A method for diagnosing bearing faults in wind turbine generators

Abstract

This invention relates to the field of wind turbine fault diagnosis technology, specifically to a method for diagnosing wind turbine bearing faults. The method constructs source domain data and target domain data using pre-collected wind turbine bearing vibration data. An initial classification network is obtained by training an AlexNet network using the source domain data, and then adjusted using the target domain data to obtain the final classification network. Real-time collected generator bearing vibration data is input into the classification network to obtain the fault category. This method reduces the training time and parameter calculations for the target domain in the AlexNet network, facilitating rapid training and use of the fault model. The local normalization in the AlexNet network is modified to weight normalization, reducing the computational load of network parameters. Furthermore, the Softsign activation function is introduced into the AlexNet network, improving the convergence speed of wind turbine fault comparison results and resolving overfitting, thereby improving fault identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine fault diagnosis, and in particular to a wind turbine bearing fault diagnosis method. BACKGROUND

[0002] The generator bearing of a doubly-fed wind turbine generator set has been running in variable working conditions and electromagnetic environment for a long time, and the bearing failure is usually caused by overload, improper lubrication, and electric corrosion. In the traditional bearing fault diagnosis method, the collected signals are determined and processed by mathematical methods to obtain the characteristics of the fault.

[0003] However, in the actual wind power generation process, the wind turbine is subjected to unbalanced wind force, and the generator bearing has the operating characteristics of variable speed, which greatly interferes with feature extraction. Moreover, in the actual working condition, the frequency of generator bearing failure in wind turbine failure is not high, and there is less data that can represent the failure of the generator bearing. In the process of algorithm learning, overfitting phenomenon is easy to occur, which leads to low recognition accuracy. In order to solve the problem of low bearing fault recognition accuracy under small sample data, the current methods include data enhancement algorithm through order tracking, hybrid self-attention prototype network, and gradient penalty distance adversarial network.

[0004] However, the above methods are all through data enhancement methods to obtain sufficient data to improve the form of fault characteristics and ensure the performance of the model. However, data enhancement can also lead to uneven data quality, and the regularization parameter in the network cannot be adjusted according to the size of the data amount, which increases the workload in the training process and leads to slow network training speed. SUMMARY

[0005] The present application aims to provide a wind turbine bearing fault diagnosis method, which solves the problem of slow model training speed and low fault recognition accuracy in the existing wind turbine fault diagnosis process.

[0006] To achieve the above-mentioned purpose, the present application provides a wind turbine bearing fault diagnosis method, comprising the following steps:

[0007] The pre-collected wind turbine bearing vibration data is converted to obtain a fault signal, and the fault signal set is divided into source domain data and target domain data;

[0008] After the source domain data is trained on the AlexNet network as a whole, an initial classification network is obtained;

[0009] The initial classification network is adjusted in parameters by the target domain data to obtain a classification network;

[0010] The real-time collected generator bearing vibration data is input into the classification network after signal processing to obtain a fault category.

[0011] Preferably, the pre-collected wind turbine bearing vibration data is subjected to data conversion to obtain a fault signal, and the fault signal set is divided into source domain data and target domain data, and the method further comprises:

[0012] The pre-collected wind turbine bearing vibration data is subjected to singular value decomposition to obtain a denoising signal.

[0013] The denoising signal is subjected to wavelet transform to obtain a fault signal.

[0014] Preferably, after the source domain data is used to train the AlexNet network as a whole to obtain an initial classification network, the method further comprises:

[0015] The local normalization calculation in the local normalization layer of the Alexnet network is adjusted to weight normalization calculation to obtain a weight normalization layer.

[0016] The Relu activation function in the Alexnet network is adjusted to a Relu-Softsign fusion function.

[0017] Preferably, the weight normalization calculation comprises the following formula:

[0018]

[0019] wherein ||V|| represents the Euclidean norm, W is a weight vector, is a unit vector, and g is a vector module.

[0020] The Relu-Softsign fusion function comprises the following formula:

[0021]

[0022] wherein x is the X-axis coordinate value in the function; a and b are softsign function constants, and s(x) is the rate of change in the softsign function.

