Offshore wind power intelligent operation and maintenance method, system and equipment based on 5G private network and medium

Abstract

The invention discloses an offshore wind power intelligent operation and maintenance method, system, equipment and medium based on a 5G private network, and belongs to the technical field of offshore wind power operation and maintenance, and the method comprises the steps: transmitting on-site real-time video data to an edge calculation platform through the 5G private network, carrying out the equipment part recognition and fault feature detection through the edge intelligent recognition processing, and carrying out the fault feature detection; the expert guidance annotation information and the spatial position of the equipment part are bound and then transmitted back to the augmented reality equipment to be displayed, multi-source data acquisition processing is adopted to adaptively select a data acquisition mode according to the interface characteristics of a service system to extract to-be-approved data, and after data conversion, the to-be-approved data are input into an approval rule engine to be judged and written back to the service system. According to the method, intelligent identification processing is locally completed on the edge platform to avoid data round-trip delay, the binding of the label and the physical position of the equipment is realized through space pose calculation, so that the label keeps accurate pointing along with the movement of the visual angle, and the cross-system data circulation without modifying the system is realized through a self-adaptive selection data acquisition mode.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power operation and maintenance technology, specifically to an intelligent operation and maintenance method, system, equipment, and medium for offshore wind power based on a 5G private network. Background Technology

[0002] Offshore wind power, as an important component of clean energy, has its operation and maintenance management directly impacting power generation efficiency and equipment lifespan. Current offshore wind power operation and maintenance methods primarily rely on on-site inspections and repairs by technical personnel, which suffers from high operational risks, long response times, and a lack of shared expert resources.

[0003] Traditional operation and maintenance methods typically use industrial WiFi or 4G public networks for communication. However, industrial WiFi has limited coverage, and signal attenuation is significant in environments with strong metal shielding, such as inside offshore wind turbine towers and nacelles, making it unsuitable for real-time transmission of high-definition video. While 4G public networks offer wider coverage, their bandwidth is limited and latency is high, making it difficult to support real-time audio and video communication and augmented reality applications during remote maintenance. Furthermore, public network transmission poses data security risks.

[0004] Regarding remote maintenance, existing technologies use video calls to allow backend experts to remotely guide on-site personnel. However, experts struggle to accurately determine the location of equipment components and fault points using 2D video feeds. The guidance annotations cannot precisely pinpoint the physical location of the equipment, and the annotations become invalid when on-site personnel move their viewpoint, requiring experts to repeatedly re-annotate, resulting in low communication efficiency. While some technologies incorporate augmented reality devices, video data and annotation processing are all performed in the cloud, leading to increased latency due to round-trip data transmission and impacting the real-time interactive experience.

[0005] In terms of operation and maintenance management, offshore wind farms involve multiple business systems, including equipment monitoring systems, asset management systems, procurement systems, and approval systems. These systems are independent of each other and have different technical architectures. Traditional methods require manual switching between multiple systems to complete the approval process. For example, approving the purchase of a spare part requires logging into multiple systems and manually entering data, which is time-consuming and prone to errors. Some systems are outdated and do not provide data interfaces, preventing automatic data transfer between systems. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention provides a method, system, equipment and medium for intelligent operation and maintenance of offshore wind power based on 5G private network.

[0007] Therefore, the technical problems solved by this invention are: how to achieve low-latency remote maintenance communication in the complex environment of offshore wind farms, and how to complete intelligent recognition and processing of video data locally through an edge computing platform to reduce data round-trip transmission delay; how to accurately bind remote guidance labeling information to the spatial position of equipment components so that the labeling can still maintain accurate pointing when the field personnel move their viewpoint; and how to adaptively select the data acquisition method according to the interface characteristics of different business systems without modifying the existing business system, so as to realize the automatic flow and approval of cross-system data.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent operation and maintenance of offshore wind power based on a 5G private network, comprising, Real-time video data of on-site operations is collected by 5G base stations deployed in wind turbine nacelles and substations, and the real-time video data is transmitted to an edge computing platform deployed in the substation equipment room via a 5G private network. In the edge computing platform, edge intelligent recognition processing is used to identify equipment components and detect fault features in the real-time video data, and the recognition and detection results are synchronously transmitted to the back-end expert terminal along with the real-time video data; The system receives guidance annotation information generated by experts in the background based on the real-time video data, binds the guidance annotation information with the spatial position of the device component, and sends the bound annotation information back to the on-site augmented reality device for overlay display. The system employs a multi-source data acquisition and processing approach, adaptively selecting the data acquisition method based on the interface characteristics of the business system to extract the data to be approved from the business system. The data to be approved is processed to obtain standardized data. The standardized data is then input into a preset approval rule engine for judgment to obtain the approval result. Based on the approval result, write-back data is generated and written into the business system.

[0009] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of using edge intelligent recognition processing to identify equipment components and detect fault features in the real-time video data includes: Video frames are extracted from the real-time video data, and image preprocessing is performed on the video frames; The preprocessed video frames are analyzed using a target detection algorithm to extract device image features; The similarity between the device image features and the pre-stored device component feature library is calculated, and the device component type is determined based on the similarity calculation results. Anomaly feature extraction is used to extract anomaly feature parameters from the real-time video data. The anomaly feature extraction method includes performing temperature distribution analysis on thermal imaging data to determine temperature anomaly areas and performing texture analysis on equipment surface images to identify cracks or deformations. The fault type is determined based on the abnormal feature parameters and the preset fault discrimination rules.

[0010] The beneficial effects of this preferred technical solution are as follows: by deploying a target detection algorithm on the edge computing platform to extract equipment image features and matching them with a pre-stored feature library to identify equipment components, and by using anomaly feature extraction to analyze thermal imaging temperature distribution and equipment surface texture to identify anomalies, the identification process is completed locally in the substation room to avoid the round-trip delay of video data transmission to the cloud and back. This enables experts to obtain equipment identification and fault detection results in real time to assist in diagnosis and decision-making, thereby improving the real-time performance and accuracy of remote maintenance.

[0011] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of calculating the similarity between the equipment image features and a pre-stored equipment component feature library includes: Extract the key point coordinates and descriptors from the features of the device image; Read the standard key point coordinates and standard descriptors corresponding to each equipment component from the equipment component feature library; Calculate the spatial distance between the coordinates of the key point and the coordinates of the standard key point, and calculate the vector distance between the descriptor and the standard descriptor; The similarity score is calculated based on the spatial distance and the vector distance, and the device component with the highest similarity score is selected as the recognition result.

