A method, system, electronic device and product for monitoring the state of a water turbine without stopping

Abstract

The present application belongs to the technical field of water turbine state monitoring, and aims to provide a water turbine non-stop state monitoring method, system, electronic device and product. The present application collects local runner images at multiple different rotation angles during the operation of the water turbine, and combines image feature extraction, image registration and splicing and other technologies to reconstruct the complete current runner image of the water turbine, thereby realizing full coverage and high-precision monitoring of the health state of the surface of the runner of the water turbine. Based on the present application, the surface damage and structural deformation and other health state information of the runner of the water turbine can be quickly identified on the basis of no sensor contact and strong anti-interference capability, thereby avoiding the problems of poor environmental adaptability, low positioning accuracy and high maintenance cost existing in the prior art.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of water turbine state monitoring, and particularly relates to a water turbine non-stop state monitoring method, system, electronic device and product. BACKGROUND

[0002] The water turbine is generally installed in a water conservancy hub such as a river or a reservoir, and has a complex operating environment and is affected by various factors such as water flow, water pressure and water quality. At the same time, the water turbine bears huge mechanical stress and hydraulic impact during operation, so that the water turbine is prone to various faults. In the prior art, the runner of the water turbine is comprehensively checked by means of regular shutdown maintenance. However, in the early stage of a fault, the operating state of the water turbine changes relatively weakly, and the fault signal is hidden in a large amount of background noise and is difficult to be directly detected. Therefore, how to use advanced signal processing and feature extraction technology to capture these weak fault signals from complex monitoring data is the key to realizing early warning of water turbine faults.

[0003] Non-stop state monitoring can effectively prevent potential faults, reduce maintenance costs, reduce downtime, and improve economic benefits. At present, there are the following non-stop state monitoring methods for water turbines: 1) laser monitoring technology, based on the characteristics of laser, using laser ranging, interference and other principles, the position, shape and other parameters of the surface of the runner of the water turbine are determined by measuring the time difference between laser emission and reflection, interference fringe changes and the like, and then the size, deformation and other information of the runner are obtained; 2) vibration monitoring technology, an acceleration sensor is installed at a key position of the runner, when the runner vibrates, the mass block in the sensor generates a force proportional to the vibration acceleration, the force is converted into an electric signal through a conversion element, so that the vibration characteristics of the runner are obtained; 3) stress and strain monitoring technology, based on the resistance strain effect of metal conductors, when the strain sheet is pasted on the surface of the runner and deforms with the runner, the resistance value of the strain sheet will change in proportion to the strain, and the strain and stress of the runner are calculated by measuring the resistance value change; 4) acoustic monitoring technology, using the characteristics that ultrasonic waves will be reflected, refracted and scattered when they encounter different medium interfaces in the medium, the information such as whether there is a defect in the runner, the position and size of the defect and the like are determined by analyzing the received ultrasonic signals.

[0004] However, in the process of using the prior art, the inventors found that at least the following problems exist in the prior art:

[0005] The effective range of laser measurement is usually short, and for large water turbine runners, multiple movements of the measuring device or the addition of measuring points may be required to obtain complete information, which is complex and time-consuming. The vibration monitoring technology can only obtain overall vibration information, and it is difficult to accurately determine the vibration source and the specific location of the fault for a runner with a complex structure. Stress and strain monitoring mainly measures the strain on the surface of the runner, and the internal stress distribution cannot be directly obtained, which needs to be inferred through complex calculation and assumption. Ultrasonic waves may not be able to effectively detect some defects in shape and direction, such as cracks parallel to the direction of ultrasonic wave propagation, and the detection sensitivity is low. Therefore, the existing monitoring technology has the problem of limited defect monitoring range and type, and it is difficult to achieve all-around monitoring of the water turbine. SUMMARY

[0006] The present application aims to at least partially solve the above technical problems, and provides a water turbine non-stop state monitoring method, system, electronic device and product.

[0007] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0008] In a first aspect, the present application provides a water turbine non-stop state monitoring method, comprising:

[0009] obtaining a plurality of local runner images at different rotation angles during the operation of the water turbine;

[0010] respectively performing feature extraction processing on each local runner image to obtain a feature descriptor of each local runner image;

[0011] performing feature matching on all adjacent local runner image combinations in terms of collection time according to the feature descriptors of the local runner images to obtain a matching point set;

[0012] obtaining a perspective transformation matrix corresponding to all adjacent local runner image combinations in terms of collection time according to the matching point set;

[0013] performing coordinate transformation processing on each local runner image according to the perspective transformation matrix corresponding to all adjacent local runner image combinations in terms of collection time to obtain a plurality of coordinate-transformed local runner images in the same coordinate system;

[0014] performing image stitching processing on the plurality of coordinate-transformed local runner images to obtain a complete current runner image of the water turbine;

[0015] obtaining health status information of the water turbine according to the current runner image.

[0016] In one possible design, the feature extraction processing on any local runner image to obtain a feature descriptor of the any local runner image comprises:

[0017] Convolving any one of the local turbine images by a Gaussian filter of different scales to generate a scale space;

[0018] Calculating a Hessian matrix of each pixel point in any one of the local turbine images in the scale space;

[0019] Obtaining a determinant response value of the Hessian matrix of each pixel point in any one of the local turbine images, and screening a candidate feature point from all pixel points in any one of the local turbine images based on the determinant response value of the Hessian matrix of each pixel point;

[0020] Using a quasi-Newton method to perform sub-pixel level optimization processing on the candidate feature point to obtain a final image feature point corresponding to the candidate feature point;

[0021] Generating a feature descriptor of the image feature point.

[0022] In one possible design, using a quasi-Newton method to perform sub-pixel level optimization processing on any one of the candidate feature points to obtain an image feature point corresponding to any one of the candidate feature points, comprising:

[0023] Obtaining an approximate matrix of the Hessian matrix of any one of the candidate feature points;

[0024] Updating the position and scale of the candidate feature point through iteration to obtain an updated feature point, so as to maximize the determinant response value of the approximate matrix corresponding to the updated feature point, and obtain a maximum value of the determinant response value of the current approximate matrix;

[0025] Taking the updated feature point corresponding to the maximum value of the determinant response value of the current approximate matrix as the image feature point corresponding to any one of the candidate feature points.

