Method for monitoring wind turbine generator, system and computer readable storage medium

Abstract

Embodiments of the present application provide a wind turbine monitoring method and system and a computer readable storage medium. The method comprises: obtaining CMS vibration data of a first wind turbine at a predetermined time sequence; obtaining feature information of at least one dimension of the first wind turbine at each timestamp based on the CMS vibration data at the predetermined time sequence; obtaining predicted feature information at each timestamp based on the feature information at each timestamp and a pre-trained unit prediction model; performing similarity evaluation on the feature information and the predicted feature information corresponding to each timestamp; and monitoring the health condition of the first wind turbine based on the result of the similarity evaluation. The present application has good universality.

Description

Technical Field

[0001] This application relates to the field of wind turbine technology, and in particular to a monitoring method and system for wind turbine generators, as well as a computer-readable storage medium. Background Technology

[0002] With the gradual depletion of energy sources such as coal and oil, humanity is increasingly emphasizing the utilization of renewable energy. Wind energy, as a clean and renewable energy source, is receiving growing attention worldwide. Along with the continuous development of wind power technology, the application of wind turbines in power systems is increasing. Wind turbines are large-scale devices that convert wind energy into electrical energy, typically installed in areas rich in wind resources. To detect potential faults in wind turbines in advance and ensure their normal operation, it is necessary to monitor their condition, especially vibration.

[0003] Currently, the vibration status of wind turbines is typically monitored and controlled through a Condition Monitoring System (CMS). The CMS system collects vibration data from the wind turbines in the field, and uses this data to monitor the turbine's operation, detect faults, and diagnose problems. However, current fault diagnosis methods lack generalization capabilities. Summary of the Invention

[0004] The purpose of this application is to provide a monitoring method and system for wind turbine generators, as well as a computer-readable storage medium, which can be widely used for predicting the operating trend and diagnosing faults of wind turbine generators and has good versatility.

[0005] One aspect of this application provides a method for monitoring a wind turbine. The monitoring method includes: acquiring CMS vibration data of a first wind turbine over a predetermined time series; acquiring feature information of at least one dimension of the first wind turbine at each time stamp based on the CMS vibration data of the predetermined time series; obtaining predicted feature information at each time stamp based on the feature information at each time stamp and a pre-trained turbine prediction model; performing a similarity assessment between the feature information at the corresponding time stamp and the predicted feature information; and monitoring the health status of the first wind turbine based on the results of the similarity assessment.

[0006] Another embodiment of this application provides a monitoring system for wind turbine generators. The monitoring system includes one or more processors for implementing the wind turbine generator monitoring method described above.

[0007] Another embodiment of this application provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the wind turbine monitoring method described above.

[0008] The wind turbine monitoring method, system, and computer-readable storage medium of this application can be widely used for wind turbine operation trend prediction and fault diagnosis, and have good versatility. Attached Figure Description

[0009] Figure 1 This is a flowchart of a wind turbine monitoring method according to an embodiment of this application;

[0010] Figure 2 for Figure 1 A specific process example of the monitoring method for wind turbine units shown;

[0011] Figure 3 The following are specific steps for obtaining the health indicators of bearing components of a first wind turbine at each time stamp based on CMS vibration data of a predetermined time series, according to one embodiment of this application.

[0012] Figure 4 The following are specific steps for obtaining gearbox health indicators of a first wind turbine at each time stamp based on CMS vibration data of a predetermined time series, according to one embodiment of this application.

[0013] Figure 5 This describes the process of establishing a unit prediction model according to one embodiment of this application;

[0014] Figure 6 This is a schematic diagram of the structure of a Conv-LSTM model according to an embodiment of this application;

[0015] Figure 7 This is a schematic block diagram of a wind turbine monitoring system according to an embodiment of this application. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses consistent with some aspects of this application as detailed in the appended claims.

[0017] The terminology used in the embodiments of this application is for the purpose of describing specific embodiments only and is not intended to limit the application. Unless otherwise defined, the technical or scientific terms used in the embodiments of this application should be understood in their ordinary sense by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are only used to distinguish different components. Similarly, the terms "a" or "one" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates two or more. Unless otherwise indicated, the terms "front," "rear," "lower," and / or "upper" and similar terms are for ease of description only and are not limited to a location or spatial orientation. The terms "comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" cover the elements or objects listed after "comprising" or "including" and their equivalents, and do not exclude other elements or objects. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections and can include electrical connections, whether direct or indirect. The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0018] One embodiment of this application provides a method for monitoring wind turbine generators. Figure 1 A flowchart illustrating a wind turbine monitoring method according to an embodiment of this application is provided. Figure 1 As shown, a wind turbine monitoring method according to one embodiment of this application may include steps S01 to S05.

[0019] In step S01, CMS vibration data of the first wind turbine unit at a predetermined time series are acquired.

[0020] In step S02, feature information of at least one dimension of the first wind turbine is obtained based on the CMS vibration data of the predetermined time series.

[0021] In step S03, the predicted feature information for each time stamp is obtained based on the feature information for each time stamp and the pre-trained unit prediction model.