[0012] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of binding the guidance and labeling information with the spatial location of the equipment components includes: Visual positioning processing is used to determine the spatial pose of the augmented reality device relative to the device components. The visual positioning processing includes acquiring environmental images and extracting environmental feature points through the augmented reality device, matching the environmental feature points with a pre-constructed three-dimensional point cloud map, and calculating the position coordinates and attitude angle of the augmented reality device based on the matching results. Based on the spatial pose and the pre-stored 3D model coordinates of the device components, calculate the anchor point coordinates of the guidance annotation information in 3D space; The anchor point coordinates are associated with the identification information of the equipment components and stored.

[0013] The beneficial effects of this preferred technical solution are as follows: by calculating the spatial pose of the augmented reality device relative to the device components through visual positioning processing, the guidance annotation information is converted into three-dimensional spatial coordinates and bound to the device components based on the spatial pose and the coordinates of the device's three-dimensional model. This overcomes the problem in traditional solutions where annotations based on two-dimensional screen coordinates become invalid after the personnel move and need to be repeatedly re-annotated. When the personnel move, the annotations automatically adjust their display position and angle according to the change in pose to maintain accurate pointing, thereby improving guidance efficiency and safety.

[0014] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of matching the environmental feature points with a pre-constructed three-dimensional point cloud map includes: Extract a local point cloud subset from the 3D point cloud map; Calculate the descriptor similarity between the environmental feature points and each point in the local point cloud subset; Point pairs with descriptor similarity greater than a preset threshold are selected as candidate matching point pairs; The candidate matching point pairs are subjected to geometric consistency checks. Matching point pairs that do not meet the geometric constraints are eliminated, and the matching point pairs that pass the check are retained for calculating the spatial pose of the augmented reality device.

[0015] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of extracting the data to be approved from the business system by adopting multi-source data acquisition and processing and adaptively selecting the data acquisition method according to the interface characteristics of the business system includes: Send an interface probe request to the business system; Parse the interface description information returned by the business system to determine whether the business system provides an application programming interface (API). When it is determined that the business system provides an application programming interface (API), the API is invoked and query parameters are passed in to obtain the data to be approved. When it is determined that the business system does not provide an application programming interface, the system checks whether the business system has database access permissions. If database access permissions are enabled, the system executes a query statement through a database connection to obtain the data to be approved. If database access permissions are not enabled, the system logs into the business system and retrieves the data to be approved by simulating a user interface.

[0016] As a preferred embodiment of the intelligent operation and maintenance method for offshore wind power based on a 5G private network described in this invention, the step of logging into the business system and retrieving the data to be approved via a simulated user interface includes: Obtain page element information for the login page of the business system; Locate the username and password input boxes based on the page element information, and input the pre-stored login credentials into the username and password input boxes; The login button element is triggered to complete the login process and enter the data display page of the business system; The page structure of the data display page is analyzed, and data records containing the pending approval identifier are extracted as the pending approval data.

[0017] This invention provides an intelligent operation and maintenance system for offshore wind power based on a 5G private network.

[0018] To solve the above technical problems, the present invention provides the following technical solution: an intelligent operation and maintenance system for offshore wind power based on a 5G private network, comprising: a data acquisition module, used to acquire real-time video data of on-site operations through 5G base stations deployed in the wind turbine nacelle and the booster station, and transmit the real-time video data to an edge computing platform deployed in the booster station equipment room via the 5G private network; The edge recognition module is used to perform device component identification and fault feature detection on the real-time video data by using edge intelligent recognition processing in the edge computing platform, and to synchronously transmit the recognition and detection results with the real-time video data to the back-end expert terminal. The annotation processing module is used to receive guidance annotation information generated by experts in the background for the real-time video data, bind the guidance annotation information with the spatial position of the device component, and send the bound annotation information back to the on-site augmented reality device for overlay display; The data acquisition module is used to extract the data to be approved from the business system by adaptively selecting the data acquisition method according to the interface characteristics of the business system through multi-source data acquisition and processing. The approval processing module is used to perform data transformation processing on the data to be approved to obtain standardized data, input the standardized data into a preset approval rule engine for judgment to obtain the approval result, generate write-back data based on the approval result and write it into the business system.

[0019] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the intelligent operation and maintenance method for offshore wind power based on a 5G private network.

[0020] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the intelligent operation and maintenance method for offshore wind power based on a 5G private network.

[0021] The beneficial effects of this invention are as follows: By employing edge intelligent recognition processing on an edge computing platform to identify equipment components and detect fault features in real-time video data, the device image features in the video frames are extracted and matched with a pre-stored feature library. Abnormal feature parameters are extracted to determine the fault type. The recognition processing is deployed locally in the substation equipment room, avoiding the round-trip delay of video data transmission to the cloud and back. This allows backend experts to obtain equipment recognition and fault detection results in real time to assist in diagnostic decision-making. By using visual positioning processing to determine the spatial pose of the augmented reality device relative to the equipment components, the expert-generated guidance annotation information is converted into three-dimensional spatial coordinates and bound to the physical position of the equipment components. When on-site personnel move their viewpoint, the annotation information automatically adjusts its display position and angle according to the pose change to maintain accurate pointing. This overcomes the problem in traditional solutions where annotations based on two-dimensional screen coordinates become invalid after personnel movement, requiring repeated re-annotation. By adopting a multi-source data acquisition and processing approach, the data acquisition method is adaptively selected based on the interface characteristics of the business system. When the business system provides an application programming interface (API), the API is called to acquire data. When no API is provided but database access is enabled, data is acquired through database query. When there is neither an API nor database access permission, data is acquired by simulating user interface operations. This achieves cross-system data extraction without modifying the existing business system. After the data to be approved is processed and transformed, it is input into the approval rule engine for automatic judgment and write-back of the results, avoiding manual switching between multiple systems. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an intelligent operation and maintenance method for offshore wind power based on a 5G private network, as provided in one embodiment of the present invention. Detailed Implementation

[0024] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0025] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for intelligent operation and maintenance of offshore wind power based on a 5G private network, including: Step 1: Collect real-time video data of the on-site operation by deploying 5G base stations in the wind turbine nacelle and substation, and transmit the real-time video data to the edge computing platform deployed in the substation equipment room via the 5G private network; Step 2: In the edge computing platform, edge intelligent recognition processing is used to identify equipment components and detect fault features in the real-time video data, and the identification and detection results are synchronously transmitted to the back-end expert terminal along with the real-time video data; Step 3: Receive guidance annotation information generated by experts in the background based on the real-time video data, bind the guidance annotation information with the spatial position of the device component, and send the bound annotation information back to the on-site augmented reality device for overlay display; Step 4: Employ multi-source data acquisition and processing, adaptively select data acquisition methods based on the interface characteristics of the business system, and extract the data to be approved from the business system; Step 5: Perform data transformation processing on the data to be approved to obtain standardized data, input the standardized data into the preset approval rule engine for judgment to obtain the approval result, generate write-back data based on the approval result and write it into the business system.

