Fault diagnosis system for wind turbine and method thereof

Abstract

The application relates to the field of wind turbine generators, and particularly discloses a fault diagnosis system and method of a wind turbine generator, which uses a convolutional neural network model based on deep learning as a feature extractor to generate a vibration waveform based on rotating speed implicit features of a rotating shaft through a waveform generator, i.e. a self-vibration signal caused by an excitation signal generated by rotation of the rotating shaft, and uses a difference matrix between the generated vibration waveform and a vibration signal detected at a head end of a generator to represent vibration features of the head end of the generator after the self-vibration signal is filtered out, and then uses a change feature between the vibration features of the head end of the generator and vibration features of a vibration signal detected at a tail end of the generator to represent working features of the wind turbine generator, so that fault diagnosis of the wind turbine generator is performed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generating sets, and more particularly to a fault diagnosis system of a wind power generating set and a method thereof. BACKGROUND

[0002] In recent years, the energy industry transformation process in China has been accelerating, and China has become the largest market in the world in terms of wind power generation scale and growth. As a renewable energy source, wind energy has the characteristics of green and clean, and is the research focus of countries around the world. Compared with onshore wind power, offshore wind resources are more abundant, have the advantages of large power generation, long power generation time, small impact on living environment, no occupation of arable land, and large-scale development, etc. Therefore, developing offshore wind power generation technology has become a new trend in the development of the wind power industry.

[0003] The wind power generating set is the core component of the offshore wind turbine, and if it fails, it will directly affect the power generation efficiency of the offshore wind turbine, so once a fault occurs, the wind power generating set needs to be repaired in a timely manner. When repairing, the fault mode of the wind power generating set needs to be determined first, so that the correct repair mode can be adopted. The existing fault mode identification of the wind power generating set is analyzed by experienced technicians, and this fault diagnosis mode not only has low efficiency, but also is prone to missed detection and false detection.

[0004] Therefore, a fault diagnosis scheme for a wind power generating set is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a fault diagnosis system of a wind power generating set and a method thereof, which uses a convolutional neural network model based on deep learning as a feature extractor to obtain a generated vibration waveform based on the rotational speed implicit features of a rotating shaft through a waveform generator, i.e. a self-vibration signal caused by an excitation signal generated by the rotation of the rotating shaft, and uses the difference matrix between the generated vibration waveform and the vibration signal detected at the head end of the generator to represent the vibration features at the head end of the generator after filtering out the self-vibration signal, and then represents the working features of the wind power generating set based on the change features between the vibration features detected at the head end of the generator and the vibration features of the vibration signal detected at the tail end of the generator, so as to perform fault diagnosis of the wind power generating set.

[0006] According to an aspect of the present application, a fault diagnosis system of a wind power generating set is provided, which comprises:

[0007] A rotational speed data acquisition module is configured to acquire rotational speed values of a generator at a plurality of predetermined time points within a predetermined time period.

[0008] The vibration data collection module is configured to acquire first vibration signals and second vibration signals collected by first vibration sensors and second vibration sensors arranged at the head end and the tail end of the generator;

[0009] The multi-scale time series data encoding module is configured to arrange the rotation speed values of the generator at a plurality of predetermined time points in the predetermined time period into a rotation speed input vector in the time dimension, and then obtain a multi-scale rotation speed feature vector through a multi-scale neighborhood feature extraction module;

[0010] The generation module is configured to obtain a generated vibration waveform diagram through a waveform generator based on the multi-scale rotation speed feature vector;

[0011] The feature extraction module is configured to obtain a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix through a convolutional neural network model serving as a feature extractor based on the generated vibration waveform diagram, a waveform diagram of the first vibration signals, and a waveform diagram of the second vibration signals;

[0012] The fusion optimization module is configured to fuse the generated vibration feature matrix and the generated vibration waveform diagram to obtain an optimized generated vibration feature matrix;

[0013] The filtering module is configured to calculate a difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix;

[0014] The transfer module is configured to calculate a transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as a classification feature matrix; and

[0015] The fault diagnosis result generation module is configured to obtain a classification result through a classifier based on the classification feature matrix, and the classification result is used to represent a probability value of the wind turbine generator to have each fault category label.

[0016] In the fault diagnosis system of the wind turbine generator set, the multi-scale time series data encoding module comprises: a vector construction unit configured to arrange speed values of the generator at a plurality of predetermined time points in a predetermined time period into a speed input vector according to a time dimension; a first convolution unit configured to input the speed input vector into a first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale speed correlation feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel with a first length; a second convolution unit configured to input the speed input vector into a second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale speed correlation feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel with a second length, and the first length is different from the second length; and a concatenation unit configured to concatenate the first neighborhood scale speed correlation feature vector and the second neighborhood scale speed correlation feature vector to obtain the multi-scale speed feature vector.

[0017] In the fault diagnosis system of the wind turbine generator set, the first convolution unit is further configured to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution coding on the speed input vector according to the following formula to obtain the first neighborhood scale speed correlation feature vector.

[0018] The formula is:

[0019]

[0020] wherein a is the width of the first convolution kernel in the x direction, F is a first convolution kernel parameter vector, G is a local vector matrix for convolution kernel function operation, w is the size of the first convolution kernel, and X represents the tension input vector; the second convolution unit is further configured to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution coding on the speed input vector according to the following formula to obtain the second neighborhood scale speed correlation feature vector.

[0021] The formula is:

[0022]

[0023] wherein b is the width of the second convolution kernel in the x direction, F is a second convolution kernel parameter vector, G is a local vector matrix for convolution kernel function operation, m is the size of the second convolution kernel, and X represents the tension input vector.

[0024] In the fault diagnosis system of the wind turbine generator set, the waveform generator is a generative adversarial network.

[0025] In the fault diagnosis system of the wind turbine set, the feature extraction module is further configured to: use each layer of the convolutional neural network model serving as a feature extractor to respectively perform, in forward transmission of the layer, on input data: convolution processing on the input data to obtain a convolution feature map; local channel dimension-based mean pooling on the convolution feature map to obtain a pooling feature map; and non-linear activation on the pooling feature map to obtain an activated feature map; and wherein an output of a last layer of the convolutional neural network model serving as the feature extractor is the generated vibration feature matrix, the first vibration waveform feature matrix, and the second vibration waveform feature matrix, and an input of a first layer of the convolutional neural network model serving as the feature extractor is the generated vibration waveform graph, the waveform graph of the first vibration signal, and the waveform graph of the second vibration signal.

[0026] In the fault diagnosis system of the wind turbine set, the fusion optimization module is further configured to: fuse the generated vibration feature matrix and the generated vibration waveform graph to obtain the optimized generated vibration feature matrix according to the following formula:

[0027] wherein the formula is:

[0028]

[0029] wherein M1 represents the generated vibration waveform graph, M2 represents the generated vibration feature matrix, f i represents a feature value at each position in the generated vibration feature matrix, represents a mean value of values at each position in the generated vibration feature matrix, N is a dimension of the generated vibration feature matrix, exp(·) represents an exponential operation on a matrix, the exponential operation on the matrix represents a natural exponential function value with a feature value at each position in the matrix as a power, and a is a hyperparameter, represents a position-wise addition of a feature matrix, and represents a position-wise point multiplication.

[0030] In the fault diagnosis system of the wind turbine set, the filtering module is further configured to: calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix according to the following formula:

[0031] wherein the formula is:

[0032]

[0033] wherein M P represents the optimized generated vibration feature matrix, M V represents the first vibration waveform feature matrix, M D represents the difference feature matrix, represents the position difference.

[0034] In the wind turbine fault diagnosis system, the transition module is further configured to calculate a transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix as the classification feature matrix according to a formula.

[0035] The formula is as follows:

[0036] M W = M*M D

[0037] wherein M W represents the second vibration waveform feature matrix, M D represents the difference feature matrix, and M represents the classification feature matrix.

