AR-VMD-based Preprocessing Method for Bearing Fault Data under Variable Working Conditions

Abstract

The present invention proposes a method for preprocessing variable-condition bearing fault data based on AR-VMD. First, to improve the adaptability of VMD, an adaptive VMD method based on the maximum ratio of adjacent modal repetitions is proposed, which solves the problem of difficult determination of the modal decomposition layer number K value. Secondly, according to the principle of maximum envelope spectrum entropy, components are selected for wavelet soft threshold denoising and then recombined to minimize the loss of effective information as much as possible. At the same time, aiming at the variable-condition characteristics of fault data, the above method is integrated with data segmentation, DC component processing, and marginal spectrum analysis to form a set of data preprocessing methods for such data. The experimental results of processing fault data on a high-speed bearing test bench show that the proposed method can effectively extract fault signals while removing noise signals, improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of data preprocessing, and in particular to a data preprocessing method based on AR-VMD, and a variable operating condition bearing fault data preprocessing method. Background Art

[0002] As a basic part of various modern industrial instruments and equipment, bearings play a vital role in the normal operation of the equipment. As we all know, bearings are lossy devices, and their working state is affected by factors such as working conditions, their own quality, and warranty status. Since bearings are widely used, once a bearing fails during operation, it will have a very adverse effect on the equipment, so bearing fault diagnosis is of great significance. Entering the era of big data, modern fault diagnosis technology faces many challenges, and simple and efficient data preprocessing methods have become one of the ways to resolve contradictions. [1] .

[0003] The main shaft bearings of aircraft engines often work under conditions of high speed, high temperature and complex loads, which makes its structure different from other bearings. The main shaft of an aircraft engine is equipped with multiple bearings, and their characteristics and functions vary depending on the installation position. During the operation of the aircraft engine, bearings at different positions may slip to varying degrees. One of the typical characteristics of bearing failure is that it is greatly affected by the working conditions, and changes in working conditions can even directly lead to the occurrence of failures. Bearings running under high-speed conditions will produce failures that are not easy to occur under other working conditions, so it is necessary to study the impact of changes in working conditions on bearing failures. When the speed or load changes, the bearing signal will have different fault information expressions in the time-frequency domain. However, constructing different data processing methods for different fault signals is labor-intensive and affects the accuracy of signal processing. Based on this situation, this paper intends to study a set of data processing algorithms for such high-speed bearing failures to achieve adaptive processing of bearing fault data.

[0004] Usually, the raw data containing fault information collected by sensors needs to be processed to filter out noise signals and highlight fault signals, and then the fault diagnosis is performed based on the processed data. The accuracy of the data processing method directly affects the accuracy of fault diagnosis to a large extent, so how to process fault data based on signal characteristics is a problem that scholars are very concerned about. He Dan et al. [2] After using SED to enhance the gear eccentricity fault features in the vibration signal, the IMF obtained by EMD decomposition uses the weighted-correlation frequency kurtosis index to adaptively extract the IMF components containing the eccentricity fault features. The experimental results show that this method can adaptively extract fault components. Zhan Yingyu et al. [3] Aiming at the modal aliasing phenomenon of EMD, a decorrelated multi-frequency empirical mode decomposition (DMFEMD) was proposed. Experiments have shown that DMFEMD can effectively separate mixed signals with different frequency ratios and improve the decomposition effect. Zhang Lizhi et al.[4] To solve the problem of rotating machinery fault diagnosis, a fault diagnosis method combining empirical mode decomposition (EMD), singular value decomposition (SVD) and deep convolutional network (CNN) is proposed, and the superiority of this method is verified through experiments. Liu Bo et al. [5] Filter through a comb filter and EMD, and after signal reconstruction of modal components according to the Pearson correlation coefficient, use the modal components obtained by VMD for secondary reconstruction according to energy entropy. Ning Yi et al. [6] After determining the boundary IMF between noise and signal according to the cross-correlation coefficient between each IMF component and the original signal, apply an improved threshold function to the noise component and the boundary component for wavelet denoising. The diagnostic results prove that this method can effectively extract the signal characteristic frequency. Yin Xunlong et al. [7] An adaptive VMD algorithm based on the sparrow search algorithm (SSA) is proposed, and the signal is reconstructed after singular value decomposition or direct elimination of components according to the Pearson correlation coefficient analysis. The experimental results show that this method has excellent noise reduction effect. Xu Yanhe et al. [8] A hybrid fault diagnosis method for rotating machinery based on variational mode decomposition energy entropy (VMD-EE) and transfer learning (TL) is proposed. The diagnostic results show that the proposed method has accuracy and robustness. Huang Yan et al. [9] Introduce the concept of scale space for frequency band division to determine the number of intrinsic mode functions, central frequency and corresponding penalty factor values required by VMD. Experiments show that this method significantly improves the adaptability and accuracy of the VMD method. Li Yuxing et al.

