CN117370879A

CN117370879A - Real-time online fault diagnosis method and system for wind turbine gearbox

Info

Publication number: CN117370879A
Application number: CN202311107990.4A
Authority: CN
Inventors: 刘琰; 何建军; 舒王咏; 吴畏; 朱祺; 廖力达; 颜景颐
Original assignee: Longyuan Jiangyong Wind Power Generation Co Ltd
Current assignee: Longyuan Jiangyong Wind Power Generation Co Ltd
Priority date: 2023-08-30
Filing date: 2023-08-30
Publication date: 2024-01-09

Abstract

The invention discloses a real-time online fault diagnosis method and a real-time online fault diagnosis system for a gearbox of a wind turbine, wherein the method comprises the steps of collecting a time domain signal of a vibration signal of the gearbox and carrying out Fourier transformation to obtain a frequency domain signal; extracting time domain features and frequency domain features; and extracting nonlinear principal component characteristics from the time domain characteristics and the frequency domain characteristics by using a pre-trained kernel principal component analysis model KPCA, calculating statistics SPE of the degree of deviation from the kernel principal component model based on the nonlinear principal component characteristics, judging that the state of the gearbox is normal if the statistics SPE of the degree of deviation from the kernel principal component model is smaller than a preset threshold, and otherwise judging that the state of the gearbox is abnormal. The invention aims to realize a real-time online fault diagnosis method of the wind turbine gearbox, which aims at the wind turbine gearbox, has the advantages of simple and easily obtained required data, high fault diagnosis efficiency, less required characteristic data sample data, and more accurate early fault signals and fault positioning, so as to avoid further deterioration of the state of the wind turbine gearbox.

Description

Real-time online fault diagnosis method and system for wind turbine gearbox

Technical Field

The invention relates to a wind turbine gearbox fault diagnosis technology in the field of wind power generation, in particular to a real-time online fault diagnosis method and system for a wind turbine gearbox.

Background

The gear box is an important mechanical connection and transmission device in the wind turbine generator, and has the characteristics of high transmission precision, fixed transmission ratio and the like. Due to the long-term complex working environment and the bearing of random variable load, the speed change gear box frequently fails and is easy to cause the shutdown of the unit, so that the service life of the unit is reduced. As the time of operation is accumulated, the reliability of the gearbox gradually decreases, and gear failure due to manufacturing errors, improper assembly, poor lubrication, overload, etc. also occurs. In addition, unstable motor voltage and time-varying rotor load under normal operation of the gearbox also cause fluctuation of the rotating speed of the gearbox, thereby affecting safe operation of the gearbox. The downtime caused by failure of each part of the gearbox transmission system accounts for about 1/3 of the total downtime of the unit, and is a main reason for influencing the reliability of the unit. Taking the case of failure overhaul of a gearbox as an example, the replacement cost of the component is usually millions of yuan/station, if the sub-health state of the equipment can be estimated in advance, the failure is changed into the advanced maintenance overhaul, the maintenance cost of the equipment can be greatly reduced, and the service life of the fan can be prolonged. The Chinese patent document with publication number of CN106777606A discloses a fault prediction diagnosis method of a wind turbine generator gearbox, which records historical sampling moment data of the temperature of the oil temperature of the gearbox, the temperature of a cabin, the temperature of the driving front end and the non-driving front end of the gearbox, the wind speed and the output power of a fan in operation, reduces the input dimension by a KPCA algorithm to extract characteristics and reuse a support vector machine for classification. However, this method has a complicated left oxygen collection, and requires the use of a Kernel Principal Component Analysis (KPCA) and a support vector machine, and requires a large amount of data to complete training of the Kernel Principal Component Analysis (KPCA) and the support vector machine, and the failure prediction diagnosis efficiency is low.

Disclosure of Invention

The invention aims to solve the technical problems: aiming at the problems in the prior art, the invention provides a real-time online fault diagnosis method and a real-time online fault diagnosis system for a wind turbine gearbox, which aim to realize the real-time online fault diagnosis method for the wind turbine gearbox, which is simple and easy to obtain required data, high in fault diagnosis efficiency, less in required characteristic data sample data, and capable of more accurately early fault signals and fault positioning, so as to avoid further deterioration of the state of the wind turbine gearbox.

In order to solve the technical problems, the invention adopts the following technical scheme:

a real-time online fault diagnosis method for a wind turbine gearbox comprises the following steps:

s101, collecting a time domain signal of a vibration signal of a gear box, and carrying out Fourier transformation to obtain a frequency domain signal;

s102, extracting time domain features from time domain signals of the vibration signals and extracting frequency domain features from frequency domain signals of the vibration signals;

s103, extracting nonlinear principal component features from the time domain features and the frequency domain features by using a pre-trained kernel principal component analysis model KPCA, calculating statistics SPE deviating from the degree of the kernel principal component model based on the nonlinear principal component features, judging that the state of the gearbox is normal if the statistics SPE deviating from the degree of the kernel principal component model is smaller than a preset threshold, and otherwise judging that the state of the gearbox is abnormal.

