A method and system for detecting wear of a coating of a hydraulic turbine

Abstract

This invention discloses a method and system for detecting wear on a turbine coating, belonging to the field of image processing technology. The invention first acquires the original RGB image of the turbine surface coating, extracts the relative color shift spectrum, color perturbation gradient, and local texture perturbation index for each pixel, and constructs a coupling tensor for each pixel. By solving the eigenvalues ​​of the coupling tensor and selecting the largest eigenvalue as the wear intensity of the corresponding pixel, a wear intensity distribution map is generated. Subsequently, the scale weight of each pixel is calculated within three scale neighborhoods, resulting in three scale weight distribution maps. Finally, a multi-scale enhanced fusion network is used to process the wear intensity distribution map and the three scale weight distribution maps to output a wear score for the turbine coating. This invention, through the collaborative design of multi-dimensional feature coupling and multi-scale weight fusion, achieves a refined characterization of the coating wear state and improves detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method and system for detecting wear on the coating of a water turbine. Background Technology

[0002] As hydro turbines operate for extended periods, the protective coating on their metal surfaces gradually wears away under the combined effects of high-speed water flow, erosion by solid particles, and cavitation. Once the coating wears away, it not only exposes the localized metal substrate but can also lead to surface corrosion and fatigue crack propagation, thereby reducing the turbine's operational reliability and energy efficiency. Therefore, conducting wear detection on hydro turbine coatings is of great significance for equipment condition assessment and maintenance.

[0003] Currently, wear detection of turbine coatings mainly relies on manual inspection or image-based judgment methods based on experience rules. Manual inspection usually judges the degree of wear by visual observation and comparing surface changes through photos. However, this method is labor-intensive, heavily dependent on human experience, and highly subjective, making it prone to missed detections or misjudgments.

[0004] Another commonly used method is image-based identification, which often distinguishes worn areas from normal areas by relying on single image grayscale changes or boundary morphology features. However, during coating wear, surface condition changes are significantly complex (such as uneven wear and blurred boundaries between worn and non-wear areas), making it difficult to accurately characterize the wear nature using only a single feature, resulting in low accuracy in turbine coating wear detection. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a method and system for detecting wear of turbine coatings, which solves the problem of low accuracy in detecting wear of turbine coatings in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a method for detecting wear of a water turbine coating, comprising the following steps:

[0007] The original RGB image of the coating on the surface of the water turbine was acquired, and the relative color shift spectrum, color perturbation gradient and local texture perturbation index were extracted for each pixel to construct the coupling tensor of the pixel.

[0008] The eigenvalues ​​of the coupled tensor of each pixel are solved, and the largest eigenvalue is selected as the wear intensity of the corresponding pixel to form a wear intensity distribution map.

[0009] Within the three scale neighborhoods of the wear intensity distribution map, the scale weight of each pixel is calculated to obtain the three scale weight distribution maps;

[0010] A multi-scale enhanced fusion network was used to process the wear intensity distribution map and the weight distribution map of the three scales to obtain the wear score of the turbine coating.

[0011] Furthermore, the process of obtaining the relative color shift spectrum includes:

[0012] Obtain the neighborhood range centered on the pixel;

[0013] Calculate the RGB mean of each pixel in the neighborhood range to obtain the neighborhood RGB mean;

[0014] Subtract the average RGB value of the neighborhood from the RGB value of the pixel to obtain the RGB offset value;

[0015] The relative color shift spectrum is obtained by dividing the RGB offset value by the mean RGB value of the neighborhood and taking the 2-norm.

[0016] Furthermore, the process of obtaining the color perturbation gradient includes:

[0017] Calculate the ratio chromaticity for each pixel:

[0018] ,

[0019] in, For the first The ratio of color intensity per pixel, For the first The R channel value of each pixel For the first The G channel value of each pixel For the first The B channel value of each pixel To avoid parameters with a denominator of 0;

[0020] The gradient operator is used to calculate the color perturbation gradient of the ratio chromaticity of each pixel.