[0023] Preferably, after the source domain data is used to train the AlexNet network as a whole to obtain an initial classification network, the method further comprises:

[0024] The source domain data is input into the AlexNet network for training to obtain optimal network parameters.

[0025] According to the optimal network parameters, the convolution layer, the pooling layer and the weight normalization layer in the AlexNet network are fixed in parameters to obtain an initial classification network.

[0026] Preferably, the initial classification network is adjusted using the target domain data to obtain a classification network. The method further includes:

[0027] The target domain data is divided into target domain training samples and target domain test samples;

[0028] Adjust the parameters of the fully connected layer, the prediction score layer, and the loss function layer in the initial classification network based on the training samples of the target domain;

[0029] The adjusted fully connected layer, prediction score layer, and loss function layer are tested using test samples from the target domain. If the result is a good fit, the classification network is obtained; otherwise, the parameters are repeatedly adjusted until the result is a good fit.

[0030] This invention discloses a method for diagnosing wind turbine bearing faults. It constructs source domain data and target domain data using pre-collected wind turbine bearing vibration data. An initial classification network is obtained by training an AlexNet network using the source domain data. This initial classification network is then adjusted using the target domain data to obtain a final classification network. Real-time collected generator bearing vibration data is input into the classification network to obtain the fault category. By utilizing transfer learning theory, the training time and parameter calculations for the target domain in the AlexNet network are reduced, facilitating rapid training and use of the fault model. Furthermore, this invention modifies the local normalization in the AlexNet network to weight normalization, reducing the computational load of network parameters. The introduction of the Softsign activation function into the AlexNet network improves the convergence speed of wind turbine fault comparison results and addresses overfitting, thereby increasing fault identification accuracy. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0032] Figure 1 This is a flowchart illustrating the steps of the wind turbine bearing fault diagnosis method of the present invention.

[0033] Figure 2 This is a schematic diagram of the transfer learning task of the present invention.

[0034] Figure 3 This is a schematic diagram illustrating the features of the Softsign activation function of the present invention.

[0035] Figure 4 This is a schematic diagram of the features corresponding to the Relu-Softsign function of the present invention.

[0036] Figure 5This is a schematic diagram of the structure of the improved AlexNet network of the present invention.

[0037] Figure 6 This is a schematic diagram of the transfer model of the AlexNet network of the present invention. Detailed Implementation

[0038] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0039] Please see Figures 1 to 6 ,in, Figure 1 This is a flowchart illustrating the steps of the wind turbine bearing fault diagnosis method of the present invention. Figure 2 This is a schematic diagram of the transfer learning task of the present invention. Figure 3 This is a schematic diagram illustrating the features of the Softsign activation function of the present invention. Figure 4 This is a schematic diagram of the features corresponding to the Relu-Softsign function of the present invention. Figure 5 This is a schematic diagram of the structure of the improved AlexNet network of the present invention. Figure 6 This is a schematic diagram of the transfer model of the AlexNet network of the present invention.

[0040] This invention provides a method for diagnosing bearing faults in wind turbine generators, comprising the following steps:

[0041] S101: Convert the pre-collected wind turbine bearing vibration data to obtain fault signals, and divide the fault signal set into source domain data and target domain data.

[0042] Specifically, the pre-collected wind turbine bearing vibration data is decomposed into singular values ​​to obtain a noise-reduced signal; then the noise-reduced signal is transformed by wavelet transform to obtain a fault signal, and the fault signal set is divided into source domain data and target domain data.

[0043] S102: After training the AlexNet network with the source domain data as a whole, an initial classification network is obtained.

[0044] Specifically, after training the AlexNet network with source domain data to obtain the initial classification network, the process includes: adjusting the local normalization calculation in the local normalization layer of the AlexNet network to the weight normalization calculation to obtain the weight normalization layer; and adjusting the ReLU activation function in the AlexNet network to the ReLU-Softsign fusion function.

[0045] The weight normalization calculation is shown in the following formula:

[0046]

[0047] Where ||V|| represents the Euclidean norm, and W is the weight vector. Let g be a unit vector, and g be the magnitude of the vector.