[0026] Offshore wind power operation and maintenance faces challenges such as a complex communication environment and low efficiency in remote maintenance. The internal structure of the wind turbine tower is a vertical metal cylinder exceeding 100 meters in height. Traditional industrial WiFi signals suffer severe attenuation due to multiple reflections and absorptions from the metal walls, resulting in insufficient coverage. The nacelle houses large metal equipment such as gearboxes, generators, and pitch control systems, creating a severe electromagnetic shielding environment that makes it difficult for wireless signals to penetrate. During offshore operations, maintenance personnel use augmented reality devices to collect on-site video and transmit it to experts at the back end. However, the long data transmission route—from the cloud for processing before returning to the site—leads to accumulated latency. When experts provide remote guidance via video, annotations are based on two-dimensional screen coordinates. When on-site personnel need to adjust their position for operation, the annotations cannot follow the equipment's movement, leading to incorrect pointing. Experts then need to redraw the annotations based on the new perspective.

[0027] This embodiment constructs a private network coverage by deploying 5G base stations in the wind turbine nacelle and tower. The 5G base stations use the 2.1GHz low-frequency band to provide stronger signal coverage, solving communication problems in the metal-shielded environment inside the tower and nacelle, and achieving stable transmission of on-site video data. An edge computing platform is deployed in the substation equipment room. After video data is transmitted to the edge platform via the 5G private network, device identification and fault detection are performed locally, avoiding round-trip delays in data transmission to the remote cloud. Identification results are synchronized to expert terminals in real time for auxiliary diagnosis. Visual positioning technology is used to calculate the spatial pose of augmented reality devices, converting expert annotations into three-dimensional spatial coordinates and binding them to device components. When on-site personnel move, the system automatically adjusts the annotation display position to maintain accurate pointing, avoiding repeated re-annotation. Multi-source data acquisition and processing adaptively selects the acquisition method based on the characteristics of business system interfaces. Data is obtained through API-provided system call interfaces, and for older systems, data is obtained through database queries or simulated interface operations. This achieves cross-system data flow without modifying existing systems. After data conversion, the data is input into the approval rule engine for automatic judgment and result writing back, avoiding manual multi-system switching operations.

[0028] Example 2, an embodiment of the present invention, provides a method for intelligent operation and maintenance of offshore wind power based on a 5G private network, based on the previous embodiment, including: Step 1: Real-time video data of on-site operations is collected by 5G base stations deployed in the wind turbine nacelle and substation, and the real-time video data is transmitted via the 5G private network to the edge computing platform deployed in the substation equipment room, including the following steps A1-A4: A1: Deploy indoor micro base stations inside the wind turbine nacelle, deploy leaky cables or pico base stations inside the wind turbine tower, and deploy an indoor distribution system at the substation. The indoor micro base stations and leaky cables use the 2.1GHz frequency band. A2: On-site maintenance personnel wear augmented reality devices, which integrate cameras, inertial measurement units, and 5G communication modules. The cameras collect real-time video data of on-site operations, and the inertial measurement units collect equipment attitude data. A3: The 5G communication module establishes a wireless connection with the 5G base station, performs video encoding and compression on the real-time video data, and transmits the encoded video data through the 5G uplink. A4: The 5G base station transmits the encoded video data to the edge computing platform deployed in the substation equipment room via the optical fiber backhaul link. The edge computing platform decodes and restores the video data to obtain the original video frame sequence.

[0029] It should be noted that the indoor micro base station uses a low-power transmission mode to cover the interior space of a single cabin. The interior of the cabin is a closed metal space, and the metal equipment and walls reflect and absorb wireless signals. Using the 2.1GHz low-frequency band can provide better diffraction and penetration capabilities, and compared with the 3.5GHz high-frequency band, the signal attenuation is smaller in the metal shielding environment, resulting in better coverage.

[0030] Furthermore, the leaky cable is laid vertically along the inner wall of the tower, covering the entire vertical space from the bottom of the tower to the nacelle. The tower is a cylindrical metal structure, and traditional antenna methods suffer from severe signal attenuation in the vertical direction and have coverage blind spots. The leaky cable radiates electromagnetic waves into the surrounding space through periodic slots on its surface, achieving continuous coverage within the tower. This ensures communication continuity for maintenance personnel moving up and down within the tower, avoiding video transmission interruptions caused by frequent switching.

[0031] It should be noted that the video encoding uses a variable bitrate encoding method, which adaptively adjusts the encoding bitrate according to the complexity of the video content. A lower bitrate is used for static scenes or scenes with small movements to reduce bandwidth consumption, while a higher bitrate is used for complex scenes or scenes with fast movement to ensure video clarity, thus optimizing the efficiency of transmission bandwidth utilization while ensuring video quality.

[0032] Furthermore, the fiber optic backhaul link uses industrial-grade fiber, which is resistant to salt spray corrosion and vibration, and adapts to harsh marine environments. The fiber optic backhaul employs a dual-route backup mechanism, with the primary fiber path and backup fiber path physically isolated. When a decrease in optical power on the primary path or a bit error rate exceeding a threshold is detected, the system automatically switches to the backup path, ensuring the reliability and continuity of video data transmission.

[0033] Step 2: In the edge computing platform, edge intelligent recognition processing is used to identify device components and detect fault features in the real-time video data. The identification and detection results are then synchronously transmitted to the backend expert terminal along with the real-time video data, including the following steps B1-B5: B1: Extract video frames from the real-time video data and perform image preprocessing on the video frames; B2: Analyze the preprocessed video frames using a target detection algorithm to extract device image features; B3: Calculate the similarity between the device image features and the pre-stored device component feature library, and determine the device component type based on the similarity calculation results; B4: Anomaly feature parameters are extracted from the real-time video data using anomaly feature extraction method. The anomaly feature extraction method includes performing temperature distribution analysis on thermal imaging data to determine temperature anomaly areas and performing texture analysis on equipment surface images to identify cracks or deformations. B5: Determine the fault type based on the abnormal feature parameters and the preset fault discrimination rules.

[0034] In this embodiment, step 2, edge intelligent recognition processing involves: deploying a target detection model based on a deep convolutional neural network on an edge computing platform. This target detection model is pre-trained using labeled images of wind turbine components. Video frames are input into the target detection model. The model's convolutional layers extract image features at different scales from the video frames. A feature pyramid network fuses multi-scale features, and a detection head network classifies and regresses bounding boxes on the fused features, outputting the category label and location coordinates of the equipment component. For thermal imaging data, the temperature value of each pixel in the temperature matrix is ​​extracted. The design operating temperature range of the equipment component is read from a pre-stored database of normal operating temperature parameters provided by the equipment manufacturer. Temperatures exceeding the upper limit or falling below the lower limit of the design operating temperature are marked as abnormal temperature areas. For the surface image of the device, a Gabor filter bank is used to extract texture features in multiple directions and at multiple scales. The Gabor filter bank contains 40 filters in 8 directions and 5 scales. The mean and standard deviation of the filter response are calculated as texture features. The Euclidean distance between the current texture features and the pre-stored normal surface texture features is calculated. When the Euclidean distance is greater than 3 times the standard deviation of the normal texture features, it is determined that there is a crack or deformation.

