A method and system for extracting fault features from bearing data

Abstract

This invention belongs to the field of fault diagnosis technology, specifically relating to a method and system for extracting fault features from bearing data. The method includes the following steps: S1, dividing the acquired bearing vibration signal into multiple frequency bands and calculating the envelope spectrum of each band; selecting the frequency band with the largest peak amplitude in the envelope spectrum; determining the main peak frequency of the envelope spectrum as the fault characteristic frequency; and converting it to obtain the fault impact period T; S2, calculating the autocorrelation function of the bearing vibration signal; determining the length L of the FIR filter based on the attenuation characteristics of the autocorrelation function and the fault impact period T; and performing periodic synchronous averaging of the bearing vibration signal based on the fault impact period T. This invention solves the problem of easily amplifying interference or losing effective information under strong noise, making the extracted fault features purer and more prominent in the envelope spectrum, thus enhancing the ability to identify early, weak faults.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a method and system for extracting fault features from bearing data. Background Technology

[0002] Rolling bearings are indispensable components in rotating machinery. Vibration signal analysis is currently the most mainstream technical method for bearing condition monitoring and fault diagnosis. However, in industrial settings, early bearing fault characteristics often manifest as weak, periodic impact signals. These signals are easily masked by strong background interference such as gear meshing, machine frame vibration, and environmental noise, resulting in a low signal-to-noise ratio and making them difficult to effectively identify directly from time-domain or frequency-domain waveforms. Therefore, accurately extracting weak, periodic fault impact information from strong noise interference is a key technical challenge in the field of bearing fault diagnosis.

[0003] To address these issues, researchers have proposed various signal processing methods, among which blind deconvolution techniques have attracted significant attention due to their ability to effectively enhance periodic impulse components. Maximum Correlation Kurtosis Deconvolution (MCKD) is a representative method in the field of blind deconvolution. It iteratively optimizes an FIR filter to maximize the correlation kurtosis of the filtered signal, selectively enhancing repetitive transient impulses associated with a preset fault period while suppressing other random noise and periodic interference. However, the effectiveness of the MCKD method is largely limited by the manual setting of key parameters, such as the length of the FIR filter and the fault impulse period. Inaccurate parameter settings can severely affect the deconvolution effect. Furthermore, under extremely low signal-to-noise ratio conditions, the optimization process of MCKD may amplify noise, resulting in a significant amount of interfering harmonics remaining in the resulting envelope spectrum, leading to insufficient clarity of fault characteristics and poor spectral sparsity. While some improved methods introduce regularization techniques to enhance noise resistance, they typically employ fixed regularization weights, failing to adaptively adjust according to changes in the signal during the iteration process, making it difficult to achieve an optimal balance between noise suppression and fault detail preservation. Summary of the Invention

[0004] This invention provides a method and system for extracting fault features from bearing data, in order to solve the technical problem that existing technologies cannot adaptively adjust according to changes in signals during the iterative process during fault diagnosis, and it is difficult to achieve the optimal balance between noise suppression and fault detail preservation.

[0005] In a first aspect, the present invention provides a method for extracting fault features from bearing data, comprising the following steps:

[0006] S1, the collected bearing vibration signal is divided into multiple frequency bands and the envelope spectrum of each frequency band is calculated. The frequency band with the largest peak amplitude in the envelope spectrum is selected, and the main peak frequency of the envelope spectrum is determined as the fault characteristic frequency. The fault impact period T is then calculated.

[0007] S2, calculate the autocorrelation function of the bearing vibration signal, and determine the length L of the FIR filter based on the attenuation characteristics of the autocorrelation function and the fault impact period T; perform periodic synchronous averaging on the bearing vibration signal based on the fault impact period T, and extract a single complete impact response part from the waveform obtained after synchronous averaging as the coefficient of the FIR filter.