[0022] In step S04, the similarity between the feature information under the corresponding timestamp and the predicted feature information is evaluated.

[0023] In step S05, the health status of the first wind turbine is monitored based on the results of the similarity assessment.

[0024] Figure 2 Revealed Figure 1 This is a specific process example of a wind turbine monitoring method. (Example:) Figure 2 As shown, in some embodiments, the wind turbine monitoring method of this application may include steps S11 to S18. In step S11, CMS vibration data of a predetermined time series of the first wind turbine can be acquired.

[0025] In step S12, multiple first features can be obtained from each time stamp of the first wind turbine based on the CMS vibration data of the predetermined time series.

[0026] In some embodiments, the multiple first features are time-frequency domain statistical features obtained by performing a feature extraction on the CMS vibration data. The time-frequency domain statistical features of the CMS vibration data may include, but are not limited to, multiple of the following: mean, peak value, peak-to-peak value, root mean square effective value, sample standard deviation, peak factor, impulse factor, margin factor, waveform factor, skewness, kurtosis, average frequency domain amplitude, sample variance of frequency domain amplitude, sample standard deviation of frequency domain amplitude, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, centroid frequency value, frequency mean square value, and root mean square value of frequency.

[0027] Table 1 below describes the time-frequency domain statistical characteristics of CMS vibration data.

[0028] Table 1

[0029]

[0030]

[0031]

[0032] Where, x n This represents the CMS vibration data obtained at a certain timestamp, where N represents the number of vibration data points, and x... max This represents the maximum value of the CMS vibration data at that timestamp, x. min This represents the minimum value of the CMS vibration data at that timestamp.

[0033] Therefore, based on Table 1 above, the corresponding statistical characteristics can be calculated from the obtained CMS vibration data. Each statistical characteristic can reflect the changes in the wind turbine from different perspectives.

[0034] In step S13, multiple second features of the first wind turbine can be obtained at each time stamp based on the CMS vibration data of the predetermined time series.

[0035] The multiple second features are mechanism-based features obtained by performing secondary feature extraction on CMS vibration data. In some embodiments, mechanism-based features may include, for example, bearing component health index (HI) and gearbox health index (HI). A bearing comprises four components: an outer ring, an inner ring, rolling elements, and a cage. Therefore, the bearing component health index may include the outer ring health index, the inner ring health index, the rolling element health index, and the cage health index.

[0036] Table 2 below shows the fault characteristics of various parts of the bearings and gearbox.

[0037] Table 2

[0038]

[0039] Where r represents the bearing speed in revolutions per minute; n represents the number of balls; d represents the diameter of the rolling elements; D represents the bearing pitch diameter; and α represents the contact angle of the rolling elements.

[0040] The following will combine Figure 3 This section details how to obtain the health indicators of bearing components at each time stamp of the first wind turbine based on CMS vibration data of a predetermined time series.

[0041] like Figure 3 As shown, in some embodiments, obtaining the bearing component health indicators of the first wind turbine at each time stamp based on CMS vibration data of a predetermined time series may include steps S21 to S25.

[0042] In step S21, a Fast Fourier Transform (FFT) is performed on the CMS vibration data at each time stamp in the predetermined time series to obtain the spectrum of the vibration data at each time stamp.

[0043] In step S22, the bearing rotation frequency f at each time stamp is obtained. speed .

[0044] Considering that sometimes the frequency conversion formula F in Table 2 can be used directly... r The rotational frequency obtained by calculating r / 60 may contain errors. Therefore, in some embodiments, the bearing rotational frequency f at any given time point can be obtained more accurately by the following method. speed .

[0045] (1.1) By measuring the rotational speed r and applying the rotational frequency formula F in Table 2 r =r / 60 to obtain the bearing's rotational frequency f speed Use this frequency as the initial frequency;

[0046] (1.2) Find the bearing rotational frequency f obtained in (1.1) in the spectrum. speed The index position corresponding to (i.e., the initial frequency)

[0047] (1.3) Using indexed location points The search range is defined by the predetermined positions to the left and right of the given location, for example... Where n1 represents the number of left and right floating positions, and the search range is found in the spectrum. Maximum amplitude within;

[0048] (1.4) Expand the search scope The maximum amplitude within is used as the frequency amplitude. With this frequency amplitude The frequency of the corresponding position point is taken as the final bearing rotational frequency f. speed This allows for the correction and updating of the bearing's initial rotational frequency, ensuring that the final obtained bearing rotational frequency f is accurate. speed The accuracy.

[0049] In step S23, based on the bearing rotational frequency f speed The failure frequency of the bearing component is calculated from the characteristic frequencies of the bearing components (as shown in Table 2), as shown in the following formula:

[0050] Failure frequency of bearing components = bearing rotational frequency × characteristic frequency of bearing components (1)

[0051] In step S24, the amplitudes of multiple harmonics of the fault frequency of the bearing component are determined in the spectrum.

[0052] With bearing rotation frequency f speed Similarly, considering that the failure frequency of the bearing component calculated directly by formula (1) in step S23 may not be accurate enough, in some embodiments, determining the amplitude of multiple harmonics of the failure frequency of the bearing component in the spectrum may include: determining the optimized amplitude of multiple harmonics of the failure frequency of the bearing component in the spectrum.