[0026] In one possible design, according to the feature descriptors of the local turbine images, performing feature matching on all combinations of local turbine images adjacent in time to obtain a matching point set, comprising:

[0027] Calculating the Euclidean distance between any two feature descriptors in the combination of local turbine images adjacent in time, and taking the image feature points corresponding to the feature descriptor pair with the closest Euclidean distance as the initial matching points of the current two local turbine images, and the initial matching points of all combinations of local turbine images adjacent in time constitute an initial matching point set;

[0028] Using a RANSAC algorithm to remove the mismatched points in the initial matching point set to obtain a final matching point set.

[0029] In one possible design, obtaining health status information of the water turbine according to the current turbine image, comprising:

[0030] perform feature extraction processing on the current runner image by using a SURF algorithm to obtain local feature information of the current runner image;

[0031] input the current runner image into a pre-trained feature extraction model to obtain global feature information of the current runner image;

[0032] fuse the local feature information and the global feature information of the current runner image to obtain fused feature information of the current runner image;

[0033] input the fused feature information of the current runner image into a pre-trained health state evaluation model to obtain health state information of the water turbine.

[0034] In one possible design, the health state evaluation model adopts a convolutional neural network.

[0035] In one possible design, after obtaining the fused feature information of the current runner image, the method further includes:

[0036] obtain a factory runner image of the water turbine, and perform feature extraction processing on the factory runner image by using a SURF algorithm to obtain local feature information of the factory runner image;

[0037] input the factory runner image into the feature extraction model to obtain global feature information of the factory runner image;

[0038] fuse the local feature information and the global feature information of the factory runner image to obtain fused feature information of the factory runner image;

[0039] perform difference analysis processing on the fused feature information of the current runner image and the fused feature information of the factory runner image to obtain a feature difference degree between the current runner image and the factory runner image;

[0040] determine whether the feature difference degree is greater than a preset feature difference threshold;

[0041] If so, input the fused feature information of the current runner image into the health state evaluation model to obtain the health state information of the water turbine.

[0042] In a second aspect, the present application provides a water turbine non-stop state monitoring system, comprising:

[0043] an image acquisition module configured to obtain a plurality of local runner images at different rotation angles during operation of the water turbine;

[0044] The feature extraction module is in communication connection with the image acquisition module, and is configured to perform feature extraction processing on each local runner image respectively to obtain a feature descriptor of each local runner image.

[0045] The feature matching module is in communication connection with the feature extraction module, and is configured to perform feature matching on all combinations of adjacent local runner images in acquisition time according to the feature descriptor of each local runner image to obtain a matching point set.

[0046] The image transformation module is in communication connection with the feature matching module, and is configured to obtain a perspective transformation matrix corresponding to each combination of adjacent local runner images in acquisition time according to the matching point set, and to perform coordinate transformation processing on each local runner image according to the perspective transformation matrix corresponding to each combination of adjacent local runner images in acquisition time to obtain a plurality of coordinate-transformed local runner images in the same coordinate system.

[0047] The image stitching module is in communication connection with the image transformation module, and is configured to perform image stitching processing on the plurality of coordinate-transformed local runner images to obtain a complete current runner image of the water turbine.

[0048] The health state detection module is in communication connection with the image stitching module, and is configured to obtain health state information of the water turbine according to the current runner image.

[0049] In a third aspect, the present application provides an electronic device, comprising:

[0050] a memory configured to store computer program instructions; and

[0051] a processor configured to execute the computer program instructions to complete the operation of the water turbine non-stop state monitoring method according to any one of the above aspects.

[0052] In a fourth aspect, the present application provides a computer program product comprising computer programs or instructions, which, when executed by a computer, implement the water turbine non-stop state monitoring method according to any one of the above aspects.

[0053] The present application has the following beneficial effects:

[0054] The application discloses a water turbine non-stop state monitoring method and system, an electronic device and a product, and can realize full coverage and high-precision monitoring of the health state of a water turbine runner surface. Specifically, in the implementation process, the application collects local runner images at multiple different rotation angles during the operation of the water turbine, and combines image feature extraction, image registration and splicing technologies to reconstruct a complete current runner image of the water turbine, so that full coverage and high-precision monitoring of the health state of the water turbine runner surface can be realized. Based on the application, the surface damage and structural deformation of the water turbine runner and other health state information can be quickly identified on the basis of no sensor contact and strong anti-interference capability, and the problems of poor environmental adaptability, low positioning accuracy and high maintenance cost in the prior art are avoided, so that the application has good application value and promotion prospect.

[0055] Other benefits of the application will be further described in the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of the water turbine non-stop state monitoring method in the embodiment;

[0057] Figure 2 is a module block diagram of the water turbine non-stop state monitoring method system in the embodiment;

[0058] Figure 3 is a module block diagram of the electronic device in the embodiment. DETAILED DESCRIPTION

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the application will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the drawings is only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings. It should be noted that the description of these embodiments is used to help understand the application, but does not constitute a limitation on the application.

[0060] Embodiment 1

[0061] The embodiment discloses a water turbine non-stop state monitoring method, which can be executed by a computer device or a virtual machine with certain computing resources, such as an electronic device such as a personal computer, a smart phone, a personal digital assistant or a wearable device, or a virtual machine.

[0062] As shown in Figure 1 , a water turbine non-stop state monitoring method can include the following steps:

[0063] S1. Obtain multiple local runner images at different rotation angles during the operation of the water turbine. It should be noted that in the embodiment, a camera combined with a stroboscope is used to collect the local runner images, wherein the camera and the stroboscope are arranged in the transparent top cover of the water turbine and face the runner, and the stroboscope emits short pulse light at a frequency that is the same as or an integer multiple of the rotation speed of the runner of the water turbine. The camera needs to have a frame rate synchronized with the stroboscope. When the stroboscope emits light, a trigger signal is sent to the camera to control it to collect images at the moment of flashing light, so as to ensure that the exposure is synchronized with the flashing light. The setting of the stroboscope enables a low-frame-rate camera to achieve high-precision imaging, greatly reduces the cost of image collection hardware, and avoids the problem of massive data caused by high-speed shooting.