[0008] S3, with the goal of maximizing the correlation kurtosis, the optimal FIR filter is obtained by iteratively optimizing the coefficients of the FIR filter. During the iterative optimization process, the weighted sum of the L1 norm and the square of the L2 norm of the coefficients of the FIR filter is used as a regularization penalty term. In each iteration, the weight coefficients of the L1 norm and the square of the L2 norm are adjusted according to the sparsity of the current filtered signal envelope spectrum. When the spectral sparsity is lower than a preset threshold, the weight of the L1 norm is increased to enhance the sparsity of the FIR filter coefficients.

[0009] S4. The optimal FIR filter obtained through optimization is used to filter the bearing vibration signal. The envelope spectrum of the filtered bearing vibration signal is then analyzed to extract bearing fault features.

[0010] Furthermore, the expression for the relevant kurtosis is:

[0011]

[0012]

[0013]

[0014] in, This indicates that the signal obtained after filtering the bearing vibration signal by the FIR filter optimized in the k-th iteration is at time [time value missing]. The amplitude, where M represents the number of cycles. This indicates the number of sampling points corresponding to the fault impact period. Let N represent the correlation kurtosis, N represent the first eigenvalue, and D represent the second eigenvalue.

[0015] Furthermore, the acquired bearing vibration signal is divided into multiple frequency bands and the envelope spectrum of each frequency band is calculated, including: decomposing the bearing vibration signal into 10 sub-frequency band signals through a set of bandpass FIR filters with continuous center frequencies; performing Hilbert envelope demodulation on each sub-frequency band signal to obtain the envelope of the sub-frequency band signal; and performing spectral analysis on the envelope of the sub-frequency band signal to obtain the envelope spectrum of the sub-frequency band signal.

[0016] Furthermore, the length L of the FIR filter is determined based on the decay characteristics of the autocorrelation function and the fault impact period T, including: calculating the envelope of the autocorrelation function of the bearing vibration signal, partially fitting an exponential decay model starting from the maximum value of the envelope; and defining the time required for the exponential decay model to decay from the initial value to a preset threshold as the time delay. The fault impact period T and the time delay All are converted to the number of sampling points, and are denoted as follows: and Set the FIR filter length L to .

[0017] Furthermore, based on the fault impact period T, the bearing vibration signal is periodically and synchronously averaged, and a single average impact response waveform is extracted as the coefficient of the FIR filter. This includes: taking the impact with the largest amplitude in the bearing vibration signal as the starting point, and according to the number of sampling points corresponding to the fault impact period T. The bearing vibration signal is divided into multiple data segments; the corresponding sampling points of all data segments are superimposed and the arithmetic mean of the sampling points is calculated to obtain a signal of length [length missing]. The average waveform is obtained; a zero vector of length D is constructed as the coefficient of the FIR filter, and the average waveform is placed at the center of the zero vector.

[0018] Furthermore, the coefficients of the FIR filter are iteratively optimized, including:

[0019] In the Kth iteration, calculate the coefficients of the current FIR filter. The gradient of the FIR filter is used to update its coefficients according to the gradient ascent method. , Represents the gradient vector. Indicates the learning rate. This represents the coefficients of the FIR filter at the Kth iteration. This represents the coefficients of the (K+1)th iteration of the FIR filter, where the learning rate is... Set a value; repeat this update step until the number of iterations reaches the preset value or the change in the objective function value is less than the preset convergence threshold.

[0020] Furthermore, based on the sparsity of the envelope spectrum of the filtered signal, the weighting coefficients of the L1 norm and the square of the L2 norm are adjusted, including: calculating the Gini exponent of the envelope spectrum of the filtered signal in the current iteration as the spectral sparsity S;

[0021] Set the L1 norm weight to λ1 and the L2 norm square weight to λ2, and λ1 + λ2 is a constant; if the spectral sparsity S is less than the preset threshold, increase the value of λ1 and decrease the value of λ2 accordingly in the next iteration until λ1 reaches the preset upper limit value.