[0021] Furthermore, the process of obtaining the local texture perturbation index includes:

[0022] Obtain the neighborhood range centered on the pixel;

[0023] Calculate the RGB mean of each pixel in the neighborhood range to obtain the neighborhood RGB mean;

[0024] Calculate the standard deviation of the neighborhood range based on the difference between the RGB value of each pixel in the neighborhood range and the mean RGB value of the neighborhood range;

[0025] The ratio of the standard deviation of the neighborhood range to the mean of the neighborhood RGB values ​​is used as the local texture perturbation index.

[0026] Furthermore, the coupling tensor is:

[0027] ,

[0028] in, For the first The coupling tensor of each pixel For the first The relative color offset spectrum of each pixel For the first Color perturbation gradient of each pixel For the first Local texture perturbation index of each pixel It is a positive integer.

[0029] Furthermore, the process of calculating the scale weight for each pixel includes:

[0030] The mean wear intensity is calculated for each pixel in the wear intensity distribution map within three scale neighborhoods, with the sizes of the three scale neighborhoods being 3×3, 5×5, and 7×7, respectively.

[0031] The standard deviation is calculated based on the difference between the wear intensity of each pixel and the mean wear intensity within the neighborhood of each scale.

[0032] Calculate the scale weight for each pixel based on the standard deviation of each scale.

[0033] Furthermore, the formula for calculating the scale weight of each pixel is as follows:

[0034] ,

[0035] in, For the first The first pixel Each scale weight, where exp is an exponential function. For the first Wear intensity per pixel For the first The first pixel Average wear strength at various scales For the first The first pixel Standard deviation at various scales To avoid parameters with a denominator of 0, and It is a positive integer.

[0036] Furthermore, the multi-scale enhancement fusion network includes: four convolutional branches, three feature enhancement branches, three CNN networks, a branch feature weighted fusion module, and a fully connected layer;

[0037] The input of the first convolutional branch is used to input the first scale weight distribution map; the input of the second convolutional branch is used to input the second scale weight distribution map; the input of the third convolutional branch is used to input the third scale weight distribution map; and the input of the fourth convolutional branch is used to input the wear intensity distribution map.

[0038] The first input of the first feature enhancement branch is connected to the output of the first convolution branch, and its output is connected to the input of the first CNN network; the first input of the second feature enhancement branch is connected to the output of the second convolution branch, and its output is connected to the input of the second CNN network; the first input of the third feature enhancement branch is connected to the output of the third convolution branch, and its output is connected to the input of the third CNN network.

[0039] The output of the fourth convolutional branch is connected to the second input of the first feature enhancement branch, the second input of the second feature enhancement branch, and the second input of the third feature enhancement branch, respectively.

[0040] The input of the branch feature weighted fusion module is connected to the output of the first CNN network, the output of the second CNN network, and the output of the third CNN network, respectively, and its output is connected to the input of the fully connected layer.

[0041] The output of the fully connected layer serves as the output of the multi-scale enhanced fusion network.

[0042] Furthermore, the first feature enhancement branch is used to element-wise multiply the weighted features output by the first convolution branch with the wear intensity features output by the fourth convolution branch to obtain the first wear intensity enhancement feature;

[0043] The second feature enhancement branch is used to multiply the weighted features output by the second convolution branch with the wear intensity features output by the fourth convolution branch element-wise to obtain the second wear intensity enhancement feature;

[0044] The third feature enhancement branch is used to multiply the weighted features output by the third convolution branch with the wear intensity features output by the fourth convolution branch element-wise to obtain the third wear intensity enhancement feature.

[0045] A water turbine coating wear detection system includes: a coupling tensor construction unit, a first distribution map construction unit, a second distribution map construction unit, and a scoring unit;

[0046] The coupling tensor construction unit is used to acquire the original RGB image of the coating on the surface of the water turbine, extract the relative color shift spectrum, color perturbation gradient and local texture perturbation index for each pixel, and construct the coupling tensor of the pixel;

[0047] The first distribution map construction unit is used to solve for the eigenvalues ​​of the coupled tensor of the pixel, and select the largest eigenvalue as the wear intensity of the corresponding pixel to form a wear intensity distribution map;

[0048] The second distribution map construction unit is used to calculate the scale weight of each pixel within the three scale neighborhoods of the wear intensity distribution map, and obtain the three scale weight distribution maps.

