Intelligent Inspection Method for Wind Turbine Blades Based on Blade Acoustic Resonance Segmentation Comparison and Defect Radial Location Mapping
By collecting and segmenting blade acoustic signatures, and combining periodic consistency verification and frequency band position correction, the accuracy problems of acoustic signature segmentation and defect location in wind turbine blade inspection have been solved, realizing high-frequency, low-cost automated inspection.
Patent Information
- Application Number
- CN202611288309.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing wind turbine blade inspection methods are unable to accurately segment blade acoustic signature segments under severe aliasing and environmental noise interference, resulting in insufficient defect location capabilities. Furthermore, traditional methods are easily affected by changes in wind speed and rotational speed, leading to frequent false alarms or missed alarms.
By collecting the blade acoustic signatures for two consecutive rotation cycles, segmented operations are performed independently to check cycle consistency and correct frequency band position. A three-blade asynchronous mutual calibration strategy is used to eliminate common-mode interference and achieve accurate mapping of defect types.
Robust segmentation and position correction of blade acoustic patterns under complex operating conditions have been achieved, improving the accuracy and robustness of defect judgment, reducing maintenance costs, and meeting the needs of unmanned inspection.
Smart Images

Figure CN123109964A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine condition monitoring technology, and in particular to an intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping. Background Technology
[0002] As the core component for capturing wind energy, wind turbine blades are subjected to alternating loads, dust erosion, lightning strikes, and fatigue stress over long periods. Their surface and internal structures are prone to damage such as cracks, pinholes, leading-edge corrosion, and trailing-edge cracking. Failure to detect and repair these damages in a timely manner can lead to blade breakage or even the collapse of the entire turbine, causing significant safety accidents and economic losses. Therefore, research into efficient, accurate, and low-cost blade inspection technologies has significant engineering value.
[0003] Currently, mainstream wind turbine blade inspection methods fall into three main categories: First, ground-based inspection using manual visual inspection or telescopes, which is inefficient, subjective, and struggles to detect minute or internal defects. Second, aerial inspection using drones equipped with visible light or infrared cameras, which, while enabling close-range imaging, is limited by flight control, endurance, weather conditions, and image post-processing algorithms, making high-frequency, fully automated online monitoring difficult. Third, contact-based monitoring using vibration accelerometers or fiber optic grating sensors, requiring pre-installation of sensors on the blades, resulting in high costs and sensor vulnerability to lightning strikes. Furthermore, the blade's rotational characteristics pose significant engineering challenges for wireless signal transmission and power supply. In recent years, acoustic signature analysis technology has been gradually introduced into the field of rotating machinery fault diagnosis due to its advantages such as non-contact operation, low sensor cost, and continuous monitoring capabilities. However, existing acoustic signature-based wind turbine blade monitoring solutions have significant shortcomings. For example, some solutions directly perform full-band energy analysis on the original acoustic signature, failing to distinguish the acoustic characteristics generated at different blade locations, resulting in a lack of defect localization capabilities. Other solutions attempt to separate aliased acoustic signatures to individual blades through blind source separation or direction-of-arrival estimation, but these methods are computationally complex and susceptible to strong interference from wind noise, tower reflections, and other factors, exhibiting poor robustness in practical engineering. More importantly, the continuous changes in wind turbine operating conditions (wind speed, rotational speed, and pitch angle) cause drastic fluctuations in the amplitude and spectral structure of blade acoustic signatures. Traditional fixed-threshold acoustic signature comparison methods are prone to false alarms or missed alarms, making it difficult to meet actual operation and maintenance needs. Summary of the Invention
[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the specification abstract and the title of the invention, to avoid obscuring the purpose of this section, the specification abstract, and the title of the invention. Such simplifications or omissions shall not be used to limit the scope of the invention.
[0005] In view of the aforementioned existing problems, the present invention is proposed.
