Motor bearing fault identification method based on variational mode decomposition denoising

Abstract

The invention discloses a motor bearing fault identification method based on variational mode decomposition denoising, and relates to the technical field of motor bearing fault identification, and the method comprises the steps: collecting a vibration signal of a motor bearing in a no-load stable state of a motor; processing the collected motor bearing vibration signal by adopting variational mode decomposition, and extracting a feature vector of the denoised motor bearing vibration signal; and inputting the motor bearing vibration signal feature vector into a pre-trained fault classification model based on a neural network, and carrying out fault classification to obtain a bearing fault identification result. According to the invention, denoising processing is carried out on the motor bearing vibration signal through a variational mode decomposition method, and a useful fault signal is recovered from the noise-containing motor bearing vibration signal. The fault classification model in the invention combines the local information extraction capability of the convolutional neural network and the time context relationship modeling capability of Transform. According to the invention, the technical effect of improving the motor bearing fault identification accuracy can be finally achieved.

Description

Technical Field

[0001] This application relates to the technical field of motor bearing fault identification, and particularly to a method for motor bearing fault identification based on variational mode decomposition denoising. Background Technique

[0002] The motor bearing is a core component in the motor system, and its performance directly affects the stability and operating efficiency of the motor. However, during long-term operation, the motor bearing may fail due to various factors, thereby affecting the normal operation of the motor. Therefore, it is necessary to perform fault prediction and health management on the motor bearing, which is of crucial significance for ensuring the reliability and safety of the motor system.

[0003] The vibration signal of the motor bearing is usually disturbed by noise, which makes the vibration signal waveform disorderly and the burr phenomenon significant, making it difficult to characterize the fault features, thus having a greater negative impact on the final fault identification result. Currently, common signal processing methods include empirical wavelet transform and empirical mode decomposition, etc., for realizing denoising processing. However, the empirical wavelet transform depends on the characteristics of the signal, has poor controllability, and is prone to modal aliasing and end effect; while when introducing white noise in empirical mode decomposition, it is easy to cause signal distortion. Therefore, there is an urgent need for an improved denoising processing method to effectively extract fault feature information.

[0004] In the field of motor bearing fault identification, existing technologies widely use neural network models to achieve. Although these models have a simple structure, they lack the ability to fully extract the fault features of the motor bearing and are difficult to meet the requirements of high-precision fault identification. Therefore, it is necessary to develop an improved neural network-based fault classification model to further improve the accuracy and reliability of fault identification. Summary of the Invention

[0005] Aiming at the problems of the existing technology, the present invention provides a method for motor bearing fault identification based on variational mode decomposition denoising, which improves the accuracy and efficiency of fault identification.

[0006] To achieve the above object, the present invention provides the following scheme, and the specific steps include:

[0007] (1) Under the stable no-load state of the motor, collect the vibration signal of the motor bearing;

[0008] (2) Use variational mode decomposition (VMD) to process the collected vibration signal of the motor bearing, and extract the feature vector of the vibration signal of the denoised motor bearing.

[0009] (3) Input the feature vector of the vibration signal of the motor bearing into a pre-trained neural network-based fault classification model for fault classification to obtain the bearing fault identification result.

[0010] Preferably, the process of (2) is:

[0011] Perform Fourier transform on the input motor bearing vibration signal to convert the signal from time domain to frequency domain for processing;

[0012] Initialize the key parameters involved in variational mode decomposition, including the mode number, center frequency and Lagrange multiplier, and use the conventional initialization method to determine an initial mode;

[0013] Iteratively update the mode in the frequency domain, and update the center frequency and Lagrange multiplier at the same time until the convergence condition is met;

[0014] After the convergence condition is determined to be met, the final frequency mode is converted back to the time domain through inverse Fourier transform to obtain the modal components (IMF);

[0015] The Pearson correlation coefficient is calculated for each modal component and the original signal to quantify the similarity between the modal component and the original signal;

[0016] The modal components with high correlation are retained, and the noise components with correlation lower than the threshold or correlation equal to 0 are eliminated. The filtered modal components are superimposed to complete the signal reconstruction and obtain the denoised motor bearing vibration signal;

[0017] The time series features of the denoised vibration signal are extracted to obtain the input vector of the neural network model.