[0049] The scoring unit is used to process the wear intensity distribution map and the three-scale weight distribution map using a multi-scale enhanced fusion network to obtain the wear score of the turbine coating.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. This invention extracts the relative color shift spectrum, color perturbation gradient, and local texture perturbation index of pixels to construct a coupling tensor. The relative color shift spectrum accurately captures the metal substrate exposure and color difference caused by coating wear; the color perturbation gradient characterizes the edge transition properties of the wear area; and the local texture perturbation index reflects the roughness changes of the wear surface. These three elements achieve comprehensive coupling of wear characteristics from three dimensions: color, gradient, and texture. By solving for the maximum eigenvalue of the coupling tensor, the wear intensity is obtained, highlighting the characteristics of the wear area. Compared to single features, this invention can more fundamentally quantify the wear degree of pixels, solving the problem of low accuracy in detecting wear on turbine coatings in existing technologies.

[0052] 2. This invention calculates the scale weight of each pixel within three scale neighborhoods to form a multi-scale weight distribution map. By allocating weights, wear characteristics at different scales can be highlighted (e.g., small-scale weights enhance the identification of microscopic local wear, while large-scale weights improve the characterization of macroscopic overall wear), avoiding the omission of local fine wear or misjudgment of the overall wear trend at a single scale, thus further improving the accuracy of detection.

[0053] 3. This invention employs a multi-scale enhanced fusion network to deeply fuse the wear intensity distribution map with the weight distribution maps of three scales, transforming multi-scale wear information into quantitative wear scores. The network can effectively integrate wear intensity information at the feature level with weight optimization information at the scale level, suppress background noise interference, enhance the response of effective wear features, and improve the accuracy of wear detection for turbine coatings. Attached Figure Description

[0054] Figure 1 A flowchart of a method for detecting wear on the coating of a water turbine.

[0055] Figure 2 This is a schematic diagram of the structure of a multi-scale enhanced fusion network;

[0056] Figure 3This is a schematic diagram of the structure of the four convolution branches. Detailed Implementation

[0057] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0058] Example 1, as Figure 1 As shown, a method for detecting wear on a turbine coating includes the following steps:

[0059] The original RGB image of the coating on the surface of the water turbine was acquired, and the relative color shift spectrum, color perturbation gradient and local texture perturbation index were extracted for each pixel to construct the coupling tensor of the pixel.

[0060] The eigenvalues ​​of the coupled tensor of each pixel are solved, and the largest eigenvalue is selected as the wear intensity of the corresponding pixel to form a wear intensity distribution map.

[0061] Within the three scale neighborhoods of the wear intensity distribution map, the scale weight of each pixel is calculated to obtain the three scale weight distribution maps;

[0062] A multi-scale enhanced fusion network was used to process the wear intensity distribution map and the weight distribution map of the three scales to obtain the wear score of the turbine coating.

[0063] Arrange the wear intensity values ​​according to their corresponding pixel positions to obtain a wear intensity distribution map; arrange the weights of the same scale according to their corresponding pixel positions to obtain a weight distribution map of the corresponding scale.

[0064] In this embodiment, the process of obtaining the relative color shift spectrum includes:

[0065] Using the pixel as the center, obtain the neighborhood range, where the neighborhood range size is 9×9;

[0066] Calculate the RGB mean of each pixel in the neighborhood range to obtain the neighborhood RGB mean;

[0067] Subtract the average RGB value of the neighborhood from the RGB value of the pixel to obtain the RGB offset value;

[0068] The relative color shift spectrum is obtained by dividing the RGB offset value by the mean RGB value of the neighborhood and taking the 2-norm.

[0069] In this embodiment, the formula for the relative color shift spectrum is:

[0070] ,

[0071] in, For the first The relative color offset spectrum of each pixel For the first The RGB values ​​of each pixel (RGB values ​​include R channel values, G channel values, and B channel values). For the first The RGB mean of the neighborhood of each pixel (the RGB mean includes the mean of the R channel, the mean of the G channel, and the mean of the B channel). To avoid parameters with a denominator of 0, This is a 2-norm operation.

