Fan rolling bearing vibration signal noise reduction method based on improved convolution auto-encoder

Abstract

The invention discloses a fan rolling bearing vibration signal noise reduction method based on an improved convolution auto-encoder. The method specifically comprises the following steps: (1) establishing a fan bearing vibration signal data set; (2) adding noise to the bearing vibration signal in the data set to form a noisy signal; (3) constructing a fan rolling bearing vibration signal noise reduction model based on multi-scale attention residual convolution self-coding; (4) training the model by adopting the training set and the verification set, and storing the trained model as an optimal model; and (5) testing the model by using the test set, and obtaining the final fan rolling bearing vibration signal noise reduction model if the test result meets the precision requirement. Compared with the prior art, the fan rolling bearing vibration signal noise reduction method based on the improved convolution auto-encoder can adaptively adjust the weight of the extracted features, improves the feature extraction and noise reduction capability of the model, and improves the convergence speed of the model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method for denoising the vibration signal of a fan rolling bearing based on an improved convolutional autoencoder. Background Art

[0002] As an important part of a wind turbine, the health status of a rolling bearing has a crucial impact on the safety and availability of a wind turbine. For this reason, most wind turbines are equipped with a fan condition monitoring system (CMS). This detection system mainly uses various monitoring technologies to manage the health of the wind turbine. For example, vibration monitoring collects the vibration data of the internal bearing through sensors, including using a low-frequency accelerometer to monitor the main bearing and a high-frequency accelerometer to monitor the generator bearing. However, the wind turbine is in a harsh working environment for a long time, and the signals detected by the sensors will be interfered by noise to varying degrees. These noise interferences will distort the signals or mask the true fault characteristics, affecting the judgment of the bearing working status and the accuracy of fault diagnosis. Therefore, denoising the vibration signal of the fan bearing is an important step in the health monitoring of the wind turbine.

[0003] Traditional signal denoising methods mainly include methods based on wavelet transform, empirical mode decomposition, and variational mode decomposition. Most traditional signal denoising methods generally have three main steps: first, the signal is decomposed into multiple components by a signal decomposition method; second, through a signal processing method, different threshold functions are used to denoise the decomposed signals; and finally, the signal components with more obvious features are selected for reconstruction. However, these denoising methods have disadvantages such as weak feature extraction ability and poor generalization. The selection of hyperparameters and threshold functions will affect the denoising performance. Therefore, a more intelligent method is needed for signal denoising.

[0004] With the rapid development of deep learning, deep learning networks such as convolutional neural networks, residual networks, and convolutional autoencoder networks are widely used in vibration signal denoising. The vibration signal denoising method based on deep learning uses a large amount of data and the powerful learning ability of the model to learn the characteristics and distributions of vibration signals, so as to obtain a denoising effect close to the original signal.

[0005] However, the current deep learning methods still have certain limitations. First, when the deep learning model uses a single-scale convolutional kernel to process vibration signals for feature extraction, some key information may be lost, and the feature extraction ability is weak. Second, the multi-scale feature extraction method does not fully consider the weight differences of different-scale features in the signal feature extraction process, affecting the final feature extraction effect and the denoising adaptability is not high. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention proposes a method for denoising the vibration signal of a fan rolling bearing based on multi-scale attention residual convolutional autoencoder. By improving the convolutional autoencoder model, using the multi-scale feature extraction method and the attention residual link, the feature weights are adaptively adjusted, multi-scale features are extracted, the problems of low self-adaptability of vibration signal denoising and weak feature extraction ability are solved, and the convergence speed of the model is accelerated at the same time.

[0007] The technical solution provided by the present invention includes the following steps:

[0008] Step 1: Obtain the vibration signal of the fan bearing to form a first data set;

[0009] Further, in the above step 1, the vibration signal of the fan bearing can be obtained from a public data set, collected by a vibration sensor for the vibration signal, or obtained from a data acquisition system.

[0010] Step 2: Add noise to the vibration signal in the first data set to form a second data set, and divide the second data set into a training set, a validation set and a test set;

[0011] Preferably, in the above step 2, the training set, the validation set and the test set can be divided according to the ratio of 7:2:1.