[0018] Preferably, the time series characteristics of the vibration signal include: variance X va , skewness X sk , Kurtosis X ku and the root mean square value X rms .

[0019] Preferably, the neural network model adopts convolutional neural network (CNN) and Transformer structure to extract local information in time series features and capture long-term dependencies of signals.

[0020] The present invention discloses the following technical effects:

[0021] Variational mode decomposition (VMD) is an adaptive and completely non-recursive signal processing method. By transforming the signal decomposition problem into a variational optimization problem of minimizing the frequency-domain bandwidth, it naturally has the characteristic of spectral aggregation. This technology can adaptively determine the number of mode decompositions and dynamically adjust the number of modes according to the characteristics of the input signal. In the denoising process of motor bearing vibration signals, VMD effectively suppresses the mode mixing phenomenon through bandwidth constraint and center frequency optimization, showing robustness to different types of noise. Especially when dealing with non-linear and non-stationary motor bearing vibration signals, VMD can significantly improve the denoising effect and recover useful fault feature signals from the noisy motor bearing vibration signals. In addition, the fault classification model combines the ability of convolutional neural network (CNN) to extract local information and the advantage of Transformer in modeling time context relationships, improving the accuracy of motor bearing fault identification and contributing to enhancing the efficiency of motor bearing fault prediction and health management. Brief Description of the Drawings

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0023] Figure 1 It is a schematic flowchart of a method for identifying motor bearing faults based on variational mode decomposition denoising provided by an embodiment of the present application.

[0024] Figure 2 It is a schematic flowchart of variational mode decomposition provided by an embodiment of the present application.

[0025] Figure 3 It is a detailed structure diagram of a motor bearing fault classification model provided by an embodiment of the present application. Detailed Embodiments

[0026] The above description is only an overview of the technical solutions of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present application more obvious and understandable, the following specifically gives the detailed embodiments of the present application.

[0027] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further describe the present application in detail with reference to the drawings. The described embodiments should not be regarded as limitations of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0028] In the following description, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that comprises a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or modules not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application pertains. The terms used herein are only for the purpose of describing the embodiments of this application.

[0029] For the convenience of explaining the inventive concept of this application, the following briefly describes the motor bearing fault identification technology.

[0030] When a fault occurs in the bearing of a motor, the degree of signal fluctuation increases, resulting in a significant increase in variance; at the same time, the impact and vibration caused by the fault will increase the energy of the signal, resulting in an increase in the root mean square value; in addition, the fault may cause changes in the skewness and kurtosis of the signal distribution, indicating that there are more impact components in the signal. Therefore, the present invention selects variance, root mean square value, skewness, and kurtosis as the time series feature parameters of the motor bearing vibration signal to form a feature vector. These parameters can intuitively reflect the fault characteristics of the bearing, are all dimensionless parameters, are sensitive to faults and defects, and are basically not affected by the bearing model.

[0031] Since the present invention relies on a pre-trained motor bearing fault classification model to implement, to more clearly explain the fault identification method, the training method of the model is first described. The training method includes:

[0032] Obtain a sufficient number of sample data. The sample data includes the time series features of the motor bearing vibration signal in the normal state and the fault state, including variance, root mean square value, skewness, and kurtosis, as well as the label information of each sample (0 for normal and 1 for fault). Divide the sample data into a training set, a validation set, and a test set. Construct a preliminary training network, select the binary cross-entropy function as the loss function, and the formula of the loss function is as follows:

[0033]

[0034] Wherein, represents the binary cross-entropy loss function, L is the number of samples; y l is the true label of the l-th sample, is the predicted value of the l-th sample, and the value ranges from 0 to 1. Set the training period and the initial learning rate, and use the Adam optimizer and the gradient descent algorithm to train the model. During the training process, adjust the hyperparameters according to the test results of the test set until the model converges or reaches the expected performance.