[0006] Therefore, the technical problem solved by this invention is: how to automatically and reliably segment the effective acoustic print segment of each blade from the acoustic print of severely aliased blades without relying on additional sensors, and eliminate the common-mode interference of environmental noise and operating condition changes on the acoustic print characteristics, thereby achieving accurate mapping of defect type to blade radial position.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an intelligent inspection method for wind turbine blades based on blade acoustic pattern segmentation comparison and defect radial position mapping, characterized by: collecting blade acoustic patterns from two consecutive complete rotation cycles, performing segmentation operations independently to obtain a first cycle segmentation point set and a second cycle segmentation point set; performing cycle consistency verification on the first and second cycle segmentation point sets, if the time interval between the two sets of segmentation points matches one rotation cycle, then using the second cycle segmentation point set as the reference segmentation point; dividing the acoustic pattern segment corresponding to the reference segmentation point into multiple frequency band sets, each frequency band set being associated with different radial positions of the blade, and performing position correction on the reference segmentation point based on the radial positions to obtain corrected segmentation points; extracting acoustic pattern segments of each stage of the current blade based on the corrected segmentation points, and comparing them with the historical corresponding stage acoustic pattern segments of two other blades of the same wind turbine, and determining the defect type based on the distribution pattern of the acoustic pattern differences among the three blades.
[0008] As a preferred embodiment of the present invention, the step of independently performing segmentation operations includes: using the low-frequency impact sound generated each time the blade passes directly in front of the tower as the initial trigger signal; identifying two adjacent trigger signals belonging to the same blade by detecting two consecutive impact sounds and using the known number of wind turbine blades; determining the time interval between the two triggers as the measured rotation cycle; starting from the first trigger signal, collecting acoustic fingerprint data; after detecting the second trigger signal of the same blade and obtaining the measured rotation cycle, continuing to collect acoustic fingerprint data for one measured rotation cycle duration to obtain an original acoustic fingerprint sequence covering two complete measured rotation cycles; segmenting the acoustic fingerprint in the first rotation cycle of the original acoustic fingerprint sequence by detecting the local minimum points of the short-time energy envelope to obtain a first cycle segmentation point set; the number of segmentation points in the first cycle segmentation point set is equal to the number of wind turbine blades; independently repeating the segmentation method for the acoustic fingerprint in the second rotation cycle of the original acoustic fingerprint sequence to obtain a second cycle segmentation point set, the number of segmentation points in the second cycle segmentation point set being the same as the first cycle segmentation point set.
[0009] In a preferred embodiment of the present invention, the period consistency verification includes: predicting a time window for each segment point in the first period segment point set using the measured rotation period, wherein the center of the time window is the time coordinate of the first period segment point plus one measured rotation period, and the window width is a preset ratio multiplied by the measured rotation period; searching within the corresponding time window for the segment point in the second period segment point set with the closest time as a pairing point, thereby pairing the segment points in the first period and the second period one by one according to the blade passage event; calculating the time difference between the second period segment point and the first period segment point in each pairing; if the deviation of each time difference relative to the measured rotation period does not exceed a preset deviation tolerance, then the two sets of segment point periods are determined to be consistent; otherwise, the current data is discarded and re-acquired; and the second period segment point set is used as the reference segment point set.
[0010] In a preferred embodiment of the present invention, the step of correcting the position of the reference segment points to obtain corrected segment points includes: taking each segment point in the set of reference segment points as the starting point, extracting a first set duration of soundprint segments from the original soundprint sequence to obtain a number of soundprint segments equal to the number of blades, wherein the first set duration is less than the measured rotation period; passing each soundprint segment through a bandpass filter with multiple preset frequency bands, each frequency band being pre-associated with a radial position of the blade, corresponding to low frequency, mid frequency, and high frequency from the blade root to the blade tip; extracting the time point corresponding to the maximum envelope value for each frequency band-filtered segment, and calculating the time offset between the time point and the corresponding reference segment point; for each soundprint segment, calculating the statistical average of the time offsets corresponding to all frequency bands, and adding the statistical average to the time coordinate of the corresponding reference segment point to obtain the corrected segment point for each blade; the corrected segment points of all blades constitute the set of corrected segment points.
