reliability and maintenance efficiency of renewable energy facilities through multi-dimensional data fusion and intelligent algorithm integration, providing important technical support for the sustainable development of green energy.

Description

Technical Field

[0001] This invention relates to the field of distributed new energy technology, and in particular discloses a distributed new energy monitoring method and system based on machine vision. Background Technology

[0002] Research in the field of distributed renewable energy is of great significance for promoting energy structure transformation and achieving sustainable development. With the widespread application of renewable energy sources such as photovoltaics and wind power, how to efficiently and accurately monitor these dispersed devices has become a key issue in ensuring the stable operation of energy systems and improving resource utilization efficiency. Research in this field not only concerns the optimization of energy output but also directly affects the balance between environmental protection and economic benefits.

[0003] However, current monitoring methods for distributed renewable energy still have significant shortcomings. Many solutions rely too heavily on manual inspections or simple sensor data collection, making it difficult to address the real-world challenges of widely distributed renewable energy equipment and complex environments. Especially when facing collaborative management of multiple devices and scenarios, existing methods often fail to achieve comprehensive perception and real-time response to equipment status, leading to delayed problem detection and low operational efficiency.

[0004] A deeper technical challenge lies in how to accurately assess the condition of core components of new energy equipment. This challenge focuses on the ability to perceive dynamic changes in the appearance and operating status of equipment. Damage or subtle changes in the appearance of equipment are often precursors to malfunctions, but these changes manifest differently under varying environmental and lighting conditions, making them extremely difficult to detect accurately. Failure to identify these subtle changes in a timely manner may result in missing the optimal repair window, leading to greater systemic risks. For example, in photovoltaic power plants, tiny cracks may appear on the surface of photovoltaic panels, invisible to the naked eye. These cracks may be overlooked under specific lighting conditions, but over time, their propagation can lead to decreased power generation efficiency or even equipment failure. Existing technologies struggle to continuously and accurately detect such problems in complex environments.

[0005] Therefore, in distributed new energy scenarios, how to achieve dynamic and accurate perception and anomaly identification of the appearance status of core equipment components has become a key issue in improving monitoring efficiency and system stability. Summary of the Invention

[0006] This invention provides a distributed new energy monitoring method and system based on machine vision, aiming to solve at least one of the defects existing in the prior art.

[0007] One aspect of the present invention relates to a distributed new energy monitoring method based on machine vision, comprising the following steps:

[0008] S100: Real-time image data of photovoltaic modules and wind power generation modules are collected by an image acquisition device installed on renewable energy facilities, and the real-time image data is processed by an anomaly detection algorithm to obtain the equipment anomaly location result;

[0009] S200. Based on the equipment anomaly location results, and combined with the component attitude detection algorithm, analyze the attitude changes of the wind power generation components to determine the energy conversion efficiency evaluation value.

[0010] S300: The energy conversion efficiency assessment value is fused with the light and shadow distribution analysis results of the photovoltaic module. The deviation after fusion processing is judged to see if it exceeds the preset threshold. If it does, it is marked as a potential abnormal area.

[0011] S400. For potential abnormal areas, extract operational data based on data communication from the facility edge computing unit to obtain a preliminary cross-facility data aggregation set, in which the operational data includes temperature parameters and vibration parameters;

[0012] S500 uses a multi-source data fusion algorithm to process the preliminary aggregation set of cross-facility data and the analysis results of the image acquisition device to determine the global status assessment index, which is used to monitor the overall health status of renewable energy facilities.

[0013] S600 generates early warning signals based on global status assessment indicators. If the global status assessment indicators are lower than the safety threshold, the maintenance process is triggered.

[0014] Further, step S100 includes:

[0015] S110. Real-time image data of photovoltaic modules and wind power generation modules are acquired through image acquisition devices installed on renewable energy facilities. The real-time image data is preprocessed, and noise is filtered out of the real-time image data using an image denoising tool to obtain the first image data after denoising.

[0016] S120. Use an edge detection tool to extract the device outline features from the first image data, obtain key area images of photovoltaic modules and wind power generation modules, determine whether there are shape anomalies in the key area images, and obtain preliminary anomaly judgment results.

[0017] S130. If the preliminary anomaly judgment result shows that there is a shape anomaly, then the pixel value distribution of the key area image is detected by grayscale analysis tool to obtain the specific location information of the anomaly distribution, to determine whether there is a surface defect at the specific location of the anomaly distribution, and to obtain detailed anomaly location information.

[0018] S140. Based on the detailed anomaly location information, a pre-established classification database is used for comparison to obtain the anomaly type and equipment component identifier, determine the anomaly category, and generate the final equipment anomaly location result.

[0019] Further, step S200 includes:

[0020] S210. Based on the abnormal positioning results of the equipment, the real-time attitude data of the wind turbine components is obtained through the attitude detection tool, and the noise interference of the real-time attitude data is removed by the smoothing tool to obtain the smoothed attitude change record.

[0021] S220. Based on the smoothed attitude change record, use an angle calculation tool to extract the blade angle deviation and determine whether the blade angle deviation exceeds the preset threshold. If it does, it is determined that there is an attitude abnormality.

[0022] S230. Obtain the vibration frequency distribution data of the wind turbine components during operation through vibration detection tools, compare and analyze the abnormal attitude conditions, and use frequency distribution tools to determine whether there are any related abnormal vibration characteristics.

[0023] S240. Based on the abnormal vibration characteristics and abnormal posture, the energy conversion efficiency assessment value is calculated using an efficiency assessment tool combined with the influence of environmental wind speed and historical operating records.

[0024] Further, step S300 includes:

[0025] S310. Real-time operating data is obtained from the photovoltaic module through data acquisition tools. Combined with monitoring indicators of the impact of ambient light, the light and shadow distribution data on the surface of the photovoltaic module is extracted using light and shadow distribution analysis tools to obtain the corresponding distribution characteristic records.

[0026] S320. Based on the energy conversion efficiency assessment values ​​recorded and obtained according to the distribution characteristics, a data fusion tool is used for comprehensive processing to generate a fused feature dataset.

[0027] S330. For the fused feature dataset, use a deviation calculation tool to compare it with a preset threshold range. If it exceeds the preset threshold range, it is marked as abnormal deviation data.

[0028] S340. Use anomaly marking tools to locate abnormal deviation data and combine the analysis results of historical records to determine the specific location distribution of potential abnormal areas.

[0029] Further, step S400 includes:

[0030] S410. For potential abnormal areas, temperature parameter data is obtained from the facility edge computing unit through data communication protocol to obtain a preliminary temperature dataset.

[0031] S420. Based on the preliminary temperature dataset, obtain vibration parameter data from the facility edge computing unit through a data communication protocol to obtain a preliminary vibration dataset.

[0032] S430. Use a data fusion tool to aggregate the preliminary temperature dataset and the preliminary vibration dataset to generate a cross-facility preliminary data subset.

[0033] S440. Use data verification tools to determine whether the preliminary cross-facility data subset meets the preset integrity threshold. If it does, determine the preliminary cross-facility data aggregation set.

[0034] Further, step S500 includes:

[0035] S510. Use a data alignment tool to time-stamp match the preliminary aggregated set of cross-facility data with the analysis results of the image acquisition device to obtain a time-aligned multi-source dataset.

[0036] S520. Based on the time-aligned multi-source dataset, use feature extraction tools to extract temperature features, vibration features, and image texture features to obtain a multi-feature vector set;

[0037] S530. Determine whether the similarity of the multiple feature vector sets is lower than a preset threshold. If the similarity is lower than the preset threshold, remove outliers using data cleaning tools to obtain the cleaned feature set.

[0038] S540. Based on the cleaned feature set, principal component analysis is used to perform fusion and determine the global state evaluation index.

[0039] Further, step S600 includes:

[0040] S610. Obtain real-time data streams from global status assessment indicators through data integration tools, compare the data streams with preset safety thresholds, determine whether they are below the thresholds, and obtain anomaly judgment results.

[0041] S620. Based on the anomaly judgment result, if it is lower than the safety threshold, generate an early warning signal for the abnormal data through the signal generation tool, and determine the signal type and priority.

[0042] S630: Obtain early warning signals through the process scheduling tool, match the pre-established maintenance process template according to the signal type and priority, determine whether it needs to be executed immediately, and obtain the process trigger instruction;

[0043] S640. Based on the process trigger instruction, use the data feedback tool to associate and update the trigger information with the global status assessment indicators, generate status monitoring logs for the updated data, and determine the direction of subsequent monitoring.

[0044] Another aspect of the present invention relates to a machine vision-based distributed new energy monitoring system, including the aforementioned machine vision-based distributed new energy monitoring method, comprising:

[0045] The real-time image data acquisition module is used to acquire real-time image data of photovoltaic modules and wind power generation modules through image acquisition devices installed on renewable energy facilities, and to process the real-time image data using anomaly detection algorithms to obtain equipment anomaly location results;

[0046] The energy conversion efficiency assessment value determination module is used to determine the energy conversion efficiency assessment value by analyzing the attitude changes of wind power generation components based on the equipment anomaly location results and the component attitude detection algorithm.

[0047] The potential anomaly region marking module is used to fuse the energy conversion efficiency assessment value with the light and shadow distribution analysis results of the photovoltaic module, and determine whether the deviation after fusion processing exceeds a preset threshold. If it does, it is marked as a potential anomaly region.

