A wind turbine bearing ring wear detection method

Abstract

The present invention relates to the field of image processing, and specifically to a method for detecting wear of a wind turbine bearing ring. A high-definition industrial camera is used to collect an all-round image of the outer ring of the bearing, and a linear neighborhood window is constructed by combining the main gradient direction and the main wear direction to generate binary coordinates. The DBSCAN clustering algorithm is used for clustering analysis to obtain an adaptive dual threshold of the Canny edge detection algorithm. The edge image is then segmented, and the texture features and geometric features of the wear area are extracted to form a comprehensive feature vector. Finally, the degree of wear of the bearing ring is classified and determined using a trained decision tree classification model. The present invention combines the main gradient direction with the main wear direction, uses the DBSCAN clustering algorithm to achieve adaptive dual threshold setting, extracts the texture and geometric features of the wear area, and finally uses a decision tree model to classify wear.

Description

Technical Field

[0001] The invention relates to the field of image processing, and in particular to a method for detecting wear of a wind turbine bearing ring. Background Art

[0002] As an important renewable energy source, wind power generation has developed rapidly around the world in recent years. One of the core components of wind turbines is the bearing ring, which is usually equipped with multiple bearing rings, such as pitch bearing rings, main shaft bearing rings, yaw bearing rings, gearbox bearing rings and generator bearing rings. Among them, the gearbox bearing ring is subjected to large friction and stress, and its working environment is complex. It is easily affected by friction, stress and environmental factors during long-term operation, resulting in wear, cracks and other damage to the bearing ring. These damages not only affect the operating efficiency of wind power equipment, but may also cause more serious mechanical failures and even equipment shutdowns. In severe cases, they will also pose a threat to the environment and personnel safety. Therefore, wear detection of wind power bearing rings has important practical significance. By disassembling the bearing ring wear detection of wind turbines that have reached the scrap age or have failed, its wear condition can be accurately evaluated and whether it has recycling value can be determined. At the same time, it can also provide strong data support for the analysis of the causes of wind power equipment failures, thereby optimizing equipment maintenance strategies, extending equipment service life, and reducing operation and maintenance costs.

[0003] However, the existing technology has many shortcomings when it comes to wind turbine bearing ring wear detection. Traditional wear detection methods mostly rely on manual inspection or image processing algorithms based on fixed thresholds. These methods are not only inefficient, but also limited by the operator's experience level, and are prone to missed detection or false detection. In addition, the fixed threshold image processing algorithm has unstable detection results when faced with complex lighting changes in wind turbine bearing rings, making it difficult to achieve high-precision and high-reliability wear assessment. Summary of the invention

[0004] In view of the problem that the above-mentioned fixed threshold detection effect is unstable, the present invention proposes a wind turbine bearing ring wear detection method, including: obtaining an outer ring image of a wind turbine gearbox bearing containing a wear area; performing Canny edge detection on the outer ring image to obtain an edge image, and performing image segmentation based on the edge image to obtain the wear area; extracting texture features and geometric features of the wear area to obtain a feature vector, and adding a wear degree label to input into a decision tree classification model for training; obtaining a real-time image of the outer ring of the wind turbine gearbox bearing, and inputting the real-time image into the trained decision tree classification model to obtain a corresponding wear degree label; judging the wear condition of the outer ring of the wind turbine gearbox bearing based on the corresponding wear degree label; the Canny edge detection also includes dual threshold detection, and the acquisition method of the adaptive high threshold and the adaptive low threshold of the dual threshold detection is specifically: constructing a vector with any pixel point as the center The first window and the second window, the direction of the first window is the main gradient direction, and the main gradient direction is the mode of the gradient directions of each pixel in the outer circle image; the direction of the second window is the main wear direction, and the angle between the main wear direction and the main gradient direction is a right angle; the gradient amplitude greater than or equal to the center pixel in the first window is recorded as 1, and the gradient amplitude less than the center pixel is recorded as 0, to obtain the gradient sequence of the first window; the grayscale value greater than or equal to the center pixel in the second window is recorded as 1, and the grayscale value less than the center pixel is recorded as 0, to obtain the grayscale sequence of the second window; the gradient sequence is used as the horizontal coordinate and the grayscale sequence is used as the vertical coordinate to form a gradient-grayscale space, and all pixels in the gradient-grayscale space are clustered; the mean grayscale value of all pixels in any cluster cluster is used as the adaptive high threshold of the corresponding cluster cluster, and 50% of the adaptive high threshold is used as the adaptive low threshold.

