Fan gear oil wear health condition evaluation method and system

Abstract

The application discloses a kind of fan gear oil wear health condition evaluation method and system, belong to wind power equipment operation safety technical field, solve the problem that different particle types cannot be distinguished in prior art for the wear condition of fan gear oil;Including: based on the feature analysis of particle in oil in fan gear oil image, including the size, quantity and kind of particle are analyzed, the weight of wear particle is calculated, the wear health grade of fan gear oil is evaluated, and operation and maintenance suggestion is output;The application analyzes and identifies the kind, size and quantity of particle in image by fan generator gear oil image, evaluates the wear health state of fan generator gear oil in combination with the morphological characteristics of particle, and then evaluates the operating condition of fan generator gear box, proposes the operation and maintenance suggestion of fan gear oil, provides protection for the safe and stable operation of fan gear box system, and has important significance for ensuring the safe and stable operation of fan.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment operation safety technology, and relates to a method and system for evaluating the wear and health status of wind turbine gear oil. Background Technology

[0002] The gearbox is one of the core components of a wind turbine (referred to as a "wind turbine"). Gearbox failures are frequent in China, with issues such as tooth surface wear, tooth surface scuffing, tooth surface pitting, and bearing damage being commonplace. These failures are closely related to the cleanliness of the gearbox oil. During wind turbine operation, friction occurs on the gear surfaces, accelerating wear. Lubricating oil is the lifeblood of mechanical equipment, primarily serving functions such as lubrication, cooling, rust prevention, cleaning, sealing, and buffering. Lubricating oil forms a continuous oil film on the surfaces of relatively moving parts to reduce friction and wear, thus providing lubrication. It also cools working parts, prevents metal surface corrosion, and washes away wear particles from mechanical parts, carrying them into the oil and preventing particle deposition between friction pairs.

[0003] However, the relative motion of friction pairs in mechanical equipment leads to wear. While lubricating oil plays a lubricating role, it also acts as a carrier of wear particles and external contaminants. These components in the lubricating oil often reflect the actual operating condition of the equipment. Gearboxes operate under complex conditions for extended periods, and pitting, fatigue spalling, and wear occur on the gear surfaces and bearings, generating metal microparticles that dissolve into the gearbox oil. Due to the viscosity, pressure, and flow characteristics of the gearbox oil, these metal microparticles cannot settle to the bottom of the tank quickly. Instead, they circulate countless times with the oil flow. During this circulation, the metal microparticles further wear down the working surfaces of the gears and bearings, creating a vicious cycle between the oil and components. This accelerates the damage to gears and bearings, as well as the deterioration of the oil, shortening their service life. Ultimately, this results in a mutually destructive situation where both suffer damage.

[0004] Existing technologies primarily evaluate the health status of wind turbine gear oils based on factors such as particle size and metal content. For example, invention patent CN120123921A discloses a diagnostic method for the degradation of wind turbine main gear oil, diagnosing its degradation based on factors such as kinematic viscosity, acid value, wear elements, and leaching elements from additives. Wear in wind turbine gearboxes is often the result of multiple mechanisms working together. However, existing diagnostic models are mostly based on feature databases of single wear types, making it difficult to diagnose wear states under coupled scenarios. Furthermore, they often focus on factors such as particle size and quantity, lacking analysis of key characteristics such as particle composition and morphology, which is detrimental to evaluating the operational status of the wind turbine gear system. Therefore, considering the impact of particle type on the wear health status of wind turbine gear oil is crucial for ensuring the safe and stable operation of wind turbine gearbox systems. Summary of the Invention

[0005] The technical problem to be solved by this invention is that the existing technology has difficulty in distinguishing the wear conditions of different particle types on the wind turbine gear oil, and it is difficult to ensure the safe and stable operation of the wind turbine gearbox system.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions: A method for evaluating the wear and health status of wind turbine gear oil includes the following steps: S1, Based on the image of the wind turbine gear oil, perform feature analysis on the particulate matter in the oil, including analysis of the size, quantity and type of the particulate matter; S2, calculate the wear particle weight based on the size, quantity and type of particles; S3 assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs operation and maintenance suggestions.

[0007] Further, S1 includes the following steps: S11, calculate the size of each particle in the figure, using the following logical representation to calculate the size of a single particle:

[0008] In the formula, P The value represents the size of a single particle, in µm; max{X2-X1, Y1-Y2} represents the maximum number of pixels in the particle, and (X1, Y1) and (X2, Y2) represent the coordinates of the top left and bottom right corners of the minimum bounding rectangle of a single particle, respectively. a This indicates the camera pixel size, in µm. b Indicates magnification; S12, Calculate the number of particulate matter in the graph, using the following logical representation:

[0009] In the formula, c Indicates the number of particles within a preset size range in the tested oil sample; Indicates the number of cameras within a certain size range i The average number of particles captured per frame; Indicates the first i The volume of the oil sample being tested in the frame, in mL; t Indicates the collection time, in units of s ; F Indicates the camera frame rate. Ft Indicates the total number of frames. ; S13, based on the improved target detection algorithm, the types of particulate matter are identified, including normal wear particles, fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, non-metallic particles, fibers, bubbles, and water droplets.

[0010] Furthermore, the particle identification process in S13 specifically involves: First, an improved target detection algorithm is used to detect the target of interest in each frame, and the location coordinates, classification and confidence of the target are obtained as the detection result. The number of detected targets is denoted as M1. Secondly, associate the detection results of the previous step with the detection results of the target of interest detected in the previous frame one by one, and record the number of targets detected in the previous frame as M2. Find the most similar pair among M1×M2 pairs. Finally, the tracked wear particle images are combined with the target detection algorithm for identification, and the types of particles are output. Combined with steps S11~S12, other feature parameters of the wear particles are extracted to complete the statistics of wear particles.

