Wind generating set fault monitoring method and system based on voiceprint recognition

Abstract

The invention provides a wind generating set fault monitoring method and system based on voiceprint recognition, and the method comprises the steps: collecting gear box voiceprint signal data in real time, recording signal fluctuation caused by gear surface wear, and obtaining an original signal data set containing a frequency spectrum high-frequency component enhancement feature; a time-frequency analysis method is adopted for the original signal data set, non-linear interference of vibration signals is recognized, multi-scale decomposition is carried out, initial wear frequency spectrum narrow-band characteristics and medium-term harmonic components are separated out, and a frequency spectrum component set is obtained; extracting characteristic parameters related to modulation depth abnormal fluctuation from the frequency spectrum component set, identifying a gear pair meshing frequency change rule, and determining a distribution mode of a tooth surface contact noise proportion; and if the feature matching result shows that the deviation between the spectrum high-frequency component diffusion distortion feature and the reference feature library exceeds a threshold value, adding the tooth surface fatigue crack noise feature into the feature library to obtain a target feature library.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a wind turbine generator set fault monitoring method and system based on voiceprint recognition. Background Art

[0002] As a key pillar of clean energy, wind power generation plays a vital role in the global energy transition. The stable operation of its equipment is directly related to the reliability and economic benefits of energy supply. However, gearboxes, as core components of wind turbines, operate under high loads for long periods of time, making them prone to wear and failure, severely impacting equipment lifespan and power generation efficiency. Therefore, research on gearbox operating condition monitoring technology has become a key area for ensuring equipment safety and improving the quality of industry development. Although various methods have been developed for gearbox fault detection, most often struggle to adapt to dynamic operating environments under complex operating conditions. In particular, the accuracy of monitoring results is insufficient when capturing subtle fault signals. These methods often overlook underlying patterns in fault signals when processing data features, making it difficult to accurately identify early faults. Against this backdrop, the acoustic signature of gearbox operating conditions has become a key area of ​​research. A prominent challenge lies in the impact of gear surface wear on acoustic signatures. Changes in gear surface wear significantly enhance the high-frequency noise components in the acoustic signature. This change further leads to a nonlinear increase in the modulation depth of the acoustic signature envelope, complicating signal feature extraction. This evolution from surface state to signal characteristics makes it difficult for traditional feature libraries to adapt to dynamically changing environments, which in turn affects the accuracy of fault identification. Therefore, how to dynamically adjust the update mechanism of the voiceprint feature library under the condition of changing gear surface wear to adapt to the signal complexity caused by high-frequency noise enhancement and nonlinear growth of modulation depth has become a key issue in improving the accuracy of early gearbox fault identification. Summary of the Invention

[0003] The present invention provides a wind turbine generator fault monitoring method based on voiceprint recognition, which mainly includes:

[0004] Collect gearbox soundprint signal data in real time, record signal fluctuations caused by gear surface wear, and obtain the original signal data set containing enhanced features of high-frequency components of the spectrum;

[0005] The time-frequency analysis method is used on the original signal data set to identify the nonlinear interference of the vibration signal, and multi-scale decomposition is performed to separate the narrowband characteristics of the spectrum in the early stage of wear and the harmonic components in the middle stage, thereby obtaining a set of spectrum components.

[0006] Extract characteristic parameters related to abnormal fluctuations in modulation depth from the set of spectral components, identify the variation pattern of the gear pair meshing frequency, and determine the distribution pattern of the tooth surface contact noise ratio;

[0007] The spectrum peak shift amplitude is obtained by monitoring the speed fluctuation, and compared with the difference in the benchmark feature library data to determine the degree of influence of the noise caused by wear particles on feature extraction. The signal-to-noise ratio deterioration index is used to obtain the feature matching comparison results;

[0008] If the feature matching result shows that the deviation between the high-frequency component diffusion distortion feature of the spectrum and the reference feature library exceeds the threshold, the tooth surface fatigue crack noise feature is added to the feature library to obtain the target feature library;

[0009] Identify the energy distribution of different wear stages based on the target feature library, perform noise filtering based on the tooth surface contact noise proportional distribution pattern, and determine the fault signal classification result by comparing the effect of wear depth on frequency offset.

[0010] Extract fault characteristic parameters from the classification results, analyze the correlation between the enhanced high-frequency components of the spectrum and the nonlinear interference of the vibration signal based on the coupling relationship between the gearbox load change and the speed fluctuation, and judge the degree of abnormality in the operating state;

[0011] Status alarm information is generated based on the degree of abnormality in the operating status. By correlating the alarm information with the background data caused by wear particles, the status assessment data of tooth surface fatigue crack noise is obtained to identify the results of minor gearbox wear faults.

[0012] The present invention provides a wind turbine generator fault monitoring system based on voiceprint recognition, which mainly includes:

[0013] The signal acquisition module is used to collect gearbox soundprint signal data in real time, record the signal fluctuations caused by gear surface wear, and obtain the original signal data set containing the enhanced characteristics of the high-frequency components of the spectrum;

[0014] The time-frequency analysis module is used to apply time-frequency analysis to the original signal data set, identify nonlinear interference in the vibration signal, perform multi-scale decomposition, separate the narrowband characteristics of the initial wear spectrum and the harmonic components of the mid-term wear spectrum, and obtain a set of spectrum components;

[0015] The feature extraction module is used to extract characteristic parameters related to abnormal fluctuations in modulation depth from the set of spectral components, identify the variation pattern of the meshing frequency of the gear pair, and determine the distribution pattern of the tooth surface contact noise ratio;

[0016] The feature matching module is used to obtain the spectrum peak offset amplitude through speed fluctuation monitoring, compare the difference with the benchmark feature library data, and determine the degree of influence of the noise caused by wear particles on feature extraction. The signal-to-noise ratio deterioration index is used to obtain the feature matching comparison results;

[0017] A feature library updating module is used to add the tooth surface fatigue crack noise feature to the feature library to obtain a target feature library if the feature matching result shows that the deviation between the spectral high-frequency component diffusion distortion feature and the reference feature library exceeds a threshold;

[0018] The fault classification module is used to identify the energy distribution of different wear stages based on the target feature library, perform noise filtering based on the proportional distribution pattern of the tooth surface contact noise, and determine the fault signal classification result by comparing the effect of wear depth on frequency offset;

[0019] The state analysis module is used to extract fault characteristic parameters from the classification results, analyze the correlation between the enhancement of the high-frequency component of the spectrum and the nonlinear interference of the vibration signal based on the coupling relationship between the gearbox load change and the speed fluctuation, and determine the degree of abnormality of the operating state;

[0020] The alarm generation module is used to generate status alarm information based on the degree of abnormality in the operating status. By correlating the alarm information with the background data caused by wear particles, the status assessment data of tooth surface fatigue crack noise is obtained and the results of minor wear faults of the gearbox are identified.

[0021] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0022] The present invention discloses a wind turbine fault monitoring method based on soundprint recognition. By collecting gearbox soundprint signal data in real time, the original signal is subjected to time-frequency analysis and multi-scale decomposition, characteristic parameters related to abnormal fluctuations in modulation depth are extracted, the variation law of the meshing frequency of the gear pair is identified, and the proportional distribution pattern of the tooth surface contact noise is determined. By combining the speed fluctuation monitoring and the comparison with the benchmark feature library, the noise influence caused by wear particles is judged, and a target feature library is established. The energy distribution of different wear stages is identified according to the target feature library, and noise filtering is performed in combination with the proportional distribution pattern of the tooth surface contact noise to determine the fault signal classification result. By analyzing the correlation between the enhancement of the high-frequency component of the spectrum and the nonlinear interference of the vibration signal, the abnormal degree of the operating state is judged, and status alarm information is generated to achieve accurate identification of minor wear faults of the gearbox, providing a basis for preventive maintenance of the gearbox. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a wind turbine generator set fault monitoring method based on voiceprint recognition according to the present invention.