[0022] Furthermore, the optimized FIR filter is used to filter the bearing vibration signal, and envelope spectrum analysis is performed on the filtered bearing vibration signal, including:

[0023] The coefficients of the optimal FIR filter obtained through iterative optimization are used to perform convolution operation on the bearing vibration signal to obtain the enhanced fault signal. Then, Hilbert transform and fast Fourier transform techniques are used to perform envelope demodulation and spectrum analysis on the enhanced fault signal to obtain the envelope spectrum. Finally, the existence of the fault is confirmed and the fault type is determined.

[0024] Furthermore, the number of cycles M is 4.

[0025] Secondly, the present invention provides a fault feature extraction system for bearing data, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned fault feature extraction method for bearing data is implemented.

[0026] The beneficial effects are as follows: This invention determines the fault impact period through multi-band envelope spectrum analysis, reducing theoretical calculation errors caused by speed fluctuations or slippage, and improving the accuracy of subsequent deconvolution analysis. In FIR filter design, this invention combines the autocorrelation characteristics of bearing vibration signals and periodic synchronous averaging technology to set the initial length and coefficients of the FIR filter, allowing the FIR filter to be optimized from an initial value close to the actual fault impact pattern, thus improving the algorithm's convergence speed and optimization accuracy. More importantly, during the iterative optimization process, a regularization strategy that combines the weights of L1 and L2 norm penalty terms is introduced. This regularization strategy can enhance the coefficient sparsity of the FIR filter to strongly filter out noise while preserving the details of the fault impact waveform, solving the problem of easily amplifying interference or losing effective information under strong noise. This makes the extracted fault features purer and more prominent in the envelope spectrum, enhancing the ability to identify early, weak faults. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method for extracting fault features from bearing data. Detailed Implementation

[0028] An embodiment of the bearing data fault feature extraction method provided by this invention:

[0029] like Figure 1As shown, the method for extracting fault features from bearing data includes the following steps:

[0030] S1, the collected bearing vibration signal is divided into multiple frequency bands and the envelope spectrum of each frequency band is calculated. The frequency band with the largest peak amplitude in the envelope spectrum is selected, and the main peak frequency of the envelope spectrum is determined as the fault characteristic frequency. The fault impact period T is then calculated.

[0031] Wavelet packet transform is used to decompose the bearing vibration signal into multiple layers, resulting in a series of narrowband sub-signals. Hilbert transform is applied to each narrowband sub-signal to obtain its analytic signal, and the magnitude of the analytic signal is calculated to obtain the envelope signal. Then, a fast Fourier transform is performed on each envelope signal to obtain the envelope spectrum. The frequency corresponding to the peak with the largest amplitude, obtained by traversing the envelope spectra of all frequency bands, is the fault characteristic frequency f. The formula T = ... The fault impact period T is calculated.

[0032] In an optional embodiment, the bearing vibration signal is decomposed into 10 sub-band signals by a set of bandpass FIR filters with continuous center frequencies;

[0033] Hilbert envelope demodulation is performed on each sub-band signal to obtain the envelope of the sub-band signal, and spectral analysis is performed on the envelope to obtain the envelope spectrum of the sub-band.

[0034] Assuming the sampling frequency of the bearing vibration signal is 25600Hz, the effective analysis frequency range of the bearing vibration signal is 0-12800Hz. This range is divided into 10 sub-band signals with equal bandwidth and continuous center frequencies, each sub-band signal having a bandwidth of approximately 1280Hz. For example, the center frequency of the first sub-band signal can be set to 640Hz, the center frequency of the second sub-band signal to 1920Hz, and so on, up to the tenth sub-band signal. In this way, the bearing vibration signal is decomposed into 10 independent sub-band signals, each representing the vibration component of the original bearing vibration signal within a specific frequency range.

