A method for diagnosing wear of a water turbine blade

Abstract

The application discloses a kind of water turbine blade wear diagnosis methods, belong to vibration test technical field.The application is by collecting the vibration signal of each blade of water turbine, and it is in the same time point to carry out amplitude superposition and average, obtains overall vibration signal.Then, using wavelet transform to the vibration signal of overall signal and blade to be detected is handled, constructs approximation and detail residual energy ratio matrix.Through extracting eigenvalue and outlier of matrix, form corresponding feature matrix and anomaly matrix.Finally, using wear diagnosis network combines these features and abnormal information, accurately assesses the wear degree of blade.The application realizes high-precision wear diagnosis, is helpful to the operation safety and maintenance management of water turbine.

Description

Technical Field

[0001] This invention relates to the field of vibration testing technology, and more specifically to a diagnostic method for wear on turbine blades. Background Technology

[0002] With the global trend towards a clean energy transition, hydropower, as a technologically mature and sustainable clean energy source, occupies a crucial position in the energy supply system. As the core equipment of a hydropower system, the operating efficiency and stability of the turbine directly affect the power plant's power generation benefits and safe operation. The blades, as key components for energy conversion in the turbine, are constantly subjected to the impact of high-speed water flow. They are inevitably subjected to the combined effects of erosion from sediment and impurities carried by the water flow, as well as cavitation, fatigue, and other factors, making them highly susceptible to wear.

[0003] Blade wear not only alters the geometry of turbine flow components, leading to decreased hydraulic performance and reduced power generation efficiency, but also causes problems such as increased unit vibration and noise. In severe cases, it can even result in blade breakage and unit shutdown, causing significant safety accidents and resulting in substantial economic losses and safety hazards for the power station. Statistics show that in hydropower stations located in river basins with high sediment loads, downtime due to blade wear accounts for more than 30% of total downtime, significantly increasing annual maintenance costs.

[0004] Currently, diagnostic methods for turbine blade wear mainly fall into two categories: offline detection and online monitoring. Offline detection requires shutting down and disassembling the equipment, assessing the wear condition through manual visual inspection, dimensional measurement, or non-destructive testing. This method not only interrupts power generation but also has a long detection cycle and poor timeliness, making it difficult to meet the requirements for continuous and stable unit operation. While online monitoring methods can achieve detection without shutting down the unit, they mostly rely on vibration signals collected by a single sensor. Vibration signal analysis is an effective fault diagnosis method, reflecting the health status of the equipment by monitoring mechanical vibration waveforms. However, turbine blade vibration signals are greatly affected by environmental noise, and the signals themselves are non-stationary, posing challenges to vibration signal-based wear diagnosis. Existing signal processing methods often struggle to extract wear-related features from complex signals, resulting in low accuracy in wear diagnosis. Summary of the Invention

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

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for diagnosing wear on turbine blades, comprising the following steps:

[0007] Vibration signals of each blade of the turbine are collected, and the amplitudes of the vibration signals of each blade in the turbine are added together at the same time point. The average amplitude is then taken to obtain the overall vibration signal.

[0008] Wavelet transforms were performed on the signal segments of the overall vibration signal and the vibration signal of the blade under test, respectively. Based on the difference between the wavelet approximation sequence and the wavelet detail sequence, the approximate residual energy ratio matrix and the detail residual energy ratio matrix were constructed.

[0009] Eigenvalues ​​and outliers are extracted from each row of the approximate residual energy ratio matrix and the detail residual energy ratio matrix, respectively, to obtain the approximate residual feature matrix, the detail residual feature matrix, the approximate residual outlier matrix, and the detail residual outlier matrix;

[0010] The wear diagnosis network is used to process the approximate residual feature matrix and the detailed residual feature matrix, and the wear value of the turbine blade is obtained based on the weights applied to the approximate residual abnormal matrix and the detailed residual abnormal matrix.

[0011] Furthermore, the process of constructing the approximate residual energy ratio matrix and the detailed residual energy ratio matrix includes:

[0012] The vibration signal of the blade to be tested is divided into multiple blade signal segments according to a fixed length;

[0013] The overall vibration signal is divided into multiple overall signal segments by dividing it into segments of fixed length;

[0014] Perform three-level wavelet transform on each blade signal segment to obtain a three-level blade wavelet approximation sequence and a three-level blade wavelet detail sequence.

[0015] Perform a 3-level wavelet transform on each overall signal segment to obtain a 3-level overall wavelet approximation sequence and a 3-level overall wavelet detail sequence;

[0016] Based on the difference between the 3-layer blade wavelet approximation sequence and the 3-layer overall wavelet approximation sequence, an approximate residual energy ratio vector is constructed.

