Rolling bearing fault monitoring method and related device

Abstract

The invention discloses a rolling bearing fault monitoring method and a related device. The method comprises the following steps: obtaining a health reference sound signal, calculating the health reference sound signal by using a feature row vector calculation method to obtain a feature row vector of the health reference sound signal, and constructing a health reference model of the rolling bearing by using the feature row vector of the health reference sound signal; obtaining a measured sound signal, and calculating the measured sound signal by using a feature row vector calculation method to obtain a feature row vector of the measured sound signal; inputting the characteristic row vector of the measured sound signal into a health reference model of the rolling bearing to calculate a health index to obtain a health index trend curve; after the health index trend curve is smoothed, a health index alarm threshold value is set in combination with the # imgabs0 # criterion and the numerical gradient of the health index trend curve; and monitoring and alarming faults of the rolling bearing in combination with the health index trend curve and the health index alarm threshold.

Description

Technical Field

[0001] The invention belongs to the field of mechanical state monitoring, and in particular relates to a rolling bearing fault monitoring method and a related device. Background Art

[0002] When a rolling bearing fails due to fatigue pitting, the continuous rolling of the rolling element will produce periodic and continuous knocking at the pitting, thereby exciting the high-frequency resonance of the entire bearing seat, and then generating a periodic impact signal in its vibration signal. This signal is always inevitably submerged in irrelevant interference quantities such as impact noise, irrelevant periodic components, background noise, etc. in the actual measured signal. Therefore, under the combined action of multiple interference sources, extracting the periodic impact signal that characterizes the fault from the measured signal and constructing a health indicator that is sensitive to the periodic impact signal is the key to realizing automatic monitoring of rolling bearing faults.

[0003] Mel-scale frequency segmentation is a method used to decompose signals when calculating Mel-frequency cepstral coefficients. This method first converts the original line frequency into a nonlinear Mel-scale, then linearly divides the Mel-frequency in the Mel-scale at equal intervals, and finally converts the aforementioned linear Mel-frequency back to a nonlinear line frequency through an inverse transform, thereby achieving nonlinear frequency segmentation of the signal spectrum. Constructing a triangular filter based on Mel-scale frequency segmentation to decompose and extract signal features has become the mainstream method for rolling bearing fault feature extraction based on sound signals. However, in practical applications, this method has shortcomings such as not considering the high-frequency resonance characteristics of rolling bearing faults and the amplitude suppression at the intersection of adjacent triangular filters.

[0004] The statistical model based on parameter estimation assumes that the measured signal obeys a certain specific or mixed distribution, and uses the training data to estimate the distribution parameters, thereby constructing a baseline distribution model of the signal, and using the deviation between the distribution of the current measured signal and the baseline distribution obtained by training as the health indicator of the rolling bearing to monitor bearing faults. As a machine learning method, it not only overcomes the problem of weak fault monitoring ability in the traditional health indicator constructed based on statistical parameters and signal processing methods, but also avoids the shortcomings of difficult training and poor interpretability of deep learning methods based on neural networks. However, this method is based on the correct assumption of signal distribution. Once the distribution assumption is wrong, it will lead to erroneous calculation results. In addition, direct distribution parameter estimation of the measured signal is also susceptible to noise interference in the measured signal.

[0005] Health indicator threshold setting is a key technology for equipment fault monitoring. Appropriate threshold setting can avoid false alarms while timely discovering equipment faults, providing a reference for maintenance personnel to formulate maintenance plans. At present, the threshold setting of health indicators mainly relies on manual experience or statistical setting based on historical benchmark data. The former requires a high level of technical ability of engineering and technical personnel, while the latter only considers the statistical characteristics of historical benchmark data, and does not consider the trend of health indicators, that is, the geometric characteristics. Therefore, the threshold set based on statistics does not fully explore the performance degradation characteristics contained in the historical benchmark data. Summary of the invention

[0006] The purpose of the present invention is to overcome the problem that equipment fault monitoring uses a triangular filter based on Mel-scale frequency segmentation to decompose and extract signal features without considering the high-frequency resonance characteristics of rolling bearing faults, and proposes a rolling bearing fault monitoring method and related devices.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme: A rolling bearing fault monitoring method comprises the following steps: Obtain a healthy reference sound signal, the healthy reference sound signal includes a sound signal when the rolling bearing is in a healthy state; calculate the characteristic row vector of the healthy reference sound signal using a characteristic row vector calculation method for the healthy reference sound signal, and use the characteristic row vector of the healthy reference sound signal to construct a healthy reference model of the rolling bearing; Obtaining a measured sound signal, the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and irrelevant interference component noise; calculating the characteristic row vector of the measured sound signal using a characteristic row vector calculation method; inputting the characteristic row vector of the measured sound signal into a health benchmark model of the rolling bearing to calculate a health index to obtain a health index trend curve; The health index trend curve is smoothed to obtain the smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the criterion; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault; The calculation method of the characteristic row vector includes the following steps: The mapping relationship curve between the Mel scale frequency and the line frequency is calculated by using the Mel scale frequency curve and the symmetric Mel scale frequency curve; the mapping relationship curve between the Mel scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency, and a filter is constructed by using the line boundary frequency; The sound signal is filtered and decomposed through a filter to obtain a narrowband filtered signal, and the square envelope of the narrowband filtered signal is calculated. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal are calculated respectively. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector.