[0024] Figure 2 The figure is a structural diagram of a wind turbine fault monitoring system based on voiceprint recognition according to the present invention. DETAILED DESCRIPTION

[0025] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0026] like Figure 1-2 In this embodiment, a method and system for monitoring wind turbine generator faults based on voiceprint recognition may specifically include:

[0027] S101. Collect gearbox soundprint signal data in real time, record signal fluctuations caused by gear surface wear, and obtain an original signal data set containing enhanced features of high-frequency components of the spectrum.

[0028] During the operation of the gearbox, acoustic sensors installed at different positions of the box continuously collect soundprint signals at a preset sampling frequency, record the sound pressure level value sequence, and convert the sound pressure level value sequence into original soundprint time series data. For the original soundprint time series data, the fast Fourier transform is used to convert the time domain signal into a frequency domain signal, extract the amplitude distribution of the spectrum components within the preset frequency range, and obtain the amplitude data corresponding to each frequency point. According to the amplitude distribution of the spectrum components, the amplitude difference between adjacent frequency points is calculated. If the cumulative sum of the amplitude differences in a certain frequency band exceeds the preset threshold, it is determined that the frequency band has a high-frequency component enhancement feature, and the starting frequency, cutoff frequency and amplitude peak of the frequency band are recorded. Based on the recorded high-frequency component enhancement feature parameters, the original soundprint time series data segments of the corresponding time period are extracted, and each data segment is associated with its high-frequency component enhancement feature parameters and stored to construct an original signal data set containing the spectrum high-frequency component enhancement feature.

[0029] Specifically, as a core component of a mechanical transmission system, the operating status of the gearbox directly impacts the performance and lifespan of the entire device. In actual industrial applications, gear surface wear is one of the most common failure modes. When gear surfaces are worn, abnormal vibrations and sounds are generated during the gear meshing process. These abnormal signals manifest themselves as increased energy in specific frequency bands on the spectrum.

[0030] Specifically, the placement of acoustic sensors is crucial to the quality of signal acquisition. On the gearbox housing, sensors are typically installed close to the gear meshing point and in a direct vibration transmission path, such as near the bearing seat or on the sidewall of the housing. The sound pressure level sequence is actually data collected by the sensor showing the time-varying acoustic pressure. This data requires analog-to-digital conversion to form a digitized time series signal. The sampling frequency must satisfy the Nyquist theorem to ensure accurate capture of the changing characteristics of high-frequency components.

[0031] In one possible implementation, a fast Fourier transform (FFT) converts the time-domain signal to the frequency domain, allowing for clear visualization of the energy distribution of different frequency components. In a properly functioning gearbox, the frequency spectrum is primarily concentrated at the gear meshing frequency and its lower harmonics. However, when gear surface wear occurs, increased tooth surface roughness and localized poor contact excite more high-frequency vibration components. These high-frequency components manifest themselves on the spectrum as a significant increase in the amplitude of previously low-energy high-frequency bands.

[0032] It's important to note that the amplitude difference calculation method is effective in identifying unusual variations in the spectrum. By calculating the amplitude differences between adjacent frequency points, the degree of localized spectrum variation can be quantified. When the cumulative sum of the amplitude differences within a frequency band exceeds a preset threshold, it indicates significant energy accumulation within that band, a typical characteristic of enhanced high-frequency components caused by gear wear. Recording the parameters of these characteristic frequency bands, including the start frequency, cutoff frequency, and peak amplitude, provides a quantitative basis for subsequent fault diagnosis.

[0033] Preferably, when constructing a dataset, the original time series data segments are associated with their corresponding high-frequency feature parameters and stored to form a structured data organization method. Each data record contains not only the original voiceprint signal but also the feature parameters extracted through analysis. This data organization method not only preserves the integrity of the original information but also provides feature tags that facilitate rapid retrieval and analysis.

[0034] S102. Using a time-frequency analysis method on the original signal data set, identify nonlinear interference in the vibration signal, perform multi-scale decomposition, separate the narrowband characteristics of the initial wear spectrum and the harmonic components in the middle stage, and obtain a spectrum component set.

[0035] For the vibration signal time series data in the original signal dataset, a short-time Fourier transform is used to divide the signal into multiple time windows. The signal within each window is Fourier transformed to obtain the spectral amplitude distribution corresponding to each time window, and a time-frequency spectrum matrix is ​​constructed. Based on the time-frequency spectrum matrix, the ratio of the amplitude difference between adjacent time windows to the average amplitude of each frequency point is calculated. If the ratio exceeds a preset threshold, it is determined that nonlinear interference exists at that frequency point. All interfering frequency points and their time positions are recorded to form an interference marker matrix. Based on the interference marker matrix, the signal components corresponding to the marked interfering time-frequency points are removed from the original vibration signal. The signal after interference removal is subjected to multi-scale decomposition using a wavelet transform to obtain signal component data at different decomposition levels. From the signal component data at different decomposition levels, components with frequency bandwidths less than a preset value are screened as narrowband characteristic signals in the early stage of wear, and components with frequency values ​​that are integer multiples of the gear meshing frequency are identified as mid-term harmonic characteristic signals. The narrowband characteristic signals and harmonic characteristic signals are stored separately to obtain a set of spectral components.

[0036] Specifically, short-time Fourier transform has unique advantages in processing gearbox vibration signals. It divides long-term signals into multiple short-time windows. The signal in each window can be approximately regarded as a stationary signal, thereby capturing the change pattern of the signal frequency component over time.

[0037] In one possible implementation, the length of the time window is usually selected to include 2-3 gear meshing cycles, which can ensure frequency resolution and track the time-varying characteristics of the frequency components.

[0038] Specifically, when constructing the time-spectrogram matrix, each matrix element represents the energy intensity of a specific frequency at a specific moment. When the gearbox is operating normally, the time-spectrogram exhibits a stable, striped distribution, with the primary energy concentrated at the gear meshing frequency and its harmonics. However, nonlinear interference, such as bearing failure, misalignment, or external shock, manifests as sudden energy clumps or irregular frequency modulation on the time-spectrogram.

[0039] It should be noted that the identification of nonlinear interference relies on the calculation of the amplitude change rate. By calculating the ratio of the amplitude difference between adjacent time windows to the average amplitude, the stability of the frequency component can be quantified.

[0040] For example, if the amplitude of a frequency point in five consecutive time windows is 100, 105, 98, 102, and 200, respectively, then the rate of change of the fifth window relative to the average value of the first four windows is approximately 96%, far exceeding the normal fluctuation range, indicating that there is abnormal interference at this frequency at that moment. The establishment of the interference marker matrix provides precise positioning information for subsequent signal processing. This matrix has the same dimensions as the time-frequency spectrum matrix, in which the time-frequency points marked as interference are represented by specific identifiers. When eliminating the interference signal component, the time-frequency domain masking method can be used, that is, the amplitude corresponding to the marked time-frequency point is set to zero or replaced with the interpolation value of the adjacent normal point, thereby obtaining a purer vibration signal.

