Rolling bearing fault diagnosis method based on wavelet packet transformation-CEEMDAN and ECA-1DCNN

Abstract

The invention discloses a rolling bearing fault diagnosis method based on wavelet packet transformation-CEEMDAN and ECA-1DCNN, and belongs to the technical field of bearing fault diagnosis. The method comprises the following steps: firstly, carrying out noise reduction processing on an original signal containing multiple noise components in a wavelet packet transformation mode, and reconstructing the signal according to the magnitude of an energy value of each frequency band; secondly, performing CEEMDAN decomposition to obtain a plurality of IMF components, and screening proper signal components in combination with a kurtosis value-correlation coefficient criterion to perform signal reconstruction again; and finally, the processed signal is put into an ECA-1DCNN model for training, and a final diagnosis result is obtained. Experiments prove that vibration signals with strong background noise can be effectively filtered through the method, and the accuracy of the bearing fault diagnosis method is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN. Background Art

[0002] Rolling bearings are important components in rotating mechanical equipment, and their normal operation is crucial to the reliability and safety of the equipment. However, during long-term use, rolling bearings are susceptible to factors such as wear, fatigue and impurities, and may fail. Rolling bearing failures can cause abnormalities such as vibration and noise, and in severe cases may also cause equipment damage and production accidents. In order to ensure the high reliability of motor bearings during operation, it is of great significance to maintain rapid and accurate detection of fault types and fault severity. Motor bearings are accompanied by friction, vibration and impact during operation, and the vibration signal exhibits nonlinear and non-stationary characteristics. Under the interference of various noises, it is difficult to accurately and effectively extract the vibration signal characteristics of motor bearings, which is not conducive to the pre-prevention and post-maintenance of rotating equipment. Therefore, before accurately extracting the fault characteristic signal of the motor bearing, the original signal must be denoised to highlight the fault characteristic information, which has also become the key to fault diagnosis.

[0003] Time domain signal analysis is one of the earliest methods in the field of fault diagnosis. It can effectively extract fault feature signals and is also the most commonly used method in nonlinear signal analysis. As a time domain analysis method, wavelet transform performs signal analysis by obtaining dimensional feature parameters and dimensionless parameters. In the process of using wavelet denoising, different wavelet bases and decomposition layers will result in different denoising results, but lack adaptability. Empirical mode decomposition (EMD) is also the most widely used analysis algorithm in nonlinear signal analysis. Traditional empirical mode decomposition is prone to problems such as mode aliasing and false intrinsic mode components during signal decomposition. To address this problem, scholars proposed ensemble empirical mode decomposition (EEMD) based on traditional mode decomposition. This method adds white noise to the input signal and has continuity at different scales, which can effectively avoid mode aliasing. However, EEMD decomposes each motor bearing vibration signal independently, which may cause different numbers of decomposition results of different constructed signals, and does not fundamentally solve the problem of mode aliasing. Torres et al. proposed a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), which not only improved the decomposition effect but also improved the completeness of the decomposition compared with the EEMD method.

[0004] WU et al. integrated the attention mechanism into the multi-scale CNN, effectively fused the multi-scale features, and improved the noise resistance of the model. WANG et al. combined the attention mechanism with the bidirectional long short-term memory neural network (BidirectionalLSTM) to achieve the temporal expression of features and assign different weights to highlight the impact of important features on the classification results. However, the attention mechanism mentioned in the above research methods is essentially a global communication dimensionality reduction attention mechanism. All its channels participate in the nonlinear transformation of the fully connected layer, which reduces the accuracy of the attention mechanism and weakens the weight allocation of important channels. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned deficiencies in the prior art, effectively filter out vibration signals with strong background noise, and improve the accuracy of bearing fault diagnosis.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN, comprising:

[0007] Step 1: Collect vibration signals of rolling bearings under different working conditions;

[0008] Step 2: Perform wavelet packet decomposition on the vibration signals under different working conditions and superimpose the signals;

[0009] Step 3: Perform CEEMDAN decomposition on the superimposed signal and screen the IMF components again for reconstruction;

[0010] Step 4: Input the processed signal into the ECA-1DCNN model for iterative training to optimize the network model parameters for outputting the diagnosis results.