[0012] Step 3: Construct a model for denoising the vibration signal of a fan rolling bearing based on multi-scale attention residual convolutional autoencoder;

[0013] Further, the above step 3 specifically includes steps 3.1 to 3.4:

[0014] Step 3.1: Convert the convolutional layer of the encoder in the convolutional autoencoder model into a one-dimensional convolution. The improved encoder consists of three one-dimensional convolutional layers, three activation functions and three max-pooling layers, and converts the input signal x=(x1, x2,... x N ) into a low-dimensional latent space representation, where x1, x2,... x N is the numerical sequence of sampling points. The 3×1 convolution operation of the convolutional layer is:

[0015] X (L) =F(∑Conv1D(x (L-1) ,W (L) )+b (L) ) (1)

[0016] where, X (L) is the output of the L-layer convolutional layer, x (L-1) is the input of the convolutional layer, W (L) is the convolutional kernel, and b (L)is the bias term, Conv1D() is the one-dimensional convolution calculation, and F() is the activation function. In the encoder, the activation function 1, activation function 2, and activation function 3 select the Relu function for non-linear transformation. The calculation formula is as follows:

[0017] F(a) = max{0, a} (2)

[0018] where a is the activation value;

[0019] The pooling layer is used for downsampling to improve the computational efficiency by reducing the data dimension. In the encoder of this model, max pooling is adopted. For a one-dimensional sequence X of length N as the input signal, with a pooling window size of k and a stride of s, the calculation formula for each element of the output sequence Y is as follows:

[0020]

[0021] where X represents the input one-dimensional signal with a length of N in , Y is the output one-dimensional signal with a length of N out , i represents the position index of the output sequence, and the value range is i = 0, 1, 2... N out - 1, and p is the padding number taken as zero;

[0022] Step 3.2: Combine the Atrous Spatial Pyramid Pooling (ASPP) module and the Channel Attention (ECA) module. Add the ECA module after each dilated convolution to form the ASPP-ECA module, and add the ASPP-ECA module between the encoder and the decoder in the convolutional autoencoder;

[0023] Furthermore, the ASPP-ECA module has a total of 5 branches. The first branch passes through the 1×1 convolutional layer 1 from the input layer of the module to extract the high-frequency components in the signal and transmit the high-frequency features of the signal to the connection layer;

[0024] The second branch, the third branch, and the fourth branch each contain a dilated convolution with three different dilation rates and an ECA module. The dilation rates of the three dilated convolutions are 4, 8, and 12 respectively. The effective coverage range of the convolution kernel can be expanded through medium dilation rates to capture features at medium time scales, and the attention module weights the extracted features at each scale to adjust the weights of different features. There is a global average pooling and an adaptive one-dimensional convolutional layer in the ECA module. The module receives the feature map of the dilated convolution. The dimension of the feature map is set as H×W×C, where C is the number of channels, and H and W are the height and width. After global average pooling, the H×W spatial information of each channel is compressed into a scalar 1×1×C feature map, and then local cross-channel interaction analysis is performed on this feature map vector through adaptive one-dimensional convolution. The convolution kernel of this convolutional layer is a dynamic convolution kernel, and the size of the one-dimensional convolution kernel is automatically determined according to the current number of channels C according to the following formula:

[0025]

[0026] where the convolution kernel size is k and the number of channels is C;

[0027] The convolution kernel performs local interaction analysis on the scalar values of k adjacent channels, learns the importance between different channels, adaptively generates the attention weights of the channels, and finally combines the features containing channel attention Figure 1 ×1×C and the original input feature map H×W×C, perform channel multiplication, multiply all spatial positions of each channel by the same weight value to achieve feature selection, and finally output the feature map H′×W′×C with channel attention. Finally, the three branches output the features with different weights to the connection layer;

[0028] The fifth branch transmits the input of the module to the pooling layer. The pooling layer compresses the time dimension of the input signal, performs global statistics, calculates the mean of each channel. The output of the pooling layer is transmitted to the 1×1 convolutional layer 2. The convolutional layer performs inter-channel interaction on the pooled features to generate channel weights. The output of the convolutional layer is transmitted to the upsampling layer to align the global features with the local convolutional features. The upsampling layer transmits the features to the connection layer. The connection layer fuses the obtained features and transmits them to the 1×1 convolutional layer 3. After calculating the features by the convolutional layer 3, they are output by the output layer;

