A wind turbine blade fault diagnosis method, device and wind turbine generator

Abstract

The application provides a wind turbine blade fault diagnosis method and device and a wind turbine, which comprises the following steps: acquiring blade multi-period vibration signals and synchronous speed signals and splicing them into full-period signals; performing time-frequency analysis to obtain a time-frequency spectrum; extracting frequency values of each time point in a preset frequency band of a first-order natural frequency of the blade to obtain an original ridge line sequence and filter the original ridge line sequence to obtain a stable ridge line sequence; screening a continuous period with a speed fluctuation not exceeding a preset value from the speed signal as a constant speed interval, and extracting a corresponding segmented frequency sequence as an evaluation sequence; taking a frequency average value of at least one segment before a current segment of the evaluation sequence as a dynamic reference to calculate a frequency change percentage of the current segment; and outputting a fault diagnosis result according to a change direction and amplitude of the percentage.

Description

Technical Field

[0001] This application relates to the field of wind power generation, specifically to a method, device, and wind turbine generator set for diagnosing wind turbine blade faults. Background Technology

[0002] Wind turbine blades are prone to structural damage such as cracks, fissures, and fractures during long-term operation. They are also susceptible to mass-increasing faults such as icing, frost accumulation, fouling, and water accumulation at the blade roots. Timely and accurate tracking of changes in the blade's natural frequency is a key technical approach to achieving early warning of these faults.

[0003] Existing technologies employ short-time Fourier transform (SFT) analysis of blade vibration signals, but these typically use default parameters or empirical window lengths, failing to optimize for the low-frequency characteristics of the blade's natural frequency (usually below 2Hz). This results in insufficient frequency resolution, making it difficult to accurately capture minute shifts in the natural frequency. Regarding frequency ridge extraction, existing methods often employ simple peak search methods, which suffer from poor noise immunity, low accuracy, and inability to operate stably in complex field environments. Furthermore, the blade's natural frequency increases with rotational speed, and existing technologies do not effectively address the modulation effect of rotational speed variations on the frequency. When frequency data from different rotational speeds are mixed, the extracted frequency trajectory fluctuates with rotational speed changes, leading to a high false alarm rate and poor robustness in damage identification. More critically, existing technologies generally only focus on diagnosing crack-related faults corresponding to frequency decreases, failing to identify mass-increasing faults such as icing and de-icing corresponding to frequency increases. Summary of the Invention

[0004] This application provides a method, device, and wind turbine for diagnosing wind turbine blade faults, which solves the technical problems in the prior art, such as low blade natural frequency tracking accuracy, single fault identification type, and high false alarm rate, caused by the failure to optimize STFT parameters for low-frequency characteristics of blades, poor noise resistance of using a single peak search method, failure to eliminate the modulation effect of rotational speed on frequency, and the inability to identify frequency rise faults while only being able to identify frequency fall faults.

[0005] The technical solution of this invention is as follows:

[0006] Firstly, this application provides a method for diagnosing faults in wind turbine blades, including:

[0007] The vibration signals of the blade during multiple time periods and the rotational speed signals synchronized with the vibration signals are acquired.

[0008] The vibration signals from each time period are spliced ​​together to obtain the full-time vibration signal;

[0009] Time-frequency analysis was performed on the vibration signal throughout the entire period to obtain the time spectrum;

[0010] Within the preset frequency band of the first natural frequency of the blade, the frequency value at each time point in the time spectrum is extracted to obtain the original ridge sequence of the first natural frequency.

[0011] The original ridge sequence is filtered to obtain a stable ridge sequence with the first natural frequency.

[0012] Select the period of continuous time in which the speed fluctuation does not exceed a first preset value from the speed signal, and use it as the constant speed interval;

[0013] Extract the segmented frequency sequence corresponding to the constant rotation speed range from the stable ridge sequence of the first natural frequency, and use it as the frequency sequence to be evaluated for the first natural frequency.

[0014] Using the average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated as a dynamic benchmark, calculate the percentage change in frequency of the current segment relative to the dynamic benchmark.

[0015] Based on the direction and magnitude of the percentage change in frequency, the fault diagnosis result of the blade is output.

[0016] Preferably, the step of performing time-frequency analysis on the full-time vibration signal to obtain the time spectrum includes:

[0017] Using a Hanning window as the window function, a short-time Fourier transform is performed on the full-time vibration signal; wherein the window length of the Hanning window includes multiple first-order natural vibration periods of the blade, the overlap rate of the short-time Fourier transform is not less than half, and the number of transform points of the short-time Fourier transform is not less than the window length.

[0018] Preferably, the step of extracting the frequency value at each time point in the time spectrum within the preset frequency band of the first-order natural frequency of the blade to obtain the original ridge sequence of the first-order natural frequency includes:

[0019] Logarithmic compression is performed on the time spectrum to obtain the logarithmic time spectrum;

[0020] For each time point in the logarithmic time spectrum, within the frequency range defined by the preset frequency band, the linear power spectral density value corresponding to each frequency point is used as the weight to perform a weighted average on all frequency points within the frequency range, and the frequency value obtained by the weighted average is used as the instantaneous estimate of the first-order natural frequency corresponding to the time point.

[0021] The instantaneous estimates of the first-order natural frequencies at all time points are arranged in chronological order to form the original ridge sequence.