[0011] Different working conditions include a normal fault-free state and a fault state, and the fault types include faults of the inner ring, outer ring and rolling element corresponding to bearings of different specifications; the diagnosis result is a fault-free state or different fault types.

[0012] The step 2 comprises:

[0013] Step 2.1: Perform one-dimensional third-order wavelet packet transform to obtain several WPT signal components;

[0014] Step 2.2: Calculate the energy value of each WPT signal component;

[0015] Step 2.3: Select signal components with energy values ​​greater than the average value for superposition reconstruction.

[0016] The screening of IMF components includes screening according to the kurtosis value K and the correlation coefficient r.

[0017] Calculate the kurtosis value K of each IMF component:

[0018]

[0019] Among them, x i is the signal value, is the signal mean, N is the sampling length, σ t is the standard deviation.

[0020] Calculate the correlation coefficient r of each IMF component:

[0021]

[0022] Where Cov(X,Y) is the covariance of the two variables X and Y, Var[X] is the variance of the X variable, and Var[Y] is the variance of the Y variable.

[0023] The IMF components with correlation coefficient greater than 0.1 and kurtosis value greater than 3 are selected for signal reconstruction.

[0024] The ECA-1DCNN model includes five sequentially connected combination modules, a fully connected layer, a relu activation function, and a softmax layer; each combination module includes a convolution layer and a pooling layer; and the convolution layers in the first four combination modules are respectively embedded with ECA models; the convolution layer slides the convolution kernel on the entire input sequence to generate a feature map, the pooling layer compresses each generated feature map, and the ECA module is added after downsampling the first four layers to further advance the features. The feature map output by the last pooling layer is connected to the fully connected layer, and then activated by the relu activation function, and then input into the softmax layer for classification to obtain a diagnosis result.

[0025] The ECA model implementation process is as follows:

[0026] The input feature map is subjected to global average pooling, and the feature map is transformed from a matrix of [h,w,c] to a vector of [1,1,c];

[0027] Calculate the adaptive one-dimensional convolution kernel size;

[0028] The results of the above convolution operation are normalized by the sigmoid activation function to obtain the attention weight of each channel;

[0029] The output of each channel is weighted and summed according to the attention weight to obtain the final feature map output.

[0030] The first convolution layer uses 16 large 64×1 convolution kernels, the second convolution layer uses 32 3×1 convolution kernels, and the third, fourth, and fifth convolution layers all use 64 3×1 convolution kernels; the pooling kernel size of the five pooling layers is 2×1, and the step size is 2.

[0031] The beneficial effects of the present invention are:

[0032] 1) The CEEMDAN decomposition used in the method provided by the present invention effectively eliminates the problem of modal aliasing, has better local characteristics at the same time and frequency scale, and adopts the overall average calculation method to effectively improve the operation efficiency of the algorithm; the wavelet packet transform used not only has high computational efficiency, but also, as an extension of the traditional wavelet transform, has a good effect of filtering out environmental background noise components while transforming the high-frequency details at the same time, thereby retaining more effective information;

[0033] 2) The ECA-Net used in the method provided by the present invention is an efficient channel attention mechanism, which is improved on the basis of SE-Net. The dimension reduction operation used in SE-Net will have a negative impact on the prediction of channel attention, and the acquisition of channel dependencies is inefficient. The ECA attention mechanism module directly uses a 1X1 convolution layer after the global average pooling layer, removing the fully connected layer. It avoids dimensionality reduction and effectively captures cross-channel interactions. It completes cross-channel interactions through one-dimensional convolution, and the size of the convolution kernel changes adaptively through a function, so that layers with larger channels can perform more cross-channel interactions.