[0029] Step 3.3: The decoder consists of a transposed convolution, an activation function, and an upsampling layer. The decoding part gradually expands the signal dimension through transposed convolution and upsampling. Let z be the latent representation output by the encoder, and use the transposed convolution kernel W (d) and the bias term b (d) to perform the deconvolution operation:

[0030] k (L) = F(∑DeConv1D(z, w (d) ) + b (d) ) (5)

[0031] where DeConv1D() is the transposed convolution operation, and k (L) is the output part of the L-layer transposed convolution. The output signal usually also undergoes a non-linear mapping through the activation function to ensure that the output of the signal is within the effective range. The activation function 4 and the activation function 5 select the Relu activation function;

[0032] Through upsampling, the network restores the time resolution of the signal during the decoding process. Finally, the LeakyRelu function improves the ability to capture the bidirectional features of the signal by retaining the negative gradient and signal information, enhancing the integrity of the denoised signal. The function formula is as follows:

[0033]

[0034] Among them, α is taken as 0.01;

[0035] Step 3.4: The main path of the residual connection is composed of convolutional layer 3, activation function 3, max pooling layer 3, ASPP-ECA module, transposed convolution 1, activation function 4, and upsampling 1. The skip connection line connects the output of the max pooling layer 2 of the second small encoding structure in the encoder to the residual connection block after the first small decoding structure in the decoder. At the residual connection block, the output of upsampling 1 at the main path end is combined with the output of the max pooling layer 2. The residual connection block is composed of a 1×1 convolution and a fusion module connected in sequence. In the residual connection block, the skip connection line is connected to the fusion module through a 1×1 convolution. The fusion module receives the output of the 1×1 convolutional layer and the output of upsampling 1, and fuses the data of the two branches, and the fused result is used as the overall output;

[0036] Step 4: Use the training set and the validation set to train the wind turbine rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder, and save the trained model as the optimal model;

[0037] Further, the step 4 specifically includes steps 4.1 to 4.4:

[0038] Step 4.1: Set the training parameters of the wind turbine rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder. The model training parameters include: learning rate, number of channels, convolutional kernel size, optimizer, number of iterations;

[0039] Step 4.2: Input the noisy signal samples of the training set and the validation set into the wind turbine bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder, and use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and adjust the model parameters by minimizing the loss function to gradually approach the optimal solution;

[0040] Step 4.3: Use the Adam optimizer to update the model parameters, so that the model parameters are updated in the direction of gradient descent until the loss functions of the training set and the validation set no longer decrease, and at the same time the evaluation index signal-to-noise ratio SNR no longer increases, and the root mean square error EMSE and the mean absolute error MAE no longer decrease;

[0041] Step 4.4: Save the trained model parameters as the optimal model.

[0042] Step 5: Use the test set to test the optimal model, evaluate the test results of the test set, and meet the accuracy requirements, that is, obtain the final wind turbine rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder;

[0043] Further, step 5 specifically includes steps 5.1 to 5.4:

[0044] Step 5.1: Input the test set into the optimal model described in step 4;

[0045] Step 5.2: Calculate the model performance metrics: signal-to-noise ratio SNR, root mean square error RMSE, and mean absolute error MAE. The specific calculation formulas are as follows:

[0046]

[0047]

[0048]

[0049] where SNR is the signal-to-noise ratio, RMSE is the root mean square error, MAE is the mean absolute error, S i is the original signal, S' i represents the signal output after noise reduction, and n is the number of signal samples;

[0050] Step 5.3: When the performance metrics meet the accuracy requirements, obtain the final noise reduction model for the vibration signal of the fan rolling bearing based on the multi-scale attention residual convolutional autoencoder.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] (1) Between the encoder and decoder of the convolutional autoencoder, the ASPP-ECA module disclosed in this invention patent is adopted, and dilated convolutions with different dilation rates are used to increase the receptive field without increasing the number of parameters and extract multi-scale features of the signal; the channels are weighted in different branches to adaptively adjust the weights of each convolutional kernel for feature extraction, which can improve the model's ability to extract signal features and its adaptability to noise intensity.