[0038] In the wind turbine fault diagnosis system, the fault diagnosis result generation module is further configured to process the classification feature matrix according to a formula to generate the classification result using the classifier.

[0039] The formula is as follows:

[0040]

[0041] wherein O is an output result matrix, W i and b i are a weight matrix and a bias matrix corresponding to the i-th classification, respectively, exp(·) represents an exponential operation on a matrix, and the exponential operation on the matrix represents a natural exponential function value with the feature value at each position in the matrix as the power.

[0042] According to another aspect of the present application, a wind turbine fault diagnosis method comprises:

[0043] obtaining a rotation speed value of a generator at a plurality of predetermined time points in a predetermined time period;

[0044] obtaining a first vibration signal and a second vibration signal collected by a first vibration sensor and a second vibration sensor arranged at a head end and a tail end of the generator;

[0045] arranging the rotation speed value of the generator at the plurality of predetermined time points in the predetermined time period in a time dimension into a rotation speed input vector, and then obtaining a multi-scale rotation speed feature vector through a multi-scale neighborhood feature extraction module;

[0046] generating a vibration waveform graph through a waveform generator based on the multi-scale rotation speed feature vector;

[0047] inputting the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal through a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix and a second vibration waveform feature matrix;

[0048] fusing the generated vibration feature matrix and the generated vibration waveform graph to obtain an optimized generated vibration feature matrix;

[0049] calculating a differential feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix;

[0050] calculating a transition matrix of the second vibration waveform feature matrix relative to the differential feature matrix as a classification feature matrix; and

[0051] inputting the classification feature matrix through a classifier to obtain a classification result, wherein the classification result is used to represent a probability value of the wind turbine to be detected having each fault category label.

[0052] In the above-mentioned wind turbine fault diagnosis method, after arranging the speed value of the generator at a plurality of predetermined time points in a predetermined time period into a speed input vector according to the time dimension, the multi-scale neighborhood feature extraction module is used to obtain a multi-scale speed feature vector, comprising: arranging the speed value of the generator at a plurality of predetermined time points in a predetermined time period into a speed input vector according to the time dimension; inputting the speed input vector into a first convolutional layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale speed correlation feature vector, wherein the first convolutional layer has a first one-dimensional convolution kernel with a first length; inputting the speed input vector into a second convolutional layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale speed correlation feature vector, wherein the second convolutional layer has a second one-dimensional convolution kernel with a second length, and the first length is different from the second length; and concatenating the first neighborhood scale speed correlation feature vector and the second neighborhood scale speed correlation feature vector to obtain the multi-scale speed feature vector.

[0053] In the above-mentioned wind turbine fault diagnosis method, inputting the speed input vector into the first convolutional layer of the multi-scale neighborhood feature extraction module to obtain the first neighborhood scale speed correlation feature vector comprises: using the first convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional coding on the speed input vector according to the following formula to obtain the first neighborhood scale speed correlation feature vector;

[0054] wherein the formula is:

[0055]

[0056] Wherein, a is the width of the first convolution kernel in the x direction, F is the first convolution kernel parameter vector, G is the local vector matrix for convolution kernel function operation, w is the size of the first convolution kernel, and X represents the tension input vector; inputting the rotation speed input vector into a second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale rotation speed correlation feature vector, including: using the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution coding on the rotation speed input vector according to the following formula to obtain the second neighborhood scale rotation speed correlation feature vector;

[0057] Wherein, the formula is:

[0058]

[0059] Wherein, b is the width of the second convolution kernel in the x direction, F is the second convolution kernel parameter vector, G is the local vector matrix for convolution kernel function operation, m is the size of the second convolution kernel, and X represents the tension input vector.

[0060] In the above fault diagnosis method of the wind turbine generator set, the waveform generator is a generative adversarial network.

[0061] In the above fault diagnosis method of the wind turbine generator set, the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal are fused through a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix and a second vibration waveform feature matrix, including: using each layer of the convolutional neural network model as the feature extractor to respectively perform the following on the input data in the forward transmission of the layer: performing convolution processing on the input data to obtain a convolution feature map; performing mean pooling based on the local channel dimension on the convolution feature map to obtain a pooling feature map; and performing nonlinear activation on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the convolutional neural network model as the feature extractor is the generated vibration feature matrix, the first vibration waveform feature matrix and the second vibration waveform feature matrix, and the input of the first layer of the convolutional neural network model as the feature extractor is the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal.

[0062] In the above fault diagnosis method of the wind turbine generator set, the generated vibration feature matrix and the generated vibration waveform graph are fused to obtain an optimized generated vibration feature matrix, including: fusing the generated vibration feature matrix and the generated vibration waveform graph according to the following formula to obtain the optimized generated vibration feature matrix;

[0063] Wherein, the formula is:

[0064]

[0065] wherein M1 represents the generated vibration waveform graph, M2 represents the generated vibration feature matrix, f i represents the eigenvalue of each position in the generated vibration feature matrix, represents the mean value of the value of each position of the generated vibration feature matrix, and N is the dimension of the generated vibration feature matrix, exp(·) represents the exponential operation of the matrix, the exponential operation of the matrix represents the natural exponential function value with the eigenvalue of each position in the matrix as the power, and a is a hyperparameter, represents the position-wise addition of the feature matrix, and represents the position-wise point multiplication.

[0066] In the fault diagnosis method of the wind turbine generator set, the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is calculated, including: calculating the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix according to the following formula;

[0067] wherein the formula is:

[0068]

[0069] wherein M P represents the optimized generated vibration feature matrix, M V represents the first vibration waveform feature matrix, M D represents the difference feature matrix, represents the position-wise difference.

[0070] In the fault diagnosis method of the wind turbine generator set, the transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix is calculated as a classification feature matrix, including: calculating the transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix as the classification feature matrix according to the following formula;

[0071] wherein the formula is:

[0072] M W = M*M D

[0073] wherein M W represents the second vibration waveform feature matrix, M D represents the difference feature matrix, and M represents the classification feature matrix.

[0074] In the fault diagnosis method of the wind turbine set, the classification feature matrix is input into a classifier to obtain a classification result, and the classification result is used to represent a probability value of the wind turbine set to be detected having a fault category label, and the method comprises: using the classifier to process the classification feature matrix according to the following formula to generate the classification result.

[0075] The formula is as follows:

[0076]

[0077] Wherein, O is an output result matrix, W i and b i are a weight and a bias matrix corresponding to the i th classification respectively, exp(·) represents an exponential operation of a matrix, and the exponential operation of the matrix represents a natural exponential function value with the feature value at each position in the matrix as the power.

[0078] Compared with the prior art, the fault diagnosis system and method of the wind turbine set provided in the application use a convolutional neural network model based on deep learning as a feature extractor to generate a vibration waveform based on the rotational speed implicit feature of a rotating shaft through a waveform generator, that is, a self-vibration signal caused by an excitation signal generated by the rotation of the rotating shaft, and use a difference matrix between the generated vibration waveform and a vibration signal detected at the head end of a generator to represent a vibration feature of the head end of the generator after the self-vibration signal is filtered out, and then represent the working feature of the wind turbine set based on the change feature between the vibration feature of the vibration signal detected at the head end of the generator and the vibration feature of the vibration signal detected at the tail end of the generator, so as to perform fault diagnosis of the wind turbine set. BRIEF DESCRIPTION OF DRAWINGS

[0079] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, taken in conjunction with the accompanying drawings. The drawings provided in the specification and the embodiments of the present application together serve to provide a further understanding that enables others skilled in the art to make or use the present application. The drawings provided are for illustrative purposes and are not intended to limit the present application. In the drawings, the same reference numerals generally indicate same components or steps throughout the several views.

[0080] Figure 1 An application scenario diagram of the fault diagnosis system of the wind turbine set according to the embodiments of the present application.