[0064] Specifically, the initially collected initial local runner image is pre-processed as follows:

[0065] The initial local runner image is subjected to a grayscale processing to convert the color image into a grayscale image, so as to simplify subsequent calculation and obtain a grayscale-processed local runner image.

[0066] The grayscale-processed local runner image is subjected to a noise reduction processing to obtain a noise-reduced local runner image. It should be noted that due to electromagnetic interference, equipment noise and other factors in the monitoring environment, the collected image can have noise. Therefore, in the embodiment, a Gaussian filter, a mean filter or a median filter is further used to remove the noise interference in the grayscale-processed local runner image, so as to improve the image quality and provide a clear image basis for subsequent analysis.

[0067] The noise-reduced local runner image is subjected to a geometric correction processing to obtain a pre-processed local runner image. It should be noted that due to the shooting angle, lens distortion and other reasons, the image can have geometric distortion. Through the geometric correction processing of the image, the shape and position of the runner can be accurately presented in the image, which is convenient for subsequent size measurement and feature analysis.

[0068] S2. Perform feature extraction processing on each local runner image to obtain a feature descriptor of each local runner image.

[0069] In step S2, a SURF (Speeded-Up Robust Features) algorithm is used for feature descriptor matching. Specifically, feature extraction processing is performed on any local runner image to obtain a feature descriptor of the any local runner image, including:

[0070] S201. Perform convolution operation on the any local runner image through a Gaussian filter of different scales to generate a scale space; wherein the scale space is:

[0071] ;

[0072] wherein, represents a function expression of a Gaussian filter with a scale of σ , represents a convolution operator, represents the any local rotation image, represents a pixel point coordinate of the any local rotation image in a planar rectangular coordinate system.

[0073] In this embodiment, the convolution operation of the any local rotation image by the Gaussian filter with different scales can realize multi-scale feature point detection, ensure that the features can be accurately extracted under different scaling conditions, and further make the subsequently obtained image feature points have scale invariance.

[0074] S202. Calculate a Hessian matrix of each pixel point in the any local rotation image in the scale space; wherein the Hessian matrix of the any local rotation image with a scale of σ is:

[0075] ;

[0076] wherein, represents a Gaussian second-order differential convolution result of the any local rotation image, represents a Gaussian second-order differential convolution result of the any local rotation image, represents a Gaussian second-order differential convolution result of the any local rotation image, represents a Gaussian second-order differential convolution result of the any local rotation image.

[0077] S203. Obtain a determinant response value of the Hessian matrix of each pixel point in the any local rotation image, and obtain a candidate feature point from all pixel points of the any local rotation image based on the determinant response value of the Hessian matrix of each pixel point; wherein the determinant response value of the Hessian matrix of any pixel point is:

[0078] ;

[0079] wherein, ω represents a preset weight value, generally set as 0.9, and used for balancing the approximation error.

[0080] ​Specifically, in the process of screening candidate feature points from all pixel points of the any local rotation wheel image, the greater the determinant response value corresponding to any pixel point in the any local rotation wheel image, the more obvious the image structure change intensity, and the more likely the pixel point is a feature point such as a corner point or an edge intersection point. In this embodiment, by non-maximum suppression, the determinant response value of the any pixel point is compared with the determinant response values of 26 neighboring points (8 neighboring points of the same scale + 9 neighboring points of each of the upper and lower adjacent scales) of the adjacent scale, and the neighboring point with the local maximum determinant response value is reserved as a candidate feature point. For example, if the determinant response value of the any pixel point is the maximum value, the any pixel point is set as a candidate feature point.

[0081] S204. Adopting the quasi-Newton method to perform sub-pixel level optimization processing on the candidate feature points to obtain final image feature points corresponding to the candidate feature points;

[0082] In this embodiment, the quasi-Newton method is adopted to perform sub-pixel level optimization processing on any candidate feature point to obtain an image feature point corresponding to the any candidate feature point, including:

[0083] S2041. Obtaining an approximate matrix of the Hessian matrix of the any candidate feature point;

[0084] S2042. Updating the position and scale of the candidate feature point through iteration to obtain an updated feature point, so as to maximize the determinant response value of the approximate matrix corresponding to the updated feature point, and obtain the maximum value of the determinant response value of the current approximate matrix;

[0085] S2043. Taking the updated feature point corresponding to the maximum value of the determinant response value of the current approximate matrix as the image feature point corresponding to the any candidate feature point.

[0086] It should be noted that the existing SURF algorithm adopts the Newton method iteration to obtain an image feature point from a candidate feature point in the feature point positioning stage, can determine the accurate position and optimal scale of the image feature point, and realizes the sub-pixel level precision positioning of the image feature point. The biggest advantage of the Newton method is that the iterative algorithm has a second-order convergence speed, but a non-positive definite Hessian matrix cannot guarantee that the iterative direction of the algorithm is the gradient direction of the objective function f ( x , y ,σ)=∣det(H( x , y ,σ))∣in the feature point position and scale x , yThe descent direction at (σ). Furthermore, when using Newton's method, the Hessian matrix needs to be calculated in each iteration. For multivariable problems, the massive computational load will consume a significant amount of time and computer resources, and the Hessian matrix may become singular during iterations, making the algorithm difficult to converge. Therefore, this embodiment uses a quasi-Newton method to obtain image feature points, employing an approximate matrix to replace the Hessian matrix. This retains the advantage of Newton's method's second-order convergence speed while resolving the drawback of needing to calculate the Hessian matrix in each iteration.

