A method and system for diagnosing vibration faults of a generator

Abstract

The application relates to the technical field of fault diagnosis, and discloses a generator vibration fault diagnosis method and system. The method comprises the following steps: obtaining original vibration signals of a generator in operation, performing standardization processing, space position deviation compensation and temperature and humidity coupling correction to obtain a preliminary corrected data set; performing discrete wavelet transform decomposition components, extracting time-frequency features to obtain layered frequency domain features; performing threshold denoising, inverse wavelet transform and boundary smoothing correction to obtain a corrected signal set; after extracting a time domain feature set, a potential fault component is located through fault mode library matching, and finally, a vibration anomaly root cause is determined through frequency anomaly matching and component correlation analysis. The method can realize accurate diagnosis and root cause positioning of generator vibration faults, and improves equipment operation safety and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method and system for diagnosing generator vibration faults. Background Technology

[0002] Currently, as the core equipment for energy conversion in the power system, the operational stability of generators directly determines the power grid's power supply security and energy supply efficiency. Abnormal vibration is the core external manifestation of mechanical and electrical faults in generators. Precise diagnostic technology for generator vibration faults is a core link in the power generation equipment fault prediction and health management system, and also a key technical support for achieving proactive equipment operation and maintenance and avoiding unplanned downtime accidents.

[0003] In existing technologies, generator vibration fault diagnosis often employs fixed vibration threshold alarms or conventional Fourier spectrum analysis schemes. These methods only perform basic noise reduction and simple feature extraction on the acquired raw vibration signals, and obtain fault diagnosis results by comparing the extracted features. However, during generator operation, vibration signals are affected by various external factors, such as load changes and ambient temperature fluctuations, resulting in highly unstable signals. This instability makes it difficult to distinguish between normal vibrations and fault precursors using traditional methods alone. Furthermore, this signal instability affects the determination of the root cause of the fault, because the specific source of abnormal vibration cannot be accurately pinpointed, making it difficult to determine whether it is caused by loose mechanical parts or wear of other components, thus preventing precise identification of the root cause of the fault.

[0004] In summary, existing technologies struggle to capture subtle characteristics of early faults from background noise in complex operating environments, making it difficult to accurately pinpoint the root cause of the fault and meet the maintenance requirements for long-term safe and stable operation of generators under complex operating conditions. Summary of the Invention

[0005] This invention provides a generator vibration fault diagnosis method and system to solve the problems of poor anti-interference ability and low feature extraction accuracy in existing generator vibration fault diagnosis, which leads to difficulty in fault tracing, and to achieve accurate diagnosis and root cause location of vibration faults under complex working conditions.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a generator vibration fault diagnosis method, comprising: The original vibration signal, real-time temperature data, and real-time humidity data during generator operation are acquired, and the original vibration signal is corrected for environmental interference to obtain a preliminary correction dataset. Perform discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components; The time-frequency feature vector is extracted from the set of detail components and the set of approximate components. The time-frequency feature vector is then recombined to obtain the hierarchical frequency domain feature, wherein the hierarchical frequency domain feature includes the set of hierarchical detail components corresponding to each decomposition level. A noise filtering operation is performed on the hierarchical detail component set to obtain a denoised detail component set. The inverse wavelet transform is performed on the denoised detail component set and the approximate component set to obtain the reconstructed vibration signal sequence, and the boundary correction is performed on the reconstructed vibration signal sequence to obtain the correction signal set; The set of correction signals is matched with operating parameters to obtain a set of fluctuation features. The set of fluctuation features is analyzed to obtain amplitude features. The amplitude features are then identified by periodic distribution to obtain a set of time-domain features. The amplitude deviation sequence is extracted from the time-domain feature set, and the amplitude deviation sequence is searched and matched with the preset fault mode library to determine the location of potential faulty components. Frequency anomaly matching is performed based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

[0007] Secondly, the present invention provides a generator vibration fault diagnosis system, comprising: The signal correction module is used to acquire the original vibration signal, real-time temperature data and real-time humidity data during generator operation, and to correct the original vibration signal for environmental interference to obtain a preliminary correction dataset. The wavelet decomposition module is used to perform discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components. The frequency domain feature generation module is used to extract time-frequency feature vectors based on the set of detail components and the set of approximate components, and to reassemble the time-frequency feature vectors to obtain hierarchical frequency domain features, wherein the hierarchical frequency domain features include hierarchical detail component sets corresponding to each decomposition level; The noise suppression module is used to perform noise filtering on the hierarchical detail component set to obtain a denoised detail component set. The signal reconstruction module is used to perform inverse wavelet transform on the denoised detail component set and the approximate component set to obtain a reconstructed vibration signal sequence, and to perform boundary correction on the reconstructed vibration signal sequence to obtain a correction signal set; The time-domain feature extraction module is used to match the operating parameters of the correction signal set to obtain a fluctuation feature set, perform amplitude analysis on the fluctuation feature set to obtain amplitude features, and identify the periodic distribution of the amplitude features to obtain a time-domain feature set. The potential fault location module is used to extract the amplitude deviation sequence from the time domain feature set, and search and match the amplitude deviation sequence with a preset fault mode library to determine the location of the potential faulty component. The fault source localization module is used to perform frequency anomaly matching based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

[0008] By acquiring the original vibration signal, real-time temperature data, and real-time humidity data during generator operation, and correcting the original vibration signal for environmental interference, a preliminary corrected dataset is obtained. Compared with existing technologies, this invention has the following advantages: (1) This invention performs standardized processing on the original vibration signal, sensor spatial offset compensation and temperature and humidity coupling correction. First, it eliminates the difference between the dimensions and the sampling frequency, and then constructs a compensation matrix through the spatial offset vector. Combined with the temperature and humidity coupling factor, it generates a comprehensive correction vector to eliminate false vibration components caused by environmental and installation deviations. This solves the problems of weak signal anti-interference ability and high false alarm rate in the existing technology, and improves the authenticity and reliability of the basic data.

[0009] (2) This invention performs multi-scale decomposition of the correction signal by discrete wavelet transform, maps the frequency band interval and calculates the energy spectral density to obtain layered frequency domain features, and then constructs an adaptive threshold based on the median to denoise layer by layer. Combined with boundary smoothing correction, the signal is reconstructed, which solves the problems of low accuracy of non-stationary signal feature extraction and easy masking of early fault features in the prior art, and improves the identification of weak fault features.