[0017] Based on the difference between the 3-layer blade wavelet detail sequence and the 3-layer global wavelet detail sequence, construct the detail residual energy ratio vector;

[0018] Each approximate residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the approximate residual energy ratio matrix.

[0019] Each detail residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the detail residual energy ratio matrix.

[0020] Furthermore, the process of constructing the approximate residual energy ratio vector includes:

[0021] The first-layer blade wavelet approximation sequence is subtracted from the first-layer global wavelet approximation sequence by position to obtain the first-layer blade residual approximation sequence.

[0022] The second-layer blade wavelet approximation sequence is obtained by subtracting the second-layer global wavelet approximation sequence position by position.

[0023] The third-layer blade wavelet approximation sequence is obtained by subtracting the third-layer global wavelet approximation sequence position by position.

[0024] The approximate residual energy ratios were calculated for the approximate residual sequences of each layer of blades, resulting in the first, second, and third approximate residual energy ratios.

[0025] The first approximate residual energy ratio, the second approximate residual energy ratio, and the third approximate residual energy ratio are used to form an approximate residual energy ratio vector.

[0026] Furthermore, the process of calculating the approximate residual energy ratio includes: taking the absolute value of the i-th element in the residual approximation sequence of the k-th layer blade, taking the ratio of the element after taking the absolute value to the i-th element in the overall wavelet approximation sequence of the k-th layer as the i-th residual energy ratio, and taking the average of all residual energy ratios corresponding to the residual approximation sequence of the k-th layer blade to obtain the approximate residual energy ratio, where k takes the values ​​1, 2, and 3, and i is a positive integer.

[0027] Furthermore, the process of constructing the detailed residual energy ratio vector includes:

[0028] The first-level blade wavelet detail sequence is subtracted from the first-level global wavelet detail sequence by position to obtain the first-level blade residual detail sequence;

[0029] The second-layer blade wavelet detail sequence is obtained by subtracting the second-layer global wavelet detail sequence position by position.

[0030] The third-layer blade wavelet detail sequence is obtained by subtracting the third-layer global wavelet detail sequence position by position.

[0031] The energy ratio of the detail residuals is calculated for each layer of leaf residual detail sequence, and the first, second, and third detail residual energy ratios are obtained respectively.

[0032] The first detail residual energy ratio, the second detail residual energy ratio, and the third detail residual energy ratio are used to construct a detail residual energy ratio vector.

[0033] Furthermore, the process of calculating the detail residual energy ratio includes: taking the absolute value of the i-th element in the residual detail sequence of the k-th layer blade, taking the ratio of the element after taking the absolute value to the i-th element in the overall wavelet detail sequence of the k-th layer as the i-th residual energy ratio, and taking the average of all residual energy ratios corresponding to the residual detail sequence of the k-th layer blade to obtain the detail residual energy ratio, where k takes the values ​​of 1, 2, and 3, and i is a positive integer.

[0034] Furthermore, the process of obtaining the approximate residual characteristic matrix, the detailed residual characteristic matrix, the approximate residual inconsistency matrix, and the detailed residual inconsistency matrix includes:

[0035] A sliding window of length M is used to slide on each row of the approximate residual energy ratio matrix and the detailed residual energy ratio matrix, with a sliding step size of K, where M is a positive integer greater than or equal to 3 and K is a positive integer.

[0036] Calculate the mean of the elements under the sliding window to obtain the approximate residual feature matrix and the detailed residual feature matrix;

[0037] On the approximate residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer's approximate residual energy ratio threshold within the sliding window to obtain approximate outliers;

[0038] Arrange the approximate outliers according to their corresponding sliding window positions to obtain the approximate residual outlier matrix;

[0039] On the detail residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer detail residual energy ratio threshold under the sliding window to obtain detail outliers;

[0040] Arrange the various detail outliers according to their corresponding sliding window positions to obtain the detail residual outlier matrix.

[0041] Furthermore, the approximate outlier is the ratio of the number of elements with an approximate residual energy ratio greater than the corresponding layer's threshold under the sliding window to the length of the sliding window;

[0042] The outlier value is the ratio of the number of elements with a detail residual energy ratio greater than the threshold of the corresponding layer under the sliding window to the length of the sliding window.

[0043] Furthermore, the wear diagnosis network includes: a first matrix feature enhancement unit, a second matrix feature enhancement unit, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an adder A1, and a fully connected layer;

[0044] The first input of the first matrix feature enhancement unit is used to input the approximate residual feature matrix, its second input is used to input the approximate residual abnormal matrix, and its output is connected to the input of the first convolutional layer.

[0045] The first input of the second matrix feature enhancement unit is used to input the detail residual feature matrix, the second input is used to input the detail residual abnormal matrix, and the output is connected to the input of the third convolutional layer.

