A bearing wear state monitoring method and system based on vibration signals

By using a bearing wear condition monitoring method based on vibration signals, and leveraging the mapping relationship of operating parameters and demodulation of multiple structural elements, the problem of resonant frequency band drift under varying operating conditions is solved. This achieves high-fidelity extraction of weak fault features and accuracy in condition assessment, while reducing false alarm rate and computational load.

CN123065758APending Publication Date: 2026-11-10JIANGXI CHANGLONG MOTOR CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611235747.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies suffer from high-frequency resonance band drift in bearings under variable speed and load conditions, leading to the failure of weak fault feature extraction. Furthermore, purely data-driven condition assessments have poor generalization ability under unseen operating conditions.

Method used

The bearing wear condition monitoring method based on vibration signals determines candidate frequency bands by utilizing the mapping relationship between operating parameters and bearing natural frequencies, accurately locates the target resonant frequency band by combining impact energy index, and generates real-time feature distribution by using multi-structural element morphological demodulation, and calculates the distribution distance to generate wear condition monitoring results.

Benefits of technology

It achieves dynamic and precise locking of the resonant frequency band under varying operating conditions, improves the accuracy of weak fault feature extraction, reduces the false alarm rate, and saves edge computing power and communication bandwidth through lightweight computing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

This application relates to the field of equipment fault diagnosis and predictive maintenance technology, and provides a method and system for monitoring bearing wear condition based on vibration signals. The method includes: acquiring vibration signals and operating parameters of the rotating equipment where the bearing is located; determining candidate frequency bands based on the mapping relationship between the operating parameters and the bearing's natural frequency, and determining a target resonant frequency band based on the impact energy index within the candidate frequency bands; performing multi-structural element morphological demodulation on the vibration signal within the target resonant frequency band to obtain an envelope signal, and constructing multi-dimensional feature samples based on the envelope signal; generating a real-time feature distribution based on the multi-dimensional feature samples, calculating the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by the operating parameters, and generating bearing wear condition monitoring results based on the distribution distance. This application can adaptively track changes in operating conditions, extract weak impact features under extremely low signal-to-noise ratios, and has strong generalization ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of equipment fault diagnosis and predictive maintenance technology, specifically to a bearing wear condition monitoring method and system based on vibration signals. Background Technology

[0002] Currently, in the field of fault diagnosis and predictive maintenance of rotating equipment, vibration signal monitoring is commonly used to assess bearing wear. Existing bearing monitoring technologies mostly employ fixed-band envelope demodulation combined with simple threshold judgment. However, under varying speed and load conditions, the high-frequency resonant band of the bearing drifts, and fixed-band filtering cannot adaptively track changes in operating conditions, leading to the failure of weak fault feature extraction. Furthermore, some existing technologies use purely data-driven black-box models for condition assessment, lacking physical constraints, exhibiting poor generalization ability under unseen operating conditions, and incurring high computational burden. Therefore, there is an urgent need for a condition assessment scheme that can adaptively track changes in operating conditions, extract weak impact features under extremely low signal-to-noise ratios, and possess clear physical and mathematical mechanisms. Summary of the Invention

[0003] To address the problems of existing technologies, such as the failure of weak fault feature extraction due to resonant frequency band drift under varying operating conditions and poor generalization ability of condition assessment, this application proposes a bearing wear condition monitoring method and system based on vibration signals. This method and system achieves dynamic and accurate locking of the resonant frequency band, high-fidelity demodulation of weak impact features, and wear condition measurement that adapts to different operating conditions.

[0004] To address the aforementioned technical problems, this application provides a method for monitoring bearing wear status based on vibration signals, comprising: acquiring vibration signals and operating parameters of the rotating equipment where the bearing is located; determining candidate frequency bands based on the mapping relationship between the operating parameters and the bearing's natural frequency, and determining a target resonance frequency band based on the impact energy index within the candidate frequency bands; performing multi-structural element morphological demodulation on the vibration signals within the target resonance frequency band to obtain an envelope signal, and constructing multi-dimensional feature samples based on the envelope signal; generating a real-time feature distribution based on the multi-dimensional feature samples, calculating the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by the operating parameters, and generating bearing wear status monitoring results based on the distribution distance.

[0005] In some embodiments, the operating condition parameters include rotational speed and load; determining the candidate frequency band based on the mapping relationship between the operating condition parameters and the bearing's natural frequency includes: determining the equivalent contact stiffness of the bearing based on the rotational speed and the load, and determining the bearing's natural frequency based on the equivalent contact stiffness; and determining the frequency range formed by the bearing's natural frequency and a preset frequency band offset as the candidate frequency band.

[0006] In some embodiments, determining the equivalent contact stiffness of the bearing based on the rotational speed and the load includes: inputting the rotational speed and the load into a preset stiffness mapping relationship to obtain the equivalent contact stiffness; determining the target resonance frequency band based on the impact energy index within the candidate frequency band includes: calculating the spectral kurtosis of each sub-band within the candidate frequency band, and determining the sub-band that satisfies the preset impact energy condition and its adjacent frequency bands as the target resonance frequency band.

[0007] In some embodiments, the multi-structural element includes a first structural element and a second structural element with different geometries; the step of performing multi-structural element morphological demodulation on the vibration signal within the target resonant frequency band to obtain an envelope signal includes: performing noise suppression processing on the vibration signal within the target resonant frequency band using the first structural element to obtain an intermediate signal; performing morphological preservation processing on the intermediate signal using the second structural element to obtain a morphologically preserved signal; performing morphological interpolation on the morphologically preserved signal to obtain an interpolation result; and determining the envelope signal based on the interpolation result.

[0008] In some embodiments, the first structural element is a flat structural element, and the second structural element is a parabolic structural element; the step of using the first structural element to perform noise suppression processing on the vibration signal within the target resonant frequency band to obtain an intermediate signal includes: using the flat structural element to sequentially perform morphological opening and morphological closing operations on the vibration signal to obtain the intermediate signal; the step of using the second structural element to perform morphological preservation processing on the intermediate signal to obtain a morphological preservation signal includes: using the parabolic structural element to perform morphological dilation and morphological erosion operations on the intermediate signal respectively to obtain the morphological preservation signal; the step of performing morphological difference operation on the morphological preservation signal to obtain a difference operation result includes: calculating the difference between the result of the morphological dilation operation and the result of the morphological erosion operation to obtain the difference operation result.

[0009] In some embodiments, the distribution distance is a manifold distance; the characteristic distribution baseline is a normal state covariance matrix based on the operating condition parameter index; generating bearing wear condition monitoring results based on the distribution distance includes: converting the distribution distance into a health index through a monotonic mapping function; and outputting a degradation indication or triggering a remaining life prediction process in response to the distribution distance meeting a preset trigger condition.

[0010] In some embodiments, the multidimensional feature samples include local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy; calculating the distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index includes: standardizing the multidimensional feature samples and constructing a real-time covariance matrix within a sliding time window; obtaining a normal state covariance matrix that matches the operating condition parameters; and determining the manifold distance based on the logarithm of the determinant of the normal state covariance matrix, the trace of the product of the inverse matrix of the normal state covariance matrix and the real-time covariance matrix, the logarithm of the determinant of the real-time covariance matrix, and the feature dimension parameter.

[0011] In some embodiments, the preset triggering condition includes the number of times the distribution distance continuously exceeds a dynamic threshold, wherein the dynamic threshold is determined based on the dispersion of the normal feature distribution corresponding to the current operating condition parameters; the triggering of the remaining life prediction process includes: acquiring a historical multidimensional feature sample sequence and inputting the historical multidimensional feature sample sequence into a time series prediction model to obtain the remaining life; generating a maintenance prompt based on the remaining life, and updating the feature distribution baseline based on the multidimensional feature samples under the current normal state; the time window length of the time series prediction model is determined based on the length of the historical multidimensional feature sample sequence and the prediction error.

[0012] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the bearing wear condition monitoring method based on vibration signals as described in the above solution.

[0013] This application also provides a bearing wear condition monitoring system based on vibration signals, comprising: a sensing module for acquiring vibration signals and operating parameters of the rotating equipment where the bearing is located; an edge computing node configured to determine candidate frequency bands based on the mapping relationship between the operating parameters and the bearing's natural frequency, determine a target resonance frequency band based on the impact energy index within the candidate frequency bands, perform multi-structural element morphological demodulation on the vibration signals within the target resonance frequency bands to obtain envelope signals, construct multi-dimensional feature samples based on the envelope signals, generate a real-time feature distribution, and calculate the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by the operating parameters to generate bearing wear condition monitoring results; when the distribution distance meets a preset trigger condition, upload a degradation feature sequence to a prediction platform; and receive baseline update data sent by the prediction platform and update the feature distribution baseline; and a prediction platform configured to, in response to the distribution distance meeting the preset trigger condition, receive the degradation feature sequence uploaded by the edge computing node, determine the remaining service life based on the degradation feature sequence, and send baseline update data to the edge computing node.