Optionally, extracting the time domain feature from the time domain signal of the vibration signal in step S102 includes: maximum value, minimum value, peak-to-peak value, mean square error, variance, square root amplitude, average amplitude, mean square amplitude, peak value, kurtosis, waveform index, peak value index, pulse index and margin index; extracting frequency domain features from frequency domain signals of the vibration signal includes: center of gravity frequency, root mean square of frequency, average frequency, and frequency variance.

Optionally, in step S103, the function expression for calculating the statistic SPE of the degree of deviation from the kernel principal component model is:

in the above formula, p is the number of characteristic data samples, gamma is the number of principal elements,is the kth ₁ Individual characteristic data samples, ++>Is the kth ₂ And the principal elements.

Optionally, step S103 is preceded by training a kernel principal component analysis model KPCA: aiming at the kernel function parameter gamma, testing the time domain characteristics and the frequency domain characteristics in a given value interval by using a designated step length, calculating the interpretation variance under different kernel function parameter gamma values, and selecting the kernel function parameter gamma with the maximum interpretation variance as the optimal kernel function parameter gamma of the kernel principal component analysis model KPCA, thereby obtaining the trained kernel principal component analysis model KPCA.

Optionally, step S103 further includes collecting an acoustic emission signal of the gearbox and performing fault diagnosis based on the acoustic emission signal, and if it is determined in step S104 that the state of the gearbox is abnormal and it is also determined that the state of the gearbox is abnormal based on the acoustic emission signal, determining that the gearbox is faulty; the fault diagnosis based on the acoustic emission signal comprises:

s201, performing multi-component modal decomposition on the acoustic emission signal to obtain a plurality of modal components;

s202, judging whether the modal components are noise or not based on variance contribution rate or correlation coefficient, and removing the noise;

s203, carrying out wavelet decomposition on the rest modal components and denoising based on wavelet threshold;

s204, recombining the modal components subjected to wavelet threshold denoising treatment to obtain a recombined signal;

s205, carrying out noise reduction processing on the recombined signal by using a graph model;

s206, performing fault diagnosis by using a fault map or a classifier according to the signals subjected to the noise reduction processing of the graph model so as to determine whether the state of the gearbox is normal or abnormal.

Optionally, in step S202, it is determined whether noise is generated based on the variance contribution rate for each modal component: and calculating variance contribution rates for the modal components respectively, and judging the modal component as noise if the variance contribution rate of one modal component is smaller than or equal to a set value.

Optionally, determining whether noise is generated for each modal component based on the correlation coefficient in step S202 includes: performing fast Fourier transform on each modal component, determining the center frequency of each modal component, sequencing the modal components according to the sequence from low center frequency to high center frequency, calculating the correlation coefficient of each modal component and an original signal, and judging that the modal component is noise if the correlation coefficient of one modal component is smaller than or equal to a set value, wherein the calculation function expression of the correlation coefficient is as follows:

in the above formula, C represents a correlation coefficient, v (t) is a modal component, x (t) is an original signal, E [ ] is a desire in mathematics, and D [ ] is a variance in mathematics.

Optionally, step S205 of performing noise reduction processing on the recombined signal using the graph model includes:

s301, exceeding the acoustic emission signal in the recombined signalA plurality of continuous measurement signals of a threshold value are used as nodes of a road map model to construct a road map model G= (V, E, W), wherein the road map model G= (V, E, W) is a simple map model with one vertex connected with the next vertex in sequence, and V= { V ₁ ,v ₂ ,v ₃ ,...,v _N The node set, v ₁ ～v _N For N nodes, e= { E (i, j) } represents an edge set, E (i, j) represents that the i node has a connection with the j node, w= { W (i, j) } is a weight set, W (i, j) represents a weight between the i node and the j node, and W (i, j) = 0 represents that no connection exists between the i node and the j node;

S302, for the road map model g= (V, E, W), filtering is performed by using a map filter represented by the following formula:

in the above formula, H represents the output signal of the graph filter, V represents the node set, H (λ ₁ )～h(λ _N ) Representing N frequency components lambda to the input ₁ ～λ _N Weighting parameters for reinforcement or weakness, Λ _h Representing a weight parameter matrix, and the transfer function of the graph filter has a functional expression of:

H(λ _k )＝1/(1+2αλ _k )

in the above formula, H (lambda) _k ) Is the transfer function of the graph filter, alpha is a coefficient, lambda _k Is the kth frequency component of the input.

In addition, the invention also provides a real-time online fault diagnosis system of the wind turbine gearbox, which comprises a microprocessor and a memory which are connected with each other, wherein the microprocessor is programmed or configured to execute the real-time online fault diagnosis method of the wind turbine gearbox.

Furthermore, the invention provides a computer readable storage medium having stored therein a computer program for being programmed or configured by a microprocessor to perform a real-time on-line fault diagnosis method of the wind turbine gearbox.

Compared with the prior art, the invention has the following advantages: the input data required by the invention is a time domain signal of the vibration signal of the gearbox, the required data is simple and easy to obtain, and the method is particularly suitable for real-time online fault diagnosis of the gearbox of the wind turbine. According to the invention, the state judgment of the gearbox can be realized on the basis of the statistics SPE only by using the nuclear principal component analysis model KPCA and calculating the statistics SPE deviating from the nuclear principal component model, so that the data processing is simple, the fault diagnosis efficiency is high, the required characteristic data sample data are few, the early fault signals and fault positioning can be more accurate, the further deterioration of the state of the gearbox of the wind turbine can be effectively avoided, the real-time accurate monitoring of the fault damage condition of the gearbox is achieved, the reliable and predictive health evaluation report is timely and effectively provided for the gearbox of the wind turbine, and the powerful support is provided for the safe and efficient operation of the wind turbine.