[0046] The output of the first convolutional layer is connected to the input of the second convolutional layer; the output of the third convolutional layer is connected to the input of the fourth convolutional layer; the input of adder A1 is connected to the outputs of the second and fourth convolutional layers respectively, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the wear diagnosis network.

[0047] Furthermore, the first matrix feature enhancement unit includes a fifth convolutional layer and a multiplier M1. The fifth convolutional layer is used to process the approximate residual feature matrix, and the multiplier M1 is used to apply weights to the output of the fifth convolutional layer using the approximate residual inconsistency matrix.

[0048] The second matrix feature enhancement unit includes a sixth convolutional layer and a multiplier M2. The sixth convolutional layer is used to process the detail residual feature matrix, and the multiplier M2 is used to apply weights to the output of the sixth convolutional layer using the detail residual inconsistency matrix.

[0049] The beneficial effects of this invention are as follows:

[0050] 1. Existing methods struggle to effectively separate wear features from noisy signals under complex operating conditions. This invention, however, obtains the overall vibration signal by summing the amplitudes of each blade's vibration signals at the same time point and taking the average. This effectively reduces the impact of common noise such as water flow pulsation and mechanical interference. Furthermore, based on the differences between the overall vibration signal and the blade under test in the wavelet approximation sequence and wavelet detail sequence, this invention constructs an approximate residual energy ratio matrix and a detail residual energy ratio matrix. This reflects anomalies in the approximation coefficient and detail coefficient, accurately capturing the differences between the blade under test and the overall state in different components. This more clearly highlights wear-related feature information and improves the accuracy of wear diagnosis.

[0051] 2. This invention extracts eigenvalues ​​and outliers from each row of the approximate residual energy ratio matrix and the detailed residual energy ratio matrix, forming an approximate residual feature matrix, a detailed residual feature matrix, and a corresponding outlier matrix. This not only fully preserves the overall distribution of wear characteristics, but also obtains the outlier weight of each element in the approximate residual feature matrix and the detailed residual feature matrix through the outlier matrix, thereby improving the accuracy of the wear diagnosis network in predicting the wear value of turbine blades. Attached Figure Description

[0052] Figure 1 A flowchart of a diagnostic method for turbine blade wear;

[0053] Figure 2 This is a schematic diagram of the wear diagnosis network structure;

[0054] Figure 3 This is a schematic diagram of the structure of the first matrix feature enhancement unit;

[0055] Figure 4 This is a schematic diagram of the structure of the second matrix feature enhancement unit. Detailed Implementation

[0056] 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.

[0057] like Figure 1 As shown, a method for diagnosing wear on turbine blades includes the following steps:

[0058] Vibration signals of each blade of the turbine are collected, and the amplitudes of the vibration signals of each blade in the turbine are added together at the same time point. The average amplitude is then taken to obtain the overall vibration signal.

[0059] Wavelet transforms were performed on the signal segments of the overall vibration signal and the vibration signal of the blade under test, respectively. Based on the difference between the wavelet approximation sequence and the wavelet detail sequence, the approximate residual energy ratio matrix and the detail residual energy ratio matrix were constructed.

[0060] Eigenvalues ​​and outliers are extracted from each row of the approximate residual energy ratio matrix and the detail residual energy ratio matrix, respectively, to obtain the approximate residual feature matrix, the detail residual feature matrix, the approximate residual outlier matrix, and the detail residual outlier matrix;

[0061] The wear diagnosis network is used to process the approximate residual feature matrix and the detailed residual feature matrix, and the wear value of the turbine blade is obtained based on the weights applied to the approximate residual abnormal matrix and the detailed residual abnormal matrix.

[0062] The overall vibration signal is obtained by arranging the average amplitude values ​​at each time point.

[0063] In this embodiment, the vibration sensor is placed at locations including: the blade root (where it connects to the hub), the blade exiting the water (the edge near the blade tip), and the back of the blade (the non-working surface). This invention captures vibration signals from the same point on each blade.

[0064] In this embodiment, the vibration signal is 1 minute long and the sampling frequency is 16kHz. The 1-minute vibration signal contains 960,000 sampling points. The segment length is 2 seconds (containing 32,000 sampling points). The 2 seconds contains 10-17 rotation cycles (300-500 r / min), reflecting the vibration pattern of the blade in multiple cycles. It is divided into 30 segments in total.

[0065] In this embodiment, the process of constructing the approximate residual energy ratio matrix and the detailed residual energy ratio matrix includes:

[0066] The vibration signal of the blade to be tested is divided into multiple blade signal segments according to a fixed length;

[0067] The overall vibration signal is divided into multiple overall signal segments by dividing it into segments of fixed length;

[0068] Perform three-level wavelet transform on each blade signal segment to obtain a three-level blade wavelet approximation sequence and a three-level blade wavelet detail sequence.