[0011] Furthermore, the improved target detection algorithm includes the following: The image acquisition step specifically involves acquiring real-time images of the wind turbine gear oil using sensors and outputting raw image data at a preset frame rate. The image data processing steps specifically include grayscale correction, image scaling, normalization, and tensor format conversion of the original image data. The model inference step specifically involves loading the SAHI-optimized YOLO11 model for detection, and outputting the particle candidate boxes, class labels, and confidence information for each frame of the image. The post-processing steps specifically involve filtering out low-confidence particle candidate boxes, removing duplicate detection boxes using an IoU threshold, converting the particle size of the pixel image to the actual physical size, and outputting the particle number and category distribution results.

[0012] Furthermore, the YOLO11 model optimized by SAHI includes a backbone network, a neck network, and a head network; The backbone network splits the input feature map along the channel dimension into two parts: a Shortcut branch and a feature extraction branch; the Shortcut branch retains... The feature map of the channel transmits shallow details such as grain edges and textures, avoiding detail loss during feature transfer; the feature extraction branch will then extract the remaining... The feature maps of the channels are used for depth feature extraction through variable convolution kernels. By adjusting the convolution kernel size k and the number of branch repetitions n, the feature extraction requirements of different scale particles can be adaptively matched. The neck network is based on an improved PANet structure. It achieves efficient fusion of shallow detail features and deep semantic features through bidirectional feature transfer and the introduction of scale-adaptive fusion weights. The head network employs a decoupled detection head design, receiving the fused feature map output by the neck network. The regression and classification branches are set in parallel. The two branches share the basic convolutional layer but are optimized independently to improve the localization accuracy of particle bounding boxes and the accuracy of category judgment.

[0013] Furthermore, the normal wear particles mentioned in S13 include scaly, elongated, and flat wear particles; wherein, all normal wear particles have smooth surfaces; The fatigue wear particles include spalling block-shaped, spherical, and flake-shaped wear particles; wherein, the surface of the spalling block-shaped wear particles is smooth and the edges are irregular, the surface of the spherical wear particles is smooth, and the surface of the flake-shaped wear particles has holes. The sliding wear particles include ordinary sliding wear particles and severe sliding wear particles; wherein, the surface of ordinary sliding wear particles is smooth, while the surface of severe sliding wear particles has obvious scratches and cracks, and both ordinary sliding wear particles and severe sliding wear particles have irregular shapes and straight edges; The cutting and abrasion particles include curved, spiral, annular, and strip-shaped abrasion particles; The non-ferrous metal particles include copper alloy and Fe2O3 oxide particles; wherein the copper alloy particles are yellow, the Fe2O3 oxide particles are red, and they have the characteristics of rust particles. The non-metallic particles are transparent particles; The fibers have a linear structure; The bubble has a circular structure; the center of the bubble has a small bright spot, while the overall brightness of the outer periphery of the bubble is relatively dark. The water droplet has a circular structure; the center of the water droplet has a large bright spot.

[0014] Furthermore, the calculation of wear particle weights in S2 specifically involves: S21, a first type of wear particles, a second type of wear particles, and a third type of wear particles are preset; the first type of wear particles includes spalling block wear particles, sliding wear particles, cutting wear particles, Fe2O3 oxide, copper alloy metal particles and non-metal particles; the second type of wear particles includes non-ferrous metal wear particles and severe sliding wear particles; and the third type of wear particles includes spherical wear particles. S22, traverse all particles in the wind turbine gear oil image, count the number and average size of different types of particles, and calculate the corresponding wear particle weights.

[0015] Furthermore, the wear particle weights are represented by the following logic: When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first jWeight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0016] Furthermore, the assessment of the wear health level of the fan gear oil based on wear particle weight in S3 specifically involves: First, the wear health level of the fan gear oil is calculated using the following logic:

[0017] In the formula, G Indicates the wear and tear health level. Indicates the first j Weight of each type of wear particle , This indicates the types and quantities of particulate matter present in the tested oil sample. Secondly, determine the wear health level of the fan gear oil. Specifically, when the wear health level... At that time, the wear of the fan gear oil was assessed as normal; when the wear health level was... When the wear is assessed as minor, the wear is rated as healthy. When the wear level is normal, it is assessed as abnormal; when the wear health level is normal... At that time, it was assessed as severely worn.

[0018] This invention also provides a wind turbine gear oil wear health status evaluation system, comprising: The feature analysis module performs feature analysis on particulate matter in the wind turbine gear oil image, including analysis of the size, quantity, and type of particulate matter. The weighting analysis module calculates the weight of wear particles based on their size, quantity, and type. The rating module assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs maintenance suggestions.

[0019] The advantages of this invention are: (1) This invention provides a method for evaluating the wear and health status of wind turbine generator gear oil. By analyzing and identifying the type, size and quantity of particles in the wind turbine generator gear oil image, the wear and health status of the wind turbine generator gear oil is evaluated in combination with the morphological characteristics of the particles. The method takes into account all factors, and then evaluates the operating status of the wind turbine generator gearbox, and puts forward maintenance suggestions for the wind turbine gear oil. This provides a guarantee for the safe and stable operation of the wind turbine gearbox system and is of great significance for ensuring the safe and stable operation of the wind turbine.

