A wind turbine yaw oil leak early warning system and method

Abstract

A wind turbine yaw oil leak early warning system and method are disclosed, addressing the problems of low accuracy and poor real-time performance of manual inspections in existing yaw oil leak monitoring, as well as the insufficient adaptability and lack of a complete early warning loop in traditional machine vision solutions. The system includes an image acquisition module, an image preprocessing module, a lightweight oil leak detection model module, an early warning judgment module, and a data storage module. The image acquisition module acquires images of the yaw's sealing area. After quality optimization by the image preprocessing module, the images are input to the lightweight oil leak detection model module, which is based on an improved YOLOv5 convolutional neural network, to identify and locate the oil leak area. The early warning judgment module determines the early warning level and generates instructions based on the detection results using preset logic or quantification formulas. The data storage module retains all data throughout the process. The method achieves real-time oil leak early warning through the steps of "image acquisition - preprocessing - oil leak detection - early warning generation," with modules working together. This invention can adapt to the complex environment of wind farms, improve the accuracy and real-time performance of oil leak detection, reduce operation and maintenance costs, and form a complete early warning loop, providing technical support for the early detection and efficient handling of yaw oil leaks.

Description

Technical Field

[0001] This invention belongs to the field of wind power operation and maintenance technology, and specifically relates to a wind turbine yaw oil leakage early warning system and method. Background Technology

[0002] The sealing parts of wind turbine yaws are prone to oil leakage due to the complex outdoor environment and aging. If not addressed promptly, this can lead to bearing wear and downtime losses. Current technologies rely primarily on manual inspections, which suffer from long cycles (1-2 weeks / time), high rates of missed or false diagnoses (minor oil leaks have a detection rate of <60%), and high maintenance costs. Some machine vision solutions have been explored, but they are not optimized for wind farm scenarios, resulting in low detection accuracy in low light and dusty environments. Furthermore, the models are bulky and difficult to deploy at edges, and they lack a closed-loop "identification-early warning-tracing" system, failing to support efficient operation and maintenance.

[0003] In summary, the existing technologies have the following technical problems: poor real-time performance and low accuracy of manual inspections, hindering early and accurate identification of oil leaks; insufficient detection accuracy of existing machine vision solutions in complex environments and difficulty in real-time model deployment; and a lack of a complete early warning loop, resulting in delayed operation and maintenance response. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a wind turbine yaw oil leakage early warning system and method. The invention can adapt to the complex environment of wind farms, improve the accuracy and real-time performance of oil leakage detection, reduce operation and maintenance costs, form a complete early warning closed loop, and provide technical support for the early detection and efficient handling of yaw oil leakage.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A wind turbine yaw oil leak early warning system includes an image acquisition module, an image preprocessing module, a lightweight oil leak detection model module, an early warning judgment module, and a data storage module. The image acquisition module is located near the yaw of the wind turbine and is used to acquire image data of the yaw's sealing part and surrounding area to obtain basic visual information on whether the yaw is leaking oil. The image preprocessing module establishes a data connection with the image acquisition module to process the image data acquired by the image acquisition module, improve image quality, and provide better image materials for subsequent oil leak identification. The lightweight oil leak detection model module is built based on an improved YOLOv5 convolutional neural network. It communicates with the image preprocessing module and is used to analyze the image data processed by the image preprocessing module, identify and locate possible oil leak areas in the image, and output oil leak-related detection results. The early warning judgment module is connected to the lightweight oil leak detection model module and has a preset oil leak early warning judgment logic. It is used to determine the early warning level corresponding to the yaw oil leak based on the detection results output by the lightweight oil leak detection model module and generate corresponding early warning information. The data storage module establishes data interaction with the image acquisition module, image preprocessing module, lightweight oil leak detection model module, and early warning judgment module, respectively, to store the image data, processing results, and early warning information generated during the operation of each module.

[0006] Preferably, the image acquisition module uses an industrial-grade image sensor. This industrial-grade image sensor has protective performance to adapt to the complex outdoor environment of wind farms. It is installed on the outside of the flange sealing surface of the wind turbine yaw, around the sealing cover of the yaw drive device, and the grease sealing part of the yaw bearing, so as to fully cover the key areas of the yaw that are prone to oil leakage and ensure that image data of each potential oil leakage point of the yaw can be acquired.

[0007] Preferably, the lightweight oil leak detection model module is based on an improved YOLOv5 convolutional neural network, which achieves lightweight optimization by reducing the number of detection heads and one feature pyramid downsampling process. This reduces the computational load of the improved YOLOv5 convolutional neural network and improves the response speed of oil leak detection while ensuring the oil leak detection effect.

[0008] Preferably, the image preprocessing module includes a noise removal submodule and an image enhancement submodule. The noise removal submodule is data-connected to the image acquisition module and is used to eliminate interference noise in the image acquired by the image acquisition module. The image enhancement submodule is connected to the noise removal submodule and is used to adjust the brightness, contrast, and enhance the details of the image after noise removal. The noise removal submodule and the image enhancement submodule work together in sequence to complete the data processing of the original image.

[0009] A method for early warning of oil leakage in a wind turbine yaw, employing the aforementioned early warning system for oil leakage in a wind turbine yaw, includes the following steps: S1: Image acquisition. The image acquisition module acquires image data of the sealing part and surrounding area of ​​the yaw of the wind turbine. The acquired image data is transmitted to the image preprocessing module and the data storage module respectively. The data storage module saves the original image data. S2: Image preprocessing. The image preprocessing module receives image data transmitted from the image acquisition module, performs preset processing operations on the image data to obtain optimized image data, transmits the optimized image data to the lightweight oil leak detection model module, and simultaneously sends the optimized image data to the data storage module for storage. S3: Oil Leak Detection. The lightweight oil leak detection model module loads pre-trained oil leak recognition parameters, receives optimized image data transmitted from the image preprocessing module, inputs the optimized image data into the lightweight oil leak detection model module loaded with pre-trained oil leak recognition parameters, and uses the lightweight oil leak detection model module to calculate and identify the oil leak area in the image, generating a detection result containing the location and features of the oil leak area. The detection result is then transmitted to the early warning judgment module and simultaneously stored in the data storage module. S4: Warning generation. The warning judgment module receives the detection results transmitted by the lightweight oil leak detection model module, analyzes the detection results according to the preset warning judgment logic, determines the warning level corresponding to the yaw oil leak, generates a warning command containing the warning level and oil leak related information, sends the warning command to the designated terminal, and stores the warning command in the data storage module.