[0048] The Relu-Softsign fusion function is shown in the following equation:

[0049]

[0050] Where x is the x-axis coordinate value of the function; a and b are the constants of the softsign function, and s(x) is the rate of change in the softsign function;

[0051] The source domain data is input into the AlexNet network for training to obtain the optimal network parameters. Based on the optimal network parameters, the convolutional layers, pooling layers, and weights in the AlexNet network are grouped into a single layer to fix the parameters, resulting in the initial classification network.

[0052] S103: Adjust the parameters of the initial classification network using the target domain data to obtain the classification network.

[0053] Specifically, step one: divide the target domain data into target domain training samples and target domain test samples; step two: adjust the parameters of the fully connected layer, prediction score layer, and loss function layer in the initial classification network according to the target domain training samples; step three: test the adjusted fully connected layer, prediction score layer, and loss function layer with the target domain test samples. If the result fits, the classification network is obtained; otherwise, continue to repeat step two until the result fits.

[0054] S104: The real-time collected generator bearing vibration data is processed and then input into the classification network to obtain the fault category.

[0055] Specifically, the real-time collected generator bearing vibration data is processed and then input into the classification network to calculate the fault category. Through the theory of transfer learning, the training time and parameter calculation of the target domain in the AlexNet network are reduced, which is beneficial to the rapid training and use of the fault model. In addition, the local normalization in the AlexNet network is modified to weight normalization, which reduces the amount of network parameter calculation. The activation function in the AlexNet network is introduced into the Softsign activation function, which improves the convergence speed of the wind turbine fault comparison results and solves the problem of overfitting.

[0056] Example 1:

[0057] This embodiment provides a wind turbine bearing fault diagnosis method based on deep transfer learning, including: converting pre-collected wind turbine bearing vibration data into fault signals, and dividing the fault signal set into source domain data and target domain data; training an AlexNet network on the source domain data to obtain an initial classification network, and adjusting the parameters of the initial classification network using the target domain data to obtain a final classification network; and inputting real-time collected generator bearing vibration data into the classification network after signal processing to obtain the fault category.

[0058] The step of converting the pre-collected wind turbine bearing vibration data into a fault signal includes: performing singular value decomposition on the pre-collected wind turbine bearing vibration data to obtain a noise-reduced signal; and performing wavelet transform on the noise-reduced signal to obtain the fault signal.

[0059] The process of training the AlexNet network with source domain data to obtain the initial classification network includes, before: adjusting the local normalization calculation in the local normalization layer of the AlexNet network to the weight normalization calculation to obtain the weight normalization layer; and adjusting the ReLU activation function in the AlexNet network to the ReLU-Softsign fusion function.

[0060] The weight normalization calculation is shown in the following formula:

[0061]

[0062] Where ||V|| represents the Euclidean norm, and W is the weight vector. Let g be a unit vector, and g be the magnitude of the vector.

[0063] The Relu-Softsign fusion function is shown in the following equation:

[0064]

[0065] Where x is the X-axis coordinate value of the function; a and b are constants of the softsign function, and s(x) is the rate of change in the softsign function.

[0066] The step of training the AlexNet network with source domain data to obtain an initial classification network includes: inputting source domain data into the AlexNet network for training to obtain optimal network parameters; and fixing the parameters of the convolutional layers, pooling layers, and weights in the AlexNet network by grouping them into a single layer based on the optimal network parameters to obtain the initial classification network.

[0067] The step of adjusting the parameters of the initial classification network using the target domain data to obtain the classification network includes:

[0068] Step 1: Divide the target domain data into target domain training samples and target domain test samples;

[0069] Step 2: Adjust the parameters of the fully connected layer, prediction score layer, and loss function layer in the initial classification network based on the target domain training samples;

[0070] Step 3: Test the adjusted fully connected layer, prediction score layer, and loss function layer with test samples from the target domain. If the result fits, the classification network is obtained; otherwise, repeat Step 2 until the result fits.

[0071] Example 2:

[0072] This embodiment describes the application of transfer learning used in the invention on the AlexNet network.

[0073] Transfer learning refers to the influence of one learning experience on another. It can transfer knowledge not only across different learning tasks but also between distribution models with similar features. In transfer learning, data is divided into a source domain and a target domain, originating from different working environments. It leverages the feature similarity between the source and target domains to transfer the relevant parameters and structure of a model trained in the source domain to the target domain, reducing training time and parameter computation in the target domain.