[0014] The beneficial effects of this application are as follows: Unlike existing technologies, this application discloses a bearing wear condition monitoring method and system based on vibration signals. It determines candidate frequency bands by mapping operating parameters to the bearing's natural frequency, and determines the target resonant frequency band by combining impact energy indicators. By constraining the search range using physical mechanisms, the resonant frequency band can dynamically drift with changes in rotational speed and load, solving the problem of fixed bandpass filter failure under varying operating conditions and improving the accuracy of weak fault feature extraction. Morphological demodulation is performed using multi-structural elements with different geometries, suppressing background noise while fully preserving the geometry of weak wear impacts, significantly improving the sensitivity of early fault features under extremely low signal-to-noise ratios. By calculating the distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition index, the purely data-driven black-box judgment is transformed into a manifold space metric with clear mathematical and physical mechanisms, greatly reducing the false alarm rate under varying operating conditions. Furthermore, the remaining life prediction process is triggered only when the distribution distance meets preset trigger conditions, allowing edge nodes to perform only lightweight calculations during normal operation, greatly saving edge computing power and communication bandwidth, and achieving seamless long-term embedding of the monitoring system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1This is a schematic flowchart of the bearing wear condition monitoring method based on vibration signals provided in this application; Figure 2 This is a schematic diagram of a predictive maintenance scenario according to an embodiment of this application; Figure 3 This is a schematic diagram of the bearing wear condition monitoring system based on vibration signals provided in this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] The terms "first," "second," and "third" used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0018] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0019] like Figure 1 As shown, this application provides a bearing wear condition monitoring method based on vibration signals. This method constrains the search range through physical mechanisms and combines morphological fidelity demodulation with adaptive distribution distance evaluation to form a complete monitoring closed loop. Specifically, it includes the following steps: Step S100: Obtain the vibration signal and operating parameters of the rotating equipment where the bearing is located.

[0020] Vibration signals can be acquired by accelerometers installed on bearing housings or rotating equipment casings to reflect the impact and vibration conditions during bearing operation.

[0021] Operating condition parameters are used to characterize the current working state of rotating equipment, and can be one or more parameters such as speed, load, torque, temperature or pressure.

[0022] For example, in the scenario of monitoring motor bearings, operating condition parameters may include the current speed of the motor and the output shaft load; in the scenario of monitoring wind turbine main shaft bearings, operating condition parameters may include the main shaft speed and the generator power.

[0023] By synchronously acquiring vibration signals and operating parameters, a data foundation is provided for subsequent frequency band searches based on physical mechanisms.

[0024] Step S200: Based on the mapping relationship between operating condition parameters and bearing natural frequency, candidate frequency bands are determined, and the target resonance frequency band is determined based on the impact energy index within the candidate frequency bands.

[0025] When a bearing experiences wear or localized defects, it generates high-frequency impacts, which excite the natural frequencies of the bearing assembly or structure to resonate. Since the natural frequencies of bearings are not fixed but drift with changes in operating parameters, this embodiment establishes a mapping relationship between operating parameters and the bearing's natural frequencies, enabling the dynamic estimation of the approximate range of the natural frequencies under current operating conditions.

[0026] The candidate frequency band is the frequency range set around the calculated inherent frequency.

[0027] The impact energy index is used to quantify the magnitude of the impact energy contained in each sub-band within a candidate frequency band. For example, it can be achieved by calculating the kurtosis, energy operator, or envelope spectrum peak value of the signal within the frequency band.

[0028] By selecting sub-bands that meet the impact energy index requirements, the target resonance band can be precisely locked to the frequency band where resonance actually occurs under the current operating conditions.

[0029] The mechanism of first narrowing the search range by utilizing physical mechanisms and then accurately locating the target by utilizing the impact energy index effectively solves the problem that fixed-band filtering cannot track the resonant band drift under varying operating conditions, and improves the accuracy of weak fault feature extraction.

[0030] Furthermore, the operating condition parameters include speed and load. Candidate frequency bands are determined based on the mapping relationship between the operating condition parameters and the bearing's natural frequency, including: determining the bearing's equivalent contact stiffness based on speed and load, and determining the bearing's natural frequency based on the equivalent contact stiffness; the frequency range formed by the bearing's natural frequency and the preset frequency band offset is determined as the candidate frequency band.

[0031] During bearing operation, the contact state between the rolling elements and raceways changes with variations in external operating conditions. According to Hertzian contact theory, the contact area between the rolling elements and raceways undergoes elastic deformation, and the degree of this deformation directly determines the equivalent contact stiffness of the bearing.

[0032] When the rotational speed of rotating equipment changes, centrifugal force causes a change in the load distribution inside the bearing; when the load changes, the normal force acting between the rolling elements and the raceway increases or decreases accordingly. Both of these changes in operating parameters will cause dynamic drift in the equivalent contact stiffness.

[0033] Since the natural frequency of a bearing is proportional to the square root of its equivalent contact stiffness, changes in stiffness inevitably lead to a shift in the natural frequency. Therefore, this embodiment introduces two key operating parameters, rotational speed and load, to establish a physical mapping relationship between equivalent contact stiffness and natural frequency, thereby enabling accurate calculation of the bearing's natural frequency under the current operating conditions.

[0034] After determining the bearing's natural frequency, this embodiment does not directly use that frequency point as the filter center, but instead combines it with a preset frequency band offset to form a candidate frequency band.

[0035] The preset frequency band offset can be reasonably set based on the bearing's structural parameters and historical test data. For example, it can be set to between 5% and 10% of the natural frequency. Subtracting the preset frequency band offset from the bearing's natural frequency as the lower limit and adding the preset frequency band offset as the upper limit, the resulting frequency range is the candidate frequency band.

[0036] By setting a preset frequency band offset, a certain frequency tolerance space can be reserved based on the physical mechanism calculation to accommodate the small frequency shifts caused by sensor sampling errors or local temperature fluctuations, ensuring that the real resonant frequency band is completely surrounded within the candidate frequency band.

[0037] To more clearly illustrate the drift of the candidate frequency band with operating conditions, a certain type of deep groove ball bearing will be used as an example. When the rotating equipment is operating at low speed and light load, the equivalent contact stiffness is small, and the calculated bearing natural frequency may be located in a lower frequency band, such as around 6 kHz. In this case, the candidate frequency band may cover the range of 5,500 Hz to 6,500 Hz.

[0038] When rotating equipment switches to high-speed and heavy-load conditions, the centrifugal force and normal contact force increase, and the equivalent contact stiffness rises significantly. The bearing's natural frequency may drift to around 9,000 Hz. At this time, the candidate frequency band is dynamically adjusted to the range of 8,500 Hz to 9,500 Hz. If the fixed frequency band filtering scheme in the prior art is used, when the operating conditions change drastically, the fixed frequency band will not be able to cover the drifted true resonant frequency band, resulting in the filtering out of impact energy and the complete loss of weak fault features. In this embodiment, the candidate frequency band is dynamically determined through physical mapping relationships, so that the search range can always adaptively drift with the changes in operating conditions. This provides a reliable physical constraint basis for accurately locking the target resonant frequency band within the candidate frequency band, effectively avoiding the problem of feature extraction failure under changing operating conditions.

[0039] Furthermore, the rotational speed and load are mapped to a preset stiffness to obtain the equivalent contact stiffness. The spectral kurtosis of each sub-band within the candidate frequency band is calculated, and the sub-band that satisfies the preset impact energy condition and its adjacent frequency bands are determined as the target resonance frequency band.

[0040] The preset stiffness mapping relationship is based on a mathematical model or lookup table pre-established according to Hertzian contact theory. Since the contact between the rolling elements and raceways inside the bearing is a nonlinear elastic contact, its equivalent contact stiffness is not a simple linear superposition, but is closely related to the rolling element diameter, contact angle, material elastic modulus, and equivalent load distribution. In actual operation, changes in rotational speed will cause changes in centrifugal force and gyroscopic torque, thus affecting the contact load distribution between the rolling elements and raceways; changes in load directly alter the normal clamping force.

[0041] By inputting real-time acquired rotational speed and load into a preset stiffness mapping relationship, the equivalent contact stiffness under the current working condition can be quickly and accurately deduced. For example, this preset stiffness mapping relationship can be represented as a multivariate nonlinear function, or a multidimensional lookup table pre-calculated through finite element simulation and stored in a database. By parameterizing the physical mechanism in this way, the real-time solution of complex partial differential equations is avoided, greatly reducing the computational load on edge computing nodes.