Drawings

FIG. 1 is a schematic diagram of a basic flow of a method according to an embodiment of the present invention.

FIG. 2 is a schematic flow chart of fault diagnosis based on acoustic emission signals comprehensively in an embodiment of the invention.

Fig. 3 is a schematic flow chart of fault diagnosis based on acoustic emission signals in an embodiment of the invention.

FIG. 4 is a graph showing the experimental results of solving the optimal kernel parameter gamma in the embodiment of the present invention.

FIG. 5 is a comparison of information contained in different principal components according to an embodiment of the present invention.

FIG. 6 is a schematic diagram illustrating classification between a data set of normal teeth and each faulty component after dimension reduction according to an embodiment of the present invention.

FIG. 7 is a graph of SPE statistic variation for a normal tooth training set in accordance with an embodiment of the invention.

FIG. 8 is a graph of variation of SPE statistics for a test dataset in accordance with an embodiment of the present invention.

Fig. 9 is a schematic structural diagram of a road map model according to an embodiment of the present invention.

Fig. 10 is a signal value corresponding to 10 threshold trigger nodes in an embodiment of the present invention.

Fig. 11 is a diagram illustrating a conventional signal-to-signal diagram in accordance with an embodiment of the present invention.

Fig. 12 is a schematic diagram of filtering functions of three types of filters commonly used in the embodiments of the present invention.

Detailed Description

As shown in fig. 1, the real-time online fault diagnosis method for the wind turbine gearbox of the embodiment includes:

In the embodiment, when the time domain signal of the vibration signal of the gear box is collected in step S101, the adopted sensor is an acceleration sensor, the detection position includes a main bearing, an input shaft and an output end of the gear box, the main bearing and the input shaft of the gear box use low-frequency acceleration sensors, and the output end of the gear box uses common-frequency acceleration sensors.

The extracting of the time domain features from the time domain signal of the vibration signal in step S102 of the present embodiment includes: maximum value, minimum value, peak-to-peak value, mean square error, variance, square root amplitude, average amplitude, mean square amplitude, peak value, kurtosis, waveform index, peak value index, pulse index and margin index; extracting frequency domain features from frequency domain signals of the vibration signal includes: center of gravity frequency, root mean square of frequency, average frequency and frequency variance. It should be noted that, the specific index items of the time domain feature and the frequency domain feature are all known index items, and the key of the method of this embodiment is that the combination of the specific index items of the time domain feature and the frequency domain feature and the error between the statistics SPE of the degree of deviation from the principal component model corresponding to the combination are found and the preset threshold value can realize the normal/abnormal resolution of the gear box state, so the calculation method of the specific index items of the time domain feature and the frequency domain feature and the calculation function expression thereof are not described in detail herein.

In step S103 of this embodiment, before extracting nonlinear principal component features from time domain features and frequency domain features by using a pre-trained kernel principal component analysis model KPCA, normalization processing is first required for the time domain features and the frequency domain features. The data normalization processing formula is as follows:

in the method, in the process of the invention,a value normalized for the data; x is x _j J data is input; x is x _max And x _min Is the maximum and minimum of the sample data.

In step S103 of this embodiment, the nonlinear principal component feature is extracted by using the kernel principal component analysis model KPCA as an existing method. The kernel principal component analysis model KPCA is an improved method for principal component analysis model PCA. Principal component analysis model PCA can remove noise, explore high-dimensional data and visualize, but still has shortcomings in processing nonlinear data. The kernel principal component analysis model KPCA can optimize redundancy and spatial correlation of PCA among elimination data, and extract nonlinear characteristic principal components containing main information. Assume that training sample T of the gearbox model is:

in the above formula, m is the number of sample features, n is the number of data samples, and each row is m features of one sample.

Passing training sample T through nonlinear mapping phi: R ⁿ Mapping to a high-dimensional feature space:

φ(x)＝{φ(x ₁ ),φ(x ₂ ),…,φ(x _n )}，

Wherein R is ⁿ Is the original space; f is the high-dimensional space after nonlinear mapping, phi (x) is the high-dimensional space integral sample, phi (x) ₁ )～φ(x _n ) Is the 1 st to n th high-dimensional space sample.

And (3) carrying out centralization treatment on the mapped data set to ensure that the data set meets the following conditions:

in the above, phi (x) _i ) Is the i-th high dimensional spatial sample.

Covariance matrix C of samples ^F The expression of the calculation function of (c) is:

covariance features decompose into:

λω＝C ^F ω，

in the above formula, lambda is covariance matrix C ^F ω is a feature vector, and has:

in the above, alpha _i For the ith high dimensional spatial sample phi (x _i ) And the jth high-dimensional spatial sample phi (x _j ) Is a correlation coefficient of phi (x) _j ) Is the j-th high-dimensional spatial sample.

