Bearing fault feature extraction method based on improved period estimation FMD
By estimating the period of bearing fault signal by using spectrum analysis and frequency selection methods in FMD, the problem of insufficient feature extraction capability in FMD in high noise environment is solved, and stronger noise robustness and fault feature extraction effect is achieved.
Patent Information
- Application Number
- CN202411047169.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-05
AI Technical Summary
The existing FMDs are difficult to accurately lock the period of bearing fault signals in high noise environments, affecting their fault feature extraction capabilities.
The fault period of the signal is estimated based on spectrum analysis and frequency selection, the original autocorrelation estimation method is replaced, and the improved FMD algorithm model is trained and iterated multiple times to extract bearing fault characteristics.
It improves FMD's fault feature extraction capability and noise robustness in high noise environments, can accurately lock the cycle of the fault signal, and effectively extract the fault feature frequency and frequency doubling.
Smart Images

Figure CN118897979A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a bearing fault feature extraction method based on improved period estimation FMD. Background Art
[0002] Under fault conditions, the bearing damage point will cause high-frequency vibration signals. In addition, the damage point will collide with other components to generate periodic impact vibration signals. These two types of vibration signals are important characteristic signals for bearing fault diagnosis and analysis. However, due to the harsh conditions of the actual industrial environment and the mutual interference between complex mechanical equipment components, the measured bearing vibration signal contains a lot of noise and exhibits nonlinear and non-stationary characteristics. In this context, fault feature extraction has become a popular direction for processing such signals.
[0003] Modal decomposition algorithm is one of the commonly used research methods for bearing fault feature extraction. Jing Zhi et al. used the empirical mode decomposition algorithm (EMD) to decompose the rolling bearing fault signal, extract effective feature data, and improve the classification accuracy of the classifier to 90%. However, EMD has problems such as endpoint effect and modal aliasing. To this end, scholars have proposed improved methods, such as ensemble empirical mode decomposition (EEMD) and variational mode decomposition. Jin Zhihao et al. proposed a rolling bearing fault diagnosis method that combines EEMD, fast spectral kurtosis frequency selection and resonance demodulation technology. The results show that EEMD suppresses modal aliasing to a certain extent, but increases the amount of calculation and cannot completely neutralize the added white noise. Dong Lunan et al. used VMD to decompose noisy vibration signals and reconstruct the signals according to the correlation, and added the reconstructed signals to the densely connected network for fault identification. The experimental results show that VMD suppresses modal aliasing and improves the noise robustness of the densely connected network to a certain extent. However, the initial center frequency of VMD is difficult to determine under strong noise background, which will greatly reduce the decomposition quality. In recent years, Pan et al. proposed the Symplectic Geometry Mode Decomposition (SGMD) method and applied it to fault diagnosis. SGMD uses the power spectral density method to adaptively determine the embedding dimension, reducing the uncertainty caused by artificial settings. However, SGMD still relies on the similarity of frequency and correlation coefficients when constructing symplectic geometry components, and requires artificial setting of iteration termination conditions, resulting in uncertainty in its analysis results. In 2022, Miao proposed a new decomposition theory, Feature Mode Decomposition (FMD), which is specifically used for feature extraction of mechanical faults. The main purpose of the proposed FMD is to decompose different modes through the designed adaptive finite impulse response (FIR) filter. FMD takes advantage of the correlation kurtosis and considers both the impulsiveness and periodicity of the fault signal. First, the FIR filter bank is designed using Hanning window initialization to provide direction for decomposition. Then, the period estimation and update process is used to lock the fault information. Finally, redundant and mixed modes are removed in the mode selection process. By using simulation and experimental data of single and combined faults of bearings, it is proved that FMD can adaptively and accurately decompose fault modes and is robust to other interferences and noise.
[0004] Although FMD performs well in fault feature extraction, its period estimation accuracy is still limited in the presence of high noise background.
[0005] Therefore, a bearing fault feature extraction method based on improved period estimation FMD is provided. Summary of the invention
[0006] The purpose of the present invention is to overcome the existing defects and provide a bearing fault feature extraction method based on improved period estimation FMD, which accurately locks the period of the fault signal, thereby improving the feature extraction capability of FMD in a high noise environment.