[0010] Preferably, in step S1, before acquiring image data, the image acquisition module first senses the ambient light conditions of the acquisition area and adjusts its acquisition parameters according to the sensed light conditions to ensure that clear images of the yaw seal can be acquired under different lighting conditions. During the acquisition process, the image acquisition module continuously acquires image data according to a preset acquisition frequency. Each frame of image data acquired is accompanied by identification information of acquisition time and acquisition location. Then, the image data with identification information is transmitted to the image preprocessing module. At the same time, the image data and identification information are synchronously transmitted to the data storage module in a structured data format. The data storage module classifies and stores the image data according to the acquisition time order.

[0011] Preferably, in step S2, after the image preprocessing module receives the image data, the noise removal submodule first processes the image data, using a filtering algorithm to filter out salt-and-pepper noise and Gaussian noise interference signals in the image to obtain denoised image data. Then, the denoised image data is transmitted to the image enhancement submodule. The image enhancement submodule adjusts the brightness distribution of the image to make the overall brightness of the image uniform, while optimizing the image contrast, enhancing the visual difference between the oil leak area and the background in the image, and enhancing the detailed features in the image to highlight the edges of the possible oil leak area. After processing, the image preprocessing module performs quality verification on the optimized image data. If the verification result meets the preset standard, the optimized image data is transmitted to the lightweight oil leak detection model module. If the verification fails, it is returned to the image acquisition module, triggering the image acquisition module to re-acquire the image data of the corresponding area.

[0012] Preferably, in step S3, the pre-trained oil leak identification parameters loaded by the lightweight oil leak detection model module are trained based on an image dataset of wind turbine yaw oil leak scenarios. This image dataset contains yaw image samples under different oil leak levels and environmental conditions. During training, Mos...A IC data augmentation technology expands the image samples and adjusts the training parameters of the lightweight oil leak detection model module to improve the accuracy and generalization ability of the lightweight oil leak detection model module in identifying yaw oil leak areas. During the operation of the lightweight oil leak detection model module, the optimized image data is divided into multiple sub-regions, and features are extracted and analyzed in each sub-region in turn to determine whether there are oil leak features in the sub-region. For the sub-regions determined to have oil leaks, their coordinate range, region size and oil leak feature matching degree in the overall image are calculated. This information is integrated to form a complete detection result, and then the detection result is transmitted to the early warning judgment module in the form of a data packet. At the same time, the detection result is associated with the corresponding optimized image data and stored in the data storage module.

[0013] Preferably, in step S4, after receiving the detection result, the early warning judgment module calls the preset early warning judgment logic, first extracts the key parameters of the oil leak area size and oil leak feature matching degree from the detection result, and compares these key parameters with the preset early warning level classification standard. If the size of the oil leak area and the degree of feature matching both meet the criteria for a Level 1 warning, then it is determined to be a Level 1 warning. If the criteria for a Level II warning are met, it will be classified as a Level II warning. If the criteria for a Level 3 warning are met, it will be classified as a Level 3 warning. After determining the warning level, the warning judgment module generates a warning command, which includes the warning level, time of oil leak occurrence, location of oil leak, and parameters of the oil leak area. Subsequently, the warning command is sent to the wind farm monitoring terminal and the mobile terminals of maintenance personnel using encrypted transmission. After receiving the warning command, the terminal notifies relevant personnel through audio-visual prompts or message push. At the same time, the warning judgment module associates the warning command with the corresponding detection results and image data and stores them in the data storage module for subsequent query and traceability.

[0014] Preferably, in step S4, the early warning judgment module calculates the early warning score using the early warning level judgment formula, and determines the early warning level corresponding to the yaw oil leak based on the early warning score, as follows: The formula for determining the warning level is as follows:

[0015] The symbols in the formula are defined as follows: S : Represents the warning score, with a value range of [0,100], used to quantify the severity of the yaw oil leak and serve as the basis for determining the warning level; : Represents the weighting coefficient for the size of the oil leak area, with a value of 0.6, used to characterize the degree of influence of the size of the oil leak area on the warning level; A: Represents the normalized value of the oil leak area size. It is a dimensionless value, ranging from [0, 100]. It is calculated from the actual area of ​​the oil leak area. The calculation method is as follows: ,in The actual oil leak area identified by the lightweight oil leak detection model module. The threshold for the maximum permissible oil leakage area at the yaw seal; : Represents the weighting coefficient of oil leak feature matching degree, with a value of 0.4, used to characterize the degree of influence of oil leak feature matching degree on the warning level; M : Oil leak feature matching degree, with a value range of [0,100], is obtained by converting the oil leak feature matching degree result output by the lightweight oil leak detection model module (multiplying the matching degree value in the [0,1] interval output by the model by 100). The criteria for determining the warning level are: if the warning score is... If it is, it is determined to be a Level 1 warning; if If it is, it is determined to be a Level II warning; if If so, it is determined to be a Level 3 warning.

[0016] The present invention can achieve the following beneficial effects: 1. This invention optimizes image quality in complex wind farm environments (low light, dust) through an image enhancement module. Combined with an improved lightweight oil leak detection model, it can accurately identify minor oil leaks in yaw bearing seals, effectively avoiding missed or misjudged minor oil traces during manual inspections. At the same time, the lightweight optimization of the model enables rapid processing of image data and real-time capture of oil leak faults, preventing the oil leak from expanding and causing secondary faults such as yaw bearing wear and gearbox lubrication failure, thus reducing losses caused by downtime due to faults.

[0017] 2. The system can automatically monitor yaw oil leaks without the need for regular on-site inspections, significantly reducing the workload of wind farm operation and maintenance. Especially for wind farms in remote areas with harsh outdoor environments, it can prevent maintenance personnel from being exposed to extreme weather or high-altitude work risks for a long time. At the same time, it saves the transportation and labor costs incurred by manual inspections, reducing the pressure on wind farm operation and maintenance.

