A fan rolling bearing vibration signal denoising method based on an improved convolutional autoencoder

Abstract

The application discloses a fan rolling bearing vibration signal denoising method based on an improved convolutional autoencoder, and specifically comprises the following steps: (1) establishing a fan bearing vibration signal dataset; (2) adding noise to the bearing vibration signals in the dataset to form noisy signals; (3) constructing a fan rolling bearing vibration signal denoising model based on a multi-scale attention residual convolutional autoencoder; (4) training the model using a training set and a validation set, and saving the trained model as an optimal model; (5) testing the model using a test set, and the test result meets the accuracy requirement, that is, the final fan rolling bearing vibration signal denoising model is obtained. Compared with the prior art, the fan rolling bearing vibration signal denoising method based on the improved convolutional autoencoder can adaptively adjust the weight of extracted features, improves the feature extraction and denoising capability of the model, and improves the convergence speed of the model.

Description

TECHNICAL FIELD

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

[0002] As an important component of wind turbines, the health status of rolling bearings has a crucial impact on the safety and availability of wind turbines. For this reason, most wind turbines have applied a fan condition monitoring system (CMS). The detection system mainly uses various monitoring technologies to manage the health of wind turbines, such as vibration monitoring, which collects internal bearing vibration data through sensors, including using low-frequency accelerometers to monitor main bearings and high-frequency accelerometers to monitor generator bearings. However, wind turbines are often in harsh working environments, and the signals detected by the sensors are subject to varying degrees of noise interference, which distorts the signal or masks the true fault characteristics, affecting the judgment of the bearing working condition and the accuracy of fault diagnosis. Therefore, denoising the fan bearing vibration signal is an important step in monitoring the health of wind turbines.

[0003] Traditional signal denoising methods mainly include wavelet transform, empirical mode decomposition, and variational mode decomposition methods. Traditional signal denoising methods mostly have three main steps: first, decompose the signal into multiple components using signal decomposition methods; second, use different threshold functions to denoise the decomposed signal; and third, select the signal component with the most obvious features to reconstruct. However, these denoising methods have weaknesses such as weak feature extraction ability and poor generalization, and the selection of hyperparameters and threshold functions can affect the denoising performance, so more intelligent methods are 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. Deep learning-based vibration signal denoising methods use large amounts of data and the powerful learning ability of models to learn the characteristics and distribution of vibration signals, thereby achieving a denoising effect close to the original signal.

[0005] However, current deep learning methods still have certain limitations. First, when using a single scale convolution kernel to process vibration signals for feature extraction, some key information may be lost, and the feature extraction ability is weak. Second, multi-scale feature extraction methods do not fully consider the weight difference of different scale features in the signal feature extraction process, affecting the final feature extraction effect and the adaptability of denoising. SUMMARY

[0006] In view of the problems in the prior art, the application provides a fan rolling bearing vibration signal denoising method based on a multi-scale attention residual convolution autoencoder, the model of the convolution autoencoder is improved, multi-scale feature extraction methods and attention residual links are used to adaptively adjust the feature weights and extract multi-scale features, the problems of low adaptability and weak feature extraction capability of the vibration signal denoising are solved, and the convergence speed of the model is accelerated.

[0007] The technical scheme provided by the application comprises the following steps:

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

[0009] Further, in step 1, the fan bearing vibration signal can be obtained by using a public data set, collecting the vibration signal by using a vibration sensor, or obtaining the vibration signal from a data acquisition system.