[0035] For each sub-band signal obtained from the above decomposition, such as the sub-band signal with a center frequency of 6400Hz, a Hilbert transform is applied to construct the analytic signal of the sub-band signal. The magnitude of the analytic signal is calculated to obtain the envelope signal of the sub-band signal. The envelope signal reflects the energy fluctuations within the sub-band signal, especially the modulation phenomenon caused by the periodic impacts due to bearing failure. A fast Fourier transform is performed on this envelope signal to obtain the envelope spectrum of the sub-band. If the bearing has a fault with a characteristic frequency of 120Hz, then obvious spectral peaks will appear at 120Hz and the harmonic frequencies of the sub-band in the envelope spectrum, while the envelope spectra of other fault-free information bands will be relatively flat.

[0036] S2, calculate the autocorrelation function of the bearing vibration signal, and determine the length L of the FIR filter based on the attenuation characteristics of the autocorrelation function and the fault impact period T; perform periodic synchronous averaging on the bearing vibration signal based on the fault impact period T, and extract a single complete impact response part from the waveform obtained after synchronous averaging as the coefficient of the FIR filter.

[0037] Calculate the autocorrelation function of the bearing vibration signal and observe its decay trend. Set the FIR filter length L to an integer multiple of the number of sampling points in the fault impact period T, such as twice or three times the period length, to ensure that the length covers the main region where the autocorrelation function decays from the peak to a stable state. Using the fault impact period T as a reference, divide the bearing vibration signal into multiple continuous segments of length T. Average these segments point by point to obtain the average waveform of one period. The average waveform is then used as the coefficient vector of the FIR filter.

[0038] In an optional embodiment, determining the length L of the FIR filter based on the attenuation characteristics of the autocorrelation function and the fault impulse period T includes:

[0039] Calculate the envelope of the autocorrelation function of the bearing vibration signal, and fit an exponential decay model to the portion of the envelope starting from the maximum point of the envelope;

[0040] The time required for the exponential decay model to decay from its initial value to 10% of its initial value is defined as the time delay. ;

[0041] The fault impact period T and the time delay All are converted to the number of sampling points, and are denoted as follows: and ;

[0042] Set the FIR filter length L to .

[0043] Determining the length L of the FIR filter is crucial to ensuring that the FIR filter can completely capture the fault impact response. The autocorrelation function of the bearing vibration signal is calculated, and its envelope is taken. The envelope reaches its maximum value at zero delay and decays with increasing delay; the decay rate reflects the duration of the fault impact response in the system. A coefficient of performance (e.g., L) is fitted to the decay portion of the envelope starting from the maximum value. The exponential function. Using this exponential function, the time required for the bearing vibration signal energy to decay from its initial peak value to 10% of its initial peak value can be calculated, denoted as the time delay. For example, this delay is calculated. It takes 0.04 seconds.

[0044] When the delay is obtained Then, the time delay The known fault impact period T is converted into the number of sampling points for use in bearing vibration signal processing. Assuming the sampling frequency of the bearing vibration signal is 25600Hz, and the known fault impact period T is 0.009s, the number of sampling points corresponding to the fault impact period is... 230 points. Latency. Corresponding number of sampling points The number of points is 1024. The length L of the FIR filter is set to three times the length of the filter. Plus This means the FIR filter has a length L of 1714 points. This length ensures that the FIR filter can not only cover the main attenuation part of the impulse response, but also include several fault impulse cycles, providing sufficient design freedom for subsequent optimization.

[0045] In an optional embodiment, the bearing vibration signal is periodically and synchronously averaged based on the fault impact period T, and a single average impact response waveform is extracted as the coefficient of the FIR filter, including:

[0046] Starting with the impact with the largest amplitude in the bearing vibration signal, the number of sampling points corresponding to the fault impact period T is determined. The bearing vibration signal is divided into multiple data segments;

[0047] The corresponding sampling points of all data segments are superimposed, and the arithmetic mean of the sampling points is calculated to obtain a value of length calculated based on the fault impact period T and the sampling frequency. The average waveform;

[0048] Construct a zero vector of length D as the coefficient of the FIR filter, and place the average waveform at the center of the zero vector.