[0022] Preferably, the step of filtering the original ridge sequence to obtain a stable ridge sequence with the first-order natural frequency includes:

[0023] The original ridge sequence is subjected to moving average filtering to obtain a smoothed ridge sequence;

[0024] Calculate the mean and standard deviation of the smoothed ridge sequence;

[0025] Points in the smooth ridge sequence whose difference from the mean exceeds three times the standard deviation are identified as outliers, and the mean is used to replace the outliers to obtain the stable ridge sequence.

[0026] Preferably, the method further includes:

[0027] The second natural frequency of the blade is monitored synchronously to obtain the frequency sequence of the second natural frequency to be evaluated and its frequency change percentage relative to the dynamic reference.

[0028] When the percentage changes in the first-order natural frequency and the second-order natural frequency change continuously in the same direction and the amplitudes both exceed the corresponding preset thresholds, the output will show an enhanced diagnostic confidence result.

[0029] Preferably, the rotational speed signal is traversed in a sliding window manner, and the longest continuous time period within the window where the difference between the maximum and minimum rotational speeds does not exceed the first preset value is extracted as the constant rotational speed interval.

[0030] Preferably, the step of outputting the fault diagnosis result of the blade based on the direction and magnitude of the change in the percentage of frequency change includes:

[0031] When the percentage change in frequency changes continuously in the same direction and the number of consecutive changes reaches a preset number, and the magnitude of each change exceeds a preset threshold for the corresponding direction, a fault type diagnosis result corresponding to the direction of change is output.

[0032] No diagnostic result will be output if the number of consecutive changes does not reach the preset number.

[0033] Preferably, the fault types corresponding to the direction of change include: when the direction of change is downward, it corresponds to a stiffness reduction fault; when the direction of change is upward, it corresponds to a mass increase fault.

[0034] The beneficial effects of this invention are as follows:

[0035] By constructing a complete technical chain consisting of low-frequency optimized short-time Fourier transform time-frequency analysis, energy-weighted centroid method for high-precision ridge extraction, constant speed range screening to exclude speed modulation, sliding window dynamic reference tracking for normal drift, continuous same-direction trend discrimination, and bidirectional fault identification, the system addresses the problems of insufficient frequency resolution, poor noise resistance, speed interference, fixed reference cumulative error, and single fault type identification in existing technologies. Each step is progressive and mutually supportive. Constant speed screening eliminates the confusion caused by centrifugal stiffening effects at the source; the energy-weighted centroid method achieves sub-frequency lattice accuracy, thus providing accuracy assurance for capturing minute changes in the natural frequency; and the continuous same-direction discrimination mechanism effectively avoids false alarms caused by single transient fluctuations. Ultimately, without the need for speed sensor calibration, the system achieves highly robust and high-precision blade natural frequency tracking and bidirectional diagnosis of stiffness reduction and mass increase faults. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the method in an embodiment of this application;

[0037] Figure 2 The image shows the STFT time spectrum and OSCF ridge extraction results.

[0038] Figure 3 A plot of the intrinsic frequency of moving average and 3σ outlier removal;

[0039] Figure 4 This is a graph showing the inherent frequency curves over the entire time period.

[0040] Figure 5 This is a graph showing the natural frequency curve after constant rotation speed screening. Detailed Implementation

[0041] Reference Figure 1 This application provides a method for diagnosing faults in wind turbine blades, including:

[0042] S101, acquire the vibration signals of the blade in multiple time periods and the rotational speed signal synchronized with the vibration signals;

[0043] S102, splicing the vibration signals from different time periods to obtain the full-time vibration signal;

[0044] S103, perform time-frequency analysis on the vibration signal throughout the entire time period to obtain the time spectrum;

[0045] S104. Within the preset frequency band of the first natural frequency of the blade, extract the frequency value of each time point in the time spectrum to obtain the original ridge sequence of the first natural frequency.

[0046] S105, the original ridge sequence is filtered to obtain a stable ridge sequence with first-order natural frequencies;

[0047] S106, Select the period of continuous time in which the speed fluctuation does not exceed the first preset value from the speed signal, and use it as the constant speed range;

[0048] S107, extract the segmented frequency sequence corresponding to the constant speed range from the stable ridge sequence of the first natural frequency, and use it as the frequency sequence to be evaluated for the first natural frequency.

[0049] S108, using the average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated as a dynamic reference, calculate the percentage change in frequency of the current segment relative to the dynamic reference;

[0050] S109 outputs the blade fault diagnosis results based on the direction and magnitude of the percentage change in frequency.

[0051] In step S101, vibration acceleration signals in the blade flapping direction and / or oscillation direction are continuously acquired at a preset sampling frequency (e.g., 2560Hz) to obtain multiple data files (each file lasting several minutes). At the same time, generator speed signals are acquired synchronously, and the timestamps of the speed signals and vibration signals are strictly aligned.

[0052] In step S102, the vibration signals of each data file are preprocessed by removing the mean and using quadratic polynomial detrending to eliminate DC bias and trend term interference. The preprocessed vibration signals from each time period are then concatenated sequentially. At the seams between adjacent data segments, a Hanning window weighted cross-gradient method is used. The L points (L is one-tenth of the window length) at the end of the previous segment and the beginning of the next segment are multiplied by a cosine gradient weight and then superimposed to eliminate abrupt shocks at the splicing boundary, resulting in a continuous full-time vibration signal. The processed full-time signal is denoted as X(t).