To solve the nonlinear mapping phi, a kernel function k in the form of a matrix is introduced, the elements of the matrix satisfying:

[k] _ij ＝(φ(x _i ),φ(x _j ))，

wherein [ k ]] _ij The ith row and jth column of kernel function k. Centralizing and simplifying the data in a high-dimensional feature space, and finally calculating a principal component t of the sample data _k ：

In the above, ω _k Alpha, which is the kth component in the feature vector _i ^k As the correlation coefficient, k is a kernel function, and in this embodiment, the kernel function k is a gaussian radial kernel function:

in the above formula, σ is a width parameter of the kernel function, and γ is a kernel function parameter gamma. The number of principal components can be obtained from the accumulated variance.

The statistics SPE of the degree of deviating from the kernel principal component model is a measure of the degree of deviating from the kernel principal component model and reflects the process of the data deviating from normal correlation change. In step S103, the function expression of the statistic SPE for calculating the degree of deviation from the kernel principal component model is:

The spatial distribution of data can be different due to different gamma values of kernel function parameters of a kernel principal component analysis model KPCA, and when the gamma values of the kernel function parameters are too small, the model can be in an under-fitting state; when the gamma value of the kernel function parameter is too large, the model is in an overfitting state. Therefore, when performing the kernel principal component analysis model KPCA processing, a proper kernel function parameter gamma value needs to be selected. The step S103 of this embodiment further includes training the kernel principal component analysis model KPCA: aiming at the kernel function parameter gamma, testing the time domain characteristics and the frequency domain characteristics in a given value interval by using a designated step length, calculating the interpretation variance under different kernel function parameter gamma values, and selecting the kernel function parameter gamma with the maximum interpretation variance as the optimal kernel function parameter gamma of the kernel principal component analysis model KPCA, thereby obtaining the trained kernel principal component analysis model KPCA.

Specifically, the present embodiment performs a test on the kernel parameter gamma value from the interval (0, 15) with a step size of 0.01, and uses the interpretation variance as a measure, where the interpretation variance is the proportion of the variance interpreted by the principal component. And selecting the gamma value of the kernel function parameter with the maximum interpretation variance as the gamma value of the kernel function parameter of the model. The results are shown in FIG. 4. The optimal kernel parameter gamma value of the model was 0.13, at which point the interpretation variance was 0.55629.

The optimal kernel function parameter gamma value is brought into a kernel principal component analysis model KPCA, the contribution rate of the first fifteen principal components is calculated, the first three principal components are selected for visual analysis, and the result is shown in figure 5, and after the kernel principal component analysis model KPCA is processed, the information contained in the first three principal components (PC 1-PC 3) is up to 80%. As shown in fig. 6, the classification between the reduced-dimension normal tooth data set and each faulty piece is obvious, and there is also an obvious cluster between each faulty gear data set.

And (3) applying an optimal kernel function parameter gamma value, calculating and selecting confidence coefficient of the SPE threshold value to be 0.95, and selecting the first 3 principal components. As a result, as shown in fig. 7, the SPE values calculated for the normal tooth training set are mostly below the threshold. As shown in fig. 8, the SPE values of the four tooth-shaped tests are shown, the SPE values of the normal tooth have individual points exceeding the SPE threshold value, and the false report exists, but the SPE values of most points are lower than the SPE threshold value, so that the actual monitoring situation is met. For fault diagnosis of three faulty tooth types, the SPE values are mostly above the calculated SPE threshold. Therefore, the gear box is monitored in real time by calculating the control limit of the SPE statistics, an early fault signal of the gear can be found in time, an alarm prompt is sent, workers are reminded to repair and replace the gear in advance, further deterioration of the gear is prevented, and larger loss is caused.

As shown in fig. 2, step S103 of the present embodiment further includes collecting an acoustic emission signal of the gear box and performing fault diagnosis based on the acoustic emission signal, and if it is determined in step S104 that the gear box is abnormal in state and it is also determined that the gear box is abnormal based on the acoustic emission signal, it is determined that the gear box is faulty. In this embodiment, an acoustic emission sensor for acquiring acoustic emission signals is mounted on the low-speed shaft bearing to acquire acoustic emission signals of the gearbox.

As shown in fig. 3, the present embodiment performs fault diagnosis based on acoustic emission signals including:

In step S202 of this embodiment, whether noise is generated or not is determined based on the variance contribution ratio for each modal component: and calculating variance contribution rates for the modal components respectively, and judging the modal component as noise if the variance contribution rate of one modal component is smaller than or equal to a set value. The variance contribution rate is an existing statistical index, specifically, refers to the ratio of the variation caused by a single common factor to the total variation, and illustrates the influence of the common factor on the dependent variable, so the specific calculation method and the functional expression thereof are not described in detail herein.