[10] Apply the genetic algorithm (GA) to optimize the parameter combination of VMD, improving the decomposition accuracy of VMD. Li Cuisheng et al.

[11] Aiming at the faults of the wheelset bearings of high-speed trains, use the envelope spectrum kurtosis value as the screening criterion for signal decomposition and reconstruction, and at the same time combine the particle swarm algorithm to construct an adaptive VMD method, obtaining good fault feature extraction effects. Feng Gang et al.

[12] Distinguish noise components according to the component energy kurtosis to achieve the purpose of signal noise reduction. This method is verified in the experiment of extracting fault signals of variable speed gearboxes.

[0005] Computer network technologies such as neural networks and deep learning have developed rapidly. Scholars incorporate such technologies into the field of fault diagnosis of mechanical equipment to improve the accuracy of fault diagnosis. Aiming at the incipient faults of sensors in the electric traction system, Hongtian Chen et al. [13，14] A data-driven sensor fault detection and diagnosis method based on PCA technology is proposed and its effectiveness is verified. The article

[15] Propose to establish a nonlinear model of photovoltaic modules using the projection method and realize the data-driven hot spot detection function. Kunpeng Zhang et al.

[16] Aiming at the time-varying characteristics of high-speed trains and the interactivity of multiple faults, a multi-fault isolation framework is developed to simplify the segmentation of the time-varying fault parameter space. At the same time, a two-level adaptive fault estimation scheme is adopted to determine the alarm priorities of different faults. However, for the bearing faults of aero-engines under complex working conditions, there is still a lack of an efficient processing and diagnosis system with functional integration.

[0006] For the large-scale raw data of bearing faults under variable working conditions, considering that there are many influencing factors and the information contained is complex, the present invention proposes a method for preprocessing bearing fault data under variable working conditions based on adaptive VMD. First, to improve the self-adaptability of VMD, an adaptive VMD method based on the maximum ratio of adjacent modal repetitions is proposed, which solves the problem of difficult determination of the modal layer number K value. Secondly, according to the principle of maximum envelope spectrum entropy, components are selected for wavelet soft threshold denoising and then recombined to minimize the loss of effective information as much as possible. At the same time, aiming at the characteristics of variable working conditions of fault data, the above method is integrated with data segmentation, DC component processing, and marginal spectrum analysis into a set of data preprocessing methods for such data. The experimental results of processing fault data on a high-speed bearing test bench show that the proposed method can effectively extract fault signals while removing noise signals, improving the accuracy of fault diagnosis.

[0007] [1] Lei Yaguo. Theory and Application of Intelligent Operation and Maintenance of Mechanical Equipment Driven by Big Data [M]. Beijing: Publishing House of Electronics Industry, 2021.

[0008] [2] He Dan, Zhang Lijie, Xiao Yuan, et al. Diagnosis of Slight Gear Eccentricity Fault of Circular Knitting Machine Based on SED-EMD [J]. Noise and Vibration Control, 2022, 42(1): 132-137.

[0009] [3] Zhan Yingyu, Cheng Lianglun, Wang Tao. Optimization Method for Fault Diagnosis Performance of Decorrelated Multi-Frequency Empirical Mode Decomposition [J]. Journal of Vibration and Shock, 2020, 39(1): 115-122, 49.

[0010] [4] Zhang Lizhi, Xu Weixiao, Jing Luyang, et al. Fault Diagnosis of Rotating Machinery Based on EMD-SVD and CNN [J]. Journal of Vibration, Measurement & Diagnosis, 2020, 40(6): 1063-1070.