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

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

[0012] Step 3: constructing a fan rolling bearing vibration signal denoising model based on a multi-scale attention residual convolution autoencoder;

[0013] Further, step 3 specifically comprises steps 3.1 to 3.4:

[0014] Step 3.1: converting the convolution layer of the encoder in the convolution autoencoder model into a one-dimensional convolution, and the improved encoder is composed of three one-dimensional convolution layers, three activation functions and three maximum pooling layers, and the input signal x=(x1, x2, …x N ) is converted into a low-dimensional latent space representation, wherein x1, x2, …x N is a sample value sequence, and the 3x1 convolution operation of the convolution layer is:

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

[0016] Wherein, X (L) is the output of the L-layer convolution layer, x (L-1) is the input of the convolution layer, W (L) is the convolution kernel, and b (L)is the bias term, Conv1D() is one-dimensional convolution calculation, F() is an activation function, and the activation functions 1, 2 and 3 in the encoder are selected to be the Relu function for nonlinear transformation, and the calculation formula is:

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

[0018] wherein a is an activation value;

[0019] The pooling layer is used for downsampling to improve the calculation efficiency by reducing the data dimension. In the encoder of the present model, the maximum pooling is used. The input signal length of a one-dimensional sequence X is N, the pooling window size is k, and the step is s. The calculation formula of each element of the output sequence Y is:

[0020]

[0021] wherein X represents an input one-dimensional signal with a length of N in , Y is an 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 which is zero.

[0022] Step 3.2: Combine the ASPP and ECA modules, add an ECA module after each dilated convolution to form an ASPP-ECA module, and add the ASPP-ECA module between the encoder and the decoder in the convolutional autoencoder;

[0023] The further ASPP-ECA module has five branches. The first branch is to extract the high-frequency components of the signal through a 1x1 convolution layer 1 from the input layer of the module, and transmit the signal high-frequency features to the connection layer.

[0024] The second branch, the third branch and the fourth branch respectively contain three dilated convolutions with different dilated rates and ECA modules. The dilated rates of the three dilated convolutions are 4, 8 and 12 respectively. The effective coverage of the convolution kernel can be expanded through the medium dilated rate to capture the characteristics of the medium time scale, and the attention module is used to weight the extracted features of each scale to adjust the weight of different features. The ECA module has a global average pooling and an adaptive one-dimensional convolution layer. The module receives the feature map of the dilated convolution, and the dimension of the feature map is set to HxWxC, C is the number of channels, H and W are the height and width. After the global average pooling, the HxW spatial information of each channel is compressed into a 1x1xC feature map, and then the local cross-channel interaction analysis is performed on the feature map vector through the adaptive one-dimensional convolution. 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 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 locally interacts with the scalar values of the k adjacent channels, learns the importance between different channels, and adaptively generates the attention weight of the channel. Finally, the feature containing the channel attention is multiplied by the original input feature map HxWxC, and all spatial positions of each channel are multiplied by the same weight value to realize feature selection. Finally, the feature map H'xW'xC with channel attention is output, and finally, the three branches output the features with different weights to the connection layer. Figure 1

[0028] The fifth branch is to transmit the input of the module to the pooling layer, which compresses the time dimension of the input signal and performs global statistics to calculate the mean of each channel. The output of the pooling layer is transmitted to the 1x1 convolution layer 2, which interacts between channels to generate channel weights. The output of the convolution layer is transmitted to the up-sampling layer to align the global features with the local convolution features. The up-sampling layer transmits the features to the connection layer, which fuses the obtained features and transmits them to the 1x1 convolution layer 3. After the convolution layer 3 calculates the features, the output layer outputs them.

[0029] Step 3.3: The decoder is composed of transpose convolution, activation function, and up-sampling layer. The decoding part gradually expands the signal dimension through transpose convolution and up-sampling. Let z be the latent representation output by the encoder. Use the transpose 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 transpose convolution operation, k (L) is the output part of the L-layer transpose convolution. The output signal is usually nonlinearly mapped by an activation function to ensure that the output of the signal is within the effective range. Activation function 4 and activation function 5 use the Relu activation function.