[0008] Furthermore, the symmetric Mel-scale frequency curve is is the axis of symmetry, is the horizontal axis of the Mel-scale frequency curve, is the signal sampling frequency; The calculation formula of the mapping relationship curve between Mel scale frequency and line frequency is as follows:

[0009] in, , , and are the line frequency, Mel scale frequency, symmetric Mel scale frequency and signal sampling frequency respectively; The mapping relationship curve between the Mel-scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency. Specifically, the mapping relationship curve between the Mel-scale frequency and the line frequency is equally spaced into N+2 Mel-scale boundary frequencies on the Mel-scale frequency, where N is the frequency band segmentation coefficient and is always a positive integer. The Mel-scale boundary frequency is converted into the line frequency domain by inverse transformation to obtain 2N+3 line boundary frequencies corresponding to the Mel-scale boundary frequency. The inverse transformation calculation formula is shown as follows:

[0011] Furthermore, the filter is constructed using the line boundary frequency, specifically: According to Littlewood-Paley wavelet theory, 2N+3 line boundary frequencies are used as upper and lower cutoff frequencies of the filter in an overlapping manner to construct 2N+1 filters, including 1 low-pass filter, 2N-1 band-pass filters and 1 high-pass filter; The mathematical expressions of various types of filters are as follows: Low pass filter:

[0012] Bandpass filter:

[0013] High Pass Filter:

[0014] in, is the angular frequency; and Respectively represent the lower cutoff frequency and the upper cutoff frequency; is the transition bandwidth coefficient, ; is a function operator, ; The filtering process of each type of filter is as follows: Low pass filtering:

[0015] Bandpass filtering:

[0016] High pass filtering:

[0017] in, , , and They are inner product operator, complex conjugate operator, Fourier transform operator and inverse Fourier transform operator; , and They are low-pass narrowband filtering signal, band-pass narrowband filtering signal and high-pass narrowband filtering signal respectively.

[0018] Furthermore, the square envelope calculation formula is as follows:

[0019]

[0020] in, is the square envelope of each narrowband filtered signal, , is 2N+1 narrowband filtered signals, is an imaginary unit, is the convolution operator, For the i The analytical signal of a narrowband filtered signal, is the modulo operator; The calculation formulas for the envelope Gini coefficient and envelope harmonic ratio are as follows:

[0021]

[0022] in, EGI i and EHR i Respectively i Envelope Gini coefficient and envelope harmonic ratio of a square envelope signal; Sort in ascending order i A discrete sequence of square envelope signals, Num Represents the total number of points in the square envelope sequence; The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector, as shown in the following formula:

[0024] Furthermore, the construction of the healthy benchmark model of the rolling bearing includes the following steps: Several groups of sound signals of rolling bearings in healthy states are collected, the characteristic row vectors of each group of sound signals of rolling bearings in healthy states are calculated to obtain several groups of characteristic row vectors, several groups of characteristic row vectors are stacked to obtain a characteristic matrix, and a kernel density estimation algorithm based on an iterative algorithm is used to estimate the normalized probability density function of each column feature in the characteristic matrix to obtain a probability density function set of the characteristic matrix, and the probability density function set of the characteristic matrix is ​​used as a healthy benchmark model of the rolling bearing.

[0025] Furthermore, the calculation formula of the health index is as follows:

[0026] in, HI Indicates health indicators, is the first feature row vector i Features f i In the health benchmark model i The probability value of the normalized probability density function.

[0027] Further, using The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing, which is as follows: Collect health indicators of rolling bearings in healthy conditions as Threshold Benchmark Dataset Calculation Criteria threshold, The calculation formula of the criterion threshold is as follows:

[0028] in, for Criteria threshold; and are the mean and standard deviation operators respectively, HI Indicates health indicators; The central difference method is used to calculate the numerical gradient at a certain point in time on the health index trend curve. The calculation formula of the numerical gradient is as follows:

[0029] in, and are the numerical gradient at time point k and the smoothed health index respectively; will satisfy both The health indicator whose criterion threshold and numerical gradient are continuously positive is set as the health indicator alarm threshold; Smoothing was performed using the Rauch–Tung–Striebel smoother.

[0030] A rolling bearing fault monitoring device includes a reference model building module, which is used to obtain a healthy reference sound signal, wherein the healthy reference sound includes a sound signal of a signal rolling bearing in a healthy state; a characteristic row vector of the healthy reference sound signal is calculated using a characteristic row vector calculation method for the healthy reference sound signal, and a healthy reference model of the rolling bearing is built using the characteristic row vector of the healthy reference sound signal; A health index calculation module is used to obtain a measured sound signal, wherein the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and an irrelevant interference component noise; the measured sound signal is used to calculate a characteristic row vector of the measured sound signal using a characteristic row vector calculation method; the characteristic row vector of the measured sound signal is input into a health benchmark model of the rolling bearing to calculate a health index to obtain a health index trend curve; The detection alarm module is used to smooth the health index trend curve to obtain a smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the health indicator trend curve; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault.

[0031] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the rolling bearing fault monitoring method when executing the computer program.

[0032] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the rolling bearing fault monitoring method is implemented.

[0033] Compared with the prior art, the present invention has the following beneficial technical effects: A rolling bearing fault monitoring method proposed in the present invention not only constructs a health indicator, but also provides a health indicator threshold setting method, implements accurate monitoring of the running health status of the rolling bearing, and uses the Mel scale-frequency curve and the symmetric Mel scale-frequency curve to perform frequency segmentation on the original signal, fully considers the characteristics of the high-frequency resonance response of the rolling bearing fault, and performs frequency division in the high frequency band in a manner equivalent to the frequency division of the original Mel scale-frequency curve, avoiding noise interference caused by the excessively wide bandwidth of the frequency division in the high frequency band, which affects the extraction of rolling bearing fault information; the envelope Gini coefficient and envelope harmonic ratio of each narrowband signal are calculated respectively through a multi-band feature extraction algorithm, so as to characterize the periodic impact characteristics of the rolling bearing fault signal.

