Wind turbine variable pitch bearing data enhancement method and device based on gan

Abstract

The application discloses a wind turbine variable pitch bearing data enhancement method and device based on GAN, vibration signals of a fault and a healthy wind turbine variable pitch bearing are collected by using a vibration signal collection system, wherein the vibration signal collection system comprises a single-shaft vibration acceleration sensor, a data collection card and an upper computer, after the collected vibration signals are segmented and denoised, a data set is constructed according to the setting of labels of the health conditions of the variable pitch bearing, a convolutional autoencoder neural network is built, and the collected vibration signals are used to train the convolutional autoencoder for extracting feature information; a generative adversarial neural network model with multiple generators is constructed, Gaussian noise is inputted for generating healthy vibration signals or fault vibration signals, and the generated healthy or fault signals are filled into an original data set to update the data set. Through the application, data enhancement of an unbalanced data set of a wind turbine variable pitch bearing can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine data processing, in particular to a wind turbine variable pitch bearing data enhancement method and device based on GAN. BACKGROUND

[0002] In the operation process of the wind turbine, the variable pitch bearing is an important supporting component for adjusting the angle of the wind turbine blade. The variable pitch bearing of the wind turbine is an important component connecting the hub and the blade of the wind turbine, and bears the function of transmitting load and torque. When the variable pitch bearing is in working condition, when each element in the variable pitch bearing is subjected to extrusion force and mutual wear, the working condition of the variable pitch bearing changes constantly, and this changing working condition will largely cause the variable pitch bearing to fail, thereby causing economic loss and manufacturing safety hazards. However, it is usually difficult to collect fault signals in a large number of running wind turbines, and when fault diagnosis is performed, the problem of unbalanced data samples often occurs, thereby causing the fault diagnosis algorithm to be not accurate enough. Data enhancement of the variable pitch bearing is an important means to improve the accuracy of fault diagnosis.

[0003] Generative Adversarial Networks (GAN) is a deep learning model proposed by Goodfellow in 2014. The structure of the standard GAN includes a generator and a discriminator, and the training is carried out in an adversarial game state. The generator generates data using noise, and the discriminator determines whether the input data is real data or generated data according to the distribution of the generated data and the distribution of the real data. The purpose of the generator is to generate data with the same distribution as the real data to "fool" the discriminator, and the purpose of the discriminator is to distinguish each generated data that attempts to "pretend" to be real data. GAN was initially applied to the fields of image and natural language processing, but in recent years, many scholars have also applied GAN to data enhancement in the field of signal processing.

[0004] At present, the wind turbine variable pitch bearing data enhancement still has the following problems: first, due to the high nonlinearity of the vibration signal, it is difficult to generate high-dimensional vibration signals using a generator composed of a standard fully connected network; second, due to the small amount of learning samples, the "mode collapse" phenomenon is prone to occur. SUMMARY

[0005] In order to solve the problems of the prior art, through data enhancement, the balance of the data is improved to overcome the situation of difficulty in generating high-dimensional signals and mode collapse, and the purpose of improving the fault diagnosis accuracy of the generator variable pitch bearing is achieved. The present application adopts the following technical solutions:

[0006] The wind turbine variable pitch bearing data enhancement method based on GAN comprises the following steps:

[0007] Step S1: Collecting vibration signals of a fault and a healthy wind turbine variable pitch bearing;

[0008] Step S2: Obtaining effective vibration signals;

[0009] Step S3: Performing noise reduction processing on the collected effective vibration signals;

[0010] Step S4: Taking the noise-reduced vibration signals as a training set, constructing an autoencoder and training it for health data and fault data of the variable pitch bearing, to obtain real vibration signals;

[0011] Step S5: Constructing a generative adversarial neural network, including a generator and a discriminator, inputting a set of encoders of the generator with Gaussian noise, training the generator to generate feature vectors conforming to the feature distribution of the vibration signals, and then using the decoder of the generator to decode the feature vectors to generate false vibration signals, the decoder of the generator being consistent with the decoder of the trained autoencoder; the set of encoders of the generator being consistent with the encoders of the autoencoder, and the discriminator including a plurality of one-dimensional convolutional layers and fully connected layers;

[0012] Step S6: Filling the false vibration signals into the real vibration signal dataset, filling the false signals generated in accordance with the real signal distribution into the unbalanced real signal dataset, and reconstructing a vibration signal dataset with balanced data volume.

[0013] In the step S1, the vibration signals are collected on the hub on the outer ring bearing load area or the inner ring bearing load area side of the variable pitch bearing, the pitch angle, the pitch rate, and the sampling frequency are set, the vibration signals are collected at the pitch rate from 0° pitch to the pitch angle, and then from the pitch angle back to 0° as one cycle, and the vibration signals are collected based on the sampling frequency.