[0048] The cross-facility data preliminary aggregation set acquisition module is used to extract operational data based on data communication from the facility edge computing unit for potential anomaly areas, and obtain a cross-facility data preliminary aggregation set, in which the operational data includes temperature parameters and vibration parameters;

[0049] The global status assessment index determination module is used to process the analysis results of the preliminary aggregation set of cross-facility data and the image acquisition device using a multi-source data fusion algorithm to determine the global status assessment index, which is used to monitor the overall health status of renewable energy facilities.

[0050] The maintenance process triggering module is used to generate early warning signals based on global status assessment indicators. If the global status assessment indicators are lower than the safety threshold, the maintenance process is triggered.

[0051] The beneficial effects achieved by this invention are as follows:

[0052] This invention provides a machine vision-based distributed new energy monitoring method and system, aiming to solve comprehensive business problems related to anomaly detection, efficiency assessment, and global status monitoring in the operation of photovoltaic and wind power generation equipment. These problems are interconnected, with the core being how to achieve precise location of equipment anomalies, dynamic assessment of energy conversion efficiency, and real-time monitoring of the overall health of the facilities through multi-source data fusion. This invention acquires real-time image data through an image acquisition device, combines anomaly detection algorithms and attitude detection algorithms to locate equipment anomalies and assess energy conversion efficiency, and simultaneously integrates light and shadow distribution analysis results to identify potential anomaly areas. For anomaly areas, operational data is extracted from edge computing units, and a multi-source data fusion algorithm is used to integrate image analysis and operational parameters to generate global status assessment indicators, triggering early warning and maintenance processes based on these indicators. The most prominent technical effect of this invention is that through multi-dimensional data fusion and intelligent algorithm integration, it significantly improves the operational reliability and maintenance efficiency of renewable energy facilities, providing important technical support for the sustainable development of green energy. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating an embodiment of the distributed new energy monitoring method based on machine vision according to the present invention.

[0054] Figure 2 This is a functional block diagram of an embodiment of the distributed new energy monitoring system based on machine vision according to the present invention.

[0055] Explanation of icon numbers:

[0056] 10. Real-time image data acquisition module; 20. Energy conversion efficiency assessment value determination module; 30. Potential abnormal area marking module; 40. Cross-facility data preliminary aggregation set acquisition module; 50. Global status assessment index determination module; 60. Maintenance process triggering module. Detailed Implementation

[0057] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0058] like Figure 1 As shown, the first embodiment of the present invention proposes a distributed new energy monitoring method based on machine vision, including the following steps:

[0059] Step S100: Collect real-time image data of photovoltaic modules and wind power generation modules through image acquisition devices installed on renewable energy facilities, and use anomaly detection algorithms to process the real-time image data to obtain equipment anomaly location results.

[0060] The monitoring targets are distributed renewable energy facilities (photovoltaic modules in photovoltaic power plants and wind turbine modules in wind farms). Image acquisition devices (such as high-definition industrial cameras, infrared thermal imaging cameras, and 360° panoramic cameras, supporting real-time continuous shooting and environmental adaptive adjustment) are installed at key locations of the facilities (such as above the photovoltaic module array and near the wind turbine nacelle / blades). Two types of real-time image data are collected at preset frequencies (such as 5 minutes / time for photovoltaic modules and 10 minutes / time for wind turbine modules, with automatic encryption during abnormal periods): one is visible light / infrared images of photovoltaic modules (reflecting the surface stains, cracks, hot spots, shading, etc. of the modules); the other is dynamic images of wind turbine modules (reflecting blade damage, nacelle misalignment, hub abnormalities, etc.).

[0061] Anomaly detection algorithms from the field of machine vision (such as the YOLO (You Only Look Once) object detection algorithm based on deep learning, the CNN (Convolutional Neural Networks) image segmentation algorithm, and the infrared thermal imaging temperature difference recognition algorithm) are used to process real-time image data. The trained model identifies abnormal features in the image (such as hot spots in photovoltaic modules, cracks in blades, and surface coverings), accurately locates the physical location of the anomaly (such as the photovoltaic module number, specific blade section, and nacelle orientation), and outputs "equipment anomaly location results" containing "anomaly type (such as hot spot, crack, occlusion), location coordinates, anomaly degree (slight / moderate / severe), and image evidence," thereby realizing the visual monitoring of the surface and appearance of the equipment.

[0062] Step S200: Based on the equipment anomaly location results, analyze the attitude changes of the wind power generation components using the component attitude detection algorithm to determine the energy conversion efficiency evaluation value.

[0063] Using the equipment anomaly location results from step S100 as a reference, the focus is on wind power generation components (core components such as blades, nacelles, and hubs). Component attitude detection algorithms (such as attitude estimation algorithms based on visual SLAM (Simultaneous Localization and Mapping), feature point matching algorithms, and IMU inertial measurement unit-assisted calibration algorithms) are employed to analyze their attitude changes. By comparing the positional deviations of key feature points of the wind power generation components (such as blade tips and nacelle edges) in real-time images with standard attitude templates, attitude parameters (such as blade yaw angle, pitch angle, nacelle horizontal offset, and rotational speed stability) are calculated to determine whether there are operational deviations caused by attitude anomalies (such as blade deformation and nacelle tilt).

[0064] By combining attitude change data with the design parameters of wind power generation components (rated power, wind energy utilization coefficient) and real-time environmental data (wind speed, wind direction), an "energy conversion efficiency assessment value" is quantified through an energy conversion efficiency calculation model (efficiency = actual output power / theoretical maximum output power × 100%). This energy conversion efficiency assessment value reflects the impact of attitude anomalies on power generation efficiency and indirectly verifies the correlation of anomaly location results (such as blade cracks causing attitude deviation, which in turn reduces conversion efficiency), forming a logical link of "anomaly location - attitude change - efficiency decline".

[0065] Step S300: The energy conversion efficiency assessment value and the light and shadow distribution analysis results of the photovoltaic module are fused together. It is determined whether the deviation after fusion processing exceeds the preset threshold. If it does, it is marked as a potential abnormal area.

[0066] On the one hand, for photovoltaic modules, light and shadow distribution analysis algorithms (such as light and shadow simulation algorithms based on solar azimuth angle and altitude angle, and image grayscale value analysis algorithms) are used to process real-time image data to obtain "light and shadow distribution analysis results"—including the shading area on the module surface (such as trees, dust, and snow shading), the degree of local irradiance unevenness caused by light and shadow projection, and the quantitative value of the impact of shading on the power generation efficiency of photovoltaic cells; on the other hand, the energy conversion efficiency assessment value (wind power generation module) obtained in step S200 and the efficiency estimate value of photovoltaic modules (derived based on light and shadow distribution) are retrieved.

[0067] The two types of data are fused: a weighted fusion algorithm (e.g., 0.6 for efficiency assessment value and 0.4 for light and shadow distribution influence value) is used to calculate a comprehensive deviation value, which reflects the combined effect of "equipment anomaly + attitude / light and shadow influence" on energy output; a preset deviation threshold is set (which can be dynamically adjusted according to facility type, operating years, and power generation target, e.g., 15% for photovoltaic modules and 20% for wind power modules). If the fused comprehensive deviation value exceeds this threshold, the corresponding equipment area (e.g., a photovoltaic module or a wind turbine blade) is marked as a "potential anomaly area"—this potential anomaly area includes both identified explicit anomalies and areas of efficiency decline caused by implicit influencing factors such as attitude deviation and light and shadow shading, achieving comprehensive screening of anomalies.

[0068] Step S400: For potential abnormal areas, extract operational data based on data communication from the facility edge computing unit to obtain a preliminary cross-facility data aggregation set, where the operational data includes temperature parameters and vibration parameters.

[0069] For the potential abnormal areas marked in step S300, the data acquisition process of the facility edge computing unit (a lightweight computing module deployed in the photovoltaic array combiner box and the wind turbine tower control cabinet) is triggered. Real-time operating data of the area and related facilities are collected through data communication methods (such as 5G (the 5th-Generation Mobile Communication Technology), WiFi (Wireless Fidelity), LoRa (Long Range Radio), and Industrial Ethernet). The collection scope covers "potential abnormal area equipment + surrounding related equipment" (such as the inverter corresponding to a photovoltaic module and the gearbox corresponding to a wind turbine) to ensure the correlation of data.

[0070] The core of the collected operational data includes two types of key parameters: one is temperature parameters (such as the backsheet temperature of photovoltaic modules, cell temperature, oil temperature of wind turbine gearbox, and temperature of nacelle control cabinet); the other is vibration parameters (such as the vibration amplitude / frequency of wind turbine blades, vibration acceleration of gearbox, and vibration data of generator bearings), along with information such as equipment number, collection timestamp, and edge node identifier.

[0071] The collected operational data from multiple devices and of various types undergoes preliminary integration and processing: unifying data formats (such as JSON (JavaScript Object Notation) and CSV (Comma-Separated Values)), standardizing field naming (such as "pv_module_temp" and "wind_gearbox_vibration"), and removing obviously invalid data (such as temperature values ​​exceeding physical limits) to form a "preliminary cross-facility data aggregation set." This preliminary cross-facility data aggregation set breaks through the limitations of isolated single-device data, realizes the centralized aggregation of operational data related to potential abnormal areas, and provides quantitative data support for multi-source fusion analysis.

[0072] Step S500: Use a multi-source data fusion algorithm to process the preliminary aggregation set of cross-facility data and the analysis results of the image acquisition device to determine the global status assessment index, which is used to monitor the overall health status of renewable energy facilities.

