Wind turbine generator gearbox fault diagnosis method and system based on time-frequency feature cross complementation and multi-domain feature fusion

Abstract

The invention discloses a wind turbine generator gearbox fault diagnosis method and system based on time-frequency feature cross complementation and multi-domain feature fusion. Collected original vibration signals are divided into a training set, a verification set and a test set; respectively utilizing FFT and CWT to obtain a one-dimensional frequency domain signal and a two-dimensional time-frequency graph of the data set; inputting the training set of the time domain signal and the frequency domain signal into a time-frequency cross attention mechanism network to obtain cross complementation characteristics of the time domain and the frequency domain; inputting the training set of the two-dimensional time-frequency graph into a two-dimensional residual CNN into which a complete attention mechanism is introduced to extract time-frequency domain features; fusing the cross complementation features of the time domain and the frequency domain and the time-frequency domain features to obtain multi-domain features, and performing fault classification by using a softmax function; and inputting test set data into the trained multi-domain feature fusion network model for testing to obtain a final fault diagnosis result. According to the method, the problem of incomplete feature extraction of deep learning fault diagnosis based on a single time domain signal and a frequency domain signal is solved, and the fault diagnosis precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent fault diagnosis of wind turbines, and in particular to a method and system for wind turbine gearbox fault diagnosis based on cross-complementarity of time-frequency features and fusion of multi-domain features. Background Art

[0002] Wind energy is a clean and renewable energy source that can be converted into electrical energy, thermal energy or mechanical energy through wind turbines. Gearboxes are an important part of wind turbines. Due to their long-term operation under fluctuating load conditions, they have a high failure rate, long downtime and serious consequences. At the same time, once a wind turbine gearbox fails, it may affect the normal and stable operation of the power system and even cause serious accidents, resulting in economic losses and casualties. Therefore, wind turbine gearbox fault diagnosis is the focus of wind turbine maintenance and operation.

[0003] At present, wind turbine gearbox fault diagnosis methods can be divided into: traditional fault diagnosis methods and intelligent fault diagnosis methods. The former is mainly achieved through algorithms such as artificial neural networks, gray models and support vector machines or combined with signal processing methods, and the expected results can be obtained under a certain amount of training sample data. In addition, it is highly dependent on expert experience in signal processing, has weak adaptive ability, and low diagnostic accuracy. Therefore, the diagnostic ability and generalization ability based on traditional fault diagnosis methods are limited in engineering. With the rapid development of computer science and technology, traditional fault diagnosis methods are gradually replaced by artificial intelligence-based methods, and deep learning models such as deep belief networks (DBNs), convolutional neural networks (CNNs), stacked autoencoders (SAEs), and long short-term memory networks (LSTMs) have become the mainstream research direction. These deep learning models can use a large amount of data to train models, suppress overfitting, and make up for the shortcomings of traditional machine learning. At the same time, the end-to-end characteristics of deep learning are used to avoid complex feature extraction signal processing processes.

[0004] However, these studies are all based on a single one-dimensional convolutional neural network or a two-dimensional convolutional neural network to achieve fault diagnosis, or they all perform fault diagnosis from the perspective of time domain, frequency domain or time-frequency domain alone, ignoring the complementary relationship between time domain, frequency domain and time-frequency domain. If the original signal is only converted into a two-dimensional image and input into CNN, the time correlation information will be lost. If only the original vibration signal is input into the deep learning model, the frequency domain features or time-frequency domain features will be ignored, resulting in incomplete feature extraction and the accuracy of fault diagnosis needs to be further improved. Summary of the invention

[0005] In order to solve the above technical problems and ensure the safe and reliable operation of wind turbines, the present invention combines the advantages of one-dimensional and two-dimensional convolution, and considers the complementary relationship between time domain, frequency domain and time-frequency domain features, and proposes a wind turbine gearbox fault diagnosis method and system based on cross-complementarity of time-frequency features and fusion of multi-domain features. It overcomes the problem of incomplete feature extraction based on single time domain signal and frequency domain signal deep learning fault diagnosis, and improves the accuracy of fault diagnosis.

[0006] The technical solution adopted by the present invention is:

[0007] The wind turbine gearbox fault diagnosis method based on cross-complementation of time-frequency features and fusion of multi-domain features includes the following steps:

[0008] Step 1: Divide the collected original vibration signal into training set, verification set and test set according to a certain length;

[0009] Step 2: Use fast Fourier transform (FFT) and continuous wavelet transform (CWT) to obtain the one-dimensional frequency domain signal and two-dimensional time-frequency diagram of the data set respectively;

[0010] Step 3: Input the original one-dimensional time domain signal and one-dimensional frequency domain signal training set into the time-frequency cross attention mechanism network to obtain the cross-complementary features of the time domain and frequency domain; input the two-dimensional time-frequency graph training set into the two-dimensional residual CNN with the full attention mechanism to extract the time-frequency domain features;

[0011] Step 4: Fuse the cross-complementary features of the time domain and frequency domain obtained in step 3 with the time-frequency domain features to obtain multi-domain features, and use the softmax function for fault classification;

[0012] Step 5: Input the test set data into the trained multi-domain feature fusion network model for testing to obtain the final fault diagnosis result.

[0013] In step 1, the original vibration signal of each health state is taken as a time domain signal sample with a length of 1024, and the sliding window is 1024, that is, there is no overlapping part between two adjacent samples, and the training set, verification set and test set are divided according to the total number of samples obtained in a ratio of 7:2:1.