In step S202 of this embodiment, determining whether noise is generated based on the correlation coefficient for each modal component includes: performing fast Fourier transform on each modal component, determining the center frequency of each modal component, sequencing the modal components according to the sequence from low center frequency to high center frequency, calculating the correlation coefficient of each modal component and an original signal, and judging that the modal component is noise if the correlation coefficient of one modal component is smaller than or equal to a set value, wherein the calculation function expression of the correlation coefficient is as follows:

In the process of acquisition and transmission of acoustic emission signals, the acoustic emission signals are easily influenced by external environment interference and self experimental machines, and noise interference is generated to influence signal analysis and recognition. If the characteristics of the signal are difficult to accurately analyze by directly researching the signal, a measure method is needed to process the data, and interference signals are filtered to achieve the effect of noise reduction. The purpose of signal noise reduction is to display more parts of the signal itself and to suppress the influence of the interference signal on the signal itself. Step S205 of the present embodiment of performing noise reduction processing on the recombined signal using a graph model includes:

s301, constructing a road map model G= (V, E, W) by taking a plurality of continuous measurement signals exceeding the threshold value of the acoustic emission signal in the recombined signal as nodes of the road map model, wherein the road map model G= (V, E, W) is a simple map model with one vertex sequentially connected with the next vertex, and V= { V ₁ ,v ₂ ,v ₃ ,...,v _N The node set, v ₁ ～v _N For N nodes, e= { E (i, j) } represents an edge set, E (i, j) represents that the i node has a connection with the j node, w= { W (i, j) } is a weight set, W (i, j) represents a weight between the i node and the j node, and W (i, j) = 0 represents that no connection exists between the i node and the j node;

H(λ _k )＝1/(1+2αλ _k )

In the embodiment, step S205 performs noise reduction processing on the recombined signal by using a graph model to achieve the effect of noise reduction, and the method is different from classical signal noise reduction analysis and is a new approach to signal processing. In classical signal processing, the signal domain is determined by equidistant time points or a set of spatially perceived points on a uniform grid. However, in the field of actual data awareness, it may be irrelevant to the physical dimensions of time or space, but exhibit various forms of irregular properties, such as irregular topologies belonging to specific objects in a social network. It is important to note that the data obtained in the determined temporal and spatial domains, by graphically introducing new relationships between signal samples, may yield new insights into the analysis and provide better data processing applications (noise reduction). Because the signal data acquired by acoustic emission are reflected by combining the attribute of the material, a road map model G= (V, E, W) can be constructed. The road map model is a simple map model with one vertex connected with the next vertex in turn, can be well matched with a network model with time translation characteristics, but the head-to-tail nodes are not connected as shown in fig. 9. In the experiment, the threshold value (47 dB) of the acoustic emission signal is set, so that the threshold trigger time point of the measurement signal is designated as a graph vertex; the connection line between the trigger point representing the threshold and the point is a graph edge and therefore represents a discrete signal. In FIG. 9, vertices of the road graph model The set may represent v= { V ₁ ,v ₂ ,v ₃ ,v ₄ ,v ₅ ,..}, the edge set of the road graph model may also be represented as e= { (v) ₁ ,v ₂ ),(v ₂ ,v ₃ ),(v ₃ ,v ₄ ),(v ₄ ,v ₅ ),...}. We choose the signals measured by ten consecutive threshold trigger nodes in the measurement process, as shown in fig. 10, and the measured signals can be converted into signal samples on the graph, as shown in fig. 11. Similar to the conventional signal processing method, various signal processing methods including noise reduction, filtering, and the like can be performed.

Signal shifting operators are key to discrete signal processing. The definition on the figure is difficult to see due to its rich underlying connectivity structure. The traditional time domain signal is analogiced, the current time node signal can only be connected with the next time node signal, and the lower layer structure of the picture signal is a high-dimensional structure and basically is in one-to-many condition. The road map structure in the acoustic emission acquisition signal is used for noise reduction, and is a one-to-two structure. From the perspective of the underlying structure of the signal, the signal shift on the graph can be seen as a movement of the nodes with which the signal samples have a weight relationship, which can be expressed as:

x _shifted ＝Sx，

in the above, x _shifted For the shifted signal, x is the input signal, and the parameter matrix S can be expressed as an adjacent matrix a, a laplace matrix L, a weighting matrix W, and a symmetric normalized laplace matrix D ^-1/2 LD ^1/2 Random walk weighting matrix D ^- ¹ W. The graph signal processing forms a linear graph shift invariant system, and converts an input graph signal into an output graph signal, and the output graph signal y can be written as:

in the above, h ₀ ,h ₁ ,h ₂ ,...,h _N-1 Is a system coefficient. For standard road map structure path diagram, the map shift operator S can be converted into weighting matrix, laplacian operator, etc., i.e. map letterThe number system is equivalent to the expression:

in the above, h _n Representing the nth system coefficient. The laplace matrix is adopted in the embodiment, and belongs to the common operator for graph signal processing, which is represented by the letter L:

L＝D-W，

where D is a degree matrix and W is a weight matrix, so the discrete signal y (n) outputting the image signal y can be rewritten as:

y(n)＝h ₀ x(n)+h ₁ Lx(n-1)+...+h _M-1 L ^M-1 x(n-M+1)，

in the above formula, M is the window size, and x (n) -x (n-M+1) are the n-th to n-M+1-th input signals.

In this embodiment, the first five nodes are selected as an example, and when only whether the nodes are connected is considered, the adjacency matrix a and the degree matrix D are expressed as follows:

when the relation of the weight is increased in consideration of the contact, multiple factors are considered in the weight setting between the two threshold nodes, and a weight standard of a relative standard can be obtained. In experimental data processing, we select the euclidean distance between two node signal values as the weight size in the road map, and the calculation formula is as follows:

w _ij ＝|x _i -x _j |，

In the above, w _ij For node signal x _i And x _j Weights between, |x _i -x _j I is node signal x _i And x _j Euclidean distance between them.