[0034] Furthermore, the present invention replaces the triangular filter used in the original Mel-scale frequency division filter with Littlewood-Paley wavelet, thereby avoiding the amplitude suppression problem existing at the intersection of the amplitude-frequency response curve of the triangular filter.

[0035] Furthermore, the construction of the health benchmark model of the rolling bearing of the present invention adopts a kernel density estimation algorithm based on an iterative algorithm to estimate the probability density function of the multi-band characteristics of the benchmark data set composed of health data, and obtain the probability density function of the feature quantity of each dimension; next, the construction of the health index is realized by calculating the normalized probability of the multi-band characteristics of the data at the current moment in the probability density function model, and the health index is constructed by the statistical model learning method based on non-parametric estimation, which avoids the statistical model health index construction method based on parameter estimation, the mismatch between the assumed distribution and the actual distribution, and the noise interference problem caused by direct parameter estimation.

[0036] Furthermore, the threshold setting proposed by the present invention combines Criteria and health indicator trend analysis take into account both the statistical and geometric characteristics of health indicators, and can provide fault alarms more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0038] Figure 1 The figure is a flow chart of a method for constructing health indicators used in an embodiment of the present invention.

[0039] FIG. 2( a ) is an external view of the rolling bearing full life test bench used in the embodiment.

[0040] FIG. 2( b ) is a cross-sectional view of a bearing box of a rolling bearing full-life test bench used in the embodiment.

[0041] Figure 3 This is a failure diagram of an experimental bearing unpacking of a rolling bearing full life test bench used in the embodiments.

[0042] Figure 4(a) is a schematic diagram of Mel-scale frequency segmentation.

[0043] Figure 4(b) is a schematic diagram of symmetric Mel-scale frequency segmentation.

[0044] Figure 5 (a) shows the amplitude-frequency characteristics of the trigonometric function filter.

[0045] Figure 5(b) shows the amplitude-frequency characteristics of the Littlewood-Paley wavelet filter.

[0046] Figure 6 Visualization of statistical model for nonparametric estimation of examples.

[0047] Figure 7 It is a health indicator trend chart and a partial enlarged picture of the embodiment.

[0048] FIG8( a ) is a square envelope spectrum of the alarm time data of the embodiment.

[0049] FIG8( b ) is a square envelope spectrum of the data at one moment before the alarm moment in the embodiment.

[0050] FIG8( c ) is a square envelope spectrum of the data at one moment after the alarm moment in the embodiment.

[0051] Fig. 9 The present invention is a flow chart of a rolling bearing fault monitoring method.

[0052] Fig.10 This is a structural diagram of a rolling bearing fault monitoring device of the present invention.

[0053] Fig.11 A schematic diagram of an electronic device for a rolling bearing fault monitoring method according to the present invention. DETAILED DESCRIPTION

[0054] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0055] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present application belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention.

[0056] Embodiment 1 See also Fig. 9 , a rolling bearing fault monitoring method, comprising the following steps: Obtain a healthy reference sound signal, which includes a sound signal of a rolling bearing in a healthy state; record the sound signal of the rolling bearing in a normal healthy state as a healthy reference; calculate the characteristic row vector of the healthy reference sound signal using a characteristic row vector calculation method for the healthy reference sound signal, and use the characteristic row vector of the healthy reference sound signal to construct a healthy reference model of the rolling bearing; Obtaining a measured sound signal, the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and irrelevant interference component noise; calculating the characteristic row vector of the measured sound signal using a characteristic row vector calculation method; inputting the characteristic row vector of the measured sound signal into a health benchmark model of the rolling bearing to calculate a health index to obtain a health index trend curve; The health index trend curve is smoothed to obtain the smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the criterion; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault; This embodiment smoothes the health indicator trend curve to reduce the impact of noise, sets a reasonable health indicator alarm threshold based on preset criteria and the numerical gradient of the health indicator trend curve, observes the health indicator trend curve in real time, and issues a fault alarm when the health indicator exceeds the alarm threshold, indicating that the rolling bearing needs to be inspected and maintained.

[0057] The calculation method of the characteristic row vector includes the following steps: The mapping relationship curve between the Mel scale frequency and the line frequency is calculated by using the Mel scale frequency curve and the symmetric Mel scale frequency curve; the mapping relationship curve between the Mel scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency, and a filter is constructed by using the line boundary frequency; This embodiment uses frequency division to decompose the measured signal into a number of narrowband signals with different bandwidths and center frequencies, and uses the Mel scale-frequency curve and the symmetric Mel scale-frequency curve to perform frequency division on the original signal, fully considering the high-frequency resonance response characteristics generated when the rolling bearing fails. In the high frequency band, a frequency division method equivalent to the frequency division method of the original Mel scale-frequency curve is used to ensure the fineness of the frequency division, effectively avoiding noise interference introduced by the excessively wide frequency division bandwidth, so that the fault information of the rolling bearing can be extracted more accurately.

[0058] The sound signal is filtered and decomposed through a filter to obtain a narrowband filtered signal, and the square envelope of the narrowband filtered signal is calculated. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal are calculated respectively. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector.

[0059] The health index and the threshold setting constructed in this embodiment can effectively extract fault information from the measured signal, thereby realizing rolling bearing fault monitoring.