[0041] The multi-scale decomposition characteristics of the wavelet transform make it particularly suitable for analyzing the characteristics of gear wear at different stages. At low decomposition levels, high-frequency details can be captured, often corresponding to the subtle narrowband features of the early stages of wear. At higher decomposition levels, low-frequency profile information, including the fundamental frequency of the gear mesh and its harmonic components, is primarily retained. By analyzing signal characteristics at different scales, it is possible to effectively separate local anomalies in the early stages of wear from overall performance degradation in the middle stages. The identification of narrowband signatures is based on frequency bandwidth calculation. When the signal energy at a certain decomposition level is concentrated within a very narrow frequency range, it indicates a distinct periodicity, a typical manifestation of localized defects on the tooth surface in the early stages of wear. Harmonic signatures are identified by the integer multiple relationship between frequencies. When multiple frequency components satisfy the relationship f, 2f, 3f, etc., it indicates that gear wear has progressed to the middle stage, affecting overall mesh quality. By extracting and storing these two types of signature signals separately, the resulting spectral component sets provide a structured feature foundation for subsequent wear assessment and fault diagnosis.

[0042] S103. Extract characteristic parameters related to abnormal fluctuations in modulation depth from the set of spectral components, identify the variation pattern of the meshing frequency of the gear pair, and determine the distribution pattern of the tooth surface contact noise ratio.

[0043] From the set of spectral components, the gear meshing frequency is identified as the carrier frequency, and the frequency components spaced on either side of the carrier frequency by integer multiples of the rotation frequency are identified as sidebands. The modulation depth is calculated as the ratio of the carrier frequency amplitude to the sum of the adjacent sideband amplitudes. The modulation depth change rate is then compared to the ratio of the modulation depth difference within consecutive time periods to the time interval. Based on the modulation depth change rate, the actual carrier frequency value at each moment is tracked. Taking the carrier frequency at the first moment as the reference, the difference between the carrier frequency and the reference value at subsequent moments is calculated as the carrier frequency offset. The ratio of the sum of the squares of the upper and lower sideband amplitudes to the square of the carrier frequency amplitude is also calculated to obtain the sideband energy distribution ratio. Based on the sequence of carrier frequency offsets and the sequence of sideband energy distribution ratios, the temporal correspondence between the two is identified. As the offset increases, the corresponding energy distribution ratio value is counted. A mapping table between offset and energy ratio is constructed to determine the energy distribution characteristics under different frequency offset conditions. By mapping the data points whose energy distribution ratio exceeds the preset threshold in the relationship table, the percentage of these points in the total data points is calculated as the tooth surface contact noise ratio. The data are grouped according to the offset size, and the noise ratio values ​​in each group are counted to determine the distribution pattern of the tooth surface contact noise ratio.

[0044] Specifically, the modulation phenomenon of the gearbox vibration signal is an important basis for fault diagnosis.

[0045] In one possible implementation, when there is local damage or wear on the gear, an impact will be generated each time the damaged part participates in meshing. This periodic impact modulates the normal meshing vibration, which is manifested in the spectrum as sideband components on both sides of the carrier frequency.

[0046] Specifically, the carrier frequency is usually the meshing frequency of the gears, which is equal to the gear speed multiplied by the number of teeth.

[0047] For example, a 30-tooth gear running at 1000 rpm has a meshing frequency of 500 Hz. Sidebands are frequency components formed on either side of the carrier frequency due to modulation, typically separated by the shaft's rotational frequency or an integer multiple thereof. When gears wear, the amplitude of the sidebands increases significantly, and this increase can be quantified by the modulation depth. Calculating the modulation depth reflects the strength of the modulating signal. By calculating the ratio of the carrier frequency amplitude to the sum of the sideband amplitudes, a dimensionless modulation depth metric can be derived. Under normal circumstances, this ratio is large, indicating that the carrier frequency dominates. However, when a fault occurs, the sideband energy increases, causing the ratio to decrease and the modulation depth to increase. By continuously monitoring the rate of change of the modulation depth, the dynamic process of fault development can be captured.

[0048] It should be noted that the occurrence of carrier frequency offset is often closely related to changes in the gearbox's operating state. When the gearbox load changes or the speed fluctuates, the actual meshing frequency will deviate from the theoretical value. This offset is not random, but rather corresponds to the operating conditions. By establishing a mapping relationship between the offset and the sideband energy distribution ratio, different types of fault modes can be distinguished. The calculation of the sideband energy distribution ratio uses the concept of energy, that is, the square of the amplitude. This is because in vibration signal analysis, energy better reflects the actual strength of the signal. The ratio of the sum of the energy of the upper and lower sidebands to the carrier energy intuitively represents the strength of the modulation effect. When this ratio exceeds a certain threshold, it indicates a significant tooth contact anomaly.

[0049] Preferably, by constructing a mapping relationship table between offset and energy ratio, fault characteristic patterns under different working conditions can be discovered.

[0050] For example, under light load, the offset is small but the energy ratio may be high, indicating that the poor tooth contact is more obvious under light load; under heavy load, the offset increases, but due to the stabilizing effect of the load, the energy ratio may decrease instead. The establishment of this mapping relationship provides an important basis for fault diagnosis. The distribution pattern of the tooth surface contact noise ratio is finally obtained through statistical analysis. All data points are grouped according to the size of the frequency offset, and the proportion of data points in each group whose energy distribution ratio exceeds the threshold is statistically calculated. This proportion is the tooth surface contact noise ratio under this working condition. By plotting the noise ratio corresponding to different offset intervals, we can clearly see the distribution pattern of faults as the working conditions change, providing a scientific basis for formulating targeted maintenance strategies.

[0051] Based on the modulation depth change rate and carrier frequency offset, the drift trajectory of the fundamental frequency harmonic components during the meshing process of the master and slave gears of the gear pair is tracked, the frequency modulation phenomenon caused by the periodic fluctuation of the meshing stiffness is identified, the meshing impact pulse interval change pattern and phase difference distribution law are extracted, and the microscopic contact state of the tooth surface is analyzed in combination with the sideband energy distribution ratio to generate a characteristic map of the gear pair operating state.

[0052] Based on the timing data of the modulation depth change rate and carrier frequency offset, the positional changes of the fundamental frequency and its integer multiples in the time-frequency domain are tracked. The frequency values ​​of each frequency multiple component are recorded over time, and periods of periodic frequency fluctuation are identified. The difference between the maximum and minimum frequency values ​​within the fluctuation period is calculated as the drift amplitude. Based on the drift amplitude data, transient impact components exceeding a preset threshold are extracted from the vibration signal as impact pulses. The time intervals between adjacent pulse peaks are calculated, and the probability distribution of the interval values ​​is calculated. Simultaneously, the phase values ​​at the peaks of each subsequent pulse are measured, using the first pulse peak as the zero-phase reference point, to obtain a phase difference sequence. The standard deviation of the phase difference sequence is used to assess tooth contact uniformity. Broadband components in the spectrum exceeding ten times the meshing frequency are identified. The noise type is determined based on the correlation between the amplitude of the broadband component and the impact pulse timing. The sum of the squares of the amplitudes of each noise component is calculated to obtain the energy of metal friction noise and aerodynamic noise. The sum of metal friction noise energy and aerodynamic noise energy is taken as the total noise energy, and the proportion of metal friction noise energy is calculated. The contact stress level is estimated based on the product of the proportion value and the frequency drift amplitude. The drift amplitude, phase difference sequence standard deviation, and the proportion of various noise energies are combined into a multidimensional feature vector to form a characteristic map of the gear pair operating status.

[0053] Specifically, during the operation of a gear pair, the frequency drift of the fundamental frequency and its harmonic components directly reflects the dynamic changes in the meshing state.

[0054] In one possible implementation, when a gearbox is subjected to varying loads, the elastic deformation of the gears causes a slight shift in the actual meshing position. This shift manifests as a frequency drift on the spectrum. By continuously recording the frequency values ​​of each harmonic component, a frequency curve can be plotted over time. These curves often exhibit periodic fluctuations.

[0055] Specifically, calculating the drift amplitude requires determining an observation period, typically a period encompassing a complete load cycle. Within this period, the maximum and minimum frequencies of each harmonic component are determined; the difference between the two is the drift amplitude for that harmonic component.