[0011] In a preferred embodiment of the present invention, the step of extracting acoustic signature segments of each stage of the current blade and comparing them with historical acoustic signature segments of the corresponding stages of the other two blades of the same wind turbine includes: selecting one of the three blades of the wind turbine as the current blade to be detected, and expanding the segment to both sides of the segment point corresponding to the blade to be detected in the corrected segment point set by a second set time period, with the segment point as the center, to obtain the current stable period segment; the second set time period is less than half a measured rotation cycle; and based on the segment points corresponding to the other two blades in the corrected segment point set, comparing them with the historical acoustic signature segments of the other two blades in the pre-constructed historical segment set. The historical stable period segments of these two blades are retrieved from the historical acoustic signature database. The historical stable period segments are those extracted in the same way during the most recent successful inspection before this acquisition, and the difference between the cumulative number of rotation cycles corresponding to the historical stable period segments and the cumulative number of rotation cycles acquired in this acquisition does not exceed a preset number of revolutions threshold. The historical acoustic signature database maintains an independent queue for each blade, and the queue length is dynamically configured according to local storage resources or set by the user. When the queue is full, an elimination strategy based on the signal-to-noise ratio of the acoustic signature segments and the stability of the operating conditions is adopted to prioritize the retention of healthy samples with high confidence.
[0012] As a preferred embodiment of the present invention, the step of determining the defect type based on the distribution pattern of the acoustic signature differences of the three blades includes: calculating the root mean square difference (RMSD) between the current stable period segment and the first historical stable period segment, calculating the RMSD between the current stable period segment and the second historical stable period segment, and calculating the RMSD between the two historical stable period segments; if the RMSD between the current stable period segment and the two historical stable period segments is greater than the RMSD between the two historical stable period segments, then the current blade is determined to have a defect; otherwise, the current blade is determined to be normal; wherein, when the current blade is determined to be normal, the RMSD between each blade is compared with a preset absolute defect threshold; if the RMSD of all three is greater than the absolute defect threshold, a multi-blade synchronous abnormal alarm is triggered, and the determination result of the current blade to be inspected is changed to a suspected defect, while it is recommended to use a backup inspection method for verification; the preset absolute defect threshold is determined by the statistical quantile of the RMSD of the acoustic signature segments during the historical healthy period of the wind turbine.
[0013] As a preferred embodiment of the present invention, the method further includes: when a defect is determined to exist, defect type localization is performed: the root mean square difference between the current stable period segment and each historical stable period segment in each preset frequency band is calculated respectively; for each frequency band, the larger of the two root mean square differences is taken as the difference value of the frequency band; the difference values of each frequency band are compared, and the frequency band with the largest difference value is determined as the target frequency band; the defect type is output according to the radial position corresponding to the target frequency band: if it is a low frequency band, it is determined to be a defect near the leaf root; if it is a mid frequency band, it is determined to be a defect near the leaf middle; if it is a high frequency band, it is determined to be a defect near the leaf tip.
[0014] On the other hand, a wind turbine blade intelligent inspection system based on blade acoustic pattern segmentation comparison and defect radial position mapping is also provided. This system further includes: a dual-cycle independent segmentation module, which collects blade acoustic patterns from two consecutive complete rotation cycles and performs segmentation operations independently to obtain a first-cycle segmentation point set and a second-cycle segmentation point set; a cycle consistency verification module, which performs cycle consistency verification on the first and second-cycle segmentation point sets. If the time interval between the two sets of segmentation points matches one rotation cycle, the second-cycle segmentation point set is used as the reference segmentation point; a frequency band-radial position correction module, which divides the acoustic pattern segment corresponding to the reference segmentation point into multiple frequency band sets, each frequency band set is associated with a different radial position of the blade, and performs position correction on the reference segmentation point based on the radial position to obtain the corrected segmentation point; and a three-blade asynchronous mutual calibration and defect mapping module, which extracts acoustic pattern segments of each stage of the current blade based on the corrected segmentation point and compares them with the historical corresponding stage acoustic pattern segments of the other two blades of the same wind turbine, determining the defect type based on the distribution pattern of the three-blade acoustic pattern differences.