[0042] After obtaining the equivalent contact stiffness and determining the candidate frequency band, this embodiment uses spectral kurtosis as an impact energy index to accurately locate the target resonance frequency band. Spectral kurtosis is a frequency domain statistic used to quantify the energy of the impact component in a signal. It is extremely sensitive to transient impacts in a signal but insensitive to stationary random noise. The specific logic for calculating the spectral kurtosis map is as follows: the candidate frequency band is divided into multiple equal-width sub-bands. For the vibration signal within each sub-band, a short-time Fourier transform or wavelet packet decomposition is performed. The normalized value of the fourth-order cumulant of each sub-band signal within different time windows is calculated, which is the spectral kurtosis value. The spectral kurtosis values ​​of all sub-bands constitute the spectral kurtosis map within the candidate frequency band.

[0043] In a spectral kurtosis map, sub-bands that meet a preset impact energy condition typically exhibit spectral kurtosis peaks. The preset impact energy condition can be a spectral kurtosis value greater than a preset threshold, or a spectral kurtosis value that is the largest among all sub-bands.

[0044] However, this embodiment does not use only a single peak sub-band as the target resonant band, but rather determines the target resonant band together with its adjacent bands. The mechanism of this design is that when the impact signal generated by bearing wear passes through the structural transmission path, its energy diffuses to both sides of the center frequency, forming a resonant band with a certain bandwidth. If only a single peak sub-band is selected for bandpass filtering, part of the impact energy will be cut off, causing distortion of the envelope signal obtained by subsequent demodulation and loss of the periodic characteristics of the weak fault.

[0045] To more clearly illustrate the limitations of taking only the peak frequency band, a comparative example is introduced.

[0046] Suppose that under certain operating conditions, the spectral kurtosis value of sub-band A within the candidate frequency band is calculated to be the maximum peak value, while the spectral kurtosis values ​​of its adjacent sub-bands B and C are slightly lower but still significantly higher than the background noise level. If, following the conventional approach of existing technology, only sub-band A is selected as the target resonant frequency band, severe edge effects will appear in the waveform of the envelope signal during subsequent morphological demodulation due to the excessively narrow bandwidth, resulting in a significant decrease in the extraction accuracy of features such as local morphological kurtosis.

[0047] The scheme in this embodiment uses sub-bands A, B, and C together as the target resonant band, which can completely cover the resonant energy distribution area and ensure high-fidelity transmission of the impact morphology. This embodiment, by combining physical mechanism mapping and spectral kurtosis local peak extension locking, ensures both the adaptability of the target resonant band search and the integrity of the impact signal energy, providing a high-quality signal input foundation for subsequent multi-structural element morphological demodulation.

[0048] Step S300: Perform multi-structural element morphological demodulation on the vibration signal within the target resonance frequency band to obtain the envelope signal, and construct multi-dimensional feature samples based on the envelope signal.

[0049] Morphological demodulation is a nonlinear filtering method based on the geometric morphological characteristics of a signal. Multiple structural elements refer to the combined processing of a signal using at least two structural elements with different geometries. Different structural elements can selectively preserve or suppress components with different morphological characteristics in the signal. For example, one structural element can be used to filter out positive and negative impulse noise in a signal, while another can be used to preserve the true geometric morphology of bearing wear impact.

[0050] By sequentially performing morphological operations on multiple structural elements of the vibration signal within the target resonant frequency band, strong background noise interference can be effectively suppressed, while the low-frequency envelope signal of weak impacts can be extracted with high fidelity. The envelope signal contains the periodic impact characteristics of bearing wear.

[0051] Constructing multidimensional feature samples based on envelope signals refers to extracting sensitive features from the envelope signal across multiple dimensions, such as impact amplitude features in the time domain, harmonic energy distribution features in the frequency domain, and signal complexity features. The construction of multidimensional feature samples enriches the representational dimensions of fault information, providing more comprehensive data support for subsequent condition assessment.

[0052] By using multi-structural element morphological demodulation, the geometry of weak wear impact can still be maintained even at extremely low signal-to-noise ratios, significantly improving the sensitivity of early fault characteristics.

[0053] Specifically, the multi-structural elements include first and second structural elements with different geometries. Multi-structural element morphological demodulation is performed on the vibration signal within the target resonant frequency band to obtain an envelope signal. This includes: using the first structural element to perform noise suppression processing on the vibration signal within the target resonant frequency band to obtain an intermediate signal; using the second structural element to perform morphological preservation processing on the intermediate signal to obtain a morphologically preserved signal; performing morphological interpolation on the morphologically preserved signal to obtain the interpolation result; and determining the envelope signal based on the interpolation result.

[0054] The core of morphological demodulation lies in using structuring elements of specific shapes to match and extract geometric features from signals. The first and second structuring elements differ significantly in geometry, and this difference determines their different functional roles in the signal processing process.

[0055] The first structural element is mainly used to perform noise suppression processing. Its geometry is usually relatively flat or regular, which can effectively filter out positive and negative pulse noise and continuous background interference in the signal. When the vibration signal in the target resonant frequency band is input, the first structural element strips off the high-frequency spikes and random disturbances superimposed on the impact features through specific morphological operations, and outputs a relatively smooth intermediate signal that retains the main impact profile.

[0056] The second structural element is primarily used to perform shape preservation processing. Its geometry typically matches the actual shape of the impact signal generated by bearing wear, such as having gradually changing edges or a specific curvature. The second structural element acts on the intermediate signal, precisely matching the peaks and troughs of the weak impact, further eliminating residual flat noise without damaging the impact shape, thus obtaining the shape preservation signal.

[0057] Subsequently, morphological difference operations are performed on the morphology-preserving signal. By calculating the difference between the results of different morphological operations, the low-frequency envelope features can be completely separated from the carrier signal, and the envelope signal can be finally determined.

[0058] To more clearly illustrate the effect of geometrical differences on the fidelity of weak impacts, a comparative example is introduced. If a single structural element is used for demodulation in the existing technology, such as using only a flat structural element, although this structural element can effectively suppress impulse noise, its geometry is too simple and cannot adaptively match the asymmetric waveform of the weak impact signal from the bearing.

[0059] During processing, a single structuring element can flatten the sharp peaks of the impact signal, causing severe distortion of the envelope signal and a significant decrease in the accuracy of subsequently extracted features such as local morphological kurtosis. This embodiment, however, achieves functional decoupling of noise suppression and morphological preservation by introducing first and second structuring elements with different geometries.

[0060] The first structural element focuses on filtering out noise, providing a clean signal noise floor for the extraction of impact features; the second structural element focuses on shape matching, ensuring that the true geometry of weak impacts is not destroyed. This functional binding mechanism can still extract the low-frequency envelope signal of weak impacts with high fidelity even at extremely low signal-to-noise ratios, significantly improving the sensitivity of early fault features.

[0061] It should be understood that the geometry of the first and second structural elements is not limited to a specific form. For example, the first structural element can be a flat, linear, or constant-shape structural element to achieve broadband noise suppression; the second structural element can be a parabolic, triangular, or sinusoidal structural element to maintain impact patterns with different curvatures. These embodiments are merely illustrative and not restrictive. Any structural element with different geometries that performs noise suppression and pattern maintenance functions respectively should be covered within the scope of this application.

[0062] Optionally, the first structural element is a flat structural element, and the second structural element is a parabolic structural element. Noise suppression processing is performed on the vibration signal within the target resonant frequency band using the first structural element to obtain an intermediate signal, including: sequentially performing morphological opening and closing operations on the vibration signal using the flat structural element to obtain the intermediate signal.

[0063] The intermediate signal is subjected to shape preservation processing using the second structuring element to obtain the shape preservation signal. This includes performing shape dilation and shape erosion operations on the intermediate signal using a parabolic structuring element to obtain the shape preservation signal.

[0064] Perform morphological interpolation on the morphology-preserving signal to obtain the interpolation result, including: calculating the difference between the result of the morphological dilation operation and the result of the morphological erosion operation.

[0065] A flat structural element refers to a geometric shape with a constant amplitude within its domain; its mathematical definition can be expressed as: ,in This is the length variable for the structure element.

[0066] In morphological operations, flat structuring elements appear as flat cut surfaces, and their main function is to filter out positive and negative pulse noise in the signal whose width is smaller than the length of the structuring element.

[0067] The morphological opening operation is defined as a process that first performs erosion and then dilation, that is:

[0068] in It is a vibration signal. For morphological opening operations. For erosion calculation, For expansion operations, It is a flat structuring element. The opening operation can eliminate isolated positive pulses in a signal.

[0069] The morphological closing operation is defined as a process that first performs dilation and then erosion, that is:

[0070] in This is a morphological closing operation. Closing operations can fill the gaps between negative pulses in a signal.

[0071] By sequentially performing shape opening and shape closing operations, the flat structural element can effectively strip away high-frequency spikes and random disturbances superimposed on the impact characteristics, outputting a relatively smooth intermediate signal that retains the main impact profile. This combination of operations suppresses impulse noise without causing destructive clipping to the overall envelope trend of the signal.