[0081] Figure 2 A block diagram of the fault diagnosis system of the wind turbine set according to the embodiments of the present application.

[0082] Figure 3 A block diagram of the multi-scale time series data encoding module in the fault diagnosis system of the wind turbine set according to the embodiments of the present application.

[0083] Figure 4 A flow chart of a fault diagnosis method for a wind turbine generator system according to an embodiment of the present application.

[0084] Figure 5 An architecture schematic diagram of a fault diagnosis method for a wind turbine generator system according to an embodiment of the present application. DETAILED DESCRIPTION

[0085] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part but not all of the embodiments of the present application, and the present application can be implemented in many different forms. Therefore, the attached drawings should not be used to limit and define the present application, and the present application should cover all changes falling within the scope of the appended claims and their equivalents.

[0086] Scenario overview

[0087] As mentioned before, in recent years, the energy industry transformation process in China is accelerating, and China has become the largest and fastest-growing market for wind power in the world. Wind energy, as a kind of renewable energy, has the characteristics of green and clean, and is the research focus of the world. Compared with land wind power, offshore wind resources are more abundant, have the advantages of large power generation, long power generation time, small impact on living environment, no occupation of arable land, and large-scale development, etc. Therefore, the development of offshore wind power technology has become a new trend in the development of wind power industry.

[0088] The wind turbine generator system is the core component of the offshore wind turbine, and if it fails, it will directly affect the power generation efficiency of the offshore wind turbine. Therefore, once a failure occurs, the wind turbine generator system needs to be repaired in time. When repairing, the fault mode of the wind turbine generator system needs to be determined first, so that the correct repair mode can be adopted. The existing fault mode identification of the wind turbine generator system is analyzed by experienced technicians. This fault diagnosis mode not only has low efficiency, but also is prone to missed detection and false detection. Therefore, a fault diagnosis scheme for the wind turbine generator system is expected.

[0089] Correspondingly, the present application inventors find that if the wind turbine of the offshore wind turbine is operating normally, the vibration signal generated thereby will be conducted in a certain specific form, so that the fault detection of the wind turbine can be performed based on the conduction of the vibration signal, and the present application inventors also consider that when the wind turbine is diagnosed based on the vibration signal, the rotation of the rotating shaft will generate an excitation signal, and if the vibration waveform generated by the excitation signal can be filtered, the fault detection accuracy can be improved. Therefore, in the technical solution of the present application, the waveform generator is used to obtain the generated vibration waveform based on the rotating speed implicit features of the rotating shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. Then, the vibration features at the head end of the generator after filtering out the self-vibration signal are represented by the difference matrix between the generated vibration waveform and the vibration signal detected at the head end of the generator. Then, the working features of the wind turbine are represented based on the change features between the vibration features detected at the head end of the generator and the vibration features of the vibration signal detected at the tail end of the generator, so as to perform fault diagnosis through the classifier.

[0090] Specifically, in the technical solution of the present application, first, the rotating speed values of the generator at a plurality of predetermined time points within a predetermined time period are obtained by a rotating speed sensor, and the first vibration signal and the second vibration signal are collected by the first vibration sensor and the second vibration sensor arranged at the head end and the tail end of the generator.

[0091] Then, for the rotating speed values of the generator at the plurality of predetermined time points, the rotating speed input vector is arranged in the time dimension to facilitate subsequent feature mining. It should be understood that the convolutional neural network is originally a model applied in the image field, but the idea of local feature extraction can also be applied to time series data analysis. For example, a time series convolution structure with a convolution kernel size of 3, for time series data input, the convolution kernel moves along the time dimension in the form of a sliding window, and outputs the weighted sum of the data in each time series segment. Each convolution unit stacks multiple convolution kernels to output multi-dimensional features. Large convolution kernels extract features from large-scale time series neighborhoods, where each value in the neighborhood has a smaller impact, thereby reducing the fluctuations in the input data and reducing the impact of noise points on the output features. However, large-scale convolution kernels reduce the difference in numerical changes, which can cause over-smoothing and loss of discriminative ability of the output features. On the contrary, small-scale convolution kernels can better preserve the information in the input data, but are more susceptible to noise points. Therefore, considering the characteristics of convolution kernels of different scales, different sizes of convolution units are combined to extract features of different time scales. Then, feature fusion is completed by feature concatenation, thereby obtaining multi-scale neighborhood features.

[0092] That is, specifically, in the technical solution of the present application, further, the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales is used to perform one-dimensional convolution coding on the rotation speed input vector, respectively, and then the feature vectors corresponding to the two one-dimensional convolution kernels of different scales are concatenated to obtain the multi-scale rotation speed input vector. In particular, in this way, the multi-scale neighborhood correlation of the dynamic change features of the rotation speed value in the time sequence dimension is extracted, so that the output features contain both smoothed features and original input features, avoiding information loss and improving the accuracy of subsequent classification. In other examples of the present application, the multi-scale neighborhood feature extraction module can also include a larger number of one-dimensional convolution layers using one-dimensional convolution kernels of different lengths for neighborhood correlation feature extraction of different scales, which is not limited by the present application.

[0093] Further, the multi-scale rotation speed feature vector is passed through a waveform generator to generate a vibration waveform based on the rotation speed implicit features of the rotating shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. In particular, in one specific example of the present application, here, the waveform generator is a generative adversarial network, which includes a generator model and a discriminator neural network model. The generator model of the waveform generator can use the dynamic features of the rotation speed value of the rotating shaft in the time sequence dimension to generate a generated vibration waveform representing the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. The discriminator neural network model of the waveform generator can make the feature information extracted from the generated vibration waveform and the reference waveform image similar by "cheating" the discriminator, so that the generated vibration waveform can more accurately represent the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft.

[0094] In this way, the generated vibration waveform, the waveform of the first vibration signal, and the waveform of the second vibration signal are subjected to feature extraction in a convolutional neural network model as a feature extractor to extract local high-dimensional implicit feature distribution information of the generated vibration waveform, the waveform of the first vibration signal, and the waveform of the second vibration signal, respectively, thereby obtaining a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix.

[0095] It can be understood that, here, the multi-scale feature vector generates a generated vibration waveform graph through the waveform generator as a pseudo-image generated by the convolutional neural network, which has a shallow feature distribution similar to that obtained based on feature extraction of the convolutional neural network, and thus, after passing through the convolutional neural network model as a feature extractor again, if the shallow feature distribution and the deep feature distribution can be better fused for the generated vibration feature matrix, the expression capability of the generated vibration feature matrix can be improved.

[0096] Therefore, in the technical solution of the present application, the generated vibration waveform graph, denoted as M1, and the generated vibration feature matrix, denoted as M2, are further subjected to attention-guided hierarchical deep simultaneous fusion, denoted as:

[0097]

[0098] wherein M1 represents the generated vibration waveform graph, M2 represents the generated vibration feature matrix, f i represents the feature value of each position in the generated vibration feature matrix, represents the mean value of the values of each position of the generated vibration feature matrix, N is the dimension of the generated vibration feature matrix, exp(·) represents the exponential operation of the matrix, the exponential operation of the matrix represents the natural exponential function value with the feature value of each position in the matrix as the power, and a is a hyperparameter, represents the position-wise addition of the feature matrix, and represents the position-wise point multiplication.

[0099] Here, the attention-guided hierarchical deep simultaneous fusion takes the sub-dimension consistency of the deep feature M2 as the attention-guided weight, applies the sub-dimension distribution consistency attention mechanism to the shallow feature M1, and thus performs volume matching between manifolds with different depths. In this way, the simultaneous distribution of the corrected generated vibration feature matrix on each sub-dimension of the matrix before fusion will have high consistency, thereby improving its expression capability and further improving the classification accuracy.

[0100] Further, a difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is calculated to represent the vibration feature information of the first end of the generator after filtering out the self-vibration signal.