[0034] 3) The processed data is passed into the convolutional neural network, with an accuracy rate of 99.80%, which is significantly higher than the fault recognition rate of the traditional convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN in the present invention;

[0036] FIG2( a ) is a wavelet packet transform frequency domain diagram of a rolling bearing inner race fault in a bearing experiment of the present invention;

[0037] FIG2( b ) is an energy diagram of each signal component of the third layer of a rolling bearing inner ring fault in a bearing experiment of the present invention;

[0038] Figure 3 It is the CEEMDAN decomposition diagram of the rolling bearing inner ring fault signal in the present invention;

[0039] Figure 4 It is the ECA-NET module diagram in the present invention;

[0040] FIG. 5( a ) is a graph showing a loss function of fault diagnosis in the present invention.

[0041] FIG5( b ) is a graph showing the loss function of fault diagnosis in the present invention. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0044] The present invention is a rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN. First, the collected vibration signal is subjected to a one-dimensional third-order wavelet packet transform to generate 8 WPT signal components, and the collected rolling bearing vibration signal is subjected to wavelet packet decomposition to obtain the energy value of each frequency band, and a reconstructed signal after preliminary noise reduction is obtained, and the reconstructed signal is subjected to CEEMDAN decomposition and combined with the kurtosis value and the correlation coefficient, and the signal components generated after the decomposition are screened and divided to obtain a denoised reconstructed signal; the processed signal is placed in the ECA-1DCNN model for training, and the diagnosis result is output. Compared with other rolling bearing fault diagnosis methods, this method effectively filters out background environmental noise components to eliminate the influence on signal analysis, while improving the accuracy of fault diagnosis and retaining more effective information.

[0045] like Figure 1 As shown, the rolling bearing fault diagnosis method based on wavelet packet transform and CEEMDAN includes:

[0046] Step 1: Use a vibration sensor to collect vibration signals of rolling bearings under different working conditions;

[0047] Different working conditions include a normal non-fault state and a fault state. The fault types include fault types of inner rings, outer rings and rolling elements of three different specifications of 0.18 mm, 0.36 mm and 0.54 mm, as well as a total of 10 types of normal states.

[0048] Step 2.1: Perform one-dimensional third-order wavelet packet transform on the rolling bearing vibration signals under different working conditions to generate several WPT signal components;

[0049] Step 2.2: Calculate the energy value of each WPT signal component respectively;

[0050] Step 2.3: Select signal components with energy values ​​greater than the average energy value from the energy map of Figure 2(b) for superposition reconstruction;

[0051] Step 3.1: Perform CEEMDAN decomposition on the reconstructed signal to obtain several IMF components, such as Figure 3 This is the breakdown diagram of the rolling bearing inner ring fault signal;

[0052] The principle of CEEMDAN is as follows: CEEMDAN is a new decomposition method proposed and improved on the basis of empirical mode decomposition EEMD. After decomposition, CEEMDAN obtains the IMF component with a finite number of adaptive white noise added, reducing the number of overall average calculations. In the decomposition process, empirical mode decomposition EEMD performs overall average calculation on the decomposed IMF components, while CEEMDAN directly performs overall average calculation after the first-order IMF is decomposed.

[0053] In the CEEMDAN algorithm, the input signal is y(t), the adaptive noise coefficient is ε, and w i (t) is white noise that obeys N(0,1) distribution, E() is defined as the kth order IMF component obtained by EMD decomposition, and IMF k It is defined as the kth order IMF component obtained by decomposition using the CEEMDAN algorithm. The signal decomposition steps are as follows:

[0054] 1) Add Gaussian white noise to the original signal for decomposition to obtain the first-order eigenmode;

[0055]

[0056] Where y(t) is the original signal; (-1) q εv i (t) is Gaussian white noise, ε is the standard deviation of noise; v i is a Gaussian white noise signal that satisfies the standard deviation; q = 1 or 2; is the first-order IMF component; r i is the residual component.

[0057] 2) Calculate the overall average value of N modal components to obtain the first eigenmodal component;

[0058]

[0059] In the formula is the overall mean of the N decomposed first-order IMF components.

