Blind deconvolution filter-based rolling bearing fault diagnosis method based on harmonic noise kurtosis
By constructing a blind deconvolution filtering framework based on harmonic noise kurtosis and periodic characteristic energy ratio, the problem of extracting early weak fault features of rolling bearings under strong noise environment is solved, accurate fault diagnosis is achieved, and equipment damage and economic losses are avoided.
Patent Information
- Application Number
- CN202410612992.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-05-17
AI Technical Summary
Existing technologies struggle to accurately extract early, subtle fault characteristics of rolling bearings in noisy environments, especially since fault signals in wheelset bearings are modulated by interference signals and drowned out by noise, making diagnosis difficult.
A blind deconvolution filtering method based on harmonic noise kurtosis and periodic characteristic energy ratio is adopted. By iteratively updating the filter coefficients, a blind deconvolution filtering framework is constructed. Fault features are extracted using relevant harmonic noise kurtosis and sparsity indices, and envelope spectrum analysis is performed.
Accurate extraction of early, subtle fault characteristics of rolling bearings in high-noise environments improves the accuracy and stability of fault diagnosis, preventing equipment damage and economic losses.
Smart Images

Figure CN118536017B_ABST
Abstract
Description
Technical Field
[0001] In the field of rolling bearing fault diagnosis, this paper specifically describes a blind deconvolution filtering method for extracting early weak fault features of rolling bearings based on the kurtosis of relevant harmonic noise. Background Technology
[0002] Rolling bearings, as crucial components of industrial equipment, are widely used in machinery, aerospace, and automotive industries. Because they play a vital role in equipment transmission and support, rolling bearings frequently operate in complex and harsh environments characterized by high speeds and heavy loads. Consequently, components such as the inner and outer rings, rolling elements, and cages of rolling bearings often experience pitting, wear, and fracture, jeopardizing the normal operation of equipment and causing significant economic losses. Therefore, fault diagnosis of rolling bearings and the ability to promptly identify existing faults are currently of paramount importance.
[0003] In practical industrial equipment, the periodic impact signals generated by early failures in rolling bearings are weak and often interfered with by strong noise during signal transmission, making it difficult to detect and extract fault features in a timely manner. Wiggins proposed Minimum Entropy Deconvolution (MED), which extracts feature information by constructing a kurtosis objective function and maximizing it through deconvolution. However, its robustness is poor and it is easily affected by other interference signals. McDonald proposed Maximum Correlation Kurtosis Deconvolution (MCKD), which extracts the periodic impacts generated by faults by constructing a correlation kurtosis index. It has been widely used in the field of rolling bearing fault diagnosis, and many related methods have been proposed. However, its extraction effect on early and weak bearing fault features is poor, and accurate fault period parameters are required to achieve good results. Therefore, inspired by this method, a blind deconvolution filtering fault diagnosis method based on correlated harmonic noise kurtosis is proposed. The correlated harmonic noise kurtosis is used as the objective function, the weight of fault features is evaluated by the periodic feature energy ratio, and the optimal deconvolution filter coefficients are continuously updated and searched through an iterative algorithm. Finally, envelope spectrum analysis is performed to extract bearing fault features and achieve accurate diagnosis of early and weak bearing faults. Summary of the Invention
[0004] The components of wheelset bearing fault signals are highly complex. Periodic impacts, representing fault characteristics, are often modulated by interference signals and submerged in strong ambient noise, posing a challenge to wheelset bearing fault diagnosis. This invention constructs an iterative framework based on blind deconvolution filtering, using the proposed harmonic noise kurtosis and periodic feature energy ratio as its core. Through in-depth analysis of the sparsity indices representing periodic impact characteristics, the accuracy and stability of traditional blind deconvolution filtering methods in fault feature identification are improved. Specifically, this invention employs the following technical solution:
[0005] A blind deconvolution filter-based method for fault diagnosis of rolling bearings based on harmonic noise kurtosis, comprising the following steps:
[0006] Step 1: Measure the signal x(n) containing fault information using an accelerometer;
[0007] Step 2: Perform an iterative blind deconvolution filtering process based on the kurtosis of relevant harmonic noise.
[0008] Step 2.1: Determine the initial values of the required parameters, sampling frequency f. s ; Filter length L; Number of iterations m; Initial fault cycle search range T f = [T0, T1]; Initial filter coefficients f0;
[0009] Step 2.2: Enter the iterative process, calculate the output signal s0(n) after filtering the initial filter coefficients f0; and calculate the given range T. f = The maximum value of TCER for all periods within [T0, T1], where the period is denoted as T. max The periodic characteristic energy ratio at this time is denoted as TCER. max1 ;
[0010] Step 2.3: Based on s0(n) and T obtained in Step 2.2 max The filter coefficients are updated, and the filter coefficients f1 based on the maximum value of the objective function are calculated according to the proposed relevant harmonic noise kurtosis.