[0014] In the step S2, the collected vibration signals are segmented, the root mean square value of the signals in each equal part is calculated, the rising and falling edges of the root mean square value are determined to determine the start-up and shutdown intervals, the vibration signals before start-up and after shutdown are removed, and the effective vibration signals are obtained.

[0015] The step S3 includes the following steps:

[0016] Step S31: Using wavelet decomposition to decompose the original effective vibration signals into low-frequency approximation signals and high-frequency detail signals;

[0017] Step S32: Calculating a threshold value for the high-frequency detail signals of each layer;

[0018] Step S33: Inhibiting high-frequency detail components through soft threshold processing;

[0019] Step S34: using wavelet reconstruction to reconstruct the low frequency approximation and the suppressed high frequency details into the denoised vibration signal.

[0020] In the step S32, the threshold value is calculated by the following method:

[0021]

[0022] wherein T represents the threshold value, m represents the sequence length, and x represents the high frequency detail component.

[0023] In the step S33, the soft threshold value is processed as follows:

[0024]

[0025] wherein x represents the high frequency detail component, and x' represents the suppressed high frequency detail component.

[0026] In the step S4, the denoised vibration signal is divided into healthy vibration signals and fault vibration signals according to the health status of the bearing, and is labeled; a convolutional autoencoder is constructed for the healthy vibration signals and the fault vibration signals, the convolutional autoencoder includes an encoder and a decoder, the encoder includes a set of convolutional layers, the decoder includes a set of deconvolutional layers corresponding to the encoder, the mean square error is used as the error function, and the denoised vibration signal is used to train the autoencoder.

[0027] In the step S5, the training process of the generative adversarial neural network includes the following steps:

[0028] Step S51: inputting Gaussian noise into the encoder of a set of generators to obtain a set of feature vectors;

[0029] Step S52: using the decoder of the generator to reconstruct the set of feature vectors into a noise reconstructed vibration signal;

[0030] Step S53: inputting the noise reconstructed vibration signal as a false vibration signal and the corresponding real vibration signal into the discriminator to calculate the discriminator loss, and performing error back propagation to optimize the discriminator parameters;

[0031] Step S54: calculating the generator error, and performing error back propagation to optimize the generator parameters;

[0032] Step S55: iterating the steps S51 to S54 until the network converges.

[0033] In the step S53, the discriminator loss function is:

[0034]

[0035] wherein D represents the discriminator function, represents a real vibration signal, represents a noise reconstructed vibration signal, m represents the number of encoders of the generator;

[0036] The multi-head generator is a plurality of convolutional networks independent of each other, wherein the loss function corresponding to the i-th generator is:

[0037]

[0038] Wherein, D represents a discriminator function, represents the feature extracted by the encoder from the real vibration signal, represents the feature generated by the generator after the Gaussian noise is encoded by the i-th generator.

[0039] In the step S6, the reconstruction of the data set, the original vibration signal data set has X fault data and Y healthy data, if the fault data is less than the healthy data, then the generative adversarial network is used to generate Y-X false fault vibration signals to fill into the original vibration signal data set, and vice versa, X-Y false healthy vibration signals are generated to fill into the original vibration signal data set.

[0040] The wind turbine variable pitch bearing data enhancement device based on GAN includes a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the wind turbine variable pitch bearing data enhancement method based on GAN.

[0041] The advantages and beneficial effects of the present application are:

[0042] (1) The present application uses a convolutional autoencoder to extract the features of the vibration signal, and then uses a generative adversarial neural network to learn the features, thereby reducing the dimensionality of the learning and improving the convergence performance of the algorithm.

[0043] (2) The present application uses multiple generators, and different loss values are used for each generator, which can reduce the probability of "mode collapse" to a certain extent.

[0044] (3) The present application has universality, and since the physical characteristics of the signal are not specifically processed, the present application can be applied to other signal enhancement of wind turbines. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a comparison diagram of each step of the method in the embodiment of the present application and the corresponding principle.

[0046] Figure 2 is a vibration acceleration sensor point position diagram in the embodiment of the present application.

[0047] Figure 3 This is a framework diagram of the convolutional autoencoder model in an embodiment of the present invention.

[0048] Figure 4 This is a schematic diagram of the generative adversarial neural network structure in an embodiment of the present invention.

[0049] Figure 5 This is a time-domain diagram of the fault vibration signal generated in an embodiment of the present invention.

