Intelligent diagnosis method for vibration of fan transmission chain based on power distribution box

Abstract

The invention discloses a fan transmission chain vibration intelligent diagnosis method based on a power sub-box, and relates to the technical field of mechanical equipment state monitoring and fault diagnosis, and the method comprises the steps: S1, obtaining a multi-channel vibration signal, collecting an original data flow from the front and rear end positions of a main shaft at the input end and the output end of a gearbox through a sensor, the method comprises the following steps: carrying out power binning on original data according to a power change range during operation of a fan, respectively storing the data in different power intervals, establishing an independent time synchronization reference, and calibrating sampling clock skew of each channel by adopting a clock synchronization algorithm to obtain a preliminary synchronization signal sequence under a unified time reference; according to the intelligent diagnosis method for the vibration of the fan transmission chain based on the power sub-boxes, the fault diagnosis accuracy, the real-time performance and the interpretability of the fan gearbox under the variable power operation condition are remarkably improved, and reliable technical support is provided for health management and intelligent operation and maintenance of wind power equipment.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring and fault diagnosis technology, specifically to a method for intelligent vibration diagnosis of a fan drive chain based on a power distribution box. Background Technology

[0002] In wind power generation equipment, the gearbox is a key component of the wind turbine drivetrain, and its operating status directly affects the energy transfer efficiency and reliability of the entire wind turbine system. During long-term operation, gearboxes are susceptible to faults such as tooth surface wear, bearing fatigue, and poor meshing due to complex loads, variable power output, and vibration shocks. To achieve early identification of equipment health status, vibration monitoring technology is widely used in the operational diagnosis of wind turbine drivetrains.

[0003] However, existing vibration monitoring methods have significant shortcomings in handling multi-channel signals and synchronous analysis under variable power conditions. Traditional vibration monitoring typically relies on a fixed sampling clock or a single trigger signal for data synchronization, which is insufficient to address sampling deviations caused by dynamic changes in wind turbine operating power over time. Especially when vibration signals are simultaneously collected from multiple locations such as the gearbox input, output, and the front and rear ends of the main shaft, the sampling clocks of each sensor are affected by differences in hardware response, signal delay, and power fluctuations, easily leading to data time misalignment and phase differences, thus reducing the accuracy of multi-channel vibration data fusion analysis. Furthermore, most existing diagnostic methods process full-power operating data uniformly, ignoring the impact of power state changes on vibration characteristics. The energy distribution, spectral morphology, and fault characteristics of gearbox vibration signals differ significantly across different power ranges. If vibration data is not managed by power state partitioning, feature aliasing and noise interference can easily occur, leading to inaccurate fault feature extraction and consequently affecting the reliability of fault location and health assessment results. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic method for vibration of the wind turbine drive chain based on power distribution boxes, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent diagnosis of vibration in a wind turbine drive train based on a power distribution box, comprising: S1. Acquire multi-channel vibration signals. Collect raw data streams from the input end to the output end of the gearbox and the front and rear ends of the main shaft using sensors. According to the power variation range during the operation of the fan, divide the raw data into power bins. Store the data of different power ranges separately and establish independent time synchronization references. Use a clock synchronization algorithm to calibrate the sampling clock deviation of each channel to obtain a preliminary synchronization signal sequence under a unified time reference. S2. Based on the preliminary synchronization signal sequence, if there is a difference in buffer delay in the sequence, the data timestamps of each channel are adjusted by the data alignment algorithm to determine the aligned vibration data group. The alignment algorithm integrates the power state change information processing step. S3. Extract dynamic power fluctuation features from the aligned vibration data set, use Fourier transform to obtain frequency domain representation, determine if the frequency peak exceeds the preset threshold, mark it as a potential fault point, and obtain a subset of signals with marked fault features. S4. For the signal subset marked with fault characteristics, cross-channel data fusion is performed. The phase consistency between channels is calculated through correlation analysis to determine the integrated vibration model after fusion. The fusion process is executed independently on a power sub-box basis. S5. Obtain the time-domain waveform change trend from the fused integrated vibration model. If the change trend shows abnormal amplitude, apply wavelet transform to decompose the multi-scale components to obtain the decomposed multi-scale signal components. The decomposition results are stored in correspondence with the power bin label. S6. Based on the decomposed multi-scale signal components, reconstruct the time-consistent vibration signal, use principal component analysis to reduce irrelevant noise, and determine the refined signal representation. S7. Extract fault location indicators from the refined signal representation. By comparing the similarity with the preset fault mode library, if the similarity is higher than the threshold, output a fault warning report divided by power gearbox intervals to obtain the gearbox health status assessment results for each power segment of the transmission chain.

[0006] Preferably, S1 includes: Multi-channel vibration signals are acquired from sensors located at the input and output ends of the gearbox and the front and rear ends of the spindle. The original data stream sequence is generated by acquiring the multi-channel vibration signals using a preset sampling frequency. The original data stream sequence is grouped according to the power variation range during wind turbine operation, and the original data stream sequence is binned by power using a preset power range threshold to obtain a binned data set; A clock synchronization algorithm is used for the binned data set. By calculating the offset of the timestamp of each channel and aligning it to a unified time base, the sampling clock deviation of the binned data set is calibrated to obtain a preliminary synchronization signal sequence. Signal data is extracted from the initial synchronization signal sequence, and the frequency domain features of each channel are generated by Fourier transform. The frequency domain features are compared with the preset synchronization threshold to determine the time synchronization accuracy and obtain the final synchronization signal sequence.

[0007] Preferably, S2 includes: The timestamp data of each channel is obtained from the initial synchronization signal sequence. If there is a difference in buffer delay, the timestamp data is adjusted by a linear interpolation algorithm to obtain a timestamp correction sequence. Based on the timestamp correction sequence, power state change information is used to calculate the power interval weight of each channel, and the power states are fused by a weighted average method to obtain a weighted synchronization sequence. For the weighted synchronization sequence, the Fast Fourier Transform is applied to generate the frequency domain features of each channel. If the frequency domain features deviate from the preset threshold, the time alignment is adjusted by a sliding window to obtain the frequency domain adjusted sequence. By adjusting the frequency domain sequence, the data of each channel is calibrated using the minimum mean square error algorithm to determine the final aligned vibration data set.

[0008] Preferably, S3 includes: The time series is obtained from the aligned vibration data set, and the power fluctuation value of each time window is calculated by short-time Fourier transform to obtain the dynamic power series. For dynamic power sequences, a fast Fourier transform is used to generate a frequency domain representation, and the oscillation frequency and frequency peak are extracted to obtain a frequency domain feature set. If the frequency peak value in the frequency domain feature set exceeds the preset threshold, it is marked as a potential fault point by the data annotation algorithm to obtain the labeled fault sequence; Based on the labeled fault sequences, a clustering algorithm is used to separate the signal subsets with abnormal frequency peaks, and to determine the signal subsets with fault characteristics.

[0009] Preferably, S4 includes: A subset of signals is acquired from a multi-channel sensor and preprocessed. The channel data in the subset of signals is then denoised and normalized to obtain the first cleaned signal dataset. Feature extraction is performed on the first signal dataset, and fault features are extracted from each channel data using fast Fourier transform to obtain a second feature dataset containing fault features. Cross-channel data fusion is performed on the second feature dataset in units of power bins. The phase consistency between channels is calculated by correlation analysis to obtain the fused third phase dataset. A comprehensive vibration model is constructed based on the phase dataset, and the phase consistency data in the third phase dataset is fitted using the least squares method to generate the final vibration model.