[0060] Embodiment 2 This embodiment selects a set of rolling bearing full life sound signals from healthy to failed in the rolling bearing full life test bench. FIG2 shows the measurement point layout of the rolling bearing full life test bench, where FIG2 (a) is the overall appearance view of the test bench, and FIG2 (b) is the cross-sectional view of the bearing box. Figure 3 As shown, after the experiment, the inner ring raceway of the rolling bearing was found to be damaged after unpacking and the characteristic frequency of the inner ring fault was 341.7 Hz according to the bearing model and speed. The data of the embodiment is collected by a microphone, and the sampling frequency, number of sampling points and sampling interval are 25.6 kHz, 51200 and 30 seconds respectively. Taking this group of full-life sound data as the analysis object, the following is an introduction to this embodiment: A rolling bearing fault monitoring method adopts a health indicator construction method for rolling bearing fault monitoring, such as Figure 1 As shown, the following steps are included: Step 1, loading and reading the rolling bearing sound signal; Step 2: Compare the Mel scale-frequency curve with the symmetric Mel scale-frequency curve (with is the axis of symmetry, is the signal sampling frequency, represent Axis) establishes the mapping relationship between Mel scale frequency and line frequency, and the calculation formula is:

[0061] in , and They are line frequency, Mel-scale frequency, and symmetric Mel-scale frequency, respectively; All data in this example ; The symmetric Mel scale-frequency curve represents the compensation for the problem of too wide frequency division in the high-frequency band of the original Mel scale. The symmetry is used to map the division characteristics of the low-frequency band of the Mel scale to the high-frequency band, ensuring that there are dense frequency divisions in both the high and low frequency bands. The reason for this is that the fault characteristics of the rolling bearing can be understood as information characterizing the fault, which is usually manifested as high-frequency resonance characteristics, especially in the early stage of the fault. Therefore, dense frequency division at both the high and low frequency bands can avoid the loss of fault information caused by the overly wide division band.

[0062] Step 3: For the Mel-scale frequency and line frequency mapping relationship curve obtained in step 2, divide it into N+2 boundary frequencies at equal intervals on the Mel-scale frequency, and convert the Mel-scale boundary frequencies obtained by segmentation into the line frequency domain through inverse transformation, so as to obtain 2N+3 line boundary frequencies corresponding to the Mel-scale boundary frequencies. The inverse transformation calculation formula is:

[0063] In this embodiment, all data N=11. Figure 4 (a) and Figure 4 (b) show the comparison of Mel scale frequency segmentation and symmetric Mel scale frequency segmentation. It can be seen that the symmetric Mel scale frequency segmentation method not only increases the density of high frequency band frequency segmentation, but also increases the density of medium frequency band frequency segmentation to a certain extent, avoiding the fault information being submerged in irrelevant interference noise due to the excessively wide frequency band of the segmentation. Step 4: Based on the Littlewood-Paley wavelet theory, the 2N+3 line frequencies obtained in step 3 are overlapped as the upper and lower cutoff frequencies of the filter to construct a low-pass filter, a 2N-1 band-pass filter and a high-pass filter. The mathematical expressions of each type of filter are as follows: ① Low-pass filter:

[0064] ②Bandpass filter:

[0065] ③ High pass filter: In the above formula is the angular frequency; and Respectively represent the lower cutoff frequency and the upper cutoff frequency; is the transition bandwidth coefficient, which should satisfy ; is a function operator, and its expression is as follows:

[0066] Figures 5(a) and 5(b) are comparison diagrams of the amplitude-frequency characteristics of the triangular filter bank and the Littlewood-Paley wavelet filter bank. It can be observed from the figures that the amplitude-frequency characteristics of each filter constructed by the Littlewood-Paley wavelet have no amplitude suppression phenomenon at the junction; Step 5: The sound signal measured in step 1 Through the 2N+1 filters constructed in step 4, a total of 2N+1 narrowband filtered signals are obtained. The mathematical expression of the filtering process is as follows: ① Low-pass filtering:

[0067] ②Bandpass filtering:

[0068] ③ High-pass filtering:

[0069] In the above formula , , and They are inner product operator, complex conjugate operator, Fourier transform operator and inverse Fourier transform operator; , and They are low-pass narrow-band filter signal, band-pass narrow-band filter signal and high-pass narrow-band filter signal respectively; Step 6: record the 2N+1 narrowband filtered signals obtained in step 5 as ; According to the definition of square envelope theory, the square envelope of each narrowband filter signal is calculated in turn, which is recorded as , the square envelope calculation formula is as follows:

[0070]

[0071] In the above formula is an imaginary unit, is the convolution operator, For the i The analytical signal of a narrowband filtered signal, is the modulo operator; Step 7: Calculate the envelope Gini coefficient representing the signal impact and the envelope harmonic ratio representing the signal periodicity of the 2N+1 square envelopes obtained in step 6. The calculation formulas of the envelope Gini coefficient and the envelope harmonic ratio are as follows:

[0072]

[0073] In the above formula EGI i and EHR i Respectively i Envelope Gini coefficient and envelope harmonic ratio of a square envelope signal; Sort in ascending order i A discrete sequence of square envelope signals, Num Represents the total number of points in the square envelope sequence; the envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector, as shown below:

[0074] Step 8: Collect L groups of sound signals of rolling bearings in a healthy state, calculate the characteristic row vectors corresponding to each group of data according to steps 2 to 7, and use the characteristic matrix formed by stacking these L groups of characteristic row vectors as the training data set for use in subsequent steps. The mathematical expression of the characteristic matrix is ​​as follows:

[0075] In this embodiment, L=100; Step 9: Treat each column of the feature matrix obtained in step 8 as a set of random variables X , using the kernel density estimation algorithm based on iterative algorithm to estimate X The probability density function of the feature matrix is ​​finally used as the health benchmark model of the rolling bearing for subsequent model reasoning, i.e., health index calculation. The specific steps are as follows: ① Establish a kernel density estimation model; random variable X The kernel density estimation model is as follows:

[0076] In the above formula is a random variable X The estimated probability density function of ; and represent bandwidth and kernel function respectively.