[0053] Vibration signals are typical non-stationary signals, with their frequency components varying over time. The Short-Time Fourier Transform (SFT) approximates a non-stationary signal as a series of stationary segments by introducing a sliding window function, thus obtaining the signal's time-frequency distribution. This application specifically optimizes the parameters of the SFT to address the low-frequency characteristics of the blade's natural frequency (the first-order natural frequency is typically below 2Hz), balancing the trade-off between frequency and time resolution and suppressing interference from rotating harmonics (1P / 3P) in the natural frequency band.

[0054] The choice of window function directly affects the degree of spectral leakage and frequency resolution. The Hanning window has a moderate main lobe width and large side lobe attenuation, which can effectively suppress the interference of strong energy components such as impeller rotation frequency and its harmonics on the frequency band of the natural frequency. Therefore, this application adopts the Hanning window as the window function for the short-time Fourier transform.

[0055] Window length is a key parameter that determines frequency resolution. A longer window length results in higher frequency resolution but lower temporal resolution; conversely, a shorter window length results in higher frequency resolution but lower temporal resolution.

[0056] The vibration signal of the blades exhibits typical low-frequency characteristics. Taking a large wind turbine blade as an example, its first-order natural frequency typically ranges from 0.3 to 0.8 Hz, and its second-order natural frequency typically ranges from 1.0 to 1.8 Hz. If the window length is too short, it is impossible to distinguish the natural frequency from adjacent rotating harmonics (1P / 3P). If the window length is too long, it is impossible to capture the trend of natural frequency variation over time.

[0057] This application sets the window length win_len = 65536 points. At a sampling frequency fs = 2560Hz, this window length corresponds to a physical duration of 25.6 seconds. This duration includes approximately 15-20 first-order natural vibration periods, while ensuring a frequency resolution Δf = fs / nfft better than 0.04Hz, sufficient to clearly separate the natural frequency from adjacent harmonic components.

[0058] The specific values ​​mentioned above represent preferred implementations based on typical blade parameters and sampling frequencies. Those skilled in the art should understand that for wind turbine blades of different specifications (such as different blade lengths and different natural frequency ranges) or different sampling frequencies, the window length can be adaptively adjusted according to the first-order natural vibration period of the actual blade, as long as the window length includes multiple first-order natural vibration periods of the blade.

[0059] The overlap rate affects the displacement step size of adjacent windows on the time axis. This application sets the number of overlap points (noverlap) to 75% of the window length, meaning that the displacement step size of adjacent windows is only one-quarter of the window length. A high overlap rate ensures a high degree of overlap between adjacent analysis windows on the time axis, making the extracted frequency trajectory continuous and smooth in the time direction, avoiding trajectory jumps caused by excessively large step sizes.

[0060] The number of transform points, nfft, determines the density of the frequency sampling grid. This application sets the number of transform points to twice the window length, i.e., nfft = 131072 points. Frequency domain interpolation is achieved through zero-padding, refining the frequency grid by a factor of two, thereby improving the sub-frequency grid accuracy of ridge extraction.

[0061] Based on the above parameter configuration, a short-time Fourier transform is performed on the full-time vibration signal X(t) to obtain the complex spectrum matrix S(t,f).

[0062] Calculate the power spectral density P(t,f) = |S(t,f)| of the complex spectral matrix S(t,f). 2 This serves as the time-spectrum data for subsequent steps.

[0063] Convert the power spectrum to logarithmic decibel values: P db(t,f)=10·log10(P(t,f)+ε), where ε is a preset minimum value (e.g., ε=10). ﹣10 ), used to avoid singularity when the logarithm is zero.

[0064] The advantage of logarithmic compression lies in the fact that the dynamic range of blade vibration signals typically exceeds 80 dB. Weak frequency sidebands caused by early cracks or minute frequency rises in the early stages of icing may be masked by strong energy components before logarithmic compression. By compressing the wide dynamic range signal into a relatively small range, weak components become visible in the time spectrum. Simultaneously, at the logarithmic scale, additive noise is approximately converted into additive noise, which is beneficial for subsequent noise reduction processing in ridge extraction.

[0065] In step S103, the time spectrum reflects the distribution of signal energy in the time-frequency plane. For the first-order natural frequency of the blade, its energy is concentrated within a preset frequency band and may drift slowly over time. To obtain a continuous trajectory of the natural frequency changing over time, it is necessary to extract the frequency value at each time point from the time spectrum to form the original ridge sequence.

[0066] Based on the blade's design parameters or historical data under healthy conditions, the theoretical value of the first-order natural frequency f1 (e.g., 0.534Hz) is preset. The search bandwidth band1 (e.g., 0.2Hz) is set to cover the possible drift range of the natural frequency during long-term operation, thereby determining the preset frequency band of the first-order natural frequency as [f1-band1, f1+band1].

[0067] For each time point tk in the time spectrum, within the preset frequency band [f1-band1, f1+band1], the weighted average frequency is calculated using the power value of each frequency point as the weight, and is used as the instantaneous estimate of the first-order natural frequency at that time point:

[0068]

[0069] Wherein, P(t) k ,f i ) represents time point t in the time spectrum. k Frequency point f i Power value at f; i is the frequency value of the i-th frequency point within the preset frequency band; N is the total number of frequency points within the preset frequency band.

[0070] The calculation of the weighted average frequency utilizes the energy distribution information of all frequency points within the preset frequency band, rather than relying solely on a single peak point. Therefore, it can effectively suppress the interference of noise and spectral leakage on frequency estimation, and the estimation result can be located at any position in the frequency sampling grid, achieving sub-frequency grid accuracy.