[0018] 3. The lightweight oil leak detection model reduces computational load and parameter quantity by optimizing the network structure. It can be directly deployed on edge devices of wind farms without relying on cloud computing power, avoiding data transmission delays or lags caused by bandwidth limitations. In addition, each module takes into account the characteristics of the outdoor environment of wind farms and has the ability to adapt to complex conditions such as extreme temperatures, dust, and rain. It can operate stably for a long time and reduce the frequency of equipment failures.

[0019] 4. The system accurately determines the severity of oil leaks through hierarchical early warning logic. The generated early warning instructions include the location of the oil leak, fault characteristics, and handling suggestions, which can be pushed to the maintenance terminal in real time to help maintenance personnel quickly grasp the fault information. At the same time, the data storage module retains complete image data, detection results, and early warning records, supporting subsequent traceability and fault analysis, which facilitates maintenance personnel to quickly formulate maintenance plans and shorten the fault response and handling cycle. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] Preferred solutions include Figure 1 As shown, a wind turbine yaw oil leak early warning system includes an image acquisition module, an image preprocessing module, a lightweight oil leak detection model module, an early warning judgment module, and a data storage module. The image acquisition module is located near the yaw of the wind turbine and is used to acquire image data of the yaw's sealing part and surrounding area to obtain basic visual information on whether the yaw is leaking oil. The image preprocessing module establishes a data connection with the image acquisition module to process the image data acquired by the image acquisition module, improve image quality, and provide better image materials for subsequent oil leak identification. The lightweight oil leak detection model module is built based on an improved YOLOv5 convolutional neural network. It communicates with the image preprocessing module and is used to analyze the image data processed by the image preprocessing module, identify and locate possible oil leak areas in the image, and output oil leak-related detection results. The early warning judgment module is connected to the lightweight oil leak detection model module and has a preset oil leak early warning judgment logic. It is used to determine the early warning level corresponding to the yaw oil leak based on the detection results output by the lightweight oil leak detection model module and generate corresponding early warning information. The data storage module establishes data interaction with the image acquisition module, image preprocessing module, lightweight oil leak detection model module, and early warning judgment module, respectively, to store the image data, processing results, and early warning information generated during the operation of each module.

[0022] Each module of this system implements oil leak early warning through a closed-loop process of "data acquisition-processing-analysis-decision-storage". The specific operating logic is as follows: After the image acquisition module captures images of the yaw's sealing area in real time, it synchronously transmits the raw image data to the image preprocessing module and the data storage module via Ethernet. The image preprocessing module first removes noise and adjusts brightness and contrast in the image, then passes the optimized image data to the lightweight oil leak detection model module. This model module extracts the texture and grayscale features of the oil leak area in the image using an improved YOLOv5 convolutional neural network, locates the oil leak location, calculates relevant parameters, and generates detection results. After receiving the detection results, the early warning judgment module calls the built-in judgment logic to compare parameters, determines the early warning level, and generates instructions. Finally, all data (raw image, optimized image, detection results, and early warning instructions) are stored in the data storage module, forming a complete data traceability chain.

[0023] The image preprocessing module improves image quality through a two-step process of "noise suppression and detail enhancement". First, a filtering algorithm is used to filter out noise caused by wind farm dust and fog. Then, a brightness equalization algorithm is used to eliminate the occlusion of details of the sealing part by backlight and low light, ensuring that the visual difference between the oil leak area and the background is amplified, providing high-quality input data for the subsequent detection model. The lightweight oil leak detection model module is based on an improved YOLOv5 convolutional neural network. By reducing the number of detection heads and optimizing the feature pyramid structure, the computational load is reduced, enabling it to run in real time on edge devices. When the model extracts features from the input optimized image, it focuses on the grayscale distribution and edge contour of the oil stains on the yaw sealing surface. By comparing with the pre-trained oil leak feature library, it achieves accurate identification and location of the oil leak area. The judgment logic of the early warning judgment module is set based on wind farm operation and maintenance experience. By associating two core parameters, namely the size of the oil leak area and the feature matching degree, it avoids misjudgment based on a single parameter (such as a minor oil leak with a high matching degree in a small area and a suspected oil leak with a low matching degree in a large area), and ensures the accuracy of the early warning level.

[0024] Furthermore, the image acquisition module adopts an industrial-grade image sensor, which has protective performance to adapt to the complex outdoor environment of wind farms. It is installed on the outside of the flange sealing surface of the wind turbine yaw, around the sealing cover of the yaw drive device, and the grease sealing part of the yaw bearing, so as to fully cover the key areas of the yaw that are prone to oil leakage and ensure that image data of each potential oil leakage point of the yaw can be collected.

[0025] Wind farms present complex environments with low temperatures, strong winds, and dust storms. The protective performance of industrial-grade image sensors directly determines the stability of data acquisition. Sensors, through sealed housings and low-temperature resistant component designs, prevent rainwater and dust from seeping into the internal circuitry and affecting its operation. They also withstand performance fluctuations caused by extreme temperatures, ensuring continuous output of clear images under various weather conditions and providing stable raw data input for subsequent modules. Yaw lever oil leaks often occur at the contact points between sealing surfaces and moving parts. Flange sealing surfaces (at the connection between upper and lower flanges), yaw drive unit sealing covers (sealing between the drive gear and shaft), and yaw bearing grease seals (bearing lubricant seals) are high-frequency leak points. Installing sensors around these locations allows for precise "point-to-point" data acquisition, avoiding image blind spots caused by installation misalignment. For example, a sensor installed on the outside of the flange sealing surface, with its lens angled towards the sealing joint, can directly capture oil seepage at the joint. A sensor at the yaw bearing grease seal can monitor the grease filling port and sealing strip for leaks in real time, ensuring no potential leaks are missed.

[0026] Furthermore, the improved YOLOv5 convolutional neural network on which the lightweight oil leak detection model module is based achieves lightweight optimization by reducing the number of detection heads and one feature pyramid downsampling process, so as to reduce the computational load of the improved YOLOv5 convolutional neural network and improve the response speed of oil leak detection while ensuring the oil leak detection effect.