[0053] (2) Residual connections are introduced into the convolutional autoencoder to reduce the model learning difficulty, effectively accelerate the model's learning of signal features, and improve the model convergence speed.

[0054] (3) The last small decoding structure of the convolutional autoencoder is improved. The original activation function is replaced with the LeakyRelu function and added after the last upsampling layer, effectively avoiding information loss, enhancing the model's non-linear expression ability, effectively restoring the original features of the signal, improving the signal-to-noise ratio of the signal, making the power of the signal in the data much higher than the power of the noise, and achieving the noise reduction effect on the signal data. Description of the Drawings

[0055] Figure 1This is the flowchart of the method for reducing the noise of the vibration signal of the fan rolling bearing based on the improved convolutional autoencoder of the present invention;

[0056] Figure 2 This is the schematic diagram of the structure of the noise reduction model for the vibration signal of the fan rolling bearing based on the multi-scale attention residual convolutional autoencoder of the present invention;

[0057] Figure 3 This is the schematic diagram of the ASPP-ECA structure;

[0058] Figure 4 This is the schematic diagram of the structure of the ECA module;

[0059] Figure 5 This is the schematic diagram of the structure of the residual connection block; Specific implementation manner

[0060] In order to make the technical solution, structural features, achieved purpose and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the specific implementation manner and with reference to the accompanying drawings. It should be noted that the specific embodiments described herein are only used to explain the present invention more clearly and are not used to limit the present invention.

[0061] Figure 1 This is the flowchart of the method for reducing the noise of the vibration signal of the fan rolling bearing based on the multi-scale attention residual convolutional autoencoder proposed by the present invention, and its implementation process is as follows:

[0062] Step 1: Obtain the vibration signal of the fan bearing to form a first data set; in the first data set, the vibration signal of the fan bearing can be collected through the network, collected by a vibration sensor for the vibration signal, or obtained from a data acquisition system;

[0063] In this embodiment, in order to better evaluate the denoising effect of the method for reducing the noise of the vibration signal of the fan rolling bearing based on the multi-scale attention residual convolutional autoencoder proposed by the present invention, the publicly available Case Western Reserve University Bearing Data Set is adopted; the Case Western Reserve University Bearing Data Set was obtained by Case Western Reserve University in the United States in a test bench composed of a 2-horsepower motor, a torque sensor, and a power tester. An acceleration sensor was placed above the bearing housing to collect the vibration acceleration signal of the faulty bearing. The sampling frequency was 12KHz. The bearing was damaged by electrical discharge machining at a single point. The data set contains damages at three different positions of the outer ring, inner ring, and ball of the bearing and damages at three different degrees of 0.007 inches (0.1778 mm), 0.014 inches (0.3556 mm), and 0.021 inches (0.5334 mm). Here, the inner ring raceway fault signal of the deep groove ball rolling bearing with a sampling frequency of 12KHz and a damage diameter of 0.007 inches (0.1778 mm) is selected to form the first data set;

[0064] Step 2: Add noise to the signals in the first dataset and divide them into small samples to form a second dataset;

[0065] Construct a noisy signal with a signal-to-noise ratio of 0 dB. First, use a function to calculate the power of the signal, which is the variance of the signal. Second, convert the signal-to-noise ratio in decibels (dB) to a linear proportional form. The calculation formula is as follows:

[0066]

[0067] where SNR Linear is the linear proportional value of the signal-to-noise ratio, and SNR dB is the decibel value of the signal-to-noise ratio;

[0068] Then, according to the definition of the signal-to-noise ratio, calculate the noise power through the signal power and the linear proportional signal-to-noise ratio. The calculation formula is as follows:

[0069]

[0070] where noise poewr is the noise power, and signal_power is the signal power; by taking the square root of the noise power, the standard deviation of the noise can be obtained. Using the noise standard deviation and the shape of the original signal as the parameters of the random number function, ensure that the noise and the signal have the same shape and can be added element by element. Use the random number function to generate a series of random numbers as the noise values. The generated random numbers follow a Gaussian distribution. Add the noise values and the original signal to obtain a noisy signal;