[0073] Using the preliminary aggregated data set (operational data) from step S400 and the image analysis results (anomaly location, light and shadow distribution) from steps S100 / S300 as input, a deep integration analysis is performed using multi-source data fusion algorithms (such as the DS (Dempster-Shafer) fusion algorithm based on evidence theory, the weighted average fusion algorithm, and machine learning fusion models (such as random forest and LSTM (Long Short-Term Memory)).

[0074] Data alignment: Image analysis results are accurately correlated with operational data by timestamp and device number (e.g., the hot spot anomaly of a photovoltaic module at a certain time point corresponds to the backsheet temperature data at the same time).

[0075] Feature extraction: Extracting abnormal feature vectors (such as abnormal area and gray value deviation) from image data, and extracting statistical features (such as mean / variance of temperature and vibration peak value) from operational data.

[0076] Fusion computing: This method uses algorithms to weightedly fuse multi-dimensional features, constructing a global state assessment model and outputting a "global state assessment index"—a comprehensive quantitative value (ranging from 0-100 points, or graded as "Excellent / Good / Average / Dangerous"), which corely covers three assessment dimensions:

[0077] 1. Equipment health dimensions (abnormality type, severity, deviation of operating parameters).

[0078] 2. Energy efficiency dimension (conversion efficiency assessment value, degree of deviation).

[0079] 3. Related impact dimensions (the scope of the anomaly's impact on surrounding equipment, and the risk of its spread).

[0080] This indicator comprehensively reflects the overall health status of renewable energy facilities, including detailed information on local anomalies as well as the overall operational status, providing a core basis for monitoring and decision-making.

[0081] Step S600: Generate an early warning signal based on the global status assessment index. If the global status assessment index is lower than the safety threshold, the maintenance process is triggered.

[0082] Based on the global state assessment index in step S500, a differentiated early warning signal is generated:

[0083] If the indicator is within a safe range (e.g., 80-100 points, "Excellent / Good" level): generate a "normal operation" signal and continuously monitor data changes.

[0084] If the indicator is close to the safety threshold (e.g., 60-80 points, "normal" level): a "mild warning" signal is generated, prompting attention to the changing trends of potential abnormal areas and increasing the frequency of data collection.

[0085] If the indicator is below the safety threshold (e.g., ≤60 points, "dangerous" level): a "severe warning" signal is generated, and the maintenance process is automatically triggered.

[0086] The maintenance process includes: pushing early warning details (global status assessment indicators, anomaly location results, operational data, and image evidence) to the operation and maintenance management platform; generating standardized maintenance work orders (clearly defining maintenance objects, priorities, and suggested maintenance solutions, such as photovoltaic module cleaning, blade crack repair, and gearbox overhaul); assigning operation and maintenance personnel (based on geographical location and skill matching); and tracking maintenance progress. After maintenance is completed, data is re-collected for evaluation to confirm that indicators have returned to a safe range, thus forming a monitoring closed loop.

[0087] Furthermore, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S100 includes:

[0088] Step S110: Acquire real-time image data of photovoltaic modules and wind power generation modules through an image acquisition device installed on the renewable energy facility, preprocess the real-time image data, and use an image denoising tool to filter noise in the real-time image data to obtain the first image data after denoising.

[0089] The pixel values ​​of the first image data after denoising are obtained using the following formula:

[0090] (1)

[0091] In formula (1), This indicates the coordinates of the first image data after denoising. Pixel value at that location, This represents the raw pixel values ​​of the real-time image data acquired from the image acquisition device. Indicates the horizontal offset index. Indicates the offset index in the vertical direction. These represent the weighting coefficients of the denoising filter. Indicates the radius of the filtering window. The weight normalization factor is represented. The control logic of formula (1) is to "use the weighted average of the local neighborhood to smooth the noise while preserving the image details as much as possible", and to balance the denoising effect and detail preservation through weight allocation.

[0092] In renewable energy facilities, high-resolution cameras installed on photovoltaic arrays and wind turbines serve as image acquisition devices, capturing real-time image data of the solar panel surfaces of photovoltaic modules and the blades of wind turbines. These image acquisition devices are typically equipped with infrared or visible light sensors to ensure clear real-time image data under varying lighting conditions, providing a reliable basis for subsequent analysis. Specifically, this acquisition process involves timed shooting, such as every 5 minutes, to monitor changes in equipment status. If the facility is located in a windy area, the camera also integrates image stabilization to reduce motion blur. In one embodiment, the acquired real-time image data undergoes preprocessing. First, an image denoising tool, such as a median filter, is used to filter noise from the data. This denoising tool removes random noise by replacing pixel values ​​with the median of their neighborhood, resulting in denoised first image data. For example, in a photovoltaic module image, if noise caused by dust or light interference exists, the median filter scans a 3x3 pixel window, calculates the median, and replaces the center pixel, thus smoothing the image without losing edge details, which helps improve the accuracy of subsequent detection.

[0093] Step S120: Use an edge detection tool to extract the device outline features from the first image data, obtain key area images of photovoltaic modules and wind power generation modules, determine whether there are shape anomalies in the key area images, and obtain preliminary anomaly judgment results.

[0094] The following formula is used to determine whether shape anomalies exist by calculating the average deviation between the actually detected shape parameters and the standard reference values, and is used to generate preliminary anomaly assessment results:

[0095] (2)

[0096] In formula (2), A quantitative indicator representing the degree of shape abnormality. This represents the total number of feature points within the critical region. Indicates the first Shape parameters of each feature point The standard reference shape parameters are represented. The control logic of formula (2) is to "use the average deviation of each feature point from the standard shape as the basis for judging shape abnormality". The greater the deviation, the higher the degree of shape abnormality.

[0097] Next, edge detection tools such as the Canny operator are used to extract the equipment contour features from the first image data. The Canny operator first smooths the image using Gaussian filtering, then calculates the gradient intensity and direction, performs non-maximum suppression, and performs double thresholding to determine the edges, thereby obtaining the rectangular border of the photovoltaic module and the blade curve of the wind turbine module as key region images. Specifically, in a real-world wind farm scenario, if the blade image shows a curved contour, the Canny operator will highlight these edge lines, facilitating the system to automatically crop the blade tip or the connection area of ​​the photovoltaic panel as key regions. This further determines whether these key region images have shape anomalies, such as blade deformation or photovoltaic panel cracks, obtaining preliminary anomaly judgment results. This judgment is achieved by comparing the similarity between the extracted contour and a standard template. If the similarity is less than 0.8, it is considered a shape anomaly, triggering more in-depth analysis. In business operations, this allows for early detection of potential faults and reduces energy output losses.

[0098] Step S130: If the preliminary anomaly judgment result shows that there is a shape anomaly, then the pixel value distribution of the key area image is detected by grayscale analysis tool to obtain the specific location information of the anomaly distribution, and to determine whether there is a surface defect at the specific location of the anomaly distribution, so as to obtain detailed anomaly location information.

[0099] The following formula is used to define the criteria for determining the presence of surface defects at specific locations of abnormal distributions:

[0100] (3)

[0101] In formula (3), Representing coordinates Surface defect assessment results at the location This indicates the grayscale value at that location. Indicates the reference standard grayscale value. This represents the defect judgment threshold. A result of 1 indicates the presence of a defect, while 0 indicates normal.

[0102] If the initial anomaly assessment results indicate the presence of shape anomalies, a grayscale analysis tool is used to detect the pixel value distribution of the key area image. This grayscale analysis tool statistically analyzes the grayscale histogram of the image to identify the specific location information of the anomaly distribution. For example, in a photovoltaic panel image, if the grayscale value changes drastically in certain areas, it indicates that there may be stains or cracks. The system will locate the coordinates of these pixels and determine whether there are surface defects, such as scratches or corrosion, at the specific location of the anomaly distribution, thus obtaining detailed anomaly location information.

[0103] Step S140: Based on the detailed anomaly location information, compare it with a pre-established classification database to obtain the anomaly type and equipment component identifier, determine the anomaly category, and generate the final equipment anomaly location result.

[0104] The final equipment anomaly location result is obtained using the following formula:

[0105] (4)

[0106] In formula (4), This represents the final set of equipment anomaly location results. Indicates the equipment component identification code. Indicates the coordinates of the abnormal location. Indicates the timestamp of the exception occurrence. This represents the confidence score of the location results. This represents a detailed anomaly location information vector from the input. This refers to a pre-established classification database. This represents the function that generates the anomaly location results.

[0107] Based on detailed anomaly location information, a pre-established classification database is used for comparison. This database contains various anomaly samples, such as hot spot defects in photovoltaic modules or icing damage to wind turbine blades. Using feature matching algorithms such as the KNN (k-Nearest Neighbor) classifier, the location information is compared with the anomaly types and equipment component identifiers in the database to determine the anomaly category. For example, linear defects on blades are classified as cracks. The final equipment anomaly location result is generated, including anomaly location coordinates, type description, and suggested maintenance measures. This enables efficient equipment maintenance in operations and optimizes the operational efficiency of renewable energy.

[0108] Preferably, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S200 includes:

[0109] Step S210: Based on the abnormal positioning results of the equipment, obtain the real-time attitude data of the wind turbine components through the attitude detection tool, and use the smoothing tool to remove noise interference from the real-time attitude data to obtain the smoothed attitude change record.

