Wind turbine generator gearbox fault diagnosis method and system based on convolution two-channel multilayer perceptron mixer and weighted voting mechanism

Abstract

The invention discloses a wind turbine generator gear box fault diagnosis method and system based on a convolution two-channel multi-layer perceptron mixer and a weighted voting mechanism. Collected original vibration signals are preprocessed and divided into a training set, a verification set and a test set; obtaining a time-frequency image of the data set by using CWT; constructing a two-dimensional convolutional neural network 2DCNN, and extracting local features of the time-frequency image; segmenting and compressing the time-frequency image of which the local features are extracted into a patch vector form, inputting the patch vector form into a dual-channel multi-layer perceptron mixer network, and further extracting global features of the time-frequency image; weighting the extracted global features by using a weighted voting mechanism to obtain final features, and training a fault diagnosis model by using the final features; and inputting the test set into the trained fault diagnosis model to obtain a final fault diagnosis result. According to the method, the global features output by the dual-channel multi-layer perceptron mixer network are weighted by using a weighted voting mechanism, and different features of each channel are used for representation, so that the limitation of a single model is reduced, and the fault diagnosis precision and robustness are 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 diagnosing wind turbine gearbox faults based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism. Background Art

[0002] Faced with increasingly serious energy shortages, environmental pollution, and global climate change issues, wind energy has become one of the most attractive forms of power generation due to its cleanliness and renewable nature. However, wind turbines (WTs) are prone to failures due to frequent exposure to extreme and harsh operating environments, resulting in unexpected downtime and economic losses. Statistics show that among all the costs of wind power generation, operation and maintenance costs account for more than 30%, of which about 70% are due to unexpected failures. Among the different components of wind turbines, planetary gearboxes are considered to be one of the most vulnerable components, accounting for the largest proportion of downtime and maintenance costs. Therefore, condition monitoring and fault diagnosis of wind turbines are considered to be effective tools for detecting faults to reduce operation and maintenance costs.

[0003] Since vibration and acoustic emission signals are sensitive to mechanical faults, condition monitoring systems based on vibration and acoustic emission have been widely used in the field of condition monitoring and fault diagnosis. In order to monitor the health of wind turbines (WTs), the wind energy industry currently uses condition monitoring systems to collect a large amount of real-time data to diagnose gearbox faults. Due to the increasing amount of data collected from gearboxes, traditional fault diagnosis methods cannot effectively analyze large amounts of data and automatically provide accurate diagnosis results, which are mainly achieved through algorithms such as artificial neural networks (ANN) and support vector machines (SVM). In addition, traditional fault diagnosis methods have limited learning capabilities in complex signals and low fault diagnosis accuracy. Therefore, fault diagnosis methods based on deep learning have attracted more and more attention. With the rapid development of computer science, deep learning has attracted widespread attention from scholars in various fields due to its powerful feature learning ability and advantages in processing massive data, such as deep belief networks (DBN), convolutional neural networks (CNN) and recurrent neural networks (RNN). Among them, convolutional neural networks (CNN) are widely used in computer vision problems due to their high efficiency and effectiveness in the field of image processing.

[0004] However, when processing complex time-frequency images, convolutional neural networks (CNN) have the problem of incomplete feature extraction, resulting in low fault diagnosis accuracy. Researchers have developed a variety of convolutional neural network (CNN) models and architectures, such as ResNet, Inception, DenseNet, etc. Although the fault diagnosis accuracy has been partially improved, the diagnosis process takes a long time due to the increasing complexity of the model, and sometimes there are diagnostic errors, so the fault diagnosis accuracy 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 proposes a wind turbine gearbox fault diagnosis method and system based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism. The global features of the dual-channel multi-layer perceptron mixer network output are weighted using a weighted voting mechanism, and the different feature representations of each channel are used to reduce the limitations of a single model and improve the accuracy and robustness of fault diagnosis.

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

[0007] A wind turbine gearbox fault diagnosis method based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism includes the following steps:

[0008] Step 1: Preprocess the collected original vibration signal and divide it into training set, verification set and test set;

[0009] Step 2: Use CWT to obtain the time-frequency image of the data set;

[0010] Step 3: Construct a two-dimensional convolutional neural network 2DCNN to extract local features of the time-frequency image;

[0011] Step 4: The time-frequency image with local features extracted in step 3 is segmented and compressed into a patch vector, and input into a dual-channel multi-layer perceptron mixer network to further extract the global features of the time-frequency image;

[0012] Step 5: Use the weighted voting mechanism to weight the global features extracted in step 4 to obtain the final features, and use the final features to train the fault diagnosis model;

[0013] Step 6: Input the test set into the trained fault diagnosis model to obtain the final fault diagnosis result.

[0014] In the step 1, the original vibration signal is taken as a time domain signal sample with a length of 1024, the sliding window is 1024, and the overall time domain signal samples are divided into a training set, a validation set and a test set in a ratio of 7:2:1.

