Machine Learning-Based Method and System for Motor Bearing Fault Diagnosis
By using machine learning-based methods, a motor bearing fault diagnosis system was constructed, which solved the problem of the motor bearing fault diagnosis process being affected by changes in equipment operating status and environmental disturbances. It achieved accurate fault identification and assessment under complex working conditions and improved the reliability of fault diagnosis.
Patent Information
- Application Number
- CN202611197332.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-17
AI Technical Summary
In the existing technology, the fault diagnosis process of motor bearings is affected by changes in equipment operating status and environmental disturbances, resulting in insufficient reliability of fault diagnosis, difficulty in accurately distinguishing fault types and severity, and easy misdiagnosis or missed diagnosis.
By using machine learning-based methods, vibration and current signals during motor operation are read, a working condition driving mapping relationship is constructed, the signals are mapped to a unified working condition reference domain, a decoupled standard response signal is established, multi-dimensional evaluation indicators are calculated to construct a frequency band-impact coupling evaluation matrix, frequency band screening and boundary adaptive reconstruction are performed, a dynamic resonance frequency band is constructed, impact structure constraint decomposition is performed, and impact evolution characteristics are extracted for fault type identification and level assessment.
It improves the reliability of motor bearing fault diagnosis, can accurately identify fault types and assess fault levels under complex operating conditions, and reduces the impact of equipment status changes and environmental disturbances.
Smart Images

Figure CN123116191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a machine learning-based method and system for diagnosing motor bearing faults. Background Technology
[0002] As a key power component in industrial equipment, the reliability of electric motors directly affects the stable operation of production systems. Bearings, in particular, are subject to rotating loads, mechanical shocks, and complex working environments over long periods, making them prone to wear, fatigue spalling, and lubrication abnormalities. Current technologies typically assess bearing operating conditions by collecting vibration and current signals from the motor during operation and combining this data with signal analysis methods. However, in actual operation, changes in motor speed, load, and external disturbances cause the monitoring signals to exhibit strong non-stationary characteristics, making fault characteristics easily affected by normal operating conditions. This makes it difficult to effectively detect subtle fault impacts. Furthermore, different fault types exhibit similar signal characteristics during their evolution, making it difficult for existing diagnostic methods to accurately distinguish fault categories and severity, leading to misdiagnosis or missed diagnosis. Summary of the Invention
[0003] This application provides a machine learning-based method and system for diagnosing motor bearing faults, which solves the technical problem that the reliability of fault diagnosis is insufficient in the existing technology due to changes in equipment operating status and environmental disturbances.
[0004] The first aspect of this application provides a machine learning-based method for diagnosing motor bearing faults, the method comprising: The original monitoring signals, including vibration and current signals during motor operation, are read. After constructing a working condition driving mapping relationship with corresponding operating condition parameters, the original monitoring signals are mapped to a unified working condition reference domain to establish a decoupled standard response signal. Multidimensional evaluation indicators are calculated based on the standard response signal, and a frequency band-impact coupling evaluation matrix is constructed using these indicators. The multidimensional evaluation indicators include impact response consistency indicators, energy accumulation indicators, and modulation synchronization indicators. Frequency band selection and adaptive reconstruction of frequency band boundaries are performed based on the frequency band-impact coupling evaluation matrix to construct a dynamic resonant frequency band. Impact structural constraint decomposition is performed on the signals within the dynamic resonant frequency band to construct an impact response sequence. Impact evolution characteristics are extracted from the impact response sequence, and fault type identification and fault level assessment are performed to generate bearing fault diagnosis results.
[0005] A second aspect of this application provides a machine learning-based motor bearing fault diagnosis system, the system comprising: Signal Condition Decoupling Module: Reads the original monitoring signals, including vibration and current signals during motor operation. After constructing a condition-driven mapping relationship with the corresponding operating parameters, the original monitoring signals are mapped to a unified operating condition reference domain to establish a decoupled standard response signal. Evaluation Matrix Construction Module: Calculates multi-dimensional evaluation indicators based on the standard response signal and constructs a frequency band-impact coupling evaluation matrix using these indicators. The multi-dimensional evaluation indicators include impact response consistency indicators, energy accumulation indicators, and modulation synchronization indicators. Resonance Frequency Band Construction Module: Performs frequency band screening and adaptive reconstruction of frequency band boundaries based on the frequency band-impact coupling evaluation matrix to construct a dynamic resonance frequency band. Response Sequence Construction Module: Performs impact structural constraint decomposition on the signals within the dynamic resonance band to construct an impact response sequence. Bearing Fault Diagnosis Module: Extracts impact evolution features from the impact response sequence, performs fault type identification and fault level assessment, and generates bearing fault diagnosis results.
[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, the original monitoring signals, including vibration and current signals during motor operation, are read. After constructing a working condition-driven mapping relationship using corresponding operating condition parameters, the original monitoring signals are mapped to a unified working condition reference domain, establishing a decoupled standard response signal. Next, multi-dimensional evaluation indicators are calculated based on the standard response signal, and a frequency band-impact coupling evaluation matrix is constructed using these indicators. These multi-dimensional evaluation indicators include impact response consistency, energy accumulation, and modulation synchronization. Further, frequency band selection and adaptive reconstruction of frequency band boundaries are performed based on the frequency band-impact coupling evaluation matrix to construct a dynamic resonant frequency band. Then, impact structural constraint decomposition is performed on the signals within the dynamic resonant band to construct an impact response sequence. Finally, impact evolution features are extracted from the impact response sequence, and fault type identification and fault level assessment are performed to generate bearing fault diagnosis results. This solves the technical problem of insufficient reliability in motor bearing fault diagnosis caused by changes in equipment operating status and environmental disturbances in the existing technology, achieving the technical effect of improving the reliability of motor bearing fault diagnosis. Attached Figure Description
[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0008] Figure 1 A schematic diagram of the machine learning-based motor bearing fault diagnosis method provided in this application embodiment; Figure 2 This application provides a comparison diagram of signal response before and after condition mapping in an embodiment of the application. Figure 3 A schematic diagram of the structure of a machine learning-based motor bearing fault diagnosis system provided in this application embodiment.
[0009] Figure labeling: Signal condition decoupling module 11, Evaluation matrix construction module 12, Resonance frequency band construction module 13, Response sequence construction module 14, Bearing fault diagnosis module 15. Detailed Implementation
[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0011] Example 1, as Figure 1 As shown, this application provides a machine learning-based method for diagnosing motor bearing faults, wherein the method includes: The original monitoring signals, including vibration signals and current signals during motor operation, are read. After constructing the working condition drive mapping relationship with the corresponding operating condition parameters, the original monitoring signals are mapped to a unified working condition reference domain to establish a decoupled standard response signal.