[0010] Preferably, S5 includes: Time-domain waveform data is obtained from the fused integrated vibration model, and the time-domain waveform data is converted into frequency-domain signals through Fourier transform to obtain the frequency-domain signal feature set. For a frequency domain signal feature set, if the feature value exceeds a preset threshold, it is judged as an abnormal amplitude, and the time domain waveform segment corresponding to the abnormal amplitude is obtained. Wavelet transform is applied to decompose the time-domain waveform segment to generate multi-scale signal components, thus obtaining a multi-scale signal feature set. Based on the multi-scale signal feature set, power bin classification is performed using the K-means clustering algorithm, and the classification results are stored in the database along with the corresponding labels.

[0011] Preferably, S6 includes: The original vibration signal is decomposed by wavelet transform algorithm to obtain features at different time and frequency scales, resulting in a multi-scale component set. Based on the multi-scale component set, a principal component analysis algorithm is used to perform a linear transformation, projecting the data onto a space with a dimension lower than a preset threshold, and determining the dimensionality-reduced feature representation.

[0012] Preferably, S6 further includes: If the variance contribution rate of the dimensionality-reduced feature representation is lower than the preset threshold, then principal component analysis is used to remove components with low contribution rates to obtain a refined signal representation. By combining multi-scale components in the refined signal representation, the signal is reconstructed according to the time sequence of the original signal to obtain a time-consistent refined vibration signal.

[0013] Preferably, S7 includes: Frequency domain features are extracted from the refined signal representation using the Fourier transform algorithm to obtain the frequency domain feature set of the vibration signal; Based on the frequency domain feature set, time series analysis is used to extract fault location indicators and determine the fault location indicator set.

[0014] Preferably, S7 further includes: If the similarity between the fault location index set and the preset fault mode library is higher than the preset threshold, then the cosine similarity algorithm is used for matching to obtain the matching result. Based on the matching results, the data is divided into power compartments using a compartmentalization algorithm to generate a fault warning report and determine the health status of the gearbox.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This intelligent vibration diagnosis method for wind turbine drivetrain based on power compartmentation effectively solves the problems of time asynchrony, feature mixing, and noise interference of multi-source signals in traditional vibration monitoring by managing wind turbine operating data according to power ranges and independently performing clock synchronization, feature extraction, and multi-channel fusion analysis within each power compartment. First, the vibration data is compartmentalized according to the wind turbine operating power variation law to ensure that data in different power ranges have independent analysis benchmarks. A clock synchronization algorithm is used to calibrate the sampling deviation of each channel, achieving high-precision time alignment within the same power compartment. Fourier transform and wavelet multi-scale analysis are combined to extract frequency and time domain features, and principal component analysis is further used for noise reduction to improve the stability and recognizability of signal features. Subsequently, cross-channel correlation analysis and association with power compartment labels achieve comprehensive modeling and feature fusion of multi-channel vibration signals. Based on this, the feature vector output by the fusion model is compared with a fault mode library to automatically identify potential fault types and locations under different power compartments and generate corresponding early warning reports. This method significantly improves the accuracy, real-time performance, and interpretability of fault diagnosis for wind turbine gearboxes under variable power operation conditions, providing reliable technical support for the health management and intelligent operation and maintenance of wind power equipment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent vibration diagnosis method for wind turbine drive chain based on power distribution box according to the present invention. Detailed Implementation

[0017] 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.

[0018] like Figure 1 As shown, the present invention provides a technical solution: a method for intelligent diagnosis of vibration in a wind turbine drive train based on a power distribution box, comprising: S1. Acquire multi-channel vibration signals. Collect raw data streams from the input end to the output end of the gearbox and the front and rear ends of the main shaft using sensors. According to the power variation range during the operation of the fan, divide the raw data into power bins. Store the data of different power ranges separately and establish independent time synchronization references. Use a clock synchronization algorithm to calibrate the sampling clock deviation of each channel to obtain a preliminary synchronization signal sequence under a unified time reference. S2. Based on the preliminary synchronization signal sequence, if there is a difference in buffer delay in the sequence, the data timestamps of each channel are adjusted by the data alignment algorithm to determine the aligned vibration data group. The alignment algorithm integrates the power state change information processing step. S3. Extract dynamic power fluctuation features from the aligned vibration data set, use Fourier transform to obtain frequency domain representation, determine if the frequency peak exceeds the preset threshold, mark it as a potential fault point, and obtain a subset of signals with marked fault features. S4. For the signal subset marked with fault characteristics, cross-channel data fusion is performed. The phase consistency between channels is calculated through correlation analysis to determine the integrated vibration model after fusion. The fusion process is executed independently on a power sub-box basis. S5. Obtain the time-domain waveform change trend from the fused integrated vibration model. If the change trend shows abnormal amplitude, apply wavelet transform to decompose the multi-scale components to obtain the decomposed multi-scale signal components. The decomposition results are stored in correspondence with the power bin label. S6. Based on the decomposed multi-scale signal components, reconstruct the time-consistent vibration signal, use principal component analysis to reduce irrelevant noise, and determine the refined signal representation. S7. Extract fault location indicators from the refined signal representation. By comparing the similarity with the preset fault mode library, if the similarity is higher than the threshold, output a fault warning report divided by power gearbox intervals to obtain the gearbox health status assessment results for each power segment of the transmission chain.

[0019] This method first collects vibration data using multi-channel sensors deployed at key locations in the wind turbine gearbox. The raw data is then binned according to the actual power changes during wind turbine operation, with each power range corresponding to a data subset and an independent time reference. A clock synchronization algorithm is introduced to calibrate the sampling time deviation between multiple channels, achieving data synchronization under different power levels. Next, an information-enhanced data alignment algorithm that integrates power state changes is employed to address timestamp inconsistencies caused by buffer differences, improving data synchronization accuracy. Subsequently, frequency domain analysis methods (such as Fourier transform) are used to extract signal frequency features, and peak frequencies are used as diagnostic criteria to identify potential fault points. For the labeled fault subsets, cross-channel phase analysis is performed to construct a unified comprehensive vibration model, thereby enhancing the overall perception of the drivetrain system's state. Furthermore, wavelet transform is used to decompose abnormal trend waveforms at multiple scales, obtaining finer-grained signal components, which are stored according to power ranges. Principal component analysis is then used to extract key components, reducing noise interference and obtaining a refined signal representation. Finally, based on the similarity matching between the constructed signal features and the predefined fault mode library, segmented diagnosis and early warning by power range are realized, and the health status assessment results of the gearbox under each power range are output.

[0020] This invention employs a power binning mechanism to finely classify wind turbine operating states, enabling the diagnostic model to adapt to different load conditions and significantly improving diagnostic accuracy. A multi-channel synchronization mechanism and an enhanced data alignment algorithm work together to overcome the information loss problem caused by time asynchrony in traditional vibration signal analysis. By combining frequency and time domain features, supplemented by multi-scale decomposition and principal component dimensionality reduction, the accuracy and robustness of fault feature extraction are effectively improved. Combined with a fault mode matching mechanism, multi-dimensional diagnostic judgments can be achieved, making it particularly suitable for early fault detection and segmented assessment of wind power generation drivetrain systems, thereby improving the system's operational intelligence and wind turbine reliability.