[0087] Specifically, in this embodiment, the following is adopted: x k Indicates the first k The position and scale of the feature points in the next iteration, i.e. x k =( u k , v k , σ k ), ( u k , v k ,) is the first k The two-dimensional coordinates of the feature point in the uv coordinate system at the next iteration (with sub-pixel precision). σ k For the first k In the next iteration, the scale of the feature points (reflecting the detection scale of the feature points) is used to initialize the approximate matrix of the Hessian matrix of the candidate feature points: B 0= I (Identity matrix), assuming the initial scale is 1 and the coordinates are the coarse localization results of the candidate feature points, the derivation process of the approximate matrix iteratively updating the position and scale of the candidate feature points is as follows:

[0088] such as objective function f ( x’ If the objective function is twice continuously differentiable on its domain, then the objective function can be obtained. f ( x’ ) in the k Position and scale of feature points at +1 iteration x k+1 The quadratic approximation model at that point is Equation (1).

[0089] (1)

[0090] In equation (1), f ( x k+1 ) indicates the first kthe objective function at the (k+1)th iteration, g k+1 denotes the objective function f ( x’ ) the position and scale of the feature point at the (k+1)th iteration k x k+1 the gradient at the (k+1)th iteration, x’ denotes the current position and scale of the feature point, T denotes the transpose symbol, G k+1 denotes the objective function f ( x’ ) the position and scale of the feature point at the (k+1)th iteration k x k+1 the Hessian matrix at the (k+1)th iteration.

[0091] Differentiate equation (1) with respect to x:

[0092] (2)

[0093] Let the gradient change be , and the feature point displacement be , then we have:

[0094] (3)

[0095] Replace B k+1 with the matrix G k+1 which approximates the Hessian matrix to obtain equation (4), which is commonly referred to as the quasi-Newton condition:

[0096] (4)

[0097] Generally B k has the following three characteristics.

[0098] (1) In a certain sense, B k ≈ G k , so that the direction generated by the algorithm approximates the Newton method, but ensures that the algorithm has a faster convergence speed.

[0099] (2) For iteration number k =1, 2, 3,..., only if the approximate matrix B k is positive definite, can the descent direction obtained by iteration be guaranteed to be the direction in which the objective function f ( x ) decreases most rapidly at the (k+1)th iteration.​​x k The downward direction is consistent at each location.

[0100] (3) Approximate matrix B k A rank-1 or rank-2 matrix is ​​usually used for correction, and the choice can be made after comparing the correction effects of each.

[0101] (5)

[0102] In equation (5), E k It is a rank-1 or rank-2 matrix. If E k Take a rank-1 matrix, then ,in, a Represents a predefined positive real number. u k Describes the rank-1 correction matrix. .

[0103] By quasi-Newtonian conditions That is, we have equation (6):

[0104] (6)

[0105] Equation (6) shows u k Collinear vectors That is, there exists a constant. β Make That is, we have equation (7):

[0106] (7)

[0107] In equation (7), B s Indicates the first s The approximate matrix of the Hessian matrix at the nth iteration. k s Indicates the relationship with the first s The constant term related to the next iteration.

[0108] Then we have equation (8).

[0109] (8)

[0110] Therefore, if Desirable That is, we have equation (9):

[0111] , (9)

[0112] Therefore, the approximate matrix is ​​obtained.B k The formula for rank-1 correction is as follows (10):

[0113] (10)

[0114] Equation (10) is the approximate matrix. B k An iterative derivation is used to replace the complex Hessian matrix, reducing computational cost. Furthermore, in noisy regions of the rotating image, the true Hessian matrix may exhibit negative eigenvalues ​​due to local interference, while the approximate matrix... B k By maintaining positive definiteness through correction, iterative stable convergence is ensured.

[0115] S205. Generate feature descriptors for image feature points. Specifically, in the process of generating feature descriptors, first, taking the image feature point as the center, calculate its radius as 6. σ The Haar wavelet features of pixels in the neighborhood are obtained in the horizontal and vertical directions. These response values ​​are then assigned Gaussian weights. Next, the responses within a 60° range are accumulated to form a new vector. Finally, the entire circular region is traversed, and the direction of the longest vector is selected as the principal direction of the image feature point. Subsequently, the coordinate axes are rotated to the principal direction, centered on the image feature point, and a side length of 20 is selected according to the principal direction. σ The square region is divided into 4×4 sub-regions, and the value of each sub-region is calculated. σ ×5 σ The horizontal and vertical Haar wavelet features of the sampling points are obtained, and the Haar wavelet features of each sub-region and their absolute values ​​are added together to form the feature vector of each sub-region. For each image feature point, a total of 64-dimensional feature vectors can be generated. These 64-dimensional feature vectors are the feature descriptors of the image feature points.

[0116] It should be noted that rotating image capture suffers from issues such as scale, rotation, lighting variations, and noise interference. The SURF algorithm, however, offers fast feature descriptor extraction, enabling rapid processing of large numbers of rotating images. Its robustness ensures stable feature point extraction and descriptor generation even under complex conditions, facilitating accurate feature matching in the subsequent process. In this embodiment, during the selection of image feature points, a quasi-Newton method is employed to obtain image feature points, using an approximate matrix instead of the Hessian matrix. This retains the advantage of Newton's second-order convergence speed while addressing the drawback of needing to calculate the Hessian matrix in each iteration.

[0117] S3. Based on the feature descriptors of each local rotary image, perform feature matching on all combinations of local rotary images acquired at adjacent times to obtain a set of matching points.

[0118] In step S3, based on the feature descriptors of each local rotary image, feature matching is performed on all combinations of local rotary images acquired at adjacent times to obtain a set of matching points, including:

[0119] S301. Calculate the Euclidean distance between each feature descriptor in a combination of local rotary images with adjacent acquisition times, and take the image feature point corresponding to the feature descriptor with the closest Euclidean distance as the initial matching point of the current two local rotary images. The initial matching points of all combinations of local rotary images with adjacent acquisition times constitute the initial matching point set.