[0012] In this embodiment, the original monitoring signals during motor operation are first synchronously acquired. These original monitoring signals include vibration acceleration signals acquired by vibration sensors installed at the motor bearing housing and stator current signals acquired by a current acquisition unit. Simultaneously, operating condition parameters corresponding to the original monitoring signals are acquired through the motor control system. These operating condition parameters include parameters characterizing the actual operating state of the motor, such as motor speed, load torque, effective current value, operating temperature, and operating time. The acquired original monitoring signals and operating condition parameters are time-synchronized, and a one-to-one correspondence between monitoring signal samples and operating condition parameter samples is established according to a preset sampling period, forming a joint sample dataset containing signal response characteristics and operating condition change characteristics. Based on this joint sample dataset, feature analysis is performed on the variation patterns of monitoring signals under different operating conditions. The correlation between vibration amplitude changes, spectral energy distribution, current fluctuation characteristics, and operating condition parameter changes is extracted, and a condition-driven mapping relationship between operating condition parameters and signal response characteristics is constructed. This condition-driven mapping relationship describes the variation patterns of monitoring signal response under different speeds, loads, and environmental conditions. After the mapping relationship is constructed, the real-time acquired raw monitoring signals are input into the working condition driving mapping relationship. The corresponding working condition influence compensation amount is calculated according to the current operating condition parameters. Based on the working condition influence compensation amount, the working condition related change components in the raw monitoring signals are corrected so that the vibration signals and current signals collected under different operating conditions are characterized under a unified reference working condition. Furthermore, amplitude normalization, time scale correction and residual deviation elimination are performed on the corrected signals so that the processed signals retain the bearing fault response characteristics while reducing the influence of non-fault response caused by factors such as speed change and load change. Finally, a decoupled standard response signal corresponding to the unified working condition reference domain is obtained.
[0013] Furthermore, mapping the raw monitoring signals to a unified operating condition reference domain includes: Based on the operating condition parameters, an operating condition feature expression is constructed; based on the operating condition feature expression, a mapping function relationship between the original monitoring signal and the target operating condition reference domain is established; the original monitoring signal is transformed using the mapping function relationship so that the signals under different operating conditions are aligned with the unified operating condition reference domain in terms of time scale and amplitude response; consistency correction processing is performed on the mapped signal to eliminate the residual coupling effect caused by operating condition changes and construct a decoupled standard response signal.
[0014] Preferably, when mapping the original monitoring signal to a unified operating condition reference domain, the operating condition parameters are first subjected to feature encoding processing. Based on the changes in parameters such as motor speed, load torque, effective current value, operating temperature, and operating time during motor operation, operating condition feature quantities that can characterize the changes in motor operating state are extracted and combined in a preset order to construct an operating condition feature expression vector. ,in, Represents the vector expressing the characteristics of the working conditions; Indicates the motor speed; Indicates load torque; Indicates the effective value of the current; Indicates the operating temperature; Indicates the runtime.
[0015] Based on the aforementioned operating condition feature expression vector, an operating condition mapping sample set is established using monitoring signal samples collected historically under different operating conditions. Correlation analysis is performed between the various operating condition feature expressions and their corresponding vibration signal and current signal characteristics to determine the influence of changes in operating conditions on the amplitude, frequency distribution, and time response characteristics of the monitoring signals. Furthermore, a regression mapping function is used to establish a mapping relationship between the original monitoring signals and the target operating condition reference domain. This mapping function is expressed as follows: ,in, This represents the response signal after mapping to the target operating condition reference domain; Indicates the original monitoring signal; Represents the vector expressing the characteristics of the working conditions; This represents a signal mapping function established based on operating condition characteristics.
[0016] During real-time diagnostics, the currently acquired raw monitoring signals and corresponding operating condition parameters are input into the mapping function. The signal transformation relationship is determined based on the current operating condition parameters, and the raw monitoring signals are corrected for time scale and compensated for amplitude response. Specifically, for the change in fault impact cycle caused by speed variation, the signal time axis is adjusted according to the ratio between the current speed and the target reference speed, so that impact events under different speed conditions correspond to each other on a unified time scale. For the change in vibration amplitude caused by load variation, the signal amplitude is normalized and corrected according to the load influence factor, so that the fault response amplitude under different load conditions is comparable, thereby obtaining a mapped signal aligned to a unified operating condition reference domain.
[0017] Furthermore, the signal after the working condition mapping is completed is subjected to consistency correction processing. The residual deviation between the mapped signal and the standard response of the target reference domain is calculated, and the signal is compensated for a second time based on the residual deviation. The consistency correction includes time offset correction, amplitude deviation correction and spectral energy distribution correction. By eliminating the residual coupling effect caused by the change in working condition, the corrected signal retains only the response characteristics related to the bearing operating state, and finally a decoupled standard response signal is established.
[0018] like Figure 2 The figure shows a comparison of signal responses before and after the operating condition mapping. In this embodiment, vibration monitoring signals collected by the motor under different operating conditions are selected as the analysis objects, where different operating conditions correspond to different speeds and load states. As can be seen from the original monitoring signals before mapping, due to the influence of changes in operating conditions, there are certain differences in the signal amplitude distribution, waveform period, and response scale under each operating condition, making it difficult to directly compare and analyze the fault response characteristics under different operating conditions. After establishing the operating condition drive mapping relationship based on the operating condition parameters, the original monitoring signals are transformed into an operating condition reference domain, so that the signal responses under different operating conditions are adjusted according to a unified reference scale to obtain standard response signals. As can be seen from the mapped signals, the amplitude response and period change trends of the signals under different operating conditions tend to be consistent, effectively reducing the influence of changes in operating conditions on the monitoring signal characteristics, and providing a stable data foundation for subsequent frequency band evaluation, impact feature extraction, and bearing fault type identification.
[0019] Multidimensional evaluation indicators are calculated based on standard response signals, and a frequency band-impact coupling evaluation matrix is constructed using these multidimensional evaluation indicators. The multidimensional evaluation indicators include impulse response consistency indicators, energy accumulation indicators, and modulation synchronization indicators.
[0020] Furthermore, multidimensional evaluation indicators are calculated based on standard response signals, including: Based on the standard response signal, the corresponding frequency band signal is extracted for each candidate frequency band; the extracted frequency band signal is envelope demodulated to establish an envelope signal; impact event monitoring of the envelope signal is performed to obtain the time sequence of impact events; the time interval between adjacent impact events is calculated using the time sequence, and the time interval is compared with the theoretical fault characteristic period determined by the operating conditions to establish an impact response consistency index; the impact response consistency index is output as a multi-dimensional evaluation index.
[0021] When calculating multidimensional evaluation indicators based on the standard response signal, a candidate frequency band set is first determined according to the sampling frequency of the standard response signal and the bearing fault characteristic frequency range. Specifically, the upper limit of the analysis frequency is determined based on the highest effective sampling frequency of the acquisition device, and the theoretical fault characteristic frequency ranges, such as the inner ring fault frequency, outer ring fault frequency, rolling element fault frequency, and cage fault frequency, are calculated in conjunction with the structural parameters of the target motor bearing and the current operating speed. Using the theoretical fault characteristic frequency range as the frequency band coverage area, the frequency band coverage area is divided into equal intervals according to a preset frequency resolution to form multiple continuous candidate frequency bands. Each candidate frequency band includes a start frequency and a cutoff frequency. For each candidate frequency band, bandpass filtering is performed on the standard response signal to extract the frequency band signal within the corresponding frequency range.
[0022] Furthermore, envelope demodulation processing is performed on the frequency band signals corresponding to each candidate frequency band. The analytic signal corresponding to the frequency band signal is obtained through Hilbert transform, and the amplitude envelope of the analytic signal is extracted to establish the envelope signal. , ,in, Indicates the first The analytical signal corresponding to each candidate frequency band; Indicates the first The frequency band signal corresponding to each candidate frequency band; Represents the Hilbert transform; Indicates the first The envelope signal corresponding to each candidate frequency band.
[0023] Impact event monitoring is performed based on the envelope signal. First, an adaptive detection threshold is calculated according to the statistical amplitude characteristics of the envelope signal. Envelope peaks that exceed the detection threshold and satisfy local extremum conditions are identified as impact events, and their corresponding time positions are recorded to obtain the impact event time sequence corresponding to the i-th candidate frequency band. ,in, This represents the time sequence of impact events corresponding to the i-th candidate frequency band; Indicates the i-th candidate frequency band. The moment of the shock event.