[0021] S1 includes acquiring multi-channel vibration signals from sensors located at the front and rear ends of the main shaft at the input and output ends of the gearbox, and generating the original data stream sequence by acquiring the multi-channel vibration signals using a preset sampling frequency; The original data stream sequence is grouped according to the power variation range during wind turbine operation, and the original data stream sequence is binned by power using a preset power range threshold to obtain a binned data set; A clock synchronization algorithm is used for the binned data set. By calculating the offset of the timestamp of each channel and aligning it to a unified time base, the sampling clock deviation of the binned data set is calibrated to obtain a preliminary synchronization signal sequence. Signal data is extracted from the initial synchronization signal sequence, and the frequency domain features of each channel are generated by Fourier transform. The frequency domain features are compared with the preset synchronization threshold to determine the time synchronization accuracy and obtain the final synchronization signal sequence.

[0022] In one possible implementation, the specific implementation process of S1 is as follows: First, vibration sensors are installed at four locations in the wind turbine drive chain: the input end, output end, front end of the main shaft, and rear end of the main shaft. Each sensor is an industrial-grade triaxial accelerometer with a measurement range of ±50 gravitational acceleration, a preset sampling frequency of 20,000 Hz, and a sampling accuracy of 16 bits. All sensors are connected to a data acquisition terminal through a multi-channel synchronous acquisition system. The acquisition terminal starts sampling with a unified trigger signal, generating a multi-channel raw data stream sequence, with each channel corresponding to the vibration data of one sensor. The acquired raw data includes two parts: a timestamp and an acceleration amplitude, where the timestamp is continuously numbered in milliseconds.

[0023] After data acquisition, the system divides the power range according to preset power interval thresholds based on the real-time power data provided by the wind turbine operation control system. The power interval thresholds are determined based on the wind turbine's rated power and actual operating fluctuation range. For example, when the wind turbine's rated power is 2000 kW, the power range can be divided into 10 intervals, each spanning 200 kW, namely 0 to 200 kW, 200 to 400 kW, 400 to 600 kW, and so on, up to the upper limit of the rated power. The power binning process involves assigning vibration signal samples within the corresponding power interval to the corresponding power binning data set to ensure that signals within the same set correspond to the same operating load state.

[0024] Subsequently, clock synchronization calibration is performed for each power sub-bin dataset. First, the original timestamp data of each channel is read, and the average time interval between adjacent sampling points is calculated to obtain the actual sampling period for each channel. Then, the channel with the smallest sampling period is used as the reference channel, and its time axis is set as a unified time reference. For other channels, the offset between their timestamps and the reference channel is calculated, expressed in milliseconds. The calibration process reallocates the data points of non-reference channels using linear interpolation, ensuring that the data points of all channels correspond consistently at the same time node. After time calibration is completed, a preliminary synchronization signal sequence is formed.

[0025] To further verify synchronization accuracy, fixed-duration data segments are extracted from the initial synchronization signal sequence, typically 10 seconds. Fourier transforms are performed on the data segments of each channel to obtain the spectral distribution results for each channel. The peak position and amplitude of the dominant frequency are extracted from each spectral distribution result, and the difference between the peak positions of the dominant frequencies of adjacent channels is calculated, with the difference measured in Hertz (Hz). A synchronization accuracy threshold is set for the system. This threshold is determined based on the sampling frequency and the gearbox's mechanical rotational speed characteristics. For example, when the sampling frequency is 20,000 Hz and the gearbox's dominant frequency is approximately 500 Hz, the allowable deviation threshold for the dominant frequency is set to 2 Hz. That is, if the difference between the dominant frequencies of any two channels exceeds 2 Hz, a time synchronization error is considered to exist. If the difference is within the threshold range, the time synchronization accuracy meets the requirements, and the signal from this power distribution data set can be retained as the final synchronization signal sequence. If the difference exceeds the threshold, the clock offset is recalculated and the time axis is adjusted until the difference meets the requirements.

[0026] The sampling frequency parameter is determined based on the maximum speed of the gearbox transmission system and the upper limit of the vibration frequency to be captured, typically ensuring that the sampling frequency is at least five times the target highest frequency. The power range threshold is divided according to the rated power of the fan and the power fluctuation range; the number of ranges can be set according to the diagnostic accuracy requirements, generally 10 to 20 ranges. The timestamp offset is determined by comparing the time difference of the first sampling point of each channel with the cumulative sampling period difference. The frequency difference threshold is calculated based on the equipment structural characteristics and the allowable time synchronization error. The conversion method is to multiply the allowable time error by the average frequency change rate, ensuring that the overall system synchronization error is less than one percent of the sampling period.

[0027] The resulting synchronization signal sequence is strictly aligned on the time axis, with consistent peak frequencies across all channels and clear power bin correspondence. Data for each power range is fully synchronized under a unified time base, providing a high-precision data foundation for subsequent fault feature extraction, spectrum analysis, and health assessment.

[0028] S2 includes obtaining timestamp data for each channel from the initial synchronization signal sequence; if there is a difference in buffer delay, the timestamp data is adjusted by a linear interpolation algorithm to obtain a timestamp correction sequence. Based on the timestamp correction sequence, power state change information is used to calculate the power interval weight of each channel, and the power states are fused by a weighted average method to obtain a weighted synchronization sequence. For the weighted synchronization sequence, the Fast Fourier Transform is applied to generate the frequency domain features of each channel. If the frequency domain features deviate from the preset threshold, the time alignment is adjusted by a sliding window to obtain the frequency domain adjusted sequence. By adjusting the frequency domain sequence, the data of each channel is calibrated using the minimum mean square error algorithm to determine the final aligned vibration data set.

[0029] In one possible implementation, the specific implementation process of S2 is as follows: Timestamp data for each channel is extracted from the preliminary synchronization signal sequence obtained in step S1. Each timestamp records the sampling time of each sampling point, with the time unit being milliseconds. First, the buffer delay difference between channels is detected. Buffer delay refers to the time lag caused by processing delays or different buffer queues between data acquisition channels. The detection method is to count the number of sampling points for each channel within the same sampling period and calculate their difference. If the difference in the number of sampling points between any channel and the reference channel exceeds 0.5% of the total number of sampling points, it is determined that the channel has buffer delay. For channels with buffer delay, the system uses a linear interpolation algorithm to correct the timestamps. The linear interpolation algorithm is executed by reading the time interval between two adjacent timestamps and their corresponding sampling point numbers, calculating the theoretical time position that each sampling point should be in within the interval, and then filling in the intermediate timestamps between the two known time points at equal intervals. The interpolation step size is determined by back-calculation from the sampling frequency. For example, when the sampling frequency is 20 kHz, the time interval between adjacent sampling points is 0.05 milliseconds, and the interpolation step size is 0.05 milliseconds. The corrected timestamps are reassigned to the sampling points of the channel, forming a timestamp correction sequence.

[0030] After generating the timestamp correction sequence, the system extracts the power status information corresponding to each timestamp from the wind turbine monitoring system. Each time point corresponds to an active power value in kilowatts. The system divides the power range according to the wind turbine's operating power range and calculates the weight of each channel in the current power range. The weight of the power range is calculated by counting the number of sampling points for each channel within that power range and dividing that number by the total number of sampling points for all channels. For example, if the number of sampling points for four channels in a certain power range are 10,000, 9,800, 10,200, and 9,900 respectively, the corresponding weight values ​​are 0.25, 0.245, 0.255, and 0.247 respectively. After calculation, the system multiplies the power value of each channel at the same time point by the corresponding weight, and then adds the weighted results to obtain a weighted synchronization sequence. Each data point in the weighted synchronization sequence represents the overall power status after power weight correction under a unified time base, reflecting the dynamic characteristics of the overall operating load of the drive train.