[0120] S302. To improve matching accuracy, the RANSAC algorithm is used to remove mismatched points from the initial matching point set, resulting in the final matching point set. Specifically, a set of initial matching points corresponding to any two adjacent local rotating images is randomly selected and assumed to be a correct match. Then, the perspective transformation model between this set of initial matching points is calculated. This perspective transformation model is then used to verify the initial matching points of other sets. Initial matching points that conform to the calculation rules of the perspective transformation model are taken as interior points, and those that do not conform are taken as exterior points. After multiple iterations, an accurate matching point set is obtained.

[0121] In this embodiment, the Euclidean distance between the descriptors of each feature point in two adjacent local wheel images is calculated, and the most similar descriptor pair is found to establish the matching relationship between the two adjacent local wheel images, so as to achieve accurate matching between the two adjacent local wheel images and provide a reliable basis for subsequent image stitching and health status detection.

[0122] S4. Based on the set of matching points, obtain the perspective transformation matrix corresponding to all combinations of local rotating images with adjacent acquisition times. Specifically, for any combination of local rotating images with adjacent acquisition times, the perspective transformation matrix between the two local rotating images in the current combination is obtained by using the least squares method based on the coordinate information of their matching points. This perspective transformation matrix can determine the transformation parameters such as rotation, scaling, and translation between the two local rotating images, so that their matching points can coincide as much as possible after image transformation.

[0123] S5. Based on the perspective transformation matrix corresponding to the combination of local wheel images acquired at adjacent times, coordinate transformation processing is performed on each local wheel image to obtain multiple coordinate-transformed local wheel images located in the same coordinate system. Specifically, in this embodiment, the cv2.warpPerspective function in Python's OpenCV library is used to transform all local wheel images to the same coordinate system for subsequent image stitching processing.

[0124] S6. Perform image stitching on the local runner images after multiple coordinate transformations to obtain a complete current runner image of the turbine. It should be noted that during the image stitching process, for pixels in overlapping areas, a weighted average is calculated based on their distance to the image center or other weighting functions, and this average is used as the pixel value of that pixel in the fused current runner image.

[0125] S7. Obtain the health status information of the turbine based on the current runner image.

[0126] In step S7, the health status information of the turbine is obtained based on the current runner image, including:

[0127] S701. The SURF algorithm is used to perform feature extraction processing on the current rotating image to obtain the local feature information of the current rotating image;

[0128] S702. Input the current rotating image into a pre-trained feature extraction model to obtain the global feature information of the current rotating image. In this embodiment, the current rotating image can also be pre-screened for feature points, retaining feature points with high response intensity, and then inputting the screened feature points into the feature extraction model. These strong response feature points are more representative of the key features of the image, which helps to reduce the computational load of the subsequent feature extraction model.

[0129] S703. The local feature information and global feature information of the current rotating image are fused to obtain the fused feature information of the current rotating image;

[0130] It should be noted that the SURF algorithm excels at capturing local details (such as feature points of blade edge cracks and wear pits) and is insensitive to noise, but it cannot characterize global structures (such as overall blade distortion) and fails in smooth curved surfaces (without significant feature points). Feature extraction models, on the other hand, can automatically learn global features (such as the curvature distribution of the blade surface and wear patterns formed by water flow impact), but their sensitivity to local minor defects (such as early microcracks) is lower than that of traditional algorithms. In this embodiment, by fusing the local feature information obtained based on the SURF algorithm and the global feature information obtained through the feature extraction model, complementary local and global features can be achieved, filling the blind spots of single algorithms. In complex curved surface scenarios of turbines, feature fusion can increase the recall rate of defect detection from 75% to 92%, avoiding missed detections due to single features.

[0131] In step S703, after obtaining the fused feature information of the current rotary image, the method further includes:

[0132] a1. Obtain the factory runner image of the water turbine, and use the SURF algorithm to perform feature extraction processing on the factory runner image to obtain the local feature information of the factory runner image;

[0133] a2. Input the factory-exported rotor image into the feature extraction model to obtain the global feature information of the factory-exported rotor image;

[0134] a3. The local feature information and global feature information of the factory-outgoing wheel image are fused to obtain the fused feature information of the factory-outgoing wheel image;

[0135] a4. Perform difference analysis on the fused feature information of the current rotary image and the fused feature information of the factory rotary image to obtain the feature difference degree between the current rotary image and the factory rotary image;

[0136] a5. Determine whether the feature difference degree is greater than a preset feature difference threshold;

[0137] a6. If so, the fused feature information of the current turbine image is input into the health status assessment model to obtain the health status information of the turbine.

[0138] It should be noted that before conducting a health status assessment of the turbine, the current turbine image is compared with the turbine image before it leaves the factory. If the difference in features is large, the health status assessment is then conducted using the health status assessment model. This can overcome the limitations of using a single health status assessment model for health status assessment.

[0139] S704. Input the fused feature information of the current turbine runner image into the pre-trained health status assessment model to obtain the health status information of the turbine.

[0140] In this embodiment, the health status assessment model uses a convolutional neural network (CNN).

[0141] It's important to note that a convolutional neural network (CNN) is a deep learning network specifically designed for processing data with a grid structure (such as images). It primarily consists of convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs convolution operations by sliding multiple convolutional kernels with different parameters across the image. Each kernel acts as a feature detector, extracting features from specific local regions of the image. Different kernels can extract different features, such as edges and textures. For example, small kernels extract low-level features like edges and corners, while large kernels extract more complex high-level features like shapes and textures. During convolution, the kernel is multiplied by the pixel values ​​of the corresponding region in the image, and the results are summed to obtain the convolutional result. This process also introduces a bias value and increases the model's non-linearity through activation functions (such as ReLU), enabling the model to learn more complex features. Pooling layers typically follow the convolutional layers, downsampling the feature maps output by the convolutional layers. Common pooling methods include max pooling and average pooling. Max pooling selects the maximum value within the pooling window as the output, while average pooling calculates the average value within the window. This reduces the amount of data and computation, lowers the risk of overfitting, and preserves key features. Fully connected layers connect to the pooling layer outputs, integrating all features and calculating the final output using a weight matrix for tasks such as classification and regression. During training, the gradient of the loss function with respect to the parameters of each layer is calculated using backpropagation, continuously updating the weights to make the model's predictions closer to the true values.