[0035] The performance of the model is evaluated using Accuracy, Precision, and Recall. The higher the values of these three metrics, the better the performance of the model. Their calculation formulas are as follows:

[0036]

[0037] Among them, TP is the number of normal samples predicted as normal, TN is the number of faulty samples predicted as faulty, FP is the number of normal samples predicted as faulty, and FN is the number of faulty samples predicted as normal. After training is completed, the model parameters at the optimal performance are retained as the pre-trained model for motor bearing fault identification.

[0038] Example 1, the present application provides a method for identifying motor bearing faults based on variational mode decomposition denoising, as Figure 1 shown, the method includes:

[0039] Step S10, collect the vibration signal of the motor bearing under the stable no-load state of the motor;

[0040] For the YBQA-3-72B04B type asynchronous motor with a rated power of 4 kw and a rated speed of 4500 rpm, the motor operates under no-load conditions. Signal collection is carried out for the normal state of the motor and the bearing fault state respectively, and 50 groups of data are collected for each state as the sample data for training the neural network model.

[0041] In the specific implementation process, the vibration acceleration sensor is fixed on the motor bearing by magnetic attraction, and the vibration signals of the motor bearing are detected along the axial direction, vertical direction, and horizontal direction respectively. At the same time, ensure that the acquisition work is carried out in a well-controlled environment to reduce the influence of temperature, humidity, and power supply fluctuations. This improves the reliability of signal acquisition.

[0042] Step S20, use variational mode decomposition to process the collected vibration signal of the motor bearing, and extract the feature vector of the denoised vibration signal of the motor bearing.

[0043] In the present invention, the variational mode decomposition (VMD) method is used to denoise the vibration signal of the motor bearing. The idea of VMD is: while ensuring that the sum of each mode approximates the original vibration signal, it is required that the sum of the bandwidths of each mode near its center frequency is minimized. This can be regarded as an optimization problem and solved by variational methods.

[0044] The bandwidth is characterized as the energy size after the Hilbert transform of the modal signal is shifted to the baseband. Let the analytic signal of a modal u k (t) be: Where denotes the Hilbert transform. Then, by multiplying by the mode is shifted to the baseband. If the mode is concentrated in the frequency domain at this time, it will have fewer high-frequency components near the baseband, so it is called "bandwidth-limited".

[0045] For the k-th mode, assume its center frequency is ω k . In the frequency domain, if U k (ω) represents the Fourier transform of u k (t), then it is expected to have the main energy near ω≈ω k . At this time, the measurement of the bandwidth can be expressed by the following formula:

[0046]

[0047] That is, to see to what extent its instantaneous energy extends to high frequencies. By summing and minimizing the above bandwidth, each mode can be gathered near its respective center frequency.

[0048] The specific process of the variational mode decomposition algorithm is as Figure 2 shown, and the relevant steps and principles are described as follows:

[0049] Perform a Fourier transform on the original signal to obtain the spectrum.

[0050] Initialize the key parameters involved in VMD. Set the number of modes K and the Lagrange multiplier λ, and use a conventional initialization method to determine an initial mode u k . In this embodiment, the initialization method used is all-zero initialization. Analyze the amplitude spectrum of the original signal, and obtain the frequencies corresponding to the K peaks as the initial center frequencies ω k .

[0051] Let be the set of center frequencies of all modes, be the set of all mode functions. The objective function to be minimized by VMD can be expressed by the following formula:

[0052]

[0053] Here is essentially the Hilbert transform of the mode function u k (t), where δ(t) is the Dirac function, which represents a unit impulse in the time domain, is the kernel function of the Hilbert transform; at the same time, it satisfies the constraint condition: where f(t) is the original motor bearing vibration signal; to achieve this constraint condition, define the Lagrangian function:

[0054]

[0055] Among them, represents the Lagrangian function, and λ(t) is the time function of the Lagrange multiplier;