The laplace matrix L is represented as follows:

thus, a graph signaling system can be represented by the following equation:

y＝H(S)x，

in the above equation, y is an output signal, x is an input signal, and H (S) is a graph transfer function.

In the conventional signal processing, the adjacent nodes are averaged by the moving average operator to achieve the effect of noise reduction, however, the adjacency relationship between the nodes in the graph signal processing system needs to be represented by considering a piece of quantitative data, so that the transformation relationship between the road graph signal and the acoustic emission signal can be established.

The road map signals corresponding to the acoustic emission acquisition signals are expressed as follows in a matrix form:

x＝[x ₀ ,x ₁ ,x ₂ ,...,x _t ] ^T ，

in the above, x ₀ ～x _t Respectively nodes in the road map signal. The acoustic emission signal collected in the experiment essentially consists of the original signal and the noise signal, and each collected signal can be expressed by the expression:

x(t)＝x ₀ (t)+φ(t)

in the above, x (t) is an acoustic emission acquisition signal at the moment t, x ₀ (t) is the original signal (true acoustic emission signal), and phi (t) is the noise signal. The signal-to-noise ratio (SNR) is the ratio of signal to noise in an acoustic emission acquisition system and is an important indicator for evaluating signal quality. When a higher ratio indicates less noise interference, the expression can be expressed as follows:

SNR＝10×lg(PS/PN)，

In the above, x ₀ (i) As original signal (true acoustic emission signal), x ₁ (i) For the noise-reduced signal, the larger the signal-to-noise ratio is when noise is reduced, the better the noise reduction effect is. Another criterion is Root Mean Square Error (RMSE):

wherein x is ₁ (i) For the noise-reduced signal, the smaller the root mean square error when the signal is noise reduced, the better the noise reduction effect is explained. Therefore, the signal-to-noise ratio and the root mean square error can be used as indexes to measure whether the purpose of noise reduction is achieved.

In the graph signaling system, when we consider only the links between nodes, when the graph move operator is a, for each node, its accumulated signal value can be expressed as:

in the above formula, x (m) is an acoustic emission acquisition signal at m time. In the road map model, since any one node only has an adjacent relationship with two nodes, it can be expressed as:

y(n)＝x(n)+x(n-1)+x(n+1)，

if the relationships of the nodes are arranged in a matrix structure:

y＝x+Ax，

thus, matrix a is referred to as a graph adjacency matrix indicating the connected structure of nodes. When the element in the matrix A is 0, the corresponding node is not associated; when the element is 1, it is stated that the corresponding node is associated. From fig. 2, it can be seen that the upper diagonal and the lower diagonal elements of the adjacency matrix are 1, and the other elements are all 0. When we consider not only the connections between nodes, but also the connection size, that is, the weight between nodes, there are:

In the above, W _mn Indicating the degree of coupling between the two nodes, i.e. the weight is small, the greater the weight the greater the correlation. W (W) _mn =0 means that there is no correlation between m node and n node, or m=n. Then there are:

y＝x+Wx

in the above formula, W is a weight matrix.

To produce an unbiased estimate, the weighting coefficient for each y (n) is kept at 1, and the number of y (n) divided by the number of correlation points is the signal value for the current n points. Normalizing the expression to obtain:

y＝0.5(x+D ^-1 Wx)，

in the above formula, D is a degree matrix,noise reduction by using the simple normalization filtering, S/N-R calculation _out Value and RMSE value, if the calculated difference satisfies:

K ₁ ＝(S/N-R _out )-(S/N-R _in )>0，

K ₂ ＝(RMSE _out )-(RMSE _in )<0，

the explanation has a certain noise reduction effect. Wherein S represents an adjacent matrix A, a Laplace matrix L, a weighting matrix W and a symmetrical normalized Laplace matrix D ^-1/2 LD ^1/2 Or random walk weighting matrix D ^-1 W. N is the length of the picture signal, R _out R is the amplitude of the output _in For the amplitude of the input, RMSE _out For the mean square error RMSE of the output _in Is the average error of the input.

In conventional signal processing, filtering is the operation of filtering out a particular band of wavelengths in a signal. By analogy to the plot signal, we increase or decrease the intensity of each frequency component in the plot signal to achieve the filtering effect. The Fourier change of the traditional signal is analogized, the frequency analysis of the traditional signal is completed in the Fourier domain, and the frequency analysis of the graph signal utilizes the characteristic frequency of the graph shifting operator S. Assuming that the graph filter is H, the output graph signal is y:

In the above formula, h (lambda) _k ) For the filter coefficients of the filter at frequency ak,is formed by linear addition of eigenvectors of Laplace matrix, and the added coefficients are transformed vectors, v ₁ ～v _N For N eigenvectors, < >>Is formed by linear addition of eigenvectors of Laplace matrix at 1-N, and the added coefficients are transformed vectors, h (lambda) ₁ )～h(λ _N ) The filter coefficients of the filter at the frequencies lambda 1-lambda N are obtained, and V is a node set in the graph;

then there are:

in the above, Λ _h Represents h (lambda) on the left side ₁ )～h(λ _N ) A diagonal matrix is formed.