[0072] Wear on turbine coatings often manifests as color differences between localized areas (such as wear points) and the surrounding normal coating. However, the RGB values ​​of a single pixel are easily affected by environmental factors such as lighting. This invention takes a neighborhood range centered on the pixel, uses the average RGB value of the neighborhood as a "normal background benchmark," and then effectively filters out the influence of global lighting fluctuations by using the difference between the pixel's own RGB value and the neighborhood average value, focusing on color shifts in localized areas (such as the difference between the color of the metal substrate and the surrounding coating after coating wear). Compared to directly using RGB values, the "relative offset" method can more accurately locate localized color anomalies caused by wear, avoiding misjudgments due to environmental interference.

[0073] In this embodiment, the process of obtaining the color perturbation gradient includes:

[0074] Calculate the ratio chromaticity for each pixel:

[0075] ,

[0076] in, For the first The ratio of color intensity per pixel, For the first The R channel value of each pixel For the first The G channel value of each pixel For the first The B channel value of each pixel To avoid parameters with a denominator of 0;

[0077] The gradient operator is used to calculate the color perturbation gradient of the ratio chromaticity of each pixel.

[0078] In this embodiment, the formula for the color perturbation gradient is:

[0079] ,

[0080] in, For the first Color perturbation gradient of each pixel For the first The horizontal gradient of the ratio of chromaticity at each pixel For the first The vertical gradient of the ratio of chromaticity at each pixel.

[0081] This invention utilizes ratio chromaticity. (By integrating the ratios of R, G, and B channels), the relative relationship of the three channels is transformed into a single feature quantity, which can more comprehensively reflect the color mode differences of pixels. Compared with single channels, this "inter-channel ratio" representation method is more in line with the color mode mutation caused by coating wear (the RGB ratio of a normal coating is stable, while the RGB ratio of a worn area will deviate significantly).

[0082] This invention uses the ratio of chromaticity of pixels as the pixel value, and calculates the horizontal and vertical gradients of each pixel using the Sobel gradient operator to obtain the color perturbation gradient of each pixel. It can accurately capture the color ratio change at the boundary of "wear area - normal area", highlight the gradient characteristics of the blurred boundary, and focus more on the color change caused by wear than directly calculating the gradient of a single RGB channel.

[0083] In this embodiment, the process of obtaining the local texture perturbation index includes:

[0084] Using the pixel as the center, obtain the neighborhood range, where the neighborhood range size is 9×9;

[0085] Calculate the RGB mean of each pixel in the neighborhood range to obtain the neighborhood RGB mean;

[0086] Calculate the standard deviation of the neighborhood range based on the difference between the RGB value of each pixel in the neighborhood range and the mean RGB value of the neighborhood range;

[0087] The ratio of the standard deviation of the neighborhood range to the mean of the neighborhood RGB values ​​is used as the local texture perturbation index.

[0088] In this embodiment, the formula for the local texture perturbation index is:

[0089] ,

[0090] ,

[0091] in, For the first Local texture perturbation index of each pixel For the first The standard deviation of the neighborhood range of each pixel For 2-norm operations, For the first Within the neighborhood of the nth pixel The RGB values ​​of each pixel For the first The average RGB value of the neighborhood of each pixel This represents the number of pixels within the neighborhood.

[0092] After the coating of a water turbine wears down, the surface will become roughened due to erosion and cavitation (e.g., the texture of the metal substrate at the point where the coating has peeled off differs from the smooth texture of the normal coating). This invention calculates the standard deviation of the RGB values ​​within the neighborhood. Quantify the degree of color fluctuation in the neighborhood area—the neighborhood color fluctuation in the normal coating area is small ( Low), with large color fluctuations in the vicinity of the wear area ( (High), thus accurately capturing the texture roughening features caused by wear.

[0093] In this embodiment, the coupling tensor is:

[0094] ,

[0095] in, For the first The coupling tensor of each pixel For the first The relative color offset spectrum of each pixel For the first Color perturbation gradient of each pixel For the first Local texture perturbation index of each pixel It is a positive integer.

[0096] This invention integrates the three dimensions of wear features—relative color shift spectrum, color perturbation gradient, and local texture perturbation index—into a symmetric matrix. Compared to using a single feature independently, this coupling method can simultaneously retain the information of each feature and also reflect the correlation between features (such as the synergistic change between color shift and texture perturbation). It comprehensively depicts the essential characteristics of wear from multiple dimensions and avoids the limitations of adapting a single feature to complex wear scenarios.