[0024] Furthermore, the time interval between adjacent impact events is calculated based on the impact event time sequence: ,in, Indicates the i-th candidate frequency band within the i-th frequency band. The time interval between adjacent impact events.
[0025] The corresponding bearing fault characteristic cycle is determined based on the current operating condition parameters. The fault characteristic cycle is calculated based on the number of bearing rolling elements, pitch diameter, rolling element diameter, contact angle, and current rotational speed. The time interval between adjacent impact events is matched with the theoretical fault characteristic cycle, and the cycle deviation between the two is calculated. ,in, This represents the period deviation corresponding to the m-th impulse interval within the i-th candidate frequency band; This represents the theoretical fault characteristic period determined under the current operating conditions.
[0026] Calculate the shock response consistency index based on the periodic deviation of all shock events: ,in, This represents the impulse response consistency index corresponding to the i-th candidate frequency band; This represents the number of effective impact event intervals. Finally, the impact response consistency index corresponding to each candidate frequency band is output as one of the multi-dimensional evaluation indicators, which is used to construct the frequency band-impact coupling evaluation matrix by combining the energy accumulation index and the modulation synchronization index.
[0027] Furthermore, calculating multidimensional evaluation indicators based on standard response signals also includes: Instantaneous energy calculation is performed on the extracted frequency band signal to obtain an energy sequence that changes with time; a time neighborhood is constructed based on the time sequence of the impact event, and the local energy in the time neighborhood is calculated using the energy sequence, while the background energy in the non-impact area is calculated simultaneously; an energy accumulation index is established based on the ratio of local energy to background energy; and the energy accumulation index is output as a multidimensional evaluation index.
[0028] Instantaneous energy calculations are performed on the frequency band signals corresponding to each candidate frequency band obtained above. The square of the frequency band signal amplitude is used as the energy representation quantity at the corresponding moment, resulting in an energy sequence that varies with time. ,in, This represents the instantaneous energy sequence corresponding to the frequency band signal; This represents the frequency band signal obtained by extracting candidate frequency bands.
[0029] Furthermore, based on the time sequence of impact events obtained during the calculation of the impact response consistency index, a corresponding time neighborhood window is established for each impact event. Taking the time of the impact event as the center, a preset time range is expanded forward and backward to form the impact response region. ,in, This represents the temporal neighborhood corresponding to the i-th impact event; Indicates the time when the i-th impact event occurs; Indicates the width of the preset time window.
[0030] Based on the aforementioned time neighborhood window, energy data within each impact response region is extracted from the energy sequence, and the energy within multiple impact response regions is accumulated to obtain the local impact energy. ,in, represents the local energy within the impact response region; m represents the number of impact events detected.
[0031] Meanwhile, to eliminate the influence of normal operating vibration energy on the evaluation results, a non-impact time region that does not contain impact events was selected as the background region in the energy sequence, and the energy within the background region was statistically analyzed to obtain the background energy: ,in, This represents the background energy in the non-impact region; This represents the time set corresponding to the non-impact region; This indicates the number of data points within the background area.
[0032] After obtaining the local energy and background energy, an energy accumulation index is established by calculating the ratio between the two: ,in, Indicates an energy accumulation index; This represents a preset small correction amount to prevent the denominator from being zero.
[0033] The energy concentration index is used to characterize the concentration of fault impact energy relative to background operating energy within the current candidate frequency band. When the candidate frequency band contains a significant fault impact response, the energy in the time region corresponding to the impact event is significantly higher than that in the non-impact region, and the energy concentration index increases. When the candidate frequency band mainly contains normal operation components or noise components, the difference between local energy and background energy is small, and the energy concentration index decreases. Finally, the calculated energy concentration index is used as one of the outputs of the multi-dimensional evaluation index, and together with the impact response consistency index and modulation synchronization index, it is used to construct the frequency band-impact coupling evaluation matrix.
[0034] Furthermore, calculating multidimensional evaluation indicators based on standard response signals also includes: After envelope demodulation of the extracted frequency band signal, spectrum analysis is performed to construct the envelope spectrum; modulation sideband components centered on the frequency shift and its harmonics are extracted from the envelope spectrum; the matching relationship between the frequency interval of the modulation sideband components and the impulse repetition frequency corresponding to the time sequence of the impulse event is calculated to establish a modulation synchronization index; the modulation synchronization index is output as a multi-dimensional evaluation index.
[0035] When calculating the modulation synchronization index based on the standard response signal, the envelope demodulation process is first performed on the frequency band signals extracted from each candidate frequency band to extract the amplitude modulation information caused by the impact of a local bearing fault. Specifically, a Hilbert transform is performed on the frequency band signals to obtain the corresponding analytic signals. ,in, This represents the analytical signal corresponding to the frequency band signal; Indicates the frequency band signal within the candidate frequency band; This represents the result after performing a Hilbert transform on the frequency band signal.
[0036] Furthermore, the envelope amplitude is calculated based on the analytic signal to obtain the envelope signal: ,in, The envelope signal is represented by this. A Fast Fourier Transform is performed on the envelope signal to obtain the corresponding frequency domain amplitude distribution, thus constructing the envelope spectrum. ,in, Represents the envelope spectrum; Represents a frequency variable.
[0037] After obtaining the envelope spectrum, the reference modulation frequency position is determined based on the rotational frequency parameters during motor operation, and modulation sideband components centered on the rotational frequency and its harmonics are extracted from the envelope spectrum. Specifically, based on the rotational frequency... As the center frequency, within the preset search range, the following is obtained: The nearby spectral peak components, among which... This represents the k-th order modulation sideband frequency; This represents the center frequency of the fault feature in the envelope spectrum; This indicates the motor's rotational frequency; k indicates the sideband order.
[0038] Furthermore, the frequency spacing of the extracted adjacent modulation sideband components is calculated to obtain the actual modulation spacing: ,in, This represents the actual frequency spacing between the modulation sideband components; and These represent the adjacent modulation sideband frequencies.
[0039] Simultaneously, based on the time sequence of impact events obtained from impact event monitoring, the average time interval between adjacent impact events is calculated: Further calculate the impact repetition frequency based on the stated time interval: ,in, Indicates the average period of a shock event; Indicates the impact repetition frequency; Indicates the first The time of occurrence of the impact event; n represents the number of impact events.
[0040] The actual frequency spacing of the modulation sideband With impact repetition frequency Perform matching calculations and establish modulation synchronization indicators: ,in, Indicates modulation synchronization index; This represents a small correction to prevent the denominator from being zero.
[0041] When the modulation band spacing is highly consistent with the impact repetition frequency, the modulation synchronization index approaches 1, indicating that there are periodic modulation characteristics corresponding to bearing fault impacts within the current candidate frequency band. When the deviation between the two is large, the modulation synchronization index decreases, indicating that the frequency band mainly contains non-fault-related vibration components. Finally, the calculated modulation synchronization index is used as one of the outputs in the multi-dimensional evaluation index, and together with the impact response consistency index and energy accumulation index, it constitutes the frequency band-impact coupling evaluation matrix, which is used for subsequent dynamic resonant frequency band screening and adaptive reconstruction of frequency band boundaries.