[0010] (3) This invention extracts time-domain features to match the fault mode library, combines sensor position and working condition data to calibrate and locate faulty components, and then associates historical data to construct a multi-dimensional vibration matrix. After spectrum refinement and pulse demodulation, the root cause of the fault is locked, solving the problem of poor fault location and tracing capabilities in the existing technology, and realizing accurate diagnosis of the entire fault process. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of a generator vibration fault diagnosis method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a generator vibration fault diagnosis system provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a generator vibration fault diagnosis method, including the following steps: S11, acquire the original vibration signal, real-time temperature data and real-time humidity data during generator operation, and perform environmental interference correction on the original vibration signal to obtain a preliminary correction dataset; S12, Perform discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components; S13, extract time-frequency feature vectors based on the set of detail components and the set of approximate components, and reconstruct the time-frequency feature vectors to obtain hierarchical frequency domain features, wherein the hierarchical frequency domain features include hierarchical detail component sets corresponding to each decomposition level; S14, Perform noise filtering on the layered detail component set to obtain a denoised detail component set; S15, perform inverse wavelet transform on the denoised detail component set and the approximate component set to obtain the reconstructed vibration signal sequence, and perform boundary correction on the reconstructed vibration signal sequence to obtain the correction signal set; S16, perform operating condition parameter matching on the correction signal set to obtain a fluctuation feature set, perform amplitude analysis on the fluctuation feature set to obtain amplitude features, and perform periodic distribution identification on the amplitude features to obtain a time domain feature set; S17, extract the amplitude deviation sequence from the time domain feature set, and search and match the amplitude deviation sequence with the preset fault mode library to determine the location of potential faulty components; S18, perform frequency anomaly matching based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

[0014] In step S11, environmental interference correction is applied to the original vibration signal to obtain a preliminary correction dataset, including: The original vibration signal includes a sampling frequency and a vibration amplitude. The vibration amplitude and the sampling frequency are standardized to obtain a standard vibration sequence. Query the position coordinates of the sensor corresponding to the standard vibration sequence, calculate the spatial offset vector of the sensor, and compare the spatial offset vector with a preset spatial offset threshold. If the spatial offset vector does not exceed the spatial offset threshold, the standard vibration sequence is directly used as the sequence to be corrected; if the spatial offset vector exceeds the spatial offset threshold, a deviation compensation matrix for geometric position is constructed, and the deviation compensation matrix is ​​used to perform spatial offset compensation on the standard vibration sequence to obtain the sequence to be corrected. A temperature-humidity coupling factor is generated based on the real-time temperature data, the real-time humidity data, and the preset temperature-humidity vibration characteristic mapping relationship. The temperature-humidity coupling factor is then used to correct environmental interference in the sequence to be corrected, resulting in a preliminary correction dataset.

[0015] It should be noted that the original vibration signal includes the sampling frequency and vibration amplitude corresponding to the data collected by each vibration sensor. The sampling frequency is the sampling clock frequency corresponding to each sensor, and the vibration amplitude is the real-time value of the vibration displacement at each sampling moment. When standardizing the vibration amplitude and sampling frequency, the vibration amplitude is first standardized dimensionlessly using the min-max standardization method. The preset dimensionless range is 0 to 1. After standardization, all vibration amplitude values ​​are mapped to this range, eliminating data skew caused by differences in physical dimensions and measurement ranges between different sensors. For example, if the vibration amplitude collected by the horizontal sensor of a bearing is 50 micrometers and the vibration amplitude collected by the vertical sensor is 0.05 millimeters, after dimensionless standardization, both sets of amplitude data are mapped to the range of 0 to 1, eliminating the dimensional difference between millimeters and micrometers.

[0016] Next, the sampling frequency alignment process uses the highest sampling frequency among the multiple original vibration signals as the target benchmark, and employs linear interpolation to complete the remaining original vibration signals with lower sampling frequencies. During the interpolation process, the time axis corresponding to the highest sampling frequency is used as a unified benchmark. Between adjacent original sampling points of the lower sampling frequency signals, the vibration amplitude values ​​of the interpolation points are calculated through linear fitting, ensuring that the sampling time interval and the number of sampling points per cycle are completely consistent across all channels. For example, if a horizontal sensor has a sampling frequency of 1000 Hz and a vertical sensor has a sampling frequency of 2000 Hz, the 1000 Hz signal is interpolated using 2000 Hz as the benchmark, ensuring complete alignment of the sampling frequencies of the two signals. This ultimately generates a standard vibration sequence with a unified sampling frequency and consistent dimensions.

[0017] It should be noted that the sensor position coordinates corresponding to the standard vibration sequence are the theoretical three-dimensional coordinates of the generator measuring point, and the spatial offset vector is the three-dimensional Euclidean distance between the actual sensor installation position and the theoretical installation position. The preset spatial offset threshold is set based on GB / T19073-2008 Rotating Machinery Vibration Monitoring Installation Specification and statistical results of 1000 sets of installation and maintenance data of the same model of generator. When the sensor installation position deviation exceeds 2 mm, the calculation error of the vibration signal transmission path will exceed the industry allowable range of ±5%. After multiple batches of on-site installation verification, the 2 mm threshold can balance installation fault tolerance and vibration signal spatial positioning accuracy. Therefore, the spatial offset threshold is fixed at 2 mm. When the spatial offset vector value does not exceed the preset spatial offset threshold, the standard vibration sequence is directly used as the sequence to be corrected. When the spatial offset vector value exceeds the preset spatial offset threshold, a 3×3 geometric position deviation compensation matrix is ​​constructed based on the sensor's three-dimensional spatial offset. The deviation compensation matrix is ​​used to perform matrix multiplication on the standard vibration sequence to complete the spatial offset compensation, restore the transmission path of the vibration signal in real space, and obtain the sequence to be corrected.

[0018] It is worth noting that the real-time temperature and humidity data of the signal acquisition environment are collected by temperature and humidity sensors deployed in the generator room, with the acquisition frequency synchronized with the highest sampling frequency of the vibration sensor. A preset temperature-humidity vibration characteristic mapping relationship is generated based on vibration characteristic calibration test data of the generator under different temperature and humidity environments, characterizing the coupling effect of environmental temperature and humidity changes on rotor stiffness and sensor sensitivity. Based on the real-time temperature and humidity data, the fitting function corresponding to the temperature-humidity vibration characteristic mapping relationship is substituted to calculate the temperature-humidity coupling factor. The sequence to be corrected is then multiplied point-by-point with the temperature-humidity coupling factor to complete environmental interference correction, eliminating spurious vibration components caused by high temperature and high humidity environments, resulting in a preliminary corrected dataset for the vibration signal. The real-time temperature and humidity data are collected by temperature and humidity sensors deployed in the generator room, with the acquisition frequency synchronized with the highest sampling frequency of the vibration sensor.

[0019] In step S12, a discrete wavelet transform is performed on the preliminary correction dataset to generate a set of detail components and a set of approximate components, including: Based on the non-stationary characteristics of the preliminary correction dataset, the decomposition level values ​​and the target wavelet basis function are determined; The target wavelet basis function is used to perform multi-level discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components corresponding to the decomposition level.

[0020] It should be noted that the decomposition level values ​​are determined based on the sampling frequency of the preliminary calibration dataset and the lower limit of the lowest fault frequency to be analyzed. The lowest fault frequency to be analyzed is the generator power frequency of 50 Hz. The decomposition level values ​​must meet the requirement that the frequency resolution in the low-frequency band can distinguish between the fundamental frequency and harmonic components. For example, when the sampling frequency of the preliminary calibration dataset is 10240 Hz, the decomposition level values ​​are set to 8 levels. The target wavelet basis function is determined based on the non-stationary characteristics of the preliminary calibration dataset. Priority is given to Daubechies wavelets, which have high similarity to fault characteristic waveforms such as generator bearing pitting and rotor misalignment, and possess good compact support and regularity. Specifically, db4 or db8 wavelets can be used to ensure accurate capture of transient impact components in the vibration signal.