[0007] The technical solution to achieve the above purpose is:
[0008] The bearing fault feature extraction method based on improved period estimation FMD includes:
[0009] Step S1, obtain a data set, label and classify the data set, divide it into a training set and a test set in proportion, and preprocess the data set;
[0010] Step S2, constructing an FMD algorithm model for improving cycle estimation;
[0011] Step S3, training the constructed FMD algorithm model for improved cycle estimation multiple times through the training set, saving the best result, and saving the training parameters;
[0012] Step S4, input the test set into the trained FMD algorithm model for improved cycle estimation to extract bearing fault features.
[0013] Preferably, in step S1, the data set adopts the Case Western Reserve University bearing data set, and the fault types of inner ring fault, rolling element fault, outer ring fault with load area at 3 o'clock, outer ring fault with load area at 6 o'clock and outer ring fault with load area at 12 o'clock are selected for feature extraction; the preprocessing method includes data enhancement and normalization processing.
[0014] Preferably, in step S2, the FMD algorithm model for improving period estimation is to replace the autocorrelation function used in the original FMD algorithm to estimate the fault period with a method based on spectrum analysis and frequency selection to estimate the fault period of the signal; the specific implementation is as follows:
[0015] By inputting the bearing fault signal and the sampling frequency f s Calculate the length N and frequency array F of the signal, and obtain the envelope by calculating the difference of the signal and applying the Hilbert transform, and then perform a fast Fourier transform to obtain the spectrum amplitude of the signal difference envelope;
[0016] Find the first local maximum point in the spectrum amplitude array, and set the spectrum components of subsequent values that are smaller than the previous value to zero;
[0017] Arrange the spectrum amplitudes in descending order and select the first ten largest frequency components as candidate frequencies F active , and remove frequencies that are very close to each other;
[0018] Create a matrix Z and a multiple relationship matrix C. If F active [j] is F active An integer multiple of [i], then Z ij =1, then C ij =F active [j] / F active [i], otherwise Z ij =0, where Z ij The element represents the candidate frequency F active [i] and F active Whether there is an integer multiple relationship between [j], C ij The element represents the candidate frequency F active [j] and F active [i] The multiple relationship between
[0019]
[0020] By calculating the modulus relationship between the candidate frequencies, check whether these candidate frequencies have an integer multiple relationship, and record the candidate frequency pairs Z that meet the conditions sum ;
[0021] Z sum =∑ j Z ij ;
[0022] On the basis of the above, calculate the sum of each row in the matrix Z, and select the frequency component corresponding to the row with the largest sum, which is the most relevant frequency Z ind ;
[0023] Z ind = arg max(Z sum );
[0024] According to the most relevant frequency components and their multiple relationships, the final fault frequency F is calculated. n and through the fault frequency F n Get the fault period T of the signal;
[0025] F res =F active ;
[0026] G res =;
[0027]
[0028] Among them, F res For Cres for.
[0029] Preferably, in step S3, decomposition and pattern extraction are performed using the input bearing fault signal and the fault period T of the signal until the maximum number of iterations is reached or the pattern quantity requirement is met, thus completing the training of the FMD algorithm model for improved period estimation.
[0030] Preferably, step S4 comprises:
[0031] Input the original vibration signal x of the bearing and initialize the parameters of the improved period estimation FMD algorithm model, including but not limited to the number of decomposed modes n and the filter length L;
[0032] Initialize the FIR filter bank using K Hanning windows, set to 5 to 10 groups, and then initialize the iteration a = 1;
[0033] Through convolution operation Get the filtered signal, that is, the decomposed modal components Where k = 1, 2, ..., K;
[0034] Estimating the fault period of a signal by spectrum analysis and frequency selection
[0035] By inputting the original vibration signal x, the decomposed modal components and estimated failure cycles Update the filter coefficients and set a=a+1 for the next iteration;
[0036] Determine whether the current iteration has reached the maximum number of iterations. If not, return to the convolution operation. Obtaining a filtered signal;
[0037] If the maximum number of iterations is reached, the correlation coefficient between each two modal components is calculated to construct a K×K correlation matrix CC (K×K) ;
[0038] Determine whether the current number of modes K reaches the set number of modes n. If not, return to the convolution operation. The filtered signal is obtained, otherwise, the iteration is stopped, and n modal components are obtained as the final decomposition results of the improved period estimation FMD algorithm model, that is, the bearing fault feature extraction results.