[0035] In an optional implementation, step 2, edge intelligent recognition processing can be achieved by deploying a two-stage object detection algorithm based on a region proposal network on an edge computing platform. The first stage generates candidate region proposals, and the second stage performs fine-grained classification and position correction on the candidate regions. Low-level feature maps of video frames are extracted, and the region proposal network generates multiple anchor boxes of different sizes and aspect ratios on the feature maps. High-quality candidate regions are selected through foreground / background classification and bounding box regression. The features corresponding to the candidate regions are fed into a classification network, which outputs the category probability distribution of the device components and the precise bounding box coordinates. For fault detection, a template matching-based method is used. Template images of the device surface under normal conditions are pre-stored. The currently acquired device surface image and the template image are then subjected to normalized cross-correlation calculation. The normalized cross-correlation calculation formula is the sum of the products of corresponding pixels in the two images divided by the square root of the sum of the squares of their respective pixels. The correlation coefficient ranges from 0 to 1. When the correlation coefficient is lower than 0.85, a surface anomaly is determined. The threshold of 0.85 is determined through statistical analysis of 100 groups of normal and abnormal samples. The mean correlation coefficient of the normal samples is 0.92 and the standard deviation is 0.02. The threshold is set as the mean minus 3 times the standard deviation.

[0036] In another optional implementation, step 2, edge intelligent recognition processing can also be achieved by: deploying a Transformer-based visual recognition model on an edge computing platform, dividing video frames into 16×16 pixel image blocks, converting each image block into a feature vector through a linear projection layer, adding positional encoding, and then inputting it into a multi-head self-attention layer for global feature association calculation. The Transformer encoder contains 12 layers, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer. The output feature vector is processed by a classification head to obtain the device component category, and by a detection head to obtain the bounding box coordinates. For fault detection, motion analysis based on optical flow is adopted. The Farneback dense optical flow algorithm is used to calculate the optical flow field between adjacent video frames. The algorithm approximates the local neighborhood of the image through polynomial expansion, calculates the displacement field of the image block, extracts the optical flow vector of each pixel, and the optical flow vector contains displacement components in the horizontal and vertical directions. The optical flow amplitude is calculated as the square root of the sum of the squares of the two components. The optical flow amplitude of all pixels in the device component area is statistically analyzed, and the average value and standard deviation of the regional optical flow amplitude are calculated. When the average value exceeds the historical average value plus twice the historical standard deviation, it is determined that there is a vibration anomaly. The historical statistical values ​​are calculated based on the normal operation data of the most recent 30 days.

[0037] In this embodiment, step B4 involves extracting abnormal features by: constructing a temperature distribution histogram from the thermal imaging data, dividing the temperature range into 50 equally spaced temperature intervals, counting the number of pixels in each temperature interval, and calculating the mean, standard deviation, and skewness of the temperature distribution. Normal temperature distribution parameters for the corresponding equipment component are read from a pre-stored normal operating condition temperature distribution parameter library. These normal temperature distribution parameters are obtained by collecting thermal imaging data of the equipment component under normal operating conditions for seven consecutive days. The Kullback-Leibler divergence between the current temperature distribution and the normal temperature distribution is calculated. The divergence calculation formula is the logarithm of the ratio of the probability of each temperature interval in the current distribution to the probability of the corresponding interval in the normal distribution, multiplied by the current distribution probability, and then summed. When the divergence exceeds 0.3, a temperature anomaly is determined. The threshold of 0.3 is determined through temperature distribution analysis of historical fault cases, with the divergence values ​​before the fault occurring concentrated in the range of 0.3 to 0.8. Morphological processing is performed on the temperature matrix, and a connected component labeling algorithm is used to extract continuous high-temperature or low-temperature regions. The connected component labeling uses 8-neighborhood connectivity to determine whether adjacent pixels belong to the same connected component. The number of pixels in a connected component is calculated as its area. When the area is greater than 100 pixels, it is considered a valid anomalous region. The threshold is determined based on the resolution of the thermal imaging camera and the actual size of the device component; 100 pixels correspond to an actual area of ​​approximately 10 square centimeters. For the device surface image, after converting the image to grayscale, Canny edge detection is performed. The Canny algorithm first smooths the image using a Gaussian filter, then calculates the magnitude and direction of the image gradient, uses non-maximum suppression to refine edges, and uses a dual-threshold method to detect strong and weak edges. The high threshold is set to the 90th percentile of the gradient magnitude distribution, and the low threshold is set to 0.4 times the high threshold. Hough line transform is performed on the edge contours, and lines with more than 50 votes in the cumulative array are identified as valid lines. When a line deviating more than 15 degrees from the main structural direction of the device component is detected, a crack is determined to exist.

[0038] In an optional implementation, in step B4, the abnormal feature extraction method can be achieved by: employing a time-series analysis method using infrared thermal imaging, continuously acquiring 30 frames of thermal imaging data to form a time series, with a frame rate of 10 frames per second; performing linear regression fitting on the temperature time series at each pixel location, and calculating the slope of the regression line as a parameter for temperature change trend. When the absolute value of the slope is greater than 0.5 degrees per second, a continuous temperature rise or fall trend is determined. This threshold is determined based on the heat capacity and heat dissipation capability of the equipment components; during normal operation, the temperature fluctuation slope is less than 0.2 degrees per second. A fast Fourier transform is performed on the temperature time series to extract frequency domain features, calculate the power spectral density, identify the main frequency components of temperature changes, and read the normal frequency range of the equipment from a pre-stored normal operation frequency feature library. When the detected main frequency deviates from the normal frequency range by more than 10%, the equipment's operating status is determined to be abnormal. For images of the device surface, a semantic segmentation model based on the UNet architecture is used to classify the images at the pixel level. The model consists of an encoder and a decoder. The encoder extracts multi-level features through convolution and pooling operations, and the decoder restores spatial resolution through upsampling and convolution operations. It outputs the probability of each pixel belonging to categories such as normal surface, crack, corrosion, and oil stains. The category with the highest probability is selected as the classification result of the pixel. The number of pixels in the crack category is counted. When the number of pixels in the crack category exceeds 2% of the total number of pixels in the image, it is determined that there is a significant surface defect.