[0011] Step 2.4: Update the filtered output signal s1(n) according to the current filter coefficient f1, and use the current f1 and s1(n) as the required parameters for the next iteration; determine whether the iteration termination condition has been met, which is reaching the maximum number of iterations m or the current TCER. max1 The maximum value during the iteration process is reached; if the iteration termination condition is met, the iteration process ends; otherwise, return to step 2 for further iteration.
[0012] Step 3: End the iteration process and obtain the filtered output signal s at this point. x (n), for s x (n) Perform envelope spectrum analysis to extract fault features and perform fault diagnosis.
[0013] Preferably, step (2) specifically includes:
[0014] The model for the blind deconvolution filtering process is constructed and defined as follows:
[0015] s(n)=x(n)*f(l) (1)
[0016] Where s(n), n = 1, 2, v3, ..., N represents the filtered output signal with a length of N; x(n), n = 1, 2, 3, ..., N represents the input signal measured by the accelerometer with a length of N; f(l), l = 1, 2, 3, ..., L represents the filter coefficient sequence with a length of L; * represents convolution operation;
[0017] Construct the objective function for blind deconvolution filtering, and determine the objective function based on the kurtosis of relevant harmonic noise:
[0018]
[0019] Where T represents the period of the input; F(f) represents the proposed objective function based on the kurtosis of related harmonic noise; f = f(l), l = 1, 2, 3, ..., L represents the filter coefficient sequence;
[0020] Solve for the derivative of the objective function with respect to the filter coefficients, and set the derivative to zero:
[0021]
[0022] Solving equation (3) yields:
[0023] p=f*Q (4)
[0024]
[0025]
[0026]
[0027] Therefore, the filter coefficients f based on the maximum value of the objective function are solved using equation (8):
[0028] f = p * Q -1 (8).
[0029] Preferably, step (2) further includes:
[0030] Construct the periodic characteristic energy ratio of the prior estimated period T, and use it to estimate the input period T of the objective function (2);
[0031] First, the output signal is filtered after the input period T0. Perform Hilbert transform to obtain the analytic signal
[0032]
[0033] Where j represents the imaginary sign, Represents an analytical signal;
[0034] Then analyze the signal Transform to the frequency domain and perform squared envelope to obtain the frequency domain signal shown in equation (10).
[0035]
[0036] Where abs represents the modulus of the complex signal, and FFT represents the Fast Fourier Transform. The frequency domain signal obtained represents the periodic characteristic energy ratio of the input period T0, which is calculated according to equation (11).
[0037]
[0038]
[0039]
[0040] in, The periodic characteristic energy ratio represents the input period T0; K represents the frequency domain signal. The corresponding input period T0 and its corresponding positive integer multiples of the period are in Quantity within the range; It represents rounding up; The Nyquist frequency is the signal sampling frequency f. s Half of; r(k) represents the weighting coefficient function, which represents the frequency domain signal corresponding to the input period T0 and its positive integer multiples of the period. The weight of the amplitude; T x The period range considering the error is represented by equation (14).
[0041]
[0042] The present invention has the following beneficial effects:
[0043] 1. This method overcomes the challenge of diagnosing early, weak faults in rolling bearings under high-noise environments. By proposing an objective function for the kurtosis of relevant harmonic noise, it accurately describes the periodicity and sparsity characteristics of the periodic impact signals caused by bearing faults. Therefore, it can accurately extract fault features through blind deconvolution filtering.
[0044] 2. This method utilizes the proposed periodic feature energy ratio to accurately describe the weights of the fault features of the filtered output signal under different input fault periods, which is beneficial for selecting the input period parameters of the objective function, thereby achieving better blind deconvolution filtering results.
[0045] 3. This method can be applied to the fault diagnosis of rolling bearing-related equipment, and can accurately analyze the bearing fault characteristic information caused by early and minor faults, so as to avoid damage to related equipment and economic losses. Attached Figure Description
[0046] Figure 1 This is a flowchart of the method of the present invention.
[0047] Figure 2 This is a time-domain waveform diagram of the signal in an embodiment.
[0048] Figure 3 These are the time-domain waveform and frequency-domain Hilbert envelope spectrum after blind deconvolution filtering.