[0015] In step 2, the time domain signal samples of the training set, the validation set and the test set are converted into two-dimensional time-frequency images using CWT;

[0016] The above CWT process is expressed as:

[0017]

[0018] Where W(s,τ) represents the coefficient of wavelet transform, which represents the local characteristics of signal x(t) at scale s and time translation τ. <x(t),ψs,τ > represents the signal x(t) and the wavelet function ψ s,τ The inner product of the signal represents the similarity between the signal and the wavelet function under a given scale s and translation τ; x(t) is the vibration signal, s is the scale parameter, τ is the translation parameter, t represents the time variable, and represents the position of the signal x(t) on the time axis, ψ s,τ is the mother wavelet function, ψ * is the complex conjugate.

[0019] Two-dimensional time-frequency images such as Figure 4 As shown in FIG. 1 , the image converted from the original vibration signal by CWT is a two-dimensional time-frequency image. In step 3, in order to better extract the local features of the two-dimensional time-frequency image, a two-dimensional convolutional neural network 2DCNN is constructed according to the two-dimensional time-frequency image. The network structure of the two-dimensional convolutional neural network 2DCNN is as follows: Figure 5 As shown, the two-dimensional convolutional neural network 2DCNN is applied before extracting the global features of the image as a preliminary feature extraction of the two-dimensional time-frequency image to extract the local features of the two-dimensional time-frequency image.

[0020] The step 4 includes the following steps: Figure 6 As shown:

[0021] Step 4.1: Divide the local feature-extracted time-frequency image (H×W×C) into N1 non-overlapping image blocks, each of which is linearly projected with the same projection matrix to the required hidden dimension C; divide the local feature-extracted time-frequency image into N1 P×P×C image blocks, where the sequence length N1 = H×W / P 2 Then, each image block is converted into a one-dimensional vector, represented as P 2 C;

[0022] Step 4.2: Use a fully connected network to process the rearranged image blocks to reduce the number of parameters and further extract features;

[0023] Step 4.3: Build a dual-channel multilayer perceptron mixer network, such as Figure 2 As shown in the figure, it mainly includes MLP module and Mixer layer. The MLP module is a standard fully connected layer structure. Its function is to perform feature mapping on the input data in linear space and map the low-dimensional input to high-dimensional representation. The main function of the Mixer layer is to mix and update features. It contains two sublayers: token mixing layer and channel mixing layer.

[0024] The processed image blocks are respectively input into a dual-channel multi-layer perceptron mixer network model to extract the shared features of the gear signal blocks. The dual-channel multi-layer perceptron mixer network includes a fully connected layer and a mixing layer.

[0025] Step 4.4: After normalizing the extracted feature layer, the feature data is averaged and rearranged into 3 dimensions in the n dimension.

[0026] In step 4.2, the image sequence is fed into the fully connected layer (FC), and U neurons are used to reduce the dimension of f. The image sequence output after processing is recorded as f k ={f k i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +}, f k i,j represents the feature value of the i-th image block after passing through the k-th fully connected layer (FC) on the j-th channel; i represents the number of the image block. U represents the number of neurons in the fully connected layer, which represents the output dimension of the model in the fully connected layer. j represents the index of the channel dimension, which represents the j-th channel in each image block. N1 represents the total number of image blocks. N + Represents the set of positive integers.

[0027] In step 4.3, the dual-channel multi-layer perceptron mixer network model includes two types of layers, namely a token mixing layer and a channel mixing layer, and each layer includes a nonlinear activation function (GULE) and Dropout to suppress overfitting of the network.

[0028] The token hybrid MLP takes the time-frequency image as input in the form of columns, and uses two fully connected layers with 4×U and U neurons to deeply extract information from the time-frequency image.

[0029] The output image set after the first layer of token processing is denoted as f kλɑ1 ={f kλɑ1 i,j |0 <i≤4U,0<j≤N1,i∈N + , j∈N +}, f kλɑ1 i,j Indicates that the jth channel of the i-th image block is in the k-th λa1 The output feature value after the fully connected layer is processed. U represents the output dimension of the fully connected layer (FC). N1 represents the total number of image blocks.

[0030] The image set output after the second layer of token processing is denoted as f kλɑ2 ={f kλɑ2 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +},

[0031] f kλɑ2 i,jIndicates that the jth channel of the i-th image block is in the k-th λa2 Output feature values ​​after processing by the fully connected layer.

[0032] Channel-mixing MLP: MLP is used to mix the channel dimension of the image, which processes the information of all image blocks on the same channel to achieve cross-channel information interaction.

[0033] The processing of channel-mixed MLP specifically refers to inputting the channel information of the image block into the CNN neural network for feature extraction. The convolution kernel size used here is 1×1 and the step size is 1, in order to perform element-by-element linear transformation on the features of each channel, and further extract and fuse the features through two layers of convolution operations. The number of channels in the two layers of convolution is 4×Q and Q respectively, which means that the first layer expands the number of channels to 4 times the original number, and the second layer compresses it back to the original number of channels Q.