[0053] In some embodiments, the optimized amplitude of multiple harmonics of the failure frequency of the bearing component in the spectrum can be determined in the following manner.

[0054] (2.1) For each of the multiple harmonics of the failure frequency of the bearing component, for example, for the first harmonic of the failure frequency of the bearing component. Locate the first harmonic of the fault frequency of the bearing component in the frequency spectrum. The corresponding index position point

[0055] (2.2) Using indexed location points The search range is defined by the predetermined positions to the left and right of the given location, for example, Where n2 represents the number of left and right floating position points, and the search range is found in the spectrum. Maximum amplitude within;

[0056] (2.3) Expand the search scope The maximum amplitude within is used as the optimized amplitude at one harmonic of the failure frequency of the bearing component.

[0057] (2.4) For the second, third, ..., nth harmonics of the failure frequency of the bearing components, i.e. Accordingly, the second, third, ..., nth harmonics of the fault frequency of the bearing component can be found in the frequency spectrum. The corresponding index position point

[0058] (2.5) Find the search range in the spectrum respectively Maximum amplitude within;

[0059] (2.6) Determine the search range respectively The maximum amplitude within is used as the optimized amplitude for the second, third, ..., nth harmonics of the failure frequency of the bearing component.

[0060] in, It can reflect changes in vibration signals caused by a fault.

[0061] In some embodiments, the multiple harmonics of the failure frequency of the bearing component may include, for example, one to six harmonics.

[0062] Furthermore, considering that the frequencies obtained by FFT transformation based on actual CMS vibration data may sometimes not contain many harmonics of the bearing component failure frequencies, in some embodiments, it can first be determined whether the spectrum contains a sixth harmonic of the bearing component failure frequency. If the spectrum does not contain a sixth harmonic, the harmonics of the bearing component failure frequency can be determined from the harmonics contained in the spectrum.

[0063] Maximum harmonics in the spectrum = x-axis of the spectrum / frequency transition

[0064] For example, if the maximum frequency value in the spectrum obtained by the Fast Fourier Transform can only include up to the 5th harmonic, then the harmonic number of the fault frequency of the bearing component can only be calculated up to 5.

[0065] In step S25, the health index of the bearing component is obtained based on the amplitude of multiple harmonics of the failure frequency of the bearing component.

[0066] If optimized amplitudes of multiple harmonics of the failure frequency of the bearing component are obtained, a bearing component health index can be obtained based on the amplitudes of these harmonics. In some embodiments, the optimized amplitudes of each harmonic of the failure frequency of the bearing component can be summed to obtain the bearing component health index.

[0067] The above description uses the example of how to obtain bearing component health indicators. Bearing components include four parts: outer ring, inner ring, rolling elements, and cage. Therefore, the health indicators for the outer ring, inner ring, rolling elements, and cage can be obtained by referring to the bearing component health indicators described above.

[0068] The following will combine Figure 4 This section details how to obtain gearbox health indicators for the first wind turbine at each time stamp based on CMS vibration data from a predetermined time series.

[0069] like Figure 4 As shown, in some embodiments, obtaining the gearbox health index of the first wind turbine at each time stamp based on the CMS vibration data of a predetermined time series may include steps S31 to S35.

[0070] In step S31, a fast Fourier transform is performed on the CMS vibration data at each time stamp in the predetermined time series to obtain the spectrum of the vibration data at each time stamp.

[0071] In step S32, the rotational frequency f of the shaft containing the gearbox gear at each time stamp is obtained. spee .

[0072] With bearing rotation frequency f speed Similarly, considering that sometimes the frequency conversion formula F in Table 2 can be used directly... r The rotational frequency obtained by calculating r / 60 may contain errors. Therefore, in some embodiments, the rotational frequency f of the shaft containing the gearbox gear at any given time point can be obtained more accurately by the following method. spee .

[0073] (3.1) By measuring the rotational speed r and applying the rotational frequency formula F in Table 2 r =r / 60 to obtain the rotational frequency f of the shaft containing the gears in the gearbox. spee This rotational frequency is used as the initial rotational frequency of the shaft where the gears of the gearbox are located;

[0074] (3.2) Search the spectrum for the f of the shaft containing the gearbox gear obtained in (3.1). speed1The index position corresponding to (i.e., the initial frequency)

[0075] (3.3) Using indexed location points The search range is defined by the predetermined positions to the left and right of the given location, for example, Where n3 represents the number of left and right floating position points, and the search range is found in the spectrum. Maximum amplitude within;

[0076] (3.4) Expand the search scope The maximum amplitude within is used as the frequency amplitude. With frequency amplitude The frequency of the corresponding position point is taken as the final gearbox rotation frequency f. speed1 This allows for the correction and updating of the initial gearbox frequency, ensuring that the final obtained gearbox frequency f is accurate. sp The accuracy.

[0077] In step S33, based on the rotational frequency f of the shaft where the gearbox gear is located... speed1 The meshing frequency of the gearbox is calculated from the number of gear teeth, as shown in the following formula:

[0078] The meshing frequency of the gearbox = the rotational frequency of the shaft containing the gear × the number of teeth of the gear (2)

[0079] In step S34, the amplitudes of multiple harmonics of the gearbox's meshing frequency are determined in the spectrum.