[0071] All time points tk The instantaneous estimates are arranged in chronological order to obtain the original ridge sequence f of the first-order natural frequencies. raw (t). For example... Figure 3 As shown, the original ridge line sequence reflects the original trajectory of the first-order natural frequency of the blade changing over time.

[0072] Original ridge sequence f raw The original ridge sequence (t) contains the true trend of frequency change over time, but also includes random fluctuations and outliers caused by environmental noise, airflow disturbances, electromagnetic interference, and other factors. To obtain a smooth, continuous, and physically meaningful intrinsic frequency trajectory, the original ridge sequence needs to be filtered.

[0073] For the original ridge sequence f raw (t) is filtered using a moving average to obtain a smoothed ridge sequence f. smooth (t).

[0074] The formula for calculating the moving average filter is:

[0075]

[0076] Among them, t k This represents the k-th time point, where W is the moving average window length (e.g., W=10), represents the number of adjacent points participating in the average, and j is the distance from the center point t within the window. k The offset index. The window length can be adaptively adjusted according to the noise level and sampling rate of the actual signal.

[0077] Moving average filtering can effectively suppress high-frequency random fluctuations and preserve the low-frequency trend components of frequency changes. It is a conventional technique for processing random noise in ridge sequences.

[0078] Calculate the smooth ridge sequence f smooth The mean μ and standard deviation σ of (t):

[0079]

[0080] Where M is the total number of sampling points in the smooth ridge sequence, and t k This represents the k-th time point.

[0081] It should be noted that, in this application, the mean μ represents the statistically significant central position of the smooth ridge sequence; the standard deviation σ is used to measure the dispersion of each data point relative to this central position. Together, they constitute the statistical basis for identifying outliers. The mean and standard deviation are merely illustrative examples, and those skilled in the art can also use the median instead of the mean, and the absolute median deviation instead of the standard deviation, to statistically determine outliers.

[0082] The smooth ridge sequence fsmooth Points in (t) that satisfy the following conditions are considered outliers:

[0083]

[0084] Among them, t k Let μ be the mean of the smoothed ridge sequence at time point k, and σ be the standard deviation of the smoothed ridge sequence. That is, points whose difference from the mean exceeds three times the standard deviation are identified as outliers.

[0085] For locations identified as outliers, replace the original value of that point with the mean μ:

[0086]

[0087] Among them, f final (t k ) represents the frequency value of the final output stable ridge sequence at the kkth time point.

[0088] It should be noted that replacing outliers with the mean is only one preferred solution in this embodiment. When replacing outliers, the mean can be used, or linear interpolation of adjacent non-outliers in the smoothed ridge sequence can be used, or the median can be used. All of these methods can achieve the technical effect of eliminating the influence of outliers on subsequent trend analysis. Those skilled in the art can choose an appropriate replacement method based on the characteristics of the actual signal.

[0089] For locations not identified as outliers, their original values ​​are retained.

[0090] like Figure 4 After moving average filtering and 3σ adaptive outlier removal, a stable ridge sequence f with first-order natural frequencies is obtained. final (t).

[0091] This stable ridge sequence eliminates the interference of high-frequency random fluctuations and outliers, and preserves the true trend of the first-order natural frequency of the blade over time. It can serve as a reliable data source for subsequent constant speed range screening and frequency change analysis.

[0092] It should be noted that the above-described processing order of moving average filtering first and 3σ outlier removal second is only a preferred embodiment of this implementation. In another embodiment, 3σ outlier removal can be performed first, followed by moving average filtering, which can also achieve the technical effects of smoothing and removing outliers. The choice of the two orders depends on the distribution characteristics of noise and outliers in the actual signal, and those skilled in the art can adjust them flexibly according to the actual situation.

[0093] The natural frequency of a blade increases with increasing rotational speed. This phenomenon is known as the centrifugal stiffening effect in rotating blade dynamics—the higher the rotational speed, the greater the apparent stiffness of the blade due to centrifugal force, and the higher the natural frequency. If frequency data extracted from different rotational speeds are mixed, the natural frequency trajectory will fluctuate with the rotational speed, making it impossible to distinguish whether the frequency change is caused by changes in rotational speed or by changes in the blade's structural state. Therefore, it is necessary to filter out a continuous period of approximately constant rotational speed from the rotational speed signal. During this period, the contribution of the centrifugal stiffening effect to the natural frequency is constant, and the frequency change uniquely reflects the change in the blade's structural stiffness or mass.

[0094] The rotational speed signal s(t) is traversed using a sliding window. The window length is set to T. win (e.g., T) win =60 seconds), sliding in preset steps (e.g., 1 second), calculating the speed fluctuation amplitude within each window one by one. The formula for calculating the speed fluctuation amplitude within a window is:

[0095]

[0096] Where, max{s(t)} t∈Twin min{s(t)} represents the maximum rotational speed within the window. t∈Twin This represents the minimum rotational speed within the window.

[0097] By continuously sliding the window, all possible start time positions can be traversed to ensure that no period of time that meets the constant speed condition is missed. The setting of the window length determines the time resolution of the constant speed interval. If the window is too short, a brief period of stable speed may be misjudged as a constant speed interval. If the window is too long, a period of brief stable speed that is sufficient for analysis may be missed.

[0098] A window whose speed fluctuation amplitude Δs(Twin) does not exceed the first preset value stol is determined to be a constant speed window.

[0099] The first preset value stol is a speed tolerance parameter, and its range can be set according to the operating characteristics and control accuracy of the unit. For example, in a preferred embodiment, stol=l=0.5rpm, that is, when the difference between the maximum speed and the minimum speed within the window does not exceed 0.5rpm, the window is determined to be in a constant speed state.