[0034] The image set output after the first layer of channel processing is denoted as f kλ ={f kλ i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +},

[0035] f kλ i,j Indicates that the jth channel of the i-th image block is in the k-th λ Output feature values ​​after processing by the fully connected layer.

[0036] The image set output after the second layer of channel processing is denoted as f kλ1 ={f kλ1 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +},

[0037] f kλ1 i,j Indicates that the jth channel of the i-th image block is in the k-th λ1 Output feature values ​​after processing by the fully connected layer.

[0038] Next, the input and output are added together using skip-connection and layer normalization.

[0039] Among them, the skip connection is a mechanism that allows the input to be directly transmitted to the subsequent layers, bypassing the intermediate layers. Through this mechanism, the input information can continue to propagate in the subsequent layers, thereby reducing the gradient disappearance and improving the performance of the model. For the input x of a certain layer and the corresponding neural network output F(x), the output using the skip connection is the sum of the input and output:

[0040] y=F(x)+x

[0041] Layer normalization is a regularization technique used to normalize the input data at each neural network layer. Suppose the input of a layer is h = [h1,h2,…,h n ],h i Represents the activation value of the i-th neuron, and layer normalization normalizes all activation values. The formula is:

[0042]

[0043] Where: is the mean activation value of this layer, is the standard deviation, is the normalized activation value.

[0044] Furthermore, the nonlinear activation function (GULE) is used to perform nonlinear transformation on the input signal to increase the nonlinear expression ability of the network. The second layer activation operation in the first conventional convolution module is described by a nonlinear activation function (GULE), and the specific formula of the GELU is:

[0045] GELU(f)=f×Φ(f)

[0046] Where f is the GULE function input; Φ(f) represents the Cumulative Distribution Function (CDF) of the standard normal distribution. For input f, the CDF of the standard normal distribution is:

[0047]

[0048] Where erf(x) is the error function, which is used to approximate the cumulative distribution of the Gaussian distribution.

[0049] Furthermore, the specific operation steps of the skip connection are shown in the formula:

[0050] U *,j =f *,j +W2σ(W1 LayerNorm(f) *,j ), for j=1…C,

[0051] Y i,* =f i,*+W4σ(W3LayerNorm(f) i,* ), for i=1…S,

[0052] In the formula, σ is the nonlinear activation layer (GELU); X is the input feature data; W is the weight matrix; U *,i It is the channel mixed information extraction method; Y j,* It is a token hybrid information extraction method;

[0053] f *,j represents the feature vector of the input data f on the jth channel. W2 represents the weight matrix of the second layer. W1 represents the weight matrix of the first layer, acting on the input feature f *,j Used to perform linear transformation. LayerNorm(f) *,j Indicates that for input f *,j Perform layer normalization. j represents the index of the channel. C represents the number of channels of the input feature.

[0054] f i,* represents the feature vector of the input data f on the i-th image block. W4 represents the weight matrix of the fourth layer, which is used to perform linear transformation on the features. W3 represents the weight matrix of the third layer, which acts on f i,* To achieve linear transformation. LayerNorm(f) i,* Indicates that for input f i,* Perform layer normalization. i represents the index number of the image block. S represents the number of image blocks. Further, the image sequence of the last layer normalization is denoted as f kλɑlt ={f kλɑlt i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +}, f kλɑlt i,j Represents the feature matrix output of the j-th channel of the i-th image block after processing.

[0055] The specific steps of layer normalization are shown in the formula:

[0056]

[0057] In the formula, f :,j Represents the input value before normalization operation. E[f] is the average value; Var[f :,j ] indicates f j The variance of f j represents the value of the jth feature. H represents the number of samples. σ represents the input f jThe standard deviation of . f is the layer normalized input; γ represents the gain, β represents the bias, and the gain γ and bias β ensure that the normalization operation does not destroy the memory information; p is a very small decimal. In step 4.4, the output spatial network form is [U, N1, B], where B is the number of images input at one time; the dataset f of the layer normalization in step 4.3 is kλɑlt ={f kλɑlt i,j,k |0 <i≤U,0<j≤N1,0<k≤B,i∈N + , j∈N + , k∈N +} Input the dimension reduction network and average along the middle dimension. The specific operations are as follows:

[0058]

[0059] In the formula, Represents the feature value after aggregation. is the input vector set; N1 is the length of the j-layer vector; i = 1…36; k = 1…16.

[0060] Extracting all features of the time-frequency image refers to inputting the time-frequency image whose local features have been extracted by 2DCNN into a dual-channel multi-layer perceptron mixer through the operation of step 4 above to extract the global features of the time-frequency image.

[0061] In step 5, a weighted voting mechanism is used to weight the global features of the extracted time-frequency image. The specific steps include:

[0062]

[0063] In the formula, ensemble_output represents the final output result after weighting the global features extracted from the time-frequency image. N is the number of classifiers, and w i is the weight of the i-th classifier, x i is the output of the i-th classifier.