[0020] (2) This invention identifies particulate matter types based on the YOLO11 model optimized by loading SAHI, wherein the backbone network retains the type of particulate matter through the Shortcut branch. The feature map of the channel conveys shallow details such as grain edges and textures, avoiding the loss of details during feature transfer. The feature extraction branch will then extract the remaining features. The feature maps of the channels are used for deep feature extraction via variable convolution kernels. By flexibly adjusting the kernel size k and the number of branch repetitions n, the feature extraction requirements of particles at different scales can be adaptively matched. The neck network is based on an improved PANet structure. Through bidirectional feature transfer and the introduction of scale-adaptive fusion weights, it achieves efficient fusion of shallow detail features and deep semantic features, which is used to solve the problems of insufficient feature fusion of multi-scale particles and difficulty in distinguishing densely adhered particles in the wind turbine gear oil scenario. The head network adopts a decoupled detection head design to receive the fused feature maps output by the neck network. By setting up regression and classification branches in parallel, and optimizing the two branches independently while sharing the basic convolutional layer, the accuracy of particle bounding box localization and category judgment can be effectively improved. Attached Figure Description

[0021] Figure 1This is a flowchart of a method for evaluating the wear and health status of wind turbine gear oil according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of typical examples of different particulate matter types in the wind turbine gear oil of Embodiment 1 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 1 As shown, specifically, a method for evaluating the wear and health status of wind turbine gear oil is disclosed, including the following steps: S1, based on the image of the wind turbine gear oil, perform feature analysis on the particulate matter in the oil, including analysis of the size, quantity, and type of the particles. Specifically, S1 includes the following steps: S11, calculate the size of each particle in the figure, using the following logical representation to calculate the size of a single particle:

[0024] In the formula, P The size of a single particle is expressed in µm. In this embodiment, a microscopic image of the gear oil sample of the tested fan (the oil sample to be tested) is first acquired using an industrial camera. On the camera target surface, the smallest bounding rectangle of a single particle is selected parallel to the horizontal and vertical axes. The coordinates of the upper left and lower right corners of the rectangle are recorded as (X1, Y1) and (X2, Y2) respectively. Then, max{X2-X1, Y1-Y2} represents the maximum number of pixels of the particle. a This indicates the camera pixel size, in µm. b Indicates the magnification factor.

[0025] S12, Calculate the number of particulate matter in the graph, using the following logical representation:

[0026] In the formula, N This indicates the number of particles in the tested oil sample within a preset size range. The preset size range is segmented according to particle size, and the number of particles in different size ranges is counted as the number of particles per 100 mL. Indicates the number of cameras within a certain size range i The average number of particles captured per frame; Indicates the first i The volume of the oil sample being tested in the frame, in mL; t Indicates the collection time, in units of s ; F Indicates the camera frame rate. Ft Indicates the total number of frames. .

[0027] S13, Based on the improved target detection algorithm, the types of particulate matter are identified. The types of particulate matter include normal wear particles, fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, non-metallic particles, fibers, bubbles, and water droplets.

[0028] In step S13, the step of identifying the type of particulate matter specifically involves: First, an improved target detection algorithm is used to detect the target of interest in each frame, and the location coordinates, classification and confidence level of the target are obtained as the detection result. The number of detected targets is denoted as M1. Secondly, associate the detection results of the previous step with the detection results of the target of interest detected in the previous frame one by one, and record the number of targets detected in the previous frame as M2. Find the most similar pair among M1×M2 pairs. Finally, the tracked wear particle images are combined with the target detection algorithm for identification, and the types of particles are output. Combined with steps S11~S12, other feature parameters of the wear particles are extracted to complete the statistics of wear particles.

[0029] In this embodiment, the improved target detection algorithm is described, including the following: The image acquisition step specifically involves acquiring real-time images of the wind turbine gear oil using a sensor and outputting the raw image data at a preset frame rate. In this embodiment, the resolution of the wind turbine gear oil image is set to 1280×720, and the preset frame rate is set to 60fps.

[0030] The image data processing steps specifically involve grayscale correction, image scaling, normalization, and tensor format conversion of the original image data. Grayscale correction is used to eliminate interference from oil reflections; image scaling is used to adapt the original image to the input size of 800×800 for the target detection algorithm; normalization is used to map the image pixel values ​​to [0,1]; and tensor format conversion is used to complete the format conversion from NHWC to NCHW through the data preprocessing interface provided by SAHI, providing data support for model inference in subsequent steps.

[0031] The model inference step specifically involves loading the SAHI-optimized YOLO11 model for detection, and outputting the particle candidate boxes, class labels, and confidence information for each frame of the image.

[0032] The post-processing steps specifically involve filtering out low-confidence particle candidate boxes, removing duplicate detection boxes using an IoU threshold, converting the particle size of the pixel image to the actual physical size, and outputting the particle number and category distribution results. In this embodiment, the low confidence level is set to below 0.2, and the IoU threshold is set to 0.3.

[0033] Furthermore, the SAHI-optimized YOLO11 model is described, which includes a backbone network, a neck network, and a head network. In this embodiment, hardware resources are initialized through the SAHI inference engine, and preprocessed tensor data is input into the inference engine. SAHI accelerates inference through operator fusion and dynamic scheduling mechanisms, and performs particle bounding box localization, category judgment, and confidence calculation in parallel, outputting particle candidate boxes, category labels, and confidence information for each frame of image. The model described in this embodiment can be deployed on the edge (ARM+NPU) or in the cloud (GPU) with a latency of ≤15ms / frame.

[0034] In a preferred embodiment, the backbone network splits the input feature map along the channel dimension into two parts: a Shortcut branch and a feature extraction branch; wherein, the backbone network input feature map is defined. H and W represent the height and width of the feature map, respectively. The number of input channels. The Shortcut branch retains... The feature map of the channel transmits shallow details such as grain edges and textures, avoiding detail loss during feature transfer; the feature extraction branch will then extract the remaining... The feature maps of the channels are used for depth feature extraction. By flexibly adjusting the kernel size k and the number of branch repetitions n, the receptive field of the module is increased. The following logic must be satisfied:

[0035] in, Based on the sensory field, The step size for the first i layers is 1 by default. The module can adaptively match the feature extraction requirements of particles of different sizes. When k=3 and n=1, the receptive field focuses on tiny particles of 5-20μm; when k=7 and n=3, the receptive field extends to larger pollutant particles of 50-100μm, effectively solving the problem of large particle size range in wind turbine gear oil. In this embodiment, k is 3~7 and n is 1~3.