[0005] The present invention introduces an adaptive dual threshold setting based on the main gradient direction and the main wear direction into the Canny edge detection algorithm, which effectively solves the problem of unstable edge detection and low accuracy of the traditional fixed threshold method under complex lighting and wear direction changes. By constructing binary coordinates in the gradient-grayscale space and using a clustering algorithm to dynamically generate adaptive thresholds, the accuracy and robustness of edge detection in wind turbine gearbox bearing ring wear detection are significantly improved. Compared with the existing technology, this method can more accurately identify and segment wear areas, reduce missed detections and false detections, improve the reliability of wear assessment, and provide more scientific and efficient technical support for the maintenance of wind power equipment.

[0006] Furthermore, obtaining an image of the outer ring of a wind turbine gearbox bearing including a wear area also includes: placing a rotatable platform and an industrial camera at the center of the wind turbine gearbox bearing; and using the industrial camera to take an RGB image of the outer ring of the wind turbine gearbox bearing every time it rotates a set angle.

[0007] Furthermore, after obtaining the outer circle image, preprocessing operations are required: grayscale the obtained RGB image using a weighted grayscale algorithm; denoise the grayscale image using bilateral filtering; and use adaptive histogram equalization to improve the local contrast of the denoised image.

[0008] Furthermore, performing image segmentation based on the edge image to obtain the worn area also includes: inputting the edge image into a Graph Cut algorithm to obtain the worn area; the Graph Cut algorithm also includes manually marking the marking points of the foreground area and the marking points of the background area; the foreground area is the worn area; and the background area is the unworn area.

[0009] By inputting the edge image into a graph cutting algorithm and combining it with manually annotated marking points of worn areas and unworn areas, the present invention can achieve more accurate image segmentation and accurately distinguish worn areas from unworn areas.

[0010] Furthermore, the decision tree classification model also includes using 80% of the sample data for training and using 20% ​​of the sample data for verification; the feature vectors corresponding to all outer ring images and the wear degree labels constitute the sample data.

[0011] Furthermore, extracting the texture features and geometric features of the wear area to obtain a feature vector also includes: using a gray level co-occurrence matrix to extract the texture features of the wear area of ​​the bearing outer ring; using the HU moment to obtain the geometric features of the wear area; using a Z-score algorithm to standardize the texture features and geometric features; the texture features and geometric features together constitute a feature vector.

[0012] By using the gray-level co-occurrence matrix to extract texture features and the HU moment to extract geometric features, the present invention can fully capture the detailed information and shape features of the wear area. Subsequently, the Z-score algorithm is applied to standardize the extracted features to eliminate the influence of different feature magnitudes and ensure the consistency and comparability of the feature vectors. The combination of texture features and geometric features makes the feature vector more comprehensive and representative, and improves the subsequent decision tree classification model's ability to identify and classify the degree of wear. Compared with existing technologies that rely only on a single feature extraction method, this method significantly improves the accuracy and reliability of wear detection through multi-dimensional feature fusion.

[0013] Furthermore, clustering all pixel points in the gradient-grayscale space also includes: obtaining the gradient-grayscale space coordinates of all pixel points and then using the DBSCAN algorithm to perform clustering; setting the neighborhood radius value range of the DBSCAN algorithm to 0.4~0.6; and setting the minimum number of points of the DBSCAN algorithm to 3~5.

[0014] The DBSCAN clustering algorithm is used to cluster binary coordinates in the gradient-grayscale space. The present invention can effectively identify groups of pixel points with similar gradient and grayscale features, which helps to accurately distinguish real edge points from noise points in complex images, improve the accuracy and robustness of clustering, ensure the stability and high accuracy of edge detection results, and provide a reliable basis for the generation of adaptive thresholds.

[0015] Furthermore, the distance calculation method of the DBSCAN algorithm is specifically as follows:

[0016] ;

[0017] in represents the distance between the pixels corresponding to the binary coordinates (z1, v1) and (z2, v2) in the gradient-grayscale space; Ham() represents the Hamming distance calculation function.

[0018] A specific distance calculation method is used, combined with the Hamming distance calculation function, to accurately measure the similarity between binary coordinates in the gradient-grayscale space. Through this unique distance calculation method, the present invention can more accurately reflect the similarity of pixel points in gradient and grayscale features, and improve the classification effect of the DBSCAN clustering algorithm in complex images. Compared with the traditional Euclidean distance or the calculation method based only on the Hamming distance, this method enhances the clustering algorithm's ability to recognize edge features by combining the characteristics of the two, and further improves the accuracy of adaptive threshold setting and the reliability of overall wear detection.