[0064] The global features are weighted to obtain the final features, which are specifically summarized by the weighted results of the dual-channel multi-layer perceptron mixer, representing a comprehensive representation of the global features of the time-frequency image.

[0065] In step 5, the fault diagnosis model combines a two-dimensional convolutional neural network (2DCNN) and a dual-channel multi-layer perceptron mixer network. Figure 3 As shown. All formulas in steps 3 and 4 are involved.

[0066] A wind turbine gearbox fault diagnosis system based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism includes: an acquisition module, a data processing module, a model training module, a real-time fault detection module, and a fault alarm module.

[0067] The acquisition module uses a vibration sensor to acquire vibration signals of the gearbox of the wind turbine;

[0068] The data processing module is used to label the health status of the data samples of the wind turbine gearbox, divide the data set according to a certain ratio, and then perform continuous wavelet transform on the data set to obtain a two-dimensional wavelet time-frequency graph to divide the training set, test set and verification set;

[0069] The model training module is used to construct a dual-channel multi-layer perceptron mixer network model, and then use the training set to train the dual-channel multi-layer perceptron mixer network model, and use the verification set to verify it, and after the training is completed, a training model of the dual-channel multi-layer perceptron mixer network is obtained;

[0070] The real-time fault detection module is used to complete the real-time fault detection of the wind turbine gearbox by using the training model of the dual-channel multi-layer perceptron mixer network;

[0071] The fault alarm module is used to quickly send out an alarm signal when the real-time fault detection module detects a fault on the wind turbine gearbox, and use 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.

[0072] The present invention provides a wind turbine gearbox fault diagnosis method and system based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism, and the technical effects are as follows:

[0073] 1) The present invention uses continuous wavelet transform to convert one-dimensional time series signals into two-dimensional time-frequency images, which better preserves the temporal nature of time domain signals and utilizes the advantages of deep neural networks in processing image classification tasks. Compared with traditional direct learning of time domain signals, it is more advanced;

[0074] 2) Compared with the wind turbine gearbox fault diagnosis model using traditional convolutional neural networks, the fault diagnosis model proposed in the present invention combines a two-dimensional convolutional neural network (2DCNN) and a dual-channel multi-layer perceptron mixer network, which can fully extract the global features of the time-frequency images of various faults and achieve a maximum diagnostic accuracy of 100%. Its effect is significantly better than the evaluation performance of traditional convolutional neural networks.

[0075] 3) The present invention uses a weighted voting mechanism to weight the global features output by the dual-channel multi-layer perceptron mixer network, and utilizes the different feature representations of each channel to reduce the limitations of a single model and improve the diagnostic accuracy and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0077] Figure 1 This is a specific flow chart of the wind turbine gearbox diagnosis method of the present invention.

[0078] Figure 2 This is a diagram of the network structure of the multilayer perceptron mixer proposed in the present invention.

[0079] Figure 3 This is the overall network structure diagram of the convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism proposed in the present invention.

[0080] Figure 4 This is a schematic diagram of the continuous wavelet transform result proposed by the present invention.

[0081] Figure 5 This is a schematic diagram of the two-dimensional convolutional network structure proposed in the present invention.

[0082] Figure 6 This is a schematic diagram of the global feature extraction of two-dimensional time-frequency images proposed in the present invention.

[0083] Figure 7 Schematic diagram of a test bench according to an embodiment of the present invention.

[0084] Figure 8 This is a T-SNE feature visualization diagram of the test data in an embodiment of the present invention.

[0085] Fig. 9 This is a confusion matrix diagram of test data in an embodiment of the present invention. DETAILED DESCRIPTION

[0086] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0087] Embodiment 1:

[0088] A wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism. First, the original vibration signal is converted into a time-frequency image through continuous wavelet transform (CWT); then, the local features of the time-frequency image are extracted through a two-dimensional convolutional network, and the image data is divided into non-overlapping patches, and a dual-channel multi-layer perceptron mixer network is constructed to extract global features; finally, the two extracted global feature vectors are weighted to obtain the final feature representation, and the final fault diagnosis result is obtained through classification through a fully connected layer.

[0089] The wind turbine gearbox fault diagnosis method based on the convolution dual-channel multi-layer perceptron mixer and weighted voting mechanism of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0090] S1: Preprocess the collected original vibration signal and divide it into training set, validation set and test set;

[0091] S2: Use CWT to obtain the two-dimensional time-frequency image of the data set;

[0092] S3: Based on the two-dimensional time-frequency image, a two-dimensional convolutional network is constructed to extract the local features of the time-frequency image;

[0093] S4: At the same time, the time-frequency image with local features extracted in step 3 is segmented and compressed into the form of patch vectors, and input into the constructed dual-channel multi-layer perceptron mixer network to further extract the global features of the time-frequency image;

[0094] S5: Use a weighted voting mechanism to weight the global features extracted in step 4 to obtain the final feature representation, and use the final feature to train the fault diagnosis model;

[0095] S6: Input the test set into the trained model to obtain the final fault diagnosis result.