[0014] In step 2, the time domain signal samples of the training set, the validation set and the test set are respectively subjected to FFT and CWT to obtain frequency domain signals and time-frequency diagrams;

[0015] The FFT process is expressed as:

[0016]

[0017] In the formula, X(k) represents the kth frequency component in the frequency domain; x(n) is the original time domain signal of the nth sampling point; N is the total sampling rate of the signal; is a complex exponential kernel representing the Fourier basis function, where j is the imaginary unit.

[0018] The CWT process is expressed as:

[0019]

[0020] In the formula, CWT x (a,τ) represents the time-frequency eigenvalue after scale transformation and translation, x(t) is the original vibration signal, τ is the translation parameter, a is the scale factor, is the wavelet mother function; Represents a wavelet mother function with scale a and translation τ. The wavelet mother function is stretched or compressed by scale a and translated by τ.

[0021] In step 3, in order to obtain cross-complementary features in the time domain and the frequency domain, a cross-attention mechanism is introduced between the CNNs that extract the time domain features and the frequency domain features;

[0022] The time-frequency cross attention mechanism network consists of a two-branch one-dimensional convolutional neural network and a cross attention mechanism. Figure 4 The dotted box shows the structure diagram of the time-frequency cross-attention mechanism network proposed in the present invention, which is used to extract cross-complementary features in the time domain and frequency domain.

[0023] Among them, the purpose of the cross-attention mechanism is to use the self-attention mechanism of the Transformer model to interactively capture the time domain features of the original vibration signal and the frequency domain feature information after FFT, so as to achieve effective feature extraction and fusion, thereby improving the fault classification effect;

[0024] The time-frequency cross attention mechanism network takes the one-dimensional signals of the time domain signal and the frequency domain signal as input respectively, first passes through a wide convolution block to extract effective features under the expanded receptive field and reduce the size of the input sequence, and then extracts the time domain and frequency domain features through three convolution blocks;

[0025] The structures of the wide convolutional block and the convolutional block are the same, both of which consist of a convolutional layer with L filters, a batch normalization BN layer, a leaky rectified linear unit LReLU and a maximum pooling layer with a kernel size of 2×1, except that the convolution kernel size of the wide convolutional block is different, which is set to 64×1, and the convolution kernel size of other convolutional blocks is 3×1; specifically, Figure 4As shown in the figure, in the two-branch one-dimensional convolutional neural network in the time-frequency cross attention mechanism network, each branch has four convolutional layers. The first convolutional layer is a wide convolutional block with a convolution kernel of 64×1, which aims to expand the receptive field of the model to extract features and extract effective features of long sequences. The convolution kernel size of the next three convolutional blocks is 3×1 to further extract deep features in the time domain and frequency domain.

[0026] Among them, the BN layer is used to alleviate the internal covariate shift and accelerate the convergence of the model, and the leaky rectified linear unit LReLU with a negative slope of 0.001 is used to activate the normalized feature map from the BN layer;

[0027] The max pooling layer is a downsampling operation in convolutional neural networks that aims to reduce the size of feature maps while retaining their important features.

[0028] The step 3 comprises the following steps:

[0029] Step 3.1: For a given time domain feature T f and frequency domain feature F f , the cross-attention mechanism first uses a feature embedding module with linear projection to obtain T q 、T k 、T v and F q 、F k 、F v ;

[0030] T q 、T k 、T v They respectively represent the features extracted from the time domain signal by the query vector Q, key vector K, and value vector V of the self-attention mechanism of the Transformer model in the cross-attention mechanism;

[0031] F q 、F k 、F v They respectively represent the features extracted from the frequency domain signal by the query vector Q, key vector K, and value vector V in the self-attention mechanism of the Transformer model in the cross-attention mechanism.

[0032] These tensors can be represented as:

[0033]

[0034] Where: W represents the learnable linear weight, ⊙ represents the element-by-element multiplication operation; They represent the parameter matrices of the linear mapping formed by the query vector Q, key vector K, and value vector V when extracting time domain features, They respectively represent the parameter matrices of the linear mapping formed by the query vector Q, the key vector K, and the value vector V when extracting frequency domain features.

[0035] Step 3.2: Substitute the T obtained in step 3.1 q 、T k 、T v and F q 、F k 、F v , the attention matrix CA is calculated by the cross attention function 1 and CA 2 , the expression is as follows:

[0036]

[0037] Where D Tk Indicates T k The dimension of Indicates F k The dimension of represents matrix multiplication operation, and the softmax function is applied to obtain the normalized weights in the interval [0,1]; T k represents the features extracted from the time domain signal by the key vector K in the self-attention mechanism of the Transformer model, F k Represents the features extracted from the frequency domain signal by the key vector K in the self-attention mechanism of the Transformer model, with CA 1 For example, the softmax function is used to calculate the relative importance of the time domain features by aggregating the scores of the frequency domain features corresponding to each attribute of the time domain features. As a result, an attention vector is obtained, which can be used to highlight more relevant attributes of the frequency domain features.

[0038] Step 3.3: Then, a nonlinear transformation block is used to calculate the output features T out and F out , the nonlinear transformation block consists of two linear layers, a normalization layer and a nonlinear activation function. The expression formula is:

[0039]

[0040] Where f(·) represents the function of nonlinear transformation, T out and F out Has the same shape as the corresponding input feature. The nonlinear activation function uses Gaussian Error Linear Unit (GELU) as the activation function instead of LReLU, because the classifier using the GELU activation function is more robust to noisy input. The GELU function can be defined by the following formula:

[0041]

[0042] Where: x is the input, erf(·) is the Gaussian error function, which can be expressed as: Step 3.4: Attention feature T out and F out It is added to the output of the second convolutional block in the corresponding branch and passes through the following two convolutional blocks to generate the final feature.