[0069] Perform a 3-level wavelet transform on each overall signal segment to obtain a 3-level overall wavelet approximation sequence and a 3-level overall wavelet detail sequence;

[0070] Based on the difference between the 3-layer blade wavelet approximation sequence and the 3-layer overall wavelet approximation sequence, an approximate residual energy ratio vector is constructed.

[0071] Based on the difference between the 3-layer blade wavelet detail sequence and the 3-layer global wavelet detail sequence, construct the detail residual energy ratio vector;

[0072] Each approximate residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the approximate residual energy ratio matrix.

[0073] Each detail residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the detail residual energy ratio matrix.

[0074] This invention uses a three-layer wavelet transform to decompose the blade vibration signal and the overall vibration signal into an approximate sequence (low-frequency trend) and a detail sequence (high-frequency fluctuation). Whether it's the subtle high-frequency impacts caused by blade wear or the low-frequency trend changes brought about by equipment aging, both can be decomposed layer by layer and accurately captured. For example, in the case of turbine blade wear, early high-frequency impacts will be highlighted in the detail sequence, while the low-frequency drift caused by long-term structural fatigue will be reflected in the approximate sequence.

[0075] In this embodiment, the wavelet basis function selected for wavelet transform is the db4 (Daubechies 4) wavelet. The first layer contains 16,000 wavelet approximation coefficients in the wavelet approximation sequence and 16,000 wavelet detail coefficients in the wavelet detail sequence. The second layer contains 8,000 wavelet approximation coefficients in the wavelet approximation sequence and 8,000 wavelet detail coefficients in the wavelet detail sequence. The third layer contains 4,000 wavelet approximation coefficients in the wavelet approximation sequence and 4,000 wavelet detail coefficients in the wavelet detail sequence.

[0076] In this embodiment, the process of constructing an approximate residual energy ratio vector includes:

[0077] The first-layer blade wavelet approximation sequence is subtracted from the first-layer global wavelet approximation sequence by position to obtain the first-layer blade residual approximation sequence.

[0078] The second-layer blade wavelet approximation sequence is obtained by subtracting the second-layer global wavelet approximation sequence position by position.

[0079] The third-layer blade wavelet approximation sequence is obtained by subtracting the third-layer global wavelet approximation sequence position by position.

[0080] The approximate residual energy ratios were calculated for the approximate residual sequences of each layer of blades, resulting in the first, second, and third approximate residual energy ratios.

[0081] The first approximate residual energy ratio, the second approximate residual energy ratio, and the third approximate residual energy ratio are used to form an approximate residual energy ratio vector.

[0082] The first approximate residual energy ratio is obtained from the residual approximation sequence of the first layer of blades, the second approximate residual energy ratio is obtained from the residual approximation sequence of the second layer of blades, and the third approximate residual energy ratio is obtained from the residual approximation sequence of the third layer of blades.

[0083] This invention subtracts the wavelet approximation sequence of each layer of blades from the corresponding overall wavelet approximation sequence position by position, highlighting the deviation of the blade approximation coefficients from the overall sequence.

[0084] In this embodiment, the process of calculating the approximate residual energy ratio includes: taking the absolute value of the i-th element in the k-th layer blade residual approximation sequence, taking the ratio of the element after taking the absolute value to the i-th element in the k-th layer overall wavelet approximation sequence as the i-th residual energy ratio, taking the average of all residual energy ratios corresponding to the k-th layer blade residual approximation sequence to obtain the approximate residual energy ratio, where k takes the values ​​1, 2, and 3, and i is a positive integer.

[0085] In this embodiment, the process of constructing the detail residual energy ratio vector includes:

[0086] The first-level blade wavelet detail sequence is subtracted from the first-level global wavelet detail sequence by position to obtain the first-level blade residual detail sequence;

[0087] The second-layer blade wavelet detail sequence is obtained by subtracting the second-layer global wavelet detail sequence position by position.

[0088] The third-layer blade wavelet detail sequence is obtained by subtracting the third-layer global wavelet detail sequence position by position.

[0089] The energy ratio of the detail residuals is calculated for each layer of leaf residual detail sequence, and the first, second, and third detail residual energy ratios are obtained respectively.

[0090] The first detail residual energy ratio, the second detail residual energy ratio, and the third detail residual energy ratio are used to construct a detail residual energy ratio vector.

[0091] The first detail residual energy ratio is obtained from the first layer of blade residual detail sequence, the second detail residual energy ratio is obtained from the second layer of blade residual detail sequence, and the third detail residual energy ratio is obtained from the third layer of blade residual detail sequence.