[0110] The real-time attitude data of the wind turbine components is obtained using the following formula:

[0111] (5)

[0112] In formula (5), Indicates the wind turbine components at any time Real-time attitude data vector, Indicates pitch angle, Indicates the roll angle. Indicates the yaw angle. This represents the transformation matrix of the pose detection tool. The vector represents the original measurement data of the sensor. The control logic of formula (5) is to "use a preset transformation matrix to solve the original measurement value of the sensor into an intuitive attitude angle (pitch, roll, yaw)", so as to realize the standardized transformation from the original data to the attitude information.

[0113] The smoothed attitude change record is obtained using the following formula:

[0114] (6)

[0115] In formula (6), This represents the set of smoothed attitude change records. Indicates the first Posture change data over a time period, This represents a sequence of attitude data after noise processing. Indicates the data sampling time interval. This represents the function that generates attitude change records.

[0116] Based on the equipment anomaly location results, real-time attitude data of wind turbine components is acquired through attitude detection tools. For example, in a wind farm, attitude detection tools are gyroscopes and accelerometers integrated on the turbine tower. These sensors monitor the tilt angle and rotational attitude of the blades in real time, collecting data including yaw and pitch angle information in a three-dimensional coordinate system. Specifically, the working principle of this attitude detection tool is to calculate the attitude by measuring gravitational acceleration and angular velocity. The process involves converting sensor signals into digital data, and then applying a Kalman filter algorithm to fuse multi-source information to ensure data accuracy, thereby providing a reliable real-time attitude data foundation for subsequent processing. For example, a smoothing tool can be used to remove noise interference from the real-time attitude data to obtain a smoothed attitude change record. In actual operation, smoothing tools such as moving average filters will process the collected attitude data sequence, for example, by taking the average of 10 consecutive data points to smooth the curve. This helps to eliminate random noise caused by wind fluctuations or sensor jitter. The process includes first identifying noise peaks and then replacing them point by point with the neighborhood average, thereby generating a continuous attitude change record curve. This is particularly useful in variable weather conditions and can reflect the true dynamic changes of the blades.

[0117] Step S220: Based on the smoothed attitude change record, use an angle calculation tool to extract the blade angle deviation and determine whether the blade angle deviation exceeds the preset threshold. If it does, it is determined that there is an attitude abnormality.

[0118] The angular deviation of the blade is determined by calculating the angle between two vectors using the following formula:

[0119] (7)

[0120] In formula (7), Indicates the first angular deviation of individual blades Indicates the first The current attitude vector of each blade. This represents the reference standard attitude vector of the blade.

[0121] The following formula is used to define the conditions for judging posture anomalies:

[0122] (8)

[0123] In formula (8), This indicates the result of the posture anomaly assessment. This indicates the actual measured blade angle. This indicates the blade angle after smoothing. This indicates a preset angle deviation threshold. When the angle deviation exceeds the threshold, output 1 to indicate an abnormality; otherwise, output 0 to indicate normality.

[0124] Based on the smoothed attitude change records, an angle calculation tool is used to extract the blade angle deviation. For example, the angle calculation tool uses trigonometric functions to calculate the difference between the current attitude and the standard attitude. The specific process is to first define the standard blade angle as 0 degrees, and then calculate the deviation through the arctangent function. For example, if the record shows that the pitch angle changes from 5 degrees to 15 degrees, the deviation is 10 degrees. It is then determined whether the blade angle deviation exceeds the preset threshold, such as 8 degrees. If it does, it is determined that there is an attitude abnormality. This can detect deviations caused by fatigue or installation problems as early as possible, and help the maintenance team to intervene in a priority manner and reduce downtime.

[0125] Step S230: Obtain the vibration frequency distribution data of the wind turbine components during operation using vibration detection tools, compare and analyze the abnormal attitude conditions, and use frequency distribution tools to determine whether there are any related abnormal vibration characteristics.

[0126] The following formula is used to calculate the vibration frequency distribution data of the wind turbine components during operation:

[0127] (9)

[0128] In formula (9), Indicates the frequency of the wind turbine components Vibration frequency distribution density at that location This indicates the length of the time window for vibration signal acquisition. Indicates time The vibration amplitude signal, Represents the imaginary unit. Represents the base of the natural constant. Represents pi (π). Indicates the vibration frequency to be analyzed. Represents a time variable.

[0129] Vibration frequency distribution data of wind turbine components during operation is obtained by vibration detection tools. In wind power generation scenarios, vibration detection tools such as piezoelectric sensors are installed at the root of the blades to collect data with a frequency range from 0 to 100 Hz. The distribution data is converted into a spectrum graph through Fourier transform to display the peak frequency, thereby capturing the vibration mode during operation.

[0130] Comparative analysis is performed on abnormal attitude conditions, and frequency distribution tools are used to determine whether there are related abnormal vibration characteristics. For example, frequency distribution tools analyze whether abnormal peaks appear in the spectrum. If a peak of 20Hz higher than the normal level is detected when the attitude deviation is large, it is considered an abnormal vibration characteristic. The process involves comparing the attitude data and vibration data in time synchronization and calculating the correlation coefficient. If the coefficient exceeds 0.7, the feature correlation is confirmed, which helps to identify vibration amplification caused by attitude problems.

[0131] Step S240: Based on the abnormal vibration characteristics and attitude conditions, use an efficiency assessment tool combined with the influence of environmental wind speed and historical operating records to calculate the energy conversion efficiency assessment value.

[0132] The energy conversion efficiency assessment value is obtained using the following formula:

[0133] (10)

[0134] In formula (10), This represents the energy conversion efficiency assessment value. This represents the reference conversion efficiency under ideal conditions. The coefficient representing the impact of vibration anomalies on efficiency. Indicates the effective value of vibration. This represents the coefficient indicating the impact of attitude anomalies on efficiency. Indicates the attitude deviation angle. This represents the environmental wind speed correction factor. This represents the correction factor for historical operation records.

[0135] Based on the abnormal vibration characteristics and the aforementioned abnormal attitude, an efficiency assessment tool is used in conjunction with the influence of ambient wind speed and historical operating records to calculate the energy conversion efficiency assessment value. The principle of the efficiency assessment tool is based on an aerodynamic model. The process includes inputting the current wind speed (e.g., 10 m / s) and the historical average efficiency (e.g., 85%), and then adjusting the coefficients to consider the impact of abnormalities. For example, if the attitude deviation causes a 5% decrease in efficiency, the vibration characteristics are reduced by another 3%. Finally, the weighted average assessment value is calculated to be 77%. This can guide the optimization of operating strategies and improve overall energy output in business operations.

[0136] Furthermore, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S300 includes:

[0137] Step S310: Obtain real-time running data from the photovoltaic module using a data acquisition tool, combine it with monitoring indicators of the impact of ambient light, and use a light and shadow distribution analysis tool to extract the light and shadow distribution data on the surface of the photovoltaic module to obtain the corresponding distribution characteristic record.

[0138] The light and shadow distribution data on the surface of a photovoltaic module is obtained using the following formula:

[0139] (11)

[0140] In formula (11), This represents the characteristic value of light and shadow distribution on the surface of a photovoltaic module. This represents the total surface area of ​​the component. Indicates the location The light intensity measured at that location, This represents the ideal uniform light intensity. The illumination threshold representing shadow detection. The step function is used to determine the shadow area. The control logic of formula (11) is to "quantify the degree of light and shadow distribution on the surface of the component by averaging the area of ​​the light loss (deviation from ideal uniform lighting) in the shadow area". The larger the value, the more significant the light deviation caused by the shadow.

[0141] Real-time operational data is acquired from photovoltaic (PV) modules using data acquisition tools. For example, in a solar power plant, these tools can be multi-channel sensors mounted on the PV panels. These sensors monitor parameters such as current, voltage, and temperature in real time. The process involves converting analog signals into digital signals, which are then transmitted wirelessly to a central processing unit, ensuring the timeliness and integrity of the data. Specifically, these data acquisition tools operate based on the photovoltaic effect, where light generates photocurrent when it shines on the module surface. The acquisition process includes periodic sampling (e.g., once per second), combined with monitoring indicators of ambient light influence such as light intensity and angle. Light and shadow distribution analysis tools are then used to extract light and shadow distribution data from the module surface. For instance, these tools utilize image processing algorithms to analyze images of the module captured by a camera. Specifically, grayscale conversion is performed first, followed by the application of edge detection methods such as the Canny algorithm to identify shadow boundaries. This allows for the calculation of the shadow coverage area percentage, resulting in a record of the corresponding distribution characteristics. This is particularly useful in cloudy weather, reflecting the impact of uneven lighting on module performance.

[0142] Step S320: Based on the energy conversion efficiency assessment values ​​recorded and obtained according to the distribution characteristics, a data fusion tool is used for comprehensive processing to generate a fused feature dataset.

[0143] The fused feature dataset is obtained using the following formula:

[0144] (12)

[0145] In formula (12), This represents the fused feature dataset. This represents the total number of records representing the distribution characteristics. Indicates the first A distribution characteristic record, Indicates the first Weight coefficients for each distribution characteristic, This represents the total number of energy conversion efficiency assessment values. Indicates the first One energy conversion efficiency assessment value, Indicates the first The weighting coefficients for each efficiency evaluation value.