[0080] With the rotational frequency f of the gearbox speed1 Similarly, considering that the gearbox meshing frequency calculated directly by formula (2) in step S33 may not be accurate enough, in some embodiments, determining the amplitude of multiple harmonics of the gearbox meshing frequency in the spectrum may include: determining the optimized amplitude of multiple harmonics of the gearbox meshing frequency in the spectrum.

[0081] In some embodiments, the optimized amplitudes of multiple harmonics of the gearbox's meshing frequency can be determined in the spectrum in the following manner.

[0082] (3.1) For each of the multiple harmonics of the gearbox's meshing frequency, for example, for the first harmonic of the gearbox's meshing frequency. Find the first harmonic of the gearbox's meshing frequency in the spectrum. The corresponding index position point

[0083] (3.2) Using the index position point The search range is defined by the predetermined positions to the left and right of the given location, for example, Where n4 represents the number of left and right floating position points, and the search range is found in the spectrum. Maximum amplitude within;

[0084] (3.3) Expand the search scope The maximum amplitude within is used as the optimized amplitude of the first harmonic of the gearbox's meshing frequency.

[0085] (3.4) For the second, ..., nth harmonics of the gearbox's meshing frequency Then, we can similarly repeat steps (3.1) to (3.3) above to obtain the optimized amplitudes of the second, ..., nth harmonics of the gearbox's meshing frequency.

[0086] According to the diagnostic mechanism, if the gearbox's meshing frequency is accompanied by sidebands of the rotational frequency on both sides, the gearbox is considered to have a fault. Therefore, when calculating the gearbox's health indicators, the optimized amplitude of each harmonic of the gearbox's meshing frequency will be considered.

[0087] In some embodiments, the multiple harmonics of the gearbox's meshing frequency may include, for example, first to third harmonics.

[0088] In step S35, a gearbox health index is obtained based on the amplitude of multiple harmonics of the gearbox's meshing frequency.

[0089] Having obtained the optimized amplitudes of multiple harmonics of the gearbox's meshing frequency, the health index of the bearing components can be obtained based on these amplitudes. In some embodiments, the optimized amplitudes of each harmonic of the gearbox's meshing frequency can be summed to obtain the gearbox health index.

[0090] Return to reference Figure 2 In step S14, a two-dimensional array is obtained based on multiple first features and multiple second features under each time stamp of the first wind turbine. For example, multiple first features can be used as the vertical coordinates of the two-dimensional array, and multiple second features can be used as the horizontal coordinates of the two-dimensional data.

[0091] In step S15, the two-dimensional array under each timestamp is converted into the actual running image under each timestamp.

[0092] In step S16, the two-dimensional arrays under each timestamp are converted into the input data format of the pre-trained unit prediction model and input into the unit prediction model to output a prediction image with timestamps.

[0093] like Figure 5 As shown, in some embodiments, the wind turbine monitoring method of this application may further include steps S41 to S43.

[0094] In step S41, a unit prediction model is established in advance.

[0095] In one embodiment, the unit prediction model of this application includes a Conv-LSTM model. Figure 6 A schematic diagram of the structure of a Conv-LSTM model according to an embodiment of this application is shown. Figure 6 As shown, the computation process of the Conv-LSTM model is similar to that of the traditional LSTM model. It consists of five parts: input gate, forget gate, memory unit, output gate, and hidden state output, expressed as follows:

[0096] i t =σ(W i *x t +U i *h t-1 +b i )

[0097] f t =σ(W f *x t +U f *h t-1 +b f )

[0098]

[0099] o t =σ(W o *x t +U o *h t-1 +b o )

[0100]

[0101] In the above formula, "*" represents the convolution operator, and the symbol... Represents the Hadamard product; i t Indicates the input gate, f t Representing the Gate of Oblivion, C t h is a memory unit. t Represents the hidden state, o t Represents the output gate; W i U i and b i Let W represent the input weight matrix, the previous time step active value weight matrix, and the bias matrix, respectively. f U f and b f W represents the input weight matrix, the previous time step activity weight matrix, and the bias matrix in the forget gate, respectively; c U c and bc W represents the input weight matrix, the hidden layer weight matrix from the previous time step, and the bias matrix in the candidate memory unit, respectively; o U o and b o These represent the input weight matrix, the hidden layer weight matrix at the previous time step, and the bias matrix in the output gate, respectively.

[0102] When performing model input calculations, the ordinary LSTM model uses the Hadamard product to calculate the weights of the input and the weights of the previous layer, while the Conv-LSTM model uses convolution calculation instead. This can better capture the spatial characteristics of the data, while also retaining the advantages of traditional LSTM in simulating time series data. It can combine historical and current data to predict the future operating status of wind turbines.