[0011] [5] LIU B, LIU C, ZHOU Y, et al. A Chatter Detection Method in Milling Based on Grey Wolf Optimization Vmd and Multi-Entropy Features [Z]. SSRN. 2022

[0012] [6] Ning Yi, Wei Zhigang, Zhou Jianxiong. Fault Diagnosis of Low-Speed and Heavy-Load Bearings in Mixing Machines Based on Improved EMD and Wavelet Thresholding [J]. Noise and Vibration Control, 2020, 40(6): 134-139.

[0013] [7] Yin Xunlong, Mou Zonglei, Wang Youqing. Fault Diagnosis of Rotating Machinery Based on DVMD Denoising [J]. Control Theory & Applications: 1-9.

[0014] [8] XU Y, LI S, JIANG W, et al. A Progressive Fault Diagnosis Method for Rolling Bearings Based on VMD Energy Entropy and a Deep Adversarial Transfer Network [J]. Measurement Science and Technology, 2022, 33(9).

[0015] [9] Huang Yan, Lin Jianhui, Liu Zechao, et al. Fault Diagnosis of Axle Box Bearings of High-Speed Trains Based on Adaptive VMD [J]. Journal of Vibration and Shock, 2021, 40(3): 240-245.

[0016]

[10] LI Y, TANG B, JIANG X, et al. Bearing Fault Feature Extraction Method Based on GA-VMD and Center Frequency [J]. Mathematical Problems in Engineering, 2022, 2022.

[0017]

[11] Li Cuisheng, Liao Yingying, Liu Yongqiang. Fault Diagnosis of Wheel Set Bearings of High-Speed Trains Based on EEMD and Parameter Adaptive VMD [J]. Journal of Vibration and Shock, 2022, 41(1): 68-77.

[0018]

[12] Feng Gang, Liu Tongtong, Cui Lingli. Research on Fault Diagnosis of Variable-Speed Gearboxes Based on Improved Order Analysis and Adaptive VMD [J]. Journal of Mechanical Transmission, 2021, 45(1): 34-39, 84.

[0019]

[13] CHEN H, JIANG B, LU N.A Multi-mode Incipient Sensor Fault Detectionand Diagnosis Method for Electrical Traction Systems[J].International Journalof Control,Automation,and Systems,2018,16(4):1783-93.

[0020]

[14] CHEN H, JIANG B, LU N,et al.Deep PCA based real-time incipientfault detection and diagnosis methodology for electrical drive in high-speedtrains[J].IEEE Transactions on Vehicular Technology,2018,67(6):4819-30.

[0021]

[15] CHEN H, YI H, JIANG B,et al.Data-Driven Detection of Hot Spots inPhotovoltaic Energy Systems[J].IEEE Trans Systems,Man,and Cybernetics:Systems,2019,49(8).

[0022]

[16] ZHANG K, JIANG B, CHEN F.Multiple-Model-Based Diagnosis of MultipleFaults With High-Speed Train Applications Using Second-Level Adaptation[J].IEEE Transactions on Industrial Electronics,2021,68(7):6257-66. Summary of the Invention

[0023] Aiming at the problem that it is difficult to remove the noise of vibration signals, resulting in low fault diagnosis accuracy, the present invention proposes an adaptive VMD method based on the adjacent modal average repetition ratio, which is used to achieve the purpose of self-determining the VMD modal stratification parameter K according to the input signal, and effectively utilizes the strong adaptability of VMD. Further considering the signal denoising performance, a denoising method for screening noisy components based on the principle of maximum envelope spectrum entropy is designed, which improves the ability to express data fault information. In addition, combining the DC component separation technology and the MAD outlier removal method, an AR-VMD data preprocessing system for bearing variable working condition fault data is constructed, realizing the automated integrated processing of large-scale fault data, and verified by using the fault data of a high-speed bearing test bench. The simulation verification and experimental test results show that the proposed method has higher fault information discrimination ability than traditional data processing methods, and can improve the accuracy of fault diagnosis.