[0146] Based on the energy conversion efficiency assessment values ​​recorded and obtained from the distribution characteristics, a data fusion tool is used for comprehensive processing. The principle of the data fusion tool is to integrate multi-source data to improve accuracy. The process includes weighted fusion of distribution characteristics such as the proportion of shadows and efficiency values ​​such as the current conversion rate of 15%. For example, the joint probability distribution is calculated using the Bayesian method. The data is first standardized, and then the weights are iteratively updated to generate the fused feature dataset, which helps to comprehensively evaluate the component status.

[0147] Step S330: For the fused feature dataset, use a deviation calculation tool to compare it with a preset threshold range. If it exceeds the preset threshold range, it is marked as abnormal deviation data.

[0148] The following formula is used to define the labeling criteria for outlier data:

[0149] (13)

[0150] In formula (13), Indicates the first Anomaly labeling results for each data point, Indicates the first One observation value, Indicates the first One reference base value, This indicates the preset upper threshold range. This indicates the preset lower threshold range. When the deviation exceeds the preset threshold range, it is marked as abnormal deviation data.

[0151] For the fused feature dataset, a deviation calculation tool is used to compare it with a preset threshold range. The deviation calculation tool calculates the difference between the dataset and the standard value using statistical methods. For example, if the proportion of shadows in the fused data exceeds 10%, while the preset threshold range is 5% to 8%, it is marked as abnormal deviation data. The process involves comparing each item and applying confidence intervals to determine whether it exceeds the limit.

[0152] Step S340: Locate the abnormal deviation data using the anomaly marking tool, and determine the specific location distribution of potential abnormal areas by combining the analysis results of historical records.

[0153] The specific location distribution of potential anomaly regions is obtained using the following formula:

[0154] (14)

[0155] In formula (14), Indicates the first The specific location coordinates of each potential anomaly region Indicates the first A set of coordinates for candidate regions Indicates historical comparison at position The analysis results value, Indicates position An abnormal threshold determination function at a certain location.

[0156] Anomaly marking tools are used to locate abnormal deviation data. These tools use spatial mapping algorithms to project data onto the component mesh model and combine the analysis results of historical records, such as similar deviation patterns over the past week, to determine the specific location distribution of potential abnormal areas. For example, if historical records show that a certain area has repeated shadow deviations, it can be located at the edge of the component. This can guide targeted inspections during maintenance.

[0157] Preferably, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S400 includes:

[0158] Step S410: For potential abnormal areas, obtain temperature parameter data from the facility edge computing unit through data communication protocol to obtain a preliminary temperature dataset.

[0159] The preliminary temperature dataset is derived using the following formula:

[0160] (15)

[0161] In formula (15), This represents the initial temperature data set. Indicates the first The original temperature measurement values ​​of each sampling point This represents the set of indexes of identified potential anomalous regions. This represents the threshold parameter for determining temperature anomalies.

[0162] For potentially abnormal areas, temperature parameter data is obtained from the facility edge computing unit via a data communication protocol to obtain a preliminary temperature dataset. Specifically, this data communication protocol can be a variant based on the Modbus protocol. In photovoltaic power plants, the facility edge computing unit, such as a smart gateway device installed near the modules, collects data from temperature sensors in real time. For example, the sensors use thermistor principles; when the surface temperature of the photovoltaic module changes, the resistance value adjusts accordingly. A query command is sent via the data communication protocol, and the edge unit responds by transmitting a data packet, including a timestamp and temperature value sequence, forming a preliminary temperature dataset. This process ensures low-latency data transmission, especially in large solar arrays, enabling monitoring of abnormal areas across multiple modules.

[0163] Step S420: Based on the preliminary temperature dataset, obtain vibration parameter data from the facility edge computing unit through a data communication protocol to obtain the preliminary vibration dataset.

[0164] The vibration parameter data obtained from the temperature dataset are derived using the following formula:

[0165] (16)

[0166] In formula (16), This represents vibration parameter data obtained based on a temperature dataset. This indicates the total number of vibration sampling points. Indicates the first Vibration amplitude at each sampling point Indicates the first The frequency corresponding to each sampling point Indicates the first Temperature values ​​corresponding to each sampling point Indicates the reference temperature value. This represents the standard deviation of temperature.

[0167] Based on the preliminary temperature dataset, vibration parameter data is obtained from the facility edge computing unit via a data communication protocol to obtain a preliminary vibration dataset. In one embodiment, this step uses temperature data as a trigger condition. If the temperature dataset shows that a certain area exceeds the normal range, such as 40 degrees Celsius, the protocol will initiate a subsequent query. The vibration parameter data comes from accelerometers, which operate based on the piezoelectric effect. When components vibrate due to wind or mechanical stress, they generate electrical signals. The edge computing unit processes these signals and packages them into a dataset via the protocol, including vibration frequency and amplitude values. The resulting preliminary vibration dataset reflects the mechanical stability of the components and complements the temperature data.

[0168] Step S430: Use a data fusion tool to aggregate the preliminary temperature dataset and the preliminary vibration dataset to generate a preliminary data subset across facilities.

[0169] The following formula is used to aggregate the temperature and vibration data using a weighted average method:

[0170] (17)

[0171] In formula (17), This represents the merged cross-facility data subset. This represents the weighting coefficient for temperature data. This represents the total number of temperature data samples. Indicates the first One temperature data point, The weighting coefficients represent the vibration data. This represents the total number of vibration data samples. Indicates the first Vibration data points.

[0172] A data fusion tool is used to aggregate the preliminary temperature and vibration datasets to generate a cross-facility preliminary data subset. Specifically, the data fusion tool can be a Kalman filter-based software module. Its principle is to integrate multi-source data through an iterative algorithm. First, the temperature and vibration datasets are time-aligned, for example, by unifying their sampling rates to once per minute. Then, a weighted average is calculated. After fusion, a subset containing comprehensive indicators such as the temperature-vibration correlation coefficient is generated. This cross-facility preliminary data subset covers different parts of the entire photovoltaic facility, helping to identify potential correlations between thermal stress and vibration anomalies.

[0173] Step S440: Use a data verification tool to determine whether the preliminary cross-facility data subset meets the preset integrity threshold. If it does, determine the preliminary cross-facility data aggregation set.

[0174] The following formula describes the logic of including each facility's data in the final aggregate set only when the data integrity requirements are met:

[0175] (18)

[0176] In formula (18), This represents the final, preliminary aggregated set of cross-facility data. Indicates the first A subset of data from each facility This represents the total number of facilities participating in the aggregation. Indicates the first Completeness score of facility data Indicates the minimum integrity requirement. This indicates that the indicator function takes the value 1 when the condition is met, and 0 otherwise.

[0177] The data verification tool determines whether the preliminary cross-facility data subset meets a preset integrity threshold. If it does, the preliminary cross-facility data aggregation set is determined. In one embodiment, the data verification tool uses parity checking and CRC (Cyclic Redundancy Check) methods. The process includes scanning each data point in the subset, checking whether the proportion of missing values ​​is lower than a preset threshold, such as 2%, and verifying data consistency, such as the logical relationship between temperature and vibration values. If the subset is complete and has no abnormal fluctuations, it is confirmed as a preliminary aggregation set. This aggregation set can be used for subsequent anomaly diagnosis, providing reliable basic data in photovoltaic maintenance.

[0178] Furthermore, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S500 includes:

[0179] Step S510: Use a data alignment tool to perform time-stamp matching between the preliminary aggregated set of cross-facility data and the analysis results of the image acquisition device to obtain a time-aligned multi-source dataset.

[0180] Time-aligned multi-source datasets are derived using the following formula:

[0181] (19)

[0182] In formula (19), This indicates a time-aligned multi-source dataset. This indicates the first data aggregation set across facilities. One data record, The first in the analysis results of the image acquisition device One data record, Indicates the timestamp of cross-facility data records. The timestamp representing the image analysis results This indicates the threshold range for timestamp matching.

[0183] A time-aligned multi-source dataset is obtained by matching the timestamps of the initial aggregated data from across facilities with the analysis results from image acquisition devices using a data alignment tool. Specifically, this data alignment tool is a software module based on a time synchronization algorithm. Its principle is to achieve matching by comparing timestamps from different data sources. For example, in a photovoltaic power plant monitoring system, the initial aggregated data from across facilities includes temperature and vibration parameters collected from multiple edge computing units. These parameters have timestamps accurate to the second. Image acquisition devices, such as cameras installed on solar arrays, analyze image results from the component surfaces, including thermal imaging or visible light image anomaly spot detection. These analysis results also have timestamps. The tool first scans each data point in the aggregated dataset, looking for the closest match to the timestamp of the image result. If the time difference is less than a preset value, such as 5 seconds, they are aligned into a pair, forming a time-aligned multi-source dataset. This process ensures the temporal consistency of the data. In large photovoltaic facilities, it can effectively integrate real-time operational data with visual information, aiding in subsequent analysis of potential faults.

[0184] Step S520: Based on the time-aligned multi-source dataset, use feature extraction tools to extract temperature features, vibration features, and image texture features to obtain a multi-feature vector set.

[0185] Multiple feature vector sets are derived using the following formula:

[0186] (20)

[0187] In formula (20), Represents a set of multiple feature vectors. This represents the temperature feature vector at the reference time point. This represents the vibration characteristic vector at the reference time point. This represents the image texture feature vector at the reference time point. It represents a unified reference time base.

[0188] The temperature eigenvector is derived using the following formula:

[0189] (twenty one)

[0190] In formula (21), Indicates the number of temperature sensors. Indicates the first The measured values ​​of a temperature sensor, Indicates the first Temperature change rate of each sensor, , , These represent the weighting coefficients for the linear, differential, and nonlinear terms, respectively.