[0036] The output feature map dimension of the backbone network Using the following logical representation:

[0037] in, This represents a variable convolution with kernel size k, number of groups g, and channel expansion rate e. For batch normalization, For activation function, This represents the feature map of the feature extraction branch after splitting; shallow details and deep semantic features are fused through cross-layer connections, and the final output is a fused feature map. , The feature map representing the Shortcut branch significantly improves the ability to capture features of tiny wear particles.

[0038] In a preferred embodiment, the neck network is based on an improved structure of PANet. Through bidirectional feature transfer and the introduction of scale-adaptive fusion weights, it achieves efficient fusion of shallow detail features and deep semantic features, which is used to solve the problems of insufficient fusion of multi-scale particle features and difficulty in distinguishing densely adhered particles in the wind turbine gear oil scenario.

[0039] The bidirectional feature transfer specifically involves the neck network receiving three scale feature maps output by the backbone network. , , Corresponding to large, medium, and small-scale particle features, bidirectional feature fusion pathways are constructed, one from top to bottom and the other from bottom to top. Specifically, the top-down fusion pathway consists of: feature maps... After upsampling and feature map A new feature map is obtained by concatenating and fusing channels, followed by optimization using a convolutional module. ,right After further upsampling and feature map The fusion yields a shallow feature map. This process injects the granular semantic features of deep networks (such as the overall shape of particles and category association information) into shallow features to improve the semantic recognition of small particles (5-20μm); in this embodiment, the upsampling scaling factor is set to 2.

[0040] The top-down fusion pathway specifically refers to: shallow feature maps. After downsampling and splicing and fusion ,right After further downsampling and Fusion yields deep feature maps This process transfers the detailed features of the shallow network (such as edge contours and texture differences) to the deeper layers, helping to distinguish the boundaries of densely clustered particles; in this embodiment, the downsampling step size is set to 2.

[0041] The introduction of scale-adaptive fusion weights specifically involves setting dynamic weights for fusion branches of different scales, taking into account the large differences in particle size of wind turbine gear oil. , The weights are obtained statistically from the particle size distribution, and the final output is a fused feature map. Using the following logical representation:

[0042] In this embodiment, when detecting tiny particles, Increased; when detecting large or dense particles. Improvements are made to ensure the targeted fusion of features from particles of different scales.

[0043] In a preferred embodiment, the head network employs a decoupled detection head design to receive the fused feature map output by the neck network. The regression and classification branches are set in parallel. The two branches share the basic convolutional layer but are optimized independently to improve the localization accuracy of particle bounding boxes and the accuracy of category judgment.

[0044] The regression branch extracts localization features through a three-layer convolutional module, ultimately outputting bounding box parameters (x, y, w, h). To adapt to the size distribution of lubricating oil particles, the regression branch introduces dynamic anchor point optimization, specifically: based on the statistical distribution of width and height of particles in the training set (w... p， h p The initial anchor points are obtained through K-means clustering. During inference, the anchor points are adaptively adjusted based on the real-time size of the particles in the input image. :

[0045] in, This is the scale adaptation coefficient, which is set to 0.8~1.2 in this embodiment. These represent the average width and height of the particles in the current frame, respectively. These represent the average width and height of the initial anchor point, respectively, used to ensure the matching degree between the anchor point and the particle size, and improve the positioning recall rate of small and irregularly shaped particles.

[0046] The classification branch, through the attention mechanism module (CBAM), strengthens the weights of particulate material-related features (such as grayscale mean and reflectivity), independently targeting metal and non-metal classification scenarios such as cutting wear, fatigue wear, sliding wear, bubbles, and water droplets, outputting multiple category probabilities. The attention weight calculation is represented by the following logic:

[0047]

[0048] in, , These are channel attention and spatial attention, respectively. Input the feature map of the attention module, For the Sigmoid function, By multiplying elements, the difference between the high reflectivity of metallic particles and the texture of non-metallic particles is highlighted through an attention mechanism.

[0049] The head network is optimized using a loss function to balance localization accuracy and class discrimination capability. This loss function includes regression loss and classification loss. Specifically, the regression loss uses CIoU, which combines bounding box overlap IoU, center point distance d, and aspect ratio. The regression loss can be represented using the following logic:

[0050] Where c is the length of the diagonal of the smallest bounding rectangle of the two frames. , ,in Using the true bounding box width and height, regression loss can accurately penalize positioning deviations and improve the accuracy of bounding boxes for tiny particles.

[0051] The classification loss adopts Focal Loss. To address the imbalance problem of particle samples in the lubricating oil scenario, the classification loss is represented by the following logic:

[0052] in, To predict class probabilities, To set the focusing parameters, take By reducing the weight of easily classified samples (such as clear large particles) and increasing the training weight of difficult-to-classify samples (such as blurry small particles and edge-adhered particles), the robustness of class prediction is improved.

[0053] The head network outputs the bounding box coordinates and confidence scores of the particles in each frame of the image. The category probability provides accurate basic data for subsequent post-processing and particle analysis.

[0054] In this embodiment, normal wear of gear oil during mechanical break-in is considered. The normal wear particles include scaly, elongated, and flat wear particles; wherein, all normal wear particles have smooth surfaces.

[0055] Furthermore, the particle size range of the normal wear particles is 0.5μm to 40μm.

[0056] like Figure 2 These are typical images of different types of particulate matter in wind turbine gear oil, among which... Figure 2 (a) represents an example of fatigue wear particles, which include spalling blocky, spherical and flake-shaped wear particles; wherein the spalling blocky wear particles have a smooth surface and irregular edges, the spherical wear particles have a smooth surface, and the flake-shaped wear particles have pores on their surface.