[0024] To achieve the above object, the present invention adopts the following technical solutions:

[0025] A data preprocessing method based on AR-VMD, the steps of determining the optimal stratification number K of VMD include: giving the initial value of K, using VMD to decompose the signal, and calculating the adjacent modal average repetition ratio AR and the modal repetition ratio threshold R of the decomposed signal T When AR is greater than R T Stop the iteration, and the K value at this time is the optimal stratification number;

[0026] The calculation formula of the adjacent modal average repetition ratio is:

[0027]

[0028] In the formula, ω i is the center frequency of the i-th order modal response, and f s is the sampling frequency of the decomposed signal;

[0029] The calculation formula of the modal repetition ratio threshold is:

[0030]

[0031] In the formula, A is the ratio amplification coefficient.

[0032] The preprocessing method for bearing variable working condition fault data applying the above method includes the following steps:

[0033] 1. DC component separation

[0034] When the working conditions change, the bearing is in an unstable state, and the information contained in the collected signal is more complex than that under stable working conditions. According to the analysis of the fault data of bearings under variable working conditions, the vibration center of the bearing signal under variable working conditions shows an obvious shift, that is, the bearing vibration signal contains a DC component at this time. When processing such bearing signals, neither the existence of the DC signal can be ignored and allowed to be confused in the oscillation signal to interfere with feature analysis, nor can it be blindly removed as a useless signal. Therefore, before processing the data of bearings under variable working conditions, the DC component in the data is first separated, and the oscillation signal continues to enter the next step of processing.

[0035] 2. Adjacent modal average repetition ratio (AR) criterion

[0036] According to the basic principle of the VMD method, the decomposition layer number K is one of the most important specified parameters of this method. The value of K determines whether VMD can correctly separate the modal responses of different frequencies. If the value of K is too small, all modal responses contained in the signal cannot be effectively separated; if the value of K is too large, the modal responses of the same frequency are easily mis-separated into different responses. At the same time, manually determining parameters cannot efficiently process large-scale data, nor can it accurately separate the modes of complex signals. Based on the above problems, Yin Hong et al. determined the optimal layer number K of the measurement signal according to the modal repetition ratio criterion to solve the order determination problem of VMD. This method improves the adaptive ability of VMD, but the measurement index for modal aliasing is relatively rough and cannot cope with more complex and extreme data signals. Therefore, following the parameter determination idea of the modal repetition ratio criterion, the adjacent modal average repetition ratio criterion is proposed. The adjacent modal average repetition ratio is calculated using formula (1).

[0037] When the center frequencies of two adjacent modes obtained by VMD processing are different, the value of AR is small; when the center frequencies of two adjacent modes are close or even the same, AR will be a larger value. According to the VMD principle, if the value of K is too large, resulting in the "modal cracking" phenomenon, the center frequencies of the cracked modes will approach the same frequency value. It can be seen from this that AR can quantify the accuracy of the modes obtained by VMD, and the boundary value of K can be determined with the help of this physical quantity. The most effective value of K should be selected as an integer value when AR approaches 0 and K itself is as large as possible. This not only ensures the complete decomposition of complete modal information but also avoids modal cracking. Therefore, as long as an appropriate threshold is selected to compare with the AR value, the appropriateness of the value of parameter K can be judged by comparing the results.

[0038] Regarding the selection principle of the comparison threshold, considering the limitation of the Shannon sampling theorem on the effective frequency of signal decomposition, combined with theoretical analysis and experimental verification, the modal repetition ratio threshold (R T ) is calculated using formula (2).

[0039] In the process of optimizing the VMD method using the adjacent modal average repetition ratio criterion, the comparison of the magnitudes of AR and R T is used as the basis for the iteration stop condition. When AR is greater than R T , the iteration stops, and the value of K at this time is the optimal number of layers.

[0040] 3. Outlier rejection based on the median absolute deviation (MAD)

[0041] In statistics, MAD is a robust measure of the sample deviation of univariate numerical data and is one of the standard features used to describe the variability of univariate samples in quantitative data, often used for outlier detection.