[0142] In this embodiment, the feature extraction model also employs a convolutional neural network (CNN). In wheel image analysis, the powerful feature learning capability of the CNN can automatically learn complex features under different health states from a large number of wheel images. For wheel images taken from different angles and under different lighting conditions, the convolution operation of the CNN can effectively extract global features from the image. Combined with the translation invariance of the pooling operation, the model exhibits robustness to geometric transformations of the image, enabling stable analysis of wheel images in complex environments. Simultaneously, the global features extracted by the CNN can complement the local features extracted by the SURF algorithm. Fusing the two provides a more comprehensive feature representation for the differential analysis of wheel images, thereby improving the accuracy of subsequent wheel health assessments.

[0143] In the wheel health status assessment task of this embodiment, by training on a large number of labeled wheel images in different states such as normal and worn, the convolutional neural network can capture subtle feature changes in the images that are difficult to detect with the naked eye, thereby accurately judging the health status of the wheel.

[0144] It's also worth noting that the Mean Square Error (MSE) loss function commonly used in convolutional neural networks for image analysis is based on the least squares method. Using the MSE loss function guides the model to learn the mapping relationship between image features and the true labels, minimizing the model's prediction error on the training data. However, the least squares method is primarily suitable for linear models or those that can be transformed into linear models. Its expressive power is limited for complex nonlinear models and models with higher-order interactions. Furthermore, as model complexity increases, such as by increasing the number of independent variables or introducing higher-order terms, the computational cost of the least squares method increases significantly, potentially leading to overfitting. Additionally, multicollinearity makes the regression coefficients estimated by the least squares method unstable, increasing their variance and making it difficult to accurately determine the individual impact of each independent variable on the dependent variable. Moreover, the model becomes highly sensitive to small changes in the data, reducing its predictive stability.

[0145] For a well-posed problem, the matrix input to the convolutional neural network... X It is a full rank matrix. X The result obtained after inputting into a convolutional neural network can be represented as:

[0146] (11)

[0147] In equation (11), the matrix X The dimension is , N For the sample size, D For the feature dimension of a single sample, θ This represents the weight vector, with dimension 1. , result vector y The dimension of ' is It can represent health status information such as rotor health status label and wear level quantification value.

[0148] If the loss function of a convolutional neural network is defined as the square of the residuals, then the loss function minimized by the least squares method can be expressed as equation (12).

[0149] (12)

[0150] Equation (12) can be solved using the gradient descent method, or it can be solved directly using equation (13).

[0151] (13)

[0152] In matrix X, the number of columns is not full. , Approaching Singularity Solving The error will be large. To solve this problem, a regularization compensation term can be added to the loss function, which is Equation (14).

[0153] (14)

[0154] Here, the regularization term is defined. ,in a For a preset positive real number, I Since it is an identity matrix, therefore:

[0155] (15)

[0156] In equation (15), θ ( a ) represents the optimal solution of the loss function after adding a regularization term.

[0157] Mathematical derivation of equation (15) yields:

[0158] (16)

[0159] (17)

[0160] Equation (17) shows yes Compression towards the origin. Therefore, it exists. This makes the mean square error Therefore, under the mean squared error criterion, this method is superior to linear regression. Mathematically, this method sacrifices the unbiasedness of least squares, at the cost of losing some information, but in return, it yields more reliable regression coefficients and more realistic regression estimates. However, it performs better than least squares in training ill-conditioned data (represented in the algorithm as ill-conditioned matrices, where even small changes in the value of a single element can cause large computational errors; such matrices are called "ill-conditioned matrices." For example, in Gaussian elimination, if the values ​​of the pivot elements are very small, then such matrices will exhibit ill-conditioned characteristics during Gaussian elimination calculations). Based on the above method, the stability of the model's predicted solution can be improved, i.e., through regularization parameters. a Shrinking the solution toward the origin reduces the impact of multicollinearity among features, lowers the model's sensitivity to noise, and avoids overfitting.

[0161] It should be noted that during the training of the health status assessment model, firstly, a large amount of wheel image data with different health status labels (such as normal, slight wear, moderate wear, severe wear, etc.) was collected and meticulously labeled. Then, the labeled data was divided into training, validation, and test sets. The convolutional neural network was trained using the training set data. By adjusting the network parameters, the network could accurately learn the feature patterns of wheel images under different health states. During training, cross-entropy loss function and other loss metrics were used, and the network weights were continuously updated using the backpropagation algorithm to improve the accuracy of the health status assessment model. Finally, the model's prediction results were validated using the test set data. Precision, recall, and other metrics were calculated to evaluate the model's performance. If the model performance was unsatisfactory, the structure of the convolutional neural network could be adjusted, such as adding or removing convolutional layers, adjusting the kernel size, or increasing the amount of training data to retrain the model, in order to optimize model performance and obtain more accurate health status assessment results.

[0162] In this embodiment, the health status information of the turbine is obtained by inputting the fused feature information of the current turbine image into the health status assessment model. This health status information can be set as a health status label, such as "moderate wear", which indicates that the current turbine is in a moderate wear health status. As another form of health status information, a health score can also be obtained by outputting the probability of each health status category through the softmax function, so as to more intuitively reflect the health level of the turbine.

[0163] This embodiment enables comprehensive and high-precision monitoring of the health status of a turbine runner surface. Specifically, during implementation, this embodiment acquires local runner images at multiple rotation angles during turbine operation. By combining image feature extraction, image registration, and stitching techniques, a complete current runner image is reconstructed, thus achieving comprehensive and high-precision monitoring of the turbine runner surface health status. Based on this embodiment, surface damage and structural deformation of the turbine runner can be quickly identified without sensor contact and with strong anti-interference capabilities. This avoids the problems of poor environmental adaptability, low positioning accuracy, and high maintenance costs inherent in existing technologies, demonstrating significant application value and promising prospects for wider adoption.