[0072] A parabolic structural element refers to a geometric shape in which the amplitude varies with the length variable in a quadratic curve. Its mathematical definition can be expressed as:

[0073] in The curvature coefficient, The total length of the structuring element. The length is a variable. The curvature characteristics of the parabolic structural element closely match the true geometry of the transient impact signal generated by bearing wear.

[0074] By performing morphological dilation and morphological erosion operations on the intermediate signal using a parabolic structuring element, a morphologically preserved signal is obtained.

[0075] The mathematical definition of morphological dilation is:

[0076] in This is an intermediate signal. For translation variables. The dilation operation can find the local maxima of the signal under the constraints of the structuring element, thus matching the peak of the impact signal.

[0077] The mathematical definition of morphological erosion operation is:

[0078] Erosion operations can find local minima of a signal under the constraints of structural elements, thus matching the troughs of the impact signal. Through the expansion and erosion operations of parabolic structural elements, the asymmetric waveform of weak impacts can be accurately matched, and residual flat noise can be further eliminated without damaging the impact morphology, resulting in a shape-preserving signal.

[0079] Subsequently, morphological interpolation is performed on the morphology-preserving signal to obtain the interpolation result. The mathematical definition of interpolation is:

[0080] in This is the result of the difference operation.

[0081] By calculating the difference between the dilation and erosion results, the low-frequency envelope features can be completely separated from the carrier signal, and the envelope signal can be determined.

[0082] The difference operation is essentially a nonlinear gradient extraction process. Since the parabolic structuring element has accurately preserved the shape of the shock wave peaks and troughs, the difference can restore the true energy distribution of the shock to the greatest extent.

[0083] To more clearly illustrate the synergistic mechanism of flat-type impulse noise suppression and parabolic-type impulse shape retention, a comparative example is introduced.

[0084] If only parabolic structuring elements are used for interpolation, although the impact shape can be preserved, the strong background impulse noise mixed in the signal will be amplified by the parabolic structuring elements due to the lack of pre-processing for noise suppression, resulting in a large number of spurious peaks in the envelope signal.

[0085] If only flat structural elements are used for interpolation, although impulse noise can be suppressed, the flat cutting characteristics of the flat structural elements will flatten the sharp peaks of the impulse signal, resulting in severe distortion of the envelope signal.

[0086] This embodiment first uses a flat structural element to perform opening and closing operations to suppress noise, and then uses a parabolic structural element to perform expansion and erosion difference operations to preserve shape. This achieves functional decoupling and cascaded synergy between noise suppression and shape preservation, enabling the high-fidelity extraction of low-frequency envelope signals with weak impacts even at extremely low signal-to-noise ratios.

[0087] It should be understood that the specific parameters and order of operations for the aforementioned structural elements are merely illustrative and not restrictive. For example, the length of a flat structural element can be adaptively adjusted based on the center frequency of the target resonant band, and the curvature coefficient of a parabolic structural element can be reasonably selected based on the type of bearing. In other embodiments, the first structural element can also be a linear or constant-type structural element, and the second structural element can also be a triangular or sinusoidal structural element. As long as the two types of structural elements have different geometric shapes and respectively perform the functions of noise suppression and shape preservation, they should be covered within the scope of protection of this application.

[0088] Step S400: Generate a real-time feature distribution based on multi-dimensional feature samples, calculate the distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index, and generate bearing wear condition monitoring results based on the distribution distance.

[0089] Real-time feature distribution refers to the statistical distribution composed of multiple multidimensional feature samples within the current monitoring time or sliding time window. The feature distribution baseline is a reference model of the feature distribution of the bearing under normal conditions, corresponding to specific operating parameters.

[0090] Since the normal feature distribution differs under different operating conditions, this embodiment uses the operating condition parameters as an index to match and obtain the feature distribution baseline under the current operating condition, thereby achieving adaptive baseline alignment. The distribution distance is used to quantify the degree of statistical difference between the real-time feature distribution and the feature distribution baseline, and can be calculated using measures such as manifold distance, Bach distance, or KL divergence.

[0091] When bearings wear, the real-time feature distribution deviates from the normal baseline, resulting in an increased distribution distance. Monitoring results generated based on this distribution distance can include directly outputting the distribution distance value as a wear quantification indicator, converting the distribution distance into a health index through a mapping function, or outputting an alarm signal when the distribution distance exceeds a threshold.

[0092] By measuring the distribution distance, the purely data-driven black-box judgment is transformed into a manifold space metric with clear mathematical and physical mechanisms, significantly reducing the false alarm rate under varying operating conditions. Furthermore, since the calculation of the distribution distance only triggers the subsequent prediction process when a deviation is confirmed, edge nodes only need to perform lightweight calculations under normal circumstances, greatly saving edge computing power and communication bandwidth, and enabling seamless long-term embedding of the monitoring system.

[0093] Optionally, the distribution distance is the manifold distance; the characteristic distribution baseline is the normal state covariance matrix based on the index of operating condition parameters; the distribution distance is converted into a health index through a monotonic mapping function; in response to the distribution distance meeting the preset triggering conditions, a degradation indication is output or the remaining life prediction process is triggered.

[0094] Manifold distance is a geometric distance used to measure the difference between two symmetric positive definite matrices. In bearing condition monitoring, both the real-time feature distribution and the feature distribution baseline can be represented as covariance matrices. Since the covariance matrix is ​​naturally a symmetric positive definite matrix, manifold distance is very suitable for quantifying the degree of shift in the feature distribution. Compared to the traditional Mahalanobis distance, manifold distance fully considers the nonlinear geometric structure of the Riemannian manifold space where the covariance matrix resides. Mahalanobis distance essentially forcibly straightens the covariance matrix into a vector in Euclidean space for distance calculation. This approach destroys the intrinsic correlation between matrix elements, easily leading to distance measurement distortion when dealing with high-dimensional and strongly correlated multidimensional feature samples. In contrast, manifold distance is calculated using the trace of the product of the logarithmic determinant and the inverse matrix of the matrix, preserving the complete information of the second-order statistical properties of the feature distribution. It can more sensitively capture subtle changes in the feature covariance structure caused by early, minor wear.

[0095] The characteristic distribution baseline is the normal state covariance matrix based on the index of operating condition parameters. Since the distribution range of normal vibration characteristics of bearings naturally differs under different speeds and loads, this embodiment pre-establishes a baseline library of covariance matrices corresponding to each operating condition interval under normal conditions, using operating condition parameters as an index. During real-time monitoring, the system matches and retrieves the corresponding normal state covariance matrix from the baseline library based on the currently acquired operating condition parameters, thereby achieving adaptive benchmark alignment and effectively avoiding false alarms caused by operating condition fluctuations.

[0096] The distribution distance is converted into a health index using a monotonic mapping function. This monotonic mapping function can be a hyperbolic tangent function, logistic function, or piecewise linear function, etc. Its purpose is to compress and map the distribution distance values, which can range from zero to infinity, to a preset finite interval, such as zero to one hundred, for easier visualization and threshold management. The health index monotonically decreases or increases with increasing distribution distance, allowing maintenance personnel to quickly determine the bearing's health status through intuitive numerical values.

[0097] In response to the distribution distance meeting a preset trigger condition, the system outputs a degradation indication or triggers a remaining lifetime prediction process. The preset trigger condition can be flexibly configured according to specific application scenarios. For example, the trigger condition can be set to the number of times the distribution distance continuously exceeds a dynamic threshold, where the dynamic threshold is determined based on the dispersion of the normal characteristic distribution corresponding to the current operating parameters. When this condition is met, the system outputs a degradation indication, such as issuing an alarm signal or highlighting it in red on the interface; simultaneously or optionally, it triggers the remaining lifetime prediction process, sending historical data to the computing unit for lifetime prediction. This distribution distance-based triggering mechanism allows edge nodes to perform only lightweight distance calculations under normal circumstances, only activating the complex prediction process when deterministic degradation is confirmed, greatly saving edge computing power and communication bandwidth.

[0098] To more clearly illustrate the advantages of manifold distance over Mahalanobis distance, a comparative example is introduced. Suppose that under a certain operating condition, a bearing experiences early, minor pitting corrosion on the outer ring. In this case, the mean of the multidimensional feature samples may not have a significant shift, but the correlation structure between features has subtly changed. If Mahalanobis distance is used for evaluation, because the mean shift is insignificant, the Mahalanobis distance value may still be within the normal threshold range, leading to missed detections. However, manifold distance is extremely sensitive to minute changes in the covariance matrix structure, accurately capturing such correlation changes. The manifold distance value rises significantly and exceeds the threshold, thus triggering a degradation indicator in a timely manner.

[0099] By introducing manifold distance and operating condition baseline indexing mechanisms, the purely data-driven black-box judgment is transformed into a manifold space metric with clear mathematical and physical mechanisms, which significantly reduces the false alarm rate and false negative rate under varying operating conditions.