[0071] Convert the signal files in.mat format in the first dataset in this embodiment to npy format signal files required by the numpy library in python. Divide the signal files in the first dataset into individual small samples. Each small sample contains 2048 consecutive sampling points and is divided into 2411 small samples in total to form a second dataset. And divide the second dataset into a training set, a validation set, and a test set according to 7:2:1; the divided training set includes 1688 signal samples, the validation set includes 482 signal files, and the test set includes 241 signal files;

[0072] Step 3: Construct a noise reduction model for the vibration signals of the fan rolling bearing based on a multi-scale attention residual convolutional autoencoder. The improved signal denoising model structure is as Figure 2 shown. The model construction process specifically includes steps 3.1 to 3.3:

[0073] Step 3.1: In the convolutional autoencoder model, convert the convolutional layers of the encoder into one-dimensional convolutions. The improved encoder consists of three one-dimensional convolutional layers, three activation functions, and three max-pooling layers;

[0074] A further encoding part converts the input signal x = (x1, x2, … x N ) into a low-dimensional latent space representation, where x1, x2, … x N is a sequence of acceleration values, and the 3×1 convolution operation of the convolutional layer is as follows:

[0075] X (L) = F(∑Conv1D(x (L-1) , W (L) ) + b (L) ) (3)

[0076] where X (L) is the output of the L-layer convolutional layer, x (L-1) is the input of the convolutional layer, W (L) is the convolutional kernel, b (L) is the bias term, Conv1D() is the one-dimensional convolution calculation, F() is the activation function. In the encoder, the activation functions 1, 2, and 3 select the RELU function as the activation function for non-linear transformation, where a is the activation value;

[0077] F(a) = max {0, a} (4)

[0078] The pooling layer is used for downsampling to improve the computational efficiency by reducing data congestion. In the encoder of this model, the max-pooling operation is adopted. For a one-dimensional sequence X of length N as the input signal, with a pooling window size of k and a stride of s, the calculation formula for each element of the output sequence Y is as follows:

[0079]

[0080] X represents the input one-dimensional signal with a length of N in , Y is the output one-dimensional signal with a length of N out , i represents the position index of the output sequence, and the value range is i = 0, 1, 2… N out -1, and p is the padding number taken as zero;

[0081] Step 3.2: Combine the atrous spatial pyramid pooling (ASPP) module and the channel attention (ECA) module. Add the ECA module after each dilated convolution to form the ASPP-ECA module as Figure 3 shown, and add the ASPP-ECA module between the encoder and the decoder in the convolutional autoencoder;

[0082] The further ASPP-ECA module has a total of 5 branches. The first branch passes through the 1×1 convolutional layer 1 from the input layer of the module to extract the high-frequency components in the signal and transmit the high-frequency features of the signal to the connection layer;

[0083] The second branch, the third branch, and the fourth branch each contain three dilated convolutions with different dilation rates and an ECA module. The dilation rates of the three dilated convolutions are 4, 8, and 12 respectively. The effective coverage range of the convolution kernel can be expanded through medium dilation rates to capture features at medium time scales, and the ECA attention module weights the features extracted at each scale to adjust the weights of different features. The ECA module is as shown in Figure 4 . It receives the feature map of the dilated convolution. The dimension of the feature map is set as H×W×C, where C is the number of channels, and H and W are the height and width. After global average pooling, the H×W spatial information of each channel is compressed into a scalar 1×1×C feature map. Then, through an adaptive one-dimensional convolution, local cross-channel interaction analysis is performed on this feature map vector. The convolution kernel of this convolution layer is a dynamic convolution kernel, and the size of the one-dimensional convolution kernel is automatically determined according to the current number of channels C by the following formula:

[0084]

[0085] where the convolution kernel size is k and the number of channels is C;

[0086] The convolution kernel performs local interaction analysis on the scalar values of k adjacent channels, learns the importance between different channels, and adaptively generates the attention weights of the channels. Finally, the feature containing channel attention Figure 1 ×1×C and the original input feature map H×W×C are multiplied channel by channel. All spatial positions of each channel are multiplied by the same weight value to achieve feature selection, and finally a feature map H with channel attention is output ′ ×W ′ ×C. Finally, the three branches output features with different weights to the connection layer;