[0042] Furthermore, a frequency band-impact coupling evaluation matrix is constructed using multidimensional evaluation indicators, including: A candidate frequency band set is established based on the standard response signal. An initial evaluation matrix is constructed with the candidate frequency bands as the row dimension and multi-dimensional evaluation indicators as the column dimension. The impulse response consistency index, energy accumulation index, and modulation synchronization index corresponding to each candidate frequency band are filled into the initial evaluation matrix. The multi-dimensional evaluation indicators in the initial evaluation matrix are normalized to eliminate the influence of differences in the dimensions and value ranges of different indicators. Based on the periodic matching relationship between the impulse response consistency index and the modulation synchronization index, and the energy constraint relationship between the energy accumulation index and the impulse event distribution, coupling constraint rules between indicators are constructed. The initial evaluation matrix is subjected to consistency screening according to the coupling constraint rules, and candidate frequency bands that do not meet the impulse physical characteristic constraints are eliminated. The candidate frequency bands that pass the consistency screening are jointly scored in the frequency band dimension and the impulse behavior dimension to obtain the frequency band-impulse coupling evaluation matrix. The joint score is used to characterize the comprehensive matching degree of the candidate frequency band with the impulse response.
[0043] Candidate analysis frequency bands are determined based on the frequency distribution range of the standard response signal, and the analysis frequency bands are divided according to a preset frequency interval to form multiple sets of continuous or non-contiguous candidate frequency bands: Where B represents the candidate frequency band set; Indicates the i-th candidate frequency band; This indicates the number of candidate frequency bands.
[0044] For each candidate frequency band, the previously calculated impulse response consistency index, energy concentration index, and modulation synchronization index are obtained respectively. An initial evaluation matrix is constructed using the candidate frequency band as the matrix row index and the multi-dimensional evaluation index as the matrix column index. ,in, Represents the initial evaluation matrix; This represents the impulse response consistency index corresponding to the i-th candidate frequency band; This represents the energy accumulation index corresponding to the i-th candidate frequency band; This represents the modulation synchronization index corresponding to the i-th candidate frequency band.
[0045] Because different evaluation indicators have different physical meanings and value ranges, to avoid the influence of numerical scale differences on the comprehensive evaluation results due to individual indicators, normalization is performed on each evaluation indicator in the initial evaluation matrix. Specifically, the maximum-minimum normalization method is used to transform each indicator to a uniform range: ,in, This represents the normalized index value; Represents the original index value; This represents the maximum value in the indicator; This represents the minimum value in the indicator.
[0046] After normalizing the indicators, coupling constraint rules between the indicators are established based on the physical correlation between them. Specifically, based on the periodic matching relationship between the impulse response consistency index and the modulation synchronization index, it is determined whether the impulse repetition period detected in the candidate frequency band simultaneously satisfies the time-domain impulse law and the frequency-domain modulation law; when the impulse response consistency index is high and the modulation synchronization index increases synchronously, the candidate frequency band is determined to satisfy the periodic coupling constraint. The periodic matching relationship is expressed as follows: ,in, Indicates periodic matching deviation; Indicates the consistency index of shock response; This indicates the modulation synchronization index.
[0047] Furthermore, based on the relationship between the energy accumulation index and the distribution of impact events, constraints are imposed on the energy of the impact region and the background region within the candidate frequency band. When the energy accumulation index reaches a preset threshold and the energy proportion of the region corresponding to the impact event is higher than that of the background region, the candidate frequency band is determined to meet the energy constraint conditions. ,in, This represents the energy constraint evaluation quantity; This indicates the local energy in the impact zone; This represents the background energy in the non-impact region; This represents a correction amount to prevent the denominator from being zero.
[0048] Based on the aforementioned periodic matching constraint and energy constraint rules, consistency screening is performed on each candidate frequency band in the initial evaluation matrix. When a candidate frequency band simultaneously satisfies both the periodic coupling constraint and energy constraint conditions, the candidate frequency band is retained; otherwise, the candidate frequency band is determined to be a non-fault response frequency band and is eliminated.
[0049] For candidate frequency bands that pass the consistency screening, a joint score is further performed on the frequency band dimension and the impact behavior dimension to comprehensively evaluate the degree of matching between the candidate frequency bands and the bearing failure impact response: ,in, This represents the joint score of the i-th candidate frequency band; This represents the consistency index of the normalized shock response. This represents the normalized energy accumulation index; This represents the normalized modulation synchronization index. , , Let represent the weight coefficients of the corresponding evaluation indicators, and satisfy the following: Based on the joint scoring results corresponding to each candidate frequency band, the candidate frequency band number, frequency band range, and joint score value are associated and stored to form a frequency band-impact coupling evaluation matrix. ,in, This represents the frequency band-impact coupling evaluation matrix; This indicates the candidate frequency bands that have passed the screening. This indicates the corresponding joint score. The frequency band-impact coupling evaluation matrix is used to characterize the comprehensive response capability of different candidate frequency bands to bearing fault impact characteristics, providing an evaluation basis for subsequent dynamic resonance frequency band screening and adaptive adjustment of frequency band boundaries.
[0050] The weighting coefficients in the joint score can be adaptively determined based on the contribution of each evaluation index to the fault impact response. Specifically, based on the candidate frequency band set after consistency screening, the dispersion of each evaluation index in different candidate frequency bands is statistically analyzed, and the distinguishing ability of the corresponding index is determined by the magnitude of index change; wherein, the more obvious the change of the index between different candidate frequency bands, the higher the contribution of the index to the fault frequency band screening, and the greater the corresponding weight.
[0051] Preferred, ,in, represents the discrete contribution of the j-th evaluation indicator; N represents the number of candidate frequency bands after consistency screening; Indicates the i-th candidate frequency band corresponding to the i-th candidate frequency band. Normalized evaluation index values; This represents the average value of the j-th evaluation indicator.
[0052] Weight normalization calculation is performed based on the discrete contribution of each evaluation indicator: ,in, This represents the weight coefficient corresponding to the j-th evaluation indicator; denoted by , indicating the discrete contribution of the corresponding evaluation index; j corresponds to the impulse response consistency index, energy accumulation index, and modulation synchronization index, respectively.
[0053] Based on the frequency band-impact coupling evaluation matrix, frequency band selection and adaptive reconstruction of frequency band boundaries are performed to construct a dynamic resonant frequency band.
[0054] Furthermore, based on the frequency band-impact coupling evaluation matrix, frequency band selection and adaptive reconstruction of frequency band boundaries are performed to construct a dynamic resonant frequency band, including: After normalizing the multidimensional evaluation indicators, a comprehensive evaluation score is configured for each candidate frequency band. All candidate frequency bands are sorted according to the comprehensive evaluation score, and candidate frequency bands with scores higher than a preset threshold are selected as the initial candidate frequency band set. Continuity discrimination is performed on adjacent frequency bands in the initial candidate frequency band set; if the difference in the comprehensive evaluation score between adjacent frequency bands is less than a set tolerance, the adjacent frequency bands are merged to form an initial continuous frequency band segment. Based on the gradient change of the frequency band-impact coupling evaluation matrix in the frequency dimension, adaptive expansion or contraction is performed on the frequency band boundary of the initial continuous frequency band segment, causing the boundary to expand in the direction of increasing comprehensive evaluation score and contract in the direction of decreasing comprehensive evaluation score, thus reconstructing the frequency band boundary. Each frequency band segment after boundary optimization is reconstructed to obtain at least one dynamic resonance frequency band covering the concentrated impact energy region and synchronized with the modulation characteristics.