[0039] Preferably, in step S4, if the maximum number of iterations is reached, the correlation coefficient between every two modal components is calculated to construct a K×K correlation matrix CC (K×K) , select the two modal components with the largest correlation coefficients and calculate the failure period by the estimated Calculate their related kurtosis, and select the modal component with larger related kurtosis as the decomposed modal component. The envelope spectrum kurtosis factor is used as a screening index. It uses the Hilbert transform to obtain the envelope of the signal, and then calculates the kurtosis of the envelope. The calculation formula is as follows:
[0040]
[0041] Among them, amp q is the envelope of the qth modal signal obtained by Hilbert transform.
[0042] The beneficial effects of the present invention are as follows: the present invention estimates the period of a signal through a method based on spectrum analysis and frequency selection, and replaces the original autocorrelation estimation period of FMD with the method to extract bearing fault features. The FMD algorithm with improved period estimation has stronger fault feature extraction capability and noise robustness in a noisy environment, can effectively extract fault feature frequencies and their multiples, and accurately locks the period of the fault signal, thereby improving the feature extraction capability of FMD in a high noise environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a flow chart of a bearing fault feature extraction method based on improved period estimation FMD of the present invention;
[0044] Figure 2 The five modal time domain diagrams are obtained by decomposing the noisy signal using the improved period estimation FMD algorithm model in the present invention;
[0045] Figure 3 The five modal frequency domain diagrams are obtained by decomposing the noisy signal using the improved period estimation FMD algorithm model in the present invention;
[0046] Figure 4 The present invention decomposes the noisy signal using the improved period estimation FMD algorithm model to obtain a histogram of the envelope spectrum kurtosis factors of the five modes;
[0047] Figure 5 The envelope spectrum of mode 1 is obtained by decomposing the noisy signal using the improved period estimation FMD algorithm model in the present invention;
[0048] Figure 6 The time-frequency diagram of mode 1 is obtained by decomposing the noisy signal using the improved period estimation FMD algorithm model in the present invention;
[0049] Figure 7 The present invention adds -14dB Gaussian white noise to the impulse signal to simulate the envelope spectrum of the impulse signal in a noise environment;
[0050] Figure 8The present invention adds -14dB Gaussian white noise to the impulse signal to simulate the time-frequency diagram of the impulse signal in a noise environment;
[0051] Fig. 9 It is the envelope spectrum of the original FMD denoising simulation signal under the premise that the parameter setting and modal screening index in the present invention remain unchanged;
[0052] Fig.10 It is a time-frequency diagram of the original FMD denoising simulation signal under the premise that the parameter setting and modal screening index in the present invention remain unchanged;
[0053] Fig.11 It is the envelope spectrum of the noisy inner ring fault signal in the present invention;
[0054] Fig.12 It is a time-frequency diagram of the noisy inner ring fault signal in the present invention;
[0055] Fig.13 It is a time domain diagram of each mode of the inner ring fault signal decomposed by the improved period estimation FMD in the present invention;
[0056] Fig.14 It is a frequency domain diagram of each mode of the inner ring fault signal decomposed by the improved period estimation FMD in the present invention;
[0057] Fig.15 It is a histogram of the kurtosis factors of each modal spectrum of the inner race fault signal decomposed by the improved period estimation FMD in the present invention;
[0058] Fig.16 is the envelope spectrum of the inner race fault reconstruction signal in the present invention;
[0059] Fig.17 It is the envelope spectrum of the outer ring fault feature extraction situation of the load area at 6 o'clock in the present invention;
[0060] Fig.18 It is the envelope spectrum of the outer ring fault feature extraction situation of the load area at 12 o'clock in the present invention;
[0061] Fig.19 It is the envelope spectrum of the outer ring fault feature extraction situation of the load area at 3 o'clock in the present invention;
[0062] Fig. 20 It is the envelope spectrum of the rolling element fault feature extraction in the present invention. DETAILED DESCRIPTION
[0063] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.