[0015] The beneficial effects of this invention are as follows: This invention utilizes the low-frequency impact sound generated by the blade itself to achieve adaptive capture of the rotation cycle and dual-cycle consistency verification, effectively eliminating wind noise and speed fluctuation interference. It achieves robust segmentation and position correction of blade acoustic events without the need for additional sensors, significantly improving data reliability under complex operating conditions. By adopting an asynchronous mutual calibration strategy for three blades of the same unit, with healthy blades as a dynamic reference baseline, it effectively cancels common-mode interference such as environmental noise, wind speed changes, and sensor gain drift, greatly improving the accuracy and robustness of defect judgment. In addition, by mapping defects to different radial positions of the blade through multi-band root mean square difference analysis, it can accurately locate damage at the blade root, blade middle, or blade tip, providing precise maintenance guidance for operation and maintenance personnel, reducing maintenance costs. Moreover, the entire process can be automatically deployed online, meeting the engineering requirements of unmanned and high-frequency inspections in wind farms. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping, as shown in this invention.
[0017] Figure 2 This is a structural diagram of the intelligent inspection system for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping, as shown in this invention. Detailed Implementation
[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0019] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.
[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0021] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates an intelligent inspection method for wind turbine blades based on segmented blade acoustic signature comparison and radial defect location mapping, which includes: S1: Collect the blade acoustic signatures for two consecutive complete rotation cycles, and perform segmentation operations independently to obtain the first cycle segmentation point set and the second cycle segmentation point set.
[0022] S1.1: Voiceprint triggering and rotation cycle determination.
[0023] A wideband microphone is installed directly in front of the wind turbine tower to collect acoustic signals generated during turbine operation in real time. When any blade rotates to the front of the tower, the combined effect of the tower's obstruction effect and the blade cutting through the airflow generates an impact sound with a significant low-frequency component. In this embodiment, this low-frequency impact sound is used as the initial trigger signal.
[0024] The specific triggering conditions are as follows: (a) Calculate the short-time energy of the bandpass filtered signal in real time, with a window length of 20ms and a step size of 10ms, to obtain the energy sequence. (b) Average background noise energy A dynamic update strategy is adopted: during periods when the unit is operating normally and no trigger signal is detected (two consecutive seconds without trigger), short-term energy is collected every 0.5 seconds and updated accordingly. .initial Data is collected and stored in advance when the unit is idling and there is no wind. This is the average background noise value from the previous storage. The current short-term energy, initial value Data is collected and stored in advance when the unit is idling and there is no wind.
[0025] (c) The trigger threshold is set to 2.5 times (configurable range 2.0–3.0, adjustable according to the on-site signal-to-noise ratio). When A valid trigger signal is considered detected when the threshold is exceeded and the duration is between 5ms and 15ms. To avoid continuous triggering, a dead time of 300ms is set after a trigger is detected, during which no new triggers are detected.
[0026] Since wind turbines typically have three blades, adjacent impact sounds originate from different blades, and the time interval between two passes of the same blade directly in front of the tower is exactly one complete rotation cycle. This embodiment utilizes this characteristic to identify two adjacent trigger signals belonging to the same blade: the times t1, t2, and t3 of three consecutive trigger signals are recorded, and the adjacent intervals are calculated. , Under stable rotational speed conditions, the interval between two triggering operations on the same blade should be approximately equal to three times the average interval, i.e. However, to avoid relying on the assumption of stable rotational speed, this embodiment adopts the following method: based on the first trigger... Starting from, within the time range The next internal search will be triggered, in which The preset rated rotation period (e.g., 4 seconds) is used. The first trigger signal detected is considered the second trigger for the same blade, and the recording time is... .Will and The difference is determined as the measured rotation period. .
[0027] S1.2 Continuously acquire the original voiceprint sequence for two complete rotation cycles.