[0056] Iteratively update the modes in the frequency domain, and simultaneously update the center frequency and the Lagrange multiplier until the convergence condition is satisfied. Denote as the frequency-domain representation of u k (t), represents the frequency-domain representation of the original signal f(t). The Lagrange multiplier corresponds to the constraint. In each iteration step, for the nth iteration of the kth mode, assuming that other modes are known and fixed, then update

[0057]

[0058] Among them, represents the nth iteration of the ith mode, and α is the penalty term coefficient. After updating , then calculate the new center frequency:

[0059]

[0060] Finally, update the Lagrange multiplier:

[0061]

[0062] where τ is the step size, which is used to control the update speed. By repeatedly and alternately updating each mode and the center frequency until convergence or reaching the iteration upper limit, each frequency mode is obtained.

[0063] After judging that the convergence condition is reached, transform the finally obtained frequency mode u k back to the time domain through inverse Fourier transform to obtain each mode component (IMF).

[0064] Calculate the Pearson correlation coefficient between each mode component and the original signal to quantify the similarity between the mode component and the original signal. The calculation formula of the Pearson correlation coefficient is as follows:

[0065]

[0066] Among them, M represents the number of time samples of the ith mode component, m is the sample index, and m ∈ [1, M]. IMF i (t) m represents the value of the mth time sample of the ith mode component, represents the mean value of the ith mode component, f(t) m represents the value of the mth time sample of the original signal, represents the mean value of the original signal data, Ri (f, IMF) represents the correlation coefficient between the i-th modal component and the original signal.

[0067] Calculate the Pearson correlation coefficient of each modal component based on the above formula, and set a specific threshold to determine whether the modal component is a noise component. In this embodiment, the threshold is set to 0.5.

[0068] Retain the modal components with correlation coefficients greater than the threshold, and eliminate the noise components with correlation coefficients lower than the threshold or close to 0. Or sort all modal components in descending order of correlation coefficients, and retain the top N modal components with the highest correlation. The value of N is determined by K, and N < K. Superimpose the remaining modal components after removing the noise components to complete signal reconstruction, and obtain the denoised motor bearing vibration signal:

[0069]

[0070] Among them, selected represents the index set of the modal components with higher correlations. The Pearson correlation coefficient is simple and easy to calculate, plays a key role in the VMD algorithm, can effectively improve the denoising effect, and retain the main features of the signal.

[0071] Extract the time-series features of the denoised vibration signal to obtain the input vector of the neural network model. In this embodiment, the variance X va , skewness X sk , kurtosis X ku and root mean square value X rms of the motor bearing vibration signal are selected as the input vector of the neural network model, and the above time-series features of the reconstructed signal are concatenated to obtain the feature vector I:

[0072] I = (X va , X sk , X ku , X rms )

[0073] Step S30, input the feature vector of the vibration signal into the pre-trained neural network-based fault classification model for fault classification to obtain the fault identification result of the bearing vibration signal.

[0074] Embodiment 2. The present application provides a method for identifying motor bearing faults based on variational mode decomposition denoising. In the method, the pre-trained fault classification model adopts a convolutional neural network (CNN) and a Transformer structure to extract local information in time series features and capture the long-term dependence relationship of signals. The convolutional neural network is a common deep learning architecture widely used in classification tasks. It can extract local features through local receptive fields and obtain key information beneficial to the task. The parameter sharing mechanism of CNN can effectively improve the calculation efficiency. The Transformer structure has powerful parallel processing capabilities and the ability to model long-distance dependence relationships. The self-attention mechanism in this structure can dynamically focus on different positions of the sequence, better understand the context information, and its pre-training and fine-tuning strategies enable the model to efficiently adapt to various tasks.

[0075] The detailed structure of the above model is as Figure 3 shown. The input feature vector of the reconstructed signal is first fed into multiple layers of CNN for extracting local feature information. Its main operation is convolution operation. By convolving the input data with a group of learnable filters (convolution kernels), feature maps are generated. Among them, the convolution kernel size of the convolution layer is 3, and the stride is 1. The average pooling method is selected for the pooling layer.