[0056] For example, the fundamental frequency is 100Hz under light load and drifts to 102Hz under heavy load, a drift amplitude of 2Hz. This seemingly minor change actually contains a wealth of fault information. Extracting impact pulses is a key step in identifying tooth surface damage. When a localized defect exists on a tooth surface, each time the defective area engages in meshing, a transient impact is generated, manifesting as a sharp pulse in the time domain signal. The time intervals between these pulses are not completely uniform, but are affected by speed fluctuations and load variations. By statistically analyzing the probability distribution of a large number of pulse intervals, it can be found that the interval distribution of normal gears is relatively concentrated, while the distribution of damaged gears is more dispersed.

[0057] It should be noted that the establishment of the phase difference sequence requires the selection of an appropriate reference point. The moment of the first detected impact pulse peak is used as the zero phase reference, and the phase value of each subsequent pulse is obtained by calculating the deviation between its moment and the theoretical period. The standard deviation of the phase difference sequence directly reflects the uniformity of the tooth surface contact. The larger the standard deviation, the more uneven the contact, which may indicate the presence of eccentric load or uneven wear of the tooth surface. The identification of broadband noise components is based on a comprehensive judgment of the frequency range and time domain characteristics. Metal friction noise usually appears synchronously with the impact pulse, and appears as a transient broadband energy surge on the spectrum; while aerodynamic noise presents a continuous distribution feature, mainly concentrated in the high frequency band. By calculating the time domain correlation of the energy of each frequency band, the two types of noise can be effectively distinguished. The energy is calculated using the sum of squares of the amplitude, which can accurately reflect the power characteristics of the signal.

[0058] Preferably, the estimation of the contact stress level is based on the physical relationship between noise energy and load. When the contact stress increases, the deformation of the microscopic roughness peaks on the tooth surface intensifies, and the friction noise energy generated increases accordingly. By multiplying the metal friction noise energy ratio by the frequency drift amplitude, a comprehensive index can be obtained, which is positively correlated with the contact stress level. The resulting characteristic map is a multi-dimensional state characterization tool. With the drift amplitude as the horizontal axis, the phase difference standard deviation as the vertical axis, and the noise energy ratio represented by the depth of color, the operating state of the gear pair can be intuitively displayed. A normally operating gear pair appears on the map as an area of ​​low drift, low phase difference, and low noise energy; as the degree of wear increases, the characteristic points will gradually migrate to the high-value area, providing a reliable basis for predictive maintenance.

[0059] S104. Obtain the spectrum peak shift amplitude through speed fluctuation monitoring, compare the difference with the benchmark feature library data, determine the impact of the noise caused by wear particles on feature extraction, and use the signal-to-noise ratio deterioration index to obtain feature matching comparison results.

[0060] By monitoring the real-time changes in the gearbox speed signal, the difference sequence between the instantaneous speed value and the average speed value is calculated. Spectral analysis of this difference sequence is performed to identify the frequency locations corresponding to each peak in the spectrum. The peak frequency offset relative to the theoretical frequency at rated speed is recorded to obtain the spectrum peak offset amplitude data. Based on the spectrum peak offset amplitude data and the current speed value, a standard spectrum feature under the same speed conditions is extracted from a pre-established benchmark feature library. The amplitude difference between the current spectrum and the standard spectrum at each frequency point is calculated. If the difference exceeds a preset threshold, the frequency point is determined to be affected by noise generated by wear particles. The number of affected frequency points is accumulated and the sum of the amplitude deviations is calculated. Based on the sum of the amplitude deviations of the affected frequency points, the sum of the squared amplitudes of the unaffected frequency points is used as the signal power, and the sum of the squared deviations of the affected frequency points is used as the noise power. The ratio of the two is calculated to obtain the current signal-to-noise ratio (SNR). This ratio is then subtracted from the standard SNR stored in the benchmark feature library to obtain the SNR degradation index. The geometric mean of the sum of the amplitude products of the current spectrum and the standard spectrum at each corresponding frequency point divided by the sum of the squared amplitudes is calculated to obtain the spectrum similarity coefficient. The noise interference level is determined by the signal-to-noise ratio deterioration index. When the deterioration index exceeds the first threshold, it is marked as mild interference. When it exceeds the second threshold, it is marked as severe interference. The spectrum similarity coefficient and peak offset amplitude are combined to generate a feature matching comparison result, which includes the similarity coefficient value, the peak offset exceeding limit mark and the noise interference level assessment result.

[0061] Specifically, speed fluctuation monitoring is an important means of gearbox fault diagnosis, which can reflect the changes in the dynamic characteristics of the transmission system.

[0062] In one possible implementation, a high-precision speed sensor collects gear shaft speed signals in real time. The sampling frequency is typically set at least 10 times the gear meshing frequency to ensure that instantaneous speed changes are captured. The difference between the instantaneous speed value and the average speed value actually reflects speed fluctuations, which can be caused by various factors such as load changes, bearing failures, and gear wear.

[0063] Specifically, when performing spectral analysis on the difference sequence, certain specific frequency components will be found. Under normal circumstances, these frequency peaks should appear at the theoretically calculated positions. For example, the gear meshing frequency should be equal to the speed multiplied by the number of teeth. However, when a gearbox fails, the actual peak frequency will deviate from the theoretical value. The magnitude of this offset directly reflects the severity of the fault. The larger the offset, the further the system state deviates from normal operating conditions. The establishment of a baseline feature library is the basis for achieving accurate fault diagnosis. This feature library contains the standard spectrum characteristics of the gearbox at different speeds and different load conditions. These characteristics are obtained through long-term monitoring and statistical analysis of healthy gearboxes. When it is necessary to evaluate the current gearbox status, the system will retrieve the corresponding standard spectrum from the feature library based on the real-time speed value. This can eliminate the impact of speed changes on the diagnostic results and improve the accuracy of the diagnosis.

[0064] It should be noted that the noise generated by wear particles has a broadband characteristic, and it will superimpose additional energy at multiple frequency points. By comparing the amplitude difference between the current spectrum and the standard spectrum at each frequency point one by one, the frequency range affected by the noise interference can be identified. The calculation of the total amplitude deviation provides a quantitative basis for the subsequent signal-to-noise ratio analysis. The larger the deviation, the more serious the noise interference. The calculation method of the signal-to-noise ratio deterioration index embodies a relative evaluation concept. The energy of the undisturbed frequency point is regarded as a useful signal, and the deviation energy of the interfered frequency point is regarded as noise. This division method can accurately reflect the impact of wear particles on signal quality. By comparing with the standard signal-to-noise ratio in the benchmark feature library, the deterioration index obtained can intuitively indicate the degree of degradation of the current signal quality.

[0065] Preferably, the calculation of the spectrum similarity coefficient adopts the normalized cross-correlation method. The amplitudes of the current spectrum and the corresponding frequency points of the standard spectrum are multiplied and summed, and then divided by the geometric mean of the sum of the squares of the two spectrum amplitudes. The coefficient value obtained is between 0 and 1. The closer the similarity is to 1, the closer the current state is to normal, and the closer it is to 0, the more serious the deviation. The generation of feature matching comparison results adopts a multi-index comprehensive evaluation method. By setting different thresholds, the signal-to-noise ratio deterioration index is divided into two levels: mild and severe. This grading method makes it easier for maintenance personnel to quickly judge the urgency of the fault. The peak offset limit mark provides another dimension of evaluation. When the offset exceeds the allowable range, even if the signal-to-noise ratio deterioration is not serious, it needs to be taken seriously.

[0066] S105. If the feature matching result shows that the deviation between the high-frequency component diffusion distortion feature of the spectrum and the reference feature library exceeds a threshold, the tooth surface fatigue crack noise feature is added to the feature library to obtain a target feature library.