[0028] This invention employs an offline architecture of first acquiring data and then processing it, but the acquisition length is determined in real-time by the period detected online. The specific process is as follows: With the first trigger signal Starting with the first signal, continuous acoustic signature acquisition is initiated and buffered in a circular buffer; during the acquisition process, the second trigger signal of the same blade is continuously detected (i.e., Once detected, the measured rotation period is immediately calculated. At this point, data collection has not yet stopped; the system continues collecting data for exactly [duration]. The voiceprint data is collected, and then the collection is stopped manually. Therefore, from A total of 2× were initially collected. The duration of the data is stored in memory to obtain the original voiceprint sequence. All subsequent steps are for It performs offline processing, independent of real-time data streams. This design avoids interference from processing delays on subsequent trigger detections, while ensuring the determinism of operations such as segmentation and correction.
[0029] As can be seen, this embodiment dynamically determines the data collection endpoint (i.e., +2× This avoids the problem of not being able to determine the acquisition length due to unknown cycle. Moreover, it does not rely on external speed sensors, but can adaptively acquire the sound pattern of the whole cycle using only the impact sound generated by the blades themselves, simplifying hardware deployment and being able to follow the rotation cycle fluctuations caused by wind speed changes.
[0030] S1.3 Segment the voiceprint independently within the first rotation cycle.
[0031] From the original voiceprint sequence Extract the subsequence corresponding to the first rotation period from the middle. (Time range) to + ).right Perform the following segmented operations: (1) A Hamming window was used, with a window length of 20 ms (corresponding to 512 sampling points) and a step size of 10 ms (corresponding to 256 sampling points). The short-time energy within each window... ,in For window length, The amplitude of the voiceprint within the window is used to obtain the energy envelope sequence. .
[0032] (2) To Perform first-order finite difference to find the inflection point where the difference changes from negative to positive. Simultaneously, the energy value at this point must be less than the energy values of the two adjacent points, and the time interval between minimum points must not be less than a preset minimum interval (set to 50ms in this embodiment). To avoid spurious minimum points caused by background noise or turbulence, the following two-stage screening is performed: Level 1: Calculate the mean and standard deviation of E(m) over the entire rotation cycle, and retain only the minimum value point where the energy value is less than 0.6 times the mean (empirical value, which can be adjusted between 0.4 and 0.8 times the noise level at the site).
[0033] Level 2: For the remaining minimum points, calculate the interval between adjacent points in chronological order. Since there are 3 leaves, theoretically there should be 3 dominant minimum points, each located after the passing event of each leaf. Points whose intervals deviate significantly from the average interval (deviation exceeding ±30%) are removed, and consecutively missing points are filled in using linear interpolation, but the total number of filled points should not exceed 1. If the final number of valid minimum points is not equal to 3, the segmentation is considered a failure.
[0034] Arrange all local minima that meet the conditions in ascending order of time.
[0035] (3) Since the wind turbine has three blades, theoretically, three segmentation points should be detected in each rotation cycle, corresponding to a fixed phase after the three blades pass directly in front of the tower (e.g., the first envelope valley value after the impact sound ends). If the number of detected local minimum points is not equal to 3, the segmentation is judged to have failed, the current data is discarded, and S1.1 to S1.3 are re-executed; if the number is equal to 3, the time coordinates corresponding to these three minimum points (relative to the time coordinates of the three minimum points) are set. The absolute time of the start moment is output as the set of segment points for the first period, denoted as... And satisfy .
[0036] The segmentation method described above divides different blade events within the rotation period by using the valleys of the short-time energy envelope. It has the advantages of strong noise resistance and simple computation. It eliminates the need for complex blind source separation or direction-of-arrival estimation, achieving coarse localization of blade events solely using time-domain energy characteristics, thus providing a reliable initial boundary for subsequent period consistency verification.
[0037] S1.4 Independently repeat segments of the voiceprint within the second rotation cycle.
[0038] From the original voiceprint sequence Extract the subsequence corresponding to the second rotation period from the middle. (Time range) + to +2 The short-time energy envelope calculation, local minimum detection, and segmentation point selection steps in S1.3 are repeated completely independently to obtain the second-period segmentation point set, denoted as... They are also arranged in ascending order of time and have a quantity of 3.