[0097] Solving for eigenvalues ​​of coupled tensors, the maximum eigenvalue can reflect the characteristic intensity of the tensor in the "most significant direction". The maximum eigenvalue can integrate the "strongest wear feature" of the three-dimensional features, highlight the core wear degree of the pixel, and make the wear intensity distribution map clearly distinguish between the "wear area" and the "normal area".

[0098] In this embodiment, the process of calculating the scale weight of each pixel includes:

[0099] The mean wear intensity is calculated for each pixel in the wear intensity distribution map within three scale neighborhoods, with the sizes of the three scale neighborhoods being 3×3, 5×5, and 7×7, respectively.

[0100] The formula for calculating the average wear intensity is:

[0101] ,

[0102] in, For the first The first pixel Average wear strength at various scales For the first The first pixel Within the neighborhood of a species scale, the first Wear intensity, For the first Scale neighborhood range For the first The number of pixels within a neighborhood of a certain scale. Take 1, 2, and 3, which correspond to 3×3, 5×5, and 7×7 respectively;

[0103] The standard deviation is calculated based on the difference between the wear intensity of each pixel and the mean wear intensity within the neighborhood of each scale.

[0104] The formula for calculating the standard deviation is:

[0105] ,

[0106] in, For the first The first pixel Standard deviation at various scales For the first The first pixel Within the neighborhood of a species scale, the first Wear intensity, For the first Scale neighborhood range For the first The number of pixels within a neighborhood of a certain scale. For the first The first pixel Average wear intensity at various scales;

[0107] Calculate the scale weight for each pixel based on the standard deviation of each scale.

[0108] In this embodiment, the formula for calculating the scale weight of each pixel is:

[0109] ,

[0110] in, For the first The first pixel Each scale weight, where exp is an exponential function. For the first Wear intensity per pixel For the first The first pixel Average wear strength at various scales For the first The first pixel Standard deviation at various scales To avoid parameters with a denominator of 0, and It is a positive integer.

[0111] In the formula ( This reflects the difference between the wear intensity of the current pixel and the mean of its corresponding scale neighborhood: if Much larger This indicates that the pixel is a "locally high-wear point" at the current scale; if near This indicates that the pixel is a "background pixel" at the current scale. This is further considered in conjunction with the standard deviation. (Quantify the overall fluctuation of wear within the neighborhood) Normalize to make the difference values ​​more comparable across scales - avoiding the interference of absolute values ​​at different scales and accurately highlighting the significant wear at each scale.

[0112] like Figure 2 As shown, the multi-scale enhancement fusion network includes: four convolutional branches, three feature enhancement branches, three CNN networks, a branch feature weighted fusion module, and a fully connected layer;

[0113] The input of the first convolutional branch is used to input the first scale weight distribution map; the input of the second convolutional branch is used to input the second scale weight distribution map; the input of the third convolutional branch is used to input the third scale weight distribution map; and the input of the fourth convolutional branch is used to input the wear intensity distribution map.

[0114] The first input of the first feature enhancement branch is connected to the output of the first convolution branch, and its output is connected to the input of the first CNN network; the first input of the second feature enhancement branch is connected to the output of the second convolution branch, and its output is connected to the input of the second CNN network; the first input of the third feature enhancement branch is connected to the output of the third convolution branch, and its output is connected to the input of the third CNN network.

[0115] The output of the fourth convolutional branch is connected to the second input of the first feature enhancement branch, the second input of the second feature enhancement branch, and the second input of the third feature enhancement branch, respectively.

[0116] The input of the branch feature weighted fusion module is connected to the output of the first CNN network, the output of the second CNN network, and the output of the third CNN network, respectively, and its output is connected to the input of the fully connected layer.

[0117] The output of the fully connected layer serves as the output of the multi-scale enhanced fusion network.

[0118] This invention designs four convolutional branches to process the "weight distribution map of three scales" and the "wear intensity distribution map" respectively. This not only preserves the feature independence of different scale weights and wear intensity (avoiding feature confusion caused by single-branch processing), but also enhances the wear intensity feature through the "cross-branch connection" of the fourth convolutional branch to the three feature enhancement branches by weight features.