[0055] Preferably, the impulse response consistency index, energy accumulation index, and modulation synchronization index in the frequency band-impulse coupling evaluation matrix are normalized to convert different evaluation indices into a unified numerical range. Specifically, a maximum-minimum normalization method is used for each evaluation index: ,in, This represents the normalized value of the i-th index. Represents the original index value; This indicates the maximum value in the current indicator; This represents the minimum value in the current indicator.
[0056] After normalizing the multidimensional evaluation indicators, corresponding weight coefficients are assigned according to the degree of influence of each indicator on the characterization capability of fault impact response, and the comprehensive evaluation score corresponding to each candidate frequency band is calculated: in, This represents the comprehensive evaluation score of the i-th candidate frequency band; This represents the consistency index of the normalized shock response. This represents the normalized energy accumulation index; This represents the normalized modulation synchronization index. , , Let represent the weight coefficients corresponding to each evaluation indicator, and satisfy the following: .
[0057] All candidate frequency bands are sorted in descending order based on the comprehensive evaluation score, and the comprehensive evaluation score is compared with a preset screening threshold. When the comprehensive evaluation score corresponding to a candidate frequency band meets the following conditions: When this happens, the candidate frequency band is retained, and candidate frequency bands that meet the conditions are combined to form an initial candidate frequency band set: ,in, Represents the initial candidate frequency band set; Indicates the i-th candidate frequency band; This indicates the preset filtering threshold.
[0058] Furthermore, a continuity determination is performed on adjacent candidate frequency bands in the initial candidate frequency band set. Specifically, the frequency connectivity relationship between adjacent candidate frequency bands and the corresponding comprehensive evaluation score differences are obtained, and the evaluation differences between adjacent frequency bands are calculated. ,in, This represents the difference in comprehensive evaluation scores between adjacent candidate frequency bands i and j.
[0059] When adjacent candidate frequency bands satisfy the frequency continuity condition, and: ,in, This indicates a preset evaluation tolerance threshold; then, if adjacent candidate frequency bands are determined to have continuous response characteristics, the corresponding frequency bands will be merged to form an initial continuous frequency band segment. ,in, Indicates the initial continuous frequency band; Indicates the lower boundary of the frequency band; It indicates the upper boundary of the frequency band.
[0060] After obtaining the initial continuous frequency bands, adaptive reconstruction of the frequency band boundaries is further performed based on the evaluation change trend of the frequency band-impact coupling evaluation matrix in the frequency dimension. Specifically, gradient calculation is performed on the comprehensive evaluation score near the frequency band boundaries: ,in, This represents the evaluation gradient at frequency position f; This indicates the comprehensive evaluation score corresponding to that frequency position; Indicates the frequency sampling interval.
[0061] When the evaluation gradient of the region outside the frequency band boundary satisfies: Furthermore, if the impulse response consistency index, energy accumulation index, and modulation synchronization index corresponding to the newly added region all meet the preset constraints, it is determined that the region still contains a valid fault response, and the frequency band boundary is extended in the direction of increasing the comprehensive evaluation score; when the evaluation gradient of the boundary region satisfies: Furthermore, if the comprehensive evaluation score of the corresponding region is lower than the preset shrinkage threshold, the region will be removed from the current frequency band to achieve frequency band boundary shrinkage.
[0062] After multiple boundary expansion and contraction adjustments, the optimized continuous frequency bands are recombined to form a dynamic resonant frequency band set: ,in, Represents the set of dynamic resonance frequency bands; This represents the range of the i-th dynamic resonance frequency band.
[0063] The dynamic resonance frequency band covers the concentrated area of bearing fault impact energy and satisfies the constraints of impact response period consistency and modulation synchronization characteristics. Compared with the fixed frequency band analysis method, it can dynamically adjust the analysis range according to the frequency response changes under different fault conditions, providing a more accurate signal input for subsequent impact structure constraint decomposition.
[0064] Furthermore, when performing adaptive expansion or contraction of the frequency band boundary, adjustment constraint management is performed based on the impact event density constraint. The impact event density constraint is that expansion is retained when the impact event density in the expansion area is higher than the background density threshold, and expansion is terminated otherwise.
[0065] Specifically, when extending the boundary of the initial continuous frequency band, the frequency band response signal corresponding to the current region to be extended is first obtained, and envelope demodulation and impulse event detection are performed based on the frequency band response signal. The number of impulse events per unit time in the extended region is counted, and the impulse event density corresponding to the extended region is calculated. ,in, This indicates the density of impact events within the extended area; This indicates the number of valid impact events detected within the extended area; This indicates the time length corresponding to the extended region. Further, a non-fault response region with the same operating conditions as the extended region is selected outside the current dynamic resonance frequency band. The number of impact events within this region is counted, and the background impact event density is calculated. ,in, Indicates the density of background impact events; This indicates the number of impact events detected within the background area; This indicates the duration of the time corresponding to the background area.
[0066] A density constraint relationship is established based on the impact event density in the extended region and the background impact event density: ,in, This represents the density ratio of impact events; This represents a small correction to prevent the denominator from being zero.
[0067] When the impact event density ratio meets a preset constraint, that is, the impact event density in the extended region is higher than the background density threshold: ,in, If the preset background density threshold is indicated, then it is determined that the current extended region contains an effective impact response related to bearing failure, the frequency band boundary extension result is retained, and the extended region is included in the dynamic resonance frequency band range.
[0068] When the impact event density within the extended region is lower than or equal to the background density threshold, it is determined that the extended region mainly contains random vibration or noise components, the frequency band extension operation in the current direction is terminated, and the current frequency band boundary remains unchanged.
[0069] During the band boundary contraction process, the impact event density constraint is also used to determine whether the area to be removed still contains a valid fault response. When the impact event density in the area to be contracted is consistently lower than the background density threshold, the corresponding frequency area is removed from the dynamic resonant band. When there is still an impact event distribution in the area to be contracted that is higher than the background density threshold, the area is retained to avoid the loss of fault feature information due to excessive contraction.
[0070] By using the boundary adjustment constraints based on the impact event density, the expansion range of the dynamic resonance frequency band is determined not only by the change in the frequency band evaluation score, but also by the actual distribution law of the fault impact. This ensures that the final constructed dynamic resonance frequency band covers the effective impact response area and reduces the impact of invalid frequency components on the subsequent fault diagnosis process.
[0071] The signal within the dynamic resonance band is subjected to impact structural constraint decomposition to construct the impact response sequence.
[0072] Specifically, a set of dynamic resonant frequency bands is obtained through adaptive reconstruction of the aforementioned frequency band boundaries, and the frequency band response signals within the range of the dynamic resonant frequency bands are extracted to form the impact response signal to be decomposed. Since the dynamic resonant frequency bands have covered the concentrated area of bearing fault impact energy, the impact response signal to be decomposed contains periodic impact components related to bearing faults and some background vibration components.
[0073] Furthermore, based on the time sequence of impact events obtained from the aforementioned impact event detection, impact structural constraints are established. These constraints include impact occurrence time constraints, impact repetition period constraints, and impact amplitude variation constraints. Specifically, the impact occurrence time constraint limits the effective impact response to be within the time neighborhood of the detected impact events; the impact repetition period constraint ensures that the time interval between consecutive impact events satisfies the periodic variation law of bearing fault characteristics; and the impact amplitude variation constraint determines whether the amplitude variation between adjacent impact events conforms to the fault development trend.