[0057] Furthermore, the size of the detached blocky wear particles is greater than 10 μm, the diameter of the spherical wear particles is 1 μm to 15 μm, and the length of the sheet-like wear particles is greater than 20 μm.

[0058] Figure 2 (b) represents an example of sliding wear particles, which include ordinary sliding wear particles and severe sliding wear particles; wherein, the surface of ordinary sliding wear particles is smooth, the surface of severe sliding wear particles has obvious scratches and cracks, and both ordinary sliding wear particles and severe sliding wear particles are irregular in shape, with straight edges, and the tempering color is brown or blue.

[0059] Furthermore, the size of the ordinary sliding wear particles ranges from several micrometers to hundreds of micrometers, while the size of the severe sliding wear particles is greater than 20 μm.

[0060] Figure 2 (c) represents an example of cutting wear particles, which include curved, spiral, annular, and strip-shaped wear particles; wherein the cutting wear particles generated by the sharp edge of the part penetrating a softer sliding surface are generally larger in size, while the cutting wear particles generated when there are hard inclusions or abrasives between the lubricated surfaces are generally smaller and linear in size.

[0061] Furthermore, the width of the cutting and abrasive particles is 2μm to 10μm, and the length is greater than 25μm. When hard inclusions or abrasives are present, the length of the cutting and abrasive particles is greater than 5μm.

[0062] Figure 2(d) represents an example where the non-ferrous metal particles are copper metal, the non-ferrous metal particles comprising copper alloy and Fe2O3 oxide particles; wherein the non-ferrous metal particles are identified by color, the copper alloy particles are yellow, the Fe2O3 oxide particles are red, and they have the characteristics of rust particles.

[0063] Figure 2 (e) represents an example of non-metallic particles, which are transparent particles; non-metallic particles are different from non-ferrous metal particles, and do not emit light or shine like non-ferrous metal particles, and are more transparent.

[0064] Figure 2 (f) represents an example of a fiber that is linear in structure.

[0065] Furthermore, the fibers can be long, thin, and partially transparent linear structures.

[0066] Figure 2 (g) represents an example of a bubble, which has a circular structure; wherein the center of the bubble has a small bright spot, and the overall brightness of the outer periphery of the bubble is relatively dark.

[0067] Figure 2 (h) represents an example of a water droplet, which is circular in shape; wherein the center of the water droplet has a large bright spot.

[0068] S2, the wear particle weights are calculated based on the size, quantity, and type of the particles, as shown in Table 1 below: Table 1. Weighting of Wear Particles in Oil Image Analysis

[0069] As shown in Table 1, the calculation of wear particle weights is specifically as follows: S21, pre-defined first type of wear particles, second type of wear particles, and third type of wear particles; the first type of wear particles includes spalling block wear particles, sliding wear particles, cutting wear particles, Fe2O3 oxide, copper alloy metal particles, and non-metallic particles; the second type of wear particles includes non-ferrous metal wear particles and severe sliding wear particles; and the third type of wear particles includes spherical wear particles.

[0070] S22, Traverse all particles in the wind turbine gear oil image, count the number and average size of different types of particles, and calculate the corresponding wear particle weights. The wear particle weights are represented by the following logic: When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0071] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0072] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0073] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0074] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0075] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0076] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0077] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0078] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0079] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0080] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0081] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0082] Furthermore, the wear particles of normal wear particles, fibers, bubbles, and water droplets have a weight of 0 and are not included in the wear health level assessment; while the average size d of flaking block wear particles, sliding wear particles, cutting wear particles, and non-metallic particles is less than 20μm, which are considered normal wear particles, and similarly have a wear particle weight of 0 and are not included in the wear health level assessment.

[0083] In this embodiment, the average size is calculated using the following logic:

[0084] In the formula, d represents the average size of the particles, and N represents the number of particles. Let u represent the size of the u-th particle. .

[0085] The remaining fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, and non-metallic particles are assigned corresponding wear particle weights based on the type, quantity, and average size d of the particles. The wear health level of the wind turbine gearbox oil is then assessed based on the sum of the types of particles in the wind turbine gear oil and the corresponding wear particle weights.

[0086] S3 assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs operation and maintenance suggestions.

[0087] In step S3, the assessment of the wear health level of the fan gear oil based on wear particle weight specifically involves: First, the wear health level of the fan gear oil is calculated using the following logic:

[0088] In the formula, G Indicates the wear and tear health level. Indicates the first j Weight of each type of wear particle , This indicates the types and quantities of particulate matter present in the tested oil sample.

[0089] Secondly, determine the wear health level of the fan gear oil. Specifically, when the wear health level... At that time, the wear of the fan gear oil was assessed as normal; when the wear health level was... When the wear is assessed as minor, the wear is rated as healthy. When the wear level is normal, it is assessed as abnormal; when the wear health level is normal... At that time, it was assessed as severely worn.

[0090] In this embodiment, when the health status of the wind turbine gear oil is assessed as normal wear, it means that the type, quantity, and size of wear particles in the oil are all within the normal range, the wind turbine gear oil lubrication is normal, there is no abnormal wear feedforward, and no maintenance recommendations are needed.

[0091] When the health status of the wind turbine gear oil is assessed as slightly worn, it indicates that the type, quantity, and size of wear particles in the oil are slightly outside the normal range, and slight wear has occurred at the lubrication points. It is necessary to monitor changes in wear particles in the oil and noise levels. When the health status of the wind turbine gear oil is assessed as normal worn, it indicates that the type, quantity, and size of wear particles in the oil are all within the normal range, the wind turbine gear oil lubrication is normal, there is no abnormal wear feedforward, and no maintenance recommendations are needed.

[0092] When the health status of the fan gear oil is assessed as abnormal wear, it indicates that the type, quantity, and size of wear particles in the oil are within an abnormal range, and abnormal wear occurs at the lubrication points. It is recommended to replace the fan gear oil and check the condition of the fan gearbox gears during maintenance.