[0060] 3) Calculate the residual after removing the first modal component

[0061]

[0062] 4) Add positive and negative paired Gaussian white noise to the residual to obtain a new signal, use the new signal as the carrier to perform EMD decomposition to obtain the first-order modal component D1, and then calculate the overall average value of D1 to obtain the second modal component

[0063]

[0064] 5) Calculate the residual after removing the second modal component

[0065]

[0066] 6) Repeat the above steps until the residual signal is a monotonic function and cannot be further decomposed. The algorithm ends. At this time, the number of intrinsic mode components obtained is K, and the original signal is decomposed into

[0067]

[0068] Step 3.2: The kurtosis value reflects the impact characteristics of the vibration signal. The kurtosis is sensitive to impact. Generally, the kurtosis value should be around 3, because the kurtosis of the normal distribution is equal to 3. If it deviates too much from 3, it means that the mechanical equipment has a certain impact vibration and there may be some hidden faults. As shown below:

[0069]

[0070] where x i is the signal value, is the signal mean, N is the sampling length, σ t is the standard deviation.

[0071] Step 3.3: Calculate the correlation coefficient r of each IMF component as follows:

[0072]

[0073] Where Cov(X,Y) is the covariance of the two variables X and Y, Var[X] is the variance of the X variable, and Var[Y] is the variance of the Y variable;

[0074] Step 3.4: Comprehensively consider and select IMF components with correlation coefficient greater than 0.1 and kurtosis value greater than 3 for signal reconstruction;

[0075] Step 4: Pre-process the denoised signal and divide it into a training set and a test set, input them into a one-dimensional convolutional neural network module, and output a diagnosis result, which is a fault-free state or different fault types.

[0076] The one-dimensional convolutional neural network includes five sequentially connected combination modules (each combination module includes a convolution layer and a pooling layer), a fully connected layer, a relu activation function, and a softmax layer. The improvement is that the ECA model is embedded in the convolution layers in the first four combination modules. The convolution layer slides the convolution kernel on the entire input sequence to generate a feature map, and the pooling layer compresses each feature map generated. The ECA module is added after the downsampling of the first four layers. The feature map output by the last pooling layer is connected to the fully connected layer, and then activated by the relu activation function, and then input to the softmax layer to obtain the diagnosis result.

[0077] Among them, ECA model is as follows Figure 4 The implementation process of the process shown is:

[0078] (1) The input feature map is subjected to global average pooling, and the feature map is converted from a matrix of [h, w, c] to a vector of [1, 1, c];

[0079] (2) Calculate the adaptive one-dimensional convolution kernel size;

[0080] (3) The result of the above convolution operation is normalized through the sigmoid activation function to obtain the attention weight of each channel;

[0081] (4) Finally, the output of each channel is weighted and summed according to the attention weight to obtain the final feature map output.

[0082] Among them, the first convolution layer uses 16 large 64×1 convolution kernels, the second convolution layer uses 32 3×1 convolution kernels, and the third, fourth, and fifth convolution layers all use 64 3×1 convolution kernels; the pooling kernel size of the 5 pooling layers is 2×1, and the step size is 2.

[0083] The experimental data in this paper comes from the Rolling Bearing Data Center of Case Western Reserve University (CWRU). The CWRU dataset is a standard experimental dataset in the field of bearing fault diagnosis and has been used by many scholars. The experimental object of this experiment is the deep groove ball bearing SKF6205. The experimental data is collected by an acceleration sensor installed at the drive end of the motor bearing. The sampling frequency is 12KHz and the motor load is 1797r / min. The test bench simulates the damage of the rolling ball, the inner ring of the bearing and the outer ring of the bearing. According to the degree of spark erosion, it is divided into three damage sizes of 0.18mm, 0.36mm and 0.54mm. Combined with the healthy bearing signal, the diagnostic task is to distinguish ten different types of bearing faults.