[0074] For the signal to be decomposed within the dynamic resonance frequency band, a sparse decomposition method based on impact structure constraints is adopted, representing the signal to be decomposed as a combination of impact response components and residual background components: ,in, This represents the input signal within the dynamic resonance frequency band; This represents the impact response component to be extracted; This represents the residual background vibration component.
[0075] During the decomposition process, the sparsity of the impact response components and the degree to which the impact structural constraints are satisfied are used as optimization objectives to iteratively solve for the impact response components: ,in, This represents the signal reconstruction error term; This represents the impact time sparse constraint term, used to restrict the impact response to occur only around the effective impact time; This represents the periodic consistency constraint term, used to ensure that the impact interval matches the theoretical failure period; This represents the amplitude evolution constraint term, used to constrain the trend of impact amplitude changes; , , These represent the weight coefficients of the corresponding constraint terms.
[0076] In the specific execution process, firstly, an impact position constraint template is generated based on the impact event time sequence. Local peak values in the dynamic resonance frequency band signal are then matched with this template. When the signal peak position coincides with the impact event position and the interval between adjacent peaks meets the fault characteristic period range, the corresponding impact component is retained. For continuous vibration components that do not meet the impact structure constraints, they are suppressed as background components. Further, the decomposed impact response components undergo amplitude correction and time sorting processing. Each effective impact response is arranged according to its occurrence time to obtain the impact response sequence. Where S represents the impulse response sequence; Indicates the time of occurrence of the i-th valid impact event; The value represents the corresponding impact amplitude; n represents the number of valid impact events. The impact response sequence preserves the periodic impact location, impact amplitude, and time evolution relationship of the bearing failure. Compared with direct analysis using the original vibration signal, it can reduce the influence of speed fluctuations, mechanical background vibration, and random noise on the extraction of fault impact features, providing a data foundation for subsequent impact evolution feature extraction, fault type identification, and fault level assessment.
[0077] Impact evolution features are extracted from the impact response sequence, and fault type identification and fault level assessment are performed to generate bearing fault diagnosis results.
[0078] Furthermore, the impact evolution characteristics are extracted from the impact response sequence to perform fault type identification and fault level assessment, including: Based on the impact response sequence, impact amplitude and impact event interval are statistically analyzed to construct impact amplitude sequences and impact interval sequences. The impact interval sequences are then frequency-converted to obtain impact repetition frequencies. These frequencies are compared with preset characteristic frequencies of various bearing faults, and the fault type is determined based on the matching results. The impact amplitude sequences are then statistically analyzed to obtain the mean and peak value trends of the impact amplitudes. Based on the mean and peak value trends of the impact amplitudes and the stability of the impact repetition frequencies, the degree of fault development is classified according to preset grading rules, establishing fault level assessment results. Finally, bearing fault diagnosis results are generated by combining the fault type and fault level assessment results.
[0079] Preferably, the impact response sequence is analyzed over time to extract the impact amplitude and occurrence time corresponding to each impact event, and impact amplitude sequence and impact interval sequence are constructed respectively. Specifically, based on the impact response sequence: Where S represents the impulse response sequence; Indicates the time when the i-th impact event occurs; represents the magnitude of the i-th impact event; n represents the number of impact events.
[0080] Extract the amplitude of each impact event to form an impact amplitude sequence: ,in, This represents the sequence of impact amplitudes. Simultaneously, the impact event interval is calculated based on the times of adjacent impact events. ,in, This represents the time interval between the i-th adjacent impact events; and These represent the times of occurrence of two adjacent impact events. An impact interval sequence is constructed based on all impact event intervals: ,in, This represents the impact interval sequence.
[0081] Furthermore, frequency conversion is performed on the impact interval sequence to calculate the impact repetition frequency, thereby characterizing the periodic characteristics of bearing failure impacts. Specifically, the average impact period is calculated based on the impact interval sequence: ,in, This represents the average impact period. The impact repetition frequency is calculated based on the average impact period. ,in, This indicates the frequency of impact repetition.
[0082] The impact repetition frequency is compared with preset characteristic frequencies of various bearing faults, including inner ring fault frequencies, outer ring fault frequencies, rolling element fault frequencies, and cage fault frequencies. Frequency matching deviations are calculated for each type of fault characteristic frequency. ,in, This represents the matching deviation corresponding to the characteristic frequency of the k-th type of fault; This represents the characteristic frequency of the k-th type of bearing fault. The fault type corresponding to the fault characteristic frequency with the smallest matching deviation is selected as the identification result: ,in, This indicates the type of bearing fault identified.
[0083] Furthermore, statistical analysis is performed on the impact amplitude sequence to calculate the mean impact amplitude: ,in, This represents the average impact amplitude. Simultaneously, the changes in impact amplitude within consecutive detection cycles are analyzed to obtain the trend of peak impact amplitude changes. Specifically, the maximum impact amplitude within each detection cycle is extracted: ,in, This indicates the peak impact value within the current detection period.
[0084] Based on the relationship between the peak impact values across multiple detection cycles, the peak value variation trend is calculated: ,in, Indicates the trend of peak impact change; This represents the maximum impact amplitude within the j-th detection period; This indicates the maximum impact amplitude within the adjacent previous detection cycle; This indicates the detection cycle time interval.
[0085] Furthermore, the degree of bearing failure development is classified according to the average impact amplitude, the trend of peak impact amplitude, and the stability of impact repetition frequency. Specifically, the degree of impact repetition frequency fluctuation is calculated based on the impact interval sequence: ,in, This indicates the fluctuation in the frequency of impact repetition. This represents the instantaneous repetition frequency calculated based on the i-th impact interval; This represents the average impact repetition frequency.
[0086] Based on preset fault level rules, the average impact amplitude, peak value variation trend, and frequency stability index are comprehensively judged. For example, when the average impact amplitude is low, the peak value variation trend is gentle, and the impact repetition frequency is stable, it is judged as a low-level fault; when the impact amplitude continues to increase, the peak value variation trend increases significantly, or the impact repetition frequency stability decreases, it is judged as a medium-to-high-level fault.
[0087] Based on the above evaluation results, a fault level assessment result is established, and the fault type identification result is correlated and fused with the fault level assessment result to generate a bearing fault diagnosis result. The bearing fault diagnosis result includes the bearing fault type, fault development level, and corresponding impact evolution characteristic parameters.
[0088] In summary, the embodiments of this application have at least the following technical effects: First, the original monitoring signals, including vibration and current signals during motor operation, are read. After constructing a working condition-driven mapping relationship using corresponding operating condition parameters, the original monitoring signals are mapped to a unified working condition reference domain, establishing a decoupled standard response signal. Next, multi-dimensional evaluation indicators are calculated based on the standard response signal, and a frequency band-impact coupling evaluation matrix is constructed using these indicators. These multi-dimensional evaluation indicators include impact response consistency, energy accumulation, and modulation synchronization. Further, frequency band selection and adaptive reconstruction of frequency band boundaries are performed based on the frequency band-impact coupling evaluation matrix to construct a dynamic resonant frequency band. Then, impact structural constraint decomposition is performed on the signals within the dynamic resonant band to construct an impact response sequence. Finally, impact evolution features are extracted from the impact response sequence, and fault type identification and fault level assessment are performed to generate bearing fault diagnosis results. This solves the technical problem of insufficient reliability in motor bearing fault diagnosis caused by changes in equipment operating status and environmental disturbances in the existing technology, achieving the technical effect of improving the reliability of motor bearing fault diagnosis.