[0043] Although the self-attention mechanism is good at capturing global dependencies, it may not be as effective as CNN in capturing local features. Especially when processing images, the self-attention mechanism may need to be combined with a convolutional network to better capture local features. In step 3, the full attention mechanism aims to better capture the global and local features of the input features. The designed full attention mechanism consists of three parts:

[0044] (1) Improve the self-attention mechanism: First, the feature vector X of the original signal is converted to [x 1 ,x 2 ,…x n ]∈R C×N , x 1 ,x 2 ,…x n Respectively represent the first sample, the second sample...the nth sample. C×N Represents the real number matrix of C rows and N columns, where N is the matrix dimension and C is the number of channels; it is input into the convolutional network, and the hidden state of each input data feature vector is calculated. The result of the hidden state is transformed into the feature space f, g and h.

[0045] f(x)=W f (x);

[0046] g(x)=W g (x);

[0047] h(x)=W h (x);

[0048] In the formula, and are all parameter matrices of linear mapping; f(x) represents the output value of the input data x after the convolution layer f, W f (x) represents the parameter matrix of the linear mapping of the convolutional layer f, g(x) represents the output value of the input data x after the convolutional layer g, W g (x) represents the parameter matrix of the linear mapping of the convolution layer g, h(x) represents the output value of the input data x after the convolution layer h, W h (x) represents the parameter matrix of the linear mapping of the convolutional layer h.

[0049] Furthermore, the traditional 1×1 convolution kernel filtering of the three convolutional networks in the above steps is extended to k×k.

[0050] (2) Apply the softmax function to each row of the product of f(x) and g(x) to generate the attention matrix, that is, the feature vector of attention in the input signal, which is expressed as:

[0051] s ij =f(x i ) T g(x j )

[0052]

[0053] Among them, x i and x j Respectively represent the i-th and j-th pixel features of the input feature map; s ij is the unnormalized attention distribution, α ij is the final attention distribution, which is used as the attention distribution for each input data x i The weight of the attention level. f(x i ) T Represents the input pixel feature x i The transpose of the output matrix after the convolution layer f; g(x j ) represents the input pixel feature x j The output value after the convolution layer g; exp(s ij ) represents the natural index calculation of the unnormalized attention distribution features.

[0054] (3) In order to fuse the output feature map obtained by traditional convolution and the features extracted by the improved self-attention mechanism, a direct weighted connection method is selected to adaptively allocate the weights of the two parts. The direct weighted formula is as follows:

[0055] f(R i ,Y i )=αR i +(1-α)Y i

[0056] Among them, f(R i ,Y i ) represents the feature output value of the full attention mechanism; R i It represents the feature output of the improved self-attention mechanism; Y i Represents the output features obtained by traditional convolution. i is the index of the feature, Y is the output of the lightweight channel attention mechanism, R is the output of the self-attention mechanism, α is a trainable parameter, and α∈(0,1), the weights of the two parts of the feature are 1-α and α respectively.

[0057] The second aspect of the present invention is a wind turbine gearbox fault diagnosis system based on cross-complementation of time-frequency features and fusion of multi-domain features, comprising: an acquisition module, a data processing module, a model training module, a real-time fault detection module, and a fault alarm module;

[0058] The acquisition module uses a vibration sensor to collect vibration signals on the gearbox of the wind turbine;

[0059] The data processing module affixes corresponding labels to the data samples according to the state type of the wind turbine gearbox, divides the data set according to a certain ratio, and then performs FFT transformation and continuous wavelet transformation on the data set respectively, and divides the obtained one-dimensional time domain signal, one-dimensional spectrum signal and two-dimensional wavelet time-frequency diagram into a training set, a test set and a validation set;

[0060] The model training module constructs a multi-domain feature fusion network model, which mainly includes three parts: a time-frequency cross-attention mechanism network for extracting cross-complementary features in the time domain and frequency domain; a two-dimensional residual CNN that introduces a full attention mechanism for extracting time-frequency domain features; and fusing and flattening the cross-complementary features in the time domain and frequency domain and the time-frequency domain features, and then classifying through a linear layer to obtain the final fault diagnosis result.

[0061] Then, the multi-domain feature fusion network model is trained using the training set and tested using the validation set. After the training is completed, the multi-domain feature fusion network training model is obtained;

[0062] The real-time fault detection module uses a multi-domain feature fusion network training model to complete real-time fault detection of the wind turbine gearbox;

[0063] The fault alarm module quickly sends out an alarm signal when the real-time fault detection module detects a fault on the wind turbine gearbox, and uses voice or display to broadcast the fault type of the wind turbine gearbox. When the wind turbine gearbox operates normally, the fault alarm module automatically enters a silent state.

[0064] The present invention provides a wind turbine gearbox fault diagnosis method and system based on cross-complementation of time-frequency features and fusion of multi-domain features, and the technical effects are as follows:

[0065] 1) The present invention designs a time-frequency cross-attention mechanism network that realizes the cross-complementarity of time domain features and frequency domain features. The time-frequency cross-attention mechanism network takes time domain signals and frequency domain signals as inputs, and improves the fault diagnosis accuracy by cross-complementarity of time domain features and frequency domain features, thus overcoming the problem of incomplete feature extraction of fault diagnosis based on deep learning of single time domain signals and frequency domain signals. At the same time, the fault diagnosis accuracy is further improved through the complementary fusion between time domain features and frequency domain features.