[0032] Through up-sampling, the network restores the time resolution of the signal during decoding. Finally, the LeakyRelu function preserves the negative gradient and signal information, improving the ability to capture bidirectional features of the signal and enhancing the integrity of the denoised signal. The function formula is as follows:

[0033] ​

[0034] wherein a takes 0.01;

[0035] Step 3.4: the main path of the residual connection is composed of convolution layer 3, activation function 3, max pooling layer 3, ASPP-ECA module, transpose convolution 1, activation function 4 and up-sampling 1, the output of the max pooling layer 2 of the second small encoding structure in the encoder is connected to the residual connection block after the first small decoding structure in the decoder through the skip connection line, at the residual connection block, the output of the up-sampling 1 at the end of the main path is combined with the output of the max pooling layer 2, the residual connection block is composed of a 1x1 convolution and a fusion module connected in turn, in the residual connection block, the skip connection line is connected to the fusion module through the 1x1 convolution, the fusion module receives the output of the 1x1 convolution layer and the output of the up-sampling 1, and the data of the two branches is fused, and the fusion is taken as the overall output;

[0036] Step 4: training the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolution auto-encoding by using the training set and the validation set, and saving the trained model as an optimal model;

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

[0038] Step 4.1: setting the training parameters of the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolution auto-encoding, the model training parameters including: learning rate, channel number, convolution kernel size, optimizer, iteration number;

[0039] Step 4.2: inputting the noisy signal samples in the training set and the validation set into the multi-scale attention residual convolution auto-encoding fan bearing vibration signal denoising model, using the back propagation algorithm to calculate the gradient of the loss function on the model parameters, and adjusting the model parameters by minimizing the loss function to gradually approach the optimal solution;

[0040] Step 4.3: using the optimizer Adam to update the model parameters, so that the model parameters are updated in the direction of gradient descent, until the loss function of the training set and the validation set no longer decreases, and the evaluation index signal-to-noise ratio SNR also no longer improves, and the root mean square error EMSE and the mean absolute error MAE also no longer become smaller;

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

[0042] Step 5: testing the optimal model by using the test set, evaluating the test set test results, meeting the accuracy requirement, and obtaining the final fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolution auto-encoding;

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

[0044] Step 5.1: input the test set into the optimal model of step 4;

[0045] 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 formula is as follows:

[0046]

[0047]

[0048]

[0049] Wherein, 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 indicates the output signal after noise reduction, n is the number of signal samples;

[0050] Step 5.3: when the performance indicators meet the accuracy requirements, the final fan rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder is obtained.

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

[0052] (1) In the encoder and decoder of the convolutional autoencoder, the ASPP-ECA module disclosed in the present application is adopted, and dilated convolution with different dilated rates is used, which can increase the receptive field without increasing the number of parameters, extract multi-scale features of the signal, and weight the channels in different branches, adaptively adjust the weight of each convolution kernel to extract features, and improve the signal feature extraction ability of the model and the adaptability to noise intensity.

[0053] (2) The residual connection is introduced on the convolutional autoencoder, which reduces the learning difficulty of the model, effectively speeds up the learning of the signal features of the model, and improves the convergence speed of the model.

[0054] (3) The last small decoding structure of the convolutional autoencoder is improved, the LeakyRelu function is used instead of the original activation function and added after the last upsampling layer, which effectively avoids information loss, enhances the nonlinear expression ability of the model, effectively restores the original features of the signal, improves the signal-to-noise ratio of the signal, and makes the power of the signal in the data much higher than the power of the noise, achieving the effect of denoising the signal data. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1A flow chart of a fan rolling bearing vibration signal denoising method based on an improved convolutional autoencoder according to the present application;

[0056] Figure 2 A fan rolling bearing vibration signal denoising model structure schematic diagram based on a multi-scale attention residual convolutional autoencoder according to the present application;

[0057] Figure 3 An ASPP-ECA structure schematic diagram;

[0058] Figure 4 An ECA module structure schematic diagram;

[0059] Figure 5 A residual connection block structure schematic diagram; DETAILED DESCRIPTION

[0060] In order to make the technical solutions, structural features, purposes achieved and advantages of the present application clearer and more apparent, the present application will be further described in detail below in conjunction with specific embodiments and the accompanying drawings. It should be noted that the specific embodiments described herein are only used to more clearly explain the present application and do not limit the present application.