[0031] After obtaining the weighted synchronization sequence, the system performs a Fast Fourier Transform (FFT) on the vibration signals of each channel to analyze its frequency domain characteristics. The FFT execution steps are as follows: a fixed-length data window is segmented for analysis, the window length being determined by the sampling frequency, typically one-tenth of the sampling frequency. For example, when the sampling frequency is 20 kHz, each window contains 2,000 sampling points. Each transform outputs the frequency domain amplitude distribution, including the dominant frequency peak and secondary frequency amplitudes. The system extracts the dominant frequency peak from the frequency domain results and compares it with the dominant frequency peak of the weighted synchronization sequence, calculating the frequency difference in Hertz. The system sets a frequency domain deviation threshold to determine whether the frequency offset of each channel exceeds the allowable range. This threshold is determined based on the gearbox's characteristic frequency range; if the gearbox's main vibration frequencies are concentrated in the range of 100 to 1000 Hz, the deviation threshold is set to 1 Hz. If the difference between the dominant frequency peak of a channel and the dominant frequency peak of the weighted synchronization sequence exceeds 1 Hz, it is determined that the channel has a time or frequency alignment error.

[0032] For channels with deviations, the system employs a sliding window approach for time alignment adjustment. The sliding window moves incrementally along the time axis, with a fixed step size of fifty sampling points. Each time it moves, the peak frequency of the corresponding window is recalculated and compared with the reference frequency until the frequency difference is less than a set threshold. After adjustment, a frequency domain adjustment sequence is generated. The step size parameter of the sliding window is determined experimentally to ensure frequency accuracy without introducing new phase transitions; it is typically set to 2.5% of the window length.

[0033] After completing the frequency domain adjustment, the system uses the minimum mean square error (MMS) algorithm to perform final calibration on each channel signal. The specific steps of the MMS algorithm are as follows: A weighted synchronization sequence is selected as the reference channel; the squared amplitude difference between each channel and the reference channel at the same time point is calculated; and the average of the squared differences of all sampling points is used to obtain the error value. Then, the time offset of that channel is adjusted to minimize the error value. The adjustment step size is 0.05 milliseconds, consistent with the sampling period. The error calculation and adjustment process is continuously iterated until the error change between two consecutive iterations is less than one ten-thousandth of the initial error, at which point the algorithm is considered converged. After all channels reach the error minimization state, the finally aligned vibration data set is output.

[0034] All parameters in the above steps have clearly defined bases. The sampling frequency is determined based on the highest vibration frequency of the wind turbine gearbox, ensuring that the sampling frequency is at least five times the target highest frequency. The buffer delay threshold is set to 0.5% of the total number of sampling points to ensure that deviations in the number of sampling points do not affect data integrity. The interpolation step size is calculated from the sampling frequency to ensure consistent time accuracy. The power interval weight is directly calculated from the proportion of the number of sampling points, without the need for empirical coefficients. The frequency domain deviation threshold is determined based on the mechanical characteristics of the transmission system, ensuring that the frequency error is less than the minimum characteristic frequency interval within the system speed fluctuation range. The sliding window step size is determined experimentally to ensure that the frequency error convergence speed is balanced with the calculation accuracy after each adjustment. The minimum mean square error convergence condition is that the error change is less than one ten-thousandth to ensure the stability of time synchronization and amplitude matching. Through the above process, the system achieves precise time alignment and frequency domain consistency correction for multi-channel vibration signals. The final generated vibration data set is strictly synchronized on the time axis and has completely consistent peak values ​​in the frequency domain.

[0035] S3 includes obtaining a time series from the aligned vibration data set, calculating the power fluctuation value of each time window using short-time Fourier transform to obtain a dynamic power series; for the dynamic power series, generating a frequency domain representation using fast Fourier transform, extracting the oscillation frequency and frequency peak to obtain a frequency domain feature set; if the frequency peak in the frequency domain feature set exceeds a preset threshold, it is marked as a potential fault point using a data labeling algorithm to obtain a labeled fault sequence; based on the labeled fault sequence, a clustering algorithm is used to separate the signal subset with abnormal frequency peaks to determine the signal subset with fault characteristics.

[0036] In one possible implementation, the specific process of S3 is as follows: Time-series signals from each channel are read from the final aligned vibration data set obtained in step S2. Each time series consists of a sampling timestamp and acceleration amplitude. The system first determines the length and overlap ratio of the time window based on the sampling frequency and gearbox rotation speed. The sampling frequency is set to 20,000 Hz, and the gearbox spindle speed is 1,000 revolutions per minute (16.67 revolutions per second), with a cycle of 0.06 seconds per revolution. To ensure that each analysis window contains a complete mechanical cycle, the time window length is set to the number of sampling points corresponding to 0.06 seconds, i.e., 1200 sampling points. A 50% overlap is set between adjacent windows, meaning a new window is started by moving 600 sampling points each time. The overlap parameter is used to ensure the continuity of power changes and the accuracy of the analysis.

[0037] Within each time window, the system performs a short-time Fourier transform to convert the signal from the time domain to the frequency domain. The transformation process is as follows: the sampled data for the current window is read and divided into fixed-length segments. Each segment is weighted to ensure a smooth transition at both ends, and then the amplitude of the frequency components is calculated. The square of the amplitude of each frequency component is multiplied by the sampling interval and summed to obtain the total power fluctuation value within that window. After calculating the power fluctuation value once for each window, the values ​​are arranged in chronological order to form a dynamic power sequence. Each value in the dynamic power sequence represents the magnitude of the signal's vibration energy during that time period, expressed as the square of acceleration multiplied by seconds. This sequence reflects the energy change trend of the wind turbine drivetrain at different times.

[0038] After obtaining the dynamic power sequence, the system performs a Fast Fourier Transform (FFT) to convert the power-time sequence into a frequency domain signal, obtaining each frequency component and its amplitude. The system extracts the main oscillation frequency (the frequency corresponding to the maximum amplitude) from the frequency domain results, and simultaneously extracts the peak amplitude of each frequency component. All extracted results form a frequency domain feature set to describe the periodic characteristics of power fluctuations. To identify abnormal fluctuations, the system sets a frequency peak threshold. This threshold is determined by statistically analyzing historical data of the wind turbine under normal operating conditions. Specifically, vibration data is continuously collected for at least 24 hours under normal wind turbine operating conditions, and the frequency domain peak amplitude is calculated for all time windows. The average and standard deviation of these amplitudes are then calculated. The average plus twice the standard deviation is then used as the threshold. For example, when the average amplitude is 0.8 gravitational acceleration squared per Hz and the standard deviation is 0.15 gravitational acceleration squared per Hz, the threshold is set to 1.1 gravitational acceleration squared per Hz. When the amplitude of any frequency in the frequency domain feature set exceeds 1.1, the system determines that point as a potential abnormal peak.

[0039] Anomaly peaks are labeled using a data annotation algorithm. The algorithm examines all frequency points in the frequency domain feature set one by one. When the amplitudes of three or more consecutive adjacent frequency points exceed a threshold, the system designates these frequency points as an abnormal frequency bandwidth and records their start frequency, end frequency, and corresponding time window number. Then, in the time series, vibration signal segments belonging to this time window are labeled with a value of 1, indicating a potential fault; unlabeled time periods are labeled with a value of 0, indicating a normal state. The resulting time series is the labeled fault sequence.