[0066] 2) Starting from the time domain, frequency domain and time-frequency domain characteristics of the signal, the present invention combines the advantages of one-dimensional and two-dimensional convolution, and takes into account the complementary fusion of the characteristics of each domain, a wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and fusion of multi-domain features is proposed, which overcomes the problem of single fault diagnosis from the perspective of time domain, frequency domain or time-frequency domain of the original signal, and ignoring the complementary relationship between time domain, frequency domain and time-frequency domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0068] Figure 1 This is a specific flow chart of the wind turbine gearbox diagnosis method of the present invention.

[0069] Figure 2 This is the network structure diagram of the time-frequency cross-attention mechanism proposed in the present invention.

[0070] Figure 3 This is a structural diagram of the full attention mechanism proposed in the present invention.

[0071] Figure 4 This is the overall network structure diagram of the multi-domain feature fusion network model proposed in the present invention.

[0072] Figure 5 Schematic diagram of a test bench according to a preferred embodiment of the present invention.

[0073] Figure 6 This is a confusion matrix diagram of test data under three different working conditions for a preferred embodiment of the present invention.

[0074] Figure 7 This is a T-SNE feature visualization diagram of test data under different working conditions in a preferred embodiment of the present invention.

[0075] Figure 8 This is a bar chart of the accuracy of test data under variable working conditions according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0076] Embodiment 1:

[0077] The present invention combines the advantages of one-dimensional and two-dimensional convolution, and considers the complementary relationship between time domain, frequency domain and time-frequency domain features, and proposes a wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and fusion of multi-domain features. The method first obtains the frequency domain and time-frequency graphs of the original vibration signal through fast Fourier transform (FFT) and continuous wavelet transform (CWT), and then inputs the time domain and frequency domain signals into the cross-attention mechanism fusion network to obtain the interactive fusion features of the time domain and frequency domain. At the same time, the time-frequency graph is input into the two-dimensional residual CNN with the introduction of the full attention mechanism to extract the time-frequency domain features. Finally, the cross-complementary features are fused with the time-frequency domain features to obtain multi-domain features, and the softmax function is used for classification to obtain the final fault diagnosis result.

[0078] The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion of the present invention has a flow chart as shown in FIG. Figure 1 As shown, the specific steps include:

[0079] S1: The collected original vibration signal is divided into training set, validation set and test set according to a certain length;

[0080] S2: Use FFT and CWT to obtain the one-dimensional frequency domain signal and two-dimensional time-frequency diagram of the data set respectively;

[0081] S3: Input the original one-dimensional time domain signal and one-dimensional frequency domain signal training set into the time-frequency cross attention mechanism network to obtain the cross-complementary features of the time domain and frequency domain;

[0082] S4: At the same time, the training set of two-dimensional time-frequency images is input into the two-dimensional residual CNN with the introduction of the full attention mechanism to extract time-frequency domain features;

[0083] S5: The cross-complementary features obtained in S3 and S4 are fused with the time-frequency domain features to obtain multi-domain features, and the softmax function is used for fault classification;

[0084] S6: Input the test set data into the trained multi-domain feature fusion network model for testing to obtain the final fault diagnosis result.

[0085] Furthermore, in S1, the original vibration signal of each health state is taken as a time domain signal sample with a length of 1024, and the sliding window is 1024, that is, there is no overlapping part between two adjacent samples, and the training set, verification set and test set are divided according to the total number of samples obtained in 7:2:1.

[0086] Further, in S2, the time domain signal samples of the training set, the validation set and the test set are respectively subjected to FFT and CWT to obtain frequency domain signals and time-frequency diagrams;

[0087] The above FFT process is expressed as:

[0088]

[0089] In the formula, X(k) represents the kth frequency component in the frequency domain; x(n) is the original time domain signal of the nth sampling point; N is the total sampling rate of the signal; is a complex exponential kernel representing the Fourier basis function, where j is the imaginary unit.

[0090] The above CWT process is expressed as:

[0091]

[0092] In the formula, x(t) is the original vibration signal, τ is the translation parameter, a is the scale factor, is the wavelet mother function.

[0093] Furthermore, the S3 can be divided into the following six steps:

[0094] S3.1: Time-frequency cross-attention mechanism network It consists of two branches of one-dimensional convolutional neural network and a cross-attention mechanism. The structure of the cross-attention mechanism is as follows Figure 2 As shown in the figure, the purpose is to use the self-attention mechanism of the Transformer model to interactively capture the time domain features of the original vibration signal and the frequency domain feature information after FFT, so as to achieve effective feature extraction and fusion, thereby improving the fault classification effect.

[0095] S3.2: Furthermore, the time-frequency cross attention mechanism network takes the one-dimensional signals of the time domain signal and the frequency domain signal as input respectively, first passes through a wide convolution block to extract effective features under the expanded receptive field and reduce the size of the input sequence, and then extracts the time domain and frequency domain features by three convolution blocks.

[0096] S3.3: Further, the structures of the wide convolutional blocks and the convolutional blocks are the same, as shown in Table 1, and are composed of a convolutional layer with L filters, a batch normalization (BN) layer, a leaky rectified linear unit (LReLU), and a maximum pooling layer with a kernel size of 2×1, except that the convolution kernel size of the wide convolution is different, set to 64×1, and the other convolutional blocks are 3×1. Among them, the BN layer is used to mitigate the internal covariate shift and accelerate the model convergence, and the LReLU with a negative slope of 0.001 is used to activate the normalized feature map from the BN layer.

[0097] Table 1 Parameters of the time-frequency cross attention network

[0098]

[0099] S3.4: For a given time domain feature T f and frequency domain feature F f , the criss-cross attention mechanism first uses an embedding block with linear projection to obtain T q 、T k 、T v and F q 、F k 、F v , these tensors can be represented as:

[0100]

[0101] Where W represents the learnable linear weights and ⊙ represents the element-wise multiplication operation.