[0064] The present invention will be further described below in conjunction with the accompanying drawings.
[0065] like Figure 1 As shown, the bearing fault feature extraction method based on the improved period estimation FMD includes:
[0066] Step S1, obtain a data set, label and classify the data set, divide it into a training set and a test set in proportion, and preprocess the data set.
[0067] In the embodiment, the data set adopts the Case Western Reserve University bearing data set, and the fault types of inner ring fault, rolling element fault, outer ring fault with load area at 3 o'clock, outer ring fault with load area at 6 o'clock and outer ring fault with load area at 12 o'clock are selected for feature extraction; the preprocessing method includes data enhancement and normalization processing.
[0068] Step S2, constructing an FMD algorithm model for improving cycle estimation.
[0069] In the embodiment, the FMD algorithm model for improving period estimation replaces the autocorrelation function used in the original FMD algorithm to estimate the fault period with a method based on spectrum analysis and frequency selection to estimate the fault period of the signal; the specific implementation is as follows:
[0070] By inputting the bearing fault signal and the sampling frequency f s Calculate the length N and frequency array F of the signal, and obtain the envelope by calculating the difference of the signal and applying the Hilbert transform, and then perform a fast Fourier transform to obtain the spectrum amplitude of the signal difference envelope;
[0071] Find the first local maximum point in the spectrum amplitude array, and set the spectrum components of subsequent values that are smaller than the previous value to zero;
[0072] Arrange the spectrum amplitudes in descending order and select the first ten largest frequency components as candidate frequencies F active , and remove frequencies that are very close to each other;
[0073] Create a matrix Z and a multiple relationship matrix C. If F active [j] is F active An integer multiple of [i], then Z ij =1, then C ij =F active [j] / F active [i], otherwise Z ij =0, where Z ij The element represents the candidate frequency F active [i] and F active Whether there is an integer multiple relationship between [j], C ij The element represents the candidate frequency F active [j] and F active [i] The multiple relationship between
[0074]
[0075] By calculating the modulus relationship between the candidate frequencies, check whether these candidate frequencies have an integer multiple relationship, and record the candidate frequency pairs Z that meet the conditions sum ;
[0076] Z sum =∑ j Z ij ;
[0077] On the basis of the above, calculate the sum of each row in the matrix Z, and select the frequency component corresponding to the row with the largest sum, which is the most relevant frequency Z ind ;
[0078] Z ind = arg max(Z sum );
[0079] According to the most relevant frequency components and their multiple relationships, the final fault frequency F is calculated. n and through the fault frequency F n Get the fault period T of the signal;
[0080] F res =F active ;
[0081] G res =;
[0082]
[0083] Among them, F res For C res for.
[0084] Step S3, training the constructed FMD algorithm model for improved cycle estimation multiple times through the training set, saving the best result, and saving the training parameters.
[0085] In the embodiment, decomposition and pattern extraction are performed through the input bearing fault signal and the fault period T of the signal until the maximum number of iterations is reached or the pattern quantity requirement is met, that is, the training of the FMD algorithm model for improving period estimation is completed.
[0086] Step S4, input the test set into the trained FMD algorithm model for improved cycle estimation to extract bearing fault features.
[0087] In the embodiment, bearing fault feature extraction is performed by improving the FMD algorithm model of cycle estimation, including:
[0088] The original vibration signal x of the bearing is input, and the parameters of the improved period estimation FMD algorithm model are initialized, and the parameters include but are not limited to the number of decomposed modes m and the filter length L.
[0089] The FIR filter bank is initialized using K Hanning windows, which are set to 5 to 10 groups, and then the iteration a=1 is initialized.
[0090] Through convolution operation Get the filtered signal, that is, the decomposed modal components Where k = 1, 2, ..., K.
[0091] Estimating the fault period of a signal by spectrum analysis and frequency selection
[0092] By inputting the original vibration signal x, the decomposed modal components and estimated failure cycles Update the filter coefficients and set a=a+1 for the next iteration.
[0093] Determine whether the current iteration has reached the maximum number of iterations. If not, return to the convolution operation. Get the filtered signal.