[0092] This invention subtracts the wavelet detail sequence of each blade layer from the corresponding overall wavelet detail sequence position by position, highlighting the deviation of the blade detail coefficients from the overall sequence.

[0093] Since there are 30 signal segments, the number of approximate residual energy ratio vectors and detail residual energy ratio vectors are both 30. Therefore, the approximate residual energy ratio matrix and the detail residual energy ratio matrix have 3 rows and 30 columns. The elements of the first row of the approximate residual energy ratio matrix are: the first approximate residual energy ratio at each time point; the elements of the second row are: the second approximate residual energy ratio at each time point; and the elements of the third row are: the third approximate residual energy ratio at each time point. Similarly, the elements of the first row of the detail residual energy ratio matrix are: the first detail residual energy ratio at each time point; the elements of the second row are: the second detail residual energy ratio at each time point; and the elements of the third row are: the third detail residual energy ratio at each time point.

[0094] In this embodiment, the process of calculating the detail residual energy ratio includes: taking the absolute value of the i-th element in the residual detail sequence of the k-th layer blade, taking the ratio of the element after taking the absolute value to the i-th element in the overall wavelet detail sequence of the k-th layer as the i-th residual energy ratio, taking the average of all residual energy ratios corresponding to the residual detail sequence of the k-th layer blade to obtain the detail residual energy ratio, where k takes the values ​​of 1, 2, and 3, and i is a positive integer.

[0095] In this embodiment, the formula for calculating the approximate residual energy ratio is: EL,k B is the approximate residual energy ratio of the k-th layer. L,k,i Let N be the i-th element in the approximate sequence of the residuals of the k-th layer of blades. L Let R be the length of the approximate sequence of blade residuals, i be a positive integer, || be the absolute value, and R be the absolute value. L,W,k,i It is the i-th element in the k-th layer global wavelet approximation sequence.

[0096] In this embodiment, the formula for calculating the detail residual energy ratio is: E H,k B is the detail residual energy ratio of the k-th layer. H,k,i For the i-th element in the residual detail sequence of the k-th layer blade, N H R is the length of the blade residual detail sequence. H,W,k,i Let i be the i-th element in the global wavelet detail sequence of the k-th layer.

[0097] This invention transforms the absolute vibration amplitude difference into a relative energy ratio, which not only eliminates the interference of "different overall vibration amplitudes" under different operating conditions (such as load changes), but also highlights the energy disturbance of blade vibration "abnormal state". For example, when the blade is slightly worn, the overall vibration amplitude may remain unchanged, but the residual energy ratio will increase significantly due to the "relative increase" of local high-frequency impact, thus detecting the signs of failure earlier.

[0098] In this embodiment, the process of obtaining the approximate residual feature matrix, the detailed residual feature matrix, the approximate residual inconsistency matrix, and the detailed residual inconsistency matrix includes:

[0099] A sliding window of length M is used to slide on each row of the approximate residual energy ratio matrix and the detailed residual energy ratio matrix, with a sliding step size of K, where M is a positive integer greater than or equal to 3 and K is a positive integer.

[0100] Calculate the mean of the elements under the sliding window to obtain the approximate residual feature matrix and the detailed residual feature matrix;

[0101] On the approximate residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer's approximate residual energy ratio threshold within the sliding window to obtain approximate outliers;

[0102] Arrange the approximate outliers according to their corresponding sliding window positions to obtain the approximate residual outlier matrix;

[0103] On the detail residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer detail residual energy ratio threshold under the sliding window to obtain detail outliers;

[0104] Arrange the various detail outliers according to their corresponding sliding window positions to obtain the detail residual outlier matrix.

[0105] In this embodiment, M is set to 5, and the sliding step size K is set to 5. That is, each row of the approximate residual energy ratio matrix and the detail residual energy ratio matrix is ​​divided into 5 groups, for a total of 6 groups. The average value of each group is taken to obtain the 3×6 approximate residual feature matrix and the detail residual feature matrix.

[0106] This invention takes the average of each set of data to reflect the overall trend of low-frequency trend anomalies and high-frequency fluctuation anomalies, while reducing the amount of data and computational pressure. It then counts how many elements in a set of data exceed a threshold and takes the ratio of the number of elements to the length of the sliding window as the outlier value to evaluate the anomaly of the mean of that set.

[0107] In this embodiment, the approximate residual energy ratio threshold and the detailed residual energy ratio threshold of each layer are both set to 0.5, which means that the average residual energy of the blade reaches 50% of the overall vibration energy. Under normal operating conditions, the blade vibration should be "highly consistent" with the overall vibration (the residual energy ratio is extremely low, such as ≤0.2); once the ratio reaches 50%, it indicates that the energy of the blade vibration "abnormal to the normal state" is very significant, and it is by no means a fluctuation in operating conditions or measurement error, but a substantial difference caused by a fault (such as abnormal vibration caused by blade wear or cracks).