[0061] Figure 1 A fan rolling bearing vibration signal denoising method flow chart based on a multi-scale attention residual convolutional autoencoder according to the present application, the implementation process is as follows:

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

[0063] In this embodiment, in order to better evaluate the denoising effect of the fan rolling bearing vibration signal denoising method based on a multi-scale attention residual convolutional autoencoder according to the present application, the public data set Case Western Reserve University bearing data set is adopted; the Case Western Reserve University bearing data set is obtained by the United States Case Western Reserve University in a test bench composed of a 2-horsepower motor, a torque sensor, and a power tester, an acceleration sensor is placed above the bearing seat to collect the vibration acceleration signal of the faulty bearing, the sampling frequency is 12KHz, the bearing is damaged by electric spark machining single point, the data set contains damage at three different positions of the bearing outer ring, inner ring and ball, and three different degrees of damage of 0.007 inches (0.1778mm), 0.014 inches (0.3556mm) and 0.021 inches (0.5334mm). Here, the sampling frequency is selected as 12KHz, the inner ring raceway fault signal of the deep groove ball rolling bearing is selected, the damage diameter is selected as 0.007 inches (0.1778mm), and the first data set is formed;

[0064] Step 2: Add noise to the signal in the first dataset and divide it into smaller samples to form the second dataset;

[0065] To construct a noisy signal with a signal-to-noise ratio (SNR) of 0 dB, firstly, the signal power, i.e., the signal variance, is calculated using a function. Secondly, the SNR, expressed in decibels (dB), is converted to a linear proportional form. The calculation formula is as follows:

[0066]

[0067] Among them, SNR Linear It is a linear scaling factor of the signal-to-noise ratio, SNR. dB It is the signal-to-noise ratio in decibels;

[0068] Then, according to the definition of signal-to-noise ratio (SNR), the noise power is calculated using the signal power and the linearly proportional SNR. The calculation formula is as follows:

[0069]

[0070] Among them, noise poewr The noise power is denoted as 'noise_power', and the signal power is denoted as 'signal_power'. By taking the square root of the noise power, we can obtain the standard deviation of the noise. Using the noise standard deviation and the shape of the original signal as parameters of the random number function, we ensure that the noise and signal have the same shape. We can add them element by element. The random number function generates a series of random numbers as noise values. The generated random numbers follow a Gaussian distribution. Adding the noise values ​​to the original signal yields the noisy signal.

[0071] The .mat format signal files in the first dataset described in this embodiment are converted to the npy format signal files required by the numpy library in Python. The signal files in the first dataset are divided into small samples, each containing 2048 continuous sampling points, for a total of 2411 small samples, forming the second dataset. The second dataset is then divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The resulting 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 signal of the wind turbine rolling bearing based on multi-scale attention residual convolutional autoencoder. The structure of the improved signal denoising model is as follows: Figure 2 As shown, the model construction process specifically includes steps 3.1 to 3.3:

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

[0074] Further encoding part converts input signal x = (x1, x2, … x N ) into a low-dimensional latent space representation, where x1, x2, … x N are acceleration value sequences, and the 3x1 convolution operation of the convolution layer is:

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

[0076] Where X (L) is the output of the L-layer convolution layer, x (L-1) is the input of the convolution layer, W (L) is the convolution kernel, b (L) is the bias term, Conv1D() is one-dimensional convolution calculation, F() is the activation function, and the activation function 1, the activation function 2, and the activation function 3 in the encoder are selected as the RELU function as the activation function for nonlinear transformation, where a is the activation value.