[0100] It should be understood that the specific form of the monotonic mapping function and the specific threshold settings for the triggering conditions described above are merely interpretive and not restrictive. For example, the monotonic mapping function could also be an exponential decay function, and the triggering condition could be that the integral value of the distribution distance over a period of time exceeds a preset energy threshold. As long as the logic of converting the distribution distance into a health index and triggering subsequent processes based on conditions is satisfied, it should be covered within the scope of protection of this application.

[0101] Specifically, the multidimensional feature samples include local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy. The distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index is calculated, including: standardizing the multidimensional feature samples and constructing a real-time covariance matrix within a sliding time window; obtaining the normal-state covariance matrix that matches the operating condition parameters; and determining the manifold distance based on the logarithm of the determinant of the normal-state covariance matrix, the trace of the product of the inverse of the normal-state covariance matrix and the real-time covariance matrix, the logarithm of the determinant of the real-time covariance matrix, and the feature dimension parameters.

[0102] Local morphological kurtosis in multidimensional feature samples is used to quantify the sharpness of impact features in the envelope signal. Compared to traditional kurtosis, local morphological kurtosis, by combining a morphological gradient operator, calculates the fourth-order cumulant only for the transient shock wave peak caused by bearing wear in the envelope signal, thus effectively suppressing the interference of background noise fluctuations on the kurtosis value. Its value increases significantly with the occurrence of early minor spalling or pitting of the bearing.

[0103] The relative energy ratio of harmonics is used to characterize the energy proportion of fault characteristic frequencies and their harmonic components in the envelope spectrum.

[0104] By extracting the sum of the first few harmonic energies of the fault characteristic frequencies of the bearing outer ring, inner ring, or rolling element in the envelope signal spectrum, and calculating its ratio to the total frequency band energy, it is possible to effectively suppress the overall energy drift caused by load fluctuations and highlight the energy concentration of periodic fault components.

[0105] Improved permutation entropy is used to measure the complexity and irregularity of an envelope signal sequence. The permutation entropy value is obtained by reconstructing the phase space and calculating the probability distribution of permutation patterns at each phase point. When the bearing is in normal condition, the vibration signal mainly exhibits random noise, resulting in a high permutation entropy. When the bearing wears and generates periodic impacts, the determinism of the signal increases, and the complexity decreases. Improved permutation entropy, by introducing weighting coefficients, suppresses the information redundancy caused by equally probable permutation patterns, making it more sensitive to early, weak periodic impacts.

[0106] Through feature extraction from the above three dimensions, multidimensional feature samples with clear physical meaning and complementarity were constructed.

[0107] After acquiring multidimensional feature samples, they are standardized to eliminate dimensional and numerical scale differences between different features. For example, a zero-mean unit variance standardization method can be used to convert local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy into dimensionless standardized values. Subsequently, a real-time covariance matrix is ​​constructed within a sliding time window.

[0108] Specifically, a sliding time window containing several continuous sampling points is defined. Multiple multidimensional feature samples within the window are arranged in matrix form, and their covariance matrix is ​​calculated as a mathematical representation of the real-time feature distribution. Since the covariance matrix contains not only the variance information of each feature but also the correlation information between features, it can more comprehensively characterize the shape of the feature distribution.

[0109] Similarly, the characteristic distribution baseline is also represented as a normal state covariance matrix, which is pre-built and stored for each operating condition interval using the operating condition parameters as an index during the initial stage of normal equipment operation.

[0110] Based on the real-time covariance matrix and the normal-state covariance matrix, the manifold distance is determined using the following formula:

[0111] in, For manifold distance, The normal state covariance matrix, For the real-time covariance matrix, It is the inverse of the normal state covariance matrix. Represents the determinant operation of a matrix. Represents the trace operation of a matrix. Represents the natural logarithm operation. This refers to the feature dimension parameter, i.e., the number of dimensions of the multidimensional feature samples. For example, when the multidimensional feature samples include local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy, The value is 3.

[0112] By calculating the logarithm of the determinant and the trace of the product of the inverse matrix, the non-Euclidean geometric properties of the covariance matrix are fully considered, avoiding the problem of destroying the correlation structure between features when forcibly straightening the matrix into a vector for Euclidean distance calculation.

[0113] To more clearly illustrate the rationality of the covariance matrix construction and the rigor of the manifold distance, a comparative example is introduced. If, following conventional methods, the local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy extracted at a specific moment are directly used to construct a feature vector, and the Euclidean or Mahalanobis distance between this feature vector and the baseline mean vector is calculated, this approach only considers the absolute offset of feature points at a single moment, ignoring the statistical characteristics of feature fluctuations over time. Under strong background noise interference, the feature values ​​at a single moment may experience drastic jumps due to the superposition of random noise, leading to unstable distance measurement results and a high likelihood of false alarms.

[0114] This embodiment constructs a real-time covariance matrix within a sliding time window, using the second-order statistical properties of the feature distribution as a metric, effectively smoothing the transient effects of random noise. Simultaneously, the manifold distance formula, through determinant and inverse matrix operations on the covariance matrix, accurately captures minute changes in the feature correlation structure. When early, minor wear occurs in the bearing, even without a significant shift in the feature mean, the correlation structure between features changes. The manifold distance can sensitively capture this change and produce a significant response, thus providing rigorous and stable mathematical support for subsequent triggering mechanisms.

[0115] It should be understood that the specific composition of the aforementioned multidimensional feature samples is interpretive, not restrictive. For example, multidimensional feature samples may also include other sensitive features such as root square amplitude, waveform factor, or envelope spectrum barycenter frequency, and feature dimension parameters. The system then adaptively adjusts. Any application that satisfies the logic of quantifying the feature distribution shift by constructing a covariance matrix and calculating it based on the manifold distance formula should be covered within the scope of protection of this application.

[0116] Furthermore, the preset triggering conditions include the number of times the distribution distance continuously exceeds the dynamic threshold, which is determined based on the dispersion of the normal feature distribution corresponding to the current operating condition parameters. The remaining useful life prediction process includes: acquiring a historical multidimensional feature sample sequence and inputting it into a time-series prediction model to obtain the remaining useful life; generating maintenance prompts based on the remaining useful life and updating the feature distribution baseline based on the multidimensional feature samples under the current normal state; the time window length of the time-series prediction model is determined based on the length of the historical multidimensional feature sample sequence and the prediction error.

[0117] The dynamic threshold is not a fixed constant, but a variable that adaptively adjusts according to the dispersion of the normal characteristic distribution corresponding to the current operating conditions. Because the dispersion of the normal vibration characteristics of a bearing naturally varies under different speeds and loads—for example, the fluctuation range of the vibration signal is usually greater under high load conditions than under no-load conditions—using a fixed threshold can easily lead to false alarms under conditions with large fluctuations, while potentially causing missed alarms under conditions with small fluctuations. This embodiment determines the dynamic threshold based on the dispersion of the normal characteristic distribution. For example, the dynamic threshold can be set to a multiple of the trace of the baseline covariance matrix of the normal characteristic distribution under the current operating conditions, allowing the threshold to adaptively match the fluctuation characteristics of the current operating conditions.

[0118] The preset trigger condition is set to the number of times the distribution distance continuously exceeds the dynamic threshold, such as exceeding the dynamic threshold three times consecutively. This continuous over-limit judgment mechanism can effectively filter transient anomalies caused by sudden external interference, ensuring that only continuous deterministic degradation will trigger subsequent processes, greatly reducing the false alarm rate.

[0119] When the preset triggering conditions are met, the system acquires the historical multidimensional feature sample sequence and inputs it into the time series prediction model to obtain the remaining service life.

[0120] The historical multidimensional feature sample sequence refers to the set of multidimensional feature samples arranged in chronological order from the start of bearing operation to the current triggering time.

[0121] Temporal prediction models can be long short-term memory networks, gated recurrent unit networks, or Transformer models, etc. These models learn the degradation trends of features over time in a sequence of historical multidimensional feature samples to output the remaining time required for a bearing to progress from its current state to failure.

[0122] The time window length of a time series prediction model is determined based on the length of the historical multidimensional feature sample sequence and the prediction error. Specifically, the time window length determines the span of historical data input to the model each time. If the historical sequence is short, an excessively long window will lead to insufficient data; if the prediction error is large, appropriately shortening the window can allow the model to focus more on recent degradation trends and improve prediction accuracy. By dynamically adjusting the time window length, the model can maintain optimal prediction performance at different degradation stages.