[0067] Based on the feature matching results, high-frequency components above ten times the meshing frequency are extracted from the current spectrum. The amplitude differences of these high-frequency components at adjacent frequency points are calculated, and the standard deviation of these differences is calculated. If the standard deviation exceeds a preset ratio of the average amplitude and the amplitude distribution width exceeds the normal range, a high-frequency component diffusion distortion feature is determined to be present. The amplitude distribution width and standard deviation of the high-frequency component diffusion distortion feature are subtracted from the corresponding parameters of the standard meshing spectrum template in the baseline feature library to obtain a deviation value. This deviation is then compared with typical fault spectrum samples and the wear stage characteristic parameter range. If the deviation value exceeds a preset threshold, the current spectrum feature is identified as a new fault feature caused by tooth surface fatigue cracks. Based on the identified tooth surface fatigue crack fault feature, the start frequency, cutoff frequency, amplitude distribution width, and standard deviation of the high-frequency component diffusion distortion are extracted from the current spectrum to form a fatigue crack noise feature vector. The vector is then annotated with the acquisition time and operating condition parameters. The fatigue crack noise feature vector and its annotated information are added to the set of fault spectrum samples in the baseline feature library. The upper and lower limits of the wear stage characteristic parameter range are also updated to obtain a target feature library containing fatigue crack fault modes.

[0068] Specifically, high-frequency component diffusion distortion is an important feature in gearbox fault diagnosis, which reflects the influence of tooth surface micro-damage on the vibration signal.

[0069] In one possible implementation, when a fatigue crack forms on a tooth surface, each time the crack edge engages the meshing surface, it generates a tiny impact that excites a wide-band vibration response. Unlike the harmonics generated by normal meshing, the vibration energy caused by the crack is diffusely distributed across a high-frequency range, rather than being concentrated at a specific frequency.

[0070] Specifically, identifying diffusion distortion characteristics requires focusing on the energy distribution pattern in the high-frequency range. Under normal circumstances, high-frequency components are primarily concentrated in the higher harmonics of the meshing frequency, while the energy in the frequency regions between the harmonics is relatively low. However, when fatigue cracks are present, the irregular contact at the crack edges produces random high-frequency vibrations, resulting in a significant energy distribution in the frequency regions between the harmonics. This degree of diffusion can be quantified by calculating the standard deviation of the amplitude differences between adjacent frequency points. The ratio of the standard deviation to the average amplitude is a key indicator for determining diffusion distortion.

[0071] For example, within the frequency range of 10kHz to 20kHz, if the standard deviation exceeds 30% of the average amplitude and the amplitude distribution width exceeds 2kHz, significant diffusion distortion can be determined. This judgment method comprehensively considers the degree of energy distribution dispersion and coverage, and can effectively distinguish normal high-frequency harmonics from abnormal diffusion characteristics.

[0072] It should be noted that the establishment of the benchmark feature library is a process of continuous accumulation. The standard meshing spectrum template records the normal spectrum characteristics of the gearbox under different working conditions. These templates are obtained through statistical analysis of a large number of healthy gearboxes. The typical fault spectrum samples contain spectrum characteristics of various known fault modes, such as pitting, spalling, broken teeth, etc. The wear stage characteristic parameter range defines the characteristic parameter value range for each stage from initial wear to severe wear. The calculation of the deviation value adopts a multi-parameter comprehensive evaluation method. The currently detected diffusion distortion characteristic parameters are compared with the corresponding parameters in the benchmark feature library. Not only the difference in single parameters is compared, but also the overall deviation of the parameter combination is considered. When the deviation value exceeds the preset threshold, it means that the current feature has exceeded the range of known fault modes and is likely to be a new form of fault such as fatigue cracks.

[0073] Preferably, the construction of the fatigue crack noise feature vector requires the extraction of parameters that best reflect the crack characteristics. The starting frequency and cutoff frequency define the frequency range of diffusion distortion, which is usually wider than other fault types. The amplitude distribution width reflects the degree of energy diffusion, and the standard deviation value quantifies the unevenness of the distribution. These parameters are combined into a feature vector and the speed, load and other operating parameters during acquisition are marked to provide complete information for subsequent fault identification. The update of the target feature library is a dynamic optimization process. The newly identified fatigue crack features are not simply added to the library, but are analyzed in association with existing features. By updating the upper and lower limits of the characteristic parameter range in the wear stage, the coverage of the feature library can be expanded to enable it to identify more types of faults.

[0074] S106. Identify the energy distribution of different wear stages based on the target feature library, perform noise filtering based on the tooth surface contact noise proportional distribution pattern, and determine the fault signal classification result by comparing the impact of wear depth on frequency offset.

[0075] Based on the wear stage characteristic parameter range in the target feature library, the total energy of the current spectrum in the three frequency bands below the meshing frequency, from the meshing frequency to ten times its frequency, and above ten times its frequency is calculated. The percentage of energy in each frequency band to the total energy is compared with the energy distribution templates corresponding to different wear stages in the feature library, and the Euclidean distance is calculated as the matching value. Based on the wear stage with the smallest matching value, the corresponding tooth surface contact noise ratio distribution data is extracted from the target feature library. The amplitude of each frequency point is attenuated according to the noise ratio. The higher the noise ratio, the greater the attenuation coefficient, and the filtered characteristic spectrum is obtained. The actual frequency values ​​of the meshing frequency and its harmonics are extracted from the filtered characteristic spectrum, and the offset from the theoretical frequency is calculated. The time series of the offset is differentiated to obtain the offset change rate. The correlation coefficient is calculated between the change rate series and the standard change rate templates corresponding to different wear degrees stored in the target feature library. Through the wear degree corresponding to the maximum value of the correlation coefficient and the obtained wear stage matching value, if both point to the early parameter range, a slight wear mark is output; if they point to the middle parameter range, a moderate wear mark is output; if they point to the late parameter range, a severe wear mark is output; if the feature library matches the fatigue crack feature, the fatigue crack fault type mark is output to form the fault signal classification result.

[0076] Specifically, frequency band division of energy distribution is a key technology in gearbox fault diagnosis.

[0077] In one possible implementation, the frequency bands are divided based on the physical characteristics of gear meshing. The low-frequency band below the meshing frequency primarily contains the rotational frequency and its lower harmonics, reflecting the overall motion of the shaft system. The mid-frequency band, from the meshing frequency to ten times its frequency, contains the meshing frequency and its main harmonics, representing the primary energy concentration of normal meshing vibrations. The high-frequency band above ten times the frequency primarily contains high-frequency components excited by local defects.

[0078] Specifically, the energy distribution in different wear stages shows obvious regular changes. In the case of mild wear, the energy is mainly concentrated in the mid-frequency band, while the low-frequency and high-frequency bands account for a smaller proportion of energy. For example, the mid-frequency band accounts for 75% of the total energy, the low-frequency band accounts for 15%, and the high-frequency band accounts for 10%. As wear intensifies, the energy in the high-frequency band gradually increases. In moderate wear, the high-frequency band may rise to 20%, and in severe wear, it may reach more than 35%. This energy migration reflects the evolution of the tooth surface state from smooth to rough. The calculation of Euclidean distance provides a quantitative indicator for matching assessment. The energy percentages of the three currently detected frequency bands are combined into a three-dimensional vector, and the distance is calculated with the standard vector of each wear stage in the feature library. The smaller the distance, the higher the matching degree and the stronger the similarity. The advantage of this method is that it comprehensively considers the energy distribution of multiple frequency bands, avoiding the one-sidedness of a single indicator.