[0042] The calculation formula for MAD is

[0043] MAD = median(|X i - median(X)|) (5)

[0044] In practical applications, generally, the consistent estimator σ of the median absolute deviation is used as the statistical standard, and its calculation formula is

[0045] σ = k·MAD (6)

[0046] where k is a proportionality factor constant, and its value depends on the distribution type. For normally distributed data, the value of k is

[0047]

[0048] That is, the reciprocal of the quantile function of the standard normal distribution (also known as the inverse cumulative distribution function).

[0049] Using the MAD statistic, outliers in the signal can be effectively removed, reducing the interference of random errors in the acquisition process on the signal accuracy. When performing outlier rejection on the signal, only need to calculate the σ value corresponding to each segment of the signal, and then compare the median absolute difference of the sample points with the σ value. If |X i - median(X)| > σ, then let X i = X i-1 . This outlier removal method can ensure the removal of outliers without introducing new outliers. At the same time, the robustness of MAD can ensure that normal samples are not misremoved during the outlier removal process.

[0050] 4. Signal denoising algorithm based on the principle of maximum envelope spectrum entropy

[0051] In information theory, entropy is a very important concept, which represents the overall information measure of a random object on average. For a discrete random variable X, its probability space is represented as

[0052]

[0053] where q represents the total number of discrete components of the variable X, and the probability p satisfies

[0054]

[0055] The information exponential entropy of the probability space is defined as

[0056]

[0057] where p(x i ) is the probability of the occurrence of the i-th element x i . The information exponential entropy can quantify the randomness of a sequence. The larger the information exponential entropy of a sequence, the more random the arrangement of the sequence. The smaller the entropy value, the stronger the periodicity of the sequence.

[0058] When processing bearing fault signals, attention should be paid to distinguishing fault signals from noise signals to prevent filtering out effective signals. The envelope spectrum is more sensitive to faults of impact signals. Compared with spectral analysis, the envelope spectrum can better highlight the fault characteristic frequencies. Generally, the envelope spectrum of the bearing fault response signal should be a discrete spectral line with a periodic distribution. The more obvious the fault, the stronger the periodicity of the spectral line and the smaller the entropy value. Therefore, using the envelope spectrum entropy as the basis for judging the signal nature can more accurately screen out effective signals. The calculation method of the envelope spectrum entropy is as follows:

[0059] Step 1: Perform Hilbert transform on each IMF

[0060]

[0061] Step 2: Take the modulus of the analytic signal and obtain the envelope spectrum through FFT

[0062]

[0063] Step 3: Calculate the energy of each sample point

[0064]

[0065] Step 4: Divide the energy levels according to the peak-to-peak value of the energy spectrum and calculate the occurrence probability of the samples corresponding to each level of energy

[0066]

[0067] Step 5: Calculate the envelope spectrum energy exponential entropy using the obtained probabilities

[0068]

[0069] Signals containing fault information are decomposed into several modal components by the VMD method. These components contain both normal intrinsic signals, fault signals, and noise signals. Correctly and effectively removing noise signals is a necessary condition to ensure the accuracy of fault diagnosis. Most VMD noise reduction methods directly eliminate the modal components most similar to the noise signals, which is likely to cause the loss of high-frequency micro fault signals. Based on the principle of maximum envelope spectrum entropy, this paper improves the traditional VMD noise reduction method and reduces the loss rate of effective signals. The specific process of the signal noise reduction algorithm is as follows:

[0070] Step 1: Decompose the original signal using the AR-VMD method;

[0071] Step 2: Calculate the envelope spectrum entropy of each obtained component;

[0072] Step 3: Compare the envelope spectrum entropy of the components in sequence from low frequency to high frequency, select the component with the maximum envelope spectrum entropy as the signal to be denoised (usually a high-frequency component), and use the wavelet soft threshold denoising method to remove high-frequency noise. If there are multiple components with the same envelope spectrum entropy, all the above components will be included in the range of the signal to be denoised;

[0073] Step 4: Perform time-domain superposition of the denoised component and the low-entropy components, reconstruct all components including the denoised component, and obtain the processed signal.

[0074] 5. Experimental analysis and verification

[0075] For the fault data processing of variable working condition bearings, a data preprocessing system based on AR-VMD is proposed to improve the ability of the signal to express its own fault information, thereby improving the accuracy of fault diagnosis.