[0123] Maintenance prompts are generated based on the remaining service life, and the feature distribution baseline is updated based on multi-dimensional feature samples under the current normal state. Maintenance prompts can include specific values ​​for the remaining service life and suggested maintenance time points, allowing maintenance personnel to plan repairs in advance. After generating maintenance prompts, the system does not stop working; instead, it updates the feature distribution baseline based on multi-dimensional feature samples under the current normal state. The mechanism of this design is that when a bearing experiences slight degradation but has not yet reached the replacement standard, its feature distribution has already drifted to some extent. If the initial baseline is continued for evaluation, the increased deviation between the baseline and the current actual state will cause the subsequent distribution distance to remain high, affecting the accuracy of monitoring. By using the currently relatively stable state features as the new baseline for updating, the system can continuously track state changes during the degradation evolution process, achieving closed-loop maintenance logic.

[0124] The baseline update mechanism enables the system to adapt to the slow degradation process of bearing performance, ensuring accurate condition assessment throughout the entire life cycle.

[0125] To more clearly illustrate the adaptive advantages of dynamic thresholds, a comparative example is introduced. Suppose that under a certain operating condition, the bearing is under high speed and heavy load, resulting in a large degree of dispersion in the normal characteristic distribution. If a fixed threshold scheme is used, this fixed threshold may be frequently exceeded, leading to continuous false alarms and degraded indications from the system. However, with the dynamic threshold scheme of this embodiment, the threshold is automatically adjusted upwards based on the high dispersion, triggering only when the distribution distance truly exceeds the normal fluctuation range under this operating condition, effectively avoiding false alarms. Conversely, under low speed and light load conditions, the dynamic threshold is automatically adjusted downwards, enabling it to sensitively detect minute abnormal deviations and avoid missed alarms.

[0126] By combining dynamic thresholds and continuous triggering mechanisms, and using a closed-loop design of time-series prediction models and baseline updates, high-reliability degradation assessment and lifetime prediction under complex operating conditions are achieved.

[0127] It should be understood that the specific values ​​of the preset number of times and the specific type of time series prediction model mentioned above are interpretive only and not restrictive. For example, the preset number of times can also be set to five or more, and the time series prediction model can also be a traditional time series model such as an autoregressive integral moving average model. As long as the logic of continuously triggering based on a dynamic threshold and inputting the time series model to predict lifetime to update the baseline is met, it should be covered within the scope of protection of this application.

[0128] To more clearly illustrate the technical effects of the bearing wear condition monitoring method and system based on vibration signals provided in this application, the following is a detailed description of the complete deployment and operation process, taking the predictive maintenance scenario of a double-row self-aligning roller bearing on the main shaft of a certain type of direct-drive wind turbine as an example.

[0129] It should be understood that this application scenario is illustrative only and not restrictive. The monitoring method and system of this application are also applicable to monitoring scenarios of bearings in other rotating equipment such as machine tool spindles, wind turbine gearbox bearings, and high-speed rail axle box bearings.

[0130] See Figure 2 In this application scenario, the wind turbine is located in a remote wind farm. The main shaft bearing is subjected to varying speed and axial load conditions that change with wind speed, and there is strong background interference from gearbox meshing noise and blade aerodynamic noise within the nacelle. The sensing module is installed on the outer wall of the main shaft bearing housing to simultaneously collect vibration acceleration signals and operating parameters such as speed and power. Power, as a proxy parameter for the load, is used to characterize the current axial load state of the main shaft bearing. The sensing module sends the collected multi-source data to the edge computing node via a message queue telemetry transmission protocol.

[0131] During operation, as the wind speed decreases from the rated rotational speed to the low wind speed range, the spindle speed and generator power also decrease. Edge computing nodes acquire vibration signals and operating parameters of the rotating equipment where the bearing is located. Candidate frequency bands are determined based on the mapping relationship between the operating parameters and the bearing's natural frequency, and the target resonant frequency band is determined based on the impact energy index within the candidate frequency bands.

[0132] Due to drastic fluctuations in wind speed, the rotational speed and load change significantly, causing dynamic drift in the equivalent contact stiffness of the bearing, which in turn leads to a shift in the bearing's natural frequency. Edge computing nodes determine the bearing's equivalent contact stiffness based on the rotational speed and load, and then determine the bearing's natural frequency based on the equivalent contact stiffness; the frequency range formed by the offset between the bearing's natural frequency and a preset frequency band is defined as a candidate frequency band.

[0133] By inputting rotational speed and load into a preset stiffness mapping relationship, the equivalent contact stiffness is obtained. The spectral kurtosis of each sub-band within the candidate frequency band is calculated, and the sub-band that meets the preset impact energy condition and its adjacent frequency bands are determined as the target resonant frequency band. For example, at rated wind speed, the candidate frequency band may be located in a higher frequency band, while when the wind speed drops to a low wind speed range, the candidate frequency band is dynamically adjusted to a lower frequency band. If the fixed frequency band threshold scheme in the prior art is used, when the above-mentioned drastic changes occur in the operating conditions, the fixed frequency band will not be able to cover the drifted true resonant frequency band, resulting in the impact energy being filtered out and weak fault features being completely lost. However, this application dynamically determines the candidate frequency band through a physical mapping relationship, so that the search range can always adaptively drift with changes in operating conditions, effectively solving the problem of feature extraction failure caused by resonant frequency band drift under changing operating conditions.

[0134] After locking onto the target resonant frequency band, the edge computing nodes perform multi-structural element morphological demodulation on the vibration signal within the target resonant frequency band to obtain the envelope signal, and construct multi-dimensional feature samples based on the envelope signal. Due to the strong interference background of gearbox meshing noise and blade aerodynamic noise within the nacelle, the signal-to-noise ratio of the vibration signal is extremely low. The multi-structural elements include a first structural element and a second structural element with different geometries. The first structural element is used to perform noise suppression processing on the vibration signal within the target resonant frequency band to obtain an intermediate signal; the second structural element is used to perform morphological preservation processing on the intermediate signal to obtain a morphologically preserved signal; morphological interpolation is performed on the morphologically preserved signal to obtain the interpolation result; and the envelope signal is determined based on the interpolation result.

[0135] Specifically, the first structural element is a flat structural element, and the second structural element is a parabolic structural element. The flat structural element is used to sequentially perform shape opening and shape closing operations on the vibration signal to obtain an intermediate signal; the parabolic structural element is then used to perform shape dilation and shape erosion operations on the intermediate signal to obtain a shape-preserving signal; the difference between the results of the shape dilation and shape erosion operations is calculated to obtain the difference operation result.

[0136] Flat structural elements can effectively suppress broadband pulse noise generated by gear meshing, while parabolic structural elements can accurately match and preserve the true geometry of weak wear impacts in bearings. Through this division of labor and cooperation among multiple structural elements, the low-frequency envelope signal of weak impacts can still be extracted with high fidelity under strong background noise, significantly improving the sensitivity of early fault characteristics.

[0137] Subsequently, the edge computing node generates a real-time feature distribution based on multi-dimensional feature samples, calculates the distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index, and generates bearing wear condition monitoring results based on the distribution distance.

[0138] The multidimensional feature samples include local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy. The multidimensional feature samples are standardized, and a real-time covariance matrix is ​​constructed within a sliding time window. A normal-state covariance matrix matching the operating parameters is obtained. Based on the logarithm of the determinant of the normal-state covariance matrix, the trace of the product of the inverse of the normal-state covariance matrix and the real-time covariance matrix, the logarithm of the determinant of the real-time covariance matrix, and the feature dimension parameters, the manifold distance is determined.

[0139] The distribution distance is the manifold distance; the characteristic distribution baseline is the normal-state covariance matrix based on the operating condition parameter index. Using wind speed ranges and power levels as operating condition indices, characteristic distribution baselines are constructed for each operating condition range under normal conditions. When gusts cause rapid switching of operating conditions, the characteristic distribution will shift. Because the manifold distance fully considers the nonlinear geometric structure of the Riemannian manifold space where the covariance matrix resides, and the characteristic distribution baseline is adaptively matched with the operating conditions, the manifold distance assessment can distinguish between characteristic shifts caused by changes in operating conditions and characteristic shifts caused by actual degradation, significantly reducing the false alarm rate under varying operating conditions.

[0140] The distribution distance is converted into a health index through a monotonic mapping function; in response to the distribution distance meeting the preset triggering conditions, a degradation indicator is output or the remaining lifespan prediction process is triggered.

[0141] Once deterministic degradation is confirmed, the edge computing node uploads the degradation feature sequence to the prediction platform. Preset triggering conditions include the number of times the distribution distance continuously exceeds a dynamic threshold. The dynamic threshold is determined based on the dispersion of the normal feature distribution corresponding to the current operating parameters.

[0142] The remaining useful life prediction process includes: acquiring historical multidimensional feature sample sequences and inputting them into a time series prediction model to obtain the remaining useful life; generating maintenance prompts based on the remaining useful life and updating the feature distribution baseline based on multidimensional feature samples under the current normal state; and determining the time window length of the time series prediction model based on the length of the historical multidimensional feature sample sequences and the prediction error.