[0039] In another optional implementation, in step B4, the abnormal feature extraction method can also be achieved by: employing multispectral imaging technology to simultaneously acquire device images in the visible light, near-infrared, and thermal infrared bands, wherein the visible light band is 400-700 nm, the near-infrared band is 700-1400 nm, and the thermal infrared band is 8-14 μm. The images of the three bands are registered and aligned using a feature point matching-based image registration method. SIFT feature points of each band image are extracted, feature point descriptors are calculated, and the correspondence between bands is established through descriptor matching. The affine transformation matrix is ​​then solved to align the images of different bands to a unified coordinate system. Band fusion is performed on the registered multispectral images using principal component analysis. The covariance matrix of the multispectral images is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​are sorted from largest to smallest, and the first three principal components are selected. The cumulative variance contribution rate of the first three principal components exceeds 95%. The original multispectral image is projected onto the principal component space to obtain a dimension-reduced fused image. Threshold segmentation is performed on the first principal component image. The Otsu automatic threshold selection method is used to calculate the threshold that maximizes the inter-class variance. The image is segmented into foreground and background. The foreground region is marked as a suspected abnormal region. Combined with the 3D model information of the device, the 2D image coordinates are converted into 3D spatial coordinates through camera calibration parameters. The actual position and area of ​​the abnormal region on the surface of the device component are calculated.

[0040] Furthermore, the step of calculating the similarity between the device image features and a pre-stored device component feature library includes: Extract the key point coordinates and descriptors from the device image features; Read the standard key point coordinates and standard descriptors corresponding to each equipment component from the equipment component feature library; Calculate the spatial distance between the keypoint coordinates and the standard keypoint coordinates, and calculate the vector distance between the descriptor and the standard descriptor; The similarity score is calculated based on the spatial distance and the vector distance, and the device component with the highest similarity score is selected as the recognition result.

[0041] It should be noted that the construction of the device component feature library was completed through offline annotation and feature extraction processes. Standard images of various device components were collected under different angles and lighting conditions, with no fewer than 50 images collected for each device component. The standard images were manually annotated, marking the device component category and bounding box. The same feature extraction algorithm as used for online recognition was used to extract the keypoint coordinates and descriptors from the standard images. The extracted features were then associated with and stored according to the device component category, forming the device component feature library. The feature library uses a KD-tree data structure for indexing, supporting fast nearest neighbor search.

[0042] Step 3: Receive guidance annotation information generated by experts in the background based on the real-time video data, bind the guidance annotation information to the spatial position of the device component, and send the bound annotation information back to the on-site augmented reality device for overlay display, including the following steps C1-C3: C1: Visual positioning processing is used to determine the spatial pose of the augmented reality device relative to the device components. The visual positioning processing includes acquiring environmental images and extracting environmental feature points through the augmented reality device, matching the environmental feature points with a pre-constructed three-dimensional point cloud map, and calculating the position coordinates and attitude angle of the augmented reality device based on the matching results. C2: Calculate the anchor point coordinates of the guidance annotation information in three-dimensional space based on the spatial pose and the pre-stored three-dimensional model coordinates of the device component; C3: Establish and store the association between the anchor point coordinates and the identification information of the equipment component.

[0043] In this embodiment, step C1 involves the following steps: The augmented reality device's camera captures environmental images within its current field of view. Feature point detection is performed on the images using the ORB feature detection algorithm. The ORB algorithm first uses a FAST corner detector to detect corners at multiple scales of the image pyramid. Harris corner response values ​​are calculated and sorted for the detected corners. The top 500 corners with the highest response values ​​are selected as feature points. A 256-bit binary descriptor is extracted for each feature point. This descriptor is generated by comparing the grayscale values ​​of randomly selected pixel pairs within the feature point's neighborhood. Point cloud data of the cabin environment is read from a pre-constructed 3D point cloud map. This point cloud map is constructed using an offline SLAM mapping process, and each point cloud data point contains XYZ 3D coordinates and a corresponding 256-bit ORB descriptor. The environmental feature point descriptors of the current image are matched with the point cloud descriptors in the point cloud map. The Hamming distance between the two descriptors is calculated. The Hamming distance is the number of different bits in two binary strings. The point cloud point with the smallest Hamming distance is selected as a candidate match. The nearest neighbor distance ratio method is used to screen reliable matches. The ratio of the smallest Hamming distance to the second smallest Hamming distance is calculated. When the ratio is less than 0.7, the match is considered reliable. The threshold of 0.7 is an empirical threshold for ORB feature matching. Based on the screened matching point pairs, which contain two-dimensional image coordinates and three-dimensional point cloud coordinates, the EPnP algorithm is used to solve the camera extrinsic parameter matrix of the augmented reality device. The EPnP algorithm represents the three-dimensional point as a linear combination of four control points. The camera pose is obtained by solving the coordinates of the control points in the camera coordinate system. The position coordinates and attitude angles of the camera are decomposed from the extrinsic parameter matrix. The attitude angles are represented by Euler angles, including the pitch angle around the X-axis, the yaw angle around the Y-axis, and the roll angle around the Z-axis.

[0044] In an optional implementation, in step C1, the visual positioning process can be performed by: using a visual-inertial odometry fusion method to combine camera images and inertial measurement unit (IMU) data from the augmented reality device for pose estimation. The IMU outputs triaxial acceleration and triaxial angular velocity data at a frequency of 100Hz. The angular velocity is integrated over time to obtain the attitude change, and the acceleration is integrated twice over time to obtain the displacement change. The integration uses the trapezoidal integral method, and the integration result is used as the predicted pose value. The camera acquires images at a frequency of 30Hz, and feature point tracking is performed on two adjacent frames using the KLT optical flow tracking algorithm. This algorithm assumes that the motion of feature points within a small range can be approximated by affine transformation. The feature point displacement is solved by minimizing the grayscale difference between image blocks before and after tracking. Based on the tracked feature point pixel displacements, the camera's rotation matrix and translation vector are calculated using epipolar geometry constraints. The predicted pose of the inertial measurement unit is used as the state prediction, and the observed pose of the visual tracking is used as the observation update. They are input into an extended Kalman filter for fusion. The state vector of the filter contains 16 dimensions, including position, velocity, attitude, and sensor bias. The prediction step updates the state estimate and covariance matrix based on the motion model and inertial measurement data. The update step corrects the state estimate based on the observation model and visual measurement data. The filter outputs the fused position coordinates and attitude angles.