[0039] In this embodiment, the segmentation operations of the two rotation cycles are independent of each other and do not borrow information from each other. This can truly reflect the differences in the position of the segmentation points caused by speed fluctuations or changes in environmental noise within the two cycles, providing an objective comparison basis for subsequent cycle consistency verification and avoiding the masking of actual cycle mismatches due to the reuse of the same segmentation results.
[0040] Furthermore, if the number of local minima detected in either S1.3 or S1.4 is not 3, the system automatically discards the two complete cycle voiceprints currently acquired and returns to S1.1 to restart triggering and acquisition until two sets of segmented points, each with 3 points, are successfully obtained. This mechanism ensures that the data input to subsequent steps has the correct structure, improving the robustness of the entire inspection method.
[0041] S2: Perform a period consistency check on the first and second period segment point sets. If the time interval between the two sets of segment points matches a rotation period, then the second period segment point set is used as the reference segment point.
[0042] S2.1 Time window prediction and cross-cycle pairing based on measured rotation period.
[0043] The first period segmentation point set is obtained from step S1. and the set of segmentation points in the second period Both sets of segmentation points are arranged in ascending order of time. However, due to minor speed fluctuations in wind turbines within two consecutive rotational cycles or local offsets introduced by the segmentation algorithm, directly pairing by sequential indexing may lead to incorrect blade event correspondences (e.g., if...). and The intervals between them were incorrectly segmented due to noise interference. (The points in the equation may be shifted by one phase overall). Therefore, this embodiment uses a method based on the measured rotation period. The prediction window method is used for pairing.
[0044] For the first cycle, the Segmentation points ( =1,2,3), calculate the predicted time within the second period:
[0045] in This refers to the measured rotation period determined in S1.1. Considering rotational speed fluctuations and segmentation errors, a time window is set. Window width set to , The value is determined based on the stability of the rotation period: the wind speed change rate (the absolute value of the linear regression slope of the wind speed over the past 10 seconds is calculated by recording the wind speed once per second through the wind turbine's SCADA system) is less than 0.5 m / s. 2 Under stable conditions, Take 0.08; when the wind speed variation rate is large, Use 0.12. If the wind speed change rate cannot be obtained in real time, use a fixed value of 0.1, ensuring that the window includes more than 95% of normal fluctuations.
[0046] by Centered on, the window range is .
[0047] In the second set of segmented points, search for segmented points falling within the aforementioned window range. Since theoretically each window should contain exactly one segmented point, and this point corresponds to the same passage event on the same blade, the searched points are denoted as... If multiple segment points are found within the window, select the one that matches the segment point. The point with the closest time distance is used as the matching point; if no segment point is found in the window, the matching is determined to be unsuccessful, the current data is discarded and the process returns to step S1 to collect data again.
[0048] Using the above method, pairing is performed for j=1, 2, 3 respectively, resulting in three pairs: .
[0049] It should be noted that, in the implementation of this invention, the obtained measured rotation period is utilized. Prediction overcomes the phase misalignment problem that may be caused by direct sequential indexing. Even when there are certain fluctuations in rotational speed or slight offsets in segmentation points, it can still reliably identify corresponding blade events across cycles, thus improving the robustness and accuracy of cycle consistency verification.
[0050] S2.2 Calculate the time difference and determine the period matching.
[0051] For each pairing, calculate the time difference between the second period segment point and the first period segment point:
[0052] Theoretically, if the voiceprint segments are completely identical within two rotational cycles and the rotational speed is constant, then each It should be strictly equal to the measured rotation period. In actual operation, due to the gradual change in rotational speed caused by wind speed variations or the limited accuracy of the piecewise algorithm, and There will be slight deviations. This embodiment sets a deviation tolerance, which is expressed as a relative proportion, and is set to... ,in The value is associated with v, and is generally taken as... In this embodiment =0.1 =0.06. This tolerance is used to determine whether the periods of the two sets of segment points are consistent: if all If the three retries fail, the system will automatically determine if they match. Increase the value to 1.5 times the original value and try again. If it still fails, log the error and exit.