[0021] It is worth noting that a multi-level discrete wavelet transform is performed on the preliminary correction dataset using the target wavelet basis function. The transform process is achieved through a series of high-pass and low-pass filters. In the first level of decomposition, the preliminary correction dataset is separated into high-frequency detail components by a high-pass filter and into low-frequency approximate components by a low-pass filter. Subsequent decomposition processes only apply the same high-pass and low-pass filtering operations to the low-frequency approximate components output from the previous level until the preset decomposition level is reached. The high-frequency components output from each level of decomposition collectively form the detail component set, and the low-frequency components output from the final level of decomposition form the approximate component set. During the multi-level discrete wavelet transform, the wavelet coefficients corresponding to each component are output synchronously for subsequent frequency domain feature calculation and noise filtering.

[0022] In step S13, time-frequency feature vectors are extracted based on the detail component set and the approximate component set. These time-frequency feature vectors are then reconstructed to obtain hierarchical frequency domain features, including: Based on the decomposition level values, calculate the vibration frequency range corresponding to each component in the detailed component set and the approximate component set; Extract the wavelet coefficient sequence corresponding to each component, calculate the sum of squares of all values ​​in the wavelet coefficient sequence, and then divide the sum of squares by the total length of the wavelet coefficient sequence to obtain the energy spectral density value of the corresponding component. Extract the vibration frequency range, energy spectral density value, and amplitude and temporal distribution features extracted from the wavelet coefficient sequence for each component to construct the time-frequency feature vector of the corresponding component. All the time-frequency feature vectors are arranged and recombined in order of decomposition level from low frequency to high frequency to obtain hierarchical frequency domain features.

[0023] It should be noted that the calculation of the vibration frequency range is based on the Nyquist sampling theorem. The Nyquist frequency is half the sampling frequency of the initial calibration dataset, representing the upper limit of the discrete signal that can be restored without distortion. The formulas for calculating the frequency bands corresponding to each decomposition level are as follows: ; in, This represents the lower limit of the frequency range corresponding to the nth layer component. This represents the upper limit of the frequency range corresponding to the nth layer component. To initially correct the sampling frequency of the dataset, n represents the decomposition level. The nth level detail component corresponds to the frequency range calculated above, and the approximate component of the highest decomposition level corresponds to the frequency range from 0 to the lower limit of the highest decomposition level. For example, if the sampling frequency of the initially corrected dataset is 10240 Hz and the decomposition level is 8, the first level detail component corresponds to a frequency range of 2560 Hz to 5120 Hz, and the eighth level approximate component corresponds to a frequency range of 0 to 20 Hz, completely covering the lower limit of the 50 Hz power frequency fault frequency to be analyzed.

[0024] It is worth noting that in the calculation of the energy spectral density, the discrete sequence of wavelet coefficients corresponding to each component is first extracted, the sum of squares of all values ​​in the sequence is calculated, and then the sum of squares is divided by the total length of the wavelet coefficient sequence for that component to obtain the energy spectral density value of the corresponding component. The calculation formula is as follows: ; In the formula, E is the energy spectral density value of the corresponding component. , where represents the i-th value in the wavelet coefficient sequence corresponding to the component, and N is the total length of the wavelet coefficient sequence, i.e., the signal length. This value directly reflects the intensity of the generator's vibration energy in the corresponding frequency band. An abnormally high energy spectral density value of the high-frequency detail component corresponds to structural cracks or component friction-related faults, while the energy concentration of the low-frequency approximate component corresponds to rotor mass imbalance-related faults.

[0025] Next, the construction of time-frequency feature vectors uses single components at the single-level decomposition as the basic unit, with each component generating an independent set of time-frequency feature vectors. The first dimension of the vector represents the upper and lower limits of the vibration frequency range corresponding to that component, the second dimension represents the energy spectral density of that component, and the remaining dimensions represent the amplitude and temporal distribution features extracted from the wavelet coefficients, specifically including the mean, peak, standard deviation, and temporal gradient of the wavelet coefficient amplitudes. All time-frequency feature vectors maintain a consistent dimension, and the values ​​of each dimension within the vector are dimensionlessly standardized to eliminate feature skew caused by differences in the magnitude of values ​​across different dimensions.

[0026] It is worth further elaborating that the arrangement and recombination of time-frequency feature vectors strictly follows the order of decomposition levels from low frequency to high frequency. First, the time-frequency feature vectors corresponding to the approximate components of the highest decomposition level are placed. Then, from the highest decomposition level to the first level, the time-frequency feature vectors corresponding to the detail components of each level are placed sequentially, ultimately forming a structured hierarchical frequency domain feature. Within the hierarchical frequency domain feature, the subset of time-frequency feature vectors corresponding to all detail components constitutes the set of hierarchical detail components for each decomposition level. This structured feature representation fully preserves the time-frequency localization characteristics of the vibration signal, forming highly discriminative feature inputs through the differences in energy distribution at different levels, adapting to subsequent noise filtering and fault mode recognition processing.

[0027] In step S14, a noise filtering operation is performed on the hierarchical detail component set to obtain a denoised detail component set, including: Extract the median value of the absolute values ​​of the wavelet coefficient sequence at each decomposition level; The total length of the wavelet coefficient sequence is extracted as the signal length. Based on the median value and the signal length of the corresponding decomposition level, the preset adaptive threshold calculation function is substituted to determine the noise filtering threshold corresponding to each decomposition level. Compare the amplitude of the wavelet coefficients in the corresponding decomposition level with the noise filtering threshold one by one; If the amplitude of the wavelet coefficient is greater than or equal to the noise filtering threshold, the set of hierarchical detail components corresponding to the wavelet coefficient is retained; if the amplitude of the wavelet coefficient is less than the noise filtering threshold, the set of hierarchical detail components corresponding to the wavelet coefficient is set to zero. The integrated set of layered detail components is used to generate a denoised set of detail components.

[0028] It should be noted that when obtaining the set of hierarchical detail components, all wavelet coefficients corresponding to each decomposition level after discrete wavelet transform are extracted simultaneously. Absolute value operations are performed on the wavelet coefficients of each decomposition level to generate a sequence of absolute coefficient values ​​for that level. After sorting the absolute coefficient value sequence in ascending order, the value at the middle position of the sequence is taken as the median value of the wavelet coefficient absolute value sequence for that decomposition level. The sorting process follows general rules of numerical statistics: when the total length of the sequence is even, the arithmetic mean of the two middle values ​​is taken as the median value; when the total length of the sequence is odd, the value at the exact middle position is taken as the median value. The calculation logic of the median value is set based on the general statistical laws of rotating machinery vibration signal processing. Verified by multiple batches of generator operation vibration data, the wavelet coefficients corresponding to real fault impacts exhibit a sparse distribution characteristic, while the wavelet coefficients corresponding to background noise exhibit a dense and uniform distribution characteristic. The median value can objectively characterize the overall intensity level of background noise within the corresponding decomposition level. After a vibration signal with a sampling frequency of 10240Hz undergoes 8 layers of discrete wavelet transform, the first layer of detail components contains 2048 wavelet coefficients. After sorting the absolute value sequence of the coefficients, the arithmetic mean of the 1024th and 1025th values ​​is taken as the median value of this layer.