[0077] ② Calculate the asymptotic mean square error ( AMISE );

[0078] ③Minimize AMISE Finding the optimal bandwidth ;

[0079] ④ Substitute the kernel function to obtain the optimal bandwidth expression; Take Gaussian kernel as kernel function and substitute it into ③ to solve The expression is

[0080] ⑤ Calculate the optimal bandwidth through iterative algorithm ; The expression contains h , which can be calculated using an iterative algorithm , the algorithm pseudo code is shown in Table 1 below: Table 1 Pseudo code of optimal bandwidth iteration algorithm

[0081] ⑥Estimation of random variables X Probability density function ; The Gaussian kernel and Bring in The expression gets the random variable X Probability density function, and normalize the probability density function according to the following formula,

[0082] ⑦ Computational Health Benchmark Model .

[0083] Repeat operations ①-⑥ for each column of features in the feature matrix, and finally obtain a healthy benchmark model of the training data.

[0084] In this embodiment, Pick ,in , , and are the length, standard deviation, and interquartile range of the random variable, respectively, so is not a fixed value, it varies with the random variable X varies depending on the type of product. Figure 6 The visualization model obtained by training the non-parametric estimation statistical model in the embodiment is shown; Step 10, bring the characteristic row vector of the rolling bearing sound signal at the current moment obtained through steps 2 to 7 into the health benchmark model trained in step 9, and use the probability value of the normalized probability density function of the characteristic value on each dimension on the corresponding dimension of the health benchmark model as the calculation basis of the health index to construct the health index. The calculation formula of the health index is as follows:

[0085] In the above formula HI Indicates health indicators, is the first feature row vector i Features f i In the health benchmark model i The probability value of the normalized probability density function; Step 11: Use the Rauch–Tung–Striebel smoother to smooth the health indicator trend graph. The Rauch–Tung–Striebel smoother includes two key steps: forward recursion and backward recursion. Its mathematical expression is as follows: ①Forward recursion:

[0086] ②Backward recursion:

[0087] In the above formula , They are k and k-1 The posterior state estimate at time t; , They are k and k+1 Smoothed results of moments; , Forward recursion k Prior state estimation and backward recursion at time k+1 A priori state estimate at time t; , The forward recursion is k and k-1 The posterior estimated covariance of the moment; , Respectively, the backward recursion is k and k+1 The posterior estimated covariance of the moment; , They are k and k+1 A priori estimate of the moment covariance; and They are k The forward Kalman gain and backward Kalman gain at time ; A, H, Q, R and I They are the state transfer matrix, observation matrix, process noise covariance, measurement noise covariance and identity matrix respectively; In this embodiment, A, H, Q and R They are taken as 1, 1, 0.1 and 1 respectively; Step 12: Combine To set thresholds for criteria and health indicator trend analysis, follow these steps: ① Calculation Criteria threshold; Collect health indicators of group L in health status as The threshold benchmark data set is calculated according to the following formula Guideline Threshold:

[0088] Among the above for Criteria threshold; and are the mean and standard deviation operators respectively.

[0089] ②Calculation of numerical gradient of health index trend curve; The central difference method is used to calculate the numerical gradient at a certain point in time on the health index trend curve. The calculation formula is as follows:

[0090] In the above formula and are the numerical gradient at time point k and the smoothed health index respectively.

[0091] ③Alarm threshold setting; The alarm thresholds must be met at the same time The criterion threshold and the numerical gradient are continuously positive, and the pseudo code of the calculation algorithm is shown in Table 2 below: Table 2 Pseudo code of threshold setting algorithm

[0092] In this embodiment, and Take them as 7 and 0.002 respectively.

[0093] Figure 7 The trend graph and local magnification graph of the health index and its smooth curve are shown. Through the threshold setting algorithm, the data at the time corresponding to file number 404 triggers an alarm.

[0094] Figures 8(a), 8(b) and 8(c) show the square envelope spectra of the data at the alarm moment and one moment before and after. It can be seen from Figures 8(a), 8(b) and 8(c) that the inner circle fault characteristic frequency does not appear in the square envelope spectrum of the data at the moment corresponding to file number 403, while the inner circle fault characteristic frequency appears in the square envelope spectrum of the data at the moment corresponding to file numbers 404 and 405, and the amplitude gradually increases, showing a trend of gradual evolution of the fault.

[0095] The above results show that the rolling bearing health indicator construction method and threshold setting method proposed in this embodiment can effectively monitor rolling bearing failures and issue alarms in a timely manner, providing a basis for operation and maintenance personnel to formulate maintenance plans.

[0096] Embodiment 3 See also Fig. 9 , a rolling bearing fault monitoring method, comprising the following steps: Obtain a healthy reference sound signal, the healthy reference sound includes a sound signal when the rolling bearing is in a healthy state; calculate the characteristic row vector of the healthy reference sound signal using a characteristic row vector calculation method, and use the characteristic row vector of the healthy reference sound signal to construct a healthy reference model of the rolling bearing; Obtaining a measured sound signal, the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and irrelevant interference component noise; calculating the characteristic row vector of the measured sound signal using a characteristic row vector calculation method; inputting the characteristic row vector of the measured sound signal into a health benchmark model of the rolling bearing to calculate a health index to obtain a health index trend curve; The health index trend curve is smoothed to obtain the smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the criterion; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault; The calculation method of the characteristic row vector includes the following steps: The mapping relationship curve between the Mel scale frequency and the line frequency is calculated by using the Mel scale frequency curve and the symmetric Mel scale frequency curve; the mapping relationship curve between the Mel scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency, and a filter is constructed by using the line boundary frequency; The sound signal is filtered and decomposed through a filter to obtain a narrowband filtered signal, and the square envelope of the narrowband filtered signal is calculated. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal are calculated respectively. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector.