[0049] To provide a good initial state for the iterative optimization of the FIR filter, a periodic synchronous averaging technique is used to extract a clean fault impulse waveform. Based on the above... Taking a sample of 230 points as an example, the sample point with the largest amplitude in the entire bearing vibration signal is found and used as the reference point for the first impact. Centered on this reference point, multiple data segments of 230 points each are extracted forward and backward at time intervals of 230 points. For example, if the total length of the bearing vibration signal is sufficient, 60 such data segments may be extracted, each roughly containing one fault impact.

[0050] After obtaining these 60 data segments, they are aligned and averaged. This involves summing the first sample points of all 60 data segments and averaging them, then summing the second sample points and averaging them, and so on, up to the 230th sample point. This process reduces random noise and other vibrational components unrelated to the fault impact period T, resulting in a well-defined average impact response waveform with a length of 230 points. A zero vector of length D (1714) is created. This 230-point average waveform is placed at the center of the zero vector, for example, starting from the 742nd point, with the remaining positions remaining zero. This constructed vector serves as the coefficients of the FIR filter; it already possesses the basic shape of the target fault waveform, significantly accelerating the convergence speed of subsequent optimizations.

[0051] S3, with the goal of maximizing the correlation kurtosis, the optimal FIR filter is obtained by iteratively optimizing the coefficients of the FIR filter. During the iterative optimization process, the weighted sum of the L1 norm and the square of the L2 norm of the coefficients of the FIR filter is used as a regularization penalty term. In each iteration, the weight coefficients of the L1 norm and the square of the L2 norm are adjusted according to the sparsity of the current filtered signal envelope spectrum. When the spectral sparsity is lower than a preset threshold, the weight of the L1 norm is increased to enhance the sparsity of the FIR filter coefficients.

[0052] Random, aperiodic, and strong impulse noise is very common, such as material falling or electromagnetic interference. If only traditional kurtosis is used as the optimization objective, the FIR filter may be optimized to amplify these random noises, leading to misjudgments. This invention aims to maximize the relevant kurtosis by iteratively optimizing the coefficients of the FIR filter. For the filtered signal... For each sampling point i, find it and its time delay. , , ...until The points. Multiply the magnitudes of these M points (the current point and M-1 delayed points), that is:

[0053] ;

[0054] Then square the product, repeat the above operation for all sampling points i, and sum all the results, then the numerator is:

[0055] ;

[0056] Calculate the filtered signal The total energy, which is the sum of the squares of the amplitudes at all sampling points: Take the total energy to the power of M, and the denominator Related kurtosis .

[0057] Iterative optimization is preferably performed using, but not limited to, the gradient ascent method. In each iteration, the filtered signal is calculated based on the coefficients of the current FIR filter. The correlation kurtosis value and the gradient with respect to the coefficients of the FIR filter are calculated based on the filtered signal. The Hoyer sparsity exponent of the envelope spectrum of the current filtered signal is calculated. A sparsity threshold is set. If the calculated sparsity exponent is lower than the sparsity threshold, the weight coefficient of the L1 norm is set to a larger value, such as 0.9, and the corresponding weight coefficient of the squared L2 norm is 0.1. Conversely, if the sparsity is higher than or equal to the sparsity threshold, the weight coefficient of the L1 norm is set to a smaller value, such as 0.2, and the weight coefficient of the squared L2 norm is 0.8. The gradient of the regularization penalty term is merged with the gradient of the correlation kurtosis to form the overall objective function gradient. The coefficients of the FIR filter are updated according to the gradient direction and the preset learning rate until the number of iterations reaches the upper limit or the coefficients of the FIR filter converge.

[0058] In an optional embodiment, the number of cycles M in the relevant kurtosis calculation is 4.