[0029] It is worth noting that the total length of the wavelet coefficient sequence of the detail components within the corresponding decomposition level is extracted as the signal length. The median value of the corresponding decomposition level and the signal length are then substituted into a preset adaptive threshold calculation function to generate a unique noise filtering threshold for each decomposition level. The adaptive threshold calculation function is set according to the signal processing standards for generator vibration fault diagnosis in the power industry. Verified through multi-condition field data, it can adapt to the noise distribution differences of different decomposition levels, avoiding the loss of effective features or noise residue caused by a fixed threshold. The function expression is: ; in, This is the noise filtering threshold corresponding to the decomposition level. The median of the absolute values ​​of the wavelet coefficient sequences corresponding to the decomposition levels is given, and N is the signal length of the corresponding decomposition level. The noise filtering threshold for different decomposition levels is dynamically adjusted as the level changes. Higher frequency decomposition levels correspond to higher noise energy and are matched with higher noise filtering thresholds, while lower frequency decomposition levels correspond to lower noise energy and are matched with lower noise filtering thresholds. After the vibration signal with a sampling frequency of 10240 Hz is transformed by 8 layers of discrete wavelet transform, the first layer detail component corresponds to the high frequency band of 2560 Hz to 5120 Hz, where the environmental background noise energy is concentrated. Statistical verification using 100 sets of field vibration data shows that setting the noise filtering threshold for this level to 0.35 can effectively filter more than 90% of random background noise. The third layer detail component corresponds to the frequency band of 640 Hz to 1280 Hz, where the noise energy is significantly reduced. The corresponding noise filtering threshold is set to 0.12, which can suppress residual noise while preserving fault characteristics.

[0030] After calculating the noise filtering threshold for each decomposition level, the amplitude of each wavelet coefficient within the corresponding decomposition level is compared with the noise filtering threshold for that level, in descending order of decomposition level. The comparison process covers all wavelet coefficients within the corresponding decomposition level, ensuring no omissions or repetitions, and guaranteeing that each coefficient completes the threshold determination. If the wavelet coefficient amplitude is greater than or equal to the noise filtering threshold, the original value and temporal position of the wavelet coefficient are fully preserved. If the wavelet coefficient amplitude is less than the noise filtering threshold, the wavelet coefficient is set to zero. This determination logic has been verified by field operation data and can effectively eliminate invalid wavelet coefficients corresponding to environmental background noise, while fully preserving the true high-frequency impact energy corresponding to fault characteristics such as early minor damage to generator bearings and slight deformation of rotor components.

[0031] Subsequently, following the decomposition hierarchy and temporal arrangement rules corresponding to the original discrete wavelet transform, the processed wavelet coefficients from each decomposition level are integrated to generate a set of denoised detail components. The integration process strictly maintains the temporal correspondence and hierarchy of the wavelet coefficients at each decomposition level, without altering the coefficient arrangement order or dimensional structure. This ensures that the set of denoised detail components can be directly matched with the unprocessed approximate component set, providing input data that meets the format requirements for subsequent inverse wavelet transform signal reconstruction.

[0032] In step S15, boundary correction is performed on the reconstructed vibration signal sequence to obtain a correction signal set, including: Extract the boundary data point set of the first and last preset lengths of the reconstructed vibration signal sequence, and apply the preset mirror continuation algorithm to the boundary data point set to generate extended boundary data; The extended boundary data is subjected to polynomial fitting and smoothing filtering to obtain boundary correction data; Calculate the amplitude jump rate of the boundary correction data and compare the amplitude jump rate with a preset jump threshold; If the amplitude jump rate does not exceed the jump threshold, then the boundary correction data is used to replace the corresponding boundary data point set of the reconstructed vibration signal sequence to obtain the correction signal set of the vibration signal. If the amplitude jump rate exceeds the jump threshold, then perform quadratic spline interpolation calculation on the boundary correction data to obtain the secondary correction data of the boundary data. Use the secondary correction data to replace the corresponding boundary data point set of the reconstructed vibration signal sequence to obtain the correction signal set of the vibration signal.

[0033] It should be noted that when performing inverse wavelet transform on the denoised detail component set and approximate component set, the target basis function and decomposition level are completely consistent with those used in the discrete wavelet transform stage. Based on the general technical specifications for generator vibration signal processing in the power industry, the target basis function is the db4 wavelet from the Daubechies series, and the decomposition level is set to 8 levels. The inverse wavelet transform is implemented through a step-by-step synthesis method. Starting from the approximate component and its corresponding detail component at the highest decomposition level, upsampling and convolution operations are performed sequentially through low-pass and high-pass reconstruction filters until a complete reconstructed vibration signal sequence is synthesized. During the upsampling process, a zero value is inserted between every two adjacent coefficients. The filter coefficients of the convolution operation are mirror-symmetric to those of the filter coefficients in the decomposition stage. Since the generator vibration signal is a finite-length sequence, the inverse wavelet transform will produce boundary distortion at the beginning and end of the signal due to the truncation effect of the filters, manifesting as artifacts or discontinuous abrupt changes at the endpoints, requiring subsequent boundary smoothing correction processing.

[0034] It is worth noting that when extracting the boundary data point sets of the first and last ends of the reconstructed vibration signal sequence, the preset length is set based on the support length of the wavelet basis function and practical experience in the power industry. The support length of the db4 wavelet is 8 coefficients. After verification through multiple batches of generator vibration signal processing, setting the preset length to 32 data points can completely cover the main influence area of ​​boundary distortion, while avoiding excessive extension and unnecessary computation. The extraction process directly extracts the first 32 data points of the reconstructed vibration signal sequence as the first boundary data point set and the last 32 data points as the last boundary data point set. When applying the preset mirror extension algorithm to extend the boundary data point set, the boundary data point set is flipped and copied with the endpoints of the boundary data point set as the axis of symmetry. For the first boundary data point set... The extended boundary data generated after the extension is The same extension process is applied to the tail boundary data point set. The length of the extended boundary data is three times the length of the original boundary data point set, i.e., 96 data points. This artificially extends the signal boundary, providing a sufficient transition interval for subsequent smoothing processing and preventing the smoothing operation from directly affecting the main signal.

[0035] It should be further explained that when performing polynomial fitting and smoothing filtering on the extended boundary data, the order of the polynomial is set according to the fluctuation characteristics of the generator vibration signal boundary. Practice has shown that third-order polynomial fitting can achieve the optimal balance between suppressing random fluctuations at the boundary and preserving the true vibration trend. Too high an order can easily lead to overfitting and introducing spurious vibration features, while too low an order results in insufficient smoothing and cannot effectively eliminate unnatural fluctuations caused by the extension. The fitting process uses the least squares method, with the time index of the extended boundary data as the independent variable and the data amplitude as the dependent variable. A third-order polynomial function is constructed, and the polynomial coefficients are determined by minimizing the sum of squared residuals between the fitted value and the actual value. After solving for the coefficients, the fitted values ​​corresponding to the middle 32 points of the extended boundary data are extracted as boundary correction data.