[0049] Figure 4 A diagram illustrating the blind deconvolution process used to obtain the actual signal. Detailed Implementation
[0050] Blind deconvolution can be viewed as the inverse process of convolution. The actual signal measured by the sensor is usually transmitted from the fault signal through paths such as the housing and shaft. Therefore, the actual measured signal can be described as the result of convolving the fault signal with the transfer function of the transmission path from the fault source to the sensor. The fault signal can be obtained by deconvolving the measured signal. Since the transfer function and frequency response of the system containing the wheelset bearings of urban rail trains are usually unknown or cannot be precisely solved in practice, this deconvolution process is blind deconvolution. Then, a filter is designed to retain the frequency band corresponding to the fault signal, thereby reducing interference and highlighting fault information. Analysis of the fault signals of urban rail train wheelset bearings shows that the signals include periodic vibration and impact signals representing the fault, as well as other interference signals and noise. Blind deconvolution filtering enhances the periodic vibration and impact signals, which is beneficial for further diagnosis of wheelset bearing faults. Figure 4 The blind deconvolution process of the measured actual signal is demonstrated.
[0051] Taking the wheelset bearing system of an urban rail train as an example, the fault characteristic signal x(t) generated by the wheelset bearing failure is coupled with interference signals s(t) and noise signals n(t), and is transmitted to the sensor through the system to obtain the measured signal y(t). To obtain the signal x(t) that represents the fault characteristics, a filter needs to be designed to perform blind deconvolution filtering on y(t), and the optimal solution of the corresponding filtering objective function needs to be found.
[0052] Currently, one of the challenges of blind deconvolution is addressing the common ill-posed problem. Because multiple filtering processes can achieve the desired result, it is difficult for the result to converge to the optimal solution. Therefore, the deconvolution process should be constrained by different prior information to determine the direction of the deconvolution result. The sparsity index of the fault signal, as the objective function in the blind deconvolution filtering process, has become a key focus of current research.
[0053] When wheelset bearings malfunction, they generate periodic, continuous impacts, producing vibration signals characterized by both periodicity and impact. Therefore, to accurately perform blind deconvolution on the bearing vibration signal to obtain the impact signal representing the bearing fault, both the periodicity and impact of the signal must be considered. Based on the kurtosis representing the impact of the fault signal and the harmonic noise ratio representing the periodicity of the fault signal, a blind deconvolution filtering objective function combining periodicity and impact is proposed.
[0054] The model for the blind deconvolution filtering process is constructed and defined as follows:
[0055] s(n)=x(n)*f(l) (1)
[0056] Where s(n), n = 1, 2, 3, ..., N represents the filtered output signal with a length of N; x(n), n = 1, 2, 3, ..., N represents the input signal measured by the accelerometer with a length of N; f(l), l = 1, 2, 3, ..., L represents the filter coefficient sequence with a length of L; * represents convolution operation.
[0057] The objective function for blind deconvolution filtering is constructed, and an objective function based on the kurtosis of related harmonic noise is proposed, defined as follows:
[0058]
[0059] Where T represents the period of the input; F(f) represents the proposed objective function based on the kurtosis of related harmonic noise; f = f(l), l = 1, 2, 3, ..., L represents the filter coefficient sequence.
[0060] To maximize the kurtosis of the correlated harmonic noise in the output signal, we solve for the derivative of the objective function with respect to the filter coefficients and set the derivative to zero:
[0061]
[0062] Solving equation (3) yields the following:
[0063] p=f*Q (4)
[0064]
[0065]
[0066]
[0067] Therefore, the filter coefficients f based on the maximum value of the objective function can be solved using equation (8):
[0068] f = p * Q -1 (8)
[0069] Accurately identifying hidden periods or frequencies in a signal has become a crucial aspect of blind deconvolution filtering. Blind deconvolution filtering requires the ability to accurately identify hidden fault information even under strong interference, such as noise or interference signals. Furthermore, the proposed sparsity metric requires an initial period estimate of the corresponding signal as input. To address these requirements, a metric for accurately identifying hidden periodic information in a signal is proposed, named the Periodic Characteristic Energy Ratio (TCER). The proposed TCER calculation method is shown below:
[0070] The periodic characteristic energy ratio of the prior estimated period T is constructed and used to estimate the input period T of the objective function (2). First, the output signal after filtering the input period T0 is... The analytic signal is obtained by performing the Hilbert transform, as shown in equation (9).
[0071]
[0072] Where j represents the imaginary sign, This represents the analytic signal. Then the analytic signal... The signal is converted to the frequency domain and squared envelope is obtained to obtain the frequency domain signal shown in equation (10).
[0073]
[0074] Where abs represents the modulus of the complex signal, and FFT represents the Fast Fourier Transform. This represents the obtained frequency domain signal. Therefore, the periodic characteristic energy ratio for the input period T0 is calculated according to equation (11).