[0045] In another optional implementation, in step C1, the visual positioning process can also be achieved by: pre-deploying visual markers in the cabin environment. These markers are AprilTag QR code labels, each 15 cm x 15 cm in size. Each label contains a unique identifier encoded with a black and white pattern. The labels are affixed to fixed positions on the surfaces of equipment components such as gearboxes and generators. The three-dimensional coordinates of the label's center point are measured using a total station and stored in a label database with millimeter-level accuracy. After the augmented reality device's camera captures an image, the AprilTag detection algorithm is used to identify the label in the image. This algorithm first performs adaptive threshold binarization on the image, extracts the contour formed by the black and white boundaries, performs polygon fitting on the contour, filters out quadrilateral contours, decodes the pattern inside the quadrilateral to obtain the label identifier, and extracts the pixel coordinates of the label's four corner points. Based on the actual size of the label (15 cm) and the pixel coordinates of the four corner points, the PnP algorithm is used to solve for the relative pose between the camera and the label. The PnP algorithm establishes the projection relationship from three-dimensional spatial points to two-dimensional image points and solves for the camera's rotation matrix and translation vector by minimizing the reprojection error. When multiple labels are detected simultaneously in an image, the camera pose is calculated independently for each label. The multiple pose results are then weighted and averaged according to the distance between the label and the camera; closer labels have higher weights. The weighting formula is: weight equals the reciprocal of the distance divided by the sum of the reciprocals of all distances. Combined with the pre-stored 3D coordinates of the labels, the camera pose relative to the labels is transformed to the global coordinate system, yielding the position coordinates and attitude angle of the augmented reality device in the global coordinate system of the cabin environment.

[0046] Furthermore, the 3D point cloud map is constructed through an offline SLAM mapping process. Handheld LiDAR or depth cameras are used to collect data while moving within the cabin environment. The LiDAR outputs 3D point cloud data, and the depth camera outputs RGB-D images. An ICP point cloud registration algorithm is used to align point clouds collected at different locations. This algorithm achieves registration by iteratively minimizing the nearest point distance between point clouds. The registered point clouds are then filtered to remove noise points. A voxel grid filtering method is used for downsampling, with the voxel size set to 5 cm. An ORB descriptor is extracted for each point cloud point. This descriptor extraction is achieved by projecting the point cloud onto a virtual camera plane to generate a depth image, and then performing ORB feature detection on the depth image. The XYZ coordinates of the point cloud and the ORB descriptor are associated and stored to form the 3D point cloud map.

[0047] Furthermore, matching the environmental feature points with the pre-constructed 3D point cloud map includes: Extract a local point cloud subset from the 3D point cloud map; Calculate the descriptor similarity between the environmental feature points and each point in the local point cloud subset; Point pairs with descriptor similarity greater than a preset threshold are selected as candidate matching point pairs; The candidate matching point pairs are subjected to geometric consistency checks. Matching point pairs that do not meet the geometric constraints are eliminated, and the matching point pairs that pass the check are retained for calculating the spatial pose of the augmented reality device.

[0048] Step 4: Employing multi-source data acquisition processing, adaptively selecting the data acquisition method based on the interface characteristics of the business system to extract the data to be approved from the business system; this includes the following steps D1-D4: D1: Send an interface probe request to the business system; D2: Parse the interface description information returned by the business system to determine whether the business system provides an application programming interface (API). D3: When it is determined that the business system provides an application programming interface, call the application programming interface and pass in the query parameters to obtain the data to be approved; D4: When it is determined that the business system does not provide an application programming interface, check whether the business system has database access permissions. If database access permissions are enabled, execute a query statement through a database connection to obtain the data to be approved. If database access permissions are not enabled, log in to the business system through a simulated user interface operation and extract the data to be approved.

[0049] In this embodiment, step 4, the multi-source data acquisition process involves: sending an HTTP GET request to the business system. The URL path of the request is the root path of the business system plus common interface description document paths such as / api / docs or / swagger.json. The Accept field in the request header is set to application / json. The HTTP response returned by the business system is parsed. When the response status code is 200 and the response body is in JSON format, the JSON data is parsed to extract the interface endpoint address, request method, parameter list, and returned data format, determining that the business system provides a RESTful API interface. An HTTP request conforming to the interface specification is constructed. Based on the interface documentation, the request method is determined to be GET or POST. An Authorization field is added to the request header, and a Bearer token is filled in for authentication. Query parameters are filled in the request body in JSON format, including filtering conditions such as time range and data type. The interface is called to obtain the data to be approved, and the data records are extracted from the JSON array in the response body. When the business system returns an HTTP 403 Forbidden or 404 Not Found error code, it is determined that the API interface is not provided, and a database connection method is attempted instead. Construct a database connection string in the format `jdbc:databasetype: / / server address:port number / database name`, where the database type includes MySQL, PostgreSQL, or Oracle. Read the database server address, port number, and database name from the configuration file. Establish a database connection using the JDBC driver, with connection parameters including username and password. After successful connection, construct an SQL query statement in the format `SELECT field list FROM table_name WHERE condition expression`. Execute the query and retrieve the result set, iterating through each row of the result set to extract field values. When a connection fails and throws an SQLException, an RPA approach is used. Selenium WebDriver is started to control the Chrome browser. The `get` method is called to open the business system login page URL. The `find_element_by_id` method is used to locate the DOM element for the username input box. This DOM element is obtained by parsing the `input` tag with the `id` attribute of `username` in the page's HTML code. The `send_keys` method is called to input the pre-stored username into the input box. The same method is used to locate the password input box and input the password. The login button element is located, and the `click` method is called to trigger a click event to complete the login. After the page redirects, the `find_elements_by_xpath` method is used to locate the data table elements. The XPath expression is ` / / table[@class=data-table] / tbody / tr`. The table row elements are traversed to extract the cell text content as the data to be approved.

[0050] In one optional implementation, multi-source data acquisition processing can be achieved by: employing an adaptive interface probing strategy and maintaining an interface type cache table for a business system. This cache table is stored in a Redis in-memory database, and its structure includes fields for system identifier, interface type, probe time, and expiration date. When data needs to be retrieved from a business system, the Redis cache is first queried based on the system identifier. If the cache hits and the current time has not expired, the corresponding data retrieval method is directly called using the interface type recorded in the cache, skipping the interface probing step. If the cache misses or expires, the interface probing process is executed. Two parallel threads are created: the first thread sends an API interface probing request, and the second thread attempts to establish a database connection. A timeout of 5 seconds is set, and a CountDownLatch synchronization mechanism is used to wait for both threads to complete. The interface type to be used is determined by which thread successfully returns the result first. After a successful probe, the Redis SET command is called to update the cache table, setting the system identifier as the key, the interface type as the value, and the expiration date to 24 hours. The EXPIRE command is then used to set the key's expiration time. When accessing the business system in the future, the interface type is read directly from the cache to avoid repeated probing. When the interface type changes or the permissions change, causing the access to fail, the corresponding cache key is cleared to trigger a re-probing.