[0036] The preset jump threshold is set based on the quality standards and historical data statistics of generator vibration signals in the power industry. One hundred sets of generator vibration signals under normal operating conditions were selected, and after undergoing the same inverse wavelet transform and boundary smoothing processing, the amplitude jump rate of their boundary correction data was calculated. Statistical results show that over 95% of normal signals have a jump rate below 0.15. Simultaneously, considering the equipment tolerance requirements for generator rotor vibration, the allowable fluctuation range of rotor vibration amplitude is ±10%, therefore, the jump threshold is set to 0.15. This threshold has been verified through multiple batches of field data, accurately identifying severe discontinuities at the boundary without misjudging normal boundary fluctuations.

[0037] Subsequently, the amplitude jump rate is compared with the preset jump threshold. If the amplitude jump rate does not exceed the threshold, it indicates that the polynomial fitting has effectively eliminated the discontinuity at the boundary. The corresponding boundary data point set of the reconstructed vibration signal sequence is directly replaced using the boundary correction data, i.e., the first two points are replaced by the first boundary correction data, and the last 32 points are replaced by the last boundary correction data, resulting in the corrected signal set of the vibration signal. If the amplitude jump rate exceeds the threshold, it indicates that polynomial fitting alone cannot completely eliminate the energy step at the boundary, and quadratic spline interpolation is required for the boundary correction data. Quadratic spline interpolation uses the 32 points of the boundary correction data as interpolation nodes to construct a piecewise quadratic polynomial function. Each segment contains three consecutive nodes, ensuring that the function values ​​and first derivatives at the connection points of adjacent segments are continuous. The condition of continuous first derivatives ensures a smooth transition of the boundary curve and avoids sharp angles. The coefficients of each piecewise quadratic polynomial are determined by solving the three bending moment equations, and the values ​​of the boundary region are recalculated to obtain the quadratic correction data of the boundary data. Finally, the corresponding boundary data point set of the reconstructed vibration signal sequence is replaced with the secondary correction data to obtain the correction signal set of the vibration signal. The amplitude jump rate of the first boundary correction data of a certain reconstructed vibration signal sequence is 0.18, which exceeds the jump threshold of 0.15. After quadratic spline interpolation, the jump rate is reduced to 0.12, effectively eliminating the energy step at the boundary.

[0038] In step S16, the operating condition parameters of the correction signal set are matched to obtain a fluctuation feature set. Amplitude analysis is performed on the fluctuation feature set to obtain amplitude features. Periodic distribution identification is then performed on the amplitude features to obtain a time-domain feature set, including: The duration of the correction signal set is obtained, and the correction signal set is divided into equal intervals based on the duration of the signal to obtain a multi-group segmented signal sequence. The segmented signal sequence is matched with the real-time operating parameters of the generator to generate a set of fluctuation features associated with the operating conditions. Calculate the change in fluctuation amplitude of the fluctuation feature set, and compare the change in fluctuation amplitude with a preset amplitude correction threshold; If the change in fluctuation amplitude does not exceed the amplitude correction threshold, the fluctuation feature set is directly used as the amplitude feature; if the change in fluctuation amplitude exceeds the amplitude correction threshold, amplitude gain correction is performed to obtain the amplitude feature. The amplitude features are periodically distributed to obtain a periodic distribution feature vector. The periodic distribution feature vector is then subjected to dimensionless mapping to obtain a time-domain feature set.

[0039] It should be noted that when dividing the calibration signal set into equal intervals based on the signal duration, the number of intervals and the duration of each interval are determined according to the general specifications for generator operation status monitoring in the power industry and the verification results of field operation data. During normal generator operation, the duration of operating condition fluctuations and transient fault characteristics is mostly concentrated in the range of 10 to 120 seconds. After verification with multiple batches of field data, setting the duration of each interval to 60 seconds achieves the optimal balance between ensuring the accuracy of transient feature capture and reducing computational load. The division process starts from the start timestamp of signal acquisition and divides the calibration signal set into non-overlapping equal intervals according to the preset duration of each interval. If the total signal duration cannot be divided evenly by the duration of each interval, the signal segments at the end that are less than the duration of each interval are padded with zeros, with the padded length not exceeding 10% of the duration of each interval, or the segment is directly discarded, ultimately obtaining multiple sets of segmented signal sequences with consistent durations. This division method can effectively isolate local vibration characteristics within different time windows and avoid the averaging effect of long-term signals masking transient vibration information caused by early faults.

[0040] It is worth noting that when matching segmented signal sequences with real-time generator operating parameters, these parameters include operating data strongly correlated with vibration characteristics, such as real-time active power, reactive power, rotor speed, excitation current, and bearing temperature. The matching process uses timestamps as a benchmark, mapping each segmented signal sequence one-to-one with the operating parameters within the corresponding time window to ensure that the time synchronization error between vibration data and operating data does not exceed 100ms. When extracting the time-domain fluctuation patterns associated with operating conditions, the time-domain statistical characteristics of each segmented signal sequence are calculated simultaneously, including basic time-domain indices such as vibration peak value, peak-to-peak value, RMS value, and mean value. A mapping relationship between time-domain statistical characteristics and changes in operating parameters is established, vibration amplitude changes caused by normal operating condition fluctuations are eliminated, and fluctuation features exceeding the normal mapping range of operating conditions are extracted to generate a set of fluctuation features associated with operating conditions. For example, during the variable operating condition process where the generator load increases from 50% to 80% of the rated load, the effective value of vibration exhibits normal fluctuations with linear growth as the load increases. These fluctuations are eliminated, while nonlinear fluctuations that exceed the linear growth range are included in the fluctuation feature set associated with the operating condition, effectively distinguishing between normal operating condition fluctuations and abnormal vibration characteristics.

[0041] Next, when calculating the amplitude change of the fluctuation feature set, the vibration reference value under the generator's rated operating conditions and in a healthy state is used as a reference. The difference between the maximum vibration amplitude value and the vibration reference value within the corresponding segmented signal sequence fluctuation feature set is divided by the vibration reference value, and the absolute value is taken as the amplitude change. The preset amplitude correction threshold is determined comprehensively based on GB / T6075.3-2011 Rotating Machinery Vibration Standard, generator equipment manufacturer's technical specifications, and historical operating data statistics. 1000 sets of healthy vibration data of the same model generator within the normal operating condition fluctuation range were selected, and their amplitude changes were calculated. The statistical results show that the amplitude changes of more than 95% of the healthy data are less than 0.2. Combined with the allowable tolerance requirements for generator bearing vibration, the allowable fluctuation range of vibration amplitude under normal operating conditions is ±20% of the rated value. Therefore, the amplitude correction threshold is set to 0.2. This threshold has been verified by multi-condition field operating data and can accurately identify abnormal amplitude fluctuations caused by drastic changes in operating conditions or potential faults, while avoiding misjudgment of normal operating condition fluctuations.