[0101] It should be understood that since the vibration signal is transmitted from the generator's head end to its tail end in a specific form, a transfer matrix is ​​used in this application to characterize this transmission characteristic. That is, the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix is ​​calculated as a classification feature matrix. This matrix characterizes the operating characteristics of the wind turbine generator set based on the change between the vibration characteristics at the generator's head end and the vibration characteristics at the generator's tail end. A classifier is then used for fault diagnosis to obtain classification results representing the probability values ​​of each fault category label for the wind turbine generator set under test.

[0102] Based on this, this application proposes a fault diagnosis system for wind turbine generator sets, comprising: a speed data acquisition module for acquiring the speed values ​​of the generator at multiple predetermined time points within a predetermined time period; a vibration data acquisition module for acquiring a first vibration signal and a second vibration signal acquired by a first vibration sensor and a second vibration sensor deployed at the front and rear ends of the generator; a multi-scale time-series data encoding module for arranging the speed values ​​of the generator at multiple predetermined time points within the predetermined time period into a speed input vector according to the time dimension, and then obtaining a multi-scale speed feature vector through a multi-scale neighborhood feature extraction module; a generation module for generating a vibration waveform diagram by passing the multi-scale speed feature vector through a waveform generator; and a feature extraction module for extracting the waveform diagram and the waveform of the first vibration signal. The waveform diagram and the waveform diagram of the second vibration signal are processed by a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix; a fusion optimization module is used to fuse the generated vibration feature matrix and the generated vibration waveform diagram to obtain an optimized generated vibration feature matrix; a filtering module is used to calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix; a transfer module is used to calculate the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as a classification feature matrix; and a fault diagnosis result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, the classification result being used to represent the probability value of the wind turbine generator set under test having each fault category label.

[0103] Figure 1 The illustration shows an application scenario of a fault diagnosis system for a wind turbine generator set according to an embodiment of this application. For example... Figure 1 As shown, in this application scenario, firstly, through a speed sensor (e.g., such as...) Figure 1 The P shown in the figure obtains generators at multiple predetermined time points within a predetermined time period (e.g., such as...). Figure 1 The rotational speed value of E shown, and through the first vibration sensors deployed at the beginning and end of the generator (e.g., such as...).Figure 1 a first vibration sensor (e.g., as shown in T1) and a second vibration sensor (e.g., as shown in T2) are arranged to collect the first vibration signal and the second vibration signal at the plurality of predetermined time points. Then, the acquired rotational speed values of the generator at the plurality of predetermined time points and the first vibration signal and the second vibration signal are input into a server (e.g., as shown in server S) deploying a fault diagnosis algorithm of the wind turbine generator system, wherein the server is capable of processing the rotational speed values of the generator at the plurality of predetermined time points and the first vibration signal and the second vibration signal with the fault diagnosis algorithm of the wind turbine generator system to generate a classification result representing probability values of the wind turbine generator system to be detected with each fault category label. Figure 1 Figure 1

[0104] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be specifically introduced below with reference to the accompanying drawings.

[0105] Exemplary system

[0106] Figure 2 Fig. 1 illustrates a block diagram of a fault diagnosis system of a wind turbine generator system according to an embodiment of the present application. As shown in Fig. 1, the fault diagnosis system of the wind turbine generator system according to the embodiment of the present application comprises a first vibration sensor (e.g., as shown in T1) and a second vibration sensor (e.g., as shown in T2) arranged to collect the first vibration signal and the second vibration signal at the plurality of predetermined time points. Then, the acquired rotational speed values of the generator at the plurality of predetermined time points and the first vibration signal and the second vibration signal are input into a server (e.g., as shown in server S) deploying a fault diagnosis algorithm of the wind turbine generator system, wherein the server is capable of processing the rotational speed values of the generator at the plurality of predetermined time points and the first vibration signal and the second vibration signal with the fault diagnosis algorithm of the wind turbine generator system to generate a classification result representing probability values of the wind turbine generator system to be detected with each fault category label. Figure 2 ​​As shown, the fault diagnosis system 200 of the wind turbine generator set according to the embodiment of the present application comprises: a rotating speed data acquisition module 210 configured to acquire rotating speed values of the generator at a plurality of predetermined time points within a predetermined time period; a vibration data acquisition module 220 configured to acquire first and second vibration signals collected by first and second vibration sensors arranged at the head end and tail end of the generator; a multi-scale time series data encoding module 230 configured to arrange the rotating speed values of the generator at a plurality of predetermined time points within the predetermined time period into a rotating speed input vector according to the time dimension, and then pass the rotating speed input vector through a multi-scale neighborhood feature extraction module to obtain a multi-scale rotating speed feature vector; a generation module 240 configured to pass the multi-scale rotating speed feature vector through a waveform generator to obtain a generated vibration waveform graph; a feature extraction module 250 configured to pass the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal through a convolutional neural network model serving as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix and a second vibration waveform feature matrix; a fusion optimization module 260 configured to fuse the generated vibration feature matrix and the generated vibration waveform graph to obtain an optimized generated vibration feature matrix; a filtering module 270 configured to calculate a differential feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix; a transfer module 280 configured to calculate a transfer matrix of the second vibration waveform feature matrix relative to the differential feature matrix as a classification feature matrix; and a fault diagnosis result generation module 290 configured to pass the classification feature matrix through a classifier to obtain a classification result, wherein the classification result is used to represent a probability value of the wind turbine generator set to be detected having each fault category label.

[0107] Specifically, in the embodiments of the present application, the rotating speed data acquisition module 210 and the vibration data acquisition module 220 are configured to acquire rotating speed values of the generator at a plurality of predetermined time points within a predetermined time period, and acquire first vibration signals and second vibration signals collected by the first vibration sensor and the second vibration sensor arranged at the head end and the tail end of the generator. As described above, since the vibration signals generated by the offshore wind turbine generator set when it is operating normally are conducted in a certain specific form, the fault detection of the wind turbine generator set can be performed based on the conduction of the vibration signals, and it is also considered that the rotating shaft rotation generates excitation signals when the wind turbine generator set is diagnosed based on the vibration signals, and the vibration waveform generated by the excitation signals can be filtered to improve the fault detection accuracy. Therefore, in the technical solution of the present application, the waveform generator is used to obtain the generated vibration waveform based on the rotating speed implicit characteristics of the rotating shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotating shaft. Then, the vibration characteristics of the head end of the generator after filtering out the self-vibration signal are represented by the difference matrix between the generated vibration waveform and the vibration signal detected at the head end of the generator. Then, the working characteristics of the wind turbine generator set are represented based on the change characteristics between the vibration characteristics of the head end of the generator and the vibration characteristics of the vibration signal detected at the tail end of the generator, so as to perform fault diagnosis through the classifier.

[0108] That is, specifically, in the technical solution of the present application, first, the rotating speed values of the generator at a plurality of predetermined time points within a predetermined time period are acquired by the rotating speed sensor, and the first vibration signals and the second vibration signals are collected by the first vibration sensor and the second vibration sensor arranged at the head end and the tail end of the generator.

[0109] Specifically, in the embodiment of the present application, the multi-scale time series data encoding module 230 is configured to arrange the speed values of the generator at the plurality of predetermined time points in the predetermined time period into a speed input vector in the time dimension, and then extract a multi-scale speed feature vector by using a multi-scale neighborhood feature extraction module. That is, for the speed values of the generator at the plurality of predetermined time points, the speed values are arranged into a speed input vector in the time dimension to facilitate subsequent feature mining. It should be understood that the convolutional neural network is originally a model applied in the field of images, but the idea of local feature extraction can also be applied to time series data analysis. For example, a time series convolution structure with a convolution kernel size of 3, for time series data input, the convolution kernel moves along the time dimension in the form of a sliding window, and outputs the weighted sum of the data in each time series segment. Each convolution unit stacks multiple convolution kernels to output multi-dimensional features. A large convolution kernel extracts features from a large-scale time series neighborhood, where each value in the neighborhood has a smaller impact, thereby reducing the fluctuations in the input data and reducing the impact of noise points on the output features. However, a large-scale convolution kernel reduces the difference in value changes, which can easily lead to over-smoothing, causing the output features to lose discriminability. On the contrary, a small-scale convolution kernel can better preserve the information in the input data, but it is also more susceptible to noise points. Therefore, considering the characteristics of convolution of different scales, different sizes of convolution units are combined to extract features of different time scales. Then, feature fusion is completed by using feature concatenation, thereby obtaining multi-scale neighborhood features.