[0075]
[0076]
[0077]
[0078] in, The periodic characteristic energy ratio represents the input period T0; K represents the frequency domain signal. The corresponding input period T0 and its corresponding positive integer multiples of the period are in Quantity within the range; It represents rounding up; The Nyquist frequency is the signal sampling frequency f. s Half of; r(k) represents the weighting coefficient function, which represents the frequency domain signal corresponding to the input period T0 and its positive integer multiples of the period. The weight of the amplitude; T xThe period range representing the error can be expressed by equation (14).
[0079]
[0080] The proposed It can effectively measure the proportion of the input period T0 and its positive integer multiples in the entire frequency domain signal.
[0081] Based on harmonic noise kurtosis and periodic feature energy ratio, a new blind deconvolution filtering method is proposed by combining period identification with sparsity index and incorporating it into the iterative framework of blind deconvolution filtering. This method is named Harmonic Noise Kurtosis-Periodic Feature Deconvolution.
[0082] The blind deconvolution filter fault diagnosis method based on correlated harmonic noise kurtosis proposed in this invention is as follows: Figure 1 As shown, the specific steps are as follows:
[0083] Step 1: Measure the signal x(n) containing fault information using an accelerometer.
[0084] Step 2: Perform a blind deconvolution filtering iterative process based on the kurtosis of relevant harmonic noise.
[0085] Step 2.1: Determine the initial values of the required parameters, sampling frequency f. s ; Filter length L; Number of iterations m; Initial fault cycle search range T f = [T0, T1]; Initial filter coefficients f0.
[0086] Step 2.2: Enter the iteration process, calculate the initial filter coefficients f0 according to equation (1) and filter the output signal s0(n); and calculate the given range T according to equation (11). f = The maximum value of TCER for all periods within [T0, T1], where the period is denoted as T. max The periodic characteristic energy ratio at this time is denoted as TCER. max1 .
[0087] Step 2.3: Based on s0(n) and T obtained in the previous step max The filter coefficients are updated, and the filter coefficients f1 based on the maximum value of the objective function are calculated using equations (5), (6), and (8) according to the proposed relevant harmonic noise kurtosis.
[0088] Step 2.4: Update the filtered output signal s1(n) based on the current filter coefficient f1, and use f1 and s1(n) as the parameters required for the next iteration. Determine whether the iteration termination condition has been met, i.e., whether the maximum number of iterations m has been reached or the current TCER has been reached. max1The iteration process reaches its maximum value. If the iteration termination condition is met, the iteration process ends; otherwise, continue to step 2 for iteration.
[0089] Step 3: End the iteration process and obtain the filtered output signal s at this point. x (n). For s x (n) Perform envelope spectrum analysis to extract fault features and perform fault diagnosis.
[0090] Simulation and Verification:
[0091] The advantages of the proposed method are illustrated through a case study of a fault simulation signal. The formula for constructing the simulation signal of a bearing outer ring fault is as follows:
[0092] x(t)=a(t)+b(t)+c(t)+n(t) (15)
[0093] Where a(t) represents the periodic impact signal generated by the bearing outer ring fault, which is simulated by equation (16).
[0094]
[0095] Where K represents the number of impacts contained in the periodic impact signal; A0 represents the amplitude of the corresponding impact; g represents the damping coefficient; f1 represents the natural frequency of the bearing; and T represents the fault characteristic period of the outer ring fault. In this simulation experiment, the values are as follows: A0 = 0.8; g = 0.1; f1 = 3000; T = 1 / 90.
[0096] b(t) represents the frequency-modulated amplitude-modulated harmonic signal, which is simulated using equation (17-18).
[0097] b(t)=P1sin(2πh1t+θ1) (17)
[0098]
[0099] In the formula, P1 represents the amplitude modulation signal, θ1 represents the frequency modulation signal; Q1 and Q2 represent the amplitudes of the amplitude modulation and frequency modulation signals, respectively; l1 and l2 represent the frequencies of the amplitude modulation and frequency modulation signals, respectively. and These represent the phases of the amplitude-modulated (AM) and frequency-modulated (FM) signals, respectively; h1 represents the frequency of the AM / FM harmonic signal. The values used in this simulation are as follows: Q1 = 0.2; Q2 = 0.1; l1 = 3; l2 = 5; h1 = 10;
[0100] c(t) represents the random impact signal, which is simulated using equation (19).