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

[0078] The pooling layer is used to reduce the sampling, and the calculation efficiency is improved by reducing the data surrounding. The maximum pooling operation is used in the encoder of the model, the input signal length is N one-dimensional sequence X, the pooling window size is k, and the step is s, then the calculation formula of each element of the output sequence Y is:

[0079]

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

[0081] Step 3.2: Combine the ASPP and ECA modules, add an ECA module after each dilated convolution to form an ASPP-ECA module as shown in Figure 3 , and add the ASPP-ECA module between the encoder and the decoder in the convolution autoencoder;

[0082] The further ASPP-ECA module has 5 branches, the first branch is from the input layer of the module through the 1x1 convolution layer 1, extracts the high-frequency components in the signal, and transmits the signal high-frequency features to the connection layer.

[0083] The second branch, the third branch and the fourth branch respectively comprise three expansion convolutions with different expansion rates and an ECA module. The three expansion convolutions have expansion rates of 4, 8 and 12 respectively, which can expand the effective coverage of the convolution kernel through the medium expansion rate, capture the characteristics of the medium time scale, and weight the extracted characteristics of each scale by the ECA attention module to adjust the weights of different characteristics. The ECA module is shown in Figure 4 The feature map received by the expansion convolution has a dimension of HxWxC, where C is the number of channels, and H and W are the height and width. The HxW spatial information of each channel is compressed into a 1x1xC feature map through global average pooling. Then, the feature map vector is analyzed for local cross-channel interaction through adaptive one-dimensional convolution. 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 according to the following formula:

[0084]

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

[0086] The convolution kernel performs local interaction analysis on the scalar values of the k adjacent channels, learns the importance between different channels, and adaptively generates the attention weights of the channels. Finally, the feature map containing channel attention Figure 1 1x1xC and the original input feature map HxWxC are multiplied channel by channel, and all spatial positions of each channel are multiplied by the same weight value to achieve feature selection. Finally, the feature map with channel attention H ′ W ′ C is output.

[0087] The fifth branch is to transmit the input of the module to the pooling layer. The pooling layer compresses the time dimension of the input signal and performs global statistics to calculate the mean value of each channel. The output of the pooling layer is transmitted to the 1x1 convolution layer 2. The convolution layer performs inter-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 1x1 convolution layer 3. After the convolution layer 3 calculates the features, the output layer outputs the features.

[0088] Step 3.3: The decoder is composed of transpose convolution, activation function and upsampling layer. The decoding part gradually expands the signal dimension through transpose convolution and upsampling. Let z be the latent representation output by the encoder. The inverse convolution operation is performed using the transpose convolution kernel W (d) and the bias term b (d) :

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

[0090] wherein DeConv1D() is a transpose convolution operation, k (L) The output signal is usually also mapped by an activation function for nonlinear mapping to ensure that the output of the signal is within the effective range. The activation function 4 and the activation function 5 are selected as a Relu activation function.

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

[0092]

[0093] wherein a is 0.01.

[0094] Step 3.4: The main path of the residual connection is composed of a convolution layer 3, an activation function 3, a max-pooling layer 3, an ASPP-ECA module, a transpose convolution 1, an activation function 4 and an upsampling 1. The output of the max-pooling layer 2 of the second small coding structure in the encoder is connected to the residual connection block after the first small decoding structure in the decoder through a skip connection line. In the residual connection block, the output of the upsampling 1 at the end of the main path is combined with the output of the max-pooling layer 2. The residual connection block is composed of a 1x1 convolution and a fusion module connected in sequence. In the residual connection block, the skip connection line is connected to the fusion module through the 1x1 convolution. The fusion module receives the output of the 1x1 convolution layer and the output of the upsampling 1, and fuses the data of the two branches. After fusion, the whole output is obtained.

[0095] Step 4: input the training set and the validation set into the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolution auto-encoding in step 3 for training, which specifically includes steps 4.1 to 4.4:

[0096] Step 4.1: set the training parameters of the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolution auto-encoding. The model training parameters include: learning rate, convolution kernel size, optimizer, and iteration round number.