[0110] That is, specifically, in the technical solution of the present application, further, the multi-scale neighborhood feature extraction module includes convolution layers with one-dimensional convolution kernels of different scales, which are used to perform one-dimensional convolution encoding on the speed input vector, respectively, and then the feature vectors corresponding to the two one-dimensional convolution kernels of different scales are concatenated to obtain the multi-scale speed input vector. In particular, in this way, the multi-scale neighborhood correlation of the dynamic change features of the speed values in the time dimension is extracted, so that the output features contain both smoothed features and original input features, avoiding information loss, and thus improving the accuracy of subsequent classification. In other examples of the present application, the multi-scale neighborhood feature extraction module can also include a larger number of one-dimensional convolution layers, which use one-dimensional convolution kernels of different lengths to extract features of neighborhood correlation in different scales. This is not limited by the present application.

[0111] More specifically, in the embodiments of the present application, the multi-scale time series data encoding module comprises: first, arranging the speed value of the generator at a plurality of predetermined time points in a predetermined time period into a speed input vector according to the time dimension. Then, inputting the speed input vector into the first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale speed correlation feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel with a first length. Accordingly, in one specific example, the first convolution layer of the multi-scale neighborhood feature extraction module is used to one-dimensionally convolve encode the speed input vector according to the following formula to obtain the first neighborhood scale speed correlation feature vector;

[0112] The formula is:

[0113]

[0114] Wherein a is the width of the first convolution kernel in the x direction, F is the first convolution kernel parameter vector, G is the local vector matrix operated with the convolution kernel function, w is the size of the first convolution kernel, and X represents the tension input vector

[0115] Then, inputting the speed input vector into the second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale speed correlation feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel with a second length, and the first length is different from the second length. Accordingly, in one specific example, the second convolution layer of the multi-scale neighborhood feature extraction module is used to one-dimensionally convolve encode the speed input vector according to the following formula to obtain the second neighborhood scale speed correlation feature vector;

[0116] The formula is:

[0117]

[0118] Wherein b is the width of the second convolution kernel in the x direction, F is the second convolution kernel parameter vector, G is the local vector matrix operated with the convolution kernel function, m is the size of the second convolution kernel, and X represents the tension input vector.

[0119] Finally, concatenating the first neighborhood scale speed correlation feature vector and the second neighborhood scale speed correlation feature vector to obtain the multi-scale speed feature vector.

[0120] Figure 3 Fig. 1 illustrates a block diagram of a multi-scale time series data encoding module in a fault diagnosis system of a wind turbine generator set according to an embodiment of the present application. As shown in the figure, Figure 3As shown, the multi-scale time-series data encoding module 230 comprises: a vector construction unit 231 configured to arrange the speed values of the generator at a plurality of predetermined time points within the predetermined time period in a time dimension as a speed input vector; a first convolution unit 232 configured to input the speed input vector into a first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale speed correlation feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel of a first length; a second convolution unit 233 configured to input the speed input vector into a second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale speed correlation feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and a concatenation unit 234 configured to concatenate the first neighborhood scale speed correlation feature vector and the second neighborhood scale speed correlation feature vector to obtain the multi-scale speed feature vector.

[0121] Specifically, in the embodiments of the present application, the generation module 240 is configured to input the multi-scale speed feature vector into a waveform generator to obtain a generated vibration waveform image. That is, in the technical solution of the present application, the multi-scale speed feature vector is further input into the waveform generator to obtain a generated vibration waveform based on the speed implicit features of the rotating shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. In particular, in one specific example, the waveform generator is a generative adversarial network, which comprises a generator model and a discriminator neural network model. The generator model of the waveform generator can utilize the dynamic features of the speed values of the rotating shaft in the time dimension to generate a generated vibration waveform representing the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. The discriminator neural network model of the waveform generator can make the feature information extracted from the generated vibration waveform and the reference waveform image similar by "cheating" the discriminator, so that the generated vibration waveform can more accurately represent the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft.

[0122] Specifically, in the embodiment of the present application, the feature extraction module 250 is configured to input the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal into the convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix and a second vibration waveform feature matrix. That is, in the technical solution of the present application, the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal are further input into the convolutional neural network model as a feature extractor for feature extraction, so as to extract the local high-dimensional implicit feature distribution information of the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal respectively, thereby obtaining the generated vibration feature matrix, the first vibration waveform feature matrix and the second vibration waveform feature matrix.

[0123] More specifically, in the embodiment of the present application, the feature extraction module is further configured to use each layer of the convolutional neural network model as a feature extractor to perform the following operations on the input data in the forward transmission of the layer: performing convolution processing on the input data to obtain a convolution feature map; performing mean pooling based on the local channel dimension on the convolution feature map to obtain a pooling feature map; and performing nonlinear activation on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the convolutional neural network model as a feature extractor is the generated vibration feature matrix, the first vibration waveform feature matrix and the second vibration waveform feature matrix, and the input of the first layer of the convolutional neural network model as a feature extractor is the generated vibration waveform graph, the waveform graph of the first vibration signal and the waveform graph of the second vibration signal.

[0124] Specifically, in the embodiment of the present application, the fusion optimization module 260 is configured to fuse the generated vibration feature matrix and the generated vibration waveform graph to obtain an optimized generated vibration feature matrix. It should be understood that here, the generated vibration waveform graph of the multi-scale feature vector obtained by the waveform generator is a pseudo image generated by the convolutional neural network, which has a shallow feature distribution similar to that obtained by the convolutional neural network based on feature extraction. Therefore, after being input into the convolutional neural network model as a feature extractor again, for the generated vibration feature matrix, if the shallow feature distribution and the deep feature distribution can be better fused, the expression ability of the generated vibration feature matrix can be improved.

[0125] Therefore, in the technical solution of the present application, the generated vibration waveform graph, denoted as M1, and the generated vibration feature matrix, denoted as M2, are further subjected to attention-guided hierarchical deep simultaneous fusion. It should be understood that, here, the attention-guided hierarchical deep simultaneous fusion takes the sub-dimension consistency of the deep feature M2 as the attention-guided weight, and applies a consistent attention mechanism of sub-dimension distribution to the shallow feature M1, so as to perform volume matching between manifolds with depth differences. In this way, the simultaneous distribution of the corrected generated vibration feature matrix on each sub-dimension of the matrix before fusion will have high consistency, thereby improving the expression ability and further improving the accuracy of classification.

[0126] More specifically, in the embodiment of the present application, the fusion optimization module is further configured to fuse the generated vibration feature matrix and the generated vibration waveform graph to obtain the optimized generated vibration feature matrix according to the following formula:

[0127] The formula is as follows:

[0128]

[0129] wherein M1 represents the generated vibration waveform graph, M2 represents the generated vibration feature matrix, f i represents the feature value of each position in the generated vibration feature matrix, represents the mean value of the values of each position in the generated vibration feature matrix, N is the dimension of the generated vibration feature matrix, exp(·) represents the exponential operation of the matrix, the exponential operation of the matrix means the natural exponential function value with the feature value of each position in the matrix as the power, and a is a hyperparameter, represents the position-wise addition of the feature matrix, and represents the position-wise point multiplication.