[0119] Each feature enhancement branch simultaneously receives "weight features of the corresponding scale" and "wear intensity features": the weight features serve as "guiding signals" to make the network prioritize significant wear areas at the corresponding scale.

[0120] Three CNN networks perform deep feature extraction on the enhanced features, deeply mining wear features from the "weight-wear intensity" fused features. The branch feature weighted fusion module integrates the deep features of the three CNN networks with weights, adaptively adjusting the contribution of features at different scales. The fully connected layer obtains the turbine coating wear score based on the output features of the branch feature weighted fusion module.

[0121] In this embodiment, the first feature enhancement branch is used to multiply the weighted feature output by the first convolution branch with the wear intensity feature output by the fourth convolution branch element-wise to obtain the first wear intensity enhancement feature;

[0122] The second feature enhancement branch is used to multiply the weighted features output by the second convolution branch with the wear intensity features output by the fourth convolution branch element-wise to obtain the second wear intensity enhancement feature;

[0123] The third feature enhancement branch is used to multiply the weighted features output by the third convolution branch with the wear intensity features output by the fourth convolution branch element-wise to obtain the third wear intensity enhancement feature.

[0124] On the one hand, this invention utilizes the quantification results of wear significance based on the weights of each scale to amplify the intensity features of the wear area with high weights and suppress the features of the background area with low weights, effectively focusing on the core wear area at different scales and avoiding background noise interference; on the other hand, it completes the feature binding of "scale weight and wear intensity" through element multiplication, allowing small, medium and large scale branches to highlight wear information of micro-details, meso-distribution and macro-trend respectively.

[0125] In this embodiment, the expression for the branch feature weighted fusion module is:

[0126] ,

[0127] in, The output of the branch feature weighted fusion module, The output features of the first CNN network, The output features of the second CNN network, The output features of the third CNN network, This is channel attention (Se-Block).

[0128] On the one hand, the channel attention adaptively learns the importance weights of different channels in each branch feature (F1 / F2 / F3), which can accurately identify and strengthen the channel features that are more critical to wear detection, while weakening the interference of redundant or noisy channels, so that the effective features of each branch are enhanced in a targeted manner. On the other hand, by weighted summation after "feature × channel weight", the enhanced features of the three branches are fused into a unified feature X, which not only retains the core wear information (microscopic details, mesoscopic distribution, macroscopic trend) of branches at different scales, but also achieves feature fusion through the attention mechanism, avoiding the limitations of single branch features. The final output fused feature can more comprehensively and accurately characterize the wear state of the turbine coating.

[0129] like Figure 3 As shown, each of the four convolutional branches consists of a first convolutional layer, a batch normalization layer, a second convolutional layer, and a ReLU layer connected in sequence. The kernel size of the first and second convolutional layers is 3×3.

[0130] Example 2: A water turbine coating wear detection system, comprising: a coupling tensor construction unit, a first distribution map construction unit, a second distribution map construction unit, and a scoring unit;

[0131] The coupling tensor construction unit is used to acquire the original RGB image of the coating on the surface of the water turbine, extract the relative color shift spectrum, color perturbation gradient and local texture perturbation index for each pixel, and construct the coupling tensor of the pixel;

[0132] The first distribution map construction unit is used to solve for the eigenvalues ​​of the coupled tensor of the pixel, and select the largest eigenvalue as the wear intensity of the corresponding pixel to form a wear intensity distribution map;

[0133] The second distribution map construction unit is used to calculate the scale weight of each pixel within the three scale neighborhoods of the wear intensity distribution map, and obtain the three scale weight distribution maps.

[0134] The scoring unit is used to process the wear intensity distribution map and the three-scale weight distribution map using a multi-scale enhanced fusion network to obtain the wear score of the turbine coating.

[0135] The specific implementation method of Example 2 is the same as that of Example 1.

[0136] In this embodiment, 0-3 points: slight wear. The coating surface has only minor local color / texture changes, with no obvious exposure of the metal substrate. It has no significant impact on the reliability and energy efficiency of the turbine operation, does not require immediate maintenance, and can be included in routine inspections.