[0053] If the above conditions are met, then the first set of periodic segment points and the second set of periodic segment points are determined to have periodic consistency, that is, the blades indicated by the two sets of segment points match a complete rotation cycle in time through events. Otherwise, the periods are determined to be inconsistent.
[0054] When a discrepancy is detected, the system automatically discards the two complete cycles of voiceprint data collected and the already obtained set of segmentation points, and returns to step S1.1 to re-perform the triggering, collection, and segmentation operations. To ensure that the system does not loop indefinitely due to occasional noise, a maximum number of retries can be set (e.g., 3 times). If this number is exceeded, an inspection failure alarm will be issued and the current inspection task will be terminated.
[0055] As can be seen, in this embodiment of the invention, by using quantitative deviation tolerance judgment, abnormal data caused by strong noise interference, sudden changes in rotation speed or segmentation failure is effectively eliminated, ensuring that the data entering the subsequent correction and comparison stages has high reliability and avoiding misjudgment caused by the propagation of incorrect segmentation.
[0056] S2.3 Determination of reference segment points.
[0057] Assuming the judgment period is consistent, the second period segment point set is copied and stored as the baseline segment point set, denoted as... ,Right now:
[0058] It should be noted that the second period segment point was chosen as the reference instead of the first period because, during the acquisition of two consecutive periods, the second period is closer in time to the subsequent voiceprint processing and correction operations. Furthermore, the second period has undergone pairing verification with the first period, resulting in higher positional accuracy. Additionally, retaining the second period as the reference facilitates subsequent segment extraction from the original voiceprint sequence (step S3.1 in S3 requires extraction from the reference segment point; choosing the second period avoids boundary extraction issues).
[0059] In this embodiment, the set of reference segment points Arranged in ascending chronological order, each point corresponds to the same fixed phase (i.e., the local minimum point of the short-time energy envelope) after the three blades pass directly in front of the tower during the second rotation cycle. Subsequent steps will perform band-radial position correction based on this reference segment point.
[0060] It should be noted that the selection of the second period segment point as the benchmark in this embodiment is a preferred solution. However, those skilled in the art will understand that the first period segment point can also be used as the benchmark after the period consistency verification, and the technical effect is essentially the same.
[0061] If no matching point is found in any window search in S2.1, the pairing is deemed to have failed, and the process returns to step S1; if any pair in S2.2... If the deviation tolerance is exceeded, the cycle is determined to be inconsistent, and the return step S1 is executed; if three consecutive retries fail, the system records the error log and exits the inspection, prompting maintenance personnel to check the acoustic sensor or the unit's operating status.
[0062] This anomaly handling mechanism prevents the system from getting stuck in an invalid loop for a long time due to brief environmental noise or wind speed fluctuations, while providing clear fault indications for maintenance personnel and improving the engineering practicality of the method.
[0063] S3: Divide the acoustic pattern segment corresponding to the reference segment point into multiple frequency band sets, and associate each frequency band set with different radial positions of the blade. Based on the radial positions, perform position correction on the reference segment point to obtain the corrected segment point.
[0064] S3.1 Extract the voiceprint segment corresponding to the reference segment point from the original voiceprint sequence.
[0065] The set of reference segment points is obtained from step S2. The points are arranged in ascending chronological order, each corresponding to the same fixed phase (a local minimum of the short-time energy envelope) after the three blades pass directly in front of the tower during the second rotation cycle. Starting from each baseline segment point, the original acoustic signature sequence... A fixed-length audioprint segment is extracted from the middle and recorded as follows: .
[0066] Capture duration Set as the measured rotation period Multiplied by a preset scaling factor , The value range is 0.15 to 0.25. In this embodiment... Take 0.2, that is = 0.2 The duration is chosen based on the following: the effective acoustic signature generated after each blade passes directly in front of the tower usually does not exceed 0.25 rotation cycles (because the sound source gradually moves away from the tower after the blade passes, and the signal attenuates). Therefore, 0.2 times the cycle can cover the main acoustic signature features and avoid the event of the next blade being mixed in.