[0102] S3.5: T obtained in S3.4 q 、T k 、T v and F q 、F k 、F v , the attention matrix CA is calculated by the cross attention function 1 and CA 2 , the expression is as follows:

[0103]

[0104] In the formula, Indicates T k The dimension of Indicates F k The dimension of represents the matrix multiplication operation, and the softmax function is applied to obtain the normalized weights in the interval [0,1]. 1 For example, the softmax function is used to calculate the relative importance of the time domain features by aggregating the scores of the frequency domain features corresponding to each attribute of the time domain features. As a result, an attention vector is obtained, which can be used to highlight more relevant attributes of the frequency domain features.

[0105] Furthermore, a nonlinear transformation block is used to calculate the output feature T out and F out , the nonlinear transformation block consists of two linear layers, a normalization layer and a nonlinear activation function. The expression formula is:

[0106]

[0107] Where f(·) represents the function of nonlinear transformation, T out and F outHas the same shape as the corresponding input feature. The nonlinear activation function uses Gaussian Error Linear Unit (GELU) as the activation function instead of LReLU, because the classifier using the GELU activation function is more robust to noisy input. The GELU function can be defined by the following formula:

[0108]

[0109] Where x is the input and erf(·) is the Gaussian error function, which can be expressed as

[0110] S3.6: Attention feature T out and F out Add it to the output of the second convolution block in the corresponding branch, and then pass it through two convolution blocks to obtain the time domain features and frequency domain features.

[0111] The network structure of the two-dimensional residual CNN used to extract time-frequency domain features in S4 is shown in Table 2.

[0112] Table 2 2D residual CNN network parameters

[0113]

[0114] Furthermore, the full attention mechanism in S4 is as follows Figure 3 As shown, the specific steps are as follows:

[0115] S4.1: First, the feature vector X of the original signal is 1 ,x 2 ,…x n ]∈R C×N Input into the convolutional network, calculate the hidden state of each input data feature vector, and the hidden state result is transformed into the feature space f, g and h.

[0116] f(x)=W f (x)

[0117] g(x)=W g (x)

[0118] h(x)=W h (x)

[0119] In the formula and are all parameter matrices of linear mapping. N is the matrix dimension and C is the number of channels.

[0120] Furthermore, unlike the traditional self-attention mechanism, the three convolutional networks in S4.1 are expanded from the traditional 1×1 convolution kernel filtering to k×k.

[0121] S4.2: Apply the softmax function to each row of the product of f(x) and g(x) to generate an attention matrix, that is, a feature vector of interest in the input signal, expressed as:

[0122] s ij =f(x i ) T g(x j )

[0123]

[0124] Among them, x i and x j Respectively represent the i-th and j-th pixel features of the input feature map. ij is the unnormalized attention distribution, α ij is the final attention distribution, which is used as the attention distribution for each input data x i The weight of the degree of attention.

[0125] S4.3: In order to fuse the output feature map obtained by traditional convolution and the features extracted by the improved self-attention mechanism, a direct weighted connection method is selected to adaptively allocate the weights of the two parts. The weighted connection formula is as follows:

[0126] f(R i ,Y i )=αR i +(1-α)Y i

[0127] Among them, i is the index of the feature, Y is the output of the lightweight channel attention mechanism, and R is the output of the self-attention mechanism. α is a trainable parameter, and α∈(0,1), the weights of the two parts of the feature are 1-α and α respectively.

[0128] S5: The time domain, frequency domain and time-frequency domain features of the extracted training set are fused into multi-domain features, and classified through linear layers and softmax activation functions to save the best training model. The overall structure of the multi-domain feature fusion network is as follows: Figure 4 As shown, specifically, the cross-complementary features of the time domain and frequency domain extracted in steps S3 and S4 are fused with the time-frequency domain features and then classified.

[0129] S6: Testing the divided test set with the best training model saved in S5 to obtain the fault diagnosis result of the method proposed in the present invention.

[0130] The following is a detailed description with reference to specific embodiments:

[0131] A preferred embodiment of the present invention is an application of a wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and fusion of multi-domain features on a gearbox dataset (WT) of Beijing University of Technology.

[0132] like Figure 5 As shown in the figure, the WT dataset is mainly composed of a motor, a planetary gearbox, a fixed-axis gearbox, and a load device. Among them, there are four planetary gears rotating around the sun gear, and the sun gear has five health states: broken teeth, healthy, missing teeth, root wear, and wear. Each health state simultaneously collects vibration signals in the x-axis and y-axis directions and encoder data of the input shaft of the planetary gearbox. Among them, the vibration signal is collected by the Sinocera CA-YD-1181 accelerometer, and the sampling frequency of all channels is 48kHz. In this dataset, each health state collects data under eight different speed conditions of 20Hz, 25Hz, 30Hz, 35Hz, 40Hz, 45Hz, 50Hz and 55Hz.

[0133] The present invention selects the vibration signal data of the x-axis direction under the conditions of 20Hz, 25Hz, and 30Hz as experimental data, and divides the signal length of 1024 in each healthy state into one sample under the three speed conditions, and the sliding window is also 1024, that is, there is no overlap between two adjacent signal samples. The experimental hyperparameters are set as follows: learning rate is 0.004, optimizer is Adam, epochs is 30, and batch size is 64. In order to reduce the training amount of the model, 500 samples of each category in the collected data set are selected as the total data set, and the sample size of the training set, verification set, and test set is evenly divided according to the ratio of 7:2:1. The specific data is shown in Table 3.