[0130] Specifically, in the embodiment of the present application, the filtering module 270 is configured to calculate a difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix. That is, in the technical solution of the present application, further, a difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is calculated to represent the vibration feature information of the first end of the generator after filtering out the self-vibration signal.

[0131] Correspondingly, in one specific example, the filtering module is further configured to calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix according to the following formula:

[0132] The formula is as follows:

[0133]

[0134] wherein M P represents that the optimization generates a vibration feature matrix, M V represents the first vibration waveform feature matrix, M D represents the difference feature matrix, represents a position difference.

[0135] Specifically, in the embodiments of the present application, the transfer module 280 and the fault diagnosis result generation module 290 are configured to calculate a transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as a classification feature matrix, and pass the classification feature matrix through a classifier to obtain a classification result, which is used to represent a probability value of the wind turbine to have each fault category label. It can be understood that, since the vibration signal is conducted in a certain specific form from the head end of the generator to the tail end of the generator, in the technical solution of the present application, a transfer matrix is used to represent the conduction feature. That is, the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix is calculated as the classification feature matrix, so as to represent the working feature of the wind turbine based on the change feature between the vibration feature of the head end of the generator and the vibration feature of the tail end of the generator, and then the classifier is used to perform fault diagnosis, so as to obtain the classification result used to represent the probability value of the wind turbine to have each fault category label.

[0136] More specifically, in the embodiments of the present application, the transfer module is further configured to calculate the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as the classification feature matrix according to the following formula:

[0137] wherein the formula is:

[0138] M W = M*M D

[0139] wherein M W represents the second vibration waveform feature matrix, M D represents the difference feature matrix, and M represents the classification feature matrix.

[0140] More specifically, in the embodiments of the present application, the fault diagnosis result generation module is further configured to use the classifier to process the classification feature matrix according to the following formula to generate the classification result:

[0141] wherein the formula is:

[0142]

[0143] where O is the output result matrix, W i and b i are the weight and bias matrix corresponding to the i-th classification respectively, exp(·) represents the exponential operation of the matrix, and the exponential operation of the matrix represents the natural exponential function value with the feature value of each position in the matrix as the power.

[0144] In summary, the fault diagnosis system 200 of the wind turbine generator set based on the embodiments of the present application is illustrated, which takes the convolutional neural network model based on deep learning as a feature extractor to obtain the generated vibration waveform based on the rotating speed implicit feature of the rotating shaft through the waveform generator, i.e. the self-vibration signal caused by the excitation signal generated due to the rotation of the rotating shaft, and takes the difference matrix between the generated vibration waveform and the vibration signal detected at the head of the generator to represent the vibration feature at the head of the generator after filtering out the self-vibration signal, and then represents the working feature of the wind turbine generator set based on the change feature between the vibration feature of the vibration signal detected at the head of the generator and the vibration feature of the vibration signal detected at the tail of the generator, so as to perform the fault diagnosis of the wind turbine generator set.

[0145] As described above, the fault diagnosis system 200 of the wind turbine generator set according to the embodiments of the present application can be implemented in various terminal devices, such as a server of the fault diagnosis algorithm of the wind turbine generator set, etc. In one example, the fault diagnosis system 200 of the wind turbine generator set according to the embodiments of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the fault diagnosis system 200 of the wind turbine generator set can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the fault diagnosis system 200 of the wind turbine generator set can also be one of the many hardware modules of the terminal device.

[0146] Alternatively, in another example, the fault diagnosis system 200 of the wind turbine generator set and the terminal device can also be separate devices, and the fault diagnosis system 200 of the wind turbine generator set can be connected to the terminal device through a wired and / or wireless network, and transmit interactive information in a conventional data format.

[0147] Exemplary method

[0148] Figure 4 A flowchart of a fault diagnosis method of a wind turbine generator set is illustrated. As Figure 4As shown, the fault diagnosis method for a wind turbine generator set according to an embodiment of this application includes the following steps: S110, acquiring the generator speed values ​​at multiple predetermined time points within a predetermined time period; S120, acquiring a first vibration signal and a second vibration signal collected by a first vibration sensor and a second vibration sensor deployed at the head and tail ends of the generator; S130, arranging the generator speed values ​​at multiple predetermined time points within the predetermined time period into a speed input vector according to the time dimension, and then obtaining a multi-scale speed feature vector through a multi-scale neighborhood feature extraction module; S140, passing the multi-scale speed feature vector through a waveform generator to obtain a generated vibration waveform diagram; S150, combining the generated vibration waveform diagram, the waveform diagram of the first vibration signal, and the... The waveform of the second vibration signal is processed by a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix; S160, the generated vibration feature matrix and the generated vibration waveform are fused to obtain an optimized generated vibration feature matrix; S170, the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is ​​calculated; S180, the transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix is ​​calculated as a classification feature matrix; and S190, the classification feature matrix is ​​processed by a classifier to obtain a classification result, the classification result being used to represent the probability value of the wind turbine generator set to be detected having each fault category label.

[0149] Figure 5 The illustration shows a schematic diagram of the architecture of a fault diagnosis method for a wind turbine generator set according to an embodiment of this application. Figure 5 As shown, in the network architecture of the fault diagnosis method for the wind turbine generator set, firstly, the obtained generator speed values ​​at multiple predetermined time points within the predetermined time period (e.g., as shown) are... Figure 5 The P1 shown is arranged in the time dimension as a rotational speed input vector (e.g., as shown in the figure). Figure 5 The V1 shown in the diagram is then processed by a multi-scale neighborhood feature extraction module (e.g., such as...). Figure 5 The MS shown is used to obtain a multi-scale rotational speed feature vector (e.g., as shown in the figure). Figure 5 The VF1 shown in the diagram is then passed through a waveform generator (e.g., as shown in the diagram). Figure 5 The WG shown is used to obtain the generated vibration waveform (e.g., as shown in the figure). Figure 5 The Q shown in the diagram); then, the generated vibration waveform diagram, the waveform diagram of the first vibration signal (e.g., as shown in the diagram); Figure 5 The waveforms of Q1 (as shown) and the second vibration signal (e.g., as shown) Figure 5 The Q2 shown is obtained through a convolutional neural network model (e.g., as shown) acting as a feature extractor.Figure 5 a generated vibration feature matrix (e.g., as MF1 shown in Figure 5 a first vibration waveform feature matrix (e.g., as MF2 shown in Figure 5 a second vibration waveform feature matrix (e.g., as MF3 shown in Figure 5 the generated vibration feature matrix and the generated vibration waveform matrix to obtain an optimized generated vibration feature matrix (e.g., as MF4 shown in Figure 5 the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix (e.g., as MF5 shown in Figure 5 a transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix as a classification feature matrix (e.g., as MF6 shown in ​ the classification feature matrix through a classifier (e.g., as the classifier shown in ​ the classifier to obtain a classification result, which is used to represent a probability value of the wind turbine to have a respective fault category label.

[0150] More specifically, in steps S110 and S120, the rotational speed values of the generator at a plurality of predetermined time points within a predetermined time period are obtained, and the first vibration signal and the second vibration signal collected by the first vibration sensor and the second vibration sensor deployed at the head end and the tail end of the generator are obtained. It should be understood that, since the vibration signal generated by the wind turbine of the offshore wind turbine will be conducted in a certain specific form if the wind turbine operates normally, the fault detection of the wind turbine can be performed based on the conduction of the vibration signal, and it is also considered that the rotational shaft rotation will generate an excitation signal when the wind turbine is diagnosed based on the vibration signal, and the vibration waveform generated by the excitation signal can be filtered to improve the fault detection accuracy. Therefore, in the technical solution of the present application, the vibration waveform generated by the rotational speed implicit feature of the rotational shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotational shaft rotation, is obtained by the waveform generator. Then, the vibration feature at the head end of the generator after filtering the self-vibration signal is represented by the difference matrix between the generated vibration waveform and the vibration signal detected at the head end of the generator. Then, the working feature of the wind turbine is represented based on the change feature between the vibration feature detected at the head end of the generator and the vibration feature of the vibration signal detected at the tail end of the generator, so as to perform fault diagnosis through the classifier.