[0059] Correlation kurtosis is an effective indicator for evaluating the strength of periodic impact signals. The calculation of correlation kurtosis involves a key parameter: the number of periods M of the periodic impact signal used to calculate the correlation. This parameter defines the number of consecutive impact periods considered when calculating the kurtosis. Choosing an appropriate value for M is crucial. If M is too small, such as 1, the periodicity of the periodic impact signal cannot be effectively utilized to suppress random noise interference; if M is too large, small fluctuations in equipment rotation speed may cause period mismatch, thus reducing the sensitivity of the indicator.

[0060] Setting the number of periods M to 4 is an optimization result obtained from extensive experiments and engineering practice. This means that when calculating the correlation kurtosis at any given time t, this setting is used to examine the correlation between the signal value at that time and the signal values ​​at the previous one, two, and three periods. For a typical rolling bearing fault signal, its impact response has a clear periodicity, but it is also affected by noise and slight speed variations. Using M=4, that is, considering four consecutive impact periods, is sufficient to effectively confirm the periodic repetitive characteristics of a typical rolling bearing fault signal, distinguishing the fault impact from isolated random pulses. This setting allows the correlation kurtosis index, as the optimization objective, to more accurately and stably reflect the intensity of the fault characteristics.

[0061] In an optional embodiment, the coefficients of the FIR filter are iteratively optimized, and in the Kth iteration, the coefficients of the complete objective function, including the regularization penalty term, are calculated with respect to the current FIR filter. The gradient;

[0062] Update the coefficients of the FIR filter using the gradient ascent method: The learning rate μ is set to 0.01.

[0063] Repeat this update step until the number of iterations reaches the preset value of 200 or the change in the objective function value is less than the preset convergence threshold.

[0064] The core of iterative optimization lies in continuously adjusting the coefficients of the FIR filter to maximize an objective function that characterizes the significance of fault features, such as correlation kurtosis. At the beginning of the k-th iteration, the coefficients of the current FIR filter are used. The bearing vibration signal is filtered. The objective function J is calculated based on the filtered signal. The objective function J not only measures the intensity of the fault characteristics but also includes L1 and L2 norm regularization penalties on the coefficients of the FIR filter to ensure the sparsity and stability of the solution. The coefficient vector of the objective function J with respect to the FIR filter is calculated. The gradient vector is obtained by taking the partial derivative of each element in the matrix. The gradient vector indicates the direction in which the objective function J increases the most.

[0065] After obtaining the gradient, the gradient ascent method is used to update the coefficients of the FIR filter. The new FIR filter coefficients are shown below. The coefficients of the current FIR filter Adding the product of the step size and the gradient direction gives us, i.e. The learning rate μ is a hyperparameter controlling the update step size, set to 0.01 here. A smaller learning rate helps ensure the stability of the optimization process. In another embodiment, the coefficients of the FIR filter are updated using the subgradient method or the proximal gradient method. The update step is repeated continuously, for example, iterating from k=1 to k=2, k=3, and so on. There are two termination conditions for iteration, and the process stops when either one is met: first, the number of iterations reaches a preset upper limit of 200; second, after a certain iteration, the increment of the objective function J compared to the previous iteration is less than a very small preset threshold, such as one ten-thousandth, indicating that the optimization process has basically converged and the coefficients of the FIR filter have reached or are close to the optimal.

[0066] In an optional embodiment, the weighting coefficients of the L1 norm and the squared L2 norm are adjusted according to the sparsity of the current filtered signal envelope spectrum, including:

[0067] Calculate the Gini exponent of the envelope spectrum of the filtered signal in the current iteration as the spectral sparsity S;

[0068] Let the L1 norm weight be λ1, the L2 norm squared weight be λ2, and λ1 + λ2 be a constant;

[0069] If the spectral sparsity S is less than the preset threshold of 0.85, in the next iteration, the value of λ1 is increased and the value of λ2 is decreased accordingly until λ1 reaches the preset upper limit value.