[0134] Table 3 WT dataset information

[0135]

[0136] In this embodiment, in order to more intuitively demonstrate the classification effect of the proposed method, T-SNE is used to perform dimensionality reduction visualization on the original samples and classification layers of the test set under three different speed conditions. Figure 6 As shown in the figure, before the training set is input into the constructed model, the samples of the five healthy categories are randomly distributed under the three speed conditions, and all categories are mixed together and cannot be distinguished well. After the model is trained, it can be seen from the dimensionality reduction visualization of the classification layer that each category can be distinguished well, and the same category is concentrated in one cluster without misclassification. Although the two categories of broken teeth and missing teeth are close to each other in the classification layer visualization under the speed condition of 20Hz, the classification boundary can be divided between the two categories to accurately distinguish different fault categories.

[0137] Furthermore, to verify the effectiveness of the proposed method, the model is trained and evaluated on the training set and validation set divided under three different working conditions, and then tested on the test set. The test results of the test set are visualized using the confusion matrix. The results are shown in Figure 2. Figure 7 The method proposed in the present invention has a high diagnostic accuracy for wind turbine gearbox fault diagnosis. Under different working conditions, all categories are accurately classified, which also reflects that the model of the present invention has good generalization ability.

[0138] Preferably, in order to verify the superiority of the proposed method, it is compared with related methods. Method 1 is a fault diagnosis method of one-dimensional CNN and original vibration signal, method 2 is a fault diagnosis method of one-dimensional CNN and original vibration signal obtained by FFT of frequency domain signal, method 3 is a fault diagnosis method of WDCNN and original vibration signal, method 4 is a fault diagnosis method of time-frequency cross attention mechanism network and original vibration signal and frequency domain signal obtained by FFT, method 5 is a fault diagnosis method of two-dimensional CNN and original vibration signal obtained by CWT of time-frequency graph, method 6 is a fault diagnosis method of two-dimensional CNN with full attention mechanism and time-frequency graph obtained by CWT, and method 7 is the method proposed in this paper. The optimizer, learning rate and batch size of all methods are kept consistent. Experiments were carried out on three different working conditions, all methods were trained 5 times, and the average test accuracy was calculated. The comparative experimental results are shown in Table 4.

[0139] Table 4 Comparative experimental results

[0140]

[0141] As shown in Table 4, the method proposed in the present invention achieves the highest diagnostic accuracy. Compared with the use of one-dimensional CNN alone, the fault diagnosis method combining one-dimensional convolution and two-dimensional convolution can improve the fault diagnosis accuracy. For example, under the speed condition of 20Hz, the diagnostic accuracy of the proposed method is 18.32% and 2.88% higher than that of method one and method two, respectively. At the same time, compared with the use of two-dimensional CNN alone, the diagnostic accuracy of the proposed method is 3.04% and 2.16% higher than that of method five and method six, respectively. Under the speed condition of 25Hz, the diagnostic accuracy of the proposed method is 16.80% and 3.84% higher than that of method one and method two, respectively, and 2.00% and 1.60% higher than that of method five and method six, respectively. Under the speed condition of 30Hz, the diagnostic accuracy of the proposed method is 18.36% and 3.92% higher than that of method one and method two, respectively, and 2.80% and 1.28% higher than that of method five and method six, respectively. In addition, the method combining time domain, frequency domain, and time-frequency domain has higher fault diagnosis accuracy than the fault diagnosis methods using time domain, frequency domain, and time-frequency domain separately. For example, the proposed method has higher diagnostic accuracy than method one, method two, and method five.

[0142] Furthermore, in order to verify the robustness and generalization ability of the proposed method, the proposed models of the seven methods were trained using the 20Hz and 25Hz training sets, and the 20Hz and 30Hz test sets were input into the trained models for testing. The test results are shown in Figure 8 As shown. Figure 8 It can be seen that the proposed method can achieve higher diagnostic accuracy than the comparison method even under variable operating conditions. Because the 30 Hz in the test set is an operating condition that does not appear in the training set, the data feature distribution is different, resulting in a slightly lower diagnostic accuracy than the single operating condition scenario.

[0143] Embodiment 2:

[0144] Secondly, this embodiment 2 provides a wind turbine gearbox fault diagnosis system based on cross-complementarity of time-frequency features and fusion of multi-domain features, including: an acquisition module, a data processing module, a model training module, a real-time fault detection module, and a fault alarm module.

[0145] The acquisition module uses a vibration sensor to collect vibration signals on the gearbox of the wind turbine;

[0146] The data processing module affixes corresponding labels to the data samples according to the state type of the gearbox, divides the data set according to a certain ratio, and then performs FFT transformation and continuous wavelet transformation on the data set to obtain a training set, a test set and a validation set consisting of a one-dimensional time domain, a one-dimensional spectrum signal and a two-dimensional wavelet time-frequency diagram;

[0147] The model training module constructs a multi-domain feature fusion network model, then uses the training set to train the parallel multi-channel convolutional neural network model, and uses the validation set to test the multi-domain feature fusion network model, and obtains the multi-domain feature fusion network training model after the training is completed;

[0148] The real-time fault detection module uses the multi-domain feature fusion network training model to complete the real-time fault detection of the wind turbine gearbox;

[0149] The fault alarm module quickly sends out an alarm signal when the above-mentioned real-time fault detection module detects a fault on the wind turbine gearbox, and uses voice or display to broadcast the fault type of the gearbox. When the wind turbine gearbox operates normally, the fault alarm module automatically enters a silent state.