[0103] The input data format of the Conv-LSTM model is a five-dimensional tensor of (sample, time, rows, cols, channels), where sample represents the number of samples, which in this embodiment is the number of two-dimensional arrays with timestamps; time is the time step, representing the number of image frames in each training sample sequence, i.e., how many timestamps of two-dimensional data are used to predict a prediction image; rows represents the number of pixels in the horizontal direction of each image frame, i.e., how many rows, such as the five health indicators in this application; col represents the number of pixels in the vertical direction of each image frame, such as the nineteen statistical features in this application; channels represent the number of color channels, the number of parameters describing the RGB value of each pixel, with 3 channels for color images and 1 channel for black and white images.

[0104] Continue to refer to Figure 5 In step S42, CMS historical vibration data of the second wind turbine over a predetermined time series is acquired. This CMS historical vibration data includes a first portion of historical vibration data and a second portion of historical vibration data. For example, the first 80% of the CMS historical vibration data of the second wind turbine over the predetermined time series can be used as the first portion of historical vibration data, and the last 20% can be used as the second portion of historical vibration data.

[0105] In step S43, the first part of historical vibration data is used as a training set and input into the unit prediction model, such as a Conv-LSTM model, to train the unit prediction model (Conv-LSTM model). Of course, the first part of historical vibration data can be normalized before being input into the Conv-LSTM model to improve model training accuracy and convergence speed. Simultaneously, to better identify faults, the first part of historical vibration data consists entirely of historical health data from the second wind turbine. That is, the trained Conv-LSTM model is a model under healthy conditions. If the actual operating trend of the second wind turbine deviates from the operating trend predicted by the model, the second wind turbine may experience a fault. The healthy model trained from the historical health data is stored. After data training, a well-trained Conv-LSTM model can finally be obtained.

[0106] In some embodiments, to effectively evaluate the accuracy of the Conv-LSTM model, the loss function of the optimizer in the Conv-LSTM model includes a cosine similarity function, which is selected to evaluate the model loss. Cosine similarity is used to evaluate the similarity between two vectors by calculating the cosine of the angle between them. For example, for an n-dimensional vector A = [A1, A2, ..., A...], ... n ] and B = [B1, B2, ..., B n The cosine of the angle θ between them is shown in the following formula:

[0107]

[0108] In the embodiments of this application, A represents the actual operating image of the first wind turbine, and B represents the predicted image of the first wind turbine. Therefore, the similarity between the two images can be determined by calculating the cosine value of the angle between the actual operating image and the predicted image output by the model. Furthermore, the model continuously iterates and corrects the model parameters to optimize the cosine value of the angle between the two images, thereby finally determining the weight matrices and bias matrices in the Conv-LSTM model structure, i.e., W in the formula (3) above. i U i b i W f U f b f W c U c b c W o U o and b o This allows us to obtain a trained aircraft prediction model, such as a Conv-LSTM model.

[0109] In some embodiments, the crew prediction model (Conv-LSTM model) outputs a predicted image for a two-dimensional array under a predetermined number of timestamps. For example, in this embodiment, the Conv-LSTM model can output a predicted image for a two-dimensional array under every five timestamps.

[0110] Table 3 below illustrates the inputs and outputs of the Conv-LSTM model.

[0111] Table 3

[0112]

[0113] In Table 3, x represents the input of the Conv-LSTM model, and y represents the output of the Conv-LSTM model. For example, the Conv-LSTM model can predict the image at the 6th timest based on a two-dimensional array of timestamps 1 to 5, predict the image at the 6th timest based on a two-dimensional array of timestamps 2 to 6, and so on, predicting the image at the 'Samples' timest based on a two-dimensional array of timestamps (Samples-5) to (Samples-1). And so on, the Conv-LSTM model makes predictions according to the time series.

[0114] After obtaining the trained Conv-LSTM model, to verify its effectiveness, the second part of the historical vibration data of the second wind turbine can be used as a test set and input into the turbine prediction model, such as the Conv-LSTM model, to validate the effectiveness of the turbine prediction model (Conv-LSTM model). The second part of the historical vibration data includes the historical health data and historical fault data of the second wind turbine.

[0115] A trained Conv-LSTM model has predictive capabilities, and theoretically, it should predict healthy states. However, if the input validation data includes fault data, this fault data will be fed into the Conv-LSTM model, causing a discrepancy between the results and the actual data. This indicates that a fault has occurred in the actual data. Therefore, when validation data is input into the Conv-LSTM model, the data trend after the fault time will deviate from the trend predicted by the Conv-LSTM model. If the Conv-LSTM model can identify this trend, it is considered effective, the validation is successful, and a trained Conv-LSTM model is obtained. A validated Conv-LSTM model can then be applied to actual wind turbine fault trend prediction and diagnosis applications.

[0116] In the embodiments of this application, the first wind turbine and the second wind turbine can be the same or different. That is, the turbine prediction model of this application can be trained and validated using historical vibration data of the wind turbine itself, and the trained model can be used for fault trend prediction and diagnosis of the wind turbine itself. Furthermore, the trained model can be used for fault trend prediction and diagnosis of other wind turbines. The turbine prediction model of this application can achieve cross-unit fault trend prediction and diagnosis, and has excellent generalization ability. Therefore, the wind turbine monitoring method of this application has good versatility.