[0094] If the maximum number of iterations is reached, the correlation coefficient between each two modal components is calculated to construct a K×K correlation matrix CC (K×K) .
[0095] In the embodiment, if the maximum number of iterations is reached, the correlation coefficient between every two modal components is calculated to construct a K×K correlation matrix CC (K×K) , select the two modal components with the largest correlation coefficients and calculate the failure period by the estimated Calculate their related kurtosis, and select the modal component with larger related kurtosis as the decomposed modal component. The envelope spectrum kurtosis factor is used as a screening index. It uses the Hilbert transform to obtain the envelope of the signal, and then calculates the kurtosis of the envelope. The calculation formula is as follows:
[0096]
[0097] Among them, amp q is the envelope of the qth modal signal obtained by Hilbert transform.
[0098] Determine whether the current number of modes K reaches the set number of modes n. If not, return to the convolution operation. The filtered signal is obtained, otherwise, the iteration is stopped, and n modal components are obtained as the final decomposition results of the improved period estimation FMD algorithm model, that is, the bearing fault feature extraction results.
[0099] Improved periodic FMD algorithm feature extraction effect:
[0100] The decomposition mode number of the improved period FMD is set to 5, the filter length is 8, and the truncation number is 6. The noisy signal is decomposed using the improved period FMD algorithm model proposed in this paper, and the following is obtained: Figure 2 , 3 Time domain and frequency domain diagrams of the five modes shown.
[0101] The envelope spectrum kurtosis factors of the five modes are obtained by the above formula. The solution is as follows: Figure 4 The envelope spectrum kurtosis factor has strong adaptability, sensitivity and noise resistance when screening modes, so the mode with the largest envelope spectrum kurtosis factor means that it contains enough fault information. Based on this, the present invention selects the mode corresponding to the largest envelope spectrum kurtosis factor as the reconstructed signal.
[0102] Furthermore, the envelope spectrum analysis and time-frequency diagram of mode 1 are performed to obtain Figure 5 and Figure 6 .contrast Figure 5 and Figure 7 It can be found that the method proposed in the present invention greatly reduces the noise information near the fault frequency, thereby more clearly extracting the 2nd to 5th times of the fault frequency. Figure 6 and Figure 8 By comparison, it is found that the method proposed in the present invention reduces the noise points of the noisy signal in the time and frequency domains, and reproduces the highlight area at the center frequency of 2500HZ, which means that the method proposed in the present invention reduces the signal noise, retains the fault information, and makes the fault frequency easier to identify.
[0103] Finally, in order to verify the superiority of the improved period estimation algorithm proposed in this paper over the autocorrelation period estimation algorithm, under the premise of keeping the above parameter settings and modal screening indicators unchanged, the original FMD is used to perform modal decomposition on the noisy signal, and the envelope spectrum and time-frequency diagram of the reconstructed signal are obtained as shown in Figures 9 and 10. It can be seen that the original FMD extracts the 2 times of the fault frequency, but the 3-5 times are submerged by the noise; Figure 6 and Fig. 9 By comparison, the reconstructed signal extracted by the original FMD has a wider highlighted frequency band in the time-frequency diagram, and there are fewer black bands in the entire frequency band, which means that the reconstructed signal contains more noise information and it is difficult to reflect the complete fault information.
[0104] In order to further quantify the difference in denoising effect between the FMD with improved period estimation and the original FMD, the signal-to-noise ratio (SNR) is selected as the evaluation index of denoising effect. Through calculation, it is known that the SNR of the noisy signal is -12dB, which means that the power of the noise signal is 101.2 times that of the fault characteristic signal, that is, 15.949 times, and the noise background is strong; after FMD denoising, the SNR is -5.79dB, and the power of the noise signal is 3.793 times that of the fault characteristic signal, reducing the noise by 6.21dB; after FMD denoising with improved period estimation, the SNR is -2.77dB, and the power of the noise signal is The fault characteristic signal is 1.892 times, reducing the noise by 9.23dB. Then, the estimated value of the period is observed. Since the fault frequency of the simulation signal is set to 160HZ and the sampling frequency is 12000HZ, the theoretical value of the fault period is 75, which is much lower than the 692 estimated by FMD. In addition, after running FMD for many times, it is found that the period value estimated by FMD is different each time and is distributed between 400-1200. This shows that the autocorrelation estimation period is affected by strong background noise, and the iterative results fail to converge, making it difficult to accurately estimate the fault period, thereby reducing the ability of FMD to extract fault features. In contrast, the method proposed in the present invention has the same running results as the theoretical values, which means that the FMD with improved period estimation has a certain noise robustness and can better extract fault features from noise.