[0097] The Mel scale-frequency curve is a curve that describes the relationship between the sound frequency (Hz) and the Mel scale (Mel). This curve is not linear, but a logarithmic relationship, that is, the Hertz change in the low-frequency part corresponds to a larger Mel change, and the Hertz change in the high-frequency part corresponds to a smaller Mel change. This relationship is more in line with the human ear's perception of pitch, that is, the human ear has a higher resolution for low-frequency sounds than for high-frequency sounds. When the frequency is low, the Mel value increases rapidly with the increase of frequency; when the frequency is high, the increase rate of the Mel value gradually slows down. This shape ensures that the pitch changes perceived by the human ear are uniform throughout the audible frequency range.

[0098] Symmetric Mel-scale frequency curves are is the symmetry axis, x is the horizontal axis of the Mel-scale frequency curve, is the signal sampling frequency; the calculation formula of the mapping relationship curve between Mel scale frequency and line frequency is as follows:

[0099] in, , , and are the line frequency, Mel scale frequency, symmetric Mel scale frequency and signal sampling frequency respectively; The mapping relationship curve between the Mel-scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency. Specifically, the mapping relationship curve between the Mel-scale frequency and the line frequency is equally spaced into N+2 Mel-scale boundary frequencies on the Mel-scale frequency, where N is the frequency band segmentation coefficient and is always a positive integer. The Mel-scale boundary frequency is converted into the line frequency domain by inverse transformation to obtain 2N+3 line boundary frequencies corresponding to the Mel-scale boundary frequency. The inverse transformation calculation formula is shown as follows:

[0101] The filter is constructed using the line boundary frequencies, specifically: According to Littlewood-Paley wavelet theory, 2N+3 line boundary frequencies are used as upper and lower cutoff frequencies of the filter in an overlapping manner to construct 2N+1 filters, including 1 low-pass filter, 2N-1 band-pass filters and 1 high-pass filter; Littlewood-Paley wavelet is a wavelet with good characteristics in the frequency domain, but its decay rate in the time domain is slow, which limits its application to a certain extent. The Littlewood-Paley wavelet theory originated in 1936, when Littlewood and Paley established a binary frequency component grouping theory for Fourier series, namely the LP theory. This theory groups the Fourier series according to the binary frequency components and points out that the phase of its Fourier transform does not affect the size and shape of the function.

[0102] The mathematical expressions of various types of filters are as follows: Low pass filter:

[0103] Bandpass filter:

[0104] High Pass Filter:

[0105] in, is the angular frequency; and Respectively represent the lower cutoff frequency and the upper cutoff frequency; is the transition bandwidth coefficient, ; is a function operator, ; The filtering process of each type of filter is as follows: Low pass filtering:

[0106] Bandpass filtering:

[0107] High pass filtering:

[0108] in, , , and They are inner product operator, complex conjugate operator, Fourier transform operator and inverse Fourier transform operator; , and They are low-pass narrowband filtering signal, band-pass narrowband filtering signal and high-pass narrowband filtering signal respectively.

[0109] The square envelope calculation formula is as follows:

[0110]

[0111] in, is the square envelope of each narrowband filtered signal, , is 2N+1 narrowband filtered signals, is an imaginary unit, is the convolution operator, For the i The analytical signal of a narrowband filtered signal, is the modulo operator; The calculation formulas for the envelope Gini coefficient and envelope harmonic ratio are as follows:

[0112]

[0113] in, EGI i and EHR i Respectively i Envelope Gini coefficient and envelope harmonic ratio of a square envelope signal; Sort in ascending order i A discrete sequence of square envelope signals, Num Represents the total number of points in the square envelope sequence; The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal constitute a characteristic row vector, as shown in the following formula:

[0115] The construction of the rolling bearing health benchmark model includes the following steps: Several groups of sound signals of rolling bearings in healthy states are collected, the characteristic row vectors of each group of sound signals of rolling bearings in healthy states are calculated to obtain several groups of characteristic row vectors, several groups of characteristic row vectors are stacked to obtain a characteristic matrix, and a kernel density estimation algorithm based on an iterative algorithm is used to estimate the normalized probability density function of each column feature in the characteristic matrix to obtain a probability density function set of the characteristic matrix, and the probability density function set of the characteristic matrix is ​​used as a healthy benchmark model of the rolling bearing.

[0116] The kernel density estimation algorithm based on iterative algorithm is a non-parametric method for estimating unknown density functions. It combines the ideas of iterative algorithm and kernel density estimation. Kernel density estimation (KDE) was proposed by Rosenblatt (1955) and Emanuel Parzen (1962) and is a non-parametric method for estimating the probability density function of random variables. It does not require assumptions about the distribution form of the data, but directly starts from the data sample and estimates the density distribution of the data through the kernel function. An iterative algorithm is an algorithm that approaches the desired target or result by repeating the feedback process. In each iteration, the algorithm calculates a new value based on the previous result and uses the value as the input for the next iteration. The iterative process continues until a certain stopping condition is met (such as reaching the maximum number of iterations, the result converges, etc.).