[0137] 4-6 points: Moderate wear. The coating shows localized flaky wear, with a small amount of exposed metal substrate, which may be accompanied by a slight risk of corrosion and will slightly reduce the turbine's energy efficiency. It is recommended to repair the coating during the next planned maintenance.

[0138] 7-10 points: Severe wear. Large areas of the coating have peeled off, exposing a large portion of the metal substrate. Obvious signs of corrosion or fatigue cracks have appeared, which will significantly reduce the reliability and energy efficiency of the turbine. It is necessary to shut down the turbine immediately for maintenance to prevent the equipment failure from escalating.

[0139] This invention extracts the relative color shift spectrum, color perturbation gradient, and local texture perturbation index of pixels to construct a coupling tensor. The relative color shift spectrum accurately captures the metal substrate exposure and color difference caused by coating wear; the color perturbation gradient characterizes the edge transition properties of the wear area; and the local texture perturbation index reflects the roughness changes of the wear surface. These three elements achieve comprehensive coupling of wear characteristics from three dimensions: color, gradient, and texture. By solving for the maximum eigenvalue of the coupling tensor, the wear intensity is obtained, highlighting the characteristics of the wear area. Compared to single features, this method more fundamentally quantifies the wear degree of pixels, solving the problem of low accuracy in detecting wear on turbine coatings in existing technologies.

[0140] This invention calculates the scale weight of each pixel within three scale neighborhoods to form a multi-scale weight distribution map. By allocating weights, wear characteristics at different scales can be highlighted (e.g., small-scale weights enhance the identification of microscopic local wear, while large-scale weights improve the characterization of macroscopic overall wear), avoiding the omission of local fine wear or misjudgment of the overall wear trend at a single scale, thus further improving the accuracy of detection.

[0141] This invention employs a multi-scale enhanced fusion network to deeply fuse the wear intensity distribution map with the weight distribution maps of three scales, transforming multi-scale wear information into a quantitative wear score. The network can effectively integrate wear intensity information at the feature level with weight optimization information at the scale level, suppress background noise interference, enhance the response of effective wear features, and improve the accuracy of wear detection for turbine coatings.

[0142] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of detecting wear of a coating of a hydraulic turbine, characterized in that, The method comprises the following steps: The original RGB image of the surface coating of the water turbine is collected, the relative color offset spectrum, the color disturbance gradient and the local texture disturbance index of each pixel point are extracted, and a coupling tensor of the pixel point is constructed; The color disturbance gradient acquisition process comprises: The ratio color is calculated for each pixel point: ， wherein, is the ratio chrominance of the th pixel point, is the R channel value of the th pixel point, is the G channel value of the th pixel point, is the B channel value of the th pixel point, is a parameter to avoid denominator zero; The color disturbance gradient of the ratio color of each pixel point is calculated through a gradient operator; The coupling tensor is: ， wherein, is the coupling tensor for the jth pixel point, is the relative color shift spectrum for the jth pixel point, is the color perturbation gradient for the jth pixel point, is the local texture perturbation index for the jth pixel point, is the color perturbation gradient for the jth pixel point, is the local texture perturbation index for the jth pixel point, is the local texture perturbation index for the jth pixel point, is the local texture perturbation index for the jth pixel point, is a positive integer; The eigenvalues of the coupling tensor of the pixel point are solved, the maximum eigenvalue is selected as the wear intensity of the corresponding pixel point, and a wear intensity distribution map is constructed; In the three scale neighborhood ranges of the wear intensity distribution map, the scale weight of each pixel point is calculated to obtain three scale weight distribution maps; The wear intensity distribution map and the three scale weight distribution maps are processed by using a multi-scale enhancement fusion network to obtain a water turbine coating wear score.

2. The method of claim 1, wherein, The relative color offset spectrum acquisition process comprises: The neighborhood range is obtained with the pixel point in the original RGB image as the center; The RGB mean values of the pixel points in the neighborhood range are calculated to obtain neighborhood RGB mean values; The RGB values of the pixel points in the original RGB image are subtracted from the neighborhood RGB mean values to obtain RGB offset values; The RGB offset values are divided by the neighborhood RGB mean values, and the 2-norm is taken to obtain the relative color offset spectrum.