[0105] The comparison table of quantitative index values of the reconstructed signals by the two methods is shown in Table 1 below.
[0106]
[0107]
[0108] Table 1
[0109] Compared with other feature extraction algorithms:
[0110] Under the premise of ensuring reasonable parameter settings and unchanged modal screening indicators, this section uses the commonly used EMD, EEMD and VMD to extract the fault information of the simulation signal respectively, and then compares the envelope spectra and time-frequency diagrams of the signals reconstructed by the three algorithms with the Figure 5 and Figure 6 In addition, SNR is used to quantitatively evaluate the extraction effects of the three algorithms.
[0111] It can be found that the FMD with improved period estimation shows significant advantages in noise reduction and feature extraction, and can better extract the fault characteristic frequency and its multiples; in contrast, the fault feature extraction ability of EMD is weak in a strong noise background, the extracted characteristic frequency multiples are less and the noise is larger; the characteristic frequency extraction ability of EEMD is slightly better than that of EMD, but there are still many noise points; VMD is also effective in noise reduction and feature extraction, second only to the FMD with improved period estimation.
[0112] The data set uses the Case Western Reserve University bearing data set. The installed bearing model is SKF6205-2RS deep groove ball bearing. Its detailed parameters are shown in Table 2. The fault is set using electrospark machining technology. The fault sizes are 0.007 inches, 0.014 inches, 0.021 inches and 0.028 inches. Considering that the smaller the fault size, the less obvious its fault characteristics are and the easier it is to be covered by noise, this section selects the signal with a fault size of 0.007 inches for feature extraction to test the effect of the method proposed in this paper; in addition, a data set with a sampling frequency of 12kHz and a motor speed of 1797 rpm is selected for experiments. The sampling points of the present invention are set to 10000, and the fault types of inner ring fault, rolling element fault, outer ring fault with a load area at 3 o'clock, outer ring fault with a load area at 6 o'clock and outer ring fault with a load area at 12 o'clock are selected for feature extraction.
[0113]
[0114] Table 2
[0115] Motor speed f r , the calculation formulas for the outer ring fault characteristic frequency BPFO, the inner ring fault characteristic frequency BPFI and the rolling element fault BSF are as follows:
[0116]
[0117] Substituting the relevant data in Table 2 into the above formula, it can be calculated that the characteristic frequency of the inner ring fault is 162.186Hz, the characteristic frequency of the outer ring fault is 107.384Hz, and the characteristic frequency of the rolling element fault is 70.584Hz. Next, the FMD feature extraction of the inner ring fault is improved by period estimation; in order to simulate the real production environment, -5dB Gaussian white noise is added to generate the envelope spectrum and time-frequency diagram of the signal as shown in the figure. Fig.11 , 12 As shown. Fig.11 It can be seen that the 1-2 times of the characteristic frequency of the inner race fault are more obvious, and the other times are submerged in the noise; in addition, there are a large number of noise points in the time-frequency diagram, so feature extraction is required.
[0118] After many experiments, the improved FMD parameters are set as follows: the number of decomposed modes is 5, the filter length is 12, and the number of truncation is 6. The improved FMD is used to perform modal decomposition on the noisy inner circle signal, and the following is obtained: Fig.13 , 14 The time domain and spectrum diagrams of each mode are shown.
[0119] Calculate the spectral kurtosis factor of each mode, and the calculation results are as follows Fig.15 As shown, the spectral kurtosis factor of mode 3 is the largest, so mode 3 is selected as the reconstructed signal.
[0120] like Fig.16 To reconstruct the envelope spectrum of the signal, it can be seen that the improved period estimation FMD can completely extract the fault feature frequency of 1-6 times of the inner race fault, and has a good fault feature extraction effect.