[0164] Example 2:

[0165] This embodiment discloses a turbine non-stop condition monitoring system for implementing the turbine non-stop condition monitoring method in Embodiment 1; such as Figure 2 As shown, the turbine non-stop status monitoring system includes:

[0166] The image acquisition module is used to acquire multiple local images of the turbine runner at different rotation angles during turbine operation;

[0167] The feature extraction module is communicatively connected to the image acquisition module and is used to perform feature extraction processing on each local rotating wheel image to obtain feature descriptors for each local rotating wheel image.

[0168] The feature matching module, which is communicatively connected to the feature extraction module, is used to perform feature matching on all combinations of local rotating wheel images with adjacent acquisition times based on the feature descriptors of each local rotating wheel image, to obtain a set of matching points;

[0169] The image transformation module, which is communicatively connected to the feature matching module, is used to obtain the perspective transformation matrix corresponding to the combination of all local rotating images with adjacent acquisition times based on the matching point set; it is also used to perform coordinate transformation processing on each local rotating image based on the perspective transformation matrix corresponding to the combination of all local rotating images with adjacent acquisition times, so as to obtain multiple coordinate-transformed local rotating images located in the same coordinate system.

[0170] The image stitching module is communicatively connected to the image transformation module and is used to perform image stitching processing on multiple coordinate-transformed local runner images to obtain a complete current runner image of the turbine.

[0171] The health status detection module is communicatively connected to the image stitching module and is used to obtain the health status information of the turbine based on the current turbine image.

[0172] It should be noted that the working process, working details and technical effects of the turbine non-stop status monitoring system provided in this embodiment 2 can be found in embodiment 1, and will not be repeated here.

[0173] Example 3:

[0174] Based on Embodiment 1 or 2, this embodiment discloses an electronic device, which may be a smartphone, tablet computer, laptop computer, or desktop computer, etc. The electronic device may be referred to as a user terminal, portable terminal, desktop terminal, etc. Figure 3 As shown, the electronic device includes:

[0175] Memory, used to store computer program instructions; and,

[0176] A processor is used to execute the computer program instructions to perform the operation of a turbine non-stop status monitoring method as described in any of Embodiment 1.

[0177] Specifically, processor 301 may include one or more processing cores, such as a quad-core processor or an octa-core processor. Processor 301 may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). Processor 301 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 301 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen.

[0178] The memory 302 may include one or more computer-readable storage media, which may be non-transitory. The memory 302 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 302 are used to store at least one instruction, which is executed by the processor 301 to implement the turbine non-stop operation monitoring method provided in Embodiment 1 of this application.

[0179] In some embodiments, the terminal may also optionally include a communication interface 303 and at least one peripheral device. The processor 301, memory 302, and communication interface 303 can be connected via a bus or signal line. Each peripheral device can be connected to the communication interface 303 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of a radio frequency circuit 304, a display screen 305, and a power supply 306.

[0180] The communication interface 303 can be used to connect at least one I / O (Input / Output) related peripheral device to the processor 301 and the memory 302. In some embodiments, the processor 301, the memory 302, and the communication interface 303 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 301, the memory 302, and the communication interface 303 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0181] The radio frequency (RF) circuit 304 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 304 communicates with communication networks and other communication devices via electromagnetic signals.

[0182] Display screen 305 is used to display the UI (User Interface). The UI may include any combination of graphics, text, icons, and video.

[0183] Power supply 306 is used to supply power to various components in electronic devices.

[0184] Example 4:

[0185] Based on any one of Embodiments 1 to 3, this embodiment discloses a computer program product, including a computer program or instructions, which, when executed by a computer, implements a method for monitoring the non-stop operation status of a hydroelectric turbine as described in any one of Embodiments 1. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0186] Obviously, those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, the present invention is not limited to any particular hardware and software combination.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the non-stop operation of a water turbine, characterized in that, include: Acquire multiple local images of the turbine runner at different rotation angles during turbine operation; A camera combined with a stroboscopic device is used to acquire images of a local turbine runner. The camera and stroboscopic device are installed inside the transparent top cover of the turbine and face the runner. The stroboscopic device emits short pulses of light at a frequency that is the same as or an integer multiple of the turbine runner's rotational speed. Feature extraction processing is performed on each local rotating wheel image to obtain the feature descriptor of each local rotating wheel image; Based on the feature descriptors of each local rotating wheel image, feature matching is performed on all combinations of local rotating wheel images acquired at adjacent times to obtain a set of matching points; Based on the matching point set, the perspective transformation matrix corresponding to the combination of local rotating images with adjacent acquisition times is obtained; Based on the perspective transformation matrix corresponding to the combination of local wheel images with adjacent acquisition times, coordinate transformation processing is performed on each local wheel image to obtain multiple coordinate-transformed local wheel images located in the same coordinate system; Image stitching is performed on the local runner images after multiple coordinate transformations to obtain the complete current runner image of the water turbine; The health status information of the turbine is obtained based on the current turbine image; Feature extraction processing is performed on any local rotating wheel image to obtain a feature descriptor for that local rotating wheel image, including: The scale space is generated by convolving any local rotating image with Gaussian filters of different scales. Calculate the Hessian matrix of each pixel in any local rotating image in the scale space; Obtain the determinant response value of the Hessian matrix of each pixel in any local rotating image, and filter candidate feature points from all pixels in any local rotating image based on the determinant response value of the Hessian matrix of each pixel. The candidate feature points are optimized at the subpixel level using the quasi-Newton method to obtain the final image feature points corresponding to the candidate feature points. Generate feature descriptors for image feature points; The quasi-Newton method is used to perform sub-pixel-level optimization processing on any candidate feature point to obtain the image feature points corresponding to any candidate feature point, including: Obtain an approximate Hessian matrix for any candidate feature point; By iteratively updating the position and scale of candidate feature points, updated feature points are obtained, so as to maximize the determinant response value of the approximate matrix corresponding to the updated feature points, and thus obtain the maximum value of the determinant response value of the current approximate matrix. The updated feature point corresponding to the maximum value of the determinant response of the current approximate matrix is ​​taken as the image feature point corresponding to any of the candidate feature points; The health status information of the turbine is obtained based on the current runner image, including: The SURF algorithm is used to extract features from the current rotary image to obtain local feature information of the current rotary image; The current wheel image is input into a pre-trained feature extraction model to obtain the global feature information of the current wheel image; The local and global feature information of the current rotating image are fused to obtain the fused feature information of the current rotating image; The fused feature information of the current turbine runner image is input into a pre-trained health status assessment model to obtain the health status information of the turbine. After obtaining the fused feature information of the current rotary image, the method further includes: The turbine's runner image is acquired, and the SURF algorithm is used to extract features from the runner image to obtain local feature information of the runner image. The factory-exported rotor image is input into the feature extraction model to obtain the global feature information of the factory-exported rotor image; The local and global feature information of the factory-exit rotary wheel image are fused to obtain the fused feature information of the factory-exit rotary wheel image; A difference analysis is performed on the fused feature information of the current rotary image and the fused feature information of the factory rotary image to obtain the feature difference degree between the current rotary image and the factory rotary image. Determine whether the feature difference degree is greater than a preset feature difference threshold; If so, the fused feature information of the current turbine runner image is input into the health status assessment model to obtain the health status information of the turbine.