[0151] That is, specifically, in the technical solution of the present application, first, the rotational speed values of the generator at a plurality of predetermined time points within a predetermined time period are obtained by a rotational speed sensor, and first and second vibration signals are collected by first and second vibration sensors arranged at the leading and trailing ends of the generator.

[0152] More specifically, in step S130, after arranging the rotational speed values of the generator at a plurality of predetermined time points within a predetermined time period into a rotational speed input vector according to the time dimension, a multi-scale neighborhood feature extraction module is used to obtain a multi-scale rotational speed feature vector. That is, for the rotational speed values of the generator at a plurality of predetermined time points, they are arranged into a rotational speed input vector according to the time dimension to facilitate subsequent feature mining. It should be understood that the convolutional neural network is originally a model applied in the image field, but its local feature extraction idea can also be applied to time series data analysis. For example, a time series convolution structure with a convolution kernel size of 3, for time series data input, the convolution kernel moves along the time dimension in the form of a sliding window, and outputs the weighted sum of the data in each time series segment. Each convolution unit stacks multiple convolution kernels to output multi-dimensional features. A large convolution kernel extracts features from a large-scale time series neighborhood, where each value in the neighborhood has a smaller impact, thereby reducing the fluctuations in the input data and reducing the impact of noise points on the output features. However, a large-scale convolution kernel reduces the difference in numerical changes, which can cause over-smoothing and make the output features lose discriminability. On the contrary, a small-scale convolution kernel can better preserve the information in the input data, but it is also more susceptible to noise points. Therefore, considering the characteristics of convolution of different scales, different sizes of convolution units are combined to extract features of different time scales. Then, feature fusion is completed by feature concatenation, thereby obtaining multi-scale neighborhood features.

[0153] That is, specifically, in the technical solution of the present application, further, the rotational speed input vector is one-dimensionally convoluted by convolution layers of the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales, and then the feature vectors corresponding to the two one-dimensional convolution kernels of different scales are concatenated to obtain the multi-scale rotational speed input vector. In particular, in this way, the multi-scale neighborhood correlation of the dynamic change features of the rotational speed values in the time series dimension is extracted, so that the output features contain both smoothed features and original input features, avoiding information loss and improving the accuracy of subsequent classification. In other examples of the present application, the multi-scale neighborhood feature extraction module can also include a larger number of one-dimensional convolution layers that use one-dimensional convolution kernels of different lengths to extract features of different scales in the neighborhood, which is not limited by the present application.

[0154] More specifically, in step S140, the multi-scale rotating speed feature vector is passed through a waveform generator to obtain a generated vibration waveform image. That is, in the technical solution of the present application, further, the multi-scale rotating speed feature vector is passed through the waveform generator to obtain a generated vibration waveform based on the rotating speed implicit feature of the rotating shaft, that is, the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. In particular, in one specific example, here, the waveform generator is a generative adversarial network, which includes a generator model and a discriminator neural network model. The generator model of the waveform generator can utilize the dynamic features of the rotating speed value of the rotating shaft in the time sequence dimension to generate a generated vibration waveform representing the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft. The discriminator neural network model of the waveform generator can make the feature information extracted from the generated vibration waveform and the reference waveform image similar by "cheating" the discriminator, so that the generated vibration waveform can more accurately represent the self-vibration signal caused by the excitation signal generated by the rotation of the rotating shaft.

[0155] More specifically, in steps S150 and S160, the generated vibration waveform image, the waveform image of the first vibration signal, and the waveform image of the second vibration signal are passed through a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix, and the generated vibration feature matrix and the generated vibration waveform image are fused to obtain an optimized generated vibration feature matrix. That is, in the technical solution of the present application, the generated vibration waveform image, the waveform image of the first vibration signal, and the waveform image of the second vibration signal are further extracted by the convolutional neural network model as a feature extractor to extract the local high-dimensional implicit feature distribution information of the generated vibration waveform image, the waveform image of the first vibration signal, and the waveform image of the second vibration signal, respectively, thereby obtaining a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix.

[0156] Then, it can be understood that here, the generated vibration waveform image obtained by the multi-scale feature vector passing through the waveform generator is a pseudo image generated by the convolutional neural network, which has a shallow feature distribution similar to that obtained based on feature extraction. Therefore, after passing through the convolutional neural network model as a feature extractor again, for the generated vibration feature matrix, if the shallow feature distribution and the deep feature distribution can be better fused, the expression ability of the generated vibration feature matrix can be improved.

[0157] Therefore, in the technical solution of this application, attention-oriented hierarchical deep joint fusion is further performed on the generated vibration waveform, for example, denoted as M1, and the generated vibration feature matrix, for example, denoted as M2. It should be understood that here, the attention-oriented hierarchical deep joint fusion uses the sub-dimensional consistency of the deep feature M2 as the attention-oriented weight, and applies a consistency attention mechanism of sub-dimensional distribution to the shallow feature M1, thereby performing volume matching between manifolds with depth differences. In this way, the joint distribution of the modified generated vibration feature matrix on each sub-dimension of the matrix before fusion will have high consistency, thereby improving its expressive power and thus improving the accuracy of classification.

[0158] More specifically, in step S170, the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is ​​calculated. That is, in the technical solution of this application, the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix is ​​further calculated to characterize the vibration feature information of the generator head end after the natural vibration signal has been filtered out.

[0159] More specifically, in steps S180 and S190, the transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix is ​​calculated as a classification feature matrix, and the classification feature matrix is ​​passed through a classifier to obtain a classification result. The classification result is used to represent the probability value of the wind turbine generator set under test having each fault category label. It should be understood that since the vibration signal is transmitted from the head end of the generator to the tail end in a specific form, the technical solution of this application uses a transition matrix to characterize this transmission characteristic. That is, the transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix is ​​calculated as a classification feature matrix to characterize the operating characteristics of the wind turbine generator set based on the change characteristics between the vibration characteristics at the head end and the vibration characteristics at the tail end of the generator. This is then used by a classifier for fault diagnosis, thereby obtaining a classification result representing the probability value of the wind turbine generator set under test having each fault category label.

[0160] In summary, the wind turbine fault diagnosis method based on the embodiments of the present application is illustrated, which takes a convolutional neural network model based on deep learning as a feature extractor to generate a vibration waveform based on the rotating speed implicit features of the rotating shaft through a waveform generator, i.e., the self-vibration signal caused by the excitation signal generated due to the rotation of the rotating shaft, and takes the difference matrix between the generated vibration waveform and the vibration signal detected at the head end of the generator to represent the vibration features at the head end of the generator after filtering out the self-vibration signal, and then represents the working features of the wind turbine based on the change features between the vibration features of the vibration signal detected at the head end of the generator and the vibration signal detected at the tail end of the generator, so as to perform the fault diagnosis of the wind turbine.

[0161] The above describes the basic principles of the present application in combination with specific embodiments, but it should be noted that the advantages, advantages, effects and the like mentioned in the present application are only examples and not limitations, and these advantages, advantages, effects and the like cannot be considered as the must-have of each embodiment of the present application. In addition, the above specific details are only for the purpose of example and for the purpose of understanding, and are not limited to the above specific details. The above specific details do not limit the present application to be necessarily implemented by the above specific details.

[0162] The block diagrams of the devices, apparatuses, equipment, systems involved in the present application are only illustrative examples and are not intended to require or imply the connection, arrangement, configuration shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner. Words such as "include", "contain", "have" and the like are open-ended words, mean "including but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0163] It should also be noted that in the devices, equipment and methods of the present application, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present application.