[0117] Continue to return to the reference Figure 2 In step S17, the actual running image in step S15 under the corresponding timestamp is compared with the predicted image output by the model in step S16 for similarity evaluation.

[0118] In some embodiments, SSIM (Structural Similarity Index) can be used to evaluate the similarity between the actual running image and the predicted image at the corresponding timestamp to obtain a structural similarity index.

[0119] Structural similarity (SSIM) is a metric for measuring the similarity between two images, first proposed in 2010 by the Image and Video Engineering Laboratory at the University of Texas at Austin. This metric effectively reflects the similarity between two images. The SSIM value ranges from -1 to 1. When SSIM = -1, it indicates that the two images are completely dissimilar; when SSIM = 1, it indicates that the two images are identical. In other words, the closer the SSIM value is to 1, the more similar the two images are.

[0120] SSIM measures the similarity between two images x and y, and consists of three contrast modules: brightness l(x,y), contrast c(x,y), and structure s(x,y). Brightness is represented by the mean, contrast by the variance after mean normalization, and structure by the correlation coefficient. The similarity S(x,y) between two images x and y is expressed as follows:

[0121] S(x,y)=f(l(x,y),c(x,y),s(x,y)) (5)

[0122] The obtained similarity follows three principles:

[0123] Symmetry: S(x,y)=S(y,x);

[0124] Boundedness: S(x,y)≤1;

[0125] The limit value is unique: S(x,y)=1 if and only if x=y.

[0126] When calculating brightness, if an image has N pixels, and the pixel value of each pixel is x... i Therefore, the average brightness of the image is:

[0127]

[0128] The formula for measuring the brightness similarity of two images (x and y) is as follows:

[0129]

[0130] C1 is used here to prevent the denominator from being zero, and C1 = (K1L) 2 ,

[0131] Where k1<<1 is a constant, usually taking the value of 0.01; L is the dynamic range of grayscale, which is determined by the data type of the image. If the data is of type uint8, then L=255. It can be seen that formula (7) is symmetrical and is always less than or equal to 1, and is 1 when x=y.

[0132] Contrast ratio is the degree of variation in brightness and darkness of an image, which is also the standard deviation of pixel values. Its calculation formula is shown below:

[0133]

[0134] The following contrast similarity formula (9) and brightness similarity formula (7) are very similar. The difference is that the mean in the following contrast similarity formula (9) is replaced by the variance, that is:

[0135]

[0136] Where: C2=(K2L) 2 ,

[0137] K2 is generally taken as 0.03. Formula (9) is also symmetrical and less than or equal to 1. The equality holds when x = y.

[0138] It's important to note that for an image, brightness and contrast are scalars, while its structure cannot be represented by a single scalar. Instead, it should be represented by a vector composed of all the pixels in the image. Therefore, when calculating structural similarity, to eliminate the influence of brightness and contrast (i.e., to eliminate the influence of the mean and standard deviation), the two vectors are usually normalized. This is used to study (x-μ)... x ) / σ x and (y-μ) y ) / σ y The relationship between these factors leads to the structural similarity formula shown below:

[0139]

[0140] In the formula σ xy The formula is defined as follows:

[0141]

[0142] To prevent the denominator from being zero, C3 is added to both the numerator and denominator. Combining formulas (7), (9), and (10), the similarity formula (5) between two images becomes the following formula:

[0143] SSIM(x,y)=l(x,y)·c(x,y)·s(x,y)) (11)

[0144] Let C3 = C2 / 2. The numerator of c(x,y) and the denominator of s(x,y) can be simplified to obtain the final similarity formula:

[0145]

[0146] Therefore, the structural similarity index of the two images x and y can be obtained by formula (12).

[0147] In step S18, the health status of the first wind turbine is monitored based on the similarity assessment results obtained in step S17.

[0148] In some embodiments, step S18 may include: predicting and diagnosing the failure trend of the first wind turbine based on the structural similarity index obtained in step S17. When the calculated structural similarity index is less than a predetermined threshold, it can be determined that the first wind turbine has failed.

[0149] The wind turbine monitoring method of this application takes into account the synergistic effect of wind turbines in time and space, and develops a Conv-LSTM deep learning model for wind turbine fault diagnosis. This Conv-LSTM model can extract the spatial features of data like a convolutional neural network, and also retains the ability of traditional LSTM to remember time series.

[0150] The wind turbine monitoring method of this application is based on the CMS vibration data of the wind turbine. It performs secondary feature extraction on the vibration data in conjunction with the vibration mechanism, and uses the extracted secondary features as input to a Conv-LSTM model. Historical operating data of a single wind turbine is selected, and the Conv-LSTM model is trained using the health data of that wind turbine. The trained Conv-LSTM model is then used to predict the health and fault data of the current wind turbine, and the accuracy of the model is evaluated using a structural similarity index. Finally, the validated model can be used for trend prediction of operating data from the current wind turbine or other wind turbines, further demonstrating the versatility of the Conv-LSTM model in cross-unit trend prediction and fault diagnosis.