[0027] The original YOLOv5 convolutional neural network contains three detection heads, corresponding to large, medium, and small-scale target detection respectively. In the context of yaw engine oil leaks, the leak areas are mostly medium and small-scale (such as oil stains of a few square centimeters to tens of square centimeters), and large-scale oil leaks are extremely rare. By reducing one detection head for large-scale targets, redundant large-scale feature extraction and calculation steps can be eliminated. Without affecting the oil leak detection effect, the model parameters can be reduced by about 30%, thus reducing the consumption of computing resources. The core function of feature pyramid downsampling is to extract image features at different levels. The original network's three downsampling operations lead to excessive compression of small-scale oil leak features. After reducing downsampling by one, the feature pyramid retains more detailed features of small-scale oil leaks (such as the edge texture of subtle oil stains), while reducing the convolution operation steps corresponding to one downsampling operation, further reducing the computational load. For example, for a slight oil leak area of ​​5 cm², the optimized feature pyramid can more clearly retain its grayscale distribution features, avoiding feature loss due to oversampling. This ensures detection accuracy and reduces the single-frame inference time from 0.25 s to 0.18 s, meeting the real-time detection requirements of wind farms.

[0028] Furthermore, the image preprocessing module includes a noise removal submodule and an image enhancement submodule. The noise removal submodule is data-connected to the image acquisition module and is used to eliminate interference noise in the image acquired by the image acquisition module. The image enhancement submodule is connected to the noise removal submodule and is used to adjust the brightness, contrast, and enhance the details of the image after noise removal. The noise removal submodule and the image enhancement submodule work together in sequence to complete the data processing of the original image.

[0029] Noise in wind farm images (salt-and-pepper noise caused by dust and Gaussian noise caused by fog) can mask the features of oil leak areas. If these images are directly fed into subsequent detection modules, they can easily lead to misjudgments or missed detections. The noise removal submodule uses filtering algorithms to process noise based on its gray-level abrupt change characteristics (salt-and-pepper noise) and distribution characteristics (Gaussian noise). For example, it uses median filtering to eliminate salt-and-pepper noise (replacing abrupt gray-level values ​​with the mean of surrounding pixels) and bilateral filtering to smooth Gaussian noise (preserving image edges while reducing gray-level fluctuations). This ensures that noise removal does not destroy the original features of the oil leak area. The denoised image may have uneven brightness and low contrast (such as images taken on cloudy days or at night), resulting in small grayscale differences between the oil leak area and the background. The image enhancement submodule adjusts the overall grayscale distribution of the image through a brightness equalization algorithm, increasing the brightness of dark areas and decreasing the brightness of bright areas to avoid local overexposure or underexposure. At the same time, it enhances the grayscale difference between the oil leak area and the background through contrast adjustment, for example, adjusting the grayscale value of the oil leak area from 100 to 80 and the grayscale value of the background from 110 to 130 to make the oil stains more obvious. Detail enhancement targets the edges of the oil leak area, highlighting the outline of the oil stains through an edge enhancement algorithm, providing clear feature input for the detection model.

[0030] Noise removal is a prerequisite for image enhancement. If the image is enhanced first, the noise will be amplified simultaneously, which will increase the difficulty of subsequent processing. The two sub-modules run in sequence, first eliminating interference and then optimizing image quality, forming a "cleaning-optimization" processing flow. This ensures that the output optimized image can accurately reflect the characteristics of the oil leak without noise interference, providing high-quality data support for the detection module.

[0031] like Figure 2 As shown, a method for early warning of oil leakage in a wind turbine yaw includes the following steps: S1: Image acquisition. The image acquisition module acquires image data of the sealing part and surrounding area of ​​the yaw of the wind turbine. The acquired image data is transmitted to the image preprocessing module and the data storage module respectively. The data storage module saves the original image data. Image acquisition is the initial data source for the method. The module acquires images at a preset frequency and transmits them synchronously to both the preprocessing module and the storage module, ensuring real-time processing and backing up the original data. For example, the acquisition frequency of a wind farm unit is set to 30fps when yawing and 10fps when running stably. At 10:00 on June 20, 2025, the unit was in a stable operating state. The image acquisition module acquired an image of the upper flange sealing surface during yaw (resolution 1920×1080), with the identification information "Unit 05_20250620_100000_Upper Flange", and transmitted it synchronously to the preprocessing module (transmission rate 100Mbps, delay <0.1s) and the data storage module (saved in JPEG format, with each image size about 2MB).

[0032] S2: Image preprocessing. The image preprocessing module receives image data transmitted from the image acquisition module, performs preset processing operations on the image data to obtain optimized image data, transmits the optimized image data to the lightweight oil leak detection model module, and simultaneously sends the optimized image data to the data storage module for storage. Image preprocessing is a crucial step connecting acquisition and detection, optimizing image quality through preset processing operations. For example, after receiving the original image (average brightness 65, signal-to-noise ratio 28dB) from step S1, the noise removal submodule uses a filtering algorithm to improve the signal-to-noise ratio to 42dB; the image enhancement submodule adjusts the average brightness to 125 and increases the contrast from 0.3 to 0.8; after processing, an optimized image (approximately 1.8MB in size) is generated, transmitted to the detection module, and simultaneously stored in the data storage module to ensure traceability of the processing.

[0033] S3: Oil Leak Detection. The lightweight oil leak detection model module loads pre-trained oil leak recognition parameters, receives optimized image data transmitted from the image preprocessing module, inputs the optimized image data into the lightweight oil leak detection model module loaded with pre-trained oil leak recognition parameters, and uses the lightweight oil leak detection model module to calculate and identify the oil leak area in the image, generating a detection result containing the location and features of the oil leak area. The detection result is then transmitted to the early warning judgment module and simultaneously stored in the data storage module. Oil leak detection achieves feature recognition by comparing pre-trained parameters with optimized images. For example, the detection module loads parameters trained based on 10,000 yaw images, inputs the optimized image from step S2, extracts image features through convolution operations, identifies an oil leak area on the upper flange sealing surface, and calculates the coordinates (x1=520, y1=380, x2=580, y2=440), area 36cm², and feature matching degree 92. The detection result is generated and transmitted to the early warning module, and the optimized image is stored synchronously.