[0070] This embodiment introduces an adaptive regularization strategy to better balance the performance of the FIR filter. In each iteration of optimization, the objective function includes an L1 norm penalty term weighted by λ1 and an L2 norm squared penalty term weighted by λ2. The L1 norm helps to produce sparse FIR filter coefficients, while the L2 norm squared helps to control the overall energy of the FIR filter coefficients. The sum of λ1 and λ2 is set to a constant, such as 1. After each iteration, the Gini index of the envelope spectrum of the filtered signal is calculated. It is an index that measures spectral sparsity, ranging from 0 to 1. A larger value indicates a sparser spectrum, i.e., a more prominent fault characteristic frequency peak.

[0071] The dynamic adjustment mechanism optimizes the regularization strength based on feedback from spectral sparsity. A target threshold for spectral sparsity is set at 0.85. At the end of an iteration, if the calculated Gini exponent S is less than 0.85, it indicates that the envelope spectrum of the current filtering result is not sparsity enough, and the fault characteristics are not clear enough. To enhance sparsity in the next iteration, the weight λ1 of the L1 norm penalty term is increased, and the weight λ2 of the L2 norm squared penalty term is correspondingly decreased. For example, if λ1 is currently 0.5 and λ2 is 0.5, it is adjusted to λ1=0.55 and λ2=0.45. The adjustment process continues until the Gini exponent S exceeds 0.85, or λ1 reaches a preset upper limit, such as 0.9, to prevent excessive pursuit of sparsity from causing other performance degradation. If the Gini exponent S is not less than 0.85, the weights remain unchanged, or λ1 is decreased.

[0072] S4. The optimal FIR filter obtained through optimization is used to filter the bearing vibration signal, and the envelope spectrum analysis of the filtered bearing vibration signal is performed to extract bearing fault features.

[0073] The coefficients of the optimal FIR filter obtained at the end of the iterative optimization are used as the parameters of the optimal FIR filter. Convolution operation is performed on the bearing vibration signal to obtain the enhanced fault signal. Hilbert transform and fast Fourier transform techniques are used again to perform envelope demodulation and spectrum analysis on the enhanced fault signal to obtain the final envelope spectrum. In the final envelope spectrum, the fault characteristic frequency and the harmonic components of the fault characteristics will appear as clear spectral peaks. By identifying these spectral peaks, the existence of the fault can be confirmed and the fault type can be determined.

[0074] An embodiment of the bearing data fault feature extraction system provided by this invention:

[0075] The bearing data fault feature extraction system includes a processor and a memory. The memory stores computer program instructions, which are executed by the processor to implement the above-mentioned bearing data fault feature extraction method.

[0076] The bearing data fault feature extraction system also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0077] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained by such a computer-readable medium.

[0078] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting fault features from bearing data, characterized in that, Includes the following steps: S1, the collected bearing vibration signal is divided into multiple frequency bands and the envelope spectrum of each frequency band is calculated. The frequency band with the largest peak amplitude in the envelope spectrum is selected, and the main peak frequency of the envelope spectrum is determined as the fault characteristic frequency. The fault impact period T is then calculated. S2, calculate the autocorrelation function of the bearing vibration signal, and determine the length L of the FIR filter based on the decay characteristics of the autocorrelation function and the fault impact period T. This includes: calculating the envelope of the bearing vibration signal autocorrelation function, partially fitting an exponential decay model starting from the maximum point of the envelope; and defining the time required for the exponential decay model to decay from the initial value to a preset threshold as the time delay. The fault impact period T and the time delay All are converted to the number of sampling points, and are denoted as follows: and Set the FIR filter length L to ; Based on the fault impact period T, the bearing vibration signal is periodically synchronously averaged. A single complete impact response portion is extracted from the waveform obtained after synchronous averaging and used as the coefficients of the FIR filter. This includes: starting from the impact with the largest amplitude in the bearing vibration signal, and according to the number of sampling points corresponding to the fault impact period T. The bearing vibration signal is divided into multiple data segments; the corresponding sampling points of all data segments are superimposed and the arithmetic mean of the sampling points is calculated to obtain a signal of length [length missing]. The average waveform; construct a zero vector of length D as the coefficients of the FIR filter, and place the average waveform at the center of the zero vector; S3, with the goal of maximizing the correlation kurtosis, the optimal FIR filter is obtained by iteratively optimizing the coefficients of the FIR filter. During the iterative optimization process, the weighted sum of the L1 norm and the square of the L2 norm of the coefficients of the FIR filter is used as a regularization penalty term. In each iteration, the weight coefficients of the L1 norm and the square of the L2 norm are adjusted according to the sparsity of the current filtered signal envelope spectrum. When the spectral sparsity is lower than a preset threshold, the weight of the L1 norm is increased to enhance the sparsity of the FIR filter coefficients. S4. The optimal FIR filter obtained through optimization is used to filter the bearing vibration signal. The envelope spectrum of the filtered bearing vibration signal is then analyzed to extract bearing fault features.