[0191] Image texture feature vectors are derived using the following formula:

[0192] (twenty two)

[0193] In formula (22), and These represent the number of rows and columns of the image, respectively. Indicates position gray-level co-occurrence matrix elements at that location, This indicates the mean characteristic of that location. Representing variance characteristics, The control logic of formula (22) is to "integrate the texture information of the image by taking the gray-level co-occurrence matrix as the core and combining the weighted sum of local statistical features", which not only reflects the spatial distribution of texture (gray-level co-occurrence matrix) but also incorporates local statistical attributes.

[0194] Based on a time-aligned multi-source dataset, a feature extraction tool is used to extract temperature, vibration, and image texture features, resulting in a multi-feature vector set. The feature extraction tool is based on a Fourier transform algorithm module, which converts raw data into feature representations. First, for temperature features, the tool calculates the average temperature, peak temperature, and temperature fluctuation rate from the time-aligned dataset. For example, in photovoltaic module monitoring, if the dataset shows that the temperature in a certain area rises from 30 degrees Celsius to 50 degrees Celsius, the tool will extract the fluctuation rate as a feature. Next, vibration features are extracted, such as the mean and peak amplitude of the vibration frequency, derived from the signal from the accelerometer. For image texture features, the tool analyzes the gray-level co-occurrence matrix in the image results and calculates indicators such as contrast and correlation. For example, it detects uneven texture due to cracks on the module surface. These features are combined into a vector set, with each vector representing a multi-dimensional description at a given time point. This multi-feature vector set comprehensively captures multiple aspects of the facility's condition.

[0195] Step S530: Determine whether the similarity of the multiple feature vector sets is lower than a preset threshold. If the similarity is lower than the preset threshold, remove outliers using a data cleaning tool to obtain a cleaned feature set.

[0196] The similarity of the entire vector set is measured by averaging the cosine similarity between all pairs of feature vectors using the following formula:

[0197] (twenty three)

[0198] In formula (23), This represents the overall similarity of multiple feature vector sets. This represents the total number of eigenvectors. Indicates the first 1 eigenvector Indicates the first 1 eigenvector.

[0199] The cleaned feature set obtained after removing outliers using statistical methods is defined by the following formula:

[0200] (twenty four)

[0201] In formula (24), Represents the feature set after cleaning. Indicates the first 1 eigenvector Represents the mean vector of a set of multiple feature vectors. Indicates standard deviation, This indicates the threshold for outlier detection.

[0202] The similarity of multiple feature vector sets is determined to be below a preset threshold. If the similarity is below the threshold, outliers are removed using data cleaning tools, resulting in a cleaned feature set. Specifically, the similarity determination uses cosine similarity calculation, which compares the similarity values ​​of adjacent vectors within the vector set. For example, if the preset threshold is 0.8, a result below this value indicates that the data may contain noise or anomalies. Then, data cleaning tools, such as modules based on the Isolation Forest algorithm, scan the vector set, identify, and remove outliers. For instance, in a photovoltaic scenario, if a temperature feature vector shows an abnormally high value, such as 100 degrees Celsius, while neighboring vectors are within the normal range, the tool will mark it as noise and delete it. The cleaned feature set is purer and more reliable. This step improves data quality and provides a foundation for subsequent fusion.

[0203] Step S540: Based on the cleaned feature set, use principal component analysis to perform fusion and determine the global state evaluation index.

[0204] The following formula is used to construct a comprehensive evaluation index by weighted fusion of multiple principal components:

[0205] (25)

[0206] In formula (25), This represents the global state assessment index. Indicates the number of principal components selected. Indicates the first The weight coefficients of each principal component, Indicates the first The standardized scores of each principal component. Indicates the first The eigenvalues ​​of the principal components This represents the total number of all principal components. The control logic of formula (25) is to "weight the scores of multiple principal components into a comprehensive index by weighting the information contribution (eigenvalue proportion) of the principal components", which not only utilizes the advantage of principal component dimensionality reduction, but also preserves the differences in information importance of each principal component.

[0207] Based on the cleaned feature set, principal component analysis (PCA) is used for fusion to determine a global state assessment index. In one embodiment, PCA is a statistical method based on the covariance matrix. Its principle is to fuse multiple features into a few principal components through dimensionality reduction. First, the covariance of the cleaned feature set is calculated, and the top few principal components, such as the combined vector of temperature-vibration-texture, are extracted. For example, in a photovoltaic power plant, these principal components may represent the overall thermomechanical stability. Then, a global index is calculated based on the principal component scores, such as an assessment value from 0 to 1. The lower the value, the higher the potential anomaly. This index can guide maintenance decisions and achieve preventative management of the facility.

[0208] Preferably, in the distributed new energy monitoring method based on machine vision proposed in this embodiment, step S600 includes:

[0209] Step S610: Obtain real-time data stream from global status assessment indicators through data integration tools, compare the data stream with preset security thresholds, determine whether it is below the threshold, and obtain an anomaly judgment result.

[0210] The following formula enables the comparison and judgment function between the data stream and the preset threshold:

[0211] (26)

[0212] In formula (26), Indicates time The abnormal judgment result, Indicates the real-time data stream at time... The value, This represents the preset lower limit of the safety threshold. When the real-time data stream value is lower than the safety threshold, the output is 1 to indicate an anomaly; otherwise, the output is 0 to indicate normal operation. The control logic of formula (26) is to determine an anomaly when the real-time data is lower than the lower limit of the safety threshold, and otherwise determine normal operation.

[0213] The process of obtaining real-time data streams from global state assessment indicators through data integration tools is a streaming-based mechanism. Its principle lies in extracting key parameters from the indicators in real time. For example, in a wind power facility monitoring system, global state assessment indicators include the overall vibration level and power output efficiency of the turbine. These global state assessment indicators are transformed into continuous data streams. Data integration tools, such as message queue systems like Apache Kafka, subscribe to indicator updates, pulling the latest values ​​every second to form a data stream. This data stream is then compared with preset safety thresholds, such as a vibration level not exceeding 5 Hz or a power efficiency not lower than 90%. The comparison process involves comparing each value in the data stream with the threshold point by point. If a value, such as a vibration frequency reaching 6 Hz, is judged to be below the threshold, an anomaly judgment result is obtained. This result is marked as an anomaly in Boolean form, helping to identify potential risks in a timely manner.

[0214] Step S620: Based on the anomaly judgment result, if it is below the safety threshold, generate an early warning signal for the abnormal data using the signal generation tool, and determine the signal type and priority.

[0215] The warning signal triggering conditions are defined using the following formula:

[0216] (27)

[0217] In formula (27), Indicates the status of the warning signal being triggered. This represents the numerical value of the anomaly detection result. This represents the safety threshold, indicating the value of the abnormal judgment result. Below the safety threshold The warning signal is triggered when the abnormal judgment result is lower than the safety threshold, otherwise it is not triggered. The control logic of formula (27) is to trigger the warning when the abnormal judgment result is lower than the safety threshold, otherwise it is not triggered.

[0218] The following formula is used to calculate priority based on the degree of deviation from the safety threshold and risk factors:

[0219] (28)

[0220] In formula (28), This indicates the priority value of the warning signal. Indicates the severity of outlier data. Indicates the safety threshold. Indicates risk factors, This represents the threshold deviation weighting coefficient. The risk weight coefficient is represented. The control logic of formula (28) is to calculate the priority of the warning signal by weighted summation of "the degree of data deviation from the safety threshold" and "risk factor". The more serious the deviation and the higher the risk, the higher the priority.

[0221] The most suitable signal type is determined by maximizing the weighted evaluation value using the following formula:

[0222] (29)

[0223] In formula (29), This indicates the classification result of a specific signal type. Indicates the first Weights for different signal types Indicates the first Evaluation functions for various signal types Represents the feature vector of an abnormal event. This indicates the current system load status. The control logic of formula (29) is based on the characteristics of abnormal events and the system load. It performs a weighted calculation of "evaluation value + weight" for each signal type and finally selects the signal type with the largest weighted evaluation value as the result.

[0224] Based on the anomaly assessment results, if the data falls below a safety threshold, a warning signal is generated for the abnormal data using a signal generation tool. This process is based on a rule-driven signal generation engine. Specifically, the signal generation tool can be a software module integrated into the monitoring platform. First, it analyzes the specific attributes of the abnormal data, such as whether the anomaly is excessive vibration or low efficiency. Then, it determines the signal type, such as "high-priority vibration anomaly" or "medium-priority efficiency decline." Priorities are graded according to severity; for example, those affecting power generation output are considered high-priority. The tool generates a signal packet containing a timestamp, anomaly description, and priority. In a wind turbine scenario, if the data stream shows abnormal bearing vibration, the signal generation tool will generate a red high-priority signal, along with a suggestion to check the bearing's metadata. This ensures the targeted nature and operability of the warning.

[0225] Step S630: Obtain the early warning signal through the process scheduling tool, match the pre-established maintenance process template according to the signal type and priority, determine whether it needs to be executed immediately, and obtain the process trigger instruction.