[0076] CNN uses a residual connection method to retain the original information, which can effectively alleviate the overfitting phenomenon. After the feature vector is processed by CNN, it is fed into the Transformer. This structure is mainly composed of a feed-forward layer and multi-head self-attention. The calculation process of multi-head self-attention is as follows:

[0077] For the output Z of the previous layer of the model, assume represents a matrix, n is the sequence length, and d is the input dimension. First, perform a linear transformation on the input data to obtain the query (Query), key (Key), and value (Value) matrices for each head, denoted by Q, K, and V respectively:

[0078] Q = ZW Q , K = ZW K , V = ZW V

[0079] where, W Q , W K , W V represents the weight matrix of the query, key, and value of the j-th head saved after pre-training, with a size of n×d k .

[0080] Split Q, K, and V into h heads, and the dimension of each head

[0081] For each head j, calculate the attention weight A j :

[0082]

[0083] where the softmax function is used to normalize the weights, is the transposed matrix of the key matrix of the j-th head, is the scaling factor used to control the distribution of the weights. The attention weights of each head are all from the pre-trained model.

[0084] For each head j, calculate the output vector O j : O j = A j V j , and then concatenate the outputs of all heads:

[0085] O = [O1 O2 ··· O h

[0086] Finally, perform a linear transformation on the concatenated output O to obtain the final multi-head self-attention output:

[0087] MultiHead(Z) = OW o

[0088] where W o is the pre-trained weight matrix with size d×d, and MultiHead represents the multi-head attention mechanism. Each head of the multi-head self-attention can learn different information of the input data, capture richer context information, generate more effective fault feature information, and thus further improve the accuracy of fault identification.

[0089] The output of the Transformer structure is mapped through a fully connected layer to obtain the final output vector, and finally the classification result is determined according to the probability distribution to determine whether there is a fault in the motor bearing.

[0090] The above specific implementation manners do not constitute a limitation on the protection scope of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of this application shall be included within the protection scope of this application. In some cases, the actions or steps recorded in this application can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or consecutive order shown to achieve the desired results. In certain embodiments, multi-task processing and parallel processing are also possible or may be advantageous.​

Claims

1. A method for identifying motor bearing faults based on variational mode decomposition denoising, characterized in that, The method includes: (1) Collect the vibration signal of the motor bearing under the no-load stable state of the motor; (2) Process the collected vibration signal of the motor bearing by variational mode decomposition, and extract the feature vector of the vibration signal of the motor bearing after denoising; (3) Input the feature vector of the vibration signal of the motor bearing into a pre-trained neural network-based fault classification model for fault classification to obtain the bearing fault identification result.

2. The method for identifying motor bearing faults based on variational mode decomposition denoising according to claim 1, wherein, In (1), the vibration acceleration sensor is fixed on the motor bearing by magnetic attraction, and the vibration signals of the motor bearing are detected along the axial direction, vertical direction, and horizontal direction respectively.

3. A method for identifying motor bearing faults based on variational mode decomposition denoising according to claim 1, characterized in that, The specific process involved in (2) includes: Perform Fourier transform on the input vibration signal of the motor bearing, and process the signal by converting it from the time domain to the frequency domain; Initialize the key parameters involved in variational mode decomposition, including the number of modes, the central frequency, and the Lagrange multiplier, and determine an initial mode using a conventional initialization method; Iteratively update the modes in the frequency domain, and update the central frequency and the Lagrange multiplier at the same time until the convergence condition is met; After judging that the convergence condition is reached, convert the finally obtained frequency modes back to the time domain through inverse Fourier transform to obtain each mode component; Calculate the Pearson correlation coefficient between each mode component and the original signal to quantify the similarity between the mode component and the original signal, and remove the noise components according to the correlation coefficient to complete signal reconstruction; Extract the time-series features of the vibration signal of the motor bearing after denoising processing to obtain the input vector of the neural network model.