2. The method for extracting fault features from bearing data according to claim 1, characterized in that, The expression for the relevant kurtosis is: in, This indicates that the signal obtained after filtering the bearing vibration signal by the FIR filter optimized in the k-th iteration is at time [time value missing]. The amplitude, where M represents the number of cycles. This indicates the number of sampling points corresponding to the fault impact period. Let N represent the correlation kurtosis, N represent the first eigenvalue, and D represent the second eigenvalue.

3. The method for extracting fault features from bearing data according to claim 1, characterized in that, The acquired bearing vibration signal is divided into multiple frequency bands and the envelope spectrum of each frequency band is calculated. This includes: decomposing the bearing vibration signal into 10 sub-frequency band signals through a set of bandpass FIR filters with continuous center frequencies; performing Hilbert envelope demodulation on each sub-frequency band signal to obtain the envelope of the sub-frequency band signal; and performing spectral analysis on the envelope of the sub-frequency band signal to obtain the envelope spectrum of the sub-frequency band signal.

4. The method for extracting fault features from bearing data according to claim 1, characterized in that, The coefficients of the FIR filter are optimized iteratively, including: In the Kth iteration, calculate the coefficients of the current FIR filter. The gradient of the FIR filter is used to update its coefficients according to the gradient ascent method. , Represents the gradient vector. Indicates the learning rate. This represents the coefficients of the FIR filter at the Kth iteration. This represents the coefficients of the (K+1)th iteration of the FIR filter, where the learning rate is... Set a value; repeat this update step until the number of iterations reaches the preset value or the change in the objective function value is less than the preset convergence threshold.

5. The method for extracting fault features from bearing data according to any one of claims 1-4, characterized in that, Based on the sparsity of the envelope spectrum of the filtered signal, the weighting coefficients of the L1 norm and the square of the L2 norm are adjusted, including: calculating the Gini exponent of the envelope spectrum of the filtered signal in the current iteration as the spectral sparsity S; Set the L1 norm weight to λ1 and the L2 norm squared weight to λ2, and λ1 + λ2 is a constant; if the spectral sparsity S is less than the preset threshold, increase the value of λ1 and decrease the value of λ2 accordingly in the next iteration until λ1 reaches the preset upper limit value.

6. The method for extracting fault features from bearing data according to claim 5, characterized in that, The optimized FIR filter is used to filter the bearing vibration signal, and the envelope spectrum of the filtered bearing vibration signal is analyzed, including: The coefficients of the optimal FIR filter obtained through iterative optimization are used to perform convolution operation on the bearing vibration signal to obtain the enhanced fault signal. Then, Hilbert transform and fast Fourier transform techniques are used to perform envelope demodulation and spectrum analysis on the enhanced fault signal to obtain the envelope spectrum. Finally, the existence of the fault is confirmed and the fault type is determined.

7. The method for extracting fault features from bearing data according to claim 2, characterized in that, The value of the number of cycles M is 4.

8. A fault feature extraction system for bearing data, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the fault feature extraction method for bearing data according to any one of claims 1-7 is implemented.

Images