[0050] Figure 6 This is a schematic diagram of the device in an embodiment of the present invention. Detailed Implementation

[0051] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0052] like Figure 1 As shown, the GAN-based data augmentation method for wind turbine pitch bearings includes the following steps:

[0053] Step S1: Use a vibration signal acquisition system to collect vibration data of the pitch bearing of the wind turbine generator for both faulty and healthy turbines.

[0054] Vibration data is collected on the hub of the pitch bearing, either on the outer ring bearing area or the inner ring bearing area. The pitch angle, pitch rate, and sampling frequency are set. One cycle is defined as the pitch angle changing from 0° to the pitch rate and then back to 0°. Vibration signals are collected based on the sampling frequency. The vibration signal acquisition system includes a vibration acceleration sensor, a data acquisition card, and a host computer.

[0055] In embodiments of the present invention, such as Figure 2 As shown, the pitch bearing includes an outer ring 1 and a bearing bearing support area 3. A single-axis vibration acceleration sensor is arranged on the bearing bearing support area 3 or on the hub 2 near the bearing bearing support area of ​​the inner ring. The sampling frequency is set to 2560Hz. The blades are locked and the pitch speed is adjusted. The pitch angle is set to 0-90°. The vibration signal is obtained by continuously pitching at a rate of 2° / s. Each measurement cycle is a pitch change from 0° to 90° at a specific rate and then back to 0°.

[0056] Step S2: Remove invalid vibration data before power-on and after power-off, and obtain valid vibration data;

[0057] The collected vibration signal is segmented, and the root mean square value of the signal in each segment is calculated. The start-up and stop intervals are determined based on the rising and falling edges of the root mean square value. Vibration data before start-up and after stop are removed to obtain valid vibration data.

[0058] In the embodiment of the present application, the vibration signal is divided into 200 parts, each part containing 1000 data points, the RMS (Root Mean Square) value of the signal in each part interval is calculated respectively, since the RMS values before starting and after stopping are close to 0, the interval after the rising edge of the RMS value is used as the starting time, and the interval before the falling edge of the RMS value is used as the stopping time, and the invalid vibration signal before the starting time and after the stopping time is discarded.

[0059] Step S3: performing noise reduction processing on the collected vibration data, including the following steps:

[0060] Step S31: using wavelet decomposition to decompose the original effective vibration signal into low-frequency approximation and high-frequency detail, in the embodiment of the present application, the wavelet function is selected as Haar wavelet, and the original effective vibration signal is decomposed for 4 times in total.

[0061] Step S32: calculating the threshold value of the high-frequency detail part of each layer; the calculation method of the threshold value is as follows:

[0062]

[0063] Wherein, T represents the threshold value, m represents the sequence length, and x represents the high-frequency detail component.

[0064] Step S33: suppressing the high-frequency detail component through soft threshold processing, that is:

[0065]

[0066] Wherein, x represents the high-frequency detail component, and x' represents the suppressed high-frequency detail component.

[0067] Step S34: using wavelet reconstruction to reconstruct the low-frequency approximation and the suppressed high-frequency detail into the noise-reduced signal.

[0068] Step S4: taking the noise-reduced vibration data as a training set to perform feature extraction.

[0069] The noise-reduced vibration signal is divided into healthy vibration data and fault vibration data according to the health status of the bearing, and is labeled; a convolutional autoencoder is constructed for the healthy data and the fault data, the convolutional autoencoder includes an encoder and a decoder, the encoder includes a group of convolutional layers, the decoder includes a group of deconvolutional layers corresponding to the encoder, and the mean square error is used as an error function to train the autoencoder.

[0070] In the embodiment of the present application, the noise-reduced vibration data is packaged into a training set to perform feature extraction on the vibration data. The vibration signal is divided into healthy vibration data and fault vibration data according to the health status of the bearing, and is labeled; a convolutional autoencoder is constructed for the healthy data and the fault data, respectively, as follows:Figure 3 As shown, the encoder part of the autoencoder is composed of 5 convolutional layers, the step of the convolutional layers is {5, 5, 5, 4, 2} respectively, the number of filters is {8, 16, 32, 128, 128} respectively, and the decoder part adopts the deconvolutional layer, the parameters correspond to the encoder. After the input layer InputLayer input, the encoder is convolved through 5 one-dimensional convolutional layers Conv1D, and then the output of the encoder is input into the corresponding 5 one-dimensional deconvolutional layers Conv1DTranspose of the decoder, and then the tensor flattening layer is used for flattening operation. The input of the autoencoder is a vibration signal with a dimension of [10000, 1], and the output is a restored signal with a dimension of [10000, 1]. The mean square error is used as the error function to train the autoencoder.