On the basis of this, h (lambda) _k ) Whether the signal has a frequency component that is increased or decreased in magnitude. The graph filter exhibits the following three features: (1) linear relationship: h (x+y) =hx+hy; (2) filtering order independence: h ₁ (H ₂ x)＝H ₂ (H ₁ x); (3) if h (λ) +.0, the filtering is reversible. Wherein H (x+y) represents a graph filter processing structure of x and y, hx represents a graph filter processing structure of x, hy represents a graph filter processing structure of y, H ₁ And H ₂ The filter h (λ) is the filter coefficients of the filter at the corresponding frequency λ for two different graphs.

Different filter response functions may result in complete failureThe same filtering effect is aimed at the treatment of practical problems, and three common types of filters are a high-pass filter, a low-pass filter and a band-pass filter, and the general trend graph is shown in fig. 12. As can be derived from fig. 12, the low-pass filter is more focused on the low-frequency part of the signal, i.e., the part where the signal is stationary, the band-pass filter is more focused on the specific frequency part of the signal, and the high-pass filter is more focused on the high-frequency part of the signal, i.e., the part where the signal oscillations are larger. The process of graph signal filtering is shown through the random sensor network node, and a low-pass filter response function 1/(1+lambda) is selected in the embodiment ² ). The experimentally obtained acoustic emission signal can be considered to consist of the original signal and the rapidly changing interference noise signal, which can be expressed by the expression:

in the above, x is an acoustic emission acquisition signal, x ₀ As the original signal (the true acoustic emission signal),is a noise signal. In the embodiment, a better-effect graph signal filter is designed to restrain partial abnormal signals and noise signals so as to reduce noise. If the graph filter output signal is y, then an optimal denoising task is defined by minimizing the cost function.

In the above formula, α is a coefficient.

In order to make the previous term in the above expression as small as possible let the observed signal x approach the output signal y, while the second term represents the smoothness of the signal as mentioned in the fourier transform of the previous section of the graph, and the coefficient α is chosen by a priori knowledge for smoothing this expression. Because of the function type presented by the above formula, when the derivative is zero, the corresponding value is minimum, and then:

ideally there should be:

y＝(I+2αL) ^-1 x，

in the above formula, L is a laplace matrix. Will thus be transformed into a matrix form, the graph filter transfer function being:

H(λ _k )＝1/(1+2αλ _k )，

according to the formula, as the coefficient alpha increases, the smoothness of the output signal increases, a good noise reduction result suitable for the coefficient alpha value is selected, in an experiment, amplitude signals of 0 to 1086 seconds in the whole stretching process are selected, a road map model algorithm is adopted for noise reduction, and the influence of different coefficients alpha on noise reduction is studied to determine the noise reduction coefficient with good effect.

Different coefficients alpha are selected to reduce noise of the signal, and the signal-to-noise ratio and the root mean square variance are calculated to obtain a better noise reduction coefficient, as shown in table 1.

Table 1 influence of different coefficients α on noise reduction effect.

Alpha coefficient	3	4	5	6	7
						Signal to noise ratio	5.5	8.3	11.2	10.5	10.8
Root mean square variance	0.225	0.231	0.210	0.235	0.218

Referring to table 1, when the parameter α is 5, the corresponding noise reduction effect is relatively most obvious, and the transfer function of the graph filter has the following function expression:

H(λ _k )＝1/(1+10λ _k )，

in the above formula, H (lambda) _k ) Lambda is the transfer function of the graph filter _k Is the kth frequency component of the input.

Through verification, the threshold value is set to be 47dB by utilizing a physical noise reduction mode, most of environmental noise can be eliminated, and in addition, friction between the clamp and a test piece can be reduced by adopting a soft pad, so that friction noise of the clamp is eliminated.

As an optional implementation manner, the embodiment further includes a step of visually outputting a fault diagnosis result of performing fault diagnosis based on the vibration signal and the acoustic emission signal, and includes some other auxiliary sensors and auxiliary diagnosis: the oil product monitoring sensor is arranged on an oil pipe between the oil return port and the filter and is used for detecting the quality of oil products so as to ensure the best working environment of the related components, detecting the quality of the oil products, wherein the quality of the oil products comprises ferromagnetic particles, nonferromagnetic particles, temperature, density, viscosity and moisture of the lubricating oil of the gear box, and the oil product monitoring sensor can be used for analyzing whether the lubricating oil needs to be replaced or not and whether the gear box is healthy and reliable or not. The camera and the infrared thermal imager are used for monitoring the fault type of the small change of the position caused by the gearbox and the fault reflected by the change of the regional temperature distribution by adopting a method of an image recognition method. When a fault occurs, the overall or local thermal balance of the gearbox device is also disrupted or affected, and the heat inside the device must gradually reach its outer surface, causing a change in the outer surface temperature field. The power equipment state monitoring sensor is used for collecting stator current signals of the wind turbine, and the characteristic frequency of gear faults is identified by using stator current signal analysis, so that fine abrasion which cannot be captured by the acceleration sensor is captured.