[0100] For a unit operating at full capacity and in a stable condition, the speed is typically maintained near the design value with minimal fluctuations. Therefore, the constant speed range corresponds to the unit operating at full capacity and in a stable condition. Under this condition, the frequency offset contributed by the centrifugal stiffening effect is a constant value, and the change in the natural frequency uniquely reflects the change in the blade's structural stiffness or mass.

[0101] Combination Figure 4 and Figure 5All constant speed windows are merged in chronological order, and the longest continuous period is extracted as the constant speed interval.

[0102] In actual operating data, there may be multiple dispersed constant speed periods. Selecting the longest duration period provides the longest stable operating condition data for subsequent frequency trend analysis, which helps improve the reliability of statistical analysis. If the shortest constant speed period duration is insufficient to meet the analysis requirements, a minimum duration threshold can be set for filtering (e.g., including at least 3 segments to be evaluated) to ensure that the filtering results are statistically significant.

[0103] The start and end times of the constant speed range are output as time masks, which are then used to extract the corresponding segmented frequency sequences from the stable ridge sequence.

[0104] It should be noted that the above-described implementation of filtering constant speed intervals using a sliding window method is only one preferred embodiment. In another embodiment, the differential method can also be used to identify the stable segment of the speed signal—calculate the first-order difference of the speed signal, and determine the period when the absolute value of the difference is continuously lower than a preset threshold as the constant speed period; in yet another embodiment, the piecewise linear fitting method can also be used to identify the constant speed interval—perform piecewise linear fitting on the speed signal, and merge continuous segments with slope absolute values ​​close to zero into a constant speed interval.

[0105] Through the aforementioned steps, a stable ridge sequence f with first-order natural frequencies has been obtained. final (t) and the time mask [T] for the constant speed interval start ,T end The stable ridge sequence covers the entire data acquisition period and includes frequency information at different rotational speeds; the constant rotational speed interval identifies periods where the rotational speed is approximately constant. This step combines the two, extracting frequency segments corresponding to the constant rotational speed intervals from the stable ridge sequence as the evaluation sequences for subsequent frequency variation analysis.

[0106] Based on the start time T of the constant speed range start and end time T end In stable ridge sequence f final Locate the corresponding time index on the time axis of (t).

[0107] Since the vibration signal and the rotational speed signal use the same time base and the timestamps are strictly aligned, the time coordinates of the constant rotational speed range can be directly mapped to the time axis of the stable ridge sequence.

[0108] All data points located within the constant rotational speed range are extracted from the stable ridge line sequence to form a segmented frequency sequence f. seg (t):

[0109]

[0110] In a preferred embodiment, the constant speed range may span multiple data files or contain multiple data segments. To facilitate subsequent sliding window dynamic benchmark comparisons, the segmented frequency sequence is further divided into several consecutive evaluation segments with fixed time lengths, each segment serving as an independent evaluation unit. The segment length can be set according to the data volume and analysis requirements, for example, each segment lasting several minutes.

[0111] Segmented frequency sequence f seg (t) is output as the first-order natural frequency of the frequency sequence to be evaluated, which is used for subsequent dynamic benchmark comparison and frequency change trend analysis.

[0112] The rotational speed corresponding to the frequency sequence to be evaluated is approximately constant. Therefore, the contribution of the centrifugal stiffening effect to the natural frequency is constant. The frequency changes in the sequence exclude the interference of rotational speed factors and only reflect the changes in the blade structural stiffness or mass. It is a reliable data source for damage diagnosis and fault identification.

[0113] In actual unit operation, the constant speed range may exist as a single continuous period or as multiple dispersed periods. When there are multiple constant speed periods, the segmented frequency sequences corresponding to each period can be extracted and arranged in chronological order as the frequency sequence to be evaluated.

[0114] The natural frequency of blades may drift slowly over long-term operation due to factors such as changes in ambient temperature, minor blade wear, and material aging. If fixed historical data from the initial commissioning of the unit is used as the health baseline, accumulated environmental and operating condition changes will cause the deviation between the baseline value and the actual health state to gradually increase, thereby reducing the accuracy of fault diagnosis. To solve the above problem, this application adopts a sliding window dynamic baseline method, which always uses the most recent historical data before the current segment as the baseline, allowing the baseline value to follow the normal aging drift of the blades, thus avoiding the cumulative error caused by a fixed baseline.

[0115] Frequency sequence f to be evaluated seg (t) is divided into several consecutive evaluation segments on the time axis, each evaluation segment is denoted as S. i The subscript i represents the sequence number of the evaluation segment on the timeline, i = 1, 2, 3, ..., with larger values ​​indicating later times. The segment currently being evaluated is denoted as the current segment S. i .

[0116] The average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated is used as the dynamic benchmark. Specifically, the frequency average of the segment preceding the current segment S in the frequency sequence to be evaluated is selected. i The preceding segment S in time i－1 The average frequency, or the first two segments of S i－1and S i－2 The average frequency is used as the dynamic reference f for the current segment. base (i).

[0117] When the preceding segment is selected as the dynamic baseline, the formula for calculating the dynamic baseline is:

[0118]

[0119] Where Q represents the preceding segment S i－1 The total number of sampling points within, f seg (t q ) represents the frequency value of the q-th sampling point within the previous segment.

[0120] When the first two segments are selected as the dynamic reference, the formula for calculating the dynamic reference is:

[0121]

[0122] Among them, S i－1 The preceding paragraph to the current paragraph, S i－2 The first two segments of the current segment are represented by Q, where Q is the total number of sampling points in each segment.