[0076] First, construct the simulation test signals required for verification to verify the noise reduction effect of the data preprocessing system based on AR-VMD. The simulation test signals are

[0077] x(t) = x g (t) + x d (t) + x n (t) (16)

[0078] where x g (t) is the effective signal with multi-component characteristics, x d (t) is the simulated vibration DC offset signal, and x g (t) is the noise signal. The specific forms of each part of the signal are

[0079]

[0080] where A k 、f k are the amplitudes and frequencies of each effective component; A d, ω d is the amplitude and period of the sawtooth DC signal; r(t) is Gaussian noise in the interval [-1, 1], (t0 < t < t1) is the instantaneous square wave signal at time [t0, t1], representing the outliers in the signal, and β1 and β2 are the amplitudes of the two signals respectively. The effective signal frequencies are preset to 4 frequency bands, namely 200Hz, 1000Hz, 2000Hz, and 4000Hz respectively, the DC component frequency is set to 5Hz, the abnormal time is set to [0.300, 0.301], and the signal acquisition frequency f s = 10000Hz.

[0081] The data is input into the AR-VMD data preprocessing system to obtain the processed data results. Compared with the Fourier spectrum, the HHT marginal spectrum has stronger ability to process non-stationary signals and can accurately reflect the actual frequency components of the fault signals. Therefore, the marginal spectrum is selected for frequency domain analysis. Four frequency components are accurately separated from the processed signal on the marginal spectrum, and the irrelevant noise signals in the low frequency band are effectively filtered. It can be seen that the AR-VMD data preprocessing system has good noise reduction effect and realizes automatic data processing.

[0082] The bearing fault data set collected by a high-speed bearing test bench is processed, and the data acquisition frequency f s = 51200. First, the normal data and fault data under the same working conditions are selected for processing and comparison to verify whether the AR-VMD data preprocessing method can accurately identify the signal modal numbers of different state data. Secondly, 3 groups of experimental data under variable speed working conditions are selected to verify the modal extraction effect of the AR-VMD data preprocessing method on non-stable signals. The experimental results prove that the AR-VMD data preprocessing method can exclude the interference of external irrelevant factors and has good signal purification effect.

[0083] The effect verification of fault diagnosis is carried out by using a fault diagnosis network based on CNN. The test data comes from an independently developed aeroengine bearing fault diagnosis test platform. The vibration signals used are normal signals and fault signals of 3 bearing parts, namely inner ring, roller, and outer ring. Among them, the fault signals with damage depths of 1mm, 2mm, and 3mm are collected for each part, with a total of 10 fault categories. From the diagnosis results, it can be seen that the fault signals preprocessed by AR-VMD have enhanced ability to express fault information. Under the same fault diagnosis network, the fault diagnosis accuracy obtained from the preprocessed data is higher, and the effectiveness of the AR-VMD preprocessing method is verified.

[0084] Beneficial effects

[0085] In view of the problem that the traditional VMD has poor noise reduction effect on bearing fault data under large-scale variable working conditions, this invention proposes a preprocessing method for bearing fault data based on AR-VMD to process bearing fault signals. The main contributions of this invention are as follows:

[0086] 1) Propose an AR-VMD method based on the adjacent mode average repetition ratio to solve the problem of difficult determination of the K value of the VMD mode stratification number;

[0087] 2) Screen the noisy components according to the principle of maximum envelope spectrum entropy and perform noise reduction processing to eliminate irrelevant noise interference and highlight the data fault information;

[0088] 3) Based on the above method, add the DC component separation technology and the MAD outlier removal method to systematically integrate the data preprocessing process for bearing variable working condition fault data;

[0089] 4) In order to verify the accuracy of the fault signal extraction of this invention, use simulation test signals and experimental data to verify the method effect at the same time, and prove that this invention can effectively improve the fault diagnosis accuracy and has certain application value in engineering. Description of the Drawings

[0090] Figure 1 Flow chart of the adaptive VMD method based on the adjacent mode average repetition ratio;

[0091] Figure 2 Original signal of DC component separation;

[0092] Figure 3 DC component of DC component separation;