[0019] Furthermore, judging the wear condition of the outer ring of the wind turbine gearbox bearing based on the corresponding wear degree label also includes: when the value of the corresponding wear degree label is in a first interval, judging that the outer ring of the wind turbine gearbox bearing is slightly worn; when the value of the corresponding wear degree label is in a second interval, judging that the outer ring of the wind turbine gearbox bearing is moderately worn; when the value of the corresponding wear degree label is in a third interval, judging that the outer ring of the wind turbine gearbox bearing is heavily worn.

[0020] The technical effects of the present invention are:

[0021] The present invention first calculates the main gradient direction and the main wear direction of the image, constructs a linear neighborhood window, and generates binary coordinates in the gradient-grayscale space. Subsequently, the DBSCAN clustering algorithm is used to perform cluster analysis on these binary coordinates, and adaptive high and low thresholds are dynamically generated based on the gradient and grayscale consistency scores of each cluster. This process significantly improves the stability and accuracy of edge detection in complex environments. In addition, the graph cut algorithm is combined to perform region segmentation, and the texture features and geometric features of the wear area are extracted to form a comprehensive feature vector, and the decision tree classification model is further used to achieve accurate classification of the degree of wear. This method not only improves the detection accuracy, but also enhances the automation and intelligence level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0023] Figure 1 is a flow chart schematically showing a method for detecting wear of a wind turbine bearing ring according to an embodiment of the present invention;

[0024] Figure 2 is a grayscale three-dimensional schematic diagram schematically showing the position of a wind turbine bearing ring in an embodiment of the present invention;

[0025] Figure 3 is a grayscale three-dimensional schematic diagram schematically showing a bearing ring of a wind turbine gearbox according to an embodiment of the present invention;

[0026] Figure 4 It is a color diagram schematically showing the wear condition of the outer ring of the wind turbine gearbox bearing in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0028] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] A wind turbine bearing ring wear detection method embodiment:

[0030] like Figure 1 As shown, a wind turbine bearing ring wear detection method of the present invention comprises:

[0031] S1. Collect and pre-process the image of the wind turbine gearbox bearing ring.

[0032] As an important renewable energy device, wind turbines have bearing rings as one of their core components. There are multiple bearing rings in a wind turbine, such as Figure 2 As shown, Q1 represents the pitch bearing ring, Q2 represents the main shaft bearing ring, Q3 represents the yaw bearing ring, Q4 represents the gearbox bearing ring, and Q5 represents the generator bearing ring. The gearbox bearing ring is used to ensure the stability and precise positioning of the rotating shaft, which is crucial to the efficiency of the gear transmission and the overall mechanical balance of the gearbox. Its structural diagram is shown in Figure 3 As shown, it is composed of an inner ring, an outer ring, and rolling bodies. In the gearbox bearing ring shown in this embodiment, the rolling bodies are cylindrical rollers.

[0033] Since the bearing rings and cylindrical rollers are affected by friction, stress and environmental factors during long-term operation, wear and cracks are likely to occur. After dismantling wind turbines that have reached their scrap age or have failed, the bearing rings are tested for wear to assess their recycling value. At the same time, the causes of wind power equipment failures can also be analyzed based on the wear conditions.

[0034] A rotatable platform is placed at the center of the bearing ring, and an industrial camera with a resolution of not less than 1920×1080 pixels is fixed at the center of the platform to take images of the outer ring of the bearing including the wear area. The rotating platform can be set to take an image with the camera every 30° rotation. For each outer ring of the bearing, there is In this embodiment, the rotation angle of the image is 30°, and the implementer may also select 20°, 45° or 60°.

[0035] After obtaining the bearing outer ring image containing the wear area, preprocessing is performed. First, the obtained RGB image is grayed using a weighted grayscale algorithm. The brightness change characteristics are retained and the interference of color information is removed by the weighted average method. Then, the grayscale image is denoised using bilateral filtering, which can effectively remove noise while retaining edge details and improve image quality. Finally, adaptive histogram equalization is used to increase the local contrast of the image and make the wear characteristics more obvious. The above-mentioned weighted grayscale algorithm, bilateral filtering and adaptive histogram equalization are all well-known technologies, and the specific implementation methods will not be repeated here.