[0101]
[0102] In the formula, K1 represents the number of random impact signals; g1 represents the damping coefficient of the random impact; f2 represents the natural frequency that generates the random impact; and t(k1) represents the time corresponding to the k1th random impact. The values used in this simulation are as follows: K1 = 3; g1 = 0.12; f2 = 2500.
[0103] n(t) represents a random Gaussian white noise signal. In this experiment, a Gaussian white noise signal with a signal-to-noise ratio of -15dB is selected.
[0104] Meanwhile, by setting the signal sampling frequency to 8192Hz and the sampling time to 0.5s, the time-domain waveform of the signal in this case can be obtained as follows: Figure 2 As shown. Early fault characteristics are weak and subject to strong noise interference, making accurate fault detection impossible. The proposed method is applied to process the signal with the following parameter settings: L = 100; m = 50; T f =[0,200]; f1=[1,-1,1,-1,…,1,-1]. Figure 3 The images show the time-domain waveform and frequency-domain Hilbert envelope spectrum after blind deconvolution filtering using the proposed method. It can be seen that the proposed method can accurately extract early and subtle bearing faults under strong noise interference.
Claims
1. A method for rolling bearing fault diagnosis based on harmonic noise kurtosis blind deconvolution filtering, characterized in that, The method includes the following steps: Step 1: Measure signal with presence of fault information by acceleration sensor ; Step 2: Perform an iterative blind deconvolution filtering process based on the kurtosis of relevant harmonic noise. Step 2.1: Determine initial values for the desired parameters, sampling frequency ; Filter length Number of iterations Initial fault period search range Initial filter coefficients ; Step 2.2: Enter iterative process, calculate initial filter coefficients Perform filtering on output signal ; and calculate maximum value of all periods within given range , the period at which the maximum value is recorded is noted as , the period characteristic energy ratio at this time is noted as ; Step 2.3: Based on the results obtained in Step 2.2 and The filter coefficients are updated, and the filter coefficients at the maximum value of the objective function are calculated based on the proposed relevant harmonic noise kurtosis. ; Step 2.4: Based on the filter coefficients at this time Update the filtered output signal and at this time and The parameters are used as needed for the next iteration; it is then determined whether the iteration termination condition has been met, which is reaching the maximum number of iterations. Or at this time The maximum value during the iteration process is reached; if the iteration termination condition is met, the iteration process ends; otherwise, return to step 2 for further iteration. Step 3: End the iteration process and obtain the filtered output signal at this point. ,right Envelope spectrum analysis is performed to extract fault features and diagnose the fault. Step (2) specifically includes: The model for the blind deconvolution filtering process is constructed and defined as follows: (1) in, This represents the filtered output signal, with a signal length of... ; This represents the input signal measured by the accelerometer, with a length of... ; This represents the filter coefficient sequence, with a length of... ; Represents convolution operation; Construct the objective function for blind deconvolution filtering, and determine the objective function based on the kurtosis of relevant harmonic noise: (2) in, Represents the input period; The proposed objective function is based on the kurtosis of relevant harmonic noise. Represents the filter coefficient sequence; Solve for the derivative of the objective function with respect to the filter coefficients, and set the derivative to zero: (3) Solving equation (3) yields: (4) (5) (6) (7) Therefore, the filter coefficients are based on the maximum value of the objective function. Solve using equation (8): (8)。 2. The blind deconvolution filter rolling bearing fault diagnosis method based on harmonic noise kurtosis as described in claim 1, characterized in that, Step (2) also specifically includes: Constructing a priori estimation period The estimated periodic characteristic energy ratio is used as the input period of the objective function (2). The estimate; First, regarding the input cycle Filtered output signal Perform Hilbert transform to obtain the analytic signal (9) in, represents the imaginary number symbol, Represents an analytical signal; Then analyze the signal Transform to the frequency domain and perform squared envelope to obtain the frequency domain signal shown in equation (10). (10) in, This represents the modulus of a complex signal. Represents Fast Fourier Transform, The resulting frequency domain signal represents the input period. The periodic characteristic energy ratio is calculated according to equation (11). (11) (12) (13) in, Represents the input cycle The periodic characteristic energy ratio below; Represents frequency domain signal The corresponding input period and its corresponding positive integer multiples of the period in Quantity within the range; It represents rounding up; The Nyquist frequency is the signal sampling frequency. Half of; The weighting coefficient function represents the input period. and the frequency domain signal corresponding to its frequency multiples and positive integer multiples of the period. The weight of the amplitude; The period range considering the error is represented by equation (14). (14)。
Citation Information
Patent Citations
Rolling bearing sound signal weak fault diagnosis method based on deconvolution enhancement
CN114124038A
Rolling bearing fault vibration signal analysis method
CN114441172A