[0034] S4: Warning generation. The warning judgment module receives the detection results transmitted by the lightweight oil leak detection model module, analyzes the detection results according to the preset warning judgment logic, determines the warning level corresponding to the yaw oil leak, generates a warning command containing the warning level and oil leak related information, sends the warning command to the designated terminal, and stores the warning command in the data storage module.

[0035] The early warning generation outputs decision instructions through logical analysis; for example, the early warning module receives the detection result of step S3 (area 36cm², matching degree 92), calls the preset judgment logic (level 1 early warning: area ≥ 20cm² and matching degree ≥ 90), and judges it as a level 1 early warning; it generates an early warning instruction (including early warning level, oil leak time, location, and parameters), pushes it to the operation and maintenance terminal, and stores it in the data storage module at the same time, forming a complete early warning closed loop.

[0036] Furthermore, in step S1, before acquiring image data, the image acquisition module first senses the ambient light conditions of the acquisition area and adjusts its acquisition parameters according to the sensed light conditions to ensure that clear images of the yaw seal can be acquired under different lighting conditions. During the acquisition process, the image acquisition module continuously acquires image data according to a preset acquisition frequency. Each frame of image data acquired is accompanied by identification information of acquisition time and acquisition location. Subsequently, the image data with identification information is transmitted to the image preprocessing module. At the same time, the image data and identification information are synchronously transmitted to the data storage module in a structured data format. The data storage module classifies and stores the image data according to the acquisition time order.

[0037] The image acquisition module has a built-in light sensor that senses the ambient light intensity in real time. It adjusts parameters according to the logic of "light intensity - acquisition parameters" to ensure image clarity. For example, when the light intensity is <200 lux (cloudy day, night), the sensor transmits the signal to the control unit, which automatically increases the ISO (from 100 to 400), extends the exposure time (from 1 / 25s to 1 / 10s), and turns on the fill light. When the light intensity is >80000 lux (strong midday light), it decreases the ISO (to 100), shortens the exposure time (to 1 / 100s), and turns off the fill light.

[0038] The acquisition frequency is dynamically adjusted according to the yaw status of the unit to avoid data redundancy or omission. For example, the acquisition frequency is 30fps when the unit is yawing (in operation state) to ensure the capture of dynamic oil leakage (such as the spread of oil stains on the sealing surface during yaw). The acquisition frequency is 10fps when the unit is running stably to reduce invalid data. The identification information (unit number, time, and location) attached to each frame of the image can achieve "precise traceability". For example, the identification "unit 08_20250620_193000_yaw bearing" directly corresponds to the specific unit, time, and acquisition location, which is convenient for subsequent query and fault tracing.

[0039] The data storage module is structured and categorized according to "time-unit-location". For example, the image of Unit 08 on June 20, 2025, is stored in paths such as "20250620 / Unit 08 / Upper Flange / " and "20250620 / Unit 08 / Yaw Bearing / ". At the same time, an index is created by associating identification information. When querying, the corresponding image can be quickly located by entering the time and unit number, which improves the efficiency of data retrieval.

[0040] Preferably, in step S2, after the image preprocessing module receives the image data, the noise removal submodule first processes the image data, using a filtering algorithm to filter out salt-and-pepper noise and Gaussian noise interference signals in the image to obtain denoised image data. Then, the denoised image data is transmitted to the image enhancement submodule. The image enhancement submodule adjusts the brightness distribution of the image to make the overall brightness of the image uniform, while optimizing the image contrast, enhancing the visual difference between the oil leak area and the background in the image, and enhancing the detailed features in the image to highlight the edges of the possible oil leak area. After processing, the image preprocessing module performs quality verification on the optimized image data. If the verification result meets the preset standard, the optimized image data is transmitted to the lightweight oil leak detection model module. If the verification fails, it is returned to the image acquisition module, triggering the image acquisition module to re-acquire the image data of the corresponding area.

[0041] In wind farm images, salt-and-pepper noise (white / black spots caused by dust) and Gaussian noise (grayscale fluctuations caused by fog) can interfere with the identification of oil leak features. The noise removal submodule adopts a combination algorithm of "median filtering + bilateral filtering". The median filtering is used for salt-and-pepper noise (3×3 window processing), and the bilateral filtering is used for Gaussian noise (spatial domain standard deviation σd=5, grayscale domain standard deviation σr=20). For example, a certain original image (yaw drive device sealing cover) has obvious salt-and-pepper noise. Before denoising, the oil stains on the edge of the sealing cover in the image are covered by white spots. After denoising, the signal-to-noise ratio is improved from 26dB to 41dB, and the edge of the oil stains is initially revealed.

[0042] The denoised image may have uneven brightness and low contrast. The image enhancement submodule adjusts the overall brightness through a brightness equalization algorithm (histogram equalization), increasing the brightness of dark areas and decreasing the brightness of bright areas. For example, it adjusts the average image brightness from 70 to 120 and the standard deviation of brightness from 35 to 20 to ensure uniform overall brightness. Simultaneously, it uses a contrast enhancement algorithm (CL... A HE) increases the grayscale difference between the oil leak area and the background from 15 to 35. For example, if the grayscale value of the oil leak area is 80 and the grayscale value of the background is 115, after processing, the grayscale value of the oil leak area is 70 and the grayscale value of the background is 105, resulting in a significant visual difference. Detail enhancement uses an edge detection algorithm (Sobel operator) to highlight the edges of the oil stains, enabling the detection model to accurately locate the oil leak area.

[0043] The quality verification is set with three core standards (resolution ≥ 300 dpi, average brightness 100~150, contrast ≥ 0.5). If the optimized image does not meet any of the standards, it means that the image quality cannot support subsequent detection. For example, if the resolution of an optimized image is 280 dpi (unacceptable), the preprocessing module sends a "re-acquisition command" to the acquisition module. The acquisition module adjusts the parameters (such as increasing the focal length and increasing the exposure time) and re-acquires the image until a qualified image is obtained, so as to avoid the detection module misjudging or missing the image due to low-quality images.