[0036] S2. Obtain the main gradient direction and the main wear direction of the image; construct a first linear neighborhood window and a second linear neighborhood window based on the main gradient direction and the main wear direction; obtain a gradient sequence according to the gradient characteristics of the first linear neighborhood window, obtain a grayscale sequence according to the grayscale characteristics of the second linear neighborhood window, and combine the gradient sequence and the grayscale sequence to obtain the gradient-grayscale space binary coordinates.

[0037] In the actual application scenario of the bearing ring of the wind turbine gearbox, due to the friction of the cylindrical roller, the surface of the area with more force becomes matte and the reflection is weaker, while the area with less force has stronger reflection, resulting in obvious uneven lighting in the image, such as Figure 4 shown.

[0038] In this embodiment, the Canny edge detection algorithm can be used to detect the edge of the wear area on the outer ring of the bearing, and the area can be segmented based on the edge to obtain the situation of the wear area. The Canny edge detection algorithm usually includes grayscale conversion, Gaussian filtering, gradient calculation, non-maximum suppression, dual threshold detection and edge connection, among which the dual threshold detection usually uses a fixed threshold or is adjusted manually, which is difficult to adapt to complex lighting changes and directional characteristics, resulting in instability and decreased accuracy of edge detection. Therefore, in this embodiment, the fixed threshold in the Canny edge detection algorithm can be improved to improve its edge detection effect in the application scenario of the present invention, so as to make the wear detection result more accurate.

[0039] observe Figure 4 It can be found that the wear area is continuous and irregular in shape, but since the cylindrical roller can only keep rolling in a fixed track, the wear area still shows a certain directionality. Therefore, in this embodiment, the main gradient direction in the image is first obtained. In the Canny edge detection algorithm, the Sobel operator is usually used to obtain the gradient of the pixel point. The specific gradient acquisition process is not repeated here. For the pixel point with coordinates (x, y), its gradient amplitude is recorded as G(x, y) and the gradient direction is recorded as θ(x, y).

[0040] In one embodiment, the main gradient direction θ is the mode of the gradient directions of all pixels in the image;

[0041] In another embodiment, the method for calculating the main gradient direction is specifically:

[0042] ;

[0043] Where Θ represents the main gradient direction of the entire image; G(x,y) represents the gradient amplitude of the pixel (x,y); θ(x,y) represents the gradient direction of the pixel (x,y); X represents the maximum value of the horizontal coordinate x of the pixel in the image; Y represents the maximum value of the vertical coordinate y of the pixel in the image; and X×Y is the total number of pixels in the image.

[0044] By calculating the main gradient direction of the image, the dominant direction of the overall gradient change in the image is reflected. Since the wear area has a certain directionality, the wear direction of the outer ring of the bearing is roughly at right angles to the dominant direction of the gradient change. Therefore, in this embodiment, the main wear direction is further obtained based on the main gradient direction θ:

[0045] ;

[0046] Where Φ represents the main wear direction; Θ represents the main gradient direction. It should be noted that when Φ>π, Φ=Φ-π, which is used to limit the value range of the main wear direction to [0,π]. So far, the main gradient direction Θ and the main wear direction Φ are obtained for the pixel point with coordinates (x, y). A linear neighborhood window is further constructed along the main wear direction. By evaluating the gradient changes in the linear neighborhood window of the pixel point, the accuracy of distinguishing the real edge from the noise edge is improved.

[0047] First, for any pixel point after maximum value suppression in the Canny edge detection algorithm, the first linear neighborhood window is constructed with the pixel point as the center, which can also be recorded as the first window. In this embodiment, the length of the first linear neighborhood window can be set to an empirical value of 7px (pixels). The implementer can also set the length to an odd value such as 9px, 11px or 15px according to actual conditions to ensure that the pixel point is in the center of the first linear neighborhood window. It should be noted that the direction of the first linear neighborhood window in this embodiment is the main wear direction. An exemplary explanation is: if in one embodiment the main wear direction of the image is 0°, then with any pixel point as the center, extend 3px in the directions of 0° and 180° respectively to form the first linear neighborhood window. The total number of linear neighborhood windows is 7. Then the pixel position of the first linear neighborhood window is expressed as:

[0048] ;

[0049] Where (x i ,y i ) represents the coordinates of the i-th pixel in the first linear neighborhood window; k represents half of the total number of pixels in the first linear neighborhood window except the central pixel. , L1 represents the total length of the first linear neighborhood window. In this embodiment, L1 is an empirical value of 7. The gradient amplitudes of all pixels in the first linear neighborhood window are further obtained to construct a binary sequence. The calculation method is:

[0050] ;

[0051] Among them B i represents the binary value corresponding to the i-th pixel in the first linear neighborhood window; G(P i ) represents the i-th pixel point P in the first linear neighborhood window i The corresponding gradient amplitude; G(P0) represents the gradient amplitude of the central pixel point P0 in the first linear neighborhood window; k represents half of the total number of pixels in the first linear neighborhood window except the central pixel point. Then each first linear neighborhood window corresponds to a binary sequence. An exemplary explanation is: there is a first linear neighborhood window with a length of 7px, and the gradient amplitudes corresponding to each pixel are arranged in order: {80,90,86,83,80,80,75}. After comparing the gradient amplitude of each pixel point with the gradient amplitude of the central pixel point, the binary sequence corresponding to the first linear neighborhood window is obtained according to the above binary sequence calculation method: 0111000.

[0052] So far, for any pixel point, there is a corresponding binary sequence, which reflects the gradient change of the pixel point in the first linear neighborhood along the main wear direction in a simpler binary form. Then the pixel points with an approximate binary sequence are more likely to be edge points of the same type. At the same time, the uneven reflection caused by the difference in reflection between the matte area and the unworn bright area after the outer ring of the bearing is worn may also cause the segmentation effect of the fixed threshold to be poor. Therefore, in this embodiment, the grayscale characteristics of the pixel point are further evaluated according to the illumination conditions of the target pixel point along the main gradient direction.

[0053] Taking any pixel as the center, a second linear neighborhood window is constructed along the main gradient direction, which can also be recorded as the second window. The length of the window is also set to an empirical value of 7px. For example, if the main gradient direction of an image is 90°, then taking any pixel as the center, the second linear neighborhood window of the pixel is obtained by extending 3px along 90° and 270°. Then the second linear neighborhood window is expressed as:

[0054] ;

[0055] Where (x j ,y j ) represents the coordinates of the jth pixel in the second linear neighborhood window; p represents half of the total number of pixels in the second linear neighborhood window except the central pixel. , L2 represents the total length of the linear neighborhood window, which is an empirical value of 7 in this embodiment. The grayscale values ​​of all pixels in the second linear neighborhood window are further obtained to construct a binary sequence, and the calculation method is:

[0056] ;

[0057] Among them, j represents the binary value corresponding to the j-th pixel in the second linear neighborhood window; I(P j ) represents the jth pixel point P in the second linear neighborhood window j The corresponding gradient amplitude; I(P0) represents the gradient amplitude of the central pixel P0 in the second linear neighborhood window; p represents half of the total number of pixels in the second linear neighborhood window except the central pixel. Similarly, each second linear neighborhood window corresponds to a binary sequence. An exemplary explanation is: there is a second linear neighborhood window with a length of 7px, and the grayscale values ​​corresponding to the pixels in the window are arranged in order: {200, 200, 200, 10, 10, 5, 5}. After comparing the grayscale value of each pixel with the grayscale value of the central pixel, the binary sequence corresponding to the second linear neighborhood window is obtained as: 1111100.

[0058] At this point, the binary sequence corresponding to the first linear neighborhood window and the binary sequence corresponding to the second linear neighborhood window of any pixel point are obtained, wherein the binary sequence corresponding to the first linear neighborhood window can also be recorded as a gradient sequence, and the binary sequence corresponding to the second linear neighborhood window can also be recorded as a grayscale sequence. After obtaining the gradient sequence and grayscale sequence of the pixel point, the gradient sequence is used as the horizontal coordinate and the grayscale sequence is used as the vertical coordinate to obtain the binary coordinates of the gradient-grayscale space. Exemplary explanation: There is a pixel point whose gradient sequence is: 0111000, and its grayscale sequence is: 1111100, then the binary coordinates of the gradient-grayscale space of the pixel point are (0111000, 1111100), which can also be recorded as binary coordinates in this embodiment.

[0059] For all pixels in the image, their gradient sequence and grayscale sequence are obtained to obtain the binary coordinates of each pixel, which are used to reflect the gradient changes along the main wear direction and the grayscale changes along the main gradient direction in its neighborhood. Any binary coordinate in the gradient-grayscale space is marked as (z, v).

[0060] S3. Use the DBSCAN clustering algorithm to cluster based on the binary coordinates of all pixels; obtain the adaptive dual threshold of the Canny edge detection algorithm for each cluster based on the gradient consistency score and grayscale consistency score of the pixels in the cluster.