4. The motor bearing fault identification method based on variational mode decomposition denoising according to claim 3, characterized in that The detailed process of initializing the key parameters involved in the variational mode decomposition is as follows: Let be the set of the central frequencies of all modes, be the set of all mode functions, and the objective function minimized by variational mode decomposition is expressed by the following formula: Here is essentially a Hilbert transform of the modal function , where is the Dirac function, representing a unit impulse in the time domain is the kernel function of the Hilbert transform, K represents the total number of modes represents the center frequency of the k-th mode represents the k-th modal component; and at the same time satisfies the constraint condition: , where f ( t ) is the original motor bearing vibration signal; to achieve this constraint condition, the Lagrange multiplier is introduced, and the Lagrangian function is defined as: Among them, represents the Lagrangian function, is a time function of the Lagrange multiplier.

5. A method for identifying motor bearing faults based on variational mode decomposition denoising according to claim 3, characterized in that, The process of iteratively updating the modes is as follows: Denote as the frequency domain representation of the mode function , denote the frequency domain representation of the original signal , while the Lagrange multiplier corresponds to the constraint denote the frequency index; in each iteration step, for the th iteration n , assume that other modes are known and fixed, and then update in the frequency domain Among them, represents the i th iteration of the n th mode, is the penalty term coefficient; after updating , then calculate the new center frequency: Finally update the Lagrange multiplier: Among them is the step size, which is used to control the update speed.

6. The motor bearing fault identification method based on variational mode decomposition denoising according to claim 3, characterized in that, The calculation formula of the Pearson correlation coefficient is as follows: Among them, M represents the number of time samples of the i-th modal component, m is the sample index, ; represents the value of the m-th time sample of the i-th modal component, represents the mean value of the i-th modal component, represents the value of the m-th time sample of the original signal, represents the mean value of the original signal data, represents the correlation coefficient between the i-th modal component and the original signal; Based on the correlation coefficient calculation formula, obtain the correlation coefficient of each mode component, set a specific threshold to judge whether the mode component is a noise component, and retain the mode components with a correlation coefficient greater than the threshold; superimpose the remaining mode components after removing the noise components to complete signal reconstruction to obtain the vibration signal of the motor bearing after denoising.

7. A method for identifying motor bearing faults based on variational mode decomposition denoising according to claim 3, characterized in that The temporal characteristics of the vibration signal include: variance , skewness , kurtosis and root mean square value .

8. A method for identifying motor bearing faults based on variational mode decomposition denoising according to claim 1, characterized in that The fault classification model in (3) adopts a convolutional neural network and a Transformer structure to extract local information in the time-series features and capture the long-term dependencies of the signal.

9. A method for identifying motor bearing faults based on variational mode decomposition denoising, as claimed in claim 8, wherein, The convolutional neural network consists of a convolutional layer, an activation layer, a normalization layer, and a pooling layer, and uses a residual connection method to alleviate the overfitting phenomenon; the Transformer structure mainly includes a feed-forward layer and a multi-head self-attention mechanism, and has the ability to model long-distance time context and the ability to process data in parallel; the convolutional neural network and the Transformer and the internal structures of the two structures are all connected in series.

10. A method for identifying motor bearing faults based on variational mode decomposition denoising, as described in claim 9, characterized in that, The calculation process of the multi-head self-attention is as follows: For the output Z of the previous layer of the model, let Z , represent matrices, n is the sequence length, and d is the input dimension; first, perform a linear transformation on the input data to obtain the query, key, and value matrices for each head, denoted by Q, K, V represent : Among them, , , represent the weight matrices of the queries, keys, and values of the j th head saved after pre-training, with a size of ; Split Q, K, V into h heads, each head having a dimension of . For each header j , calculate the attention weights : Among them, the softmax function is used to normalize the weights, is the transposed matrix of the key matrix of the j th head, is the scaling factor used to control the distribution of the weights; the attention weights of each head are all from the pre-trained model; For each header j compute the output vector : and then concatenate the outputs of all headers: Finally, perform a linear transformation on the concatenated output to obtain the final multi-head self-attention output: Among them, is the pre-trained weight matrix, with a size of , represents the multi-head attention mechanism.

Images