[0108] In this embodiment, in order to improve detection accuracy, the approximate residual energy ratio threshold and the detail residual energy ratio threshold of each layer can also be set to values ​​between 0.3 and 0.5.

[0109] In this embodiment, the approximate outlier value is the ratio of the number of elements with an approximate residual energy ratio greater than the corresponding layer's threshold to the length of the sliding window.

[0110] The outlier value is the ratio of the number of elements with a detail residual energy ratio greater than the threshold of the corresponding layer under the sliding window to the length of the sliding window.

[0111] In this embodiment, the process of obtaining approximate outliers includes: in the first row of the approximate residual energy ratio matrix, counting the number of elements in the sliding window of the first row that are greater than the first layer approximate residual energy ratio threshold, and taking the ratio of the number of elements to the length of the sliding window as the approximate outlier at the corresponding sliding window position in the first row.

[0112] In the second row of the approximate residual energy ratio matrix, count the number of elements in the sliding window of the second row that are greater than the second layer approximate residual energy ratio threshold, and use the ratio of the number of elements to the length of the sliding window as the approximate outlier at the corresponding sliding window position in the second row.

[0113] In the third row of the approximate residual energy ratio matrix, count the number of elements in the sliding window of the third row that are greater than the third layer approximate residual energy ratio threshold. Use the ratio of the number of elements to the length of the sliding window as the approximate outlier at the corresponding sliding window position in the third row.

[0114] The process of obtaining detail outliers includes: in the first row of the detail residual energy ratio matrix, counting the number of elements in the sliding window of the first row that are greater than the first layer detail residual energy ratio threshold, and taking the ratio of the number of elements to the length of the sliding window as the detail outlier at the corresponding sliding window position in the first row.

[0115] In the second row of the detail residual energy ratio matrix, count the number of elements in the sliding window of the second row that are greater than the second layer detail residual energy ratio threshold, and use the ratio of the number of elements to the length of the sliding window as the detail outlier at the corresponding sliding window position in the second row.

[0116] In the third row of the detail residual energy ratio matrix, count the number of elements in the sliding window of the third row that are greater than the detail residual energy ratio threshold of the third layer. The ratio of the number of elements to the length of the sliding window is taken as the detail outlier at the corresponding sliding window position in the third row.

[0117] like Figure 2 As shown, the wear diagnosis network includes: a first matrix feature enhancement unit, a second matrix feature enhancement unit, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an adder A1, and a fully connected layer;

[0118] The first input of the first matrix feature enhancement unit is used to input the approximate residual feature matrix, its second input is used to input the approximate residual abnormal matrix, and its output is connected to the input of the first convolutional layer.

[0119] The first input of the second matrix feature enhancement unit is used to input the detail residual feature matrix, the second input is used to input the detail residual abnormal matrix, and the output is connected to the input of the third convolutional layer.

[0120] The output of the first convolutional layer is connected to the input of the second convolutional layer; the output of the third convolutional layer is connected to the input of the fourth convolutional layer; the input of adder A1 is connected to the outputs of the second and fourth convolutional layers respectively, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the wear diagnosis network.

[0121] The kernel size of the first, second, third, and fourth convolutional layers is 3×3.

[0122] This invention employs an approximate residual inconsistency matrix to apply weights to the features corresponding to the approximate residual feature matrix, and a detail residual inconsistency matrix to apply weights to the features corresponding to the detail residual feature matrix. This approach emphasizes key features and weakens non-key features. Features are then extracted from the output of the first matrix feature enhancement unit through the first and second convolutional layers, and from the output of the second matrix feature enhancement unit through the third and fourth convolutional layers. Finally, these features are added together by adder A1, allowing the fully connected layer to synthesize the outputs from both aspects to predict the wear value of the turbine blades.

[0123] like Figure 3 As shown, the first matrix feature enhancement unit includes a fifth convolutional layer and a multiplier M1. The fifth convolutional layer is used to process the approximate residual feature matrix, and the multiplier M1 is used to apply weights to the output of the fifth convolutional layer using the approximate residual abnormal matrix.

[0124] like Figure 4 As shown, the second matrix feature enhancement unit includes a sixth convolutional layer and a multiplier M2. The sixth convolutional layer is used to process the detail residual feature matrix, and the multiplier M2 is used to apply weights to the output of the sixth convolutional layer using the detail residual abnormal matrix.