[0044] Furthermore, in step S3, the pre-trained oil leak identification parameters loaded into the lightweight oil leak detection model module are trained based on an image dataset of wind turbine yaw oil leak scenarios. This image dataset contains yaw image samples under different oil leak levels and environmental conditions. During training, Mos... A IC data augmentation technology expands the image samples and adjusts the training parameters of the lightweight oil leak detection model module to improve the accuracy and generalization ability of the lightweight oil leak detection model module in identifying yaw oil leak areas. During the operation of the lightweight oil leak detection model module, the optimized image data is divided into multiple sub-regions, and features are extracted and analyzed in each sub-region in turn to determine whether there are oil leak features in the sub-region. For the sub-regions determined to have oil leaks, their coordinate range, region size and oil leak feature matching degree in the overall image are calculated. This information is integrated to form a complete detection result, and then the detection result is transmitted to the early warning judgment module in the form of a data packet. At the same time, the detection result is associated with the corresponding optimized image data and stored in the data storage module.

[0045] Pre-trained parameters serve as "feature templates" for the detection model to identify oil leaks. The training dataset needs to cover different scenarios to improve generalization ability; for example, the dataset contains 20,000 images, including 5,000 images of minor oil leaks (area 5~10cm²), 5,000 images of moderate oil leaks (10~20cm²), 5,000 images of severe oil leaks (>20cm²), and 5,000 normal images, covering environments such as cloudy days, nighttime, and dust storms; during training, MOS is used. A For IC data augmentation, four images from different scenarios are randomly stitched together each time (e.g., a minor oil leak on a cloudy day + a moderate oil leak at night + a normal image of a sandstorm + a severe oil leak on a sunny day), expanding the dataset to 80,000 images; training parameters are adjusted to ensure the accuracy of identifying oil leaks in different scenarios.

[0046] For images with large sizes (e.g., 1920×1080), direct analysis can lead to high computational costs and inaccurate feature extraction. The detection module divides the image into 112×112 sub-regions (324 in total) and extracts the grayscale distribution and edge contour features of each sub-region in turn. For example, a certain sub-region (coordinates x1=600, y1=400, x2=712, y2=512) corresponds to the yaw flange sealing surface. The grayscale average value of this region is extracted to be 82 (lower than the background average value of 110), and the edge is an irregular ellipse (oil stain feature). After comparison with the pre-training parameters, it is determined to be an oil leak sub-region. The actual area of ​​this sub-region is calculated to be 12cm² (1 pixel corresponds to 0.1cm), and the feature matching degree is 89 (similarity with the pre-trained oil leak features).

[0047] If multiple oil leak sub-regions exist in the image, the module needs to integrate them to form a complete result. For example, if an optimized image detects two oil leak sub-regions (Region 1: area 12cm², matching degree 89; Region 2: area 10cm², matching degree 85), the module converts the sub-region coordinates into overall image coordinates (Region 1: x1=600, y1=400, x2=712, y2=512; Region 2: x1=800, y1=500, x2=912, y2=612), integrates them into a detection result containing the number of oil leaks and the parameters of each region, generates a data message (approximately 500B in size) in JSON format, transmits it to the early warning module, and simultaneously stores the associated optimized image in the data storage module.

[0048] Preferably, in step S4, after receiving the detection result, the early warning judgment module calls the preset early warning judgment logic, first extracts the key parameters of the oil leak area size and oil leak feature matching degree from the detection result, and compares these key parameters with the preset early warning level classification standard. If the size of the oil leak area and the degree of feature matching both meet the criteria for a Level 1 warning, then it is determined to be a Level 1 warning. If the criteria for a Level II warning are met, it will be classified as a Level II warning. If the criteria for a Level 3 warning are met, it will be classified as a Level 3 warning. After determining the warning level, the warning judgment module generates a warning command, which includes the warning level, time of oil leak occurrence, location of oil leak, and parameters of the oil leak area. Subsequently, the warning command is sent to the wind farm monitoring terminal and the mobile terminals of maintenance personnel using encrypted transmission. After receiving the warning command, the terminal notifies relevant personnel through audio-visual prompts or message push. At the same time, the warning judgment module associates the warning command with the corresponding detection results and image data and stores them in the data storage module for subsequent query and traceability.

[0049] Early warning determination requires simultaneous consideration of both the size of the leaking area (reflecting the severity of the leak) and the feature matching degree (reflecting the reliability of the leak determination) to avoid misjudgment based on a single parameter. For example, preset standards include: Level 1 warning (area ≥ 20cm² and matching degree ≥ 90), Level 2 warning (10 ≤ area < 20cm² and 80 ≤ matching degree < 90), and Level 3 warning (area < 10cm² and 70 ≤ matching degree < 80). If a detection result is "area 18cm², matching degree 88", after extracting the parameters and comparing them with the standards, it meets the conditions for a Level 2 warning and is determined to be a Level 2 warning. If only the area (18cm² is close to the Level 1 standard) or the matching degree (88 is close to the Level 1 standard) is considered, it is easy to misjudge as a Level 1 warning. Comparison of two parameters ensures the accuracy of the determination.

[0050] Warning commands must contain key information to facilitate rapid response by maintenance personnel; for example, the command content could be: "Warning Level: Level 2; Unit Number: 12; Oil Leak Time: 2025-06-20 14:30:00; Oil Leak Location: Yaw Lower Flange Sealing Surface; Oil Leak Parameters: Area 18cm², Matching Degree 88; Handling Recommendation: On-site inspection within 24 hours"; transmission should be encrypted (e.g., ...). A ES encryption is used to prevent commands from being tampered with or leaked; push notifications are sent to the monitoring terminal (with a pop-up alert window and audio-visual prompts) and the mobile phones of maintenance personnel. A PP (Push notifications + vibration) ensures that relevant personnel receive messages in a timely manner.

[0051] The warning instructions are stored in association with the detection results and image data, forming a complete traceability chain of "instruction-result-image". For example, the instruction is named "Unit 12_20250620_143000_Level 2 Warning", and the corresponding detection result JSON file and optimized image JPEG file are stored in association. The storage path is "20250620 / Unit 12 / Warning Instruction / ". When maintenance personnel query the warning later, they can directly view the original image and detection parameters to analyze the cause of the oil leak.