[0084] The environmental background noise of the experimental bearing fault sample data is filtered out by the rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN. Figure 2(a) and Figure 2(b) are the wavelet packet transform frequency domain diagram of the inner ring fault of the rolling bearing in the bearing experiment of the Case Western Reserve University in the United States and the energy diagram of each signal component in the third layer in the present invention.

[0085] Figure 5(a) and Figure 5(b) show the changing trends of accuracy and loss function values ​​during the experiment. The network weights are updated through each iteration to minimize the value of the objective function, thereby achieving the best recognition accuracy. After 80 iterations, the entire training becomes stable and converges to the final result.

[0086] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. Rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN, It is characterized in that include: Step 1: Collect vibration signals of rolling bearings under different working conditions; Step 2: Perform wavelet packet decomposition on the vibration signals under different working conditions and superimpose the signals; Step 3: Perform CEEMDAN decomposition on the superimposed signal and screen the IMF components again for reconstruction; Step 4: Input the processed signal into the ECA-1DCNN model for iterative training to optimize the network model parameters for outputting the diagnosis results.

2. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN according to claim 1, It is characterized in that Different working conditions include a normal fault-free state and a fault state, and the fault types include faults of the inner ring, outer ring and rolling element corresponding to bearings of different specifications; the diagnosis result is a fault-free state or different fault types.

3. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-IDCNN according to claim 1, It is characterized in that The step 2 comprises: Step 2.1: Perform one-dimensional third-order wavelet packet transform to obtain several WPT signal components; Step 2.2: Calculate the energy value of each WPT signal component; Step 2.3: Select signal components with energy values ​​greater than the average value for superposition reconstruction.

4. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN according to claim 1, It is characterized in that The screening of IMF components includes screening according to the kurtosis value K and the correlation coefficient r.

5. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-IDCNN according to claim 5, It is characterized in that Calculate the kurtosis value K of each IMF component: Among them, x i is the signal value, is the signal mean, N is the sampling length, σ t is the standard deviation.

6. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN according to claim 5, It is characterized in that Calculate the correlation coefficient r of each IMF component: Where Cov(X,Y) is the covariance of the two variables X and Y, Var[X] is the variance of the X variable, and Var[Y] is the variance of the Y variable.

7. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-IDCNN according to claim 5, It is characterized in that The IMF components with correlation coefficient greater than 0.1 and kurtosis value greater than 3 are selected for signal reconstruction.

8. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-IDCNN according to claim 1, Features: The ECA-1DCNN model includes five combined modules, a fully connected layer, a relu activation function, and a softmax layer connected in sequence; each combined module includes a convolutional layer and a pooling layer; and the ECA model is embedded in the convolutional layers of the first four combined modules; the convolutional layer slides the convolutional kernel over the entire input sequence to generate feature maps, the pooling layer compresses each generated feature map, the ECA module is added after the downsampling of the first four layers to further advance the features, the feature map output by the last pooling layer is connected to the fully connected layer, and after being activated by the relu activation function, it is input to the softmax layer for classification to obtain the diagnosis result.

9. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN according to claim 8, characterized in that, the implementation process of the ECA model is as follows: The input feature map is subjected to global average pooling, and the feature map changes from a matrix of [h, w, c] to a vector of [1, 1, c]; Calculate the size of the adaptive one-dimensional convolutional kernel kernel size; Normalize the result of the above convolutional operation through the sigmoid activation function to obtain the attention weight of each channel; Weighted sum the outputs of each channel according to the attention weight to obtain the final feature map output.

10. The rolling bearing fault diagnosis method based on wavelet packet transform-CEEMDAN and ECA-1DCNN according to claim 8, characterized in that, The first convolutional layer uses 16 large convolutional kernels of 64×1, the second convolutional layer uses 32 convolutional kernels of 3×1, and the third, fourth, and fifth convolutional layers all use 64 convolutional kernels of 3×1; the pooling kernel sizes of the five pooling layers are all 2×1, and the strides are all 2.

Images