[0151] This application also provides a monitoring system 200 for wind turbine generators. Figure 7 A schematic block diagram of a wind turbine monitoring system 200 according to an embodiment of this application is shown. Figure 7 As shown, the wind turbine monitoring system 200 may include one or more processors 201 for implementing the wind turbine monitoring method described in any of the above embodiments. In some embodiments, the wind turbine monitoring system 200 may include a computer-readable storage medium 202, which may store a program that can be called by the processor 201, and may include a non-volatile storage medium. In some embodiments, the wind turbine monitoring system 200 may include memory 203 and an interface 204. In some embodiments, the wind turbine monitoring system 200 of this application embodiment may also include other hardware depending on the actual application.

[0152] The wind turbine monitoring system 200 of this application embodiment has similar beneficial technical effects to the wind turbine monitoring method described above, so it will not be repeated here.

[0153] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a program that, when executed by a processor, implements the wind turbine monitoring method described in any of the above embodiments.

[0154] The embodiments of this application may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to: novel memories such as phase-change memory / resistive random access memory / magnetic memory / ferroelectric memory (PRAM / RRAM / MRAM / FeRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only optical disc read-only memory (CD-ROM), digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0155] The monitoring method, system, and computer-readable storage medium for wind turbine generators provided in this application have been described in detail above. Specific examples have been used to illustrate the monitoring method, system, and computer-readable storage medium for wind turbine generators in this application. The descriptions of the embodiments above are only for helping to understand the core ideas of this application and are not intended to limit this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the spirit and principles of this application, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A monitoring method for wind turbine generators, characterized in that: It includes: Obtain CMS vibration data of the first wind turbine unit at a predetermined time series; Based on the predetermined time series CMS vibration data, at least one dimension of feature information of the first wind turbine is obtained at each time stamp, including: obtaining multiple first features and multiple second features of the first wind turbine at each time stamp based on the predetermined time series CMS vibration data; and the multiple first features and multiple second features of the first wind turbine at each time stamp constitute a two-dimensional array at each time stamp, wherein the multiple first features are time-frequency domain statistical features of CMS vibration data, and the multiple second features are mechanism-based features, including bearing component health indicators and gearbox health indicators; Based on the feature information at each timestamp and the pre-trained unit prediction model, the prediction feature information at each timestamp is obtained; The similarity assessment is performed between the feature information under the corresponding timestamp and the predicted feature information; and The health status of the first wind turbine is monitored based on the results of the similarity assessment.

2. The monitoring method as described in claim 1, characterized in that: The method of obtaining the prediction feature information for each time stamp based on the feature information for each time stamp and the pre-trained unit prediction model includes: The two-dimensional arrays under each timestamp are converted into actual running images under each timestamp; and The two-dimensional arrays under each timestamp are converted into the input data format of a pre-trained unit prediction model and input into the unit prediction model to output a prediction image with timestamps. The similarity assessment of the feature information and the predicted feature information under the corresponding timestamp includes: assessing the similarity between the actual running image and the predicted image under the corresponding timestamp.

3. The monitoring method as described in claim 1, characterized in that: The time-frequency domain statistical features include multiple values ​​from the following: mean, peak value, peak-to-peak value, root mean square effective value, sample standard deviation, peak factor, impulse factor, margin factor, waveform factor, skewness, kurtosis, average frequency domain amplitude, sample variance of frequency domain amplitude, sample standard deviation of frequency domain amplitude, skewness index of frequency domain amplitude, kurtosis index of frequency domain amplitude, centroid frequency value, frequency mean square value, and root mean square value of frequency.

4. The monitoring method as described in claim 1, characterized in that: The health indicators of the bearing components include the health indicators of the outer ring, the health indicators of the inner ring, the health indicators of the rolling elements, and the health indicators of the cage.

5. The monitoring method as described in claim 1, characterized in that: The health indicators of the bearing components of the first wind turbine at each time stamp are obtained based on the CMS vibration data of the predetermined time series, including: A fast Fourier transform is performed on the CMS vibration data at each time stamp in the predetermined time series to obtain the spectrum of the vibration data at each time stamp; Obtain the bearing rotation frequency at each timestamp; The failure frequency of the bearing component is calculated based on the bearing's rotational frequency and the characteristic frequency of the bearing component. The amplitudes of multiple harmonics of the fault frequency of the bearing component are determined in the spectrum; and The health index of the bearing component is obtained based on the amplitude of multiple harmonics of the failure frequency of the bearing component.

6. The monitoring method as described in claim 5, characterized in that: The process of obtaining the bearing rotation frequency at each timestamp includes: The initial rotational frequency of the bearing is obtained by measuring the rotational speed; Locate the index position point corresponding to the initial rotational frequency of the bearing in the spectrum; Using predetermined positions to the left and right of the index position as the search range, find the maximum amplitude within the search range in the spectrum; and The maximum amplitude within the search range is taken as the rotational frequency amplitude, and the frequency of the position point corresponding to the rotational frequency amplitude is taken as the rotational frequency of the bearing.