2. The method for monitoring the non-stop operation of a water turbine according to claim 1, characterized in that, Based on the feature descriptors of each local rotor image, feature matching is performed on all combinations of local rotor images acquired at adjacent times to obtain a set of matching points, including: Calculate the Euclidean distance between each feature descriptor in a combination of local rotary images with adjacent acquisition times, and take the image feature point corresponding to the feature descriptor with the closest Euclidean distance as the initial matching point of the current two local rotary images. The initial matching points of all combinations of local rotary images with adjacent acquisition times constitute the initial matching point set. The RANSAC algorithm is used to remove mismatched points from the initial matching point set to obtain the final matching point set.

3. The method for monitoring the non-stop operation of a water turbine according to claim 1, characterized in that, The health status assessment model uses a convolutional neural network.

4. A monitoring system for the non-stop operation status of a water turbine, characterized in that, include: The image acquisition module is used to acquire multiple local images of the turbine runner at different rotation angles during turbine operation; The feature extraction module is communicatively connected to the image acquisition module and is used to perform feature extraction processing on each local rotating wheel image to obtain feature descriptors for each local rotating wheel image. The feature matching module, which is communicatively connected to the feature extraction module, is used to perform feature matching on all combinations of local rotating wheel images with adjacent acquisition times based on the feature descriptors of each local rotating wheel image, to obtain a set of matching points; The image transformation module, which is communicatively connected to the feature matching module, is used to obtain the perspective transformation matrix corresponding to the combination of all local rotating images with adjacent acquisition times based on the matching point set; it is also used to perform coordinate transformation processing on each local rotating image based on the perspective transformation matrix corresponding to the combination of all local rotating images with adjacent acquisition times, so as to obtain multiple coordinate-transformed local rotating images located in the same coordinate system. The image stitching module is communicatively connected to the image transformation module and is used to perform image stitching processing on multiple coordinate-transformed local runner images to obtain a complete current runner image of the turbine. The health status detection module is communicatively connected to the image stitching module and is used to obtain the health status information of the turbine based on the current turbine image; Feature extraction processing is performed on any local rotating wheel image to obtain a feature descriptor for that local rotating wheel image, including: The scale space is generated by convolving any local rotating image with Gaussian filters of different scales. Calculate the Hessian matrix of each pixel in any local rotating image in the scale space; Obtain the determinant response value of the Hessian matrix of each pixel in any local rotating image, and filter candidate feature points from all pixels in any local rotating image based on the determinant response value of the Hessian matrix of each pixel. The candidate feature points are optimized at the subpixel level using the quasi-Newton method to obtain the final image feature points corresponding to the candidate feature points. Generate feature descriptors for image feature points; The quasi-Newton method is used to perform sub-pixel-level optimization processing on any candidate feature point to obtain the image feature points corresponding to any candidate feature point, including: Obtain an approximate Hessian matrix for any candidate feature point; By iteratively updating the position and scale of candidate feature points, updated feature points are obtained, so as to maximize the determinant response value of the approximate matrix corresponding to the updated feature points, and thus obtain the maximum value of the determinant response value of the current approximate matrix. The updated feature point corresponding to the maximum value of the determinant response of the current approximate matrix is ​​taken as the image feature point corresponding to any of the candidate feature points; The health status information of the turbine is obtained based on the current runner image, including: The SURF algorithm is used to extract features from the current rotary image to obtain local feature information of the current rotary image; The current wheel image is input into a pre-trained feature extraction model to obtain the global feature information of the current wheel image; The local and global feature information of the current rotating image are fused to obtain the fused feature information of the current rotating image; The fused feature information of the current turbine runner image is input into a pre-trained health status assessment model to obtain the health status information of the turbine. After obtaining the fused feature information of the current rotating image, the method further includes: The turbine's runner image is acquired, and the SURF algorithm is used to extract features from the runner image to obtain local feature information of the runner image. The factory-exported rotor image is input into the feature extraction model to obtain the global feature information of the factory-exported rotor image; The local and global feature information of the factory-exit rotary wheel image are fused to obtain the fused feature information of the factory-exit rotary wheel image; A difference analysis is performed on the fused feature information of the current rotary image and the fused feature information of the factory rotary image to obtain the feature difference degree between the current rotary image and the factory rotary image. Determine whether the feature difference degree is greater than a preset feature difference threshold; If so, the fused feature information of the current turbine runner image is input into the health status assessment model to obtain the health status information of the turbine.

5. An electronic device, characterized in that, include: Memory is used to store computer program instructions; as well as, A processor is configured to execute the computer program instructions to perform the operation of a method for monitoring the non-stop status of a water turbine as described in any one of claims 1 to 3.

6. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement a method for monitoring the non-stop operation of a water turbine as described in any one of claims 1 to 3.

Images