[0051] In another alternative implementation, multi-source data acquisition processing can also be achieved by: adopting a pluggable data acquisition adapter architecture, defining a unified adapter interface, which includes three abstract methods: connect, query, and close. Specific adapter classes are implemented for different types of business systems. For example, the RestApiAdapter class implements API interface call logic; its connect method establishes an HTTP connection pool, its query method constructs and sends an HTTP request, and its close method releases connection resources. The DatabaseAdapter class implements database access logic; its connect method establishes a JDBC connection, its query method executes SQL statements, and its close method closes the database connection. The RpaAdapter class implements user interface operation logic; its connect method launches a browser, its query method executes an automation script, and its close method closes the browser. Upon system startup, Java reflection is used to scan all adapter classes under a specified package path, the Class.forName method is called to load the class, the newInstance method is called to create an adapter instance, and the adapter instance is registered in the adapter factory's HashMap, with the business system identifier as the key and the adapter instance as the value. When data needs to be retrieved from a business system, the `getAdapter` method of the adapter factory is called. The corresponding adapter instance is then retrieved from the HashMap based on the system identifier. The adapter's `connect` method is called to establish a connection, the `query` method is called with the query parameters to retrieve the data, and the `close` method is called to release resources. When a new business system is added, a new adapter class that inherits from the adapter interface is developed, implementing the `connect`, `query`, and `close` methods. This new adapter class is placed in the scanning path, and it is automatically loaded and registered after the system restarts. No modification to the main program code is required, thus achieving a scalable multi-source data retrieval architecture.

[0052] Furthermore, the step of logging into the business system and retrieving the data to be approved via a simulated user interface includes: Obtain page element information for the login page of the business system; Locate the username and password input boxes based on the page element information, and input the pre-stored login credentials into the username and password input boxes; The login button element is triggered to complete the login process and enter the data display page of the business system; The page structure of the data display page is analyzed, and data records containing the pending approval identifier are extracted as the pending approval data.

[0053] Step 5: The data to be approved is processed to obtain standardized data. The standardized data is then input into a preset approval rule engine for judgment to obtain the approval result. Based on the approval result, write-back data is generated and written into the business system, including the following steps E1-E5: E1: Parse the data structure of the data to be approved, identify the field types and field values ​​in the data, and establish a mapping table between the source system fields and the target standard fields; E2: Based on the mapping relationship table, perform field mapping transformation on the data to be approved, convert the field names of the source system into standard field names, unify the data format of the field values, standardize the date format fields, and unify the precision of the numeric format fields to obtain standardized data; E3: Extract the key approval parameters from the standardized data. The key approval parameters include data priority identifier, data type identifier, data source identifier, and associated process identifier. Pass the key approval parameters as input to the preset approval rule engine. E4: The approval rule engine performs logical judgments based on preset approval rules, determines the approval routing strategy based on the data priority identifier, matches the corresponding approval process template based on the data type identifier, determines whether to trigger cascading approval based on the associated process identifier, and generates approval result data. E5: Generate write-back data formats corresponding to each business system based on the approval result data, and write the write-back data to each business system through the corresponding data writing method in the multi-source data acquisition and processing, and update the approval status field and approval time field in the business system.

[0054] It should be noted that the mapping table is pre-configured and stored in the database, recording the correspondence between field names of each business system and standard field names. Different business systems use different field names for the same business concept; the mapping table achieves semantic alignment of fields, eliminating naming differences between heterogeneous systems. The mapping table supports dynamic updates; when a new business system is added or field definitions change, the mapping relationship can be updated through the configuration interface without modifying the program code.

[0055] Furthermore, the date format standardization conversion process addresses the different date representation methods used by various business systems. Some systems use a hyphen as the year-month-day separator, some use a slash separator, and some use a timestamp format. The process unifies the conversion to a standard format that includes date, time, and time zone information, eliminating data parsing errors and time calculation discrepancies caused by inconsistent date formats.

[0056] It should be noted that the data priority identifier is divided into three levels: high priority, medium priority, and low priority. Priority determination is based on the timeliness requirements and business impact of the data. High-priority data triggers an expedited approval process, with the approval rule engine sending an immediate notification to the approver via a push notification mechanism. Medium-priority data is processed according to the standard process, and low-priority data is processed periodically in a batch processing queue. The priority identifier is automatically determined by parsing the urgency or fault impact level fields in the data to be approved.

[0057] Furthermore, the data type identifier is used to distinguish approval data for different business types, including equipment maintenance, material procurement, personnel scheduling, and security management. Different data types correspond to different approval process templates and approval permission configurations. The approval rule engine matches the corresponding process template based on the data type identifier and determines parameters such as the number of approval nodes, approver roles, and approval timeout.

[0058] It should be noted that the associated process identifier is used to determine whether the current approval data is associated with other business processes. When an associated process identifier is detected, the approval rule engine queries the execution status of the associated process to determine whether the associated process has completed the preceding approval node. If the preceding node has not been completed, the current approval is suspended and waited for; if the preceding node has been completed, the current approval continues to be executed, thus realizing cross-process approval dependency management and cascading triggering.

[0059] Furthermore, the data format conversion process for writing back is the reverse of the data acquisition process. For systems providing API interfaces, standardized approval result data is converted into the data format required by the API interface, and the update method of the interface is called to write the data. For systems supporting database access, approval results are converted into database update statements to directly modify the data table. For systems that only support interface operations, data writing back is completed by simulating interface element operations to ensure consistency and synchronization of approval results across systems.

[0060] It should be noted that the rules of the approval rule engine are preset through a visual configuration interface. (Administrator) Log in to the configuration interface and select the data type to create an approval process template. The template contains a sequence of approval nodes, each configured with the approver's role, approval timeout, and approval conditions. Approval conditions are expressed using IF-THEN rules. Conditional expressions support logical operators AND, OR, NOT, and comparison operators greater than, less than, and equal to. For example, `IF Data Priority = High AND Data Type = Equipment Maintenance THEN` routes to the operations supervisor. After configuration, the rules are stored in the database in JSON format. The rules engine loads JSON rules files at runtime, parses the rule expressions, and executes logical judgments.

[0061] Example 3 is an embodiment of the present invention. This embodiment provides an intelligent operation and maintenance system for offshore wind power based on a 5G private network, including: a data acquisition module, used to collect real-time video data of on-site operations through 5G base stations deployed in the wind turbine nacelle and the substation, and transmit the real-time video data to an edge computing platform deployed in the substation equipment room via the 5G private network; The edge recognition module is used to perform device component identification and fault feature detection on the real-time video data by using edge intelligent recognition processing in the edge computing platform, and to synchronously transmit the recognition and detection results with the real-time video data to the back-end expert terminal. The annotation processing module is used to receive guidance annotation information generated by experts in the background for the real-time video data, bind the guidance annotation information with the spatial position of the device component, and send the bound annotation information back to the on-site augmented reality device for overlay display; The data acquisition module is used to extract the data to be approved from the business system by adaptively selecting the data acquisition method according to the interface characteristics of the business system through multi-source data acquisition and processing. The approval processing module is used to perform data transformation processing on the data to be approved to obtain standardized data, input the standardized data into a preset approval rule engine for judgment to obtain the approval result, generate write-back data based on the approval result and write it into the business system.