7. The monitoring method as described in claim 5, characterized in that: The amplitudes of the multiple harmonics used to determine the fault frequency of the bearing component in the spectrum include: Determining the optimized amplitude of multiple harmonics of the fault frequency of the bearing component in the spectrum, including: For each of the multiple harmonics of the fault frequency of the bearing component, find the index position point corresponding to that harmonic of the fault frequency of the bearing component in the spectrum; Using predetermined positions to the left and right of the index position as the search range, find the maximum amplitude within the search range in the spectrum; and The maximum amplitude within the search range is used as the optimized amplitude of that octet of the fault frequency of the bearing component.

8. The monitoring method as described in claim 6, characterized in that: The method of obtaining the health index of the bearing component based on the amplitude of multiple harmonics of the failure frequency of the bearing component includes: The optimized amplitudes of each harmonic of the failure frequency of the bearing component are summed to obtain the health index of the bearing component.

9. The monitoring method as described in claim 1, characterized in that: The gearbox health indicators of the first wind turbine at each time stamp are obtained based on the CMS vibration data of the predetermined time series, including: A fast Fourier transform is performed on the CMS vibration data at each time stamp in the predetermined time series to obtain the spectrum of the vibration data at each time stamp; Obtain the rotational frequency of the shaft containing the gear in the gearbox at each timestamp; The meshing frequency of the gearbox is calculated based on the rotational frequency of the shaft where the gear is located and the number of teeth of the gear. The amplitudes of multiple harmonics of the gearbox's meshing frequency are determined in the spectrum; and The gearbox health index is obtained based on the amplitude of multiple harmonics of the gearbox's meshing frequency.

10. The monitoring method as described in claim 8, characterized in that: The process of obtaining the gearbox frequency at each timestamp includes: The initial rotational frequency of the shaft containing the gearbox gears is obtained by measuring the rotational speed. Locate the index position point in the spectrum corresponding to the initial rotational frequency of the shaft where the gear of the gearbox is located; Using predetermined positions to the left and right of the index position as the search range, find the maximum amplitude within the search range in the spectrum; and The maximum amplitude within the search range is taken as the rotational frequency amplitude, and the frequency of the position point corresponding to the rotational frequency amplitude is taken as the rotational frequency of the gearbox.

11. The monitoring method as described in claim 8 or 9, characterized in that: The amplitudes of the multiple harmonics determining the meshing frequency of the gearbox in the spectrum include: Determining the optimized amplitudes of multiple harmonics of the gearbox's meshing frequency in the spectrum includes: For each of the multiple harmonics of the gearbox's meshing frequency, find the index position point corresponding to that harmonic of the gearbox's meshing frequency in the spectrum; Using predetermined positions to the left and right of the index position as the search range, find the maximum amplitude within the search range in the spectrum; and The maximum amplitude within the search range is used as the optimized amplitude of that octave of the gearbox's meshing frequency.

12. The monitoring method as described in claim 10, characterized in that: The method of obtaining the gearbox health index based on the amplitude of multiple harmonics of the gearbox's meshing frequency includes: The optimized amplitudes of each harmonic of the gearbox's meshing frequency are summed to obtain the gearbox's health index.

13. The monitoring method as described in claim 1, characterized in that: Also includes: The prediction model for the unit is established in advance; Acquire CMS historical vibration data of a second wind turbine unit over a predetermined time series, wherein the CMS historical vibration data includes a first part of historical vibration data and a second part of historical vibration data; and The first part of the historical vibration data is used as the training set to train the unit prediction model, and the second part of the historical vibration data is used as the test set to verify the unit prediction model so as to finally obtain the trained unit prediction model.

14. The monitoring method as described in claim 13, characterized in that: The first part of the historical vibration data consists entirely of the historical health data of the second wind turbine, while the second part of the historical vibration data includes the historical health data and historical fault data of the second wind turbine.

15. The monitoring method as described in claim 13, characterized in that: The first wind turbine may be the same as or different from the second wind turbine.

16. The monitoring method as described in claim 13, characterized in that: The unit prediction model includes the Conv-LSTM model.

17. The monitoring method as described in claim 16, characterized in that: The loss function of the optimizer in the Conv-LSTM model includes the cosine similarity function.

18. The monitoring method as described in claim 2, characterized in that: The step of converting the two-dimensional array under each timestamp into the input data format of the pre-trained crew prediction model and inputting it into the crew prediction model to output a predicted image with timestamps includes: For each predetermined number of timestamps in the two-dimensional array, the unit prediction model outputs a prediction image.

19. The monitoring method as described in claim 2, characterized in that: The step of evaluating the similarity between the actual running image and the predicted image under the corresponding timestamp includes: SSIM is used to evaluate the similarity between the actual running image and the predicted image at the corresponding timestamp to obtain a structural similarity index. The monitoring of the health status of the first wind turbine based on the results of the similarity assessment includes: predicting and diagnosing the failure trend of the first wind turbine based on the obtained structural similarity index.

20. The monitoring method as described in claim 19, characterized in that: The structural similarity index varies within the range of [-1, 1]. When the calculated structural similarity index is less than a predetermined threshold, it is determined that the first wind turbine unit has malfunctioned.

21. A monitoring system for a wind turbine generator, characterized in that: It includes one or more processors for implementing the monitoring method for wind turbines as described in any one of claims 1-20.

22. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the monitoring method for wind turbines as described in any one of claims 1-20.

Images