[0164] The above description of the disclosed aspects is provided so that any person skilled in the art can make or use the present application. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0165] The foregoing description has been presented for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the application to the forms disclosed herein. Although various example aspects and embodiments have been discussed above, those of ordinary skill in the art will appreciate a variety of modifications, alternatives, permutations, additions, and sub-combinations of the described aspects and embodiments.

Claims

1. A fault diagnosis system for wind turbine generator sets, characterized in that, include: The speed data acquisition module is used to acquire the generator speed values ​​at multiple predetermined time points within a predetermined time period; The vibration data acquisition module is used to acquire the first vibration signal and the second vibration signal collected by the first vibration sensor and the second vibration sensor deployed at the head and tail ends of the generator. The multi-scale time series data encoding module is used to arrange the generator speed values ​​at multiple predetermined time points within the predetermined time period into a speed input vector according to the time dimension, and then obtain a multi-scale speed feature vector through the multi-scale neighborhood feature extraction module. The generation module is used to generate a vibration waveform by passing the multi-scale rotational speed feature vector through a waveform generator. The feature extraction module is used to process the generated vibration waveform, the waveform of the first vibration signal, and the waveform of the second vibration signal through a convolutional neural network model as a feature extractor to obtain a generated vibration feature matrix, a first vibration waveform feature matrix, and a second vibration waveform feature matrix. The fusion optimization module is used to fuse the generated vibration feature matrix and the generated vibration waveform to obtain an optimized generated vibration feature matrix. The filtering module is used to calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix; The transfer module is used to calculate the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as the classification feature matrix; as well as The fault diagnosis result generation module is used to pass the classification feature matrix through a classifier to obtain classification results, which are used to represent the probability values ​​of the wind turbine generator set under test having each fault category label.

2. The fault diagnosis system for wind turbine generator sets according to claim 1, characterized in that, The multi-scale time-series data encoding module includes: A vector construction unit is used to arrange the generator speed values ​​at multiple predetermined time points within the predetermined time period into a speed input vector according to the time dimension. The first convolutional unit is used to input the rotational speed input vector into the first convolutional layer of the multi-scale neighborhood feature extraction module to obtain the rotational speed associated feature vector at the first neighborhood scale, wherein the first convolutional layer has a first one-dimensional convolutional kernel of a first length; The second convolutional unit is used to input the rotational speed input vector into the second convolutional layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood-scale rotational speed associated feature vector, wherein the second convolutional layer has a second one-dimensional convolutional kernel of a second length, and the first length is different from the second length; and The cascade unit is used to cascade the first neighborhood scale rotational speed associated feature vector and the second neighborhood scale rotational speed associated feature vector to obtain the multi-scale rotational speed feature vector.

3. The fault diagnosis system for wind turbine generator sets according to claim 2, characterized in that, The first convolutional unit is further configured to: use the first convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional encoding on the rotational speed input vector using the following formula to obtain the first neighborhood scale rotational speed associated feature vector; The formula is as follows: Where a is the width of the first convolution kernel in the x direction, F is the parameter vector of the first convolution kernel, G is the local vector matrix that operates with the convolution kernel function, w is the size of the first convolution kernel, and X represents the tension input vector; The second convolutional unit is further configured to: use the second convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional encoding on the rotational speed input vector using the following formula to obtain the second neighborhood scale rotational speed associated feature vector; The formula is as follows: Where b is the width of the second convolution kernel in the x direction, F is the parameter vector of the second convolution kernel, G is the local vector matrix of the operation with the convolution kernel function, m is the size of the second convolution kernel, and X represents the tension input vector.

4. The fault diagnosis system for wind turbine generator sets according to claim 3, characterized in that, The waveform generator is a generative adversarial network.

5. The fault diagnosis system for wind turbine generator sets according to claim 4, characterized in that, The feature extraction module is further configured to: use each layer of the convolutional neural network model, which serves as the feature extractor, to process the input data during the forward propagation of the layer: The input data is processed by convolution to obtain a convolutional feature map; The convolutional feature map is subjected to mean pooling based on the local channel dimension to obtain a pooled feature map; as well as The pooled feature map is nonlinearly activated to obtain an activated feature map; The output of the last layer of the convolutional neural network model, which serves as the feature extractor, is the generated vibration feature matrix, the first vibration waveform feature matrix, and the second vibration waveform feature matrix. The input of the first layer of the convolutional neural network model, which serves as the feature extractor, is the generated vibration waveform diagram, the waveform diagram of the first vibration signal, and the waveform diagram of the second vibration signal.

6. The fault diagnosis system for wind turbine generator sets according to claim 5, characterized in that, The fusion optimization module is further configured to: fuse the generated vibration feature matrix and the generated vibration waveform diagram according to the following formula to obtain the optimized generated vibration feature matrix; The formula is as follows: Where M1 represents the generated vibration waveform, M2 represents the generated vibration feature matrix, and f i This represents the eigenvalue at each position in the generated vibration feature matrix. The value represents the mean of the values ​​at each position in the generated vibration feature matrix, and N is the scale of the generated vibration feature matrix. exp(·) represents the exponential operation on the matrix, where exponential operation on the matrix represents the natural exponential function value raised to the power of the eigenvalues ​​at each position in the matrix. α is a hyperparameter. ⊙ represents positional addition of the characteristic matrix, and ⊙ represents positional dot product.

7. The fault diagnosis system for wind turbine generator sets according to claim 6, characterized in that, The filtering module is further configured to: calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix using the following formula; The formula is as follows: Where M P M represents the optimized generation of the vibration feature matrix. V M represents the first vibration waveform feature matrix. D This represents the difference feature matrix. This indicates differentiation based on position.

8. The fault diagnosis system for wind turbine generator sets according to claim 7, characterized in that, The transfer module is further configured to: calculate the transfer matrix of the second vibration waveform feature matrix relative to the difference feature matrix as the classification feature matrix using the following formula; The formula is as follows: M W =M*M D Where M W M represents the characteristic matrix of the second vibration waveform. D Let M represent the difference feature matrix, and let M represent the classification feature matrix.

9. The fault diagnosis system for wind turbine generator sets according to claim 8, characterized in that, The fault diagnosis result generation module is further configured to: process the classification feature matrix using the classifier according to the following formula to generate the classification result; The formula is as follows: Where O is the output matrix, W i and b i Let be the weight and bias matrices corresponding to the i-th category, respectively. exp(·) represents the exponential operation of the matrix, which means the natural exponential function value is the power of the eigenvalues ​​at each position in the matrix.

10. A fault diagnosis method for a wind turbine generator set, characterized in that, include: Obtain the generator speed values ​​at multiple predetermined time points within a predetermined time period; Acquire a first vibration signal and a second vibration signal from a first vibration sensor and a second vibration sensor deployed at the head and tail ends of the generator; The generator speed values ​​at multiple predetermined time points within the predetermined time period are arranged into a speed input vector according to the time dimension and then processed by a multi-scale neighborhood feature extraction module to obtain a multi-scale speed feature vector. The multi-scale rotational speed feature vector is passed through a waveform generator to obtain a generated vibration waveform. The generated vibration waveform, the waveform of the first vibration signal, and the waveform of the second vibration signal are processed by a convolutional neural network model as a feature extractor to obtain the generated vibration feature matrix, the first vibration waveform feature matrix, and the second vibration waveform feature matrix. The generated vibration feature matrix and the generated vibration waveform are fused to obtain an optimized generated vibration feature matrix; Calculate the difference feature matrix between the optimized generated vibration feature matrix and the first vibration waveform feature matrix; The transition matrix of the second vibration waveform feature matrix relative to the difference feature matrix is ​​calculated as the classification feature matrix; as well as The classification feature matrix is ​​passed through a classifier to obtain classification results, which are used to represent the probability values ​​of the wind turbine generator set under test having each fault category label.

Images