[0040] To further extract abnormal signals, the system performs cluster analysis on the labeled fault sequences. The clustering algorithm proceeds as follows: First, all time periods labeled as 1 are extracted, and the amplitude of the corresponding frequency peaks is calculated. These amplitudes are standardized to values ​​between 0 and 1 to eliminate amplitude differences between different power gearboxes. Then, the absolute value of the amplitude difference between any two samples is calculated as a distance metric. The system sets a clustering number parameter, which is determined based on the number of main components of the gearbox. When the gearbox includes a first-stage planetary gear and a first-stage high-speed gear, the clustering number is set to 2. The system uses the minimum distance criterion for grouping, classifying samples whose distance is less than a set distance threshold into the same cluster. The distance threshold is determined by statistically analyzing the standard deviation of the labeled samples, taking 50% of the standard deviation as the threshold. For example, if the standard deviation of the labeled samples is 0.2, the distance threshold is 0.1. After clustering, the system calculates the average amplitude of each cluster as the cluster center value, and selects the cluster with the highest average amplitude as the subset of fault feature signals.

[0041] In the above process, the sampling frequency of 20,000 Hz is determined based on the highest frequency of gearbox vibration to ensure that the Nyquist sampling conditions are met; the time window length of 1,200 sampling points is determined by the spindle speed; the overlap ratio of 50% is a fixed value to ensure the continuity of analysis; the frequency peak threshold is automatically calculated based on the statistical results of the peak distribution under normal operating conditions; three consecutive points exceeding the threshold are fixed conditions for anomaly judgment to prevent misjudgment due to single-point noise; the cluster quantity parameter is determined by the number of layers in the gearbox structure; the distance threshold is taken as 50% of the sample standard deviation to ensure the stability and reliability of the cluster boundary.

[0042] Finally, through continuous processing including short-time Fourier transform, fast Fourier transform, data annotation, and cluster analysis, the system accurately extracts power fluctuation characteristics from the aligned vibration signals, identifies abnormal energy frequency bands, and separates abnormal signal segments into independent fault feature signal subsets. These signal subsets contain time-varying information about sudden increases in vibration energy or abnormal frequency changes, directly reflecting fault symptoms such as gear meshing imbalance, bearing wear, or structural loosening within the transmission chain system. This provides a high-precision data foundation for subsequent location and qualitative diagnosis.

[0043] S4 includes acquiring a subset of signals from a multi-channel sensor and performing data preprocessing, and then processing the channel data in the subset of signals by denoising and standardizing to obtain a cleaned first signal dataset. Feature extraction is performed on the first signal dataset, and fault features are extracted from each channel data using fast Fourier transform to obtain a second feature dataset containing fault features. Cross-channel data fusion is performed on the second feature dataset in units of power bins. The phase consistency between channels is calculated by correlation analysis to obtain the fused third phase dataset. A comprehensive vibration model is constructed based on the phase dataset, and the phase consistency data in the third phase dataset is fitted using the least squares method to generate the final vibration model.

[0044] In one possible implementation, the specific implementation process of S4 is as follows: The system first reads the raw signal data of the multi-channel sensors from the fault feature signal subset obtained in the aforementioned steps. Each channel corresponds to the sensor signals at the input end, output end, front end of the spindle, and rear end of the spindle. To ensure the accuracy of subsequent analysis results, the system performs data preprocessing on the signal subset, including denoising and standardization. The denoising process uses a fixed window averaging filter method, replacing the original value of the center point with the average of the amplitudes of five consecutive sampling points to eliminate high-frequency interference components. If the sampling frequency is 20000 Hz, the time length corresponding to the five-point window is 0.25 milliseconds, which can effectively smooth high-frequency noise. For low-frequency trend interference, the system uses a bandpass filter to retain signal components in the range of 100 to 1000 Hz, which covers the gear meshing and bearing rotation frequencies. The standardization process calculates the mean and standard deviation of each channel signal, subtracts the mean from the signal amplitude, and divides by the standard deviation to make the mean of all channel signals 0 and the standard deviation 1, thus obtaining the cleaned first signal dataset.

[0045] Subsequently, feature extraction is performed on the first signal dataset. The system performs a Fast Fourier Transform (FFT) on the standardized signal for each channel, converting the time-domain signal into a frequency-domain representation. The transform result includes amplitude information for each frequency component. The system identifies the main energy concentration region in the spectrum and extracts parameters such as the peak amplitude of the dominant frequency, the harmonic energy ratio, and the bandwidth energy distribution. The peak amplitude of the dominant frequency is the value at the frequency point corresponding to the largest amplitude; the harmonic energy ratio is the ratio of the amplitude at frequencies two and three times the dominant frequency to the dominant frequency amplitude; and the bandwidth energy distribution is the percentage of energy within ±10 Hz of the dominant frequency relative to the total energy. All extracted results constitute the second feature dataset, used to represent the fault characteristics of different channels.

[0046] Next, the system performs cross-channel data fusion on a power bin-by-bin basis. Each power bin contains data within the same operating power range. For each power range, the system calculates the correlation between the characteristics of each channel. The correlation calculation is performed using the Pearson correlation coefficient method. This method compares the numerical change trends of the channel signal feature sequences pairwise to calculate their degree of coordination. The system converts the correlation coefficient values ​​into a phase consistency index, which is expressed as an absolute value ranging from 0 to 1, where 1 represents complete consistency and 0 represents complete inconsistency. After the calculation is completed, the phase consistency results between all channels are stored according to power range, forming the fused third phase dataset.

[0047] Finally, the system constructs a comprehensive vibration model based on the third phase dataset. This comprehensive vibration model represents the overall phase coupling characteristics of multi-channel signals under different power bins. During model building, the average phase coherence of each power bin is used as input data, and the corresponding average vibration energy is used as output data. The least squares method is employed for fitting. The specific implementation of the least squares method is as follows: all data points from all power bins are input into the algorithm, the residual value for each data point is calculated (i.e., the square of the difference between the actual vibration energy and the fitted value), and then all squared residuals are summed. The fitting curve parameters are adjusted to minimize the sum of squared residuals. The system uses a linear fitting method, where the slope of the fitting function represents the influence of phase coherence changes on vibration energy, and the intercept represents the baseline vibration energy value under no phase difference condition. After fitting, the final comprehensive vibration model is generated.

[0048] In the above process, the filter window length is determined based on the sampling frequency and noise frequency. When the sampling frequency is 20,000 Hz, a window of 5 points can effectively suppress high-frequency interference without losing mechanical characteristics. The bandpass filtering range of 100 to 1,000 Hz is determined by the vibration characteristics of the gearbox structure, covering the main mechanical vibration frequency bands. The standardization parameters are calculated from the actual signal mean and standard deviation, requiring no manual setting. The frequency resolution of the Fast Fourier Transform is determined by the sampling frequency and the number of sampling points; for example, with 2,000 sampling points, the resolution is 10 Hz. Correlation analysis uses a fixed calculation formula and does not rely on empirical parameters. The phase consistency threshold is set to 0.8 as the criterion for high consistency judgment. This value is experimentally verified and can effectively distinguish the phase synchronization state between normal and abnormal channels. The convergence condition for least squares fitting is that the sum of squared residuals of two consecutive iterations changes by less than 0.0001, ensuring the stability of the fitting results.

[0049] The above process enables the system to extract features from multi-channel vibration signals and fuse phase correlations within power ranges, ultimately forming a comprehensive vibration model that quantitatively reflects the overall dynamic characteristics of the system. This model can accurately reflect the phase coordination and energy distribution patterns of the gearbox transmission chain within different power ranges, providing a precise analytical basis for subsequent health assessments and intelligent diagnostics.

[0050] S5 includes obtaining time-domain waveform data from the fused integrated vibration model, converting the time-domain waveform data into a frequency-domain signal through Fourier transform, and obtaining a frequency-domain signal feature set; For a frequency domain signal feature set, if the feature value exceeds a preset threshold, it is judged as an abnormal amplitude, and the time domain waveform segment corresponding to the abnormal amplitude is obtained. Wavelet transform is applied to decompose the time-domain waveform segment to generate multi-scale signal components, thus obtaining a multi-scale signal feature set. Based on the multi-scale signal feature set, power bin classification is performed using the K-means clustering algorithm, and the classification results are stored in the database along with the corresponding labels.