[0150] The system is used to implement the above-mentioned wind turbine gearbox fault diagnosis method, so the specific implementation method of the system can refer to the implementation example of the above-mentioned wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and multi-domain feature fusion. For example, the acquisition module, the data processing module, the model training module, and the real-time fault detection module are respectively used to implement steps S1-S6 in the above-mentioned wind turbine gearbox fault diagnosis method, and the fault alarm module in the system displays the working status of the gearbox to the outside world in real time after executing the above-mentioned fault diagnosis process. The specific implementation method of the system can refer to the detailed description of the corresponding embodiments of each part of the above-mentioned wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and multi-domain feature fusion.

Claims

1. A wind turbine gearbox fault diagnosis method based on cross-complementarity of time-frequency features and fusion of multi-domain features, characterized by The following steps are involved: Step 1: Divide the collected original vibration signal into training set, verification set and test set according to a certain length; Step 2: Use fast Fourier transform (FFT) and continuous wavelet transform (CWT) to obtain the one-dimensional frequency domain signal and two-dimensional time-frequency diagram of the data set respectively; Step 3: Input the original one-dimensional time domain signal and one-dimensional frequency domain signal training set into the time-frequency cross attention mechanism network to obtain the cross-complementary features of the time domain and frequency domain; input the two-dimensional time-frequency graph training set into the two-dimensional residual CNN with the full attention mechanism to extract the time-frequency domain features; Step 4: Fuse the cross-complementary features of the time domain and frequency domain obtained in step 3 with the time-frequency domain features to obtain multi-domain features, and use the softmax function for fault classification; Step 5: Input the test set data into the trained multi-domain feature fusion network model for testing to obtain the final fault diagnosis result.

2. The wind turbine gearbox fault diagnosis method based on cross-complementation of time-frequency features and fusion of multi-domain features according to claim 1 is characterized in that: In step 1, the original vibration signal of each health state is taken as a time domain signal sample with a length of 1024, and the sliding window is 1024, that is, there is no overlapping part between two adjacent samples, and the training set, verification set and test set are divided according to the total number of samples obtained in a ratio of 7:2:

1.

3. The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 1 is characterized in that: In step 2, the time domain signal samples of the training set, the validation set and the test set are respectively subjected to FFT and CWT to obtain frequency domain signals and time-frequency diagrams; The FFT process is expressed as: In the formula, X(k) represents the kth frequency component in the frequency domain; x(n) is the original time domain signal of the nth sampling point; N is the total sampling rate of the signal; is the complex exponential kernel, representing the Fourier basis function, where j is the imaginary unit; The CWT process is expressed as: In the formula, CWT x (a,τ) represents the time-frequency eigenvalue after scale transformation and translation, x(t) is the original vibration signal, τ is the translation parameter, a is the scale factor, is the wavelet mother function; Represents a wavelet mother function with scale a and translation τ; the wavelet mother function is stretched or compressed by scale a and translated by τ.

4. The wind turbine gearbox fault diagnosis method based on cross-complementation of time-frequency features and fusion of multi-domain features according to claim 1 is characterized in that: In step 3, in order to obtain cross-complementary features in the time domain and the frequency domain, a cross-attention mechanism is introduced between the CNNs that extract the time domain features and the frequency domain features; Time-frequency cross-attention mechanism network It consists of a two-branch one-dimensional convolutional neural network and a cross-attention mechanism; it is used to extract cross-complementary features in the time domain and frequency domain; Among them, the cross-attention mechanism uses the self-attention mechanism of the Transformer model to interactively capture the time domain features of the original vibration signal and the frequency domain feature information after FFT; The time-frequency cross-attention mechanism network takes the one-dimensional signals of the time domain signal and the frequency domain signal as input respectively. It first passes through a wide convolution block to extract effective features under the expanded receptive field and reduce the size of the input sequence. Then, three convolution blocks are used to extract time domain and frequency domain features.

5. The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 4 is characterized in that: The structures of the wide convolution block and the convolution block are the same, both consisting of a convolution layer with L filters, a batch normalization BN layer, a leaky rectified linear unit LReLU and a maximum pooling layer with a kernel size of 2×1. The convolution kernel size of the wide convolution block is set to 64×1, and the convolution kernel size of other convolution blocks is 3×1.

6. The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 6 is characterized in that: In the two-branch one-dimensional convolutional neural network in the time-frequency cross attention mechanism network, each branch has four convolutional layers. The first convolutional layer is a wide convolutional block with a convolution kernel of 64×1, which aims to expand the receptive field of the model to extract features and extract effective features of long sequences. The convolution kernel size of the following three convolutional blocks is 3×1 to further extract deep features in the time domain and frequency domain. The BN layer is used to alleviate the internal covariate shift and accelerate the model convergence, while the leaky rectified linear unit LReLU with a negative slope of 0.001 is used to activate the normalized feature map from the BN layer.