[0125] In this embodiment, the kernel size of the fifth and sixth convolutional layers is 1×1, keeping the matrix size unchanged. Element-wise multiplication is performed by multipliers M1 and M2 to accurately enhance these local burst features and improve diagnostic accuracy.

[0126] In this embodiment, the wear value of the turbine blades ranges from 0 to 10, where 0 corresponds to no wear and 10 corresponds to severe wear. If the blade wear value is 8 to 10, the blade is heavily worn; if the blade wear value is 6 to 8, the blade is moderately worn; if the blade wear value is 4 to 6, the blade is slightly worn; and if the blade wear value is 0 to 4, the blade is considered to be without wear.

[0127] Existing methods struggle to effectively separate wear features from noisy signals under complex operating conditions. This invention, however, obtains the overall vibration signal by summing the amplitudes of each blade's vibration signals at the same time point and taking the average. This effectively reduces the impact of common noise such as water flow pulsation and mechanical interference. Furthermore, based on the differences between the overall vibration signal and the blade under test in the wavelet approximation sequence and wavelet detail sequence, this invention constructs an approximate residual energy ratio matrix and a detail residual energy ratio matrix. This reflects anomalies in the approximation coefficients and detail coefficients, accurately capturing the differences between the blade under test and the overall state in different components. This more clearly highlights wear-related feature information and improves the accuracy of wear diagnosis.

[0128] This invention extracts eigenvalues ​​and outliers from each row of the approximate residual energy ratio matrix and the detailed residual energy ratio matrix, forming an approximate residual feature matrix, a detailed residual feature matrix, and a corresponding outlier matrix. This not only fully preserves the overall distribution of wear characteristics, but also obtains the outlier weight of each element in the approximate residual feature matrix and the detailed residual feature matrix through the outlier matrix, thereby improving the accuracy of the wear diagnosis network in predicting the wear value of turbine blades.

[0129] 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 for diagnosing wear on turbine blades, characterized in that, Includes the following steps: Vibration signals of each blade of the turbine are collected, and the amplitudes of the vibration signals of each blade in the turbine are added together at the same time point. The average amplitude is then taken to obtain the overall vibration signal. Wavelet transforms are performed on signal segments of both the overall vibration signal and the vibration signal of the blade under test. Based on the difference between the wavelet approximation sequence and the wavelet detail sequence: the wavelet approximation sequence of each layer of blades is subtracted positionally from the corresponding layer's overall wavelet approximation sequence, and the wavelet detail sequence of each layer of blades is subtracted positionally from the corresponding layer's overall wavelet detail sequence. The absolute value of the i-th element in the residual approximation sequence of the k-th layer of blades is taken, and the ratio of the absolute value of this element to the i-th element in the overall wavelet approximation sequence of the k-th layer is used as the i-th residual energy ratio. The approximate residual energy ratio is obtained by averaging all residual energy ratios corresponding to the approximate residual sequence of the k-th layer blade, where k takes the values ​​of 1, 2, and 3, and i is a positive integer. The absolute value of the i-th element in the residual detail sequence of the k-th layer blade is taken, and the ratio of the absolute value of this element to the i-th element in the overall wavelet detail sequence of the k-th layer is taken as the i-th residual energy ratio. The average of all residual energy ratios corresponding to the residual detail sequence of the k-th layer blade is then used to obtain the detail residual energy ratio. The approximate residual energy ratio matrix and the detail residual energy ratio matrix are then constructed. Eigenvalues ​​and outliers are extracted from each row of the approximate residual energy ratio matrix and the detail residual energy ratio matrix, respectively, to obtain the approximate residual feature matrix, the detail residual feature matrix, the approximate residual outlier matrix, and the detail residual outlier matrix; The wear diagnosis network is used to process the approximate residual feature matrix and the detailed residual feature matrix, and the wear value of the turbine blade is obtained based on the weights applied to the approximate residual abnormal matrix and the detailed residual abnormal matrix.

2. The method for diagnosing turbine blade wear according to claim 1, characterized in that, The process of constructing the approximate residual energy ratio matrix and the detailed residual energy ratio matrix includes: The vibration signal of the blade to be tested is divided into multiple blade signal segments according to a fixed length; The overall vibration signal is divided into multiple overall signal segments by dividing it into segments of fixed length; Perform three-level wavelet transform on each blade signal segment to obtain a three-level blade wavelet approximation sequence and a three-level blade wavelet detail sequence. Perform a 3-level wavelet transform on each overall signal segment to obtain a 3-level overall wavelet approximation sequence and a 3-level overall wavelet detail sequence; Based on the difference between the 3-layer blade wavelet approximation sequence and the 3-layer overall wavelet approximation sequence, an approximate residual energy ratio vector is constructed. Based on the difference between the 3-layer blade wavelet detail sequence and the 3-layer global wavelet detail sequence, construct the detail residual energy ratio vector; Each approximate residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the approximate residual energy ratio matrix. Each detail residual energy ratio vector is used as a column vector, and the column vectors are arranged in chronological order to obtain the detail residual energy ratio matrix.