[0062] This embodiment also provides an electronic device applicable to a smart operation and maintenance method for offshore wind power based on a 5G private network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the smart operation and maintenance method for offshore wind power based on a 5G private network as proposed in the above embodiment.

[0063] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements a smart operation and maintenance method for offshore wind power based on a 5G private network as proposed in the above embodiment.

[0064] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing intelligent operation and maintenance of offshore wind power based on 5G private network proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0065] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent operation and maintenance of offshore wind power based on a 5G private network, characterized in that: include, Real-time video data of on-site operations is collected by 5G base stations deployed in wind turbine nacelles and substations, and the real-time video data is transmitted to an edge computing platform deployed in the substation equipment room via a 5G private network. In the edge computing platform, edge intelligent recognition processing is used to identify equipment components and detect fault features in the real-time video data, and the recognition and detection results are synchronously transmitted to the back-end expert terminal along with the real-time video data; The system receives guidance annotation information generated by experts in the background based on the real-time video data, binds the guidance annotation information with the spatial position of the device component, and sends the bound annotation information back to the on-site augmented reality device for overlay display. The system employs a multi-source data acquisition and processing approach, adaptively selecting the data acquisition method based on the interface characteristics of the business system to extract the data to be approved from the business system. The data to be approved is processed to obtain standardized data. The standardized data is then input into a preset approval rule engine for judgment to obtain the approval result. Based on the approval result, write-back data is generated and written into the business system.

2. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 1, characterized in that: The step of using edge intelligent recognition processing to identify equipment components and detect fault features in the real-time video data includes: Video frames are extracted from the real-time video data, and image preprocessing is performed on the video frames; The preprocessed video frames are analyzed using a target detection algorithm to extract device image features; The similarity between the device image features and the pre-stored device component feature library is calculated, and the device component type is determined based on the similarity calculation results. Anomaly feature extraction is used to extract anomaly feature parameters from the real-time video data. The anomaly feature extraction method includes performing temperature distribution analysis on thermal imaging data to determine temperature anomaly areas and performing texture analysis on equipment surface images to identify cracks or deformations. The fault type is determined based on the abnormal feature parameters and the preset fault discrimination rules.

3. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 2, characterized in that: The step of calculating the similarity between the device image features and a pre-stored device component feature library includes: Extract the key point coordinates and descriptors from the device image features; Read the standard key point coordinates and standard descriptors corresponding to each equipment component from the equipment component feature library; Calculate the spatial distance between the keypoint coordinates and the standard keypoint coordinates, and calculate the vector distance between the descriptor and the standard descriptor; The similarity score is calculated based on the spatial distance and the vector distance, and the device component with the highest similarity score is selected as the recognition result.

4. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 3, characterized in that: The step of binding the guidance labeling information with the spatial location of the device component includes: Visual positioning processing is used to determine the spatial pose of the augmented reality device relative to the device components. The visual positioning processing includes acquiring environmental images and extracting environmental feature points through the augmented reality device, matching the environmental feature points with a pre-constructed three-dimensional point cloud map, and calculating the position coordinates and attitude angle of the augmented reality device based on the matching results. Based on the spatial pose and the pre-stored 3D model coordinates of the device components, calculate the anchor point coordinates of the guidance annotation information in 3D space; The anchor point coordinates are associated with the identification information of the equipment components and stored.

5. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 4, characterized in that: The step of matching the environmental feature points with a pre-constructed 3D point cloud map includes: Extract a local point cloud subset from the 3D point cloud map; Calculate the descriptor similarity between the environmental feature points and each point in the local point cloud subset; Point pairs with descriptor similarity greater than a preset threshold are selected as candidate matching point pairs; The candidate matching point pairs are subjected to geometric consistency checks. Matching point pairs that do not meet the geometric constraints are eliminated, and the matching point pairs that pass the check are retained for calculating the spatial pose of the augmented reality device.

6. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 5, characterized in that: The step of using multi-source data acquisition and processing to adaptively select data acquisition methods based on the interface characteristics of the business system to extract data to be approved from the business system includes: Send an interface probe request to the business system; Parse the interface description information returned by the business system to determine whether the business system provides an application programming interface (API). When it is determined that the business system provides an application programming interface (API), the API is invoked and query parameters are passed in to obtain the data to be approved. When it is determined that the business system does not provide an application programming interface, the system checks whether the business system has database access permissions. If database access permissions are enabled, the system executes a query statement through a database connection to obtain the data to be approved. If database access permissions are not enabled, the system logs into the business system and retrieves the data to be approved by simulating a user interface.

7. The intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in claim 6, characterized in that: The step of logging into the business system and retrieving the data to be approved via a simulated user interface includes: Obtain page element information for the login page of the business system; Locate the username and password input boxes based on the page element information, and input the pre-stored login credentials into the username and password input boxes; The login button element is triggered to complete the login process and enter the data display page of the business system; The page structure of the data display page is analyzed, and data records containing the pending approval identifier are extracted as the pending approval data.

8. A smart operation and maintenance system for offshore wind power based on a 5G private network, employing the smart operation and maintenance method for offshore wind power based on a 5G private network as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect real-time video data of on-site operations through 5G base stations deployed in the wind turbine nacelle and the substation, and transmit the real-time video data to the edge computing platform deployed in the substation equipment room via the 5G private network. The edge recognition module is used to perform device component identification and fault feature detection on the real-time video data by using edge intelligent recognition processing in the edge computing platform, and to synchronously transmit the recognition and detection results with the real-time video data to the back-end expert terminal. The annotation processing module is used to receive guidance annotation information generated by experts in the background for the real-time video data, bind the guidance annotation information with the spatial position of the device component, and send the bound annotation information back to the on-site augmented reality device for overlay display; The data acquisition module is used to extract the data to be approved from the business system by adaptively selecting the data acquisition method according to the interface characteristics of the business system through multi-source data acquisition and processing. The approval processing module is used to perform data transformation processing on the data to be approved to obtain standardized data, input the standardized data into a preset approval rule engine for judgment to obtain the approval result, generate write-back data based on the approval result and write it into the business system.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent operation and maintenance method for offshore wind power based on a 5G private network as described in any one of claims 1 to 7.

Images