[0071] Step S5: Constructing a generative adversarial neural network, inputting Gaussian noise as input, training the generator to generate feature vectors conforming to the feature distribution of the vibration signal, and then using the decoder to decode the feature vectors to generate false vibration signals.

[0072] The GAN network as shown in Figure 4 The network includes three modules, which are a trained convolutional autoencoder, a generator based on a convolutional layer, and a discriminator based on a convolutional layer. The generator has a total of 5 encoders for generating different vibration signals. The training process of the generative adversarial network includes the following steps:

[0073] Step S51: input Gaussian noise into the 5 generators, and the encoder of the ith generator obtains a feature vector { };

[0074] Step S52: use the decoder to reconstruct the feature vectors { } of the encoders of the 5 generators into vibration signals { };

[0075] Step S53: input the m real vibration signals { } and the noise reconstructed vibration signals { } into the discriminator, calculate the discriminator loss, and perform error back propagation to optimize the discriminator parameters;

[0076] The discriminator loss function is:

[0077]

[0078] where D represents the discriminator function, is the real vibration data, is the noise reconstructed vibration data.

[0079] The loss function of the ith generator is:

[0080]

[0081] wherein D is a discriminator function, is a feature extracted from the real vibration signal by the encoder, is a feature generated by the i-th generator from the Gaussian noise.

[0082] Step S54: Calculate the generator error, and back-propagate the error to optimize the generator parameters.

[0083] Step S55: Iterate the steps S51 to S54 until the network converges.

[0084] Step S6: Fill the false vibration signal into the real vibration signal dataset, fill the false signal generated according to the distribution of the real signal into the unbalanced real signal dataset, and reconstruct the vibration signal dataset with balanced data quantity.

[0085] In the original vibration signal dataset, there are X fault data and Y health data, if the fault data is less than the health data, then generate Y-X false fault vibration signals using the generative adversarial network, and fill them into the original vibration signal dataset, otherwise generate X-Y false health vibration signals and fill them into the original vibration signal dataset.

[0086] In the embodiment of the application, 240 sets of health data and 120 sets of fault data are used, except for the validation set and the test set, the training set contains 200 sets of health data and 80 sets of fault data, 120 sets of Gaussian noise signals are generated using Gaussian distribution with mean 0 and variance 0.1, the generated fault signals are as shown in Figure 5 The generative adversarial neural network trained using the fault data is used to generate 120 sets of fault data to fill into the original dataset, and the updated dataset contains 240 sets of health data and 240 sets of fault data, forming a balanced dataset.

[0087] Corresponding to the foregoing embodiment of the GAN-based wind turbine variable pitch bearing data enhancement method, the application also provides an embodiment of a GAN-based wind turbine variable pitch bearing data enhancement device.

[0088] Referring to Figure 6 The GAN-based wind turbine variable pitch bearing data enhancement device provided in the embodiment of the application includes a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the GAN-based wind turbine variable pitch bearing data enhancement method in the foregoing embodiments.

[0089] The embodiment of the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be implemented by software, hardware or a combination of software and hardware. For example, as a logical device, the computer program instructions corresponding to the non-volatile memory are read into the memory by the processor of the device and run to form. From the hardware level, as shown in Figure 6 The embodiment of the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be implemented by software, hardware or a combination of software and hardware. For example, as a logical device, the computer program instructions corresponding to the non-volatile memory are read into the memory by the processor of the device and run to form. From the hardware level, as shown in Figure 6 In addition to the processor, memory, network interface and non-volatile memory shown in the above, the device in the embodiment can also include other hardware according to the actual function of the device with data processing capability, which will not be described here.

[0090] The implementation process of the functions and roles of each unit in the above device is described in detail in the implementation process of the corresponding steps in the above method, which will not be described here.

[0091] For the device embodiment, it basically corresponds to the method embodiment, so the relevant part is described in the method embodiment. The above described device embodiment is only schematic, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. According to the actual needs, part or all of the modules can be selected to achieve the purpose of the application. Those skilled in the art can understand and implement without creative labor.

[0092] The embodiment of the application further provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the GAN-based wind turbine variable pitch bearing data enhancement method in the above embodiment.

[0093] The computer readable storage medium can be an internal storage unit of any of the aforementioned data processing capable devices, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any of the aforementioned data processing capable devices, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc. Further, the computer readable storage medium can include both an internal storage unit and an external storage device of any of the aforementioned data processing capable devices. The computer readable storage medium is used to store the computer program and other programs and data required by the aforementioned data processing capable devices, and can also be used to temporarily store data that has been output or will be output.