[0093] When the health status of the wind turbine gear oil is assessed as severely worn, it indicates that the type, quantity, and size of wear particles in the oil are abnormal, abnormal wear occurs at the lubrication points, and there is a risk of lubrication failure in the equipment components. It is recommended to replace the wind turbine gear oil immediately, arrange maintenance in a timely manner, check the gear condition during maintenance, shorten the operating cycle after commissioning, and conduct regular follow-up tests.

[0094] This invention also provides an evaluation system applying the above-mentioned method for evaluating the wear and health status of wind turbine gear oil, comprising: The feature analysis module performs feature analysis on particulate matter in the wind turbine gear oil image, including analysis of the size, quantity, and type of particulate matter; specifically, the feature analysis module includes: The size analysis unit is used to calculate the size of each particle in the image, using the following logic to calculate the size of a single particle:

[0095] In the formula, P The size of a single particle is expressed in µm. In this embodiment, a microscopic image of the gear oil sample of the tested fan (the oil sample to be tested) is first acquired using an industrial camera. On the camera target surface, the smallest bounding rectangle of a single particle is selected parallel to the horizontal and vertical axes. The coordinates of the upper left and lower right corners of the rectangle are recorded as (X1, Y1) and (X2, Y2) respectively. Then, max{X2-X1, Y1-Y2} represents the maximum number of pixels of the particle. a This indicates the camera pixel size, in µm. b Indicates the magnification factor.

[0096] The quantity analysis unit is used to calculate the number of particles in the graph, using the following logical representation:

[0097] In the formula, N This indicates the number of particles in the tested oil sample within a preset size range. The preset size range is segmented according to particle size, and the number of particles in different size ranges is counted as the number of particles per 100 mL. Indicates the number of cameras within a certain size range i The average number of particles captured per frame; Indicates the first i The volume of the oil sample being tested in the frame, in mL; t Indicates the collection time, in units of s ; F Indicates the camera frame rate. Ft Indicates the total number of frames. .

[0098] The type identification unit identifies the types of particulate matter based on an improved target detection algorithm. The types of particulate matter include normal wear particles, fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, non-metallic particles, fibers, bubbles, and water droplets.

[0099] In the category identification unit, the step of identifying the category of particulate matter specifically includes: First, an improved target detection algorithm is used to detect the target of interest in each frame, and the location coordinates, classification and confidence level of the target are obtained as the detection result. The number of detected targets is denoted as M1. Secondly, associate the detection results of the previous step with the detection results of the target of interest detected in the previous frame one by one, and record the number of targets detected in the previous frame as M2. Find the most similar pair among M1×M2 pairs. Finally, the tracked wear particle images are combined with the target detection algorithm for identification, and the type of particles is output. Other feature parameters of the wear particles are extracted by combining the size analysis unit and the quantity analysis unit to complete the statistics of wear particles.

[0100] In this embodiment, normal wear of gear oil during mechanical break-in is considered. The normal wear particles include scaly, elongated, and flat wear particles; wherein, all normal wear particles have smooth surfaces.

[0101] Furthermore, the particle size range of the normal wear particles is 0.5μm to 40μm.

[0102] like Figure 2 These are typical images of different types of particulate matter in wind turbine gear oil, among which... Figure 2 (a) represents an example of fatigue wear particles, which include spalling blocky, spherical and flake-shaped wear particles; wherein the spalling blocky wear particles have a smooth surface and irregular edges, the spherical wear particles have a smooth surface, and the flake-shaped wear particles have pores on their surface.

[0103] Furthermore, the size of the detached blocky wear particles is greater than 10 μm, the diameter of the spherical wear particles is 1 μm to 15 μm, and the length of the sheet-like wear particles is greater than 20 μm.

[0104] Figure 2 (b) represents an example of sliding wear particles, which include ordinary sliding wear particles and severe sliding wear particles; wherein, the surface of ordinary sliding wear particles is smooth, the surface of severe sliding wear particles has obvious scratches and cracks, and both ordinary sliding wear particles and severe sliding wear particles are irregular in shape, with straight edges, and the tempering color is brown or blue.

[0105] Furthermore, the size of the ordinary sliding wear particles ranges from several micrometers to hundreds of micrometers, while the size of the severe sliding wear particles is greater than 20 μm.

[0106] Figure 2 (c) represents an example of cutting wear particles, which include curved, spiral, annular, and strip-shaped wear particles; wherein the cutting wear particles generated by the sharp edge of the part penetrating a softer sliding surface are generally larger in size, while the cutting wear particles generated when there are hard inclusions or abrasives between the lubricated surfaces are generally smaller and linear in size.

[0107] Furthermore, the width of the cutting and abrasive particles is 2μm to 10μm, and the length is greater than 25μm. When there are hard inclusions or abrasives, the length of the cutting and abrasive particles is greater than 5μm.

[0108] Figure 2 (d) represents an example where the non-ferrous metal particles are copper metal, the non-ferrous metal particles comprising copper alloy and Fe2O3 oxide particles; wherein the non-ferrous metal particles are identified by color, the copper alloy particles are yellow, the Fe2O3 oxide particles are red, and they have the characteristics of rust particles.

[0109] Figure 2 (e) represents an example of non-metallic particles, which are transparent particles; non-metallic particles are different from non-ferrous metal particles, and do not emit light or shine like non-ferrous metal particles, and are more transparent.

[0110] Figure 2 (f) represents an example of a fiber that is linear in structure.

[0111] Furthermore, the fibers can be long, thin, and partially transparent linear structures.

[0112] Figure 2 (g) represents an example of a bubble, which has a circular structure; wherein the center of the bubble has a small bright spot, and the overall brightness of the outer periphery of the bubble is relatively dark.