[0097] In this embodiment, the initial learning rate is 0.0003, the optimizer optimizer is Adam, the convolution kernel size is 3x1, and the iteration round number Epoch is 360.

[0098] Step 4.2: input the training set and validation set noise-containing signal samples into the fan bearing vibration signal denoising model of the multi-scale attention residual convolutional autoencoder, and use the back propagation algorithm to calculate the gradient of the loss function with respect to the model parameters, which is a core algorithm for training neural networks and is mainly used to calculate the gradient of the loss function with respect to the model parameters (such as the weights and biases of the convolution kernel). In the convolutional autoencoder model, the input signal is first extracted by the encoder, and then the output signal is reconstructed by the decoder. Then the difference between the reconstructed signal and the original input signal is calculated by the loss function. The back propagation algorithm calculates the gradient of each parameter from the output layer to the input layer, which reflects the influence of each parameter on the loss function, so that the model can update the parameters in the direction of reducing the loss;

[0099] Step 4.3: after calculating the gradient of the model parameters, use the optimizer Adam to update the model parameters. According to the gradient information of the parameters, the optimizer updates the parameters in the opposite direction of the gradient. By continuously iterating 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 function of the training set and the validation set no longer decreases, and the signal-to-noise ratio SNR, the root mean square error EMSE and the mean absolute error MAE no longer improve or decrease;

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

[0101] Step 5: test the optimal model of step 4 using the test set, evaluate the test set test results, and meet the accuracy requirements, i.e. obtain the final fan rolling bearing vibration signal denoising model based on multi-scale attention residual convolutional autoencoder, which includes steps 5.1 to 5.4:

[0102] Step 5.1: input the test set into the optimal model of 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, with the specific calculation formulas as follows:

[0104]

[0105]

[0106]

[0107] Wherein, 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 after noise reduction, n is the number of signal samples;

[0108] Step 5.3: When the performance index meets the accuracy requirement, the final fan rolling bearing vibration signal denoising model based on multi-scale attention residual convolution autoencoder is obtained.

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

[0110] Table 1 Comparison of experimental results

[0111]

[0112] As shown in Table 1, compared with the original convolutional autoencoder model, the fan rolling bearing vibration signal denoising model disclosed in the application reduces the mean absolute error MAE by 0.0103, reduces the root mean square error RMSE by 0.0372, and increases the signal-to-noise ratio SNR by 1.248. The model method can better remove noise and restore signal characteristics, which is conducive to accurately judging the health state of the fan rolling bearing.