[0087] The fifth branch transmits the input of the module to the pooling layer. The pooling layer compresses the time dimension of the input signal for global statistics and calculates the mean value of each channel. The output of the pooling layer is transmitted to the 1×1 convolution layer 2. The convolution layer performs cross-channel interaction on the pooled features to generate channel weights. The output of the convolution layer is transmitted to the upsampling layer to align the global features with the local convolution features. The upsampling layer transmits the features to the connection layer. The connection layer fuses the obtained features and transmits them to the 1×1 convolution layer 3. After calculating the features by the convolution layer 3, they are output by the output layer;

[0088] Step 3.3: The decoder consists of a transposed convolution, an activation function, and an upsampling layer. The decoding part gradually expands the signal dimension through transposed convolution and upsampling. Let z be the latent representation output by the encoder, and use the transposed convolution kernel W (d) and the bias term b (d) to perform deconvolution operations:

[0089] k (L) =F(∑DeConv1D(z,w(d) ) + b (d) ) (7)

[0090] Among them, DeConv1D() is the transposed convolution operation, and k (L) is the output part of the transposed convolution of the Lth layer. The output signal usually also undergoes a non-linear mapping through an activation function to ensure that the output of the signal is within the effective range. The activation function 4 and the activation function 5 select the Relu activation function;

[0091] Through upsampling, the network restores the time resolution of the signal during the decoding process. Finally, the LeakyRelu function improves the ability to capture the bidirectional features of the signal and enhances the integrity of the denoised signal by retaining the negative gradient and signal information. The function formula is as follows:

[0092]

[0093] Among them, α takes 0.01;

[0094] Step 3.4: The main path of the residual connection is composed of convolutional layer 3, activation function 3, max pooling layer 3, ASPP-ECA module, transposed convolution 1, activation function 4, and upsampling 1. The skip connection line connects the output of the max pooling layer 2 of the second small encoding structure in the encoder to the residual connection block after the first small decoding structure in the decoder. At the residual connection block, the output of upsampling 1 at the main path end is combined with the output of the max pooling layer 2. The residual connection block is composed of a 1×1 convolution and a fusion module connected in sequence. In the residual connection block, the skip connection line is connected to the fusion module through a 1×1 convolution. The fusion module receives the output of the 1×1 convolutional layer and the output of upsampling 1, and fuses the data of the two branches, and the fused result is used as the overall output;

[0095] Step 4: Input the training set and the validation set into the wind turbine rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder described in Step 3 for training, which specifically includes Steps 4.1 to 4.4:

[0096] Step 4.1: Set the training parameters of the wind turbine rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder. The model training parameters include: learning rate, convolution kernel size, optimizer, and number of iteration rounds;

[0097] In this embodiment, the initial learning rate is 0.0003, the optimizer optimizer is Adam, the convolution kernel size is 3×1, and the number of iteration rounds Epoch is 360;

[0098] Step 4.2: Input the noisy signal samples of the training set and the validation set into the wind turbine bearing vibration signal denoising model of the multi-scale attention residual convolutional autoencoder. Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters. It is a core algorithm for training neural networks, mainly used to calculate the gradient of the loss function with respect to the model parameters (such as the weights and biases of the convolutional kernels, etc.). In the convolutional autoencoder model, first, the input signal is subjected to feature extraction through the encoder, then reconstructed into the output signal through the decoder, and then the difference between the reconstructed signal and the original input signal is calculated through the loss function. The backpropagation algorithm will, based on this loss function, start from the output layer and calculate the gradient of each parameter layer by layer in the reverse direction. These gradients reflect the influence degree of each parameter on the loss function, enabling the model to update the parameters in the direction of reducing the loss;

[0099] Step 4.3: After calculating the gradients of the model parameters, use the Adam optimizer to update the model parameters. According to the gradient information of the parameters, the optimizer will update the parameters in the opposite direction of the gradient. By continuously iteratively updating the model parameters, the value of the loss function can be gradually reduced. By minimizing the loss function, the values of the model parameters are adjusted to gradually approach the optimal solution until the loss functions of the training set and the validation set no longer decrease, and at the same time, the evaluation index signal-to-noise ratio SNR no longer increases, and the root mean square error RMSE and the mean absolute error MAE no longer decrease;

[0100] Step 4.4: Save the trained model parameters as the optimal model;