[0123] When i=1, meaning there is no data in the previous segment or the first two segments, the current segment is the first segment. It is impossible to calculate the percentage change in frequency relative to historical data. Therefore, it is used as the initial baseline segment, and no percentage change result is output.

[0124] In one preferred embodiment, the average frequency of the preceding segment is selected as the dynamic benchmark. This approach can respond most quickly to recent changes in blade condition and has high sensitivity to sudden faults. In another preferred embodiment, the average frequency of the first two segments is selected as the dynamic benchmark. This approach can suppress random fluctuations that may exist in a single segment of data and has better stability in field environments with large data fluctuations. The selection of the two approaches depends on the quality of the field data and diagnostic needs, and those skilled in the art can flexibly determine the approach based on the actual situation.

[0125] Calculate the current segment S i Frequency average value and dynamic reference f base Percentage change between (i):

[0126]

[0127] in, For the current segment S i The average frequency, , t q ∈S i ;f base (i) is the dynamic baseline corresponding to the current segment.

[0128] Percentage change in frequency Δf i A positive value indicates that the natural frequency of the current segment is increasing relative to the reference, while a negative value indicates that the natural frequency of the current segment is decreasing relative to the reference.

[0129] Arrange the percentage changes in frequency for each evaluation segment in chronological order to obtain a sequence of percentage changes in frequency:

[0130]

[0131] This sequence reflects the segmented variation trend of the first-order natural frequency of the blade relative to the dynamic reference, providing a data basis for determining the direction and magnitude of subsequent changes.

[0132] In a preferred embodiment, when the dynamic benchmark uses the average frequency of the previous segment, if the data interval between the current segment and the previous segment is long (e.g., the interval exceeds a preset time threshold), the timeliness of the dynamic benchmark decreases. In this case, the dynamic benchmark can be corrected. Specifically, the correction can be achieved by using the average frequency of consecutive valid segments preceding the current segment as the dynamic benchmark, or by using linear extrapolation to adjust the frequency value of the previous segment to the estimated value corresponding to the current time point. If the timeliness requirement is still not met after correction, the current segment will not output a percentage change result, awaiting subsequent valid data.

[0133] Through the aforementioned steps, a frequency change percentage sequence ΔF={Δf1,Δf2,Δf3,…} for each evaluation segment has been obtained. This sequence reflects the segment-by-segment change trend of the blade's first-order natural frequency relative to the dynamic baseline. Based on the direction and magnitude of this sequence's change, changes in the blade's structural state can be identified, and corresponding fault diagnosis results can be output.

[0134] Wind turbine blades may experience two types of failures with different physical characteristics during operation: one type is stiffness reduction failure, such as blade cracks, fractures, and composite material delamination. This type of failure leads to a decrease in the blade's structural stiffness, and consequently a decrease in its natural frequency, which is shown as a continuous negative change in the percentage frequency change sequence. The other type is mass increase failure, such as blade icing, ice accumulation, fouling, and water accumulation at the blade root. This type of failure leads to an increase in the blade's mass, and consequently an increase in its natural frequency, which is shown as a continuous positive change in the percentage frequency change sequence.

[0135] Therefore, by monitoring the direction and magnitude of the percentage change in frequency, the two types of faults mentioned above can be distinguished, thus achieving bidirectional fault identification.

[0136] First, determine whether the percentage change in frequency sequence ΔF changes continuously in the same direction.

[0137] The specific judgment method is as follows: for adjacent evaluation segments S i－1 and Si Calculate the corresponding percentage change in frequency Δf. i－1 and Δf i Determine whether the two signs are the same. If the signs are the same, the changes are in the same direction; if the signs are different, the direction of change has changed.

[0138] The number of consecutive changes in the same direction, C, is defined as the number of consecutive evaluation segments appearing in the current direction. For example, if Δf1 is negative, Δf2 is negative, and Δf3 is negative, then the number of consecutive decreases, C, is... down =3; if Δf4 turns positive, the number of consecutive increases is reset to C. up =1.

[0139] Only when the number of consecutive unidirectional changes reaches the preset number C thr Only when the number of consecutive changes in the same direction does it qualify as a valid trend, proceeding to the next step of magnitude assessment. If the number of consecutive changes in the same direction does not reach C... thr If the condition is not met, it is determined to be a random fluctuation or transient interference, and no diagnostic result is output.

[0140] Set preset number of times C thr The technical advantage is that a single change exceeding the threshold may be caused by non-fault factors such as aerodynamic transients, sudden changes in wind speed, and measurement noise. It requires multiple consecutive changes in the same direction to trigger the diagnosis, which can effectively avoid false alarms caused by a single fluctuation.

[0141] When the number of consecutive changes in the same direction reaches the preset number C thr Then, it is further determined whether the magnitude of each change exceeds the preset threshold in the corresponding direction.

[0142] For the descent direction, set a descent warning threshold T. hwarn_down and the drop alarm threshold T halarm_down (T) halarm_dow n>T hwarn_down Furthermore, both the decline warning threshold and the decline alarm threshold are negative. When the magnitude of each consecutive decline is lower than (i.e., the absolute value is greater than) the decline alarm threshold T... halarm_down When the amplitude of each consecutive decrease is lower than the decrease warning threshold T, an alarm-level diagnostic result is output. hwarn_down But it is higher than (i.e., its absolute value is less than) the drop alarm threshold T halarm_down At that time, output the early warning level diagnostic results.