[0079] It's important to note that the tooth contact noise distribution data reflects the contribution of noise to different frequency components. During gear meshing, in addition to regular meshing vibrations, there's also random noise caused by factors like surface roughness and lubrication conditions. Through extensive experimental analysis, we can determine the typical noise proportions at each frequency point under varying wear conditions.

[0080] For example, at the third harmonic of the meshing frequency, the noise contribution may be only 5% with mild wear, but may rise to 30% with severe wear. Filtering employs an adaptive attenuation method. The attenuation coefficient is set based on the noise contribution. For a 10% noise contribution, the attenuation coefficient can be set to 0.9, while for a 30% noise contribution, the attenuation coefficient drops to 0.7. This processing method preserves useful fault characteristic information while effectively suppressing noise interference, improving the accuracy of subsequent analysis.

[0081] Preferably, the rate of change of the frequency offset is calculated using the method of numerical differentiation. For offset data collected continuously at multiple time points, the offset difference between adjacent time points is calculated and divided by the time interval to obtain a rate of change sequence. Under normal circumstances, the rate of change should fluctuate within a small range; when abnormal wear occurs, the rate of change will show a sudden increase or a continuous upward trend. The comprehensive judgment of fault classification makes full use of multi-dimensional information. The energy distribution reflects the overall severity of the fault, while the frequency offset change rate reveals the speed of fault development. When the two indicators are inconsistent, the system will give priority to the more serious situation.

[0082] For example, if the energy distribution indicates moderate wear, but the rate of change in the offset reaches the level of severe wear, the wear is ultimately determined to be severe. This conservative strategy helps to promptly detect potential serious failures. In particular, sudden failures such as fatigue cracks have characteristics that differ significantly from those of conventional wear, and accurate identification can be achieved through precise matching of the feature library.

[0083] S107. Extract fault characteristic parameters from the classification results, analyze the correlation between the enhancement of the high-frequency component of the spectrum and the nonlinear interference of the vibration signal based on the coupling relationship between the gearbox load change and the speed fluctuation, and determine the abnormality degree of the operating state.

[0084] The corresponding spectral feature data is extracted from the fault signal classification results. The surface roughness equivalent value is calculated by dividing the high-frequency energy fraction by the mid-frequency energy fraction and then multiplying it by the baseline roughness value. The wear depth estimate is obtained by summing the frequency offsets at each moment and dividing them by the sampling interval and the number of teeth. The crack growth trend indicator is determined from the high-frequency diffusion width change rate corresponding to the fatigue crack fault type. Based on the three fault characteristic parameters obtained, real-time data from the gearbox load sensor and speed sensor are simultaneously collected. The load change rate is calculated as the ratio of the load difference to the load mean at adjacent moments, and the speed fluctuation rate is calculated as the ratio of the speed standard deviation to the speed mean. The two are multiplied to obtain the coupling coefficient. Based on the coupling coefficient, when it exceeds a preset threshold, the amplitudes of all frequency components above ten times the meshing frequency are extracted from the current spectrum. The average growth rate of these amplitudes is calculated as the high-frequency enhancement degree. The instantaneous frequency of the vibration signal is extracted through a Hilbert transform, and the variance of the instantaneous frequency is calculated as the nonlinear interference intensity. The strength of the correlation is determined by calculating the Pearson correlation coefficient between the high-frequency enhancement and the nonlinear interference intensity. The crack propagation trend index, coupling coefficient, and correlation coefficient are multiplied by preset weights and then summed to obtain the abnormal operation status index. The degree of abnormal operation status is judged based on the abnormal index.

[0085] Specifically, the calculation of the surface roughness equivalent value reflects the intrinsic connection between the spectrum energy distribution and the tooth surface micromorphology.

[0086] In one possible implementation, as tooth wear intensifies, high-frequency vibration components increase significantly, while normal meshing vibration in the mid-frequency range remains relatively stable. This change can be quantitatively characterized by calculating the energy ratio between the high-frequency and mid-frequency bands. A baseline roughness value is typically obtained through field measurements on new gears, for example, an Ra value of 0.8 microns. When the energy ratio reaches 2, the equivalent roughness may reach 1.6 microns, indicating significant wear on the tooth surface.

[0087] Specifically, the estimation of wear depth needs to take into account the cumulative effect of frequency offset. Each time the gears mesh, the change in tooth profile caused by wear will cause the actual meshing point to deviate from the theoretical position, thereby causing an offset in the meshing frequency. By accumulating the continuously collected frequency offsets, the cumulative degree of wear can be reflected. Dividing by the sampling time interval is to obtain the wear rate per unit time, and then dividing by the number of teeth is to average the total wear on each tooth to obtain the average wear depth of a single tooth. The determination of the crack growth trend indicator is based on the dynamic changes of the high-frequency diffusion phenomenon. During the growth of fatigue cracks, the stress concentration at the crack tip will excite an increasingly wider frequency range. By monitoring the rate of change of the diffusion width, the crack growth rate can be determined.

[0088] For example, the initial crack propagation width may be only 500Hz, and after a period of time it expands to 1500Hz, with a change rate of 200Hz per hour. This value can be used as a quantitative indicator of the crack propagation trend.

[0089] It's important to note that the coupling effect between load and speed can amplify fault characteristics. The load variation rate reflects the degree of load fluctuation on the gear. When the load suddenly increases, the contact stress on the tooth surface rises instantaneously, potentially accelerating the propagation of existing defects. The speed fluctuation rate affects the intensity of the meshing shock. Unstable speed can cause additional dynamic loads during gear meshing. The product of these two factors, serving as the coupling coefficient, comprehensively reflects the severity of the operating conditions. The Hilbert transform offers unique advantages in extracting instantaneous frequency. By applying the Hilbert transform to the vibration signal, we can obtain the signal's analytical form, allowing us to calculate the instantaneous phase and frequency at each moment. Under normal circumstances, the instantaneous frequency should fluctuate slightly around the average frequency; however, in the presence of nonlinear interference, the instantaneous frequency can experience large jumps. Calculating the variance of the instantaneous frequency can quantify this instability; a larger variance indicates stronger nonlinear interference.

[0090] Preferably, the calculation of the Pearson correlation coefficient reveals the statistical relationship between high-frequency enhancement and nonlinear interference. When the correlation coefficient is close to 1, it means that the two are strongly positively correlated, that is, the more obvious the high-frequency enhancement, the more serious the nonlinear interference, which usually means that the fault has developed to a more serious stage. A correlation coefficient close to 0 indicates that the two are relatively independent, which may be a manifestation of different fault mechanisms. The comprehensive evaluation of the operating status abnormality index adopts a weighted summation method, and the setting of the weight reflects the importance of different indicators. The crack propagation trend indicator has a higher weight because the danger of crack failure is the greatest; the coupling coefficient is second, reflecting the influence of working conditions; the correlation coefficient has a relatively small weight and serves as an auxiliary criterion. Through this multi-indicator fusion method, the degree of abnormality in the operating status of the gearbox can be comprehensively and accurately evaluated.

[0091] S108. Generate status alarm information based on the degree of abnormality in the operating state, obtain status assessment data of tooth surface fatigue crack noise by correlating the alarm information with background data caused by wear particles, and identify minor wear fault results of the gearbox.