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

Claims

1. A method for denoising vibration signals of rolling bearings of a fan based on an improved convolutional autoencoder, characterized in that, Specifically comprising the following steps: Step 1: Obtain the fan bearing vibration signal to form a first data set; In the first data set, the fan bearing vibration signal can be acquired from a public data set, collected by a vibration sensor, or obtained from a data acquisition system; 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; Step 3: Construct a fan rolling bearing vibration signal denoising model based on a multi-scale attention residual convolution autoencoder, the model comprising an encoder, an ASPP-ECA module, a decoder and a residual connection block, the construction of the model further comprising steps 3.1 to 3.4: Step 3.1: The encoder is composed of sequentially connected convolution layer 1, activation function 1, max pooling layer 1, convolution layer 2, activation function 2, max pooling layer 2, convolution layer 3, activation function 3 and max pooling layer 3, wherein the convolution layer, the activation function and the max pooling layer are sequentially connected to form a small encoder, and there are three small encoders with the same structure in the encoder; wherein 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) Wherein, a is the activation value; Step 3.2: The ASPP-ECA module is between the encoder and the decoder, which is a combination of the Atrous Spatial Pyramid Pooling module ASPP and the Channel Attention ECA module, and the ASPP-ECA module includes an input layer, a 1×1 convolution layer 1, three 1×3 dilated convolutions with different dilated rates, an attention module after each dilated convolution, a pooling layer, a 1×1 convolution layer 2, an up-sampling layer, a connection layer, a 1×1 convolution layer 3 and an output layer; The ASPP-ECA module is divided into five branches, wherein, The first branch transmits from the module input layer through the 1×1 convolution layer 1 to the connection layer; The second branch, the third branch and the fourth branch respectively pass through the dilated convolution with three different dilated rates from the module input layer, and then enter the ECA module respectively, and the ECA module adaptively adjusts the weights of different features, and then transmits to the connection layer; The fifth branch transmits from the module input layer through the pooling layer, the 1×1 convolution layer 2 and the up-sampling layer, and then transmits the channel weight to the connection layer; The connection layer receives the output of the five branches and transmits the data to the 1×1 convolution layer 3, and then transmits to the output layer for output; Step 3.3: The decoder is composed of sequentially connected transposed convolution 1, activation function 4, up-sampling 1, residual connection block, transposed convolution 2, activation function 5, up-sampling 2, transposed convolution 3, up-sampling 3 and LeakyRelu function, wherein the transposed convolution, the activation function and the up-sampling layer are sequentially connected to form a small decoder, and there are two small decoders with the same structure in the decoder, and the third small decoder is composed of transposed convolution 3, up-sampling 3 and LeakyRelu function; The formula of LeakyRelu function is as follows: Wherein, α is 0.01; The main path of the residual connection is composed of a convolutional layer 3, an activation function 3, a maximum pooling layer 3, an ASPP-ECA module, a transpose convolution 1, an activation function 4 and an up-sampling 1; the output of the maximum pooling layer 2 of the second small coding structure in the encoder is connected to the residual connection block after the first small decoding structure in the decoder through a skip connection line; at the residual connection block, the output of the up-sampling 1 at the end of the main path is combined with the output of the maximum 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 the 1*1 convolution; the fusion module receives the output of the 1*1 convolution layer and the output of the up-sampling 1, and fuses the data of the two branches; after fusion, the output is taken as the overall output; Step 4: training the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolutional auto-encoder using the training set and the validation set, saving the trained model as an optimal model, further comprising steps 4.1 to 4.4: Step 4.1: setting the training parameters of the fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolutional auto-encoder, the model training parameters including: learning rate, convolution kernel size, optimizer, and iteration number; Step 4.2: inputting the noisy signal samples in the training set and the validation set into the multi-scale attention residual convolutional auto-encoder fan bearing vibration signal denoising model, using the back propagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and adjusting the model parameters by minimizing the loss function to gradually approach the optimal solution; Step 4.3: using the optimizer Adam to update the model parameters, updating the model parameters in the direction of gradient descent until the loss function of the training set and the validation set no longer decreases, and the evaluation indicators signal-to-noise ratio SNR, root mean square error EMSE and mean absolute error MAE no longer improve or decrease; Step 4.4: saving the trained model parameters as an optimal model; Step 5: testing the optimal model using the test set, evaluating the test set test results, and meeting the accuracy requirements to obtain the final fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolutional auto-encoder.

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

1.

3. The fan rolling bearing vibration signal denoising method based on the improved convolutional autoencoder according to claim 1, characterized in that, Step 5 further comprises steps 5.1 to 5.3: Step 5.1: inputting the test set into the optimal model of step 4; Step 5.2: calculating the model performance indicators: signal-to-noise ratio SNR, root mean square error RMSE, and mean absolute error MAE, with the specific calculation formulas as follows: Wherein, SNR is signal noise ratio, RMSE is root mean square error, MAE is mean absolute error, S i is the original signal, S' i represents the output signal after noise reduction, n is the number of signal samples; Step 5.3: when the performance indicators meet the accuracy requirements, the final fan rolling bearing vibration signal denoising model based on the multi-scale attention residual convolutional auto-encoder is obtained.

Images