[0117] In the kernel density estimation algorithm based on iterative algorithm, iteration is mainly used to optimize the bandwidth h or kernel function K (⋅) is selected to improve the accuracy and efficiency of density estimation. The convergence and stability of the iterative algorithm are the key to algorithm design. It is necessary to ensure that the iterative process can stably converge to the optimal solution or an acceptable solution. The choice of bandwidth h has an important impact on the results of kernel density estimation. Too large a bandwidth will cause the estimation result to be too smooth and lose the detailed information of the data; too small a bandwidth will cause the estimation result to be too rough and have too much fluctuation. The choice of kernel function K(⋅) will also affect the estimation result. Different kernel functions have different shapes and properties, and are suitable for different data distributions and estimation requirements.

[0118] The calculation formula of health index is as follows:

[0119] in, HI Indicates health indicators, is the first feature row vector i Features f i In the health benchmark model i The probability value of the normalized probability density function.

[0120] use The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing, which is as follows: Collect health indicators of rolling bearings in healthy conditions as Threshold Benchmark Dataset Calculation Criteria threshold, The calculation formula of the criterion threshold is as follows:

[0121] in, for Criteria threshold; and are the mean and standard deviation operators respectively, HI Indicates health indicators; The criterion is also called the Laida criterion. It is a commonly used method in statistics and signal processing to identify and remove outliers or noise in data. The "3" here refers to the standard deviation ( ), which is a quantitative indicator of the fluctuation range or dispersion of the data. In a normal distribution (Gaussian distribution), about 99.73% of the data will fall within the mean (μ) plus or minus 3 standard deviations ( ). This means that if a data point has a value outside this range, i.e., less than μ- or greater than μ+ , then it is considered an outlier or noise and may be removed or given special treatment.

[0122] The central difference method is used to calculate the numerical gradient at a certain point in time on the health index trend curve. The calculation formula of the numerical gradient is as follows:

[0123] in, and are the numerical gradient at time point k and the smoothed health index respectively; will satisfy both The health indicator whose criterion threshold and numerical gradient are continuously positive is set as the health indicator alarm threshold; Smoothing was performed using the Rauch–Tung–Striebel smoother.

[0124] The Rauch–Tung–Striebel (RTS) smoother is a state smoothing algorithm for nonlinear systems, also known as RTS smoothing or RTS smoother. The algorithm is extended on the basis of Kalman filtering to use the entire observation sequence to estimate the system state, thereby improving the accuracy of state estimation. The Rauch–Tung–Striebel smoother is based on Bayesian state estimation theory, in particular, after all observation data are obtained, the system state is estimated a posteriori. It utilizes the forward prediction and update steps of the Kalman filter and adds a backward smoothing step to correct the past state estimate in combination with future observation information.

[0125] RTS smoother usually includes the following steps: Forward filtering: Use the Kalman filter algorithm for forward filtering to obtain the state estimate and covariance matrix at each time point.

[0126] Backward smoothing: Starting from the last time point, the smoothed state estimate and covariance matrix of each time point are calculated backward recursively using the results of forward filtering and future observation information.

[0127] See also Fig.10 A rolling bearing fault monitoring device includes a reference model building module for obtaining a healthy reference sound signal, wherein the healthy reference sound includes a sound signal of a rolling bearing in a healthy state; a characteristic row vector of the healthy reference sound signal is calculated using a characteristic row vector calculation method for the healthy reference sound signal, and a healthy reference model of the rolling bearing is built using the characteristic row vector of the healthy reference sound signal; A health index calculation module is used to obtain a measured sound signal, wherein the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and an irrelevant interference component noise; the measured sound signal is used to calculate a characteristic row vector of the measured sound signal using a characteristic row vector calculation method; the characteristic row vector of the measured sound signal is input into a health benchmark model of the rolling bearing to calculate a health index to obtain a health index trend curve; The detection alarm module is used to smooth the health index trend curve to obtain a smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the health indicator trend curve; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault.

[0128] See also Fig.11 , an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the rolling bearing fault monitoring method when executing the computer program.

[0129] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the rolling bearing fault monitoring method is implemented.

[0130] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A rolling bearing fault monitoring method, characterized in that: The following steps are involved: Obtaining a healthy reference sound signal, wherein the healthy reference sound signal includes a sound signal of the rolling bearing in a healthy state; The characteristic row vector of the healthy reference sound signal is calculated by using the characteristic row vector calculation method, and the healthy reference sound signal is used to construct a healthy reference model of the rolling bearing; Obtaining a measured sound signal, the measured sound signal includes a rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes a fatigue pitting fault periodic impact signal and irrelevant interference component noise; Calculating the characteristic row vector of the measured sound signal using a characteristic row vector calculation method; Inputting the characteristic row vector of the measured sound signal into the health benchmark model of the rolling bearing to calculate the health index and obtain the health index trend curve; The health index trend curve is smoothed to obtain the smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing and the criterion; the health indicator trend curve and the health indicator alarm threshold are used to monitor and alarm the rolling bearing fault; The calculation method of the characteristic row vector includes the following steps: The mapping relationship curve between the Mel scale frequency and the line frequency is calculated by using the Mel scale frequency curve and the symmetric Mel scale frequency curve; the mapping relationship curve between the Mel scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency, and a filter is constructed by using the line boundary frequency; The sound signal is filtered and decomposed through a filter to obtain a narrowband filtered signal, and the square envelope of the narrowband filtered signal is calculated. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal are calculated respectively. The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector.