[0051] In one possible implementation, the specific implementation process of S5 is as follows: The system first extracts time-domain waveform data from the comprehensive vibration model generated in step S4. Each waveform data corresponds to a vibration response sequence under a power sub-bin interval, including a timestamp and acceleration amplitude. For frequency domain analysis, the system performs a Fourier transform on the time-domain waveform data, converting the time-domain signal into a frequency-domain signal. The Fourier transform calculation process is as follows: the time-domain waveform is divided into multiple time periods of fixed length, the length of each time period is determined by the sampling frequency. When the sampling frequency is 20000 Hz, each segment length takes 2048 sampling points to ensure a frequency resolution of approximately 9.77 Hz. After transformation, each signal segment outputs frequency components and their corresponding amplitudes. The system extracts the amplitude of each frequency component to form a frequency domain signal feature set. Each element of the frequency domain signal feature set consists of a frequency value, amplitude, and energy percentage, used to characterize the energy distribution characteristics of the time-domain waveform at different frequencies.

[0052] The system identifies signal anomalies based on a frequency domain signal feature set. To this end, a feature value threshold is set to determine abnormal amplitudes. The feature value threshold is determined as follows: under normal wind turbine operation, at least 24 hours of historical vibration data are collected, and the mean and standard deviation of the frequency domain peak amplitude are calculated for all power distribution intervals. The mean is then multiplied by twice the standard deviation to obtain the threshold. For example, if the mean frequency domain peak amplitude during normal operation is 0.85 gravitational acceleration squared per Hz and the standard deviation is 0.1 gravitational acceleration squared per Hz, then the threshold is set to 1.05 gravitational acceleration squared per Hz. When the amplitude of any frequency component exceeds 1.05, the system determines that the amplitude corresponding to that frequency is abnormal and extracts a time-domain waveform segment within the time window containing the abnormal frequency. The length of the time-domain waveform segment is set to 2048 sampling points, corresponding to a time length of 0.1024 seconds, to ensure that the complete abnormal signal cycle is included.

[0053] For the extracted time-domain waveform segments, the system performs wavelet transform decomposition. The wavelet transform process is as follows: a fixed mother wavelet function is selected to decompose the signal into several sub-signals of different scales. The mother wavelet function adopts the Daubechies 4 type, with a decomposition scale of 4 levels. The sub-signals obtained from each level of decomposition represent signal components in different frequency ranges, with the first level containing high-frequency detail signals and the fourth level containing low-frequency trend signals. The system calculates the average energy, energy proportion, and standard deviation of each level of signal, forming a multi-scale signal feature set. Each element of the multi-scale signal feature set includes a scale number, corresponding energy value, and energy distribution ratio, which can reflect the energy variation law of the vibration signal at different frequency levels.

[0054] After obtaining the multi-scale signal feature set, the system performs power bin classification based on the K-means clustering algorithm. The clustering process first determines the number of clusters K, which is equal to the number of power bins. For example, when the wind turbine operating power range is divided into 10 intervals, K is set to 10. The system uses the multi-scale feature vector of each sample as input data, calculates the Euclidean distance between the sample and each cluster center, and assigns the sample to the cluster center with the smallest distance. The initial centroids of the clustering process are set by randomly selecting sample features from different power intervals, and then iteratively updated. Each iteration recalculates the feature mean of each cluster center and reclassifies the clusters until the cluster labels of all samples no longer change or the change in centroids between two consecutive iterations is less than one ten-thousandth. After clustering is complete, the system maps the classification results to power bin labels and stores each classification result in a database. Each record in the database contains a power interval number, multi-scale feature value, anomaly marker, and cluster category label.

[0055] In the above process, each parameter has a clearly defined basis. The sampling frequency of 20,000 Hz is determined by the upper limit of the gearbox vibration frequency, ensuring that the sampling rate is at least five times the highest characteristic frequency. The Fourier transform window length of 2048 sampling points ensures that the frequency resolution is higher than 10 Hz. The eigenvalue threshold is determined by statistically analyzing the peak distribution under normal operating conditions, with the mean plus twice the standard deviation as the boundary. The time-domain waveform segment length of 2048 sampling points is consistent with the transform window to ensure the integrity of the signal period. The wavelet decomposition level is set to 4 levels, determined by the signal bandwidth, so that the lowest level frequency after decomposition covers low-frequency features below 100 Hz. The mother wavelet type is fixed as Daubechies 4, because its smoothness and tight support characteristics are suitable for mechanical vibration signals. The number of clusters K is consistent with the number of power bins to ensure that the classification results correspond to the power range. The cluster iteration convergence condition is that the center variation is less than one ten-thousandth to ensure computational stability.

[0056] Through the above process, the system can extract time-domain waveform features from the comprehensive vibration model. After Fourier transform, anomaly detection, wavelet multi-scale decomposition, and clustering classification, the vibration signal features under different power ranges are systematically organized. The final generated database records contain multi-scale energy distribution and abnormal vibration characteristics under each power sub-box, providing accurate and quantifiable basic data support for the health status assessment of the wind turbine drivetrain.

[0057] S6 includes decomposing the original vibration signal using a wavelet transform algorithm to obtain features at different time and frequency scales, resulting in a multi-scale component set; Based on the multi-scale component set, a linear transformation is performed using the principal component analysis algorithm, projecting the data onto a space with a dimension lower than a preset threshold, and determining the dimensionality-reduced feature representation. If the variance contribution rate of the dimensionality-reduced feature representation is lower than the preset threshold, then principal component analysis is used to remove components with low contribution rates to obtain a refined signal representation. By combining multi-scale components in the refined signal representation, the signal is reconstructed according to the time sequence of the original signal to obtain a time-consistent refined vibration signal.

[0058] In one possible implementation, the specific implementation process of S6 is as follows: The system first extracts the original vibration signal data from the database obtained in step S5. Each signal contains sampling time series and acceleration amplitude data, with a sampling frequency of 20,000 Hz. To extract features at different time and frequency scales, the system uses a wavelet transform algorithm to decompose the original signal. Specifically, the Daubechies 4 type mother wavelet function is selected to decompose the original signal into 6 layers of wavelet coefficients. Each layer of wavelet components corresponds to a different frequency range, where the first layer represents high-frequency detail signals from 1000 to 2000 Hz, the second layer represents sub-high-frequency signals from 500 to 1000 Hz, the third layer represents mid-frequency signals from 250 to 500 Hz, the fourth layer represents low-mid-frequency signals from 125 to 250 Hz, the fifth layer represents low-frequency signals from 62 to 125 Hz, and the sixth layer represents trend signals below 62 Hz. The system calculates the average energy, standard deviation, and kurtosis of each layer of wavelet components to form a multi-scale component set. Each element in this set consists of a layer number, energy value, standard deviation, and kurtosis, used to describe the characteristic distribution of the signal at different frequency levels.