[0101] Step 5: Use the test set to test the optimal model described in Step 4, and evaluate the test results of the test set. If the accuracy requirements are met, the final wind turbine rolling bearing vibration signal denoising model based on the multi-scale attention residual convolutional autoencoder is obtained, specifically including Steps 5.1 to 5.4:

[0102] Step 5.1: Input the test set into the optimal model described in Step 4;

[0103] Step 5.2: Calculate the model performance indicators: signal-to-noise ratio SNR, root mean square error RMSE, and mean absolute error MAE. The specific calculation formulas are as follows:

[0104]

[0105]

[0106]

[0107] where SNR is the signal-to-noise ratio, RMSE is the root mean square error, MAE is the mean absolute error, S i is the original signal, S′i denotes the signal output after noise reduction, and n is the number of signal samples;

[0108] Step 5.3: When the performance index meets the accuracy requirement, obtain the final noise reduction model of the fan rolling bearing vibration signal based on the multi-scale attention residual convolutional autoencoder.

[0109] In this embodiment, in order to verify the effect of the improved model proposed by the present invention, the RAECAE model, basic autoencoder (AE), sparse autoencoder (SAE), denoising autoencoder (DAE), and convolutional autoencoder (CAE) proposed in this invention patent are used to test on the bearing dataset of Case Western Reserve University, and the data in the dataset is constructed into a noisy signal of 0 dB. Calculate the evaluation indexes for the five models respectively. The evaluation index data is shown in Table 1. The smaller the mean absolute error and root mean square error, the less different from the original signal, and the higher the signal-to-noise ratio, the better the noise removal effect.

[0110] Table 1 Comparison experiment results

[0111]

[0112] As can be seen from Table 1, compared with the original convolutional autoencoder model, the noise reduction model of the fan rolling bearing vibration signal disclosed in the present invention reduces by 0.0103 in terms of the mean absolute error MAE, reduces by 0.0372 in terms of the root mean square error RMSE, and increases by 1.248 in terms of the signal-to-noise ratio SNR. The method of this model can better remove noise and restore signal features, which is beneficial to accurately judge the health status of the fan rolling bearing subsequently.