2. A rolling bearing fault monitoring method according to claim 1, characterized in that: Symmetric Mel-scale frequency curves are is the axis of symmetry, is the horizontal axis of the symmetric Mel-scale frequency curve, is the signal sampling frequency; The calculation formula of the mapping relationship curve between Mel scale frequency and line frequency is as follows: in, , , and are the line frequency, Mel scale frequency, symmetric Mel scale frequency and signal sampling frequency respectively; The mapping relationship curve between the Mel-scale frequency and the line frequency is segmented and then inversely transformed to obtain the line boundary frequency. Specifically, the mapping relationship curve between the Mel-scale frequency and the line frequency is equally spaced into N+2 Mel-scale boundary frequencies on the Mel-scale frequency, where N is the frequency band segmentation coefficient and is always a positive integer. The Mel-scale boundary frequency is converted into the line frequency domain by inverse transformation to obtain 2N+3 line boundary frequencies corresponding to the Mel-scale boundary frequency. The inverse transformation calculation formula is shown as follows: 。 3. A rolling bearing fault monitoring method according to claim 2, characterized in that: The filter is constructed using the line boundary frequencies, specifically: According to Littlewood-Paley wavelet theory, 2N+3 line boundary frequencies are used as upper and lower cutoff frequencies of the filter in an overlapping manner to construct 2N+1 filters, including 1 low-pass filter, 2N-1 band-pass filters and 1 high-pass filter; The mathematical expressions of various types of filters are as follows: Low pass filter: Bandpass filter: High Pass Filter: in, is the angular frequency; and Respectively represent the lower cutoff frequency and the upper cutoff frequency; is the transition bandwidth coefficient, ; is a function operator, ; The filtering process of each type of filter is as follows: Low pass filtering: Bandpass filtering: High pass filtering: in, , , and They are inner product operator, complex conjugate operator, Fourier transform operator and inverse Fourier transform operator; , and They are low-pass narrowband filtering signal, band-pass narrowband filtering signal and high-pass narrowband filtering signal respectively.

4. A rolling bearing fault monitoring method according to claim 3, characterized in that: The square envelope calculation formula is as follows: in, is the square envelope of each narrowband filtered signal, , is 2N+1 narrowband filtered signals, is an imaginary unit, is the convolution operator, For the i The analytical signal of a narrowband filtered signal, is the modulo operator; The calculation formulas for the envelope Gini coefficient and envelope harmonic ratio are as follows: in, EGI i and EHR i Respectively i Envelope Gini coefficient and envelope harmonic ratio of a square envelope signal; Sort in ascending order i A discrete sequence of square envelope signals, Num Represents the total number of points of the square envelope sequence; The envelope Gini coefficient and envelope harmonic ratio of each square envelope signal form a characteristic row vector, as shown in the following formula: 。 5. A rolling bearing fault monitoring method according to claim 4, characterized in that: The construction of the rolling bearing health benchmark model includes the following steps: Several groups of sound signals of rolling bearings in healthy states are collected, the characteristic row vectors of each group of sound signals of rolling bearings in healthy states are calculated to obtain several groups of characteristic row vectors, several groups of characteristic row vectors are stacked to obtain a characteristic matrix, and a kernel density estimation algorithm based on an iterative algorithm is used to estimate the normalized probability density function of each column feature in the characteristic matrix to obtain a probability density function set of the characteristic matrix, and the probability density function set of the characteristic matrix is ​​used as a healthy benchmark model of the rolling bearing.

6. A rolling bearing fault monitoring method according to claim 5, characterized in that: The calculation formula of health index is as follows: in, HI Indicates health indicators, is the first feature row vector i Features f i In the health benchmark model i The probability value of the normalized probability density function.

7. A rolling bearing fault monitoring method according to claim 6, characterized in that: use The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after smoothing, which is as follows: Collect health indicators of rolling bearings in healthy conditions as Threshold Benchmark Dataset Calculation Criteria threshold, The calculation formula of the criterion threshold is as follows: in, for Criteria threshold; and are the mean and standard deviation operators respectively, HI Indicates health indicators; The central difference method is used to calculate the numerical gradient at a certain point in time on the health index trend curve. The calculation formula of the numerical gradient is as follows: in, and are the numerical gradient at time point k and the smoothed health index respectively; will satisfy both The health indicator whose criterion threshold and numerical gradient are continuously positive is set as the health indicator alarm threshold; Smoothing was performed using the Rauch–Tung–Striebel smoother.

8. A rolling bearing fault monitoring device, characterized in that: It includes a reference model construction module, which is used to obtain a healthy reference sound signal, wherein the healthy reference sound includes a sound signal when the rolling bearing is in a healthy state; the characteristic row vector of the healthy reference sound signal is calculated using a characteristic row vector calculation method, and the characteristic row vector of the healthy reference sound signal is used to construct a healthy reference model of the rolling bearing; The health index calculation module is used to obtain the measured sound signal, the measured sound signal includes the rolling bearing sound signal at the current moment; the rolling bearing sound signal at the current moment includes the fatigue pitting fault periodic impact signal and irrelevant interference component noise; Calculating the characteristic row vector of the measured sound signal using a characteristic row vector calculation method; Inputting the characteristic row vector of the measured sound signal into the health benchmark model of the rolling bearing to calculate the health index and obtain the health index trend curve; The detection alarm module is used to smooth the health index trend curve to obtain a smoothed health index trend curve. The health indicator alarm threshold is set based on the numerical gradient of the health indicator trend curve after the criterion and smoothing; The health indicator trend curve and health indicator alarm threshold are used to monitor and alarm the rolling bearing faults.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a rolling bearing fault monitoring method as claimed in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements a rolling bearing fault monitoring method as claimed in any one of claims 1 to 7.

Images