[0089] Example 2, based on the same inventive concept as the machine learning-based motor bearing fault diagnosis method in the foregoing examples, such as... Figure 3 As shown, this application provides a machine learning-based motor bearing fault diagnosis system, wherein the system includes: Signal operating condition decoupling module 11: Reads the original monitoring signals, including vibration signals and current signals during motor operation. After constructing the operating condition drive mapping relationship with the corresponding operating condition parameters, the original monitoring signals are mapped to a unified operating condition reference domain to establish a decoupled standard response signal. Evaluation matrix construction module 12: Calculates multi-dimensional evaluation indicators based on the standard response signal and constructs a frequency band-impact coupling evaluation matrix using the multi-dimensional evaluation indicators. The multi-dimensional evaluation indicators include impact response consistency indicators, energy accumulation indicators, and modulation synchronization indicators. Resonance frequency band construction module 13: Performs frequency band screening and adaptive reconstruction of frequency band boundaries based on the frequency band-impact coupling evaluation matrix to construct a dynamic resonance frequency band. Response sequence construction module 14: Performs impact structure constraint decomposition on the signals within the dynamic resonance band to construct an impact response sequence. Bearing fault diagnosis module 15: Extracts impact evolution features from the impact response sequence, performs fault type identification and fault level assessment, and generates bearing fault diagnosis results.
[0090] Furthermore, the evaluation matrix construction module 12 is used to perform the following method: Based on the standard response signal, the corresponding frequency band signal is extracted for each candidate frequency band; the extracted frequency band signal is envelope demodulated to establish an envelope signal; impact event monitoring of the envelope signal is performed to obtain the time sequence of impact events; the time interval between adjacent impact events is calculated using the time sequence, and the time interval is compared with the theoretical fault characteristic period determined by the operating conditions to establish an impact response consistency index; the impact response consistency index is output as a multi-dimensional evaluation index.
[0091] Furthermore, the evaluation matrix construction module 12 is used to perform the following method: Instantaneous energy calculation is performed on the extracted frequency band signal to obtain an energy sequence that changes with time; a time neighborhood is constructed based on the time sequence of the impact event, and the local energy in the time neighborhood is calculated using the energy sequence, while the background energy in the non-impact area is calculated simultaneously; an energy accumulation index is established based on the ratio of local energy to background energy; and the energy accumulation index is output as a multidimensional evaluation index.
[0092] Furthermore, the evaluation matrix construction module 12 is used to perform the following method: After envelope demodulation of the extracted frequency band signal, spectrum analysis is performed to construct the envelope spectrum; modulation sideband components centered on the frequency shift and its harmonics are extracted from the envelope spectrum; the matching relationship between the frequency interval of the modulation sideband components and the impulse repetition frequency corresponding to the time sequence of the impulse event is calculated to establish a modulation synchronization index; the modulation synchronization index is output as a multi-dimensional evaluation index.
[0093] Furthermore, the evaluation matrix construction module 12 is used to perform the following method: A candidate frequency band set is established based on the standard response signal. An initial evaluation matrix is constructed with the candidate frequency bands as the row dimension and multi-dimensional evaluation indicators as the column dimension. The impulse response consistency index, energy accumulation index, and modulation synchronization index corresponding to each candidate frequency band are filled into the initial evaluation matrix. The multi-dimensional evaluation indicators in the initial evaluation matrix are normalized to eliminate the influence of differences in the dimensions and value ranges of different indicators. Based on the periodic matching relationship between the impulse response consistency index and the modulation synchronization index, and the energy constraint relationship between the energy accumulation index and the impulse event distribution, coupling constraint rules between indicators are constructed. The initial evaluation matrix is subjected to consistency screening according to the coupling constraint rules, and candidate frequency bands that do not meet the impulse physical characteristic constraints are eliminated. The candidate frequency bands that pass the consistency screening are jointly scored in the frequency band dimension and the impulse behavior dimension to obtain the frequency band-impulse coupling evaluation matrix. The joint score is used to characterize the comprehensive matching degree of the candidate frequency band with the impulse response.
[0094] Furthermore, the resonant frequency band construction module 13 is used to perform the following method: After normalizing the multidimensional evaluation indicators, a comprehensive evaluation score is configured for each candidate frequency band. All candidate frequency bands are sorted according to the comprehensive evaluation score, and candidate frequency bands with scores higher than a preset threshold are selected as the initial candidate frequency band set. Continuity discrimination is performed on adjacent frequency bands in the initial candidate frequency band set; if the difference in the comprehensive evaluation score between adjacent frequency bands is less than a set tolerance, the adjacent frequency bands are merged to form an initial continuous frequency band segment. Based on the gradient change of the frequency band-impact coupling evaluation matrix in the frequency dimension, adaptive expansion or contraction is performed on the frequency band boundary of the initial continuous frequency band segment, causing the boundary to expand in the direction of increasing comprehensive evaluation score and contract in the direction of decreasing comprehensive evaluation score, thus reconstructing the frequency band boundary. Each frequency band segment after boundary optimization is reconstructed to obtain at least one dynamic resonance frequency band covering the concentrated impact energy region and synchronized with the modulation characteristics.
[0095] Furthermore, the resonant frequency band construction module 13 is used to perform the following method: When performing adaptive expansion or contraction of the frequency band boundary, adjustment constraint management is performed based on the impact event density constraint. The impact event density constraint is that the expansion is retained when the impact event density in the expansion area is higher than the background density threshold, and the expansion is terminated otherwise.
[0096] Furthermore, the signal condition decoupling module 11 is used to perform the following method: Based on the operating condition parameters, an operating condition feature expression is constructed; based on the operating condition feature expression, a mapping function relationship between the original monitoring signal and the target operating condition reference domain is established; the original monitoring signal is transformed using the mapping function relationship so that the signals under different operating conditions are aligned with the unified operating condition reference domain in terms of time scale and amplitude response; consistency correction processing is performed on the mapped signal to eliminate the residual coupling effect caused by operating condition changes and construct a decoupled standard response signal.
[0097] Furthermore, the bearing fault diagnosis module 15 is used to perform the following methods: Based on the impact response sequence, impact amplitude and impact event interval are statistically analyzed to construct impact amplitude sequences and impact interval sequences. The impact interval sequences are then frequency-converted to obtain impact repetition frequencies. These frequencies are compared with preset characteristic frequencies of various bearing faults, and the fault type is determined based on the matching results. The impact amplitude sequences are then statistically analyzed to obtain the mean and peak value trends of the impact amplitudes. Based on the mean and peak value trends of the impact amplitudes and the stability of the impact repetition frequencies, the degree of fault development is classified according to preset grading rules, establishing fault level assessment results. Finally, bearing fault diagnosis results are generated by combining the fault type and fault level assessment results.
[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A machine learning-based method for diagnosing motor bearing faults, characterized in that, The method includes: Read the original monitoring signals, which include vibration signals and current signals during motor operation. After constructing the working condition drive mapping relationship with the corresponding operating condition parameters, map the original monitoring signals to a unified working condition reference domain and establish a decoupled standard response signal. Multidimensional evaluation indicators are calculated based on standard response signals, and a frequency band-impact coupling evaluation matrix is constructed using these multidimensional evaluation indicators. The multidimensional evaluation indicators include impulse response consistency indicators, energy accumulation indicators, and modulation synchronization indicators. Based on the frequency band-impact coupling evaluation matrix, frequency band selection and adaptive reconstruction of frequency band boundaries are performed to construct a dynamic resonant frequency band. Impact structural constraint decomposition is performed on the signal within the dynamic resonance flat band to construct the impact response sequence; Impact evolution features are extracted from the impact response sequence, and fault type identification and fault level assessment are performed to generate bearing fault diagnosis results.