[0113] Figure 2 (h) represents an example of a water droplet, which is circular in shape; wherein the center of the water droplet has a large bright spot.

[0114] The weighting analysis module calculates the weight of wear particles based on their size, quantity, and type; specifically, the calculation of wear particle weights involves: The type classification unit pre-defines three types of wear particles: a first type, a second type, and a third type. The first type of wear particles includes spalling block wear particles, sliding wear particles, cutting wear particles, Fe2O3 oxide, copper alloy metal particles, and non-metallic particles. The second type of wear particles includes non-ferrous metal wear particles and severe sliding wear particles. The third type of wear particles includes spherical wear particles.

[0115] The weight calculation unit traverses all particles in the wind turbine gear oil image, counts the number and average size of different types of particles, and calculates the corresponding wear particle weights. The wear particle weights are represented by the following logic: When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0116] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0117] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0118] When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0119] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0120] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0121] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0122] When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0123] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0124] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0125] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0126] When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

[0127] Furthermore, the wear particles of normal wear particles, fibers, bubbles, and water droplets have a weight of 0 and are not included in the wear health level assessment; while the average size d of flaking block wear particles, sliding wear particles, cutting wear particles, and non-metallic particles is less than 20μm, which are considered normal wear particles, and similarly have a wear particle weight of 0 and are not included in the wear health level assessment.

[0128] In this embodiment, the average size is calculated using the following logic:

[0129] In the formula, d represents the average size of the particles, and N represents the number of particles. Let u represent the size of the u-th particle. .

[0130] The remaining fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, and non-metallic particles are assigned corresponding wear particle weights based on the type, quantity, and average size d of the particles. The wear health level of the wind turbine gearbox oil is then assessed based on the sum of the types of particles in the wind turbine gear oil and the corresponding wear particle weights.

[0131] The rating module assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs maintenance suggestions.

[0132] The specific method for assessing the wear health level of fan gear oil based on wear particle weight is as follows: First, the wear health level of the fan gear oil is calculated using the following logic:

[0133] In the formula, G Indicates the wear and tear health level. Indicates the first j Weight of each type of wear particle , This indicates the types and quantities of particulate matter present in the tested oil sample.

[0134] Secondly, determine the wear health level of the fan gear oil. Specifically, when the wear health level... At that time, the wear of the fan gear oil was assessed as normal; when the wear health level was... When the wear is assessed as minor, the wear is rated as healthy. When the wear level is normal, it is assessed as abnormal; when the wear health level is normal... At that time, it was assessed as severely worn; In this embodiment, when the health status of the wind turbine gear oil is assessed as normal wear, it means that the type, quantity, and size of wear particles in the oil are all within the normal range, the wind turbine gear oil lubrication is normal, there is no abnormal wear feedforward, and no maintenance recommendations are needed.

[0135] When the health status of the wind turbine gear oil is assessed as slightly worn, it indicates that the type, quantity, and size of wear particles in the oil are slightly outside the normal range, and slight wear has occurred at the lubrication points. It is necessary to monitor changes in wear particles in the oil and noise levels. When the health status of the wind turbine gear oil is assessed as normal worn, it indicates that the type, quantity, and size of wear particles in the oil are all within the normal range, the wind turbine gear oil lubrication is normal, there is no abnormal wear feedforward, and no maintenance recommendations are needed.

[0136] When the health status of the fan gear oil is assessed as abnormal wear, it indicates that the type, quantity, and size of wear particles in the oil are within an abnormal range, and abnormal wear occurs at the lubrication points. It is recommended to replace the fan gear oil and check the condition of the fan gearbox gears during maintenance.

[0137] When the health status of the wind turbine gear oil is assessed as severely worn, it indicates that the type, quantity, and size of wear particles in the oil are abnormal, abnormal wear occurs at the lubrication points, and there is a risk of lubrication failure in the equipment components. It is recommended to replace the wind turbine gear oil immediately, arrange maintenance in a timely manner, check the gear condition during maintenance, shorten the operating cycle after commissioning, and conduct regular follow-up tests.

[0138] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the wear and health status of wind turbine gear oil, characterized in that, Includes the following steps: S1, Based on the image of the wind turbine gear oil, perform feature analysis on the particulate matter in the oil, including analysis of the size, quantity and type of the particulate matter; S2, calculate the wear particle weight based on the size, quantity and type of particles; S3 assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs operation and maintenance suggestions.

2. The method for evaluating the wear and health status of wind turbine gear oil according to claim 1, characterized in that, S1 includes the following steps: S11, calculate the size of each particle in the figure, using the following logical representation to calculate the size of a single particle: In the formula, P The value represents the size of a single particle, in µm; max{X2-X1, Y1-Y2} represents the maximum number of pixels in the particle, and (X1, Y1) and (X2, Y2) represent the coordinates of the top left and bottom right corners of the minimum bounding rectangle of a single particle, respectively. a This indicates the camera pixel size, in µm. b Indicates magnification; S12, Calculate the number of particulate matter in the graph, using the following logical representation: In the formula, c Indicates the number of particles within a preset size range in the tested oil sample; Indicates the number of cameras within a certain size range i The average number of particles captured per frame; Indicates the first i The volume of the oil sample being tested in the frame, in mL; t Indicates the collection time, in units of s ; F Indicates the camera frame rate. Ft Indicates the total number of frames. ; S13, based on the improved target detection algorithm, the types of particulate matter are identified, including normal wear particles, fatigue wear particles, sliding wear particles, cutting wear particles, non-ferrous metal particles, non-metallic particles, fibers, bubbles, and water droplets.