[0143] The prediction platform responds to the distribution distance meeting the preset triggering conditions by receiving the degradation feature sequence uploaded by the edge computing node, determining the remaining lifetime based on the degradation feature sequence, and sending baseline update data to the edge computing node.

[0144] Edge computing nodes receive baseline update data from the prediction platform and update the feature distribution baseline. The prediction platform combines this data with the wind farm's operation and maintenance plan to generate maintenance recommendations. For example, when the predicted remaining service life is less than a preset number of days, it automatically issues a preventative maintenance work order to the wind farm management system.

[0145] The edge-cloud collaborative architecture enables edge nodes to perform only lightweight distance calculations under normal circumstances, and only activates the complex prediction process in the cloud when deterministic degradation is confirmed. This greatly saves edge computing power and communication bandwidth, and has important practical significance for the remote deployment of remote wind farms.

[0146] In summary, under the triple challenges of variable operating conditions, strong noise, and remote deployment, this application achieves dynamic and accurate tracking of the resonant frequency band, high-fidelity extraction of weak impact characteristics under extremely low signal-to-noise ratio, and wear state measurement that adapts to different operating conditions through the synergistic effect of physical mapping frequency band dynamic locking, multi-structural element morphology fidelity demodulation, and operating condition baseline manifold distance assessment.

[0147] Compared with the fixed frequency band threshold scheme, this application can adaptively track the resonant frequency band drift and avoid feature extraction failure; compared with the pure data-driven scheme, this application has clear physical and mathematical mechanism constraints, stronger generalization ability in unseen operating conditions, and significantly reduces false alarm rate and local computing load.

[0148] This application also provides an electronic device. The electronic device includes a processor and a memory, the memory storing a computer program, and the processor, when executing the computer program, implements the steps of the above-described bearing wear condition monitoring method based on vibration signals.

[0149] Electronic devices, as hardware carriers, can be edge computing terminals deployed near rotating equipment, industrial personal computers, programmable logic controllers, or status monitoring modules integrated within the equipment controller. The processor and memory are interconnected via an internal bus to exchange instructions and data. The processor executes computer programs stored in memory to coordinate the various hardware modules within the electronic device to perform functions such as data acquisition, signal processing, feature extraction, and status assessment.

[0150] In terms of hardware selection, the processor can be a microcontroller, digital signal processor, field-programmable gate array (FPGA), or application-specific integrated circuit (ASIC). For example, in space-constrained or power-sensitive edge deployment scenarios, the processor can be an STM32H743 microcontroller based on the ARM Cortex-M7 core, which has a hardware floating-point unit and can efficiently perform complex mathematical operations such as matrix inversion and manifold distance calculation.

[0151] The memory can include random access memory and read-only memory. For example, static random access memory can be used as a cache to temporarily store real-time acquired vibration signals and intermediate calculation variables; flash memory or electrically erasable programmable read-only memory can be used as non-volatile storage to persistently store computer programs, preset stiffness mapping relationships, feature distribution baseline libraries, and historical multidimensional feature sample sequences.

[0152] In terms of the internal data processing flow of the equipment, the processor connects to the sensor via a peripheral interface to acquire the vibration signal and operating parameters of the rotating equipment where the bearing is located. Subsequently, the processor calls computer program instructions from memory to determine candidate frequency bands based on the mapping relationship between the operating parameters and the bearing's natural frequency, and then determines the target resonant frequency band based on the impact energy index within the candidate frequency bands. Next, the processor performs multi-structural element morphological demodulation on the vibration signal within the target resonant frequency band to obtain the envelope signal, and constructs multi-dimensional feature samples based on the envelope signal. Finally, the processor generates a real-time feature distribution based on the multi-dimensional feature samples, calculates the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by the operating parameters, and generates bearing wear condition monitoring results based on the distribution distance. The entire data processing flow is completed locally on the electronic device, achieving real-time condition monitoring and degradation indication output without relying on a remote server, greatly improving the real-time performance and reliability of the monitoring system.

[0153] It should be understood that the specific model of the microcontroller and the specific type of memory mentioned above are merely illustrative and not restrictive. For example, in industrial personal computers with more abundant computing resources, the processor can also be a general-purpose central processing unit, such as the Intel Core series or AMD Ryzen series, and the memory can be dynamic random access memory and solid-state drives. Any hardware architecture capable of executing the above monitoring method steps should be covered within the scope of this application. By embedding the monitoring method into the hardware execution logic within the electronic device, the device can operate stably under strong electromagnetic interference and complex operating conditions, achieving closed-loop control of physical mechanism-constrained frequency band search, morphological fidelity demodulation, and distribution distance assessment.

[0154] like Figure 3 As shown, this application also provides a bearing wear condition monitoring system based on vibration signals. This system adopts an edge-cloud collaborative architecture, including a sensing module, edge computing nodes, and a prediction platform.

[0155] The sensing module is used to collect vibration signals and operating parameters of the rotating equipment where the bearing is located.

[0156] The sensing module, acting as the perception layer, is responsible for acquiring raw physical data. It includes vibration sensors and operating parameter sensors. Vibration sensors can be microelectromechanical systems (MEMS) accelerometers, piezoelectric accelerometers, or velocity sensors. For example, in space-constrained or power-sensitive deployment scenarios, a triaxial MEMS accelerometer with a noise density below 25 micrograms per ohm can be selected. This accelerometer is mounted on the outer wall of the rotating equipment end cover or bearing housing using flexible printed circuit board technology and is cured and encapsulated with thermally conductive epoxy resin to ensure rigid coupling between the sensor and the measured structure under strong vibration, avoiding attenuation of high-frequency signals due to installation resonance. Operating parameter sensors can include magnetoelectric encoders, Hall effect speed sensors, torque sensors, or power transmitters, used to synchronously acquire speed, load, or motor power as a load proxy parameter. The sensing module amplifies and anti-aliasing the acquired analog signals through signal conditioning circuitry, converts them into digital signals through an analog-to-digital converter, and then sends them to the edge computing node.

[0157] The edge computing node is configured to determine candidate frequency bands based on the mapping relationship between operating condition parameters and the bearing's natural frequency, determine the target resonance frequency band based on the impact energy index within the candidate frequency band, perform multi-structural element morphological demodulation on the vibration signal within the target resonance frequency band to obtain the envelope signal, construct multi-dimensional feature samples based on the envelope signal, generate a real-time feature distribution, and calculate the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by operating condition parameters to generate bearing wear condition monitoring results; when the distribution distance meets the preset triggering conditions, upload the degradation feature sequence to the prediction platform; and receive baseline update data sent by the prediction platform and update the feature distribution baseline.

[0158] Edge computing nodes, acting as the computing layer, are responsible for executing lightweight real-time signal processing and state assessment logic. These nodes are based on microcontrollers or digital signal processors. For example, an edge computing node can use a microcontroller based on the ARM Cortex-M7 core, with a built-in hardware floating-point unit and a high-precision analog-to-digital converter, enabling efficient execution of complex mathematical operations such as matrix inversion and manifold distance calculation. After receiving multi-source data from the sensing module, the edge computing node invokes internally stored computer program instructions to sequentially execute band locking, morphological demodulation, feature construction, and distribution distance calculation. When the distribution distance meets preset trigger conditions, the edge computing node uploads the degradation feature sequence to the prediction platform via a communication protocol. This communication protocol can be a lightweight IoT protocol such as a message queue telemetry transmission protocol, a restricted application protocol, or an advanced message queue protocol. For example, the edge computing node encapsulates the degradation feature sequence into a data packet of a preset format and publishes it to a specific topic via a message queue telemetry transmission protocol for the prediction platform to subscribe to. Simultaneously, the edge computing node listens for and receives baseline update data sent by the prediction platform, replacing the locally stored feature distribution baseline with the new baseline data to achieve closed-loop maintenance logic. This edge computing architecture allows the system to perform only lightweight distance calculations under normal circumstances, and only wakes up cloud communication when deterministic degradation is confirmed, which greatly saves edge computing power and communication bandwidth.

[0159] The prediction platform is configured to receive degradation feature sequences uploaded by edge computing nodes in response to a preset trigger condition that the distribution distance meets the requirement, determine the remaining lifetime based on the degradation feature sequences, and send baseline update data to the edge computing nodes.

[0160] As a cloud service layer, the prediction platform is responsible for performing computationally demanding time-series prediction and global management functions. The prediction platform can be deployed on a graphics processing unit (GPU) server cluster or a cloud computing center. Internally, the prediction platform includes a time-series database, a model training engine, and a prediction network model library. When the prediction platform receives a degradation feature sequence uploaded by an edge computing node, it inputs the sequence into the time-series prediction model. By learning the degradation trend of features over time in historical multi-dimensional feature sample sequences, it outputs the remaining service life required for the bearing to progress from its current state to failure. Subsequently, the prediction platform generates maintenance prompts based on the prediction results and generates baseline update data by combining multi-dimensional feature samples under the current normal state. The prediction platform sends the baseline update data to the edge computing nodes via a communication protocol to guide the edge computing nodes in updating the baseline. For example, when the prediction platform determines that the remaining service life is less than a preset number of days, it automatically issues a preventative maintenance work order to the equipment management system. Through this edge-cloud collaborative architecture, high-frequency signal processing and low-frequency life prediction are physically decoupled, ensuring real-time status monitoring while utilizing the powerful computing power of the cloud for complex life prediction, thus optimizing the overall system performance.