[0052] Preferably, in step S4, the early warning judgment module calculates the early warning score using the early warning level judgment formula, and determines the early warning level corresponding to the yaw oil leak based on the early warning score, as follows: The formula for determining the warning level is as follows:

[0053] The symbols in the formula are defined as follows: S : Represents the warning score, with a value range of [0,100], used to quantify the severity of the yaw oil leak and serve as the basis for determining the warning level; : Represents the weighting coefficient for the size of the oil leak area, with a value of 0.6, used to characterize the degree of influence of the size of the oil leak area on the warning level; A : Represents the normalized value of the oil leak area size. It is a dimensionless value, ranging from [0, 100]. It is calculated from the actual area of ​​the oil leak area. The calculation method is as follows: ,in The actual oil leak area identified by the lightweight oil leak detection model module. The threshold for the maximum permissible oil leakage area at the yaw seal; : Represents the weighting coefficient of oil leak feature matching degree, with a value of 0.4, used to characterize the degree of influence of oil leak feature matching degree on the warning level; M : Oil leak feature matching degree, with a value range of [0,100], is obtained by converting the oil leak feature matching degree result output by the lightweight oil leak detection model module (multiplying the matching degree value in the [0,1] interval output by the model by 100). The criteria for determining the warning level are: if the warning score is... If it is, it is determined to be a Level 1 warning; if If it is, it is determined to be a Level II warning; if If so, it is determined to be a Level 3 warning.

[0054] The weighting coefficient α (0.6) is greater than β (0.4) because the size of the oil leak area directly reflects the severity of the fault (the larger the area, the higher the risk of the fault), and therefore needs to be given a higher weight. A max (Maximum permissible oil leakage area for yaws) is set based on wind farm operation and maintenance experience, for example, 50cm² (exceeding this area can easily lead to bearing wear), ensuring the normalized result. A The value can accurately reflect the actual percentage of oil leakage; M The value is converted from the [0,1] matching degree output by the detection model (×100) to unify the parameter units and facilitate calculation.

[0055] Qualitative parameters (size, matching degree) are converted into quantitative scores using formulas to avoid subjective judgment bias; for example: 1. Case 1 (Level 1 Warning): A act =40cm² (actual oil leakage area) A max =50cm², calculate A = (40 / 50) × 100 = 80; M = 95 (model output matching degree 0.95 × 100); S = 0.6 × 80 + 0.4 × 95 = 48 + 38 = 86; Since S = 86 ≥ 80, it is judged as a level 1 warning; 2. Case 2 (Level 2 Warning): A act =18cm² A= (18 / 50) × 100 = 36; M =88; S=0.6×36+0.4×88=21.6+35.2=56.8; Since 50≤56.8<80, it is determined to be a level II warning; 3. Case 3 (Level 3 Warning): A act =8cm², A = (8 / 50) × 100 = 16; M = 75; S = 0.6 × 16 + 0.4 × 75 = 9.6 + 30 = 39.6; Since 39.6 < 50, it is determined to be a level three warning; Quantitative scoring makes the determination of warning levels more objective and avoids subjective differences in the understanding of "minor / moderate / severe" among different maintenance personnel.

[0056] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A wind turbine yaw oil leakage early warning system, characterized in that, It includes an image acquisition module, an image preprocessing module, a lightweight oil leak detection model module, an early warning judgment module, and a data storage module; The image acquisition module is located near the yaw of the wind turbine and is used to acquire image data of the yaw's sealing parts and surrounding areas to obtain basic visual information on whether the yaw is leaking oil. The image preprocessing module establishes a data connection with the image acquisition module to process the image data acquired by the image acquisition module and improve image quality. The lightweight oil leak detection model module is built based on an improved YOLOv5 convolutional neural network. It communicates with the image preprocessing module and is used to analyze the image data processed by the image preprocessing module, identify and locate possible oil leak areas in the image, and output oil leak-related detection results. The early warning judgment module is connected to the lightweight oil leak detection model module and has a preset oil leak early warning judgment logic. It is used to determine the early warning level corresponding to the yaw oil leak based on the detection results output by the lightweight oil leak detection model module and generate corresponding early warning information. The data storage module establishes data interaction with the image acquisition module, image preprocessing module, lightweight oil leak detection model module, and early warning judgment module, respectively, to store the image data, processing results, and early warning information generated during the operation of each module.

2. The wind turbine yaw oil leakage early warning system according to claim 1, characterized in that, The image acquisition module uses an industrial-grade image sensor, which has protective performance to adapt to the complex outdoor environment of wind farms. It is installed on the outside of the flange sealing surface of the wind turbine yaw, around the sealing cover of the yaw drive device, and the grease sealing part of the yaw bearing to fully cover the key areas of the yaw that are prone to oil leakage, ensuring that image data of each potential oil leakage point of the yaw can be collected.

3. The wind turbine yaw oil leakage early warning system according to claim 1, characterized in that, The lightweight oil leak detection model module is based on an improved YOLOv5 convolutional neural network, which achieves lightweight optimization by reducing the number of detection heads and one feature pyramid downsampling process. This reduces the computational load of the improved YOLOv5 convolutional neural network while ensuring the oil leak detection effect, thereby improving the response speed of oil leak detection.

4. The wind turbine yaw oil leakage early warning system according to claim 1, characterized in that, The image preprocessing module includes a noise removal submodule and an image enhancement submodule. The noise removal submodule is connected to the image acquisition module and is used to eliminate interference noise in the image acquired by the image acquisition module. The image enhancement submodule is connected to the noise removal submodule and is used to adjust the brightness, contrast and enhance the details of the image after noise removal. The noise removal submodule and the image enhancement submodule work together in sequence to complete the data processing of the original image.