[0113] The above is only one embodiment of the present invention, and it does not limit the patent scope of the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A noise reduction method for vibration signals of a fan rolling bearing based on an improved convolutional autoencoder, characterized in that, Specifically, it includes the following steps: Step 1: Obtain the vibration signal of the fan bearing to form a first data set; in the first data set, the vibration signal of the fan bearing can be obtained from a public data set, collected by a vibration sensor for the vibration signal, or obtained from a data acquisition system; Step 2: Add noise to the vibration signals in the first data set to form a second data set, and divide the second data set into a training set, a validation set, and a test set; Step 3: Construct a noise reduction model for the vibration signal of the fan rolling bearing based on a multi-scale attention residual convolutional autoencoder. The model includes an encoder, an ASPP-ECA module, a decoder, and a residual connection block. The construction of the model further includes steps 3.1 to 3.4: Step 3.1: The encoder consists of a convolutional layer 1, an activation function 1, a max pooling layer 1, a convolutional layer 2, an activation function 2, a max pooling layer 2, a convolutional layer 3, an activation function 3, and a max pooling layer 3 connected in sequence. Among them, the convolutional layer, the activation function, and the max pooling layer are connected in sequence to form a small encoder layer, and there are three small encoders with the same structure connected in sequence in the encoder; among them, the activation function 1, the activation function 2, and the activation function 3 select the Relu function, and the calculation formula is as follows: F(a) = max{0, a} (1) where a is the activation value; Step 3.2: The ASPP-ECA module is between the encoder and the decoder. This module combines the atrous spatial pyramid pooling module ASPP and the channel attention ECA module. The ASPP-ECA module includes an input layer, a 1×1 convolutional layer 1, three 1×3 dilated convolutions with different dilation rates, an attention module after each dilated convolution, a pooling layer, a 1×1 convolutional layer 2, an upsampling layer, a connection layer, a 1×1 convolutional layer 3, and an output layer; The ASPP-ECA module is divided into 5 branches, among which, The first branch passes through the 1×1 convolutional layer 1 from the module input layer and is transmitted to the connection layer; The second branch, the third branch, and the fourth branch respectively pass through the dilated convolutions with three different dilation rates from the module input layer, and then enter the ECA module respectively. The ECA module adaptively adjusts the weights of different features and then transmits them to the connection layer; The fifth branch passes through the pooling layer, the 1×1 convolutional layer 2, and the upsampling layer in sequence from the module input layer, and then transmits the channel weights to the connection layer; After the connection layer receives the outputs of the five branches, it transmits the data to the 1×1 convolutional layer 3, and then transmits it to the output layer for output; Step 3.3: The decoder consists of a transposed convolution 1, an activation function 4, an upsampling 1, a residual connection block, a transposed convolution 2, an activation function 5, an upsampling 2, a transposed convolution 3, an upsampling 3, and a LeakyRelu function connected in sequence. Among them, the transposed convolution, the activation function, and the upsampling layer are connected in sequence to form a small decoder layer, and there are two small decoders with the same structure connected in sequence in the decoder. The third small decoder is composed of the transposed convolution 3, the upsampling 3, and the LeakyRelu function connected in sequence; the LeakyRelu function formula is as follows: where α takes 0.01; The main path of the residual connection is composed of convolutional layer 3, activation function 3, max pooling layer 3, ASPP-ECA module, transposed convolution 1, activation function 4, and upsampling 1. The skip connection line connects the output of the max pooling layer 2 in the second small encoding structure in the encoder to the residual connection block after the first small decoding structure in the decoder. At the residual connection block, the output of upsampling 1 at the main path end is combined with the output of the max pooling layer 2. The residual connection block consists of a 1×1 convolution and a fusion module connected in sequence. In the residual connection block, the skip connection line is connected to the fusion module through a 1×1 convolution. The fusion module receives the output of the 1×1 convolutional layer and the output of upsampling 1, fuses the data of the two branches, and outputs the result as a whole; Step 4: Use the training set and the validation set to train the noise reduction model of the fan rolling bearing vibration signal based on multi-scale attention residual convolution auto-encoding, and save the trained model as the optimal model. It further includes steps 4.1 to 4.4: Step 4.1: Set the training parameters of the noise reduction model of the fan rolling bearing vibration signal based on multi-scale attention residual convolution auto-encoding. The model training parameters include: learning rate, convolution kernel size, optimizer, number of iterations; Step 4.2: Input the noisy signal samples of the training set and the validation set into the noise reduction model of the fan bearing vibration signal based on multi-scale attention residual convolution auto-encoding. Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and adjust the model parameters by minimizing the loss function to gradually approach the optimal solution; Step 4.3: Use the Adam optimizer to update the model parameters, so that the model parameters are updated in the direction of gradient descent until the loss functions of the training set and the validation set no longer decrease, and at the same time, the signal-to-noise ratio SNR of the evaluation index no longer increases, and the root mean square error RMSE and the mean absolute error MAE no longer decrease; Step 4.4: Save the trained model parameters as the optimal model; Step 5: Use the test set to test the optimal model, evaluate the test results of the test set, and if the accuracy requirements are met, the final noise reduction model of the fan rolling bearing vibration signal based on multi-scale attention residual convolution auto-encoding is obtained.

2. A noise reduction method for the vibration signal of a fan rolling bearing based on an improved convolutional autoencoder according to claim 1, characterized in that In step 2, the training set, validation set, and test set are divided in the ratio of 7:2:

1.

3. A noise reduction method for the vibration signal of a fan rolling bearing based on multi-scale attention residual convolution auto-encoder according to claim 1, characterized in that, Step 5 further includes steps 5.1 to 5.3: Step 5.1: Input the test set into the optimal model described in step 4; Step 5.2: Calculate the model performance indicators: signal-to-noise ratio SNR, root mean square error RMSE, mean absolute error MAE. The specific calculation formulas are as follows: Among them, SNR is the signal-to-noise ratio, RMSE is the root mean square error, MAE is the mean absolute error, S i is the original signal, S′ i represents the signal output after noise reduction, and n is the number of signal samples; Step 5.3: When the performance indicators meet the accuracy requirements, the final noise reduction model of the fan rolling bearing vibration signal based on multi-scale attention residual convolution auto-encoding is obtained.

Images