[0059] Subsequently, the system performs principal component analysis (PCA) on the multi-scale component set to achieve dimensionality reduction. The PCA algorithm works as follows: first, it calculates the covariance matrix among the features in the multi-scale component set; then, it obtains the eigenvalues ​​and eigenvectors of this covariance matrix. The magnitude of the eigenvalue represents the contribution of the corresponding principal component to the overall variance, and the eigenvector represents the direction of the principal component in the original feature space. The system sorts all eigenvalues ​​from largest to smallest and calculates the cumulative variance contribution rate sequentially. The variance contribution rate threshold is set to 95%, meaning the number of principal components corresponding to a cumulative variance contribution rate of 95% is determined as the final number of retained principal components. If a principal component in the dimensionality-reduced feature representation has a variance contribution rate below 5%, the system removes that principal component to avoid low-contribution features affecting the overall signal accuracy. The retained principal components after removal constitute the dimensionality-reduced feature representation. This dimensionality reduction result can retain the main variation features of the signal to the greatest extent while significantly reducing redundant dimensions.

[0060] To ensure temporal consistency in signal reconstruction, the system combines the dimensionality-reduced principal components to recover a refined signal representation. The reconstruction process follows the time sequence of the original signal: first, based on the transformation matrix of principal component analysis, the retained principal components are linearly combined back into time-domain signal components; then, each component is superimposed according to its corresponding frequency band position at the wavelet decomposition level to obtain the complete time-domain reconstructed signal. The reconstructed signal maintains a one-to-one correspondence with the original sampling points in time, and its amplitude is normalized and inversely calculated to restore it to the actual acceleration unit. The final output signal is the time-consistent refined vibration signal.

[0061] All parameters in the above process have clear determination criteria. The sampling frequency of 20,000 Hz is determined by the highest characteristic frequency of the gearbox vibration signal, ensuring that the sampling rate is not less than five times the target frequency. The number of wavelet decomposition layers is set to 6, determined by the frequency bandwidth distribution range and calculation accuracy, so that the frequency band of each layer is approximately half that of the previous layer. The mother wavelet is selected as Daubechies 4 type, because its smoothness and tight support characteristics are suitable for the decomposition of mechanical vibration signals. The variance contribution rate threshold of 95% is a conventional empirical value for principal component analysis, which can effectively balance the dimensionality reduction effect and information preservation. The low contribution rate removal threshold of 5% ensures that only components with minimal impact on signal characteristics are deleted. During the time series reconstruction process, the linear combination of principal components uses the forward reconstruction method to ensure that the signal time sequence and sampling interval are completely consistent.

[0062] Through the above process, feature extraction, dimensionality reduction analysis, and refined reconstruction of multi-scale vibration signals were achieved. The refined signal effectively removed redundant and noise components while preserving the key features of the original signal. The resulting time-consistent refined vibration signal has a high signal-to-noise ratio and structural integrity, providing a high-quality input data foundation for subsequent fault location and health assessment.

[0063] S7 includes extracting frequency domain features from the refined signal representation using a Fourier transform algorithm to obtain a set of frequency domain features of the vibration signal; Based on the frequency domain feature set, time series analysis is used to extract fault location indicators and determine the fault location indicator set. If the similarity between the fault location index set and the preset fault mode library is higher than the preset threshold, then the cosine similarity algorithm is used for matching to obtain the matching result. Based on the matching results, the data is divided into power compartments using a compartmentalization algorithm to generate a fault warning report and determine the health status of the gearbox.

[0064] In one possible implementation, the specific implementation process of S7 is as follows: The system first extracts frequency domain features from the time-consistent refined vibration signal obtained in step S6. To this end, the system performs a Fourier transform on the refined signal under each power sub-bin, converting the time-domain signal into a frequency-domain signal. The Fourier transform process is as follows: Vibration sampling data within each power sub-bin is read. Assuming a sampling frequency of 20000 Hz, 4096 sampling points are taken as the analysis window, with each window corresponding to a time length of 0.2048 seconds. After performing a Fourier transform on each window, frequency components and their corresponding amplitudes are obtained. The square of the amplitude of each frequency component is calculated and multiplied by the sampling interval to obtain the energy spectrum. Parameters such as the dominant frequency peak, harmonic amplitude, spectral bandwidth energy ratio, and energy concentration are statistically analyzed. The dominant frequency peak represents the frequency at which the signal energy is highest; the harmonic amplitude is the amplitude at an integer multiple of the dominant frequency; the bandwidth energy ratio is the proportion of energy within ±10 Hz of the dominant frequency to the total energy; and the energy concentration is the percentage of the dominant frequency energy to the total energy. All extracted results form a frequency domain feature set, which is used to reflect the energy distribution characteristics of vibration signals in the frequency domain under different power ranges.

[0065] After acquiring the frequency domain feature set, the system uses time series analysis to extract fault location indicators. This process includes two parts: signal autocorrelation calculation and power spectral density analysis. Autocorrelation calculation measures the periodic repetition of the signal over time. The system calculates the autocorrelation coefficient curve by multiplying the refined signal point-by-point with its own time-delayed version and averaging the results. The peak position of the autocorrelation curve corresponds to the dominant period of the signal. Power spectral density analysis calculates the distribution ratio of frequency domain energy across different frequency ranges, extracting representative frequency peaks and harmonic ratios. Subsequently, the system generates a fault location indicator set from four features: autocorrelation peak spacing, power spectral density peak position, harmonic ratio, and energy concentration. This set reflects the correspondence between vibration signals and mechanical characteristics such as gear meshing frequencies and bearing defect frequencies.

[0066] The system performs similarity matching between the set of fault location indicators and a pre-defined fault mode library. The fault mode library consists of a large number of frequency and time domain features of known fault conditions, with each fault mode containing typical dominant frequency, harmonic characteristics, and energy distribution patterns. The system first calculates the cosine similarity between the current fault location indicator and each mode vector in the library. The cosine similarity algorithm calculates the similarity by multiplying the two feature vectors to be matched component by component, summing the results, and then dividing by the product of the magnitudes of the two vectors. The result ranges from 0 to 1. The closer the similarity is to 1, the more similar the current signal feature is to that mode. The system sets a similarity threshold of 0.9 as the judgment standard. This threshold was determined experimentally and can effectively distinguish between normal operating conditions, minor faults, and severe faults. If the similarity of any fault mode is greater than 0.9, the system determines that a fault of the corresponding type exists and records the matching result, including the fault type, matching mode number, and similarity value.

[0067] After obtaining the matching results, the system uses a binning algorithm to classify the fault information according to power ranges. The execution process of the binning algorithm is as follows: read the timestamp information in the matching results, determine the corresponding power range based on the wind turbine operating power record, and aggregate the matching results within the same power bin. The system counts the frequency of fault occurrence, average similarity, and highest similarity for each power range, and calculates the health index for that power range. The health index is defined as 1 minus the average similarity value, with a value ranging from 0 to 1. The closer the value is to 0, the healthier the equipment; the closer the value is to 1, the higher the fault risk. The system generates a fault warning report based on the health index of each power bin. The report includes the power range number, corresponding fault type, frequency of occurrence, average similarity, health index, and diagnostic level. The diagnostic level is divided into three levels: healthy, minor fault, and severe fault, with thresholds set as follows: a health index less than 0.3 is healthy, 0.3 to 0.6 is minor fault, and greater than 0.6 is severe fault.

[0068] All parameters in the above process have clear determination criteria. The sampling frequency of 20,000 Hz ensures that the spectral resolution can cover the main vibration frequency range of the gearbox; the Fourier transform window length of 4,096 points ensures that the frequency resolution is better than 10 Hz; the similarity threshold of 0.9 was determined by comparing a large number of normal and faulty samples, which can accurately identify abnormal patterns; the power distribution interval is derived from the division of the wind turbine's rated power into 10 or 20 equal parts; the health index calculation formula is fixed, and the diagnostic level threshold is calibrated through empirical statistics.