[0074] Transfer learning has a weak ability to extract features from wind turbine data, and deep learning has a weak ability to transform data imbalance problems. Therefore, a model that complements deep learning and transfer learning should be established to improve the feature extraction and transformation capabilities of wind turbine fault data.

[0075] First, the pre-collected wind turbine data is used to train a training model A. When a new training task B is encountered, the training can be completed using only the relevant parameters of the training model A and a small number of samples from the domain of B. If necessary, the relevant parameters of task B need to be fine-tuned.

[0076] Deep learning typically requires that the training and test sets have identical data distributions and be of sufficient quantity, thus consuming a significant amount of time for training from scratch. Transfer learning, on the other hand, involves transferring knowledge from a pre-trained model to two similar sets of data, saving training time. The principle of transfer learning is as follows.

[0077] First, define the source domain D1:

[0078]

[0079] At the same time, define an unlabeled target domain D2:

[0080]

[0081] Where x represents the input, y represents the label, and n and m represent the number of data samples, a classifier is trained from the source domain to predict the label in the target domain. Assuming their feature spaces and class spaces are the same but their feature distributions differ, a classification model is trained using data from the source domain, and this model is then transferred to the target domain to perform classification. However, not all domain data can be used for transfer learning training; it requires a certain error range. The learning error is expressed as:

[0082]

[0083] H represents a hypothetical class, and sup represents the minimum upper bound of the function. The formula expresses the class error between two datasets. If it is a symmetric class, it can be determined using the following formula:

[0084]

[0085] Here, I represents the indicator function; the smaller the calculated error, the stronger the transferability. Therefore, a design such as... Figure 2 The transfer learning task shown is continuously optimized to obtain the final transfer fault diagnosis model.

[0086] Furthermore, this invention also improves the AlexNet network model by normalizing the weights.

[0087] Traditional AlexNet networks employ local normalization, but this can easily lead to local optima. Therefore, weight normalization is proposed as a replacement for local normalization. The formula for weight normalization is shown below.

[0088]

[0089] Where ||V|| represents the Euclidean norm, and W is the weight vector. Let g be a unit vector, and g be the vector magnitude.

[0090] To verify the effectiveness of weight normalization, three classic datasets were used for recognition analysis. The results after weight normalization were compared with those after local normalization, as shown in Table 1. The recognition results after adding the weight normalization layer are significantly better than those after adding the local normalization layer. The function of the pooling layer is to reduce the feature dimension extracted by the convolutional layer. In this invention, the weight normalization layer is placed after the pooling layer, which can reduce the computational cost of network parameters. Adding a convolutional layer after the weight normalization layer can increase the number of neurons, avoiding insufficient computational cost of neurons due to the reduction of parameters, thus avoiding affecting the network's recognition accuracy.

[0091] Table 1 Comparative Analysis of Classic Dataset Recognition

[0092]

[0093] In this invention, the activation function of the AlexNet network is also improved. Considering the properties of the activation function itself and the characteristics of the function derivative in the AlexNet network, the derivative of the first half of the activation function should be large, so as to make the network more effective in parameter optimization. The derivative value of the second half should be gradually reduced to 0, so as to facilitate parameter adjustment.

[0094] AlexNet networks commonly use the ReLU function, a linear function. When the function value is not less than 0, its slope is 1, avoiding the problem of gradient vanishing on the positive x-axis. However, its derivative is always 0 on the negative x-axis, meaning the output is always 0 during data processing, which can easily lead to overfitting. The Softsign function has the advantage of fast convergence and a non-zero derivative on the negative x-axis, which helps solve the gradient vanishing problem on the negative x-axis. However, it does not meet the requirement of a sufficiently large derivative in the first half of the activation function when processing the positive x-axis. Therefore, combining these two activation functions can achieve complementary advantages and improve the network's recognition accuracy.

[0095] By combining the ReLU and Softsign functions piecewise along the positive and negative x-axis, we obtain the combined activation function ReLU-Softsign. This function has a constant derivative of 1 along the positive x-axis, which accelerates convergence, and its derivative gradually decreases to 0 along the negative x-axis, enhancing the network's stability in processing data along the negative x-axis. The corresponding characteristics of the Softsign function are as follows: Figure 3 As shown, the Relu-Softsign function has the following characteristics: Figure 4 As shown, the Relu-Softsign function is represented by the following equation.