[0042] Subsequently, the change in fluctuation amplitude is compared with a preset amplitude correction threshold. If the change in fluctuation amplitude does not exceed the amplitude correction threshold, it indicates that the vibration amplitude fluctuation is within the normal operating condition fluctuation range, and the operating condition-related fluctuation feature set is directly used as the amplitude feature. If the change in fluctuation amplitude exceeds the amplitude correction threshold, it indicates that there is amplitude distortion in the signal caused by drastic changes in operating conditions, or energy mutation caused by potential faults. Amplitude gain correction is then performed to obtain the amplitude feature. The amplitude gain correction is adjusted using a dynamic gain coefficient, which is calculated as follows: ; in, This is the dynamic gain coefficient. This represents the change in the amplitude of the fluctuation. An amplitude correction threshold is used. When the fluctuation amplitude exceeds the threshold, a dynamic gain coefficient is applied to smooth out the amplitude fluctuations exceeding the threshold. Simultaneously, weak fault characteristics masked by noise are linearly enhanced. The correction process strictly preserves the relative temporal relationship and frequency distribution characteristics of the vibration signal, avoiding alteration of the essential attributes of the fault characteristics. For example, the amplitude variation of a segmented signal sequence is 0.25, exceeding the amplitude correction threshold of 0.2. The calculated dynamic gain coefficient is 0.95. This coefficient is used to correct the vibration amplitude, smoothing out excessive amplitude fluctuations and obtaining the amplitude characteristics.

[0043] It is worth further explaining that when identifying periodic distribution patterns based on amplitude characteristics, an autocorrelation analysis method is used to perform autocorrelation calculations on the time-domain signal corresponding to the amplitude characteristics. The formula for calculating the autocorrelation function is as follows: ; in, For time delay The autocorrelation function value at time N, where N is the total length of the segmented signal sequence. This represents the segmented time-domain signal sequence. By analyzing the peak distribution of the autocorrelation function, the implicit periodicity in the signal is identified, determining the fundamental and harmonic distributions of the vibration signal. Parameters such as fundamental frequency amplitude, harmonic amplitude, peak interval, and periodic stability are extracted to obtain the periodic distribution feature vector. When the generator rotor's rated speed is 3000 r / min, the corresponding power frequency is 50 Hz. The peak interval of the autocorrelation function of the normal vibration signal is stable at 0.02 seconds, corresponding to the 50 Hz power frequency characteristic. If there is a bearing fault or rotor misalignment, additional peaks will appear at the corresponding fault frequency. When performing dimensionless mapping on the periodic distribution feature vector, dimensionless indices commonly used in rotating machinery vibration fault diagnosis are selected, including waveform factor, peak factor, impulse factor, margin factor, and kurtosis index. The periodic distribution feature vector is mapped to the corresponding set of dimensionless indices, obtaining the time-domain feature set of the vibration signal. Dimensionless indices are unaffected by sensor gain, absolute signal amplitude, or operating condition fluctuations, exhibiting high sensitivity to early impact faults and effectively improving the distinguishability of fault characteristics under different operating conditions.

[0044] In step S17, an amplitude deviation sequence is extracted from the time-domain feature set, and the amplitude deviation sequence is searched and matched with a preset fault mode library to determine the location of potential faulty components, including: The difference between the actual observed amplitude and the healthy baseline amplitude is calculated from the time-domain feature set to obtain the amplitude deviation sequence; The amplitude deviation sequence is compared with a preset error threshold. If the amplitude deviation sequence does not exceed the error threshold, the generator vibration state is determined to be normal. If the amplitude deviation sequence exceeds the error threshold, the amplitude deviation sequence is searched and matched with a preset fault mode library to determine the preliminary fault type and component identifier. Based on the pre-acquired generator equipment topology, the location of the potential faulty component corresponding to the component identifier is determined.

[0045] It should be noted that, from the periodic distribution feature vector of the time-domain feature set, the actual observed amplitude at each frequency point within the corresponding time window is extracted, matched with the health benchmark amplitude corresponding to the generator's health state under the same operating conditions, and the difference between the two is calculated point by point to generate a continuous difference sequence. The fluctuation amplitude change is calculated based on the extreme values ​​and mean values ​​of the difference sequence. The health benchmark amplitude comes from the factory calibration data and historical health operation data of the same model generator under rated operating conditions. A benchmark matrix is ​​established according to operating parameters such as active power and rotor speed to ensure the matching degree between the benchmark value and the real-time operating conditions.

[0046] It is worth noting that the preset error threshold is set based on the GB / T6075.3-2011 standard for vibration intensity of rotating machinery, generator equipment operation and maintenance specifications, and statistical results of historical fault data. Vibration data from 1000 generators of the same model under healthy operating conditions were selected, and the 95th percentile of the amplitude variation was statistically analyzed. Combined with the equipment's allowable vibration tolerance, the error threshold was set to 0.2. The amplitude variation was compared with the error threshold. If the amplitude variation did not exceed the threshold, the generator vibration state was considered normal. If it exceeded the threshold, the trend of the difference sequence and the slope of amplitude growth were searched and matched against a preset fault mode library. The library contains the difference sequence evolution paths of typical faults such as bearing wear, rotor misalignment, rotor imbalance, and uneven air gap. The matching degree was calculated using cosine similarity, and the fault type with the highest matching degree was determined as the preliminary fault type.

[0047] Next, real-time equipment operation data is acquired. The sensor position coordinates are the spatial coordinates of each measuring point within the generator's three-dimensional topology. The real-time output power includes parameters such as active power, excitation current, and rotor speed. Based on the preliminary fault type, a linear mapping model between operating parameters and vibration components is established. Synchronous vibration components caused by changes in operating conditions are eliminated, and non-fault vibration components caused by electromagnetic excitation and load fluctuations are removed, resulting in calibrated vibration components that only reflect mechanical damage characteristics. Using the pre-stored generator equipment topology, the calibrated vibration components are mapped to unique component identifiers of corresponding physical components according to the vibration transmission path and the spatial location of the measuring points. These component identifiers cover core components such as drive-end bearings, non-drive-end bearings, rotor, stator, and frame. Based on the component identifiers, a pre-set equipment health record is retrieved. This record contains information such as the cumulative runtime, historical maintenance records, and fault frequency of the corresponding component. Combining the vibration transmission characteristics and fault occurrence probability, the location of the potential faulty component corresponding to the abnormal vibration is finally determined.

[0048] In step S18, frequency anomaly matching is performed based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly, including: Based on the index of the location of the potential faulty component, retrieve the historical vibration amplitude records within the corresponding preset period, and combine them with the current real-time vibration data to construct a multi-dimensional vibration state matrix; The multidimensional vibration state matrix is ​​subjected to spectral refinement to identify the frequency range corresponding to abnormal vibrations; The spectral components within the frequency range are separated into asynchronous components to obtain abnormal spectral components. Extract the nonlinear harmonic component sequence from the abnormal spectral components, and demodulate the impulse pulse sequence from the nonlinear harmonic component sequence; The impact pulse sequence is compared with a preset fault mode fingerprint database to identify the abnormal root cause component of the vibration anomaly in the physical entity that generates the impact.

[0049] It should be noted that, based on the unique component identifier index of the potential faulty component location, historical vibration amplitude records for the corresponding component over 30 consecutive operating cycles are retrieved. These records cover vibration data in the axial, radial, and tangential dimensions, along with corresponding phase angles and operating parameters. These records are then timestamped and aligned with the current real-time vibration data in the same dimension. A multi-dimensional vibration state matrix is ​​constructed by reorganizing the data according to time and space dimensions, fully preserving the temporal evolution and spatial distribution characteristics of the vibration signal. This matrix is ​​then subjected to spectral refinement using the Zoom-FFT algorithm, with the refinement factor set at 100 times based on the fault identification accuracy requirements. This improves the spectral analysis resolution to 0.1Hz. By calculating the energy spectral density at each frequency point, the frequency bands where energy anomalies are concentrated are located, accurately identifying the frequency range corresponding to the abnormal vibration.