[0096] Furthermore, in S1, the original vibration signal is taken as a time domain signal sample with a length of 1024, the sliding window is 1024, and the overall sample is divided into a training set, a validation set and a test set in a ratio of 7:2:1.

[0097] Further, in S2, the time domain signal samples of the training set, the validation set and the test set are transformed into a two-dimensional time-frequency image using CWT;

[0098] The above CWT process is expressed as:

[0099]

[0100] Where x(t) is the vibration signal, s is the scale parameter, τ is the translation parameter, and ψ s,τis the mother wavelet function, ψ * is the complex conjugate.

[0101] Furthermore, in S3, in order to better extract the local features of the two-dimensional time-frequency image, a 2DCNN is constructed according to the image, and is applied before extracting the global features of the image as a preliminary feature extraction.

[0102] Furthermore, the multilayer perceptron mixer network structure in S4 is as follows: Figure 2 As shown, S4 may include the following four steps:

[0103] S4.1: Divide the time-frequency image (H×W×C) from which local features have been extracted into N non-overlapping images. Each block is linearly projected with the same projection matrix to the required hidden dimension C. Divide the image into N1 P×P×C image blocks, where the sequence length N1 = H×W / P 2 , and then each tile is converted into a one-dimensional vector, represented as P 2 C;

[0104] S4.2: Input the vector set into the fully connected layer (FC) and use U neurons to reduce the dimension of f. The output image set after processing is recorded as f k ={f k i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +};

[0105] S4.3: The dual-channel multilayer perceptron mixer network model contains two types of layers, namely the token mixing layer and the channel mixing layer. Each layer contains a nonlinear activation function (GULE) and Dropout to suppress overfitting of the network. The token mixing MLP takes the time-frequency image as input in the form of columns, and the image is deeply extracted by two layers of neurons, 4×U and U, respectively. The output image set after the first layer of token processing is denoted as f kλɑ1 ={f kλɑ1 i,j |0 <i≤4U,0<j≤N1,i∈N + , j∈N +}, the image set output after the second layer of token processing is denoted as f kλɑ2 ={f kλɑ2 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +}. The channel mixing MLP converts f kTaking the form of rows as input, a two-layer CNN neural network with a convolution kernel size of 1×1, a step size of 1, and a number of 4×Q and Q respectively is used to extract information from the image. The image set output after the first layer of channel processing is denoted as f kλ ={f kλ i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +}, the image set output after the second layer channel processing is recorded as f kλ1 ={f k λ1 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N +}. Then use skip-connection and layer normalization to add the input and output.

[0106] Furthermore, the nonlinear activation layer (GELU) is used to perform nonlinear transformation on the input signal to increase the nonlinear expression ability of the network. The second layer activation operation in the first conventional convolution module is described by the GELU function, and the specific formula of the GELU is:

[0107] GELU(f)=f×Φ(f)

[0108] Where f is the GULE function input; Φ(x) is the normal distribution probability function.

[0109] Furthermore, the specific operation steps of the skip connection are shown in the formula:

[0110] U *,j =f *,j +W2σ(W1 LayerNorm(f) *,j ), for j=1…C,

[0111] Y i,* =f i,* +W4σ(W3LayerNorm(f) i,* ), for i=1…S,

[0112] Where σ is the nonlinear activation layer (GELU); X is the input feature data; W is the weight matrix; U *,i It is the channel mixed information extraction method; Y j,* It is a token hybrid information extraction method.

[0113] Furthermore, denote the image set after the last layer of normalization as \(f_{k\lambda alt}=\{f_{k\lambda alti,j}|0 < i\leq U, 0 < j\leq N1, i\in N^+, j\in N^+\}\), and the specific operation steps of layer normalization (Layer Norm) are shown in the formula:

[0114]

[0115]

[0116] In the formula, \(f\) is the input of layer normalization; \(E[f]\) is the average value; the gain \(\gamma\) and bias \(\beta\) ensure that the normalization operation does not destroy the memory information; \(p\) is a very small decimal.

[0117] S4.4: The spatial network form of the output is \([U, N1, B]\), where \(B\) is the number of pictures input at one time. Input the dataset \(f\) kλɑlt =\(\{f\) kλɑlt i,j,k |0 < i\leq U, 0 < j\leq N1, 0 < k\leq B, i\in N + , j\in N + , k\in N + \} into the dimensionality reduction network and average along the middle dimension. The specific operation steps are shown in the formula:

[0118]

[0119] In the formula, is the set of input vectors; \(N1\) is the length of the vector at layer \(j\).

[0120] S5: Use the weighted voting mechanism to weight the global features of the extracted time-frequency images and classify them through the fully connected layer, and save the best training model. The overall network structure diagram of the convolutional dual-channel multi-layer perceptron mixer and the weighted voting mechanism is as Figure 3 shown, and the weighting formula is:

[0121]

[0122] In the formula, \(N\) is the number of classifiers, \(w\) i is the weight of the \(i\)-th classifier, and \(x\) i is the output of the \(i\)-th classifier.