5. A method for early warning of oil leakage in a wind turbine yaw, characterized in that, A wind turbine yaw oil leak early warning system according to any one of claims 1-4 is adopted, comprising the following steps: S1: Image acquisition. The image acquisition module acquires image data of the sealing part and surrounding area of ​​the yaw of the wind turbine. The acquired image data is transmitted to the image preprocessing module and the data storage module respectively. The data storage module saves the original image data. S2: Image preprocessing. The image preprocessing module receives image data transmitted from the image acquisition module, performs preset processing operations on the image data to obtain optimized image data, transmits the optimized image data to the lightweight oil leak detection model module, and simultaneously sends the optimized image data to the data storage module for storage. S3: Oil Leak Detection. The lightweight oil leak detection model module loads pre-trained oil leak recognition parameters, receives optimized image data transmitted from the image preprocessing module, inputs the optimized image data into the lightweight oil leak detection model module loaded with pre-trained oil leak recognition parameters, and uses the lightweight oil leak detection model module to calculate and identify the oil leak area in the image, generating a detection result containing the location and features of the oil leak area. The detection result is then transmitted to the early warning judgment module and simultaneously stored in the data storage module. S4: Warning generation. The warning judgment module receives the detection results transmitted by the lightweight oil leak detection model module, analyzes the detection results according to the preset warning judgment logic, determines the warning level corresponding to the yaw oil leak, generates a warning command containing the warning level and oil leak related information, sends the warning command to the designated terminal, and stores the warning command in the data storage module.

6. The method for early warning of oil leakage in a wind turbine yaw as described in claim 5, characterized in that, In step S1, before acquiring image data, the image acquisition module first senses the ambient light conditions of the acquisition area and adjusts its acquisition parameters according to the sensed light conditions to ensure that clear images of the yaw seal can be acquired under different lighting conditions. During the acquisition process, the image acquisition module continuously acquires image data at a preset acquisition frequency. Each frame of image data acquired is accompanied by identification information of acquisition time and acquisition location. The image data with identification information is then transmitted to the image preprocessing module. At the same time, the image data and identification information are synchronously transmitted to the data storage module in a structured data format. The data storage module classifies and stores the image data according to the acquisition time order.

7. A method for early warning of oil leakage in a wind turbine yaw as described in claim 5, characterized in that, In step S2, after receiving the image data, the image preprocessing module first processes the image data using the noise removal submodule. The filtering algorithm filters out salt-and-pepper noise and Gaussian noise interference signals in the image to obtain denoised image data. Then, the denoised image data is transmitted to the image enhancement submodule. The image enhancement submodule adjusts the brightness distribution of the image to make the overall brightness of the image uniform, while optimizing the image contrast, enhancing the visual difference between the oil leak area and the background in the image, and enhancing the detailed features in the image to highlight the edges of the possible oil leak area. After processing, the image preprocessing module performs quality verification on the optimized image data. If the verification result meets the preset standard, the optimized image data is transmitted to the lightweight oil leak detection model module. If the verification fails, it is returned to the image acquisition module, which triggers the image acquisition module to re-acquire image data of the corresponding area.

8. A method for early warning of oil leakage in a wind turbine yaw as described in claim 5, characterized in that, In step S3, the lightweight oil leak detection model module loads pre-trained oil leak recognition parameters. The training process is based on an image dataset of wind turbine yaw leak scenarios, which includes yaw image samples under different oil leak levels and environmental conditions. During training, Mos... A IC data augmentation technology expands the image samples and adjusts the training parameters of the lightweight oil leak detection model module to improve its accuracy and generalization ability in identifying yaw oil leak areas. During the operation of the lightweight oil leak detection model module, the optimized image data is divided into multiple sub-regions. Features are extracted and analyzed for each sub-region in turn to determine whether there are oil leak features. For sub-regions that are determined to have oil leaks, their coordinate range, region size, and oil leak feature matching degree in the overall image are calculated. The information, including the coordinate range, region size, and oil leak feature matching degree, is integrated to form a complete detection result. The detection result is then transmitted to the early warning judgment module in the form of a data packet. At the same time, the detection result is associated with the corresponding optimized image data and stored in the data storage module.

9. A method for early warning of oil leakage in a wind turbine yaw as described in claim 5, characterized in that, In step S4, after receiving the detection results, the early warning judgment module calls the preset early warning judgment logic, extracts the key parameters of the oil leak area size and oil leak feature matching degree from the detection results, and compares the key parameters with the preset early warning level classification standard. If the size of the oil leak area and the degree of feature matching both meet the criteria for a Level 1 warning, then it is determined to be a Level 1 warning. If the criteria for a Level II warning are met, it will be classified as a Level II warning. If the criteria for a Level 3 warning are met, it will be classified as a Level 3 warning. After determining the warning level, the warning judgment module generates a warning command, which includes the warning level, time of oil leak occurrence, location of oil leak, and parameters of the oil leak area. Subsequently, the warning command is sent to the wind farm monitoring terminal and the mobile terminal of the operation and maintenance personnel using encrypted transmission. After receiving the warning command, the terminal notifies the relevant personnel through sound and light prompts or message push. At the same time, the warning judgment module associates the warning command with the corresponding detection results and image data and stores them in the data storage module for subsequent query and traceability.

10. A method for early warning of oil leakage in a wind turbine yaw as described in claim 9, characterized in that, In step S4, the early warning judgment module calculates the early warning score using the early warning level determination formula, and determines the early warning level corresponding to the yaw oil leak based on the early warning score, as follows: The formula for determining the warning level is as follows: The symbols in the formula are defined as follows: S : Represents the warning score, with a value range of [0,100], used to quantify the severity of the yaw oil leak and serve as the basis for determining the warning level; : Represents the weighting coefficient for the size of the oil leak area, with a value of 0.6, used to characterize the degree of influence of the size of the oil leak area on the warning level; A : Represents the normalized value of the oil leak area size, ranging from [0, 100]. It is calculated from the actual area of ​​the oil leak area, and the calculation method is as follows: ,in The actual oil leak area identified by the lightweight oil leak detection model module. The threshold for the maximum permissible oil leakage area at the yaw seal; : Represents the weighting coefficient of oil leak feature matching degree, with a value of 0.4, used to characterize the degree of influence of oil leak feature matching degree on the warning level; M Oil leak feature matching degree The value range is [0, 100], and it is obtained by converting the oil leak feature matching degree result output by the lightweight oil leak detection model module; The criteria for determining the warning level are: if the warning score is... If it is, it is determined to be a Level 1 warning; if If it is, it is determined to be a Level II warning; if If so, it is determined to be a Level 3 warning.

Images