[0050] Using the fundamental frequency calculated from the real-time rotor speed of the generator as a reference, the spectral components within the frequency range are checked for synchronization. Synchronous spectral components that are positive integer multiples of the fundamental frequency are filtered out, and vibration components caused by non-fault factors such as normal rotor rotation and electromagnetic excitation are eliminated, separating out asynchronous abnormal spectral components. An analytical signal is constructed by performing a Hilbert transform on the abnormal spectral components, and the signal envelope is extracted to obtain a nonlinear harmonic component sequence. This sequence is then subjected to further spectral refinement analysis to demodulate the periodic impact pulse sequence hidden in the high-frequency carrier, fully preserving key characteristics such as the impact time interval and amplitude attenuation characteristics corresponding to local damage to mechanical components.

[0051] It should be noted that the impact pulse sequence is compared with a preset fault mode fingerprint database. The database contains theoretical fault characteristic frequencies of key components of the generator calculated based on geometric parameters, as well as measured fault fingerprint features obtained from historical fault cases of the same model and full life cycle tests. The matching degree between the two is calculated using a cosine similarity algorithm. When the matching degree reaches 0.85 or higher, the physical entity associated with the corresponding fingerprint feature is identified. Combining the installation location, operating status, and equipment health record of this physical entity, the root cause of the abnormal vibration is finally determined, such as fatigue spalling of the bearing inner ring raceway, pitting of rolling elements, and wear of the rotor journal.

[0052] In summary, this invention discloses a generator vibration fault diagnosis method, comprising: acquiring the original vibration signal, real-time temperature data, and real-time humidity data during generator operation; correcting the original vibration signal for environmental interference to obtain a preliminary correction dataset; performing discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components; extracting time-frequency feature vectors from the set of detail components and the set of approximate components, and reconstructing the time-frequency feature vectors to obtain layered frequency domain features; performing noise filtering on the layered detail component set to obtain a denoised detail component set; and further processing the denoised detail component set... The reconstructed vibration signal sequence is obtained by performing inverse wavelet transform on the quantity set and the approximate component set, and boundary correction is performed on the reconstructed vibration signal sequence to obtain a correction signal set. Operating condition parameter matching is performed on the correction signal set to obtain a fluctuation feature set, amplitude analysis is performed on the fluctuation feature set to obtain amplitude features, and periodic distribution identification is performed on the amplitude features to obtain a time-domain feature set. An amplitude deviation sequence is extracted from the time-domain feature set, and the amplitude deviation sequence is searched and matched with a preset fault mode library to determine the location of potential faulty components. Frequency anomaly matching is performed based on the location of the potential faulty components to determine the abnormal root cause component of the vibration anomaly. This invention achieves accurate diagnosis and root cause location of generator vibration faults through vibration signal standardization correction, multi-scale time-frequency decomposition denoising, operating condition correlation feature extraction, and fault mode matching, thereby improving equipment operation safety and maintenance efficiency.

[0053] Reference Figure 2 The second embodiment of the present invention provides a generator vibration fault diagnosis system, comprising: The signal correction module is used to acquire the original vibration signal, real-time temperature data and real-time humidity data during generator operation, and to correct the original vibration signal for environmental interference to obtain a preliminary correction dataset. The wavelet decomposition module performs discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components. The frequency domain feature generation module is used to extract time-frequency feature vectors based on the set of detail components and the set of approximate components, and to reassemble the time-frequency feature vectors to obtain hierarchical frequency domain features, wherein the hierarchical frequency domain features include hierarchical detail component sets corresponding to each decomposition level; The noise suppression module is used to perform noise filtering on the hierarchical detail component set to obtain a denoised detail component set. The signal reconstruction module is used to perform inverse wavelet transform on the denoised detail component set and the approximate component set to obtain a reconstructed vibration signal sequence, and to perform boundary correction on the reconstructed vibration signal sequence to obtain a correction signal set; The time-domain feature extraction module is used to match the operating parameters of the correction signal set to obtain a fluctuation feature set, perform amplitude analysis on the fluctuation feature set to obtain amplitude features, and identify the periodic distribution of the amplitude features to obtain a time-domain feature set. The potential fault location module is used to extract the amplitude deviation sequence from the time domain feature set, and search and match the amplitude deviation sequence with a preset fault mode library to determine the location of the potential faulty component. The fault source localization module is used to perform frequency anomaly matching based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

[0054] It should be noted that the generator vibration fault diagnosis system provided in this embodiment of the invention is used to perform all the process steps of the generator vibration fault diagnosis method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0055] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0056] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for diagnosing generator vibration faults, characterized in that, include: The original vibration signal, real-time temperature data, and real-time humidity data during generator operation are acquired, and the original vibration signal is corrected for environmental interference to obtain a preliminary correction dataset. Perform discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components; The time-frequency feature vector is extracted from the set of detail components and the set of approximate components. The time-frequency feature vector is then recombined to obtain the hierarchical frequency domain feature, wherein the hierarchical frequency domain feature includes the set of hierarchical detail components corresponding to each decomposition level. A noise filtering operation is performed on the hierarchical detail component set to obtain a denoised detail component set. The inverse wavelet transform is performed on the denoised detail component set and the approximate component set to obtain the reconstructed vibration signal sequence, and the boundary correction is performed on the reconstructed vibration signal sequence to obtain the correction signal set; The set of correction signals is matched with operating parameters to obtain a set of fluctuation features. The set of fluctuation features is analyzed to obtain amplitude features. The amplitude features are then identified by periodic distribution to obtain a set of time-domain features. The amplitude deviation sequence is extracted from the time-domain feature set, and the amplitude deviation sequence is searched and matched with the preset fault mode library to determine the location of potential faulty components. Frequency anomaly matching is performed based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

2. The generator vibration fault diagnosis method according to claim 1, characterized in that, The environmental interference correction of the original vibration signal yields a preliminary correction dataset, including: The original vibration signal includes a sampling frequency and a vibration amplitude. The vibration amplitude and the sampling frequency are standardized to obtain a standard vibration sequence. Query the position coordinates of the sensor corresponding to the standard vibration sequence, calculate the spatial offset vector of the sensor, and compare the spatial offset vector with a preset spatial offset threshold. If the spatial offset vector does not exceed the spatial offset threshold, the standard vibration sequence is directly used as the sequence to be corrected; if the spatial offset vector exceeds the spatial offset threshold, a deviation compensation matrix for geometric position is constructed, and the deviation compensation matrix is ​​used to perform spatial offset compensation on the standard vibration sequence to obtain the sequence to be corrected. A temperature-humidity coupling factor is generated based on the real-time temperature data, the real-time humidity data, and the preset temperature-humidity vibration characteristic mapping relationship. The temperature-humidity coupling factor is then used to correct environmental interference in the sequence to be corrected, resulting in a preliminary correction dataset.

3. The generator vibration fault diagnosis method according to claim 1, characterized in that, The step of performing discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components includes: Based on the non-stationary characteristics of the preliminary correction dataset, the decomposition level values ​​and the target wavelet basis function are determined; The target wavelet basis function is used to perform multi-level discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components corresponding to the decomposition level.