[0226] The following formula is used to define the signal type and priority matching function for warning signals:

[0227] (30)

[0228] In formula (30), This represents the signal type and priority matching function for warning signals. Indicates the type of signal input for the warning signal. Indicates the priority of the input warning signal. Indicates the first A pre-established maintenance process template, Weighting coefficients representing signal type Weighting coefficients representing priority This represents the similarity matching function. template Corresponding signal type, template Corresponding priority, This represents the set of all preset maintenance process templates. The control logic of formula (30) is to select the template with the largest matching value from the preset templates by using the comprehensive matching value of "signal type similarity (weighted) + priority similarity (weighted)" as the matching result of the current warning signal.

[0229] The following formula is used to define the conditions for immediate execution of a warning signal:

[0230] (31)

[0231] In formula (31), This indicates that the conditional function will be executed immediately. This indicates the priority value of the current warning signal. Indicates the current timestamp. Indicates the timestamp of signal detection. This indicates the priority threshold for immediate execution. The maximum allowable delay time is indicated by the immediate execution judgment function. An output of 1 indicates that immediate execution is required, and an output of 0 indicates that immediate execution is not required. The control logic of formula (31) is that when the priority of the warning signal meets the standard and the current delay time detected is within the allowable range, it is determined that immediate execution is required; otherwise, it is not required.

[0232] The process trigger instruction is derived using the following formula:

[0233] (32)

[0234] In formula (32), This indicates the final generated process trigger instruction. Indicates the instruction generation function. This represents the feature vector of the acquired warning signal. This indicates the identifier of the matched maintenance process template. This indicates that the judgment result will be executed immediately. The feature fusion operator is indicated. The control logic of formula (32) is to fuse the three pieces of information—the early warning signal features, the matched maintenance process template, and the immediate execution judgment result—to obtain the final process trigger instruction.

[0235] The process of obtaining the early warning signal through a process scheduling tool and matching it with a pre-established maintenance process template based on the signal type and priority is an automated workflow management method. Its principle is to associate the signal with the template through a matching algorithm. For example, a process scheduling tool such as a BPMN (Business Process Model and Notation) engine will receive the signal and parse its type and priority. The template library has pre-stored flowcharts such as "Immediate Stop Inspection" or "Planned Maintenance". If the signal is a high-priority vibration anomaly, the tool will match the "Emergency Response Template", determine that it needs to be executed immediately, and generate a process trigger instruction. This process trigger instruction includes the execution time and the allocation of responsible persons. In actual wind farms, such matching can respond quickly and prevent small problems from evolving into major failures.

[0236] Step S640: Based on the process trigger instruction, use a data feedback tool to associate and update the trigger information with the global status assessment indicators, generate status monitoring logs for the updated data, and determine the direction of subsequent monitoring.

[0237] The updated global state evaluation metric is derived using the following formula:

[0238] (33)

[0239] In formula (33), This represents the updated global state assessment metric. This indicates the trigger information at the current moment. This represents the global state evaluation index at the current moment. This represents the weighting coefficient of the triggering information. The retention coefficient represents the historical state. This indicates the adjustment parameter for correlation updates. The function represents the association between the trigger information and the global state. The control logic of formula (33) is to calculate the updated global state evaluation index by combining the "weighted value of the current trigger information, the retained value of the historical global state, and the association adjustment value between the trigger information and the current state".

[0240] The generated values ​​of the status monitoring logs are obtained using the following formula:

[0241] (34)

[0242] In formula (34), Indicates the first The generated values ​​of each status monitoring log. Indicates the number of dimensions in the updated data. Indicates the first Importance weights of dimensional data Indicates the first Updated data in the first The values ​​of each monitoring point This represents a tiny constant to prevent the logarithmic operation from producing a zero value. The control logic of formula (34) is to perform "adding a tiny constant + taking the logarithm + multiplying the weight" on the updated data of each dimension, and then sum the processing results of all dimensions to obtain the generated value of the status monitoring log.

[0243] The direction of subsequent monitoring is determined by the following formula:

[0244] (35)

[0245] In formula (35), This indicates the confirmed direction for subsequent monitoring. An index indicating the optional monitoring direction. This represents the total number of monitoring and evaluation factors. Indicates the first The priority weight of each evaluation factor Indicates the first The monitoring direction is in the first The control logic of formula (35) is to multiply the "weight × score" results of each evaluation factor for each monitoring direction to obtain a comprehensive value, and select the monitoring direction with the largest comprehensive value as the subsequent monitoring direction.

[0246] Based on the process trigger command, a mechanism is used to associate and update the trigger information with global status evaluation indicators using data feedback tools. The principle is a closed-loop feedback loop. Specifically, the data feedback tool, such as a database update module, associates the maintenance action in the trigger command, such as "maintenance team dispatched," with the indicator, updates the indicator's status field, and then generates a status monitoring log based on the updated data. The log record is as follows: "2023-10-01 14:00 Vibration abnormality triggered maintenance, estimated recovery time 2 hours." Based on log analysis, the subsequent monitoring direction is determined, such as increasing the monitoring frequency of bearing temperature. This process can achieve continuous optimization in wind power facilities to ensure system health.

[0247] Please see Figure 2 This embodiment provides a distributed new energy monitoring system based on machine vision, including the aforementioned distributed new energy monitoring method based on machine vision. It includes a real-time image data acquisition module 10, an energy conversion efficiency assessment value determination module 20, a potential anomaly area marking module 30, a cross-facility data preliminary aggregation set acquisition module 40, a global status assessment index determination module 50, and a maintenance process triggering module 60. The real-time image data acquisition module 10 is used to acquire real-time image data of photovoltaic modules and wind power generation modules through image acquisition devices installed on renewable energy facilities, and uses anomaly detection algorithms to process the real-time image data to obtain equipment anomaly location results. The energy conversion efficiency assessment value determination module 20 is used to analyze the attitude changes of wind power generation modules based on the equipment anomaly location results and a module attitude detection algorithm to determine the energy conversion efficiency assessment value. The potential anomaly area marking module... The module 30 is used to fuse the energy conversion efficiency assessment value with the light and shadow distribution analysis results of the photovoltaic module, and determine whether the deviation after fusion processing exceeds a preset threshold. If it does, it is marked as a potential abnormal area. The cross-facility data preliminary aggregation set acquisition module 40 is used to extract the operation data based on the data communication method from the facility edge computing unit for potential abnormal areas to obtain the cross-facility data preliminary aggregation set, where the operation data includes temperature parameters and vibration parameters. The global status assessment index determination module 50 is used to process the cross-facility data preliminary aggregation set and the analysis results of the image acquisition device using a multi-source data fusion algorithm to determine the global status assessment index, where the global status assessment index is used to monitor the overall health status of the renewable energy facility. The maintenance process triggering module 60 is used to generate an early warning signal based on the global status assessment index. If the global status assessment index is lower than the safety threshold, the maintenance process is triggered.

[0248] The distributed new energy monitoring method and system based on machine vision provided in this embodiment, compared with existing technologies, acquires real-time image data through an image acquisition device, combines anomaly detection algorithms and attitude detection algorithms to locate equipment anomalies and assess energy conversion efficiency, and simultaneously integrates light and shadow distribution analysis results to identify potential anomaly areas. For anomaly areas, operational data is extracted from edge computing units, and a multi-source data fusion algorithm is used to integrate image analysis and operational parameters to generate a global status assessment index, which triggers early warning and maintenance processes. The most prominent technical effect of this embodiment is that through multi-dimensional data fusion and intelligent algorithm integration, it significantly improves the operational reliability and maintenance efficiency of renewable energy facilities, providing important technical support for the sustainable development of green energy.

[0249] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A distributed new energy monitoring method based on machine vision, characterized in that, Includes the following steps: S100. Real-time image data of photovoltaic modules and wind power generation modules are acquired by an image acquisition device installed on renewable energy facilities, and the real-time image data is processed by an anomaly detection algorithm to obtain the equipment anomaly location result. S200. Based on the abnormal positioning results of the equipment, and combined with the component attitude detection algorithm, analyze the attitude change of the wind power generation component to determine the energy conversion efficiency evaluation value. The energy conversion efficiency assessment value is obtained using the following formula: ； in, This represents the energy conversion efficiency assessment value. This represents the reference conversion efficiency under ideal conditions. The coefficient representing the impact of vibration anomalies on efficiency. Indicates the effective value of vibration. This represents the coefficient indicating the impact of attitude anomalies on efficiency. Indicates the attitude deviation angle. This represents the environmental wind speed correction factor. This represents the correction factor for historical operation records; S300: The energy conversion efficiency assessment value and the light and shadow distribution analysis result of the photovoltaic module are fused together. It is determined whether the deviation after fusion processing exceeds a preset threshold. If it does, it is marked as a potential abnormal area. S400. For the potential abnormal area, extract the operation data based on the data communication method from the facility edge computing unit to obtain a preliminary cross-facility data aggregation set, wherein the operation data includes temperature parameters and vibration parameters; S500. The analysis results of the preliminary aggregated set of cross-facility data and the image acquisition device are processed by a multi-source data fusion algorithm to determine the global status assessment index, wherein the global status assessment index is used to monitor the overall health status of renewable energy facilities. S600. Generate an early warning signal based on the global status assessment index. If the global status assessment index is lower than the safety threshold, trigger the maintenance process.

2. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S100 includes: S110. Real-time image data of photovoltaic modules and wind power generation modules are acquired through an image acquisition device installed on renewable energy facilities. The real-time image data is preprocessed, and noise is filtered out of the real-time image data using an image denoising tool to obtain the first image data after denoising. S120. Use an edge detection tool to extract the device outline features in the first image data, obtain key area images of photovoltaic modules and wind power generation modules, determine whether there are shape anomalies in the key area images, and obtain preliminary anomaly judgment results. S130. If the preliminary anomaly judgment result shows that there is a shape anomaly, then the pixel value distribution of the key area image is detected by grayscale analysis tool to obtain the specific location information of the anomaly distribution, to determine whether there is a surface defect at the specific location of the anomaly distribution, and to obtain detailed anomaly location information. S140. Based on the detailed anomaly location information, a pre-established classification database is used for comparison to obtain the anomaly type and equipment component identifier, determine the anomaly category, and generate the final equipment anomaly location result.

3. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S200 includes: S210. Based on the abnormal positioning results of the equipment, the real-time attitude data of the wind turbine component is obtained through the attitude detection tool, and the noise interference of the real-time attitude data is removed by the smoothing tool to obtain the smoothed attitude change record. S220. Based on the smoothed attitude change record, use an angle calculation tool to extract the blade angle deviation, and determine whether the blade angle deviation exceeds a preset threshold. If it does, then it is determined that there is an attitude abnormality. S230. Obtain the vibration frequency distribution data of the wind turbine component during operation using a vibration detection tool, compare and analyze the abnormal attitude conditions, and use a frequency distribution tool to determine whether there are any related abnormal vibration characteristics. S240. Based on the abnormal vibration characteristics and the abnormal posture, an energy conversion efficiency assessment value is calculated using an efficiency assessment tool combined with the influence of environmental wind speed and historical operating records.

4. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S300 includes: S310. Real-time operating data is obtained from the photovoltaic module through a data acquisition tool. Combined with monitoring indicators of the impact of ambient light, the light and shadow distribution data on the surface of the photovoltaic module is extracted using a light and shadow distribution analysis tool to obtain the corresponding distribution characteristic record. S320. Based on the energy conversion efficiency evaluation values ​​recorded and obtained according to the distribution characteristics, a data fusion tool is used for comprehensive processing to generate a fused feature dataset. S330. For the fused feature dataset, use a deviation calculation tool to compare it with a preset threshold range. If it exceeds the preset threshold range, it is marked as abnormal deviation data. S340. The abnormal deviation data is located by using an anomaly marking tool, and the specific location distribution of potential abnormal areas is determined by combining the analysis results of historical records.

5. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S400 includes: S410. For the potential abnormal area, temperature parameter data is obtained from the facility edge computing unit through a data communication protocol to obtain a preliminary temperature dataset. S420. Based on the preliminary temperature dataset, obtain vibration parameter data from the facility edge computing unit through the data communication protocol to obtain a preliminary vibration dataset; S430. Use a data fusion tool to aggregate the preliminary temperature dataset and the preliminary vibration dataset to generate a cross-facility preliminary data subset; S440. Use a data verification tool to determine whether the preliminary cross-facility data subset meets a preset integrity threshold. If it does, determine the preliminary cross-facility data aggregation set.

6. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S500 includes: S510. The data alignment tool is used to time-stamp match the preliminary aggregated set of cross-facility data with the analysis results of the image acquisition device to obtain a time-aligned multi-source dataset. S520. Based on the time-aligned multi-source dataset, use a feature extraction tool to extract temperature features, vibration features, and image texture features to obtain a multi-feature vector set; S530. Determine whether the similarity of the multi-feature vector set is lower than a preset threshold. If the similarity is lower than the preset threshold, remove outliers using a data cleaning tool to obtain a cleaned feature set. S540. Based on the cleaned feature set, principal component analysis is used to perform fusion and determine the global state evaluation index.

7. The distributed new energy monitoring method based on machine vision according to claim 1, characterized in that, Step S600 includes: S610. Obtain real-time data stream from global status assessment indicators through data integration tools, compare the data stream with a preset security threshold, determine whether it is below the threshold, and obtain an anomaly judgment result. The following formula enables the comparison and judgment function between the data stream and the preset threshold: ； in, Indicates time The abnormal judgment result, Indicates the real-time data stream at time... The value, This indicates the preset lower limit of the safety threshold. When the real-time data stream value is lower than the safety threshold, output 1 to indicate an abnormality; otherwise, output 0 to indicate normality. S620. Based on the anomaly judgment result, if it is lower than the safety threshold, a warning signal is generated for the abnormal data using a signal generation tool, and the signal type and priority are determined. The warning signal triggering conditions are defined using the following formula: ； in, Indicates the status of the warning signal being triggered. This represents the numerical value of the anomaly detection result. This represents the safety threshold, indicating the value of the abnormal judgment result. Below the safety threshold A warning signal is generated only if the warning is triggered on time; otherwise, it is not triggered. The following formula is used to calculate priority based on the degree of deviation from the safety threshold and risk factors: ； in, This indicates the priority value of the warning signal. Indicates the severity of outlier data. Indicates the safety threshold. Indicates risk factors, This represents the threshold deviation weighting coefficient. This represents the risk weighting coefficient; The most suitable signal type is determined by maximizing the weighted evaluation value using the following formula: ； in, This indicates the classification result of a specific signal type. Indicates the first Weights for different signal types Indicates the first Evaluation functions for various signal types Represents the feature vector of an abnormal event. Indicates the current system load status; S630. Obtain the warning signal through the process scheduling tool, match the pre-established maintenance process template according to the signal type and priority, determine whether it needs to be executed immediately, and obtain the process triggering instruction. S640. Based on the process trigger instruction, use a data feedback tool to associate and update the trigger information with the global status evaluation index, generate a status monitoring log for the updated data, and determine the subsequent monitoring direction.

8. The distributed new energy monitoring method based on machine vision according to claim 7, characterized in that, In step S630, the signal type and priority matching function of the warning signal is defined by the following formula: ； in, This represents the signal type and priority matching function for warning signals. Indicates the signal type of the input warning signal. Indicates the priority of the input warning signal. Indicates the first A pre-established maintenance process template, Weighting coefficients representing signal type Weighting coefficients representing priority This represents the similarity matching function. template Corresponding signal type, template Corresponding priority, This represents the collection of all preset maintenance process templates; The following formula is used to define the conditions for immediate execution of a warning signal: ； in, This indicates that the conditional function will be executed immediately. This indicates the priority value of the current warning signal. Indicates the current timestamp. Indicates the timestamp of signal detection. This indicates the priority threshold for immediate execution. This indicates the maximum allowed delay time. The immediate execution judgment function outputs 1 to indicate that immediate execution is required, and outputs 0 to indicate that immediate execution is not required. The process trigger instruction is derived using the following formula: ； in, This indicates the final generated process trigger instruction. Indicates the instruction generation function. This represents the feature vector of the acquired warning signal. This indicates the identifier of the matched maintenance process template. This indicates that the judgment result will be executed immediately. This indicates the feature fusion operator.

9. The distributed new energy monitoring method based on machine vision according to claim 8, characterized in that, In step S640, the updated global state evaluation metric is obtained using the following formula: ； in, This represents the updated global state assessment metric. This indicates the trigger information at the current moment. This represents the global state evaluation index at the current moment. This represents the weighting coefficient of the triggering information. The retention coefficient represents the historical state. This indicates the adjustment parameter for correlation updates. A function that represents the association between trigger information and the global state; The generated values ​​of the status monitoring logs are obtained using the following formula: ； in, Indicates the first The generated values ​​of each status monitoring log. Indicates the number of dimensions in the updated data. Indicates the first Importance weights of dimensional data Indicates the first Updated data in the first The values ​​of each monitoring point This represents a tiny constant that prevents logarithmic operations from producing a value of zero. The direction of subsequent monitoring is determined by the following formula: ； in, This indicates the confirmed direction for subsequent monitoring. An index indicating the optional monitoring direction. This represents the total number of monitoring and evaluation factors. Indicates the first The priority weight of each evaluation factor Indicates the first The monitoring direction is in the first The scores on each evaluation factor.

10. A distributed new energy monitoring system based on machine vision, comprising the distributed new energy monitoring method based on machine vision as described in any one of claims 1 to 9, characterized in that, include: The real-time image data acquisition module (10) is used to acquire real-time image data of photovoltaic modules and wind power generation modules through an image acquisition device installed on renewable energy facilities, and to process the real-time image data using an anomaly detection algorithm to obtain the equipment anomaly location result; The energy conversion efficiency assessment value determination module (20) is used to determine the energy conversion efficiency assessment value by analyzing the attitude change of the wind power generation component based on the abnormal positioning result of the equipment and the component attitude detection algorithm. The potential anomaly region marking module (30) is used to fuse the energy conversion efficiency assessment value with the light and shadow distribution analysis result of the photovoltaic module, and determine whether the deviation after fusion processing exceeds a preset threshold. If it exceeds the threshold, it is marked as a potential anomaly region. The cross-facility data preliminary aggregation set acquisition module (40) is used to extract operational data based on data communication from the facility edge computing unit for the potential abnormal area to obtain a cross-facility data preliminary aggregation set, wherein the operational data includes temperature parameters and vibration parameters; The global status assessment index determination module (50) is used to process the analysis results of the preliminary aggregation set of cross-facility data and the image acquisition device using a multi-source data fusion algorithm to determine the global status assessment index, wherein the global status assessment index is used to monitor the overall health status of renewable energy facilities. The maintenance process triggering module (60) is used to generate an early warning signal based on the global status assessment index. If the global status assessment index is lower than the safety threshold, the maintenance process is triggered.

Images