3. The method for diagnosing turbine blade wear according to claim 2, characterized in that, The process of constructing an approximate residual energy ratio vector includes: The first-layer blade wavelet approximation sequence is subtracted from the first-layer global wavelet approximation sequence by position to obtain the first-layer blade residual approximation sequence. The second-layer blade wavelet approximation sequence is obtained by subtracting the second-layer global wavelet approximation sequence position by position. The third-layer blade wavelet approximation sequence is obtained by subtracting the third-layer global wavelet approximation sequence position by position. The approximate residual energy ratios were calculated for the approximate residual sequences of each layer of blades, resulting in the first, second, and third approximate residual energy ratios. The first approximate residual energy ratio, the second approximate residual energy ratio, and the third approximate residual energy ratio are used to form an approximate residual energy ratio vector.

4. The method for diagnosing turbine blade wear according to claim 2, characterized in that, The process of constructing the detailed residual energy ratio vector includes: The first-level blade wavelet detail sequence is subtracted from the first-level global wavelet detail sequence by position to obtain the first-level blade residual detail sequence; The second-layer blade wavelet detail sequence is obtained by subtracting the second-layer global wavelet detail sequence position by position. The third-layer blade wavelet detail sequence is obtained by subtracting the third-layer global wavelet detail sequence position by position. The energy ratio of the detail residuals is calculated for each layer of leaf residual detail sequence, and the first, second, and third detail residual energy ratios are obtained respectively. The first detail residual energy ratio, the second detail residual energy ratio, and the third detail residual energy ratio are used to construct a detail residual energy ratio vector.

5. The method for diagnosing turbine blade wear according to claim 1, characterized in that, The process of obtaining the approximate residual characteristic matrix, the detailed residual characteristic matrix, the approximate residual inconsistency matrix, and the detailed residual inconsistency matrix includes: A sliding window of length M is used to slide on each row of the approximate residual energy ratio matrix and the detailed residual energy ratio matrix, with a sliding step size of K, where M is a positive integer greater than or equal to 3 and K is a positive integer. Calculate the mean of the elements under the sliding window to obtain the approximate residual feature matrix and the detailed residual feature matrix; On the approximate residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer's approximate residual energy ratio threshold within the sliding window to obtain approximate outliers; Arrange the approximate outliers according to their corresponding sliding window positions to obtain the approximate residual outlier matrix; On the detail residual energy ratio matrix, count the elements in each row that are greater than the corresponding layer detail residual energy ratio threshold under the sliding window to obtain detail outliers; Arrange the various detail outliers according to their corresponding sliding window positions to obtain the detail residual outlier matrix.

6. The method for diagnosing turbine blade wear according to claim 5, characterized in that, The approximate outlier is the ratio of the number of elements with an approximate residual energy ratio greater than the corresponding layer's threshold under the sliding window to the length of the sliding window. The outlier value is the ratio of the number of elements with a detail residual energy ratio greater than the threshold of the corresponding layer under the sliding window to the length of the sliding window.

7. The method for diagnosing turbine blade wear according to claim 1, characterized in that, The wear diagnosis network includes: a first matrix feature enhancement unit, a second matrix feature enhancement unit, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an adder A1, and a fully connected layer; The first input of the first matrix feature enhancement unit is used to input the approximate residual feature matrix, its second input is used to input the approximate residual abnormal matrix, and its output is connected to the input of the first convolutional layer. The first input of the second matrix feature enhancement unit is used to input the detail residual feature matrix, the second input is used to input the detail residual abnormal matrix, and the output is connected to the input of the third convolutional layer. The output of the first convolutional layer is connected to the input of the second convolutional layer; the output of the third convolutional layer is connected to the input of the fourth convolutional layer; the input of adder A1 is connected to the outputs of the second and fourth convolutional layers respectively, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the wear diagnosis network.

8. The method for diagnosing turbine blade wear according to claim 7, characterized in that, The first matrix feature enhancement unit includes a fifth convolutional layer and a multiplier M1. The fifth convolutional layer is used to process the approximate residual feature matrix, and the multiplier M1 is used to apply weights to the output of the fifth convolutional layer using the approximate residual inconsistency matrix. The second matrix feature enhancement unit includes a sixth convolutional layer and a multiplier M2. The sixth convolutional layer is used to process the detail residual feature matrix, and the multiplier M2 is used to apply weights to the output of the sixth convolutional layer using the detail residual inconsistency matrix.

Images