[0061] After obtaining the binary coordinates of all pixels, the DBSCAN algorithm is used for clustering. The input of the algorithm includes a point set and a distance metric. Since the coordinates in this embodiment are in binary form, the distance metric is calculated based on the Hamming distance. An exemplary explanation is as follows: there are two binary coordinates (z1, v1) and (z2, v2). Then the Hamming distance between z1 and z2 and the Hamming distance between v1 and v2 are calculated respectively, and the distance between (z1, v1) and (z2, v2) is obtained by taking the square root of the product. The specific calculation method is as follows:

[0062] ;

[0063] in Represents the distance between the pixels corresponding to the binary coordinates (z1, v1) and (z2, v2); Ham () represents the Hamming distance calculation function. The input of the DBSCAN algorithm also includes the neighborhood radius Eps and the minimum number of points MinPts, which can be set to the empirical values ​​Eps=0.5 and MinPts=4 in this embodiment; the output of the algorithm includes clustering labels, core points, boundary points and noise points, and the clustering of all pixels in the image is completed based on the clustering label corresponding to each pixel. The above DBSCAN clustering algorithm belongs to the well-known technology, and the specific implementation method will not be repeated here.

[0064] A specific distance calculation method is used, combined with the Hamming distance calculation function, to accurately measure the similarity between binary coordinates in the gradient-grayscale space. Through this unique distance calculation method, the present invention can more accurately reflect the similarity of pixel points in gradient and grayscale features, and improve the classification effect of the DBSCAN clustering algorithm in complex images. Compared with the traditional Euclidean distance or the calculation method based only on the Hamming distance, this method enhances the clustering algorithm's ability to recognize edge features by combining the characteristics of the two, and further improves the accuracy of adaptive threshold setting and the reliability of overall wear detection.

[0065] The above steps cluster all the pixels after maximum value suppression based on gradient features and grayscale features. For the pixels in the same cluster, the edge features of their locations have great similarity. Therefore, adaptive dual threshold detection is performed on the pixels in each cluster. In one embodiment, the grayscale mean of all the pixels in the cluster can be used as the adaptive high threshold T t,high , the adaptive high threshold T t,high 50% of the adaptive high threshold T t,low .

[0066] In another embodiment, the method for obtaining the adaptive dual threshold is specifically as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] Where S t represents the gradient consistency score of all pixels in the tth cluster; M represents the total number of pixels in the tth cluster; G(P m ) represents the mth pixel point P in the tth cluster m The gradient amplitude of represents the mean gradient amplitude of all pixels in the tth cluster; D t represents the gray consistency score of all pixels in the tth cluster; I(P m ) represents the mth pixel point P in the tth cluster m Gray value of represents the grayscale mean of all pixels in the tth cluster; T t,high represents the adaptive high threshold of the t-th cluster; I t,max represents the maximum grayscale value of all pixels in the tth cluster; T t,low Represents the adaptive low threshold for the t-th cluster.

[0072] The above method is used for all clusters to obtain the adaptive dual threshold value. Correspondingly, the dual threshold value used for dual threshold detection of any pixel after non-maximum suppression is the adaptive dual threshold value corresponding to the cluster to which it belongs.

[0073] So far, the adaptive dual threshold improvement of the Canny edge detection algorithm has been completed. The final steps of the Canny edge detection algorithm are: grayscale conversion, Gaussian filtering (the preprocessing step of this embodiment has been grayscaled, and the first two steps can be omitted), gradient calculation, non-maximum suppression, pixel clustering based on binary coordinates, obtaining adaptive dual thresholds for each cluster, dual threshold detection, and edge connection. The input of the above-mentioned Canny edge detection algorithm includes a grayscale image of the outer ring of the bearing, and the output is an edge image. The specific implementation method belongs to the known technology and will not be repeated here.

[0074] S4. Obtain edge images based on the adaptive dual-threshold Canny edge detection algorithm; obtain the wear area through the graph cut algorithm and obtain the texture features and geometric features of the wear area to obtain the feature vector, and manually annotate the labels to obtain the sample set; train a decision tree classification model based on the sample set, and perform wear detection on the real-time image of the outer ring of the gearbox bearing based on the trained decision tree classification model.