[0121] Other fault feature extraction situations:
[0122] Next, the feature extraction of the other four faults is performed, and each fault signal is still added with -5dB Gaussian white noise to simulate the actual fault signal. After multiple experiments, the improved FMD parameter setting table for different faults is obtained, as shown in Table 3.
[0123]
[0124] Table 3
[0125] After performing improved FMD feature extraction on each fault, the corresponding reconstructed signal is obtained and the envelope spectrum is drawn respectively, as shown in Figure 17-20As shown, it can be seen that the outer ring fault signal at 6 o'clock in the load area is extracted by the improved FMD with 1-8 times the frequency, and the feature extraction is good; the outer ring fault at 12 o'clock in the load area is extracted with 1-4 times the frequency; the outer ring fault at 3 o'clock in the load area is extracted with 1 times and 4 times the frequency, and the 2 times and 3 times the frequency are not obvious. Considering that the above three are all outer ring faults, their fault characteristic frequency is 107.384, and the fault cycle can be obtained as 112 (retain integers). The outer ring fault cycle of the load area at 6 o'clock is estimated to be 111, which is basically close to the theoretical value, and the extraction effect is good; the cycle at 12 o'clock is estimated to be 178, and at 3 o'clock is 441, both of which deviate from the theoretical value, and the corresponding extraction results are also poor; this is because the fault characteristics at 6 o'clock and 12 o'clock are not obvious, and the -5dB Gaussian white noise makes it difficult for the improved FMD to extract fault characteristics. For the final rolling element failure, no obvious characteristic frequency was extracted after multiple tests, which once again verified the deficiency that the improved FMD was difficult to extract weak fault features under a strong noise background.
[0126] The present invention proposes a method based on spectrum analysis and frequency selection to estimate the period of a signal, and uses the method to replace the original autocorrelation estimation period of FMD.
[0127] In order to verify the advantages of this improvement, a single bearing periodic impact fault simulation signal is constructed, and the feature extraction of the simulation signal is performed using the improved periodic FMD algorithm, and the feature extraction effect is compared with that of the original FMD; the simulation results show that the improved periodic estimation FMD algorithm has stronger fault feature extraction ability and noise robustness in a noisy environment, and can effectively extract the fault characteristic frequency and its multiples. Then, under the premise of ensuring reasonable parameter settings and unchanged modal screening indicators, the present invention uses the commonly used EMD, EEMD and VMD to extract the fault information of the simulation signal respectively, and compares it with the improved periodic estimation FMD; the results show that the improved periodic estimation FMD shows significant advantages in noise reduction and feature extraction, and can better extract the fault characteristic frequency and its multiples. In contrast, the fault feature extraction ability of EMD is weak in a strong noise background; the feature frequency extraction ability of EEMD is slightly better than that of EMD, but there are still many noise points; VMD has significant effects in noise reduction and feature extraction, second only to the improved periodic estimation FMD.
[0128] Finally, in order to verify the feature extraction performance of the improved FMD in actual situations, this paper applies it to the Case Western Reserve University bearing dataset; the experimental results show that the FMD with improved period estimation can effectively extract the inner race fault and the outer race fault with the load zone at 6 o'clock under the noise background.
[0129] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A bearing fault feature extraction method based on improved period estimation FMD, characterized in that: include: Step S1, obtain a data set, label and classify the data set, divide it into a training set and a test set in proportion, and preprocess the data set; Step S2, constructing an FMD algorithm model for improving cycle estimation; Step S3, training the constructed FMD algorithm model for improved cycle estimation multiple times through the training set, saving the best result, and saving the training parameters; Step S4, input the test set into the trained FMD algorithm model for improved cycle estimation to extract bearing fault features.
2. The bearing fault feature extraction method based on improved period estimation FMD according to claim 1 is characterized in that: In step S1, the data set uses the Case Western Reserve University bearing data set, and the fault types of inner ring fault, rolling element fault, outer ring fault with load area at 3 o'clock, outer ring fault with load area at 6 o'clock, and outer ring fault with load area at 12 o'clock are selected for feature extraction; the preprocessing method includes data enhancement and normalization processing.