2. The machine learning-based motor bearing fault diagnosis method as described in claim 1, characterized in that, Multidimensional evaluation indices are calculated based on standard response signals, including: Based on the standard response signal, extract the corresponding frequency band signal for each candidate frequency band; The extracted frequency band signal is envelope demodulated to establish an envelope signal; Execute the impact event monitoring of the envelope signal to obtain the time sequence of impact events; The time interval between adjacent impact events is calculated using the time series, and the time interval is compared with the theoretical fault characteristic period determined by the operating conditions to establish an impact response consistency index. The shock response consistency index is output as a multidimensional evaluation index.
3. The machine learning-based motor bearing fault diagnosis method as described in claim 2, characterized in that, The calculation of multidimensional evaluation indicators based on standard response signals also includes: Instantaneous energy calculation is performed on the extracted frequency band signal to obtain the energy sequence that changes with time; A time neighborhood is constructed based on the time sequence of the impact event. After calculating the local energy within the time neighborhood using the energy sequence, the background energy of the non-impact region is calculated simultaneously. An energy accumulation index is established based on the ratio of local energy to background energy; The energy accumulation index is output as a multidimensional evaluation index.
4. The machine learning-based motor bearing fault diagnosis method as described in claim 3, characterized in that, The calculation of multidimensional evaluation indicators based on standard response signals also includes: After envelope demodulation of the extracted frequency band signal, spectral analysis is performed to construct the envelope spectrum; Extract the modulation sideband components centered on the frequency shift and its harmonics from the envelope spectrum; Calculate the matching relationship between the frequency interval of the modulation side band component and the impulse repetition frequency corresponding to the time sequence of the impulse event, and establish a modulation synchronization index; The modulation synchronization index is output as a multi-dimensional evaluation index.
5. The machine learning-based motor bearing fault diagnosis method as described in claim 4, characterized in that, A frequency band-impact coupling evaluation matrix is constructed using multidimensional evaluation indicators, including: A candidate frequency band set is established based on the standard response signal, and an initial evaluation matrix is constructed with the candidate frequency bands as the row dimension and the multi-dimensional evaluation index as the column dimension. Fill the initial evaluation matrix with the impulse response consistency index, energy accumulation index and modulation synchronization index corresponding to each candidate frequency band; After normalizing the multidimensional evaluation indicators in the initial evaluation matrix to eliminate the influence of differences in the dimensions and value ranges of different indicators, coupling constraint rules between indicators are constructed based on the periodic matching relationship between the impact response consistency index and the modulation synchronization index, and the energy constraint relationship between the energy accumulation index and the impact event distribution. The initial evaluation matrix is subjected to consistency screening based on the coupling constraint rules, and candidate frequency bands that do not meet the impact physical characteristic constraints are eliminated. Candidate frequency bands that pass the consistency screening are jointly scored in the frequency band dimension and the impact behavior dimension to obtain the frequency band-impact coupling evaluation matrix, wherein the joint score is used to characterize the comprehensive matching degree of the candidate frequency band to the impact response.
6. The machine learning-based motor bearing fault diagnosis method as described in claim 5, characterized in that, Based on the frequency band-impact coupling evaluation matrix, frequency band selection and adaptive reconstruction of frequency band boundaries are performed to construct a dynamic resonant frequency band, including: After normalizing the multidimensional evaluation indicators, configure the comprehensive evaluation score corresponding to each candidate frequency band; All candidate frequency bands are sorted according to the comprehensive evaluation score, and candidate frequency bands with scores higher than a preset threshold are selected as the initial candidate frequency band set. Continuity discrimination is performed on adjacent frequency bands in the initial candidate frequency band set. When the difference in the comprehensive evaluation score of an adjacent frequency band is less than a set tolerance, the adjacent frequency bands are merged to form an initial continuous frequency band segment. Based on the gradient change of the frequency band-impact coupling evaluation matrix in the frequency dimension, adaptive expansion or contraction is performed on the frequency band boundary of the initial continuous frequency band, so that the boundary expands in the direction of increasing comprehensive evaluation score and contracts in the direction of decreasing comprehensive evaluation score, thus reconstructing the frequency band boundary. By reconstructing each frequency band after boundary optimization, at least one dynamic resonant frequency band covering the concentrated area of impact energy and synchronized with the modulation characteristics is obtained.
7. The machine learning-based motor bearing fault diagnosis method as described in claim 6, characterized in that, When performing adaptive expansion or contraction of the frequency band boundary, adjustment constraint management is performed based on the impact event density constraint. The impact event density constraint is that the expansion is retained when the impact event density in the expansion area is higher than the background density threshold, and the expansion is terminated otherwise.
8. The machine learning-based motor bearing fault diagnosis method as described in claim 1, characterized in that, Mapping the raw monitoring signals to a unified operating condition reference domain includes: A condition feature representation is constructed based on the aforementioned operating condition parameters; Based on the aforementioned operating condition characteristics, a mapping function relationship is established between the original monitoring signal and the target operating condition reference domain; The original monitoring signal is transformed using the mapping function relationship to align the signals under different operating conditions with the unified operating condition reference domain in terms of time scale and amplitude response. The mapped signal is subjected to consistency correction to eliminate the residual coupling effects caused by changes in operating conditions, and a decoupled standard response signal is constructed.
9. The machine learning-based motor bearing fault diagnosis method as described in claim 1, characterized in that, Impact evolution features are extracted from impact response sequences to perform fault type identification and fault level assessment, including: Based on the statistical analysis of impact amplitude and impact event intervals using impact response sequences, impact amplitude sequences and impact interval sequences are constructed. The impact repetition frequency is obtained by frequency conversion of the impact interval sequence. The impact repetition frequency is compared with the preset characteristic frequencies of various bearing faults, and the fault type is determined according to the matching result. The distribution statistics of the impact amplitude sequence are performed to obtain the mean and peak value variation trends of the impact amplitude; Based on the mean and peak value variation trends of the impact amplitude and the stability of the impact repetition frequency, the degree of fault development is classified according to the preset classification rules, and a fault level assessment result is established. The bearing fault diagnosis results are generated by combining the assessment results of fault type and fault level.
10. A machine learning-based motor bearing fault diagnosis system, characterized in that, The system is used to implement the machine learning-based motor bearing fault diagnosis method according to any one of claims 1-9, the system comprising: Signal operating condition decoupling module: Reads the original monitoring signals, which include vibration signals and current signals during motor operation. After constructing the operating condition drive mapping relationship with the corresponding operating condition parameters, the original monitoring signals are mapped to a unified operating condition reference domain to establish a decoupled standard response signal. Evaluation matrix construction module: Calculates multi-dimensional evaluation indicators based on standard response signals, and constructs a frequency band-impact coupling evaluation matrix using the multi-dimensional evaluation indicators. The multi-dimensional evaluation indicators include impulse response consistency indicators, energy accumulation indicators, and modulation synchronization indicators. Resonance band construction module: Based on the band-impact coupling evaluation matrix, band selection and adaptive reconstruction of band boundaries are performed to construct a dynamic resonance band; Response sequence construction module: Performs impact structural constraint decomposition on the signal within the dynamic resonance flat band to construct the impact response sequence; Bearing fault diagnosis module: Extracts impact evolution features from impact response sequence, performs fault type identification and fault level assessment, and generates bearing fault diagnosis results.