[0093] Figure 4 Oscillation component of DC component separation;

[0094] Figure 5 Data preprocessing process for bearing variable working condition fault data;

[0095] Figure 6 Time domain curve of the simulation test signal;

[0096] Figure 7 Frequency domain curve of the simulation test signal;

[0097] Figure 8 Time domain diagram of the processing result of the test signal;

[0098] Figure 9 HHT marginal spectrum of the processing result of the test signal;

[0099] Figure 10 Time domain diagram of the same working condition normal data experiment;

[0100] Figure 11Time-domain diagram of fault data under the same working conditions experiment;

[0101] Figure 12 Marginal spectrum diagram of normal data experiment under the same working conditions;

[0102] Figure 13 Marginal spectrum diagram of fault data experiment under the same working conditions;

[0103] Figure 14 Time-domain diagram of experimental data under the constant speed condition of 1000 rpm;

[0104] Figure 15 Time-domain diagram of experimental data under the constant speed condition of 2000 rpm;

[0105] Figure 16 Time-domain diagram of experimental data under the speed-up condition from 1000 rpm to 2000 rpm;

[0106] Figure 17 Marginal spectrum diagram of experimental data under the constant speed condition of 1000 rpm;

[0107] Figure 18 Marginal spectrum diagram of experimental data under the constant speed condition of 2000 rpm;

[0108] Figure 19 Marginal spectrum diagram of experimental data under the speed-up condition from 1000 rpm to 2000 rpm;

[0109] Figure 20 Fault diagnosis network structure;

[0110] Figure 21 Confusion matrix of the fault diagnosis network results for the original data;

[0111] Figure 22 Confusion matrix of the fault diagnosis network results for the preprocessed data. Detailed implementation method

[0112] Based on the adjacent mode average repetition ratio, the process of designing the adaptive VMD method is as Figure 1 shown.

[0113] Considering Figure 2 the bearing fault data shown, it can be seen that it contains a certain DC component. After separating the DC component, Figure 3 the DC component and Figure 4 the oscillation component are obtained. The modal decomposition of the oscillation component will be clearer and more effective. Therefore, the separation of the DC component can improve the accuracy of signal frequency domain analysis.

[0114] The data preprocessing process for the bearing fault data under variable working conditions is as Figure 5As shown in the figure. The specific process of the data preprocessing system for variable working condition bearing fault data is as follows: After the DC component is separated from the original signal, the oscillating signal enters the AR-VMD iterative process, and the signal with multi-component characteristics is decomposed into multiple effective components; after all components remove outliers according to MAD, the envelope spectrum entropy is calculated respectively, and the noisy components are screened for noise reduction. Finally, the data preprocessing effect is analyzed through the HHT marginal spectrum.

[0115] The time-frequency domain form of the simulation test signal is as Figures 6 - 7 shown. The effective signal frequencies are preset to 4 frequency bands, which are 200Hz, 1000Hz, 2000Hz, and 4000Hz respectively. The DC component frequency is set to 5Hz, the abnormal time is set to [0.300, 0.301], and the signal acquisition frequency f s = 10000Hz.

[0116] Input the data into the AR-VMD data preprocessing system, and the processed data is as Figures 8 - 9 shown. Figures 8 - 9 It shows that the DC component in the time domain diagram is screened out, and the abnormal segment response is also suppressed to a certain extent. The four frequency components are accurately separated on the marginal spectrum of the processed signal, and the irrelevant noise signals in the low frequency band are effectively filtered.

[0117] Process the bearing fault data set collected by the high-speed bearing test bench. The data acquisition frequency f s = 51200. First, select the normal data and fault data under the same working condition for processing and comparison. The experimental results are as Figures 10 - 13 shown. The figure shows that the AR-VMD data preprocessing method can accurately identify the number of signal modes of data in different states, achieving the purpose of modal decomposition adaptability.

[0118] Select 3 groups of experimental data under variable speed working conditions. The corresponding working conditions are 1000 rpm constant speed, 2000 rpm constant speed, and 1000 → 2000 rpm speed increase. The data processing results are as Figures 14 - 19 . The figure shows that for the modal extraction of non-stable signals, the AR-VMD data preprocessing method can exclude the interference of external irrelevant factors, and the signal purification effect is good.