[0123] S6: Test the test set divided by the best training model saved in S5 to obtain the fault diagnosis result of the method proposed in this patent.

[0124] The following is a detailed description in combination with specific embodiments:

[0125] The embodiment of the present invention is an application of a wind turbine gearbox fault diagnosis method based on a convolutional dual-channel multilayer perceptron mixer and a weighted voting mechanism on a University of Connecticut gearbox dataset.

[0126] like Figure 7 As shown in Figure 1, a university gearbox dataset is mainly a benchmark two-stage gearbox with replaceable gears. The speed of the gears is controlled by a motor, and the torque is provided by a magnetic brake that is adjusted by changing the input voltage. The first-stage input shaft has 32 and 80 gears, and the second stage has 48 and 64 gears. The speed of the input shaft is measured by a toothed tachometer, where the vibration signal is recorded by the dSPACE system at a sampling frequency of 20KHz. The gears on the input shaft introduce 9 different health states, including healthy, missing teeth, tooth root cracks, spalling and cracking, among which there are 5 cracking faults with different severity.

[0127] The present invention selects five different fault states in a university gearbox: healthy, missing teeth, tooth root cracks, spalling and cracking as experimental data, divides the signal length of 1024 in each healthy state into one sample, and the sliding window is also 1024. The vibration signals under different fault states are converted into time-frequency images through CWT to form a diagnostic data set, where the total number of samples in each state is 200, and the training set, validation set and test set are divided according to the ratio of 7:2:1. The specific experimental data information is shown in Table 1. The experimental hyperparameters are set as follows: learning rate is 0.001, optimizer is SGD, epochs is 20, and batch size is 16.

[0128] Table 1 Experimental data information

[0129]

[0130] In this embodiment, in order to intuitively demonstrate the fault classification effect of the proposed method, T-SNE is used to reduce the dimension of the original data and classification results in the test set. Figure 8 As shown in the figure, before model training, the sample data of the five health categories are mixed together and the distribution is messy, and the samples of different health categories cannot be distinguished well. After model training, it can be seen that each category can be distinguished well, and the same type of data is concentrated in one cluster, and there is no misclassification.

[0131] After using the training set and validation set to train and validate the proposed model, the diagnostic experiment is conducted on the test set, and the diagnosis results are visualized through the confusion matrix, such as Fig. 9 As shown. Fig. 9 It can be seen that the proposed model accurately identifies the data of five health states and has good diagnostic accuracy.

[0132] To demonstrate the superiority of the proposed method, it is compared with four related methods:

[0133] Method 1 is a fault diagnosis method for a dual-channel multi-layer perceptron mixer;

[0134] Method 2 is a fault diagnosis method based on dual-channel multi-layer perceptron mixer and voting integration;

[0135] Method 3 is a fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer;

[0136] Method 4 is a fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and voting integration;

[0137] Method 5 is the method proposed by the present invention. The parameters such as the number of iterations, batch size, learning rate and optimizer of all methods are kept consistent. The diagnostic results of different methods are shown in Table 2.

[0138] Table 2 Diagnostic results of different methods

[0139]

[0140]

[0141] From the comparison experimental results in the table, it can be seen that the method proposed in the present invention has achieved the highest diagnostic accuracy of 100%, which is 7.82%, 4.43%, 4.17%, and 1.05% higher than the other four methods, respectively. Compared with the method of using a dual-channel multilayer perceptron mixer alone or in combination with voting integration, the method proposed in the present invention can more effectively improve the fault diagnosis accuracy. In addition, the average training time required for the five methods is about 400s, with little fluctuation.

[0142] Embodiment 2:

[0143] The second aspect of the present invention is a wind turbine gearbox fault diagnosis system based on a convolutional dual-channel multi-layer perceptron mixer and a weighted voting mechanism, comprising: an acquisition module, a data processing module, a model training module, a real-time fault detection module, and a fault alarm module.

[0144] The acquisition module uses a vibration sensor to acquire vibration signals of the gearbox of the wind turbine;

[0145] The data processing module is used to label the health status of the data samples of the wind turbine gearbox, divide the data set according to a certain ratio, and then perform continuous wavelet transform on the data set to obtain a two-dimensional wavelet time-frequency graph to divide the training set, test set and validation set;

[0146] The model training module is used to construct a dual-channel multi-layer perceptron mixer network model, and then use the training set to train the dual-channel multi-layer perceptron mixer network model, and use the verification set to verify it, and after the training is completed, a training model of the dual-channel multi-layer perceptron mixer network is obtained;

[0147] The real-time fault detection module is used to complete the real-time fault detection of the wind turbine gearbox by using the training model of the dual-channel multi-layer perceptron mixer network;

[0148] The fault alarm module is used to quickly send 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 wind turbine gearbox. When the wind turbine gearbox operates normally, the fault alarm module automatically enters a silent state.