7. The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 6 is characterized by: The step 3 comprises the following steps: Step 3.1: For a given time domain feature T f and frequency domain feature F f , the cross-attention mechanism first uses a feature embedding module with linear projection to obtain T q , T k , T v and F q 、F k 、F v ; T q , T k , T v They respectively represent the features extracted from the time domain signal by the query vector Q, key vector K, and value vector V of the self-attention mechanism of the Transformer model in the cross-attention mechanism; F q 、F k 、F v They respectively represent the features extracted from the frequency domain signal by the query vector Q, key vector K, and value vector V in the self-attention mechanism of the Transformer model in the cross-attention mechanism; These tensors are represented as: Among them: ⊙ represents element-by-element multiplication operation; They represent the parameter matrices of the linear mapping formed by the query vector Q, the key vector K, and the value vector V when extracting time domain features, Respectively represent the parameter matrices of the linear mapping formed by the query vector Q, the key vector K, and the value vector V when extracting frequency domain features; Step 3.2: Substitute the T obtained in step 3.1 q , T k , T v and F q 、F k 、F v , the attention matrices CA1 and CA2 are calculated by the cross attention function, and the expression is as follows: Where D Tk Indicates T k The dimension, D Fk Indicates F k The dimension of represents matrix multiplication operation, and the softmax function is applied to obtain the normalized weights in the interval [0,1]; T k represents the features extracted from the time domain signal by the key vector K in the self-attention mechanism of the Transformer model, F k The key vector K in the self-attention mechanism of the Transformer model represents the features extracted from the frequency domain signal. Taking CA1 as an example, the softmax function is used to aggregate the scores of the frequency domain features corresponding to each attribute of the time domain features to calculate the relative importance of the time domain features. As a result, an attention vector can be obtained, which can be used to highlight more relevant attributes of the frequency domain features. Step 3.3: Then, a nonlinear transformation block is used to calculate the output features T out and F out , the nonlinear transformation block consists of two linear layers, a normalization layer and a nonlinear activation function; the expression formula is: Where f(·) represents the function of nonlinear transformation, T out and F out It has the same shape as the corresponding input feature; the nonlinear activation function uses the Gaussian Error Linear Unit (GELU) as the activation function, and the GELU function is defined by the following formula: Where: x is the input, erf(·) is the Gaussian error function, expressed as: Step 3.4: Attention feature T out and F out It is added to the output of the second convolutional block in the corresponding branch and passes through the following two convolutional blocks to generate the final feature.

8. The wind turbine gearbox fault diagnosis method based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 7 is characterized in that: In step 3, the full attention mechanism consists of three parts: (1) Improve the self-attention mechanism: First, the feature vector X of the original signal is converted into [x1, x2, …x n ]∈R C×N x1,x2,…x n Respectively represent the first sample, the second sample...the nth sample; R C×N Represents the real number matrix of C rows and N columns, where N is the matrix dimension and C is the number of channels; it is input into the convolutional network, and the hidden state of each input data feature vector is calculated. The result of the hidden state is transformed into the feature space f, g and h; f(x)=W f (x); g(x)=W g (x); h(x)=W h (x); In the formula, and are all parameter matrices of linear mapping; f(x) represents the output value of the input data x after the convolution layer f, W f (x) represents the parameter matrix of the linear mapping of the convolutional layer f, g(x) represents the output value of the input data x after the convolutional layer g, W g (x) represents the parameter matrix of the linear mapping of the convolution layer g, h(x) represents the output value of the input data x after the convolution layer h, W h (x) represents the parameter matrix of the linear mapping of the convolutional layer h; Furthermore, the traditional 1×1 convolution kernel filtering of the above three convolutional networks is extended to k×k; (2) Apply the softmax function to each row of the product of f(x) and g(x) to generate the attention matrix, that is, the feature vector of attention in the input signal, which is expressed as: s ij =f(x i ) T g(x j ) Among them, x i and x j Respectively represent the i-th and j-th pixel features of the input feature map; s ij is the unnormalized attention distribution, α ij is the final attention distribution, which is used as the attention distribution for each input data x i The weight of the degree of attention; f(x i ) T Represents the input pixel feature x i The transpose of the output matrix after the convolution layer f; g(x j ) represents the input pixel feature x j The output value after the convolution layer g; exp(s ij ) represents the natural index calculation of the unnormalized attention distribution features; (3) In order to fuse the output feature map obtained by traditional convolution and the features extracted by the improved self-attention mechanism, a direct weighted connection method is selected to adaptively allocate the weights of the two parts. The direct weighted formula is as follows: f(R i ,Y i )=αR i +(1-α)Y i Among them, f(R i ,Y i ) represents the feature output value of the full attention mechanism; R i It represents the feature output of the improved self-attention mechanism; Y i Represents the output features obtained by traditional convolution; i is the index of the feature, α is a trainable parameter, and α∈(0,1), the weights of the two parts of the feature are 1-α and α respectively.

9. A wind turbine gearbox fault diagnosis system based on cross-complementation of time-frequency features and fusion of multi-domain features, characterized by: include: Acquisition module, data processing module, model training module, real-time fault detection module, fault alarm module; The acquisition module uses a vibration sensor to acquire vibration signals on the gearbox of the wind turbine; The data processing module affixes corresponding labels to the data samples according to the state type of the wind turbine gearbox, divides the data set according to a certain ratio, and then performs FFT transformation and continuous wavelet transformation on the data set respectively, and divides the obtained one-dimensional time domain signal, one-dimensional spectrum signal and two-dimensional wavelet time-frequency diagram into a training set, a test set and a validation set; The model training module constructs a multi-domain feature fusion network model. Then, the multi-domain feature fusion network model is trained using the training set and tested using the validation set. After the training is completed, the multi-domain feature fusion network training model is obtained; The real-time fault detection module uses a multi-domain feature fusion network training model to complete real-time fault detection of the wind turbine gearbox; The fault alarm module quickly sends out an alarm signal when the real-time fault detection module detects a fault on the wind turbine gearbox, and uses voice or display to broadcast the fault type of the wind turbine gearbox. When the wind turbine gearbox operates normally, the fault alarm module automatically enters a silent state.

10. The wind turbine gearbox fault diagnosis system based on time-frequency feature cross-complementation and multi-domain feature fusion according to claim 9 is characterized in that: The multi-domain feature fusion network model consists of three parts: a time-frequency cross-attention mechanism network for extracting cross-complementary features in the time domain and frequency domain; a two-dimensional residual CNN with a full attention mechanism for extracting time-frequency domain features; and a fusion and flattening of the cross-complementary features in the time domain and frequency domain with the time-frequency domain features, followed by classification through a linear layer to obtain the final fault diagnosis result.

Images