[0119] Use the fault diagnosis network based on CNN for diagnosis verification. The network structure is as Figure 20As shown below. The fault diagnosis network structure is as follows: First, the one-dimensional vibration signal is randomly sampled to generate training and test samples, and then converted into a two-dimensional feature map of 32×32; then the two-dimensional features are input into the three channels of the multi-scale feature extraction module for convolutional fusion, and an attention mechanism module is added to highlight the effective information; then the result is input into the Inception module to achieve dimensionality reduction convolution, and finally the final output is obtained through the fully connected layer. The original signal and the preprocessed signal are respectively input into the diagnosis network. The diagnosis accuracy is shown in Table 1, and the confusion matrix diagram is as Figures 21 - 22 shown.

[0120] Table 1 Comparison of Fault Diagnosis Accuracy

[0121]

Claims

1. A preprocessing method for bearing fault data under variable working conditions, characterized in that, It includes the following steps: collect the variable working condition bearing fault signal. After the DC component of the signal is separated, the oscillating signal therein enters the AR-VMD iterative process, and the signal with multi-component characteristics is decomposed into multiple effective components; after removing outliers from all components using the absolute median difference algorithm, calculate the envelope spectrum entropy of the components after removing outliers respectively, select the component with the largest envelope spectrum entropy for wavelet soft threshold denoising processing, then perform time-domain superposition of the denoised component and the low-entropy component, and reconstruct all components including the denoised component to obtain the processed signal. The AR-VMD iterative process is as follows: The steps to determine the optimal number of VMD layers K include: Given the initial value of K, use VMD to decompose the signal, and calculate the adjacent mode average repetition ratio AR and the mode repetition ratio threshold R of the decomposed signal T , when AR is greater than R T , stop the iteration, and the value of K at this time is the optimal number of layers; The calculation formula of the adjacent mode average repetition ratio is: where ω i is the center frequency of the i-th order modal response, and f s is the sampling frequency of the decomposed signal; The calculation formula of the mode repetition ratio threshold is: In the formula, A is the ratio amplification coefficient.

2. The variable operating condition bearing fault data preprocessing method according to claim 1, wherein The absolute median difference algorithm is used to remove outliers. The specific steps include: calculating the corresponding absolute median consistent estimator σ for each component X after signal decomposition, and then calculating the sampling value X of each component at each moment. i The absolute difference of the median value of is compared with the σ value. If the difference exceeds the σ value, the sample value X at the previous moment is used. i-1 Overwrite the current sample value, and so on to complete the elimination operation of a component.

3. The variable operating condition bearing fault data preprocessing method according to claim 1, wherein, Perform denoising processing using the signal denoising algorithm based on the principle of the largest envelope spectrum entropy. The specific steps include: calculate the envelope spectrum entropy of the obtained components respectively, select the component with the largest envelope spectrum entropy for wavelet soft threshold denoising processing, and reconstruct all components including the denoised component to obtain the processed signal.

4. The variable working condition bearing fault data preprocessing method according to claim 3, wherein, The calculation method of the envelope spectrum entropy is as follows: Step 1: Perform Hilbert transform on each IMF component where: n = 0, 1, 2, …, N - 1, N is the number of signal samples, j is the imaginary symbol, τ is the integral quantity, and imf(n) is the IMF component of the nth signal sample; Step 2: Take the modulus of the analytical signal and obtain the envelope spectrum through FFT Step 3: Calculate the energy of each sample point Step 4: Divide the energy levels according to the peak-to-peak value of the energy spectrum and calculate the occurrence probability of the samples corresponding to each level of energy where i is the number of sample points, D is the total number of energy levels, and d is the order corresponding to the currently calculated level; Step 5: Calculate the envelope spectrum energy index entropy using the obtained probability 5. The method for preprocessing variable operating condition bearing fault data according to claim 3, characterized in that, Perform time-domain superposition of the denoised component and the low-entropy component, and reconstruct all components including the denoised component.

6. A variable working condition bearing fault diagnosis method applying the preprocessing method described in any one of claims 2 - 5.

Images