[0143] For the upward direction, set the upward warning threshold Thrwarn_up and the upward alarm threshold T. halarm_up (T) halarm_up >T hwarn_up >0). When the magnitude of each consecutive increase is higher than the alarm threshold T. halarm_upWhen the alarm level diagnostic result is output, and the amplitude of each consecutive increase is higher than the rising warning threshold T, the alarm level diagnostic result is output. hwarn_up But below the rising alarm threshold T halarm_up At that time, output the early warning level diagnostic results.

[0144] In a preferred embodiment, the preset number of times C thr =2, meaning that a diagnosis is triggered only when there are two consecutive changes in the same direction, and the magnitude of each change exceeds the corresponding threshold. The requirement of two consecutive determinations can effectively exclude single, accidental fluctuations, and at the same time, it has sufficient response speed in engineering, so as not to cause diagnostic lag due to requiring too many consecutive times.

[0145] Based on the above judgment, output the corresponding fault diagnosis result:

[0146] When the percentage change in frequency decreases continuously and each decrease is less than the alarm threshold T, halarm_down When the blade stiffness is significantly reduced, an alarm-level diagnostic result is output, corresponding to the risk of serious structural damage such as blade crack propagation and breakage.

[0147] When the percentage change in frequency decreases continuously and each decrease is less than the warning threshold T, hwarn_down But it is higher than the alarm threshold T. halarm_down At that time, it outputs a warning-level diagnostic result for a slight decrease in blade stiffness, corresponding to early structural damage such as early cracks and composite material delamination;

[0148] When the percentage change in frequency increases continuously and each increase exceeds the alarm threshold T, halarm_up When the output is significant, an alarm-level diagnostic result is given, corresponding to serious mass increase faults such as severe icing, large water accumulation at the blade root, and large-area foreign object adhesion.

[0149] When the percentage change in frequency increases continuously and each increase exceeds the warning threshold T, hwarn_up But below the rising alarm threshold T halarm_up At that time, the output will show a warning-level diagnostic result indicating a slight increase in blade mass, corresponding to initial mass increase faults such as slight icing and fouling accumulation.

[0150] If the number of consecutive unidirectional changes does not reach the preset number CthrCthr, no diagnostic result will be output, and it will be judged as speed fluctuation, aerodynamic transient or measurement noise.

[0151] In a preferred embodiment, the second natural frequency of the blade is monitored simultaneously to obtain the frequency sequence of the second natural frequency to be evaluated and its frequency change percentage relative to a dynamic reference.

[0152] When the percentage change of the first-order natural frequency and the second-order natural frequency both show a consistent trend in the same direction, that is, when the first-order and second-order frequencies both continuously decrease or both continuously increase, the output diagnostic confidence is enhanced.

[0153] The advantage of cross-validation lies in the fact that changes in blade structural stiffness or mass are systemic physical changes that simultaneously affect both the first and second natural frequencies. If only the first frequency changes while the second frequency remains stable, it may indicate a change in local boundary conditions or a sensor malfunction, rather than a global change in the blade's state. Two-stage cross-validation significantly improves diagnostic reliability and reduces false alarm rates.

[0154] By constructing a complete technical chain consisting of low-frequency optimized short-time Fourier transform time-frequency analysis, energy-weighted centroid method for high-precision ridge extraction, constant speed range screening to exclude speed modulation, sliding window dynamic reference tracking for normal drift, continuous same-direction trend discrimination, and bidirectional fault identification, the system addresses the problems of insufficient frequency resolution, poor noise resistance, speed interference, fixed reference cumulative error, and single fault type identification in existing technologies. Each step is progressive and mutually supportive. Constant speed screening eliminates the confusion caused by centrifugal stiffening effects at the source; the energy-weighted centroid method achieves sub-frequency lattice accuracy, thus providing accuracy assurance for capturing minute changes in the natural frequency; and the continuous same-direction discrimination mechanism effectively avoids false alarms caused by single transient fluctuations. Ultimately, without the need for speed sensor calibration, the system achieves highly robust and high-precision blade natural frequency tracking and bidirectional diagnosis of stiffness reduction and mass increase faults.

[0155] Secondly, this application provides a wind turbine blade fault diagnosis device, comprising:

[0156] The vibration signals of the blade during multiple time periods and the rotational speed signals synchronized with the vibration signals are acquired.

[0157] The vibration signals from each time period are spliced ​​together to obtain the full-time vibration signal;

[0158] Time-frequency analysis was performed on the vibration signal throughout the entire period to obtain the time spectrum;

[0159] Within the preset frequency band of the first natural frequency of the blade, the frequency value at each time point in the time spectrum is extracted to obtain the original ridge sequence of the first natural frequency.

[0160] The original ridge sequence is filtered to obtain a stable ridge sequence with the first natural frequency.

[0161] Select the period of continuous time in which the speed fluctuation does not exceed a first preset value from the speed signal, and use it as the constant speed interval;

[0162] Extract the segmented frequency sequence corresponding to the constant rotation speed range from the stable ridge sequence of the first natural frequency, and use it as the frequency sequence to be evaluated for the first natural frequency.

[0163] Using the average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated as a dynamic benchmark, calculate the percentage change in frequency of the current segment relative to the dynamic benchmark.

[0164] Based on the direction and magnitude of the percentage change in frequency, the fault diagnosis result of the blade is output.

[0165] Thirdly, this application also provides a wind turbine generator set, including the aforementioned wind turbine blade fault diagnosis device.