[0149] 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 convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism. 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 convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism.

Claims

1. A wind turbine gearbox fault diagnosis method based on a convolutional dual-channel multilayer perceptron mixer and a weighted voting mechanism, characterized in that The following steps are involved: Step 1: Preprocess the collected original vibration signal and divide it into training set, verification set and test set; Step 2: Use CWT to obtain the time-frequency image of the data set; Step 3: Construct a two-dimensional convolutional neural network 2DCNN to extract local features of the time-frequency image; Step 4: The time-frequency image with local features extracted in step 3 is segmented and compressed into a patch vector, and input into a dual-channel multi-layer perceptron mixer network to further extract the global features of the time-frequency image; Step 5: Use the weighted voting mechanism to weight the global features extracted in step 4 to obtain the final features, and use the final features to train the fault diagnosis model; Step 6: Input the test set into the trained fault diagnosis model to obtain the final fault diagnosis result.

2. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism 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 converted into two-dimensional time-frequency images using CWT; The above CWT process is expressed as: Where W(s,τ) represents the coefficient of wavelet transform, which represents the local characteristics of signal x(t) at scale s and time translation τ; <x(t),ψ s,τ > represents the signal x(t) and the wavelet function ψ s,τ The inner product of the signal represents the similarity between the signal and the wavelet function under a given scale s and translation τ; x(t) is the vibration signal, s is the scale parameter, τ is the translation parameter, t represents the time variable, and represents the position of the signal x(t) on the time axis, ψ s,τ is the mother wavelet function, ψ * is the complex conjugate.

3. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 1 is characterized in that: In the step 3, in order to better extract the local features of the two-dimensional time-frequency image, a two-dimensional convolutional neural network 2DCNN is constructed according to the two-dimensional time-frequency image. The two-dimensional convolutional neural network 2DCNN is applied before extracting the global features of the image as a preliminary feature extraction of the two-dimensional time-frequency image to extract the local features of the two-dimensional time-frequency image.

4. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 1 is characterized in that: The step 4 comprises the following steps: Step 4.1: Divide the time-frequency image from which local features have been extracted into N1 non-overlapping image blocks, each of which is linearly projected with the same projection matrix to the required hidden dimension C; divide the time-frequency image from which local features have been extracted into N1 P×P×C image blocks, where the sequence length N1 = H×W / P 2 Then, each image block is converted into a one-dimensional vector, represented as P 2 C; Step 4.2: Use a fully connected network to process the rearranged image blocks to reduce the number of parameters and further extract features; Step 4.3: Construct a dual-channel multi-layer perceptron mixer network, which includes an MLP module and a Mixer layer. The MLP module is a standard fully connected layer structure, which is used to perform feature mapping on the input data in a linear space and map low-dimensional input to a high-dimensional representation. The main function of the Mixer layer is to mix and update features, and it includes two sublayers: a token mixing layer and a channel mixing layer. The processed image blocks are respectively input into a dual-channel multi-layer perceptron mixer network model to extract the shared features of the gear signal blocks. The dual-channel multi-layer perceptron mixer network includes a fully connected layer and a mixing layer. Step 4.4: After normalizing the extracted feature layer, the feature data is averaged and rearranged into 3 dimensions in the n dimension.

5. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 4 is characterized in that: In step 4.2, the image sequence is fed into the fully connected layer, and U neurons are used to reduce the dimension of f. The image sequence output after processing is recorded as f k ={f k i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N + }, f k i,j Represents the feature value of the i-th image block after passing through the k-th fully connected layer (FC) on the j-th channel; i represents the number of the image block; U represents the number of neurons in the fully connected layer, indicating the output dimension of the model in the fully connected layer; j represents the index of the channel dimension, indicating the jth channel in each image block; N1 represents the total number of image blocks; N + Represents the set of positive integers.

6. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 4 is characterized in that: In step 4.3, the dual-channel multi-layer perceptron mixer network model includes two types of layers, namely, a token mixing layer and a channel mixing layer, and each layer includes a nonlinear activation function and Dropout to suppress overfitting of the network; The token hybrid MLP takes the time-frequency image as input in the form of columns, and uses two fully connected layers with 4×U and U neurons to extract information from the time-frequency image in depth; The output image set after the first layer of token processing is denoted as f kλɑ1 ={f kλɑ1 i,j |0 <i≤4U,0<j≤N1,i∈N + , j∈N + }, f kλɑ1 i,j Indicates that the jth channel of the i-th image block is in the k-th λa1 The output feature value after the fully connected layer is processed; U represents the output dimension of the fully connected layer (FC); N1 represents the total number of image blocks; The image set output after the second layer of token processing is denoted as f kλɑ2 ={f kλɑ2 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N + }, f kλɑ2 i,j Indicates that the jth channel of the i-th image block is in the k-th λa2 Output feature value after processing by the fully connected layer; Channel Mixing MLP: MLP is used to mix the channel dimension of the image. It processes the information of all image blocks on the same channel to achieve cross-channel information interaction. The image set output after the first layer of channel processing is denoted as f kλ ={f kλ i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N + }, f kλ i,j Indicates that the jth channel of the i-th image block is in the k-th λ Output feature value after processing by the fully connected layer; The image set output after the second layer of channel processing is denoted as f kλ1 ={f kλ1 i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N + }, f kλ1 i,j Indicates that the jth channel of the i-th image block is in the k-th λ1 Output feature value after processing by the fully connected layer; Next, the input and output are added together using skip-connection and layer normalization.

7. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 6 is characterized by: The nonlinear activation function (GULE) is used to perform nonlinear transformation on the input signal to increase the nonlinear expression ability of the network; the second layer activation operation in the first conventional convolution module is described by the nonlinear activation function (GULE), and the specific formula of the GELU is: GELU(f)=f×Φ(f) Where f is the GULE function input; Φ(f) represents the cumulative distribution function (CDF) of the standard normal distribution; for input f, the CDF of the standard normal distribution is: Where erf(x) is the error function, which is used to approximate the cumulative distribution of the Gaussian distribution; The specific operation steps of jump connection are shown in the formula: U *,j =f *,j +W2σ(W1 LayerNorm(f) *,j ),for j=1…C, Y i,* =f i,* +W4σ(W3LayerNorm(f) i,* ),for i=1…S, Where σ is the nonlinear activation layer (GELU); U *,i It is the channel mixed information extraction method; Y j,* It is a token hybrid information extraction method; f *,j represents the feature vector of the input data f on the jth channel; W2 represents the weight matrix of the second layer; W1 represents the weight matrix of the first layer, acting on the input feature f *,j Used to perform linear transformation; LayerNorm(f) *,j Indicates that for input f *,j Perform layer normalization operation; j represents the index number of the channel; C represents the number of channels of the input feature; f i,* represents the feature vector of the input data f on the i-th image block; W4 represents the weight matrix of the fourth layer, which is used to perform linear transformation on the features; W3 represents the weight matrix of the third layer, which acts on f i,* To achieve linear transformation; LayerNorm(f) i,* Indicates that for input f i,* Perform layer normalization operation; i represents the index number of the image block; S represents the number of image blocks; Furthermore, the image sequence of the last layer normalization is recorded as f kλɑlt ={f kλɑlt i,j |0 <i≤U,0<j≤N1,i∈N + , j∈N + }, f kλɑlt i,j Represents the feature matrix output by the jth channel of the i-th image block after processing; The specific steps of layer normalization are shown in the formula: In the formula, f :,j represents the input value before normalization operation; E[f] is the average value; Var[f :,j ] indicates f j The variance of j represents the value of the jth feature; H represents the number of samples; σ represents the input f j The standard deviation of ; f is the normalized input of the layer; γ represents the gain, β represents the bias, the gain γ and bias β ensure that the normalization operation does not destroy the memory information; p is a very small decimal.

8. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 4 is characterized in that: In step 4.4, the output spatial network form is [U, N1, B], where B is the number of images input at one time; the normalized data set f in step 4.3 is kλɑlt ={f kλɑlt i,j,k |0 <i≤U,0<j≤N1,0<k≤B,i∈N + , j∈N + , k∈N + } Input the dimension reduction network and average along the middle dimension. The specific operations are as follows: In the formula, Represents the feature value after aggregation; is the input vector set; N1 is the length of the j-layer vector; i = 1…36; k = 1…16.

9. The wind turbine gearbox fault diagnosis method based on convolutional dual-channel multi-layer perceptron mixer and weighted voting mechanism according to claim 4 is characterized in that: In step 5, a weighted voting mechanism is used to weight the global features of the extracted time-frequency image. The specific steps include: In the formula, ensemble_output represents the final output result after weighting the global features extracted from the time-frequency image; N is the number of classifiers, and w i is the weight of the i-th classifier, x i is the output of the i-th classifier.

10. A wind turbine gearbox fault diagnosis system based on a convolutional dual-channel multilayer perceptron mixer and a weighted voting mechanism, characterized in that 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 of the gearbox of the wind turbine; The data processing module is used to label the health status of the data samples of the wind turbine gearbox, divide the data set according to a certain ratio, and then perform continuous wavelet transform on the data set to obtain a two-dimensional wavelet time-frequency graph to divide the training set, test set and verification set; The model training module is used to construct a dual-channel multi-layer perceptron mixer network model, and then use the training set to train the dual-channel multi-layer perceptron mixer network model, and use the verification set to verify it, and after the training is completed, a training model of the dual-channel multi-layer perceptron mixer network is obtained; The real-time fault detection module is used to complete the real-time fault detection of the wind turbine gearbox by using the training model of the dual-channel multi-layer perceptron mixer network; The fault alarm module is used to quickly send out an alarm signal when the real-time fault detection module detects a fault on the wind turbine gearbox, and use 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.

Images