[0161] It should be understood that the aforementioned hardware models, communication protocols, and deployment details are illustrative only and not restrictive. For example, the sensing module can also be directly connected to the edge computing node via an industrial Ethernet or controller area network bus; the edge computing node can also be a programmable logic controller or an edge gateway device; the prediction platform can also be an industrial personal computer deployed within a local area network. Any data flow and functional configuration logic that satisfies edge-cloud collaboration should be covered within the scope of this application.

[0162] Unlike existing technologies, this application addresses the problems of feature extraction failure due to resonance band drift under varying operating conditions, loss of weak features at extremely low signal-to-noise ratios, and poor generalization ability in state assessment. It determines the equivalent contact stiffness of the bearing based on real-time rotational speed and load, then calculates the natural frequency to dynamically delineate candidate frequency bands, and finally precisely locks the target resonance band by combining impact energy indicators such as spectral kurtosis. By utilizing physical mechanisms to constrain the search range, the filter band can adaptively follow the operating condition drift, solving the problem of fixed-band filtering failure and improving the accuracy of weak fault feature extraction. This application abandons traditional single filtering, employing first and second structural elements with different geometries to sequentially perform noise suppression and shape preservation processing. The flat structural element effectively filters out broadband impulse noise, while the parabolic structural element accurately matches and preserves the true geometric shape of weak impacts. The two work together to decouple the functions of noise suppression and shape preservation, significantly improving the sensitivity of early fault features at extremely low signal-to-noise ratios. By extracting local morphological kurtosis, harmonic relative energy ratio, and improved permutation entropy to construct multidimensional features, the Riemannian distance between the real-time feature distribution and the baseline adaptively matched to different operating conditions is calculated. This transforms black-box judgment into a metric with a clear mathematical mechanism, significantly reducing the false alarm rate under varying operating conditions. Simultaneously, cloud-based lifetime prediction is triggered only when the distribution distance continuously exceeds limits, and the local baseline is dynamically updated. This ensures early diagnostic accuracy under extremely low signal-to-noise ratios while greatly saving edge computing power and communication bandwidth through edge-cloud collaboration, achieving seamless long-term embedding of the monitoring system and highly reliable prediction.

[0163] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for monitoring bearing wear condition based on vibration signals, characterized in that, include: Acquire vibration signals and operating parameters of the rotating equipment where the bearing is located; Candidate frequency bands are determined based on the mapping relationship between the operating condition parameters and the bearing's natural frequency, and the target resonance frequency band is determined based on the impact energy index within the candidate frequency bands. Multi-structural element morphological demodulation is performed on the vibration signal within the target resonance frequency band to obtain the envelope signal, and multi-dimensional feature samples are constructed based on the envelope signal; A real-time feature distribution is generated based on the multi-dimensional feature samples. The distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index is calculated. The bearing wear condition monitoring result is generated based on the distribution distance.

2. The bearing wear condition monitoring method based on vibration signals according to claim 1, characterized in that, The operating parameters include speed and load; The determination of candidate frequency bands based on the mapping relationship between the operating condition parameters and the bearing's natural frequency includes: The equivalent contact stiffness of the bearing is determined based on the rotational speed and the load, and the natural frequency of the bearing is determined based on the equivalent contact stiffness. The frequency range formed by the bearing's natural frequency and the preset frequency band offset is determined as the candidate frequency band.

3. The bearing wear condition monitoring method based on vibration signals according to claim 2, characterized in that, Determining the equivalent contact stiffness of the bearing based on the rotational speed and the load includes: The equivalent contact stiffness is obtained by mapping the rotational speed and the load input to a preset stiffness relationship. The determination of the target resonance frequency band based on the impact energy index within the candidate frequency band includes: Calculate the spectral kurtosis of each sub-band within the candidate frequency band, and determine the sub-band that satisfies the preset impact energy condition and its adjacent frequency bands as the target resonance frequency band.

4. The bearing wear condition monitoring method based on vibration signals according to claim 1, characterized in that, The multi-structural elements include first and second structural elements with different geometric shapes; The process of performing multi-structural element morphological demodulation on the vibration signal within the target resonant frequency band to obtain the envelope signal includes: The first structural element is used to perform noise suppression processing on the vibration signal within the target resonant frequency band to obtain an intermediate signal; The intermediate signal is subjected to shape preservation processing using the second structuring element to obtain a shape-preserving signal; Perform morphological interpolation on the morphology-preserving signal to obtain the interpolation result; The envelope signal is determined based on the result of the difference operation.

5. The bearing wear condition monitoring method based on vibration signals according to claim 4, characterized in that, The first structural element is a flat structural element, and the second structural element is a parabolic structural element; The step of using the first structural element to perform noise suppression processing on the vibration signal within the target resonant frequency band to obtain an intermediate signal includes: The vibration signal is subjected to shape opening and shape closing operations sequentially using the flat structural element to obtain the intermediate signal; The step of performing shape-preserving processing on the intermediate signal using the second structuring element to obtain a shape-preserving signal includes: The intermediate signal is subjected to morphological dilation and morphological erosion operations using the parabolic structuring element to obtain the morphologically preserved signal. The step of performing morphological interpolation on the morphology-preserving signal to obtain the interpolation result includes: The difference between the result of the morphological dilation operation and the result of the morphological erosion operation is calculated to obtain the difference operation result.

6. The bearing wear condition monitoring method based on vibration signals according to claim 1, characterized in that, The distribution distance is the manifold distance; the characteristic distribution baseline is the normal state covariance matrix based on the operating condition parameter index; The generation of bearing wear condition monitoring results based on the distribution distance includes: The distribution distance is converted into a health index using a monotonic mapping function; In response to the distribution distance meeting the preset triggering conditions, a degradation indication is output or the remaining lifetime prediction process is triggered.

7. The bearing wear condition monitoring method based on vibration signals according to claim 6, characterized in that, The multidimensional feature samples include local morphological kurtosis, relative harmonic energy ratio, and improved permutation entropy; The calculation of the distribution distance between the real-time feature distribution and the feature distribution baseline based on the operating condition parameter index includes: The multidimensional feature samples are standardized, and a real-time covariance matrix is ​​constructed within a sliding time window; Obtain the normal state covariance matrix that matches the operating condition parameters; The manifold distance is determined based on the logarithm of the determinant of the normal state covariance matrix, the trace of the product of the inverse of the normal state covariance matrix and the real-time covariance matrix, the logarithm of the determinant of the real-time covariance matrix, and the feature dimension parameter.

8. The bearing wear condition monitoring method based on vibration signals according to claim 6, characterized in that, The preset triggering condition includes the number of times the distribution distance continuously exceeds the dynamic threshold, and the dynamic threshold is determined based on the dispersion of the normal characteristic distribution corresponding to the current operating condition parameters. The process of triggering remaining lifetime prediction includes: Obtain a historical multidimensional feature sample sequence and input the historical multidimensional feature sample sequence into a time series prediction model to obtain the remaining lifetime; A maintenance prompt is generated based on the remaining service life, and the feature distribution baseline is updated based on the multidimensional feature samples under the current normal state; The time window length of the time series prediction model is determined based on the length of the historical multidimensional feature sample sequence and the prediction error.

9. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the bearing wear condition monitoring method based on vibration signals as described in any one of claims 1-8.

10. A bearing wear condition monitoring system based on vibration signals, characterized in that, include: The sensing module is used to collect vibration signals and operating parameters of the rotating equipment where the bearing is located; An edge computing node is configured to determine candidate frequency bands based on the mapping relationship between the operating condition parameters and the bearing's natural frequency, determine a target resonance frequency band based on the impact energy index within the candidate frequency bands, perform multi-structural element morphological demodulation on the vibration signal within the target resonance frequency band to obtain an envelope signal, construct multi-dimensional feature samples based on the envelope signal, generate a real-time feature distribution, and calculate the distribution distance between the real-time feature distribution and the feature distribution baseline indexed by the operating condition parameters to generate bearing wear condition monitoring results; when the distribution distance meets a preset trigger condition, upload a degradation feature sequence to the prediction platform; and receive baseline update data sent by the prediction platform and update the feature distribution baseline. The prediction platform is configured to receive the degradation feature sequence uploaded by the edge computing node in response to the distribution distance meeting a preset trigger condition, determine the remaining lifetime based on the degradation feature sequence, and send baseline update data to the edge computing node.