[0166] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Those skilled in the art can understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this application still fall within the scope of this application.

Claims

1. A method for diagnosing faults in wind turbine blades, characterized in that, include: The vibration signals of the blade during multiple time periods and the rotational speed signals synchronized with the vibration signals are acquired. The vibration signals from each time period are spliced ​​together to obtain the full-time vibration signal; Time-frequency analysis was performed on the vibration signal throughout the entire period to obtain the time spectrum; Within the preset frequency band of the first natural frequency of the blade, the frequency value at each time point in the time spectrum is extracted to obtain the original ridge sequence of the first natural frequency. The original ridge sequence is filtered to obtain a stable ridge sequence with the first natural frequency. Select the period of continuous time in which the speed fluctuation does not exceed a first preset value from the speed signal, and use it as the constant speed interval; Extract the segmented frequency sequence corresponding to the constant rotation speed range from the stable ridge sequence of the first natural frequency, and use it as the frequency sequence to be evaluated for the first natural frequency. Using the average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated as a dynamic benchmark, calculate the percentage change in frequency of the current segment relative to the dynamic benchmark. Based on the direction and magnitude of the percentage change in frequency, the fault diagnosis result of the blade is output.

2. The method according to claim 1, characterized in that, The steps for performing time-frequency analysis on the full-time vibration signal to obtain the time spectrum include: Using a Hanning window as the window function, a short-time Fourier transform is performed on the full-time vibration signal; wherein the window length of the Hanning window includes multiple first-order natural vibration periods of the blade, the overlap rate of the short-time Fourier transform is not less than half, and the number of transform points of the short-time Fourier transform is not less than the window length.

3. The method according to claim 1, characterized in that, The step of extracting the frequency value at each time point in the time spectrum within the preset frequency band of the first-order natural frequency of the blade to obtain the original ridge sequence of the first-order natural frequency includes: Logarithmic compression is performed on the time spectrum to obtain the logarithmic time spectrum; For each time point in the logarithmic time spectrum, within the frequency range defined by the preset frequency band, the linear power spectral density value corresponding to each frequency point is used as the weight to perform a weighted average on all frequency points within the frequency range, and the frequency value obtained by the weighted average is used as the instantaneous estimate of the first-order natural frequency corresponding to the time point. The instantaneous estimates of the first-order natural frequencies at all time points are arranged in chronological order to form the original ridge sequence.

4. The method according to claim 1, characterized in that, The steps of filtering the original ridge sequence to obtain a stable ridge sequence with the first-order natural frequency include: The original ridge sequence is subjected to moving average filtering to obtain a smoothed ridge sequence; Calculate the mean and standard deviation of the smoothed ridge sequence; Points in the smooth ridge sequence whose difference from the mean exceeds three times the standard deviation are identified as outliers, and the mean is used to replace the outliers to obtain the stable ridge sequence.

5. The method according to claim 1, characterized in that, The method further includes: The second natural frequency of the blade is monitored synchronously to obtain the frequency sequence of the second natural frequency to be evaluated and its frequency change percentage relative to the dynamic reference. When the percentage changes in the first-order natural frequency and the second-order natural frequency change continuously in the same direction and the amplitudes both exceed the corresponding preset thresholds, the output will show an enhanced diagnostic confidence result.

6. The method according to claim 1, characterized in that, The rotational speed signal is traversed using a sliding window method, and the longest continuous time period within the window where the difference between the maximum and minimum rotational speeds does not exceed the first preset value is extracted as the constant rotational speed interval.

7. The method according to claim 1, characterized in that, The steps for outputting the fault diagnosis result of the blade based on the direction and magnitude of the change in the percentage of frequency change include: When the percentage change in frequency changes continuously in the same direction and the number of consecutive changes reaches a preset number, and the magnitude of each change exceeds a preset threshold for the corresponding direction, a fault type diagnosis result corresponding to the direction of change is output. No diagnostic result will be output if the number of consecutive changes does not reach the preset number.

8. The method according to claim 7, characterized in that, The fault types corresponding to the direction of change include: when the direction of change is downward, it corresponds to a stiffness reduction fault; when the direction of change is upward, it corresponds to a mass increase fault.

9. A wind turbine blade fault diagnosis device, characterized in that, include: The vibration signals of the blade during multiple time periods and the rotational speed signals synchronized with the vibration signals are acquired. The vibration signals from each time period are spliced ​​together to obtain the full-time vibration signal; Time-frequency analysis was performed on the vibration signal throughout the entire period to obtain the time spectrum; Within the preset frequency band of the first natural frequency of the blade, the frequency value at each time point in the time spectrum is extracted to obtain the original ridge sequence of the first natural frequency. The original ridge sequence is filtered to obtain a stable ridge sequence with the first natural frequency. Select the period of continuous time in which the speed fluctuation does not exceed a first preset value from the speed signal, and use it as the constant speed interval; Extract the segmented frequency sequence corresponding to the constant rotation speed range from the stable ridge sequence of the first natural frequency, and use it as the frequency sequence to be evaluated for the first natural frequency. Using the average frequency of at least one segment preceding the current segment in the frequency sequence to be evaluated as a dynamic benchmark, calculate the percentage change in frequency of the current segment relative to the dynamic benchmark. Based on the direction and magnitude of the percentage change in frequency, the fault diagnosis result of the blade is output.

10. A wind turbine generator set, characterized in that, It includes the wind turbine blade fault diagnosis device as described in claim 9.

Images