[0069] Through the above process, the system realizes a complete analysis workflow from refining vibration signals to fault identification and health assessment. By extracting multidimensional features through Fourier transform and time series analysis, and combining cosine similarity matching and power-level clustering, the system can automatically generate segmented health assessment results for the transmission chain gearbox under different power conditions. The final output fault warning report not only clearly indicates the fault type and severity, but also quantifies the health level of the equipment in each power range, providing accurate and traceable diagnostic basis for wind turbine operation and maintenance.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent diagnosis of vibration in a wind turbine drive train based on a power distribution box, characterized in that, include: S1. Acquire multi-channel vibration signals. Collect raw data streams from the input end to the output end of the gearbox and the front and rear ends of the main shaft using sensors. According to the power variation range during the operation of the fan, divide the raw data into power bins. Store the data of different power ranges separately and establish independent time synchronization references. Use a clock synchronization algorithm to calibrate the sampling clock deviation of each channel to obtain a preliminary synchronization signal sequence under a unified time reference. S2. Based on the preliminary synchronization signal sequence, if there is a difference in buffer delay in the sequence, the data timestamps of each channel are adjusted by the data alignment algorithm to determine the aligned vibration data group. The alignment algorithm integrates the power state change information processing step. S3. Extract dynamic power fluctuation features from the aligned vibration data set, use Fourier transform to obtain frequency domain representation, determine if the frequency peak exceeds the preset threshold, mark it as a potential fault point, and obtain a subset of signals with marked fault features. S4. For the signal subset marked with fault characteristics, cross-channel data fusion is performed. The phase consistency between channels is calculated through correlation analysis to determine the integrated vibration model after fusion. The fusion process is executed independently on a power sub-box basis. S5. Obtain the time-domain waveform change trend from the fused integrated vibration model. If the change trend shows abnormal amplitude, apply wavelet transform to decompose the multi-scale components to obtain the decomposed multi-scale signal components. The decomposition results are stored in correspondence with the power bin label. S6. Based on the decomposed multi-scale signal components, reconstruct the time-consistent vibration signal, use principal component analysis to reduce irrelevant noise, and determine the refined signal representation. S7. Extract fault location indicators from the refined signal representation. By comparing the similarity with the preset fault mode library, if the similarity is higher than the threshold, output a fault warning report divided by power gearbox intervals to obtain the gearbox health status assessment results for each power segment of the transmission chain.

2. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box as described in claim 1, characterized in that: S1 includes: Multi-channel vibration signals are acquired from sensors located at the input and output ends of the gearbox and the front and rear ends of the spindle. The original data stream sequence is generated by acquiring the multi-channel vibration signals using a preset sampling frequency. The original data stream sequence is grouped according to the power variation range during wind turbine operation, and the original data stream sequence is binned by power using a preset power range threshold to obtain a binned data set; A clock synchronization algorithm is used for the binned data set. By calculating the offset of the timestamp of each channel and aligning it to a unified time base, the sampling clock deviation of the binned data set is calibrated to obtain a preliminary synchronization signal sequence. Signal data is extracted from the initial synchronization signal sequence, and the frequency domain features of each channel are generated by Fourier transform. The frequency domain features are compared with the preset synchronization threshold to determine the time synchronization accuracy and obtain the final synchronization signal sequence.

3. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box as described in claim 1, characterized in that: S2 includes: The timestamp data of each channel is obtained from the initial synchronization signal sequence. If there is a difference in buffer delay, the timestamp data is adjusted by a linear interpolation algorithm to obtain a timestamp correction sequence. Based on the timestamp correction sequence, power state change information is used to calculate the power interval weight of each channel, and the power states are fused by a weighted average method to obtain a weighted synchronization sequence. For the weighted synchronization sequence, the Fast Fourier Transform is applied to generate the frequency domain features of each channel. If the frequency domain features deviate from the preset threshold, the time alignment is adjusted by a sliding window to obtain the frequency domain adjusted sequence. By adjusting the frequency domain sequence, the data of each channel is calibrated using the minimum mean square error algorithm to determine the final aligned vibration data set.

4. The intelligent vibration diagnosis method for wind turbine drive train based on power distribution box according to claim 1, characterized in that: S3 includes: The time series is obtained from the aligned vibration data set, and the power fluctuation value of each time window is calculated by short-time Fourier transform to obtain the dynamic power series. For dynamic power sequences, a fast Fourier transform is used to generate a frequency domain representation, and the oscillation frequency and frequency peak are extracted to obtain a frequency domain feature set. If the frequency peak value in the frequency domain feature set exceeds the preset threshold, it is marked as a potential fault point by the data annotation algorithm to obtain the labeled fault sequence; Based on the labeled fault sequences, a clustering algorithm is used to separate the signal subsets with abnormal frequency peaks, and to determine the signal subsets with fault characteristics.

5. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box as described in claim 1, characterized in that: S4 includes: A subset of signals is acquired from a multi-channel sensor and preprocessed. The channel data in the subset of signals is then denoised and normalized to obtain the first cleaned signal dataset. Feature extraction is performed on the first signal dataset, and fault features are extracted from each channel data using fast Fourier transform to obtain a second feature dataset containing fault features. Cross-channel data fusion is performed on the second feature dataset in units of power bins. The phase consistency between channels is calculated by correlation analysis to obtain the fused third phase dataset. A comprehensive vibration model is constructed based on the phase dataset, and the phase consistency data in the third phase dataset is fitted using the least squares method to generate the final vibration model.

6. The intelligent vibration diagnosis method for wind turbine drive train based on power distribution box according to claim 1, characterized in that: S5 includes: Time-domain waveform data is obtained from the fused integrated vibration model, and the time-domain waveform data is converted into frequency-domain signals through Fourier transform to obtain the frequency-domain signal feature set. For a frequency domain signal feature set, if the feature value exceeds a preset threshold, it is judged as an abnormal amplitude, and the time domain waveform segment corresponding to the abnormal amplitude is obtained. Wavelet transform is applied to decompose the time-domain waveform segment to generate multi-scale signal components, thus obtaining a multi-scale signal feature set. Based on the multi-scale signal feature set, power bin classification is performed using the K-means clustering algorithm, and the classification results are stored in the database along with the corresponding labels.

7. The intelligent vibration diagnosis method for wind turbine drive train based on power distribution box according to claim 1, characterized in that: S6 includes: The original vibration signal is decomposed by wavelet transform algorithm to obtain features at different time and frequency scales, resulting in a multi-scale component set. Based on the multi-scale component set, a principal component analysis algorithm is used to perform a linear transformation, projecting the data onto a space with a dimension lower than a preset threshold, and determining the dimensionality-reduced feature representation.

8. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box as described in claim 7, characterized in that: S6 further includes: If the variance contribution rate of the dimensionality-reduced feature representation is lower than the preset threshold, then principal component analysis is used to remove components with low contribution rates to obtain a refined signal representation. By combining multi-scale components in the refined signal representation, the signal is reconstructed according to the time sequence of the original signal to obtain a time-consistent refined vibration signal.

9. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box according to claim 1, characterized in that: S7 includes: Frequency domain features are extracted from the refined signal representation using the Fourier transform algorithm to obtain the frequency domain feature set of the vibration signal; Based on the frequency domain feature set, time series analysis is used to extract fault location indicators and determine the fault location indicator set.

10. The intelligent vibration diagnosis method for a wind turbine drive train based on a power distribution box according to claim 9, characterized in that: The S7 also includes: If the similarity between the fault location index set and the preset fault mode library is higher than the preset threshold, then the cosine similarity algorithm is used for matching to obtain the matching result. Based on the matching results, the data is divided into power compartments using a compartmentalization algorithm to generate a fault warning report and determine the health status of the gearbox.

Images