[0075] First, the edge image of the outer ring of the bearing is obtained based on the improved dual-threshold Canny edge detection algorithm in step S3; then, the edge detection result is segmented by the Graph Cut algorithm to divide the image into a worn area and an unworn area. The input of the algorithm includes the edge image; the manually annotated foreground and background marking points, in this embodiment, the foreground and background refer to the worn area and the unworn area respectively; the input of the algorithm also includes the edge weight and the smoothing item weight. In this embodiment, the edge weight can be set as a weight based on the spatial distance, and the smoothing item weight can be set as an empirical value of 0.5; further, features are extracted based on the segmented worn area image. The specific steps of feature extraction are as follows:

[0076] First, the gray-level co-occurrence matrix is ​​used to extract the texture features of the outer ring wear area of ​​the bearing, including contrast, correlation and energy; then the HU moment is used to obtain the geometric features of the wear area, including 7 moments. The Z-score algorithm is used to standardize all the above texture features and geometric features. The texture features and geometric features corresponding to the nth wear area together form a feature vector, denoted as V n The above-mentioned graph cut algorithm, gray level co-occurrence matrix and HU moment belong to well-known technologies, and the specific implementation methods will not be described in detail here.

[0077] By using the gray-level co-occurrence matrix to extract texture features and the HU moment to extract geometric features, the present invention can fully capture the detailed information and shape features of the wear area. Subsequently, the Z-score algorithm is applied to standardize the extracted features to eliminate the influence of different feature magnitudes and ensure the consistency and comparability of the feature vectors. The combination of texture features and geometric features makes the feature vector more comprehensive and representative, and improves the subsequent decision tree classification model's ability to identify and classify the degree of wear. Compared with existing technologies that rely only on a single feature extraction method, this method significantly improves the accuracy and reliability of wear detection through multi-dimensional feature fusion.

[0078] Finally, after obtaining the feature vector of the wear area, a wear degree label is manually added to each wear area. In this implementation, the label can be set to any decimal between [0,1] according to the wear degree, such as 0.1, 0.2 or 0.3, etc.; and the wear degree label corresponding to the nth wear area is recorded as U n , then the nth wear region has a characteristic vector F n , and there is: F n ={V n ,U n}.

[0079] After obtaining edge images of all images containing worn areas according to the method of step S3, the feature vectors of each worn area are extracted, and all feature vectors constitute a sample set, and 80% of the sample set is input into the decision tree for training, and the remaining 20% ​​of the sample set is used for verification, so as to obtain a trained decision tree classification model. The above decision tree algorithm belongs to the well-known technology, and the specific training process will not be described in detail here.

[0080] The method of step S1 is used to capture a real-time image of the outer ring of the gearbox bearing, and then the edge detection algorithm after the improved double threshold is used to obtain its edge image, and then its feature vector is obtained. Then the feature vector is input into the trained decision tree classification model to obtain the wear degree value of each wear area in the image, and finally the wear degree of the outer ring of the gearbox bearing is judged based on the sum of the wear degrees in the real-time image. In this embodiment, when the sum of the wear degrees is in the interval [0, 0.2], it is considered that the outer ring of the gearbox bearing is slightly worn, and the above interval [0, 0.2] can also be recorded as the first interval; when the sum of the wear degrees is in the interval (0. 2,0.5], the outer ring of the gearbox bearing is considered to be moderately worn. A specific analysis is conducted based on the wear area to evaluate its recovery value, and the cause of wear is analyzed to optimize the equipment maintenance strategy. The above interval (0.2,0.5] can also be recorded as the second interval; when the total wear degree is in the interval (0.5,1], the outer ring of the gearbox bearing is considered to be severely worn. Further analysis is conducted on the cause of severe bearing wear to prevent other wind power equipment in the same area from experiencing the same severe wear and causing equipment failure. The above interval (0.5,1] can also be recorded as the third interval. The implementers of the above interval thresholds can also set them according to actual conditions.

[0081] The present invention introduces an adaptive dual threshold setting based on the main gradient direction and the main wear direction into the Canny edge detection algorithm, which effectively solves the problem of unstable edge detection and low accuracy of the traditional fixed threshold method under complex lighting and wear direction changes. By constructing binary coordinates in the gradient-grayscale space and using a clustering algorithm to dynamically generate adaptive thresholds, the accuracy and robustness of edge detection in wind turbine gearbox bearing ring wear detection are significantly improved. Compared with the existing technology, this method can more accurately identify and segment wear areas, reduce missed detections and false detections, improve the reliability of wear assessment, and provide more scientific and efficient technical support for the maintenance of wind power equipment.