[0092] According to the numerical range of the abnormal operating status index, a mild alarm message is generated when the index exceeds the first preset threshold, a moderate alarm message is generated when it exceeds the second preset threshold, and a severe alarm message is generated when it exceeds the third preset threshold. Each alarm message contains the abnormal index value, the fault type identifier, and the acquisition time. Based on the acquisition time and abnormal index value of the alarm message, the spectrum data of the corresponding time period is extracted, the broadband noise component generated by the collision of wear particles in the spectrum is identified, the ratio of the broadband noise energy to the characteristic frequency energy corresponding to the fault type in the alarm message is calculated, and the correlation coefficient of the ratio sequence at consecutive moments and the standard sequence is calculated as the correlation value. Through the alarm records with correlation values ​​greater than the preset value, the high-frequency diffusion characteristic component of the fatigue crack in the corresponding spectrum is extracted, and the ratio of the characteristic component energy to the background noise energy is calculated as the signal-to-noise ratio. At the same time, the amplitude variation coefficient of the characteristic component in continuous sampling is calculated as the stability index. The signal-to-noise ratio, stability index and crack extension trend index are combined to form accurate status assessment data. Accurate state assessment data is used to determine the degree of fault. If the signal-to-noise ratio exceeds the minor wear threshold and the coefficient of variation of the stability index is less than the preset upper limit, and the crack growth trend index is in the initial range, it is identified as a minor wear fault of the gearbox.

[0093] Specifically, in gearbox vibration monitoring, an abnormality index of 0.3 triggers a minor alarm, 0.6 generates a moderate alarm, and 0.8 generates a severe alarm. Each alarm message precisely records the abnormality index value of 0.65, the fault type identifier "bearing wear," and the acquisition time "2024-03-15 14:30:25," forming a complete fault traceability chain. This hierarchical alarm mechanism beneficially avoids false alarms and missed alarms, ensuring that operations and maintenance personnel can promptly respond to real fault threats. Spectral data extraction and broadband noise identification are key steps in fault diagnosis.

[0094] In one possible implementation, the system extracts spectrum data for the five minutes before and after the alarm at 14:30:25 and analyzes the vibration signal using a fast Fourier transform. The broadband noise generated by wear particle collisions typically occurs in the 500-2000 Hz frequency range and exhibits a continuous energy distribution.

[0095] Specifically, the system calculated the total broadband noise energy to be 120 dB, while the energy at the bearing's characteristic frequency of 1200 Hz was 90 dB, resulting in a ratio of 1.33. This ratio reflects the severity of wear, with higher values ​​indicating more intense wear particle activity. Correlation analysis quantifies fault development trends through interrelationships.

[0096] For example, the ratio sequence of 10 consecutive moments is

[0097] The correlation coefficient calculated with the standard wear development sequence is 0.85. When the correlation exceeds the preset value of 0.8, indicating a high degree of consistency between the fault development pattern and typical wear characteristics, the system automatically triggers the in-depth analysis process. Identifying the high-frequency propagation characteristics of fatigue cracks relies on precise spectrum analysis technology.

[0098] It should be noted that fatigue cracks appear as diffusion of high-frequency components in the frequency spectrum, usually in the 3000-8000 Hz frequency band.

[0099] For example, a characteristic component near 5000 Hz was extracted, with an energy of 45 dB and a background noise energy of 25 dB, resulting in a calculated signal-to-noise ratio of 1.8. Furthermore, the amplitudes of this characteristic component over 20 consecutive sampling times were [0.8, 0.82, 0.79, 0.85, 0.81, ...], with a coefficient of variation of 0.12, which was used as a stability indicator. The comprehensive evaluation of precise condition assessment data demonstrates the advantages of multi-dimensional fault diagnosis.

[0100] Specifically, when the signal-to-noise ratio (SNR) of 1.8 exceeds the minor wear threshold of 1.5, the stability index's coefficient of variation (CVO) of 0.12 is less than the preset upper limit of 0.15, and the crack growth trend index (CTI) of 0.25 is within the initial range of 0.2-0.3, the system comprehensively identifies a minor gearbox wear fault. This multi-index fusion approach effectively avoids the limitations of a single indicator and improves the accuracy and reliability of fault identification.

[0101] The present invention provides a wind turbine generator fault monitoring system based on voiceprint recognition, which mainly includes:

[0102] The signal acquisition module is used to collect gearbox soundprint signal data in real time, record the signal fluctuations caused by gear surface wear, and obtain the original signal data set containing the enhanced characteristics of the high-frequency components of the spectrum;

[0103] The time-frequency analysis module is used to apply time-frequency analysis to the original signal data set, identify nonlinear interference in the vibration signal, perform multi-scale decomposition, separate the narrowband characteristics of the initial wear spectrum and the harmonic components of the mid-term wear spectrum, and obtain a set of spectrum components;

[0104] The feature extraction module is used to extract characteristic parameters related to abnormal fluctuations in modulation depth from the set of spectral components, identify the variation pattern of the meshing frequency of the gear pair, and determine the distribution pattern of the tooth surface contact noise ratio;

[0105] The feature matching module is used to obtain the spectrum peak offset amplitude through speed fluctuation monitoring, compare the difference with the benchmark feature library data, and determine the degree of influence of the noise caused by wear particles on feature extraction. The signal-to-noise ratio deterioration index is used to obtain the feature matching comparison results;

[0106] A feature library updating module is used to add the tooth surface fatigue crack noise feature to the feature library to obtain a target feature library if the feature matching result shows that the deviation between the spectral high-frequency component diffusion distortion feature and the reference feature library exceeds a threshold;

[0107] The fault classification module is used to identify the energy distribution of different wear stages based on the target feature library, perform noise filtering based on the proportional distribution pattern of the tooth surface contact noise, and determine the fault signal classification result by comparing the effect of wear depth on frequency offset;

[0108] The state analysis module is used to extract fault characteristic parameters from the classification results, analyze the correlation between the enhancement of the high-frequency component of the spectrum and the nonlinear interference of the vibration signal based on the coupling relationship between the gearbox load change and the speed fluctuation, and determine the degree of abnormality of the operating state;

[0109] The alarm generation module is used to generate status alarm information based on the degree of abnormality in the operating status. By correlating the alarm information with the background data caused by wear particles, the status assessment data of tooth surface fatigue crack noise is obtained and the results of minor wear faults of the gearbox are identified.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wind turbine fault monitoring method based on voiceprint recognition, characterized in that: The method comprises: Gearbox soundprint signals are collected, signal fluctuations caused by gear surface wear are recorded, and an original signal data set containing enhanced features of high-frequency components of the spectrum is obtained; time-frequency analysis is performed on the original signal data set to decompose the vibration signal, separate the narrowband characteristics of the initial wear stage and the harmonic components of the mid-term wear stage, and obtain a spectrum component set; the modulation depth change rate, carrier frequency offset, and sideband energy distribution ratio are extracted from the spectrum component set to determine the tooth surface contact noise distribution ratio; the spectrum peak offset amplitude is obtained by speed fluctuation monitoring, compared with the benchmark feature library, and the spectrum similarity coefficient and signal-to-noise ratio deterioration index are calculated to obtain feature matching comparison results; according to the The feature matching and comparison results are used to identify the diffusion distortion characteristics of the high-frequency components of the spectrum, and the baseline feature library is updated as the target feature library; based on the target feature library, noise filtering is performed in combination with the tooth surface contact noise ratio distribution, the similarity of the influence of wear depth on frequency offset is calculated, and the fault classification results of slight wear, moderate wear, severe wear, and fatigue cracks are determined; the surface roughness equivalent value, wear depth estimation value, and crack propagation trend index are extracted from the fault classification results, the coupling relationship between load change and speed fluctuation is analyzed, and the operating status abnormality index is calculated; based on the operating status abnormality index, an alarm message containing a fault type identifier is generated.

2. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The method of collecting the gearbox soundprint signal, recording the signal fluctuation caused by gear surface wear, and obtaining the original signal data set containing the enhanced features of the high-frequency components of the spectrum includes: The gearbox sound pressure level numerical sequence is collected by an acoustic sensor and converted into soundprint time series data; the soundprint time series data is transformed into the frequency domain to extract the spectral component amplitude distribution in a preset frequency range; based on the spectral component amplitude distribution, the amplitude difference between adjacent frequency points is calculated, the high-frequency component enhancement feature is identified, and the starting frequency, cutoff frequency, and amplitude peak are recorded; the soundprint time series data segments associated with the high-frequency component enhancement feature are stored to construct the original signal data set.

3. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The time-frequency analysis is performed on the original signal data set to decompose the vibration signal, separate the narrowband characteristics of the initial wear stage and the harmonic components of the mid-term wear stage, and obtain a set of spectral components, including: The vibration signal of the original signal data set is transformed in the short-time frequency domain, the time window is divided, the spectrum amplitude distribution of each window is obtained, and a time-frequency spectrum matrix is ​​constructed; according to the time-frequency spectrum matrix, the frequency point amplitude change ratio is calculated, the nonlinear interference frequency point is identified, and an interference marker matrix is ​​generated; according to the interference marker matrix, the interference signal component is eliminated, the remaining signal is decomposed at multiple scales, and the narrowband characteristic signal with a frequency bandwidth less than a preset value and the harmonic characteristic signal with an integer multiple of the meshing frequency are extracted and stored as the spectrum component set.

4. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The step of extracting the modulation depth change rate, the carrier frequency offset, and the sideband energy distribution ratio from the set of spectrum components to determine the tooth surface contact noise distribution ratio includes: The meshing frequency and the sideband frequency are identified from the set of spectral components, the ratio of the carrier frequency amplitude to the sum of the sideband amplitudes is calculated as the modulation depth, and the ratio of the modulation depth difference of consecutive time periods to the time interval is calculated as the modulation depth change rate; based on the modulation depth change rate, the difference between the carrier frequency and the reference frequency is calculated as the carrier frequency offset, and the ratio of the sum of the squares of the sideband amplitudes to the square of the carrier frequency amplitude is calculated as the sideband energy distribution ratio; based on the carrier frequency offset and the sideband energy distribution ratio, a mapping relationship table is constructed, the noise ratio is counted, and the tooth surface contact noise ratio distribution is generated.

5. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The step of extracting the modulation depth change rate, the carrier frequency offset, and the sideband energy distribution ratio from the set of spectrum components to determine the tooth surface contact noise distribution ratio includes: According to the modulation depth change rate and the carrier frequency offset, the frequency drift trajectory of the fundamental frequency multiplication component is tracked and the drift amplitude is calculated; the transient impact pulse is extracted from the vibration signal, and the pulse interval distribution and phase difference sequence are calculated; according to the phase difference sequence, the broadband noise component is identified, and the energy proportion of metal friction noise and aerodynamic noise is calculated; according to the drift amplitude, the phase difference sequence, and the noise energy proportion, a gear pair operating state characteristic map including meshing frequency drift, phase jump, and noise energy distribution is generated.

6. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The spectrum peak offset amplitude is obtained by monitoring the speed fluctuation, compared with the reference feature library, and the spectrum similarity coefficient and the signal-to-noise ratio deterioration index are calculated to obtain the feature matching comparison results, including: According to the gearbox speed signal, a speed difference sequence is calculated, and the spectrum peak frequency offset is extracted as the spectrum peak offset amplitude; according to the spectrum peak offset amplitude, a standard spectrum is extracted from a reference feature library, the amplitude difference is calculated, and the noise interference frequency point is identified; according to the interference frequency point, the ratio of signal power to noise power is calculated as the signal-to-noise ratio, and the ratio of the spectrum amplitude product to the square sum is calculated as the spectrum similarity coefficient, and the feature matching comparison result including the spectrum similarity coefficient and the signal-to-noise ratio deterioration index is generated.

7. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The step of identifying the diffusion distortion feature of the high frequency component of the spectrum according to the feature matching and comparison result, and updating the reference feature library to the target feature library, comprises: Based on the feature matching and comparison results, the standard deviation of the high-frequency component amplitude difference is extracted to identify the diffusion distortion feature; based on the diffusion distortion feature, the deviation from the baseline feature library is calculated, the starting frequency, cutoff frequency, and amplitude distribution width are extracted, and the fatigue crack noise feature vector is generated; the fatigue crack noise feature vector is added to the baseline feature library, the feature parameter range is updated, and the target feature library is generated.

8. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The method of performing noise filtering based on the target feature library and the tooth surface contact noise ratio distribution, calculating the similarity of the effect of wear depth on frequency offset, and determining fault classification results of slight wear, moderate wear, severe wear, and fatigue cracks includes: Based on the target feature library, the energy proportion of each frequency band of the spectrum is calculated and compared with the wear stage template to obtain the matching degree; based on the proportional distribution of the tooth surface contact noise, the spectrum amplitude is attenuated to generate a filtered spectrum; the meshing frequency offset is extracted from the filtered spectrum, and the correlation coefficient between the offset change rate and the standard template is calculated to determine the fault classification result.

9. The wind turbine generator fault monitoring method based on voiceprint recognition according to claim 1, characterized in that: The extracting of the surface roughness equivalent value, the wear depth estimation value, and the crack growth trend index from the fault classification result, analyzing the coupling relationship between load change and speed fluctuation, and calculating the operation state abnormality index includes: According to the fault classification result, the ratio of high-frequency to medium-frequency energy is calculated to generate the surface roughness equivalent value, the sum of frequency offsets is calculated to generate the wear depth estimation value, and the high-frequency diffusion width change rate is calculated to generate the crack propagation trend index; based on the obtained surface roughness equivalent value, wear depth estimation value, and crack propagation trend index, the coupling coefficient is calculated according to the load change rate and the speed fluctuation rate, the high-frequency enhancement degree and the nonlinear interference intensity are extracted, the correlation coefficient is calculated, and the operating status abnormality index is generated.

10. A wind turbine fault monitoring system based on voiceprint recognition, characterized in that: The system comprises: The signal acquisition module is used to collect gearbox soundprint signal data in real time, record the signal fluctuations caused by gear surface wear, and obtain the original signal data set containing the enhanced characteristics of the high-frequency components of the spectrum; The time-frequency analysis module is used to apply time-frequency analysis to the original signal data set, identify nonlinear interference in the vibration signal, perform multi-scale decomposition, separate the narrowband characteristics of the initial wear spectrum and the harmonic components of the mid-term wear spectrum, and obtain a set of spectrum components; The feature extraction module is used to extract characteristic parameters related to abnormal fluctuations in modulation depth from the set of spectral components, identify the variation pattern of the meshing frequency of the gear pair, and determine the distribution pattern of the tooth surface contact noise ratio; The feature matching module is used to obtain the spectrum peak offset amplitude through speed fluctuation monitoring, compare the difference with the benchmark feature library data, and determine the degree of influence of the noise caused by wear particles on feature extraction. The signal-to-noise ratio deterioration index is used to obtain the feature matching comparison results; A feature library updating module is used to add the tooth surface fatigue crack noise feature to the feature library to obtain a target feature library if the feature matching result shows that the deviation between the spectral high-frequency component diffusion distortion feature and the reference feature library exceeds a threshold; The fault classification module is used to identify the energy distribution of different wear stages based on the target feature library, perform noise filtering based on the proportional distribution pattern of the tooth surface contact noise, and determine the fault signal classification result by comparing the effect of wear depth on frequency offset; The state analysis module is used to extract fault characteristic parameters from the classification results, analyze the correlation between the enhancement of the high-frequency component of the spectrum and the nonlinear interference of the vibration signal based on the coupling relationship between the gearbox load change and the speed fluctuation, and determine the degree of abnormality of the operating state; The alarm generation module is used to generate status alarm information based on the degree of abnormality in the operating status. By correlating the alarm information with the background data caused by wear particles, the status assessment data of tooth surface fatigue crack noise is obtained and the results of minor wear faults of the gearbox are identified.

Images