3. The bearing fault feature extraction method based on improved period estimation FMD according to claim 1 is characterized in that: In step S2, the improved FMD algorithm model for period estimation replaces the autocorrelation function used in the original FMD algorithm to estimate the fault period with a method based on spectrum analysis and frequency selection to estimate the fault period of the signal; the specific implementation is as follows: By inputting the bearing fault signal and the sampling frequency f s Calculate the length N and frequency array F of the signal, and obtain the envelope by calculating the difference of the signal and applying the Hilbert transform, and then perform a fast Fourier transform to obtain the spectrum amplitude of the signal difference envelope; Find the first local maximum point in the spectrum amplitude array, and set the spectrum components of subsequent values that are smaller than the previous value to zero; Arrange the spectrum amplitudes in descending order and select the first ten largest frequency components as candidate frequencies F active , and remove frequencies that are very close to each other; Create a matrix Z and a multiple relationship matrix C. If F active [j] is F active An integer multiple of [i], then Z ij =1, then C ij =F active [j] / F active [i], otherwise Z ij =0, where Z ij The element represents the candidate frequency F active [i] and F acttiv Whether there is an integer multiple relationship between [j], C ij The element represents the candidate frequency F active [j] and F active [i] The multiple relationship between By calculating the modulus relationship between the candidate frequencies, check whether these candidate frequencies have an integer multiple relationship, and record the candidate frequency pairs Z that meet the conditions sum ; WITH sum =∑ j WITH ij ; On the basis of the above, calculate the sum of each row in the matrix Z, and select the frequency component corresponding to the row with the largest sum, which is the most relevant frequency Z ind ; WITH ind =argmax(Z sum ); According to the most relevant frequency components and their multiple relationships, the final fault frequency F is calculated. n and through the fault frequency F n Get the fault period T of the signal; F res =F active ; C res =; Among them, F res For C res for.
4. The bearing fault feature extraction method based on improved period estimation FMD according to claim 3 is characterized in that: In step S3, decomposition and pattern extraction are performed using the input bearing fault signal and the fault period T of the signal until the maximum number of iterations is reached or the pattern quantity requirement is met, thus completing the training of the FMD algorithm model for improving period estimation.
5. The bearing fault feature extraction method based on improved period estimation FMD according to claim 4 is characterized in that: The step S4 comprises: Input the original vibration signal x of the bearing and initialize the parameters of the improved period estimation FMD algorithm model, including but not limited to the number of decomposed modes n and the filter length L; Initialize the FIR filter bank using K Hanning windows, set to 5 to 10 groups, and then initialize the iteration a = 1; Through convolution operation Get the filtered signal, that is, the decomposed modal components Wherein, k=1, 2, ..., K; Estimating the fault period of a signal by spectrum analysis and frequency selection By inputting the original vibration signal x, the decomposed modal components and estimated failure cycles Update the filter coefficients and set a=a+1 for the next iteration; Determine whether the current iteration has reached the maximum number of iterations. If not, return to the convolution operation. Obtaining a filtered signal; If the maximum number of iterations is reached, the correlation coefficient between each two modal components is calculated to construct a K×K correlation matrix CC (K×K) ; Determine whether the current number of modes K reaches the set number of modes n. If not, return to the convolution operation. The filtered signal is obtained, otherwise, the iteration is stopped, and n modal components are obtained as the final decomposition results of the improved period estimation FMD algorithm model, that is, the bearing fault feature extraction results.
6. The bearing fault feature extraction method based on improved period estimation FMD according to claim 5 is characterized in that: In step S4, if the maximum number of iterations is reached, the correlation coefficient between every two modal components is calculated to construct a K×K correlation matrix CC (K×K) , select the two modal components with the largest correlation coefficients and calculate the failure period by the estimated Calculate their related kurtosis, and select the modal component with larger related kurtosis as the decomposed modal component. The envelope spectrum kurtosis factor is used as a screening index. It uses the Hilbert transform to obtain the envelope of the signal, and then calculates the kurtosis of the envelope. The calculation formula is as follows: Among them, amp q is the envelope of the qth modal signal obtained by Hilbert transform.