[0082] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0083] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A method for detecting wear of a wind turbine bearing ring, characterized in that: The method comprises: Obtain an outer ring image of a wind turbine gearbox bearing containing a wear area; perform Canny edge detection on the outer ring image to obtain an edge image, and perform image segmentation based on the edge image to obtain the wear area; extract texture features and geometric features of the wear area to obtain a feature vector, and add a wear degree label to input into a decision tree classification model for training; Acquire a real-time image of the outer ring of the wind turbine gearbox bearing, input the real-time image into a trained decision tree classification model to obtain a corresponding wear degree label; and judge the wear condition of the outer ring of the wind turbine gearbox bearing based on the corresponding wear degree label; The Canny edge detection also includes dual threshold detection, and the adaptive high threshold and the adaptive low threshold of the dual threshold detection are obtained in the following manner: a first window and a second window are constructed with any pixel point as the center, the direction of the first window is the main gradient direction, and the main gradient direction is the mode of the gradient directions of each pixel point in the outer circle image; the direction of the second window is the main wear direction, and the angle between the main wear direction and the main gradient direction is a right angle; The gradient amplitude greater than or equal to the central pixel point in the first window is recorded as 1, and the gradient amplitude less than the central pixel point is recorded as 0, so as to obtain the gradient sequence of the first window; the grayscale value greater than or equal to the central pixel point in the second window is recorded as 1, and the grayscale value less than the central pixel point is recorded as 0, so as to obtain the grayscale sequence of the second window; the gradient sequence is used as the abscissa and the grayscale sequence is used as the ordinate to form a gradient-grayscale space, and all the pixels in the gradient-grayscale space are clustered; The mean grayscale value of all pixels in any cluster is used as the adaptive high threshold of the corresponding cluster, and 50% of the adaptive high threshold is used as the adaptive low threshold.

2. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: Acquire images of the outer race of a wind turbine gearbox bearing including the worn area, including: A rotatable platform and an industrial camera are placed at the center of the wind turbine gearbox bearing; An industrial camera is used to capture an RGB image of the outer ring of the wind turbine gearbox bearing every time the bearing rotates by a set angle.

3. A wind turbine bearing ring wear detection method according to claim 2, characterized in that: After obtaining the outer ring image, preprocessing operations are required: Use the weighted grayscale algorithm to grayscale the acquired RGB image; Use bilateral filtering to reduce the noise of grayscale images; Adaptive histogram equalization is used to improve the local contrast of the denoised image.

4. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: Performing image segmentation based on the edge image to obtain the wear area includes: Inputting the edge image into a Graph Cut algorithm to obtain a wear area; The Graph Cut algorithm also includes manually marking the marker points of the foreground area and the marker points of the background area; The foreground area is a worn area; The background area is an unworn area.

5. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: The decision tree classification model also includes using 80% of the sample data for training and using 20% ​​of the sample data for verification; The feature vectors corresponding to all outer ring images and the wear degree labels constitute the sample data.

6. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: Extracting the texture features and geometric features of the wear area to obtain a feature vector includes: The texture features of the wear area of ​​the outer ring of the bearing are extracted using the gray level co-occurrence matrix; Use HU moment to obtain the geometric characteristics of the wear area; The texture features and geometric features are standardized using a Z-score algorithm; The texture features and geometric features together constitute a feature vector.

7. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: Clustering all pixels in the gradient-grayscale space includes: After obtaining the gradient-grayscale space coordinates of all pixels, use the DBSCAN algorithm for clustering; Set the neighborhood radius of the DBSCAN algorithm to 0.4~0.6; Set the minimum number of points for the DBSCAN algorithm to 3~5.

8. A wind turbine bearing ring wear detection method according to claim 7, characterized in that: The distance calculation method of the DBSCAN algorithm is specifically as follows: ； in represents the distance between the pixels corresponding to the binary coordinates (z1, v1) and (z2, v2) in the gradient-grayscale space; Ham() represents the Hamming distance calculation function.

9. A wind turbine bearing ring wear detection method according to claim 1, characterized in that: Judging the wear condition of the outer ring of the wind turbine gearbox bearing based on the corresponding wear degree label includes: When the value of the corresponding wear degree label is in the first interval, it is determined that the outer ring of the wind turbine gearbox bearing is slightly worn; When the value of the corresponding wear degree label is in the second interval, it is determined that the outer ring of the wind turbine gearbox bearing is moderately worn; When the value of the corresponding wear degree label is in the third interval, it is determined that the outer ring of the wind turbine gearbox bearing is severely worn.

Images