3. The method of claim 1, wherein, The local texture disturbance index acquisition process comprises: The neighborhood range is obtained with the pixel point in the original RGB image as the center; The RGB mean values of the pixel points in the neighborhood range are calculated to obtain neighborhood RGB mean values; The standard deviation of the neighborhood range is calculated according to the difference between the RGB values of each pixel point in the neighborhood range and the neighborhood RGB mean values; The ratio of the standard deviation of the neighborhood range to the neighborhood RGB mean values is taken as the local texture disturbance index.

4. The method of claim 1, wherein, The process of calculating the scale weight of each pixel point comprises: The wear intensity mean values in the three scale neighborhood ranges of each pixel point of the wear intensity distribution map are calculated, and the sizes of the three scale neighborhood ranges are 3*3, 5*5 and 7*7 respectively; The standard deviations are calculated according to the difference between the wear intensity of each pixel point in each scale neighborhood range and the wear intensity mean value; The scale weight of each pixel point is calculated according to the standard deviation of each scale.

5. The method of claim 1 or 4, wherein, The formula for calculating the scale weight of each pixel point is: ， in, For the first The first pixel Each scale weight, where exp is an exponential function. For the first Wear intensity per pixel For the first The first pixel Average wear strength at various scales For the first The first pixel Standard deviation at various scales To avoid parameters with a denominator of 0, and It is a positive integer.

6. The method of claim 1, wherein, The multi-scale enhancement fusion network comprises four convolution branches, three feature enhancement branches, three CNN networks, a branch feature weighted fusion module and a full connection layer; The input end of the first convolution branch is used for inputting the first scale weight distribution map; the input end of the second convolution branch is used for inputting the second scale weight distribution map; the input end of the third convolution branch is used for inputting the third scale weight distribution map; and the input end of the fourth convolution branch is used for inputting the wear intensity distribution map; The first input end of the first feature enhancement branch is connected with the output end of the first convolution branch, and the output end thereof is connected with the input end of the first CNN network; the first input end of the second feature enhancement branch is connected with the output end of the second convolution branch, and the output end thereof is connected with the input end of the second CNN network; and the first input end of the third feature enhancement branch is connected with the output end of the third convolution branch, and the output end thereof is connected with the input end of the third CNN network; The output ends of the fourth convolution branch are connected with the second input ends of the first, second and third feature enhancement branches respectively; The input ends of the branch feature weighting and fusion module are connected with the output ends of the first, second and third CNN networks respectively, and the output end thereof is connected with the input end of the full connection layer; The output end of the full connection layer is the output end of the multi-scale enhancement fusion network.

7. The method of claim 6, wherein, The first feature enhancement branch is configured to multiply the weight features output by the first convolution branch with the abrasion intensity features output by the fourth convolution branch to obtain first abrasion intensity enhancement features; The second feature enhancement branch is configured to multiply the weight features output by the second convolution branch with the abrasion intensity features output by the fourth convolution branch to obtain second abrasion intensity enhancement features; The third feature enhancement branch is configured to multiply the weight features output by the third convolution branch with the abrasion intensity features output by the fourth convolution branch to obtain third abrasion intensity enhancement features.

8. A system for detecting wear of a coating of a hydraulic turbine, implemented on the basis of the method for detecting wear of a coating of a hydraulic turbine according to any one of claims 1 to 7, characterized in that, The method comprises the following steps: a coupling tensor construction unit, a first distribution map construction unit, a second distribution map construction unit and a scoring unit; The coupling tensor construction unit is configured to collect original RGB images of the surface coating of the water turbine, extract a relative color offset spectrum, a color disturbance gradient and a local texture disturbance index for each pixel point, and construct a coupling tensor of the pixel point; The first distribution map construction unit is configured to solve eigenvalues of the coupling tensor of the pixel point, and select a maximum eigenvalue as an abrasion intensity of the corresponding pixel point to form an abrasion intensity distribution map; The second distribution map construction unit is configured to calculate a scale weight of each pixel point in a three-scale neighborhood range of the abrasion intensity distribution map to obtain three scale weight distribution maps; The scoring unit is configured to process the abrasion intensity distribution map and the three scale weight distribution maps by using the multi-scale enhancement fusion network to obtain a water turbine coating abrasion score.

Images