4. The generator vibration fault diagnosis method according to claim 3, characterized in that, The step of extracting time-frequency feature vectors based on the set of detail components and the set of approximate components, and reconstructing the time-frequency feature vectors to obtain hierarchical frequency domain features includes: Based on the decomposition level values, calculate the vibration frequency range corresponding to each component in the detailed component set and the approximate component set; Extract the wavelet coefficient sequence corresponding to each component, calculate the sum of squares of all values ​​in the wavelet coefficient sequence, and then divide the sum of squares by the total length of the wavelet coefficient sequence to obtain the energy spectral density value of the corresponding component. Extract the vibration frequency range, energy spectral density value, and amplitude and temporal distribution features extracted from the wavelet coefficient sequence for each component to construct the time-frequency feature vector of the corresponding component. All the time-frequency feature vectors are arranged and recombined in order of decomposition level from low frequency to high frequency to obtain hierarchical frequency domain features.

5. The generator vibration fault diagnosis method according to claim 4, characterized in that, The step of performing noise filtering on the hierarchical detail component set to obtain a denoised detail component set includes: Extract the median value of the absolute values ​​of the wavelet coefficient sequence at each decomposition level; The total length of the wavelet coefficient sequence is extracted as the signal length. Based on the median value and the signal length of the corresponding decomposition level, the preset adaptive threshold calculation function is substituted to determine the noise filtering threshold corresponding to each decomposition level. Compare the amplitude of the wavelet coefficients in the corresponding decomposition level with the noise filtering threshold one by one; If the amplitude of the wavelet coefficient is greater than or equal to the noise filtering threshold, the set of hierarchical detail components corresponding to the wavelet coefficient is retained; if the amplitude of the wavelet coefficient is less than the noise filtering threshold, the set of hierarchical detail components corresponding to the wavelet coefficient is set to zero. The integrated set of layered detail components is used to generate a set of denoised detail components.

6. The generator vibration fault diagnosis method according to claim 1, characterized in that, The boundary correction of the reconstructed vibration signal sequence to obtain a correction signal set includes: Extract the boundary data point set of the first and last preset lengths of the reconstructed vibration signal sequence, and apply the preset mirror continuation algorithm to the boundary data point set to generate extended boundary data; The extended boundary data is subjected to polynomial fitting to obtain boundary correction data; Calculate the amplitude jump rate of the boundary correction data and compare the amplitude jump rate with a preset jump threshold; If the amplitude jump rate does not exceed the jump threshold, the boundary correction data is used to replace the corresponding boundary data point set of the reconstructed vibration signal sequence to obtain the correction signal set. If the amplitude jump rate exceeds the jump threshold, then perform quadratic spline interpolation on the boundary correction data to obtain secondary correction data. Use the secondary correction data to replace the corresponding boundary data point set of the reconstructed vibration signal sequence to obtain the correction signal set.

7. The generator vibration fault diagnosis method according to claim 1, characterized in that, The process involves matching the calibration signal set with operating condition parameters to obtain a fluctuation feature set, performing amplitude analysis on the fluctuation feature set to obtain amplitude features, and identifying the periodic distribution of the amplitude features to obtain a time-domain feature set, including: The duration of the correction signal set is obtained, and the correction signal set is divided into equal intervals based on the duration of the signal to obtain a multi-group segmented signal sequence. The segmented signal sequence is matched with the real-time operating parameters of the generator to generate a set of fluctuation features associated with the operating conditions. Calculate the change in fluctuation amplitude of the fluctuation feature set, and compare the change in fluctuation amplitude with a preset amplitude correction threshold; If the change in fluctuation amplitude does not exceed the amplitude correction threshold, the fluctuation feature set is directly used as the amplitude feature; if the change in fluctuation amplitude exceeds the amplitude correction threshold, amplitude gain correction is performed to obtain the amplitude feature. The amplitude features are periodically distributed to identify periodic distribution feature vectors. The periodic distribution feature vectors are then dimensionlessly mapped to obtain a time-domain feature set.

8. The generator vibration fault diagnosis method according to claim 1, characterized in that, The step of extracting the amplitude deviation sequence from the time-domain feature set, and searching and matching the amplitude deviation sequence with a preset fault mode library to determine the location of potential faulty components includes: The difference between the actual observed amplitude and the healthy baseline amplitude is calculated from the time-domain feature set to obtain the amplitude deviation sequence; The amplitude deviation sequence is compared with a preset error threshold. If the amplitude deviation sequence does not exceed the error threshold, the generator vibration state is determined to be normal. If the amplitude deviation sequence exceeds the error threshold, the amplitude deviation sequence is searched and matched with a preset fault mode library to determine the preliminary fault type and component identifier. Based on the pre-acquired generator equipment topology, the location of the potential faulty component corresponding to the component identifier is determined.

9. The generator vibration fault diagnosis method according to claim 1, characterized in that, The step of performing frequency anomaly matching based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly includes: Based on the index of the location of the potential faulty component, retrieve the historical vibration amplitude records within the corresponding preset period, and combine them with the current real-time vibration data to construct a multi-dimensional vibration state matrix; The multidimensional vibration state matrix is ​​subjected to spectral refinement to obtain the frequency range corresponding to the abnormal vibration; The spectral components within the frequency range are separated into asynchronous components to obtain abnormal spectral components. Extract the nonlinear harmonic component sequence from the abnormal spectral components, and demodulate the impulse pulse sequence from the nonlinear harmonic component sequence; The impact pulse sequence is compared with a preset fault mode fingerprint database to identify the abnormal root cause component of the vibration anomaly in the physical entity that generates the impact.

10. A generator vibration fault diagnosis system, characterized in that, include: The signal correction module is used to acquire the original vibration signal, real-time temperature data and real-time humidity data during generator operation, and to correct the original vibration signal for environmental interference to obtain a preliminary correction dataset. The wavelet decomposition module performs discrete wavelet transform on the preliminary correction dataset to generate a set of detail components and a set of approximate components. The frequency domain feature generation module is used to extract time-frequency feature vectors based on the set of detail components and the set of approximate components, and to reassemble the time-frequency feature vectors to obtain hierarchical frequency domain features, wherein the hierarchical frequency domain features include hierarchical detail component sets corresponding to each decomposition level; The noise suppression module is used to perform noise filtering on the hierarchical detail component set to obtain a denoised detail component set. The signal reconstruction module is used to perform inverse wavelet transform on the denoised detail component set and the approximate component set to obtain a reconstructed vibration signal sequence, and to perform boundary correction on the reconstructed vibration signal sequence to obtain a correction signal set; The time-domain feature extraction module is used to match the operating parameters of the correction signal set to obtain a fluctuation feature set, perform amplitude analysis on the fluctuation feature set to obtain amplitude features, and identify the periodic distribution of the amplitude features to obtain a time-domain feature set. The potential fault location module is used to extract the amplitude deviation sequence from the time domain feature set, and search and match the amplitude deviation sequence with a preset fault mode library to determine the location of the potential faulty component. The fault source localization module is used to perform frequency anomaly matching based on the location of the potential faulty component to determine the abnormal root cause component of the vibration anomaly.

Images