In summary, the real-time online fault diagnosis method of the embodiment has the advantages of simple data processing, high fault diagnosis efficiency, less required characteristic data sample data, more accurate early fault signals and fault positioning, and can effectively avoid further deterioration of the state of the gearbox of the wind turbine, achieve real-time accurate monitoring of the fault damage condition of the gearbox, timely and effectively provide reliable and predictive health assessment reports for the gearbox of the wind turbine, and provide powerful support for safe and efficient operation of the wind turbine.

In addition, the embodiment also provides a real-time online fault diagnosis system of the wind turbine gearbox, which comprises a microprocessor and a memory which are connected with each other, wherein the microprocessor is programmed or configured to execute the real-time online fault diagnosis method of the wind turbine gearbox. The present embodiment also provides a computer readable storage medium having a computer program stored therein for being programmed or configured by a microprocessor to perform a real-time on-line fault diagnosis method of the wind turbine gearbox.

It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein. The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the present invention may occur to one skilled in the art without departing from the principles of the present invention and are intended to be within the scope of the present invention.

Claims

1. A real-time online fault diagnosis method for a wind turbine gearbox is characterized by comprising the following steps:

2. The real-time online fault diagnosis method of a wind turbine gearbox of claim 1, wherein extracting the time domain features from the time domain signals of the vibration signal in step S102 comprises: maximum value, minimum value, peak-to-peak value, mean square error, variance, square root amplitude, average amplitude, mean square amplitude, peak value, kurtosis, waveform index, peak value index, pulse index and margin index; extracting frequency domain features from frequency domain signals of the vibration signal includes: center of gravity frequency, root mean square of frequency, average frequency, and frequency variance.

3. The real-time online fault diagnosis method of a wind turbine gearbox according to claim 1, wherein the statistics SPE of the degree of deviation from the kernel principal component model in step S103 has a functional expression as follows:

in the above formula, p is the number of characteristic data samples, n is the number of principal elements,is the kth ₁ Individual characteristic data samples, ++>Is the kth ₂ And the principal elements.

4. The real-time online fault diagnosis method of a wind turbine gearbox according to claim 1, further comprising training a kernel principal component analysis model KPCA before step S103: aiming at the kernel function parameter gamma, testing the time domain characteristics and the frequency domain characteristics in a given value interval by using a designated step length, calculating the interpretation variance under different kernel function parameter gamma values, and selecting the kernel function parameter gamma with the maximum interpretation variance as the optimal kernel function parameter gamma of the kernel principal component analysis model KPCA, thereby obtaining the trained kernel principal component analysis model KPCA.

5. The real-time online fault diagnosis method of a wind turbine gearbox according to claim 1, wherein step S103 further comprises collecting acoustic emission signals of the gearbox and performing fault diagnosis based on the acoustic emission signals, and if the gearbox state is determined to be abnormal in step S104 and the gearbox state is also determined to be abnormal based on the acoustic emission signals, determining that the gearbox is faulty; the fault diagnosis based on the acoustic emission signal comprises:

6. The real-time online fault diagnosis method of a wind turbine gearbox according to claim 5, wherein in step S202, whether noise is determined based on variance contribution ratio for each modal component: and calculating variance contribution rates for the modal components respectively, and judging the modal component as noise if the variance contribution rate of one modal component is smaller than or equal to a set value.

7. The real-time online fault diagnosis method of a wind turbine gearbox according to claim 5, wherein the determining whether noise is based on the correlation coefficient for each modal component in step S202 comprises: performing fast Fourier transform on each modal component, determining the center frequency of each modal component, sequencing the modal components according to the sequence from low center frequency to high center frequency, calculating the correlation coefficient of each modal component and an original signal, and judging that the modal component is noise if the correlation coefficient of one modal component is smaller than or equal to a set value, wherein the calculation function expression of the correlation coefficient is as follows:

8. The real-time online fault diagnosis method of a wind turbine gearbox of claim 5, wherein step S205 of performing noise reduction processing on the recombined signal using a graph model comprises:

s301, constructing a road map model G= (V, E, W) by taking a plurality of continuous measurement signals exceeding the threshold value of the acoustic emission signal in the recombined signal as nodes of the road map model, wherein the road map model G= (V, E, W) is a simple map model with one vertex sequentially connected with the next vertex, and V= { V ₁ ,v ₂ ,v ₃ ,...,v _N Is represented by }Node set, v ₁ ～v _N For N nodes, e= { E (i, j) } represents an edge set, E (i, j) represents that the i node has a connection with the j node, w= { W (i, j) } is a weight set, W (i, j) represents a weight between the i node and the j node, and W (i, j) = 0 represents that no connection exists between the i node and the j node;

H(λ _k )＝1/(1+2αλ _k )

9. A real-time on-line fault diagnosis system for a wind turbine gearbox comprising a microprocessor and a memory connected to each other, characterized in that the microprocessor is programmed or configured to perform a real-time on-line fault diagnosis method for a wind turbine gearbox according to any of claims 1-8.

10. A computer readable storage medium having a computer program stored therein, wherein the computer program is for being programmed or configured by a microprocessor to perform the real-time on-line fault diagnosis method of a wind turbine gearbox according to any one of claims 1-8.