3. The method for evaluating the wear and health status of wind turbine gear oil according to claim 2, characterized in that, The specific process of particle identification in S13 is as follows: First, an improved target detection algorithm is used to detect the target of interest in each frame, and the location coordinates, classification and confidence of the target are obtained as the detection result. The number of detected targets is denoted as M1. Secondly, associate the detection results of the previous step with the detection results of the target of interest detected in the previous frame one by one, and record the number of targets detected in the previous frame as M2. Find the most similar pair among M1×M2 pairs. Finally, the tracked wear particle images are combined with the target detection algorithm for identification, and the types of particles are output. Combined with steps S11~S12, other feature parameters of the wear particles are extracted to complete the statistics of wear particles.

4. The method for evaluating the wear and health status of wind turbine gear oil according to claim 3, characterized in that, The improved target detection algorithm includes the following: The image acquisition step specifically involves acquiring real-time images of the wind turbine gear oil using sensors and outputting raw image data at a preset frame rate. The image data processing steps specifically include grayscale correction, image scaling, normalization, and tensor format conversion of the original image data. The model inference step specifically involves loading the SAHI-optimized YOLO11 model for detection, and outputting the particle candidate boxes, class labels, and confidence information for each frame of the image. The post-processing steps specifically involve filtering out low-confidence particle candidate boxes, removing duplicate detection boxes using an IoU threshold, converting the particle size of the pixel image to the actual physical size, and outputting the particle number and category distribution results.

5. The method for evaluating the wear and health status of wind turbine gear oil according to claim 4, characterized in that, The YOLO11 model optimized by SAHI includes a backbone network, a neck network, and a head network. The backbone network splits the input feature map along the channel dimension into two parts: a Shortcut branch and a feature extraction branch; the Shortcut branch retains... The feature map of the channel transmits shallow details such as grain edges and textures, avoiding detail loss during feature transfer; the feature extraction branch will then extract the remaining... The feature maps of the channels are used for deep feature extraction through variable convolution kernels. By adjusting the convolution kernel size k and the number of branch repetitions n, the feature extraction requirements of different scale particles can be adaptively matched. The neck network is based on an improved PANet structure. It achieves efficient fusion of shallow detail features and deep semantic features through bidirectional feature transfer and the introduction of scale-adaptive fusion weights. The head network employs a decoupled detection head design, receiving the fused feature map output by the neck network. The regression and classification branches are set in parallel. The two branches share the basic convolutional layer but are optimized independently to improve the localization accuracy of particle bounding boxes and the accuracy of category judgment.

6. The method for evaluating the wear and health status of wind turbine gear oil according to claim 3, characterized in that, The normal wear particles described in S13 include scaly, elongated, and flat wear particles; all of which have smooth surfaces. The fatigue wear particles include spalling block-shaped, spherical, and flake-shaped wear particles; wherein, the surface of the spalling block-shaped wear particles is smooth and the edges are irregular, the surface of the spherical wear particles is smooth, and the surface of the flake-shaped wear particles has holes. The sliding wear particles include ordinary sliding wear particles and severe sliding wear particles; wherein, the surface of ordinary sliding wear particles is smooth, while the surface of severe sliding wear particles has obvious scratches and cracks, and both ordinary sliding wear particles and severe sliding wear particles have irregular shapes and straight edges; The cutting and abrasion particles include curved, spiral, annular, and strip-shaped abrasion particles; The non-ferrous metal particles include copper alloy and Fe2O3 oxide particles; wherein the copper alloy particles are yellow, the Fe2O3 oxide particles are red, and they have the characteristics of rust particles. The non-metallic particles are transparent particles; The fibers have a linear structure; The bubble has a circular structure; the center of the bubble has a small bright spot, while the overall brightness of the outer periphery of the bubble is relatively dark. The water droplet has a circular structure; the center of the water droplet has a large bright spot.

7. The method for evaluating the wear and health status of wind turbine gear oil according to claim 1, characterized in that, The calculation of wear particle weights in S2 is specifically as follows: S21, a first type of wear particles, a second type of wear particles, and a third type of wear particles are preset; the first type of wear particles includes spalling block wear particles, sliding wear particles, cutting wear particles, Fe2O3 oxide, copper alloy metal particles and non-metal particles; the second type of wear particles includes non-ferrous metal wear particles and severe sliding wear particles; and the third type of wear particles includes spherical wear particles. S22, traverse all particles in the wind turbine gear oil image, count the number and average size of different types of particles, and calculate the corresponding wear particle weights.

8. The method for evaluating the wear and health status of wind turbine gear oil according to claim 7, characterized in that, The wear particle weights are represented by the following logic: When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the first type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the second type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but ; When the number N of the third type of wear particles satisfies hour: If the average particle size d satisfies Then the first j Weight of each type of wear particle If d satisfies ,but If d satisfies ,but If d satisfies ,but .

9. The method for evaluating the wear and health status of wind turbine gear oil according to claim 1, characterized in that, The specific method for assessing the wear health level of the fan gear oil based on wear particle weighting in S3 is as follows: First, the wear health level of the fan gear oil is calculated using the following logic: In the formula, G Indicates the wear and tear health level. Indicates the first j Weight of each type of wear particle , This indicates the types and quantities of particulate matter present in the tested oil sample. Secondly, determine the wear health level of the fan gear oil. Specifically, when the wear health level... At that time, the wear of the fan gear oil was assessed as normal; when the wear health level was... When the wear is assessed as minor, the wear is rated as healthy. When the wear level is normal, it is assessed as abnormal; when the wear health level is normal... At that time, it was assessed as severely worn.

10. A system for evaluating the wear and health status of wind turbine gear oil, characterized in that, include: The feature analysis module performs feature analysis on particulate matter in the wind turbine gear oil image, including analysis of the size, quantity, and type of particulate matter. The weighting analysis module calculates the weight of wear particles based on their size, quantity, and type. The rating module assesses the wear health level of the wind turbine gear oil based on the weight of wear particles and outputs maintenance suggestions.

Images