[0094] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A GAN-based method for data augmentation of wind turbine pitch bearings, characterized in that... Includes the following steps: Step S1: Collect vibration signals of the pitch bearing of the wind turbine generator in both faulty and healthy states; Step S2: Acquire valid vibration signals; Step S3: Perform noise reduction processing on the collected valid vibration signals; Step S4: Using the noise-reduced vibration signal as a training set, construct an autoencoder based on the health and fault data of the pitch bearing and train it to obtain the real vibration signal. The noise-reduced vibration signals are divided into healthy vibration signals and fault vibration signals according to the bearing's health condition, and then labeled. A convolutional autoencoder is constructed for the healthy vibration signals and the fault vibration signals. The convolutional autoencoder includes an encoder and a decoder. The encoder includes a set of convolutional layers, and the decoder includes a set of deconvolutional layers corresponding to the encoder. The mean square error is used as the error function, and the noise-reduced vibration signals are used to train the autoencoder. Step S5: Construct a multi-generator generative adversarial neural network, including a multi-head generator and a discriminator. Noise is input into a set of encoders in the generator to train it to generate feature vectors that conform to the characteristic distribution of vibration signals. Then, the generator's decoder is used to decode the feature vectors and generate fake vibration signals. The generator's decoder is consistent with the decoder of the trained autoencoder. The training process of the generative adversarial neural network includes the following steps: Step S51: Input noise into the encoder of a set of generators to obtain a set of feature vectors; Step S52: Using the decoder of the generator, reconstruct a set of feature vectors into a noise-reconstructed vibration signal; Step S53: The noise-reconstructed vibration signal is used as a spurious vibration signal, and input into the discriminator along with the corresponding real vibration signal. The discriminator loss is calculated, and the error is backpropagated to optimize the discriminator parameters. The discriminator loss function is: Where D represents the discriminator function. Represents the actual vibration signal, This represents the noise-reconstructed vibration signal, where m represents the number of encoders in the generator, and i represents the i-th generator; The multi-head generator consists of multiple independent convolutional networks, where the loss function for the i-th generator is: Where D represents the discriminator function, This represents the features extracted from the actual vibration signal by the encoder. This represents the feature generated by the generator after the noise is encoded by the i-th generator; Step S54: Calculate the generator error, and backpropagate the error to optimize the generator parameters; Step S55: Iterate through steps S51 to S54 until the network converges; Step S6: Fill the real vibration signal dataset with the fake vibration signals to reconstruct a vibration signal dataset with balanced data volume.

2. The method for data augmentation of wind turbine pitch bearings based on GAN according to claim 1, characterized in that: In step S1, vibration signals are collected on the hub (2) on the outer ring bearing area (3) or the inner ring bearing area of ​​the pitch bearing. The pitch angle, pitch rate and sampling frequency are set. The pitch rate is adjusted from 0° to pitch angle and then back to 0° as one cycle. The vibration signals are collected based on the sampling frequency.

3. The method for data augmentation of wind turbine pitch bearings based on GAN according to claim 1, characterized in that: In step S2, the collected vibration signal is segmented, the root mean square value of the signal in each segment is calculated, the start-up and stop intervals are determined based on the rising and falling edges of the root mean square value, and the vibration signals before start-up and after stop are removed to obtain the effective vibration signal.

4. The method for data augmentation of wind turbine pitch bearings based on GAN according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Use wavelet decomposition to decompose the original effective vibration signal into a low-frequency approximate signal and a high-frequency detail signal; Step S32: Calculate the threshold for the high-frequency detail signal of each layer; Step S33: Suppress high-frequency detail components through soft thresholding; Step S34: Use wavelet reconstruction to reconstruct the low-frequency approximation and suppressed high-frequency details into a noise-reduced vibration signal.

5. The GAN-based data augmentation method for wind turbine pitch bearings according to claim 4, characterized in that: In step S32, the threshold is calculated as follows: Where T represents the threshold, m represents the sequence length, and x represents the high-frequency detail component; In step S33, the soft threshold processing is as follows: Where x represents the high-frequency detail component, and x' represents the suppressed high-frequency detail component.

6. The method for data augmentation of wind turbine pitch bearings based on GAN according to claim 1, characterized in that: In step S6, the original vibration signal dataset contains X fault data points and Y healthy data points. If there are fewer fault data points than healthy data points, a generative adversarial network is used to generate YX false fault vibration signals to fill the original vibration signal dataset. Conversely, XY false healthy vibration signals are generated to fill the original vibration signal dataset.

7. A GAN-based data augmentation device for wind turbine pitch bearings, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the GAN-based wind turbine pitch bearing data augmentation method according to any one of claims 1-6.

Images