[0096]

[0097] Where x is the X-axis coordinate value of the function; a and b are constants of the softsign function, and s(x) is the rate of change in the softsign function.

[0098] Therefore, the local normalization layer is replaced by the weight normalization layer, and the weight normalization layer is placed after the pooling layer.

[0099] The activation function in the activation layer is changed to the ReLU-Softsign activation function. The improved AlexNet network structure is shown below. Figure 5 As shown. The transfer model is constructed by combining the model transfer method with an improved AlexNet network. The entire structure is divided into retention layers and replacement layers, as shown below. Figure 6As shown, the proposed improved AlexNet network is first trained using source domain data with sufficient data volume. Through transfer learning, the parameters of the trained convolutional layers, pooling layers, and weight normalization layers are fixed. Then, the parameters of the last three layers of the network are fine-tuned for the target domain data to achieve the classification objective.

[0100] The main steps of the proposed method are as follows:

[0101] Step 1: After denoising the pre-collected wind turbine fault vibration data, perform wavelet transform, label the time-frequency graphs according to type, and divide them into training samples and test samples;

[0102] Step 2: Input the source domain data into the network for training and output the optimal network parameters;

[0103] Step 3: Fix the parameters of the convolutional layer, pooling layer, and weight normalization layer, and fine-tune the parameters of the last three layers according to the training samples of the target domain to complete the construction of the target domain classification network.

[0104] Step 4: Test the target domain classification network using target domain test samples to complete the construction of the fault diagnosis model.

[0105] Step 5: Input the real-time collected wind turbine fault vibration data into the fault section model, and obtain the fault category through fitting and comparison.

[0106] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art will understand that all or part of the processes for implementing the above embodiments and equivalent variations made in accordance with the claims of this application are still within the scope of this application.

Claims

1. A wind turbine generator bearing fault diagnosis method characterized by, The method comprises the following steps: The wind turbine bearing vibration data collected in advance is subjected to data conversion to obtain a fault signal, and the fault signal set is divided into source domain data and target domain data; The local normalization calculation in the local normalization layer of the Alexnet network is adjusted to weight normalization calculation to obtain a weight normalization layer; The Relu activation function in the Alexnet network is adjusted to a Relu-Softsign fusion function, wherein the weight value normalization calculation comprises the following formula: (1), wherein, represents the Euclidean norm, is a weight vector, is a unit vector, is a vector module; the Relu-Softsign fusion function comprises the following formula: (2), wherein, is the X-axis coordinate value in the function; , is a softsign function constant, is a change rate in the softsign function; After the source domain data is used to train the AlexNet network as a whole, an initial classification network is obtained, wherein the source domain data is input into the AlexNet network for training to obtain optimal network parameters; according to the optimal network parameters, the convolution layer, the pooling layer and the weight normalization layer in the AlexNet network are parameter-fixed to obtain the initial classification network; The target domain data is used to adjust the parameters of the initial classification network to obtain a classification network; The real-time collected generator bearing vibration data is subjected to signal processing and then input into the classification network to obtain a fault category.

2. The wind turbine generator bearing fault diagnostic method of claim 1, wherein, The wind turbine bearing vibration data collected in advance is subjected to data conversion to obtain a fault signal, and the fault signal set is divided into source domain data and target domain data, and the method further comprises: The wind turbine bearing vibration data collected in advance is subjected to singular value decomposition to obtain a denoising signal; The denoising signal is subjected to wavelet transform to obtain a fault signal.

3. The wind turbine generator bearing fault diagnostic method of claim 1, wherein, The target domain data is used to adjust the parameters of the initial classification network to obtain a classification network, and the method further comprises: The target domain data is divided into target domain training samples and target domain test samples; The parameters of the full connection layer, the prediction score layer and the loss function layer in the initial classification network are adjusted according to the target domain training samples; The adjusted full connection layer, the prediction score layer and the loss function layer are tested by the target domain test samples, and if the result is fitted, the classification network is obtained; otherwise, the parameters are repeatedly adjusted until the result is fitted.

Images