[0067] The truncation operation employs precise time alignment: based on the sampling rate. (In this embodiment, the frequency is set to 25.6kHz), The time coordinates are converted to sampling point indices and then offset backwards. × Each sampling point yields a fragment. The endpoint. If the endpoint exceeds... If the data exceeds the end of the second rotation cycle, it is determined that the data is out of bounds, the current data is discarded, and the process returns to step S1 to re-acquire data. This out-of-bounds check ensures the integrity of subsequent processing.
[0068] This step yields three voiceprint fragments. The length of each segment is × Each sampling point corresponds to a voiceprint feature segment after the three blades pass through.
[0069] As can be seen, this invention uses the reference segmentation point as the starting point to truncate backwards, avoiding cross-cycle aliasing that may be caused by forward truncation. At the same time, the length control ensures that each segment contains only the event soundprint of a single blade, providing a clean input for subsequent frequency band decomposition.
[0070] S3.2 Perform multi-band bandpass filtering on each voiceprint segment and associate it with radial position.
[0071] For each voiceprint segment The filters are passed through three preset frequency band digital bandpass filters. The filter design parameters are as follows: The low-frequency bandpass filter has a passband frequency range of 50Hz to 200Hz, with a center frequency of 125Hz. This frequency band mainly includes airflow disturbances, structural vibrations, and tower-reflected sound near the blade root, which is related to the radial relative position of the blade, r / R = 0.6 (i.e., near the blade root). Here, r is the radial distance from a point on the blade to the blade root (where the blade connects to the hub), in meters (m). For an actual blade, r ranges from 0 (blade root) to the total blade length (blade tip). R is the total blade length (i.e., the distance from the blade root to the blade tip), in meters (m). For a given wind turbine model, R is a known design parameter.
[0072] The passband frequency range of the mid-frequency bandpass filter is 200Hz to 500Hz, with a center frequency of 350Hz. This frequency band mainly includes eddy current shedding and surface friction noise in the middle region of the blade, associated with a radial position r / R = 0.7 (near the middle of the blade).
[0073] The passband frequency range of the high-frequency bandpass filter is 500Hz to 1000Hz, with a center frequency of 750Hz. This band mainly includes turbulent boundary layer noise and small-scale vortices in the leading edge and tip regions of the blade, associated with a radial position r / R = 0.8 (near the blade tip).
[0074] The filter type used is a fourth-order Butterworth infinite impulse response filter, which has the advantages of a flat passband and low phase distortion, making it suitable for acoustic envelope analysis. The segment is first processed before filtering. Zero-phase forward-backward filtering (i.e., FLTFFIT processing) is performed to eliminate the group delay caused by the filter and ensure that the envelope peak time position of the filtered signal does not undergo systematic shift.
[0075] After filtering, each original segment Three sub-segments are generated: (Low frequency) (Medium frequency) (High frequency). Each sub-segment represents the acoustic signature of the same blade at different radial positions.
[0076] As can be seen, separating the acoustic responses of different regions of the blade through frequency band division allows subsequent offset calculations to reflect local differences in different parts of the blade, providing a physical basis for radial position correction. Compared to full-band processing, the multi-band decomposition of this invention significantly improves the sensitivity to local defects.
[0077] S3.3 Calculate the envelope peak time offset for each filtered band segment.
[0078] For each filtered sub-segment (in terms of...) Taking this as an example, the Hilbert transform is used to extract the envelope. Specifically, the Hilbert transform is applied to the sub-segment signal x(t) to obtain the analytic signal.
[0079]
[0080] in, It is a continuous time variable, with the unit being seconds (s). This is the original real signal, i.e., the audioprint segment after bandpass filtering. for The Hilbert transform, in the frequency domain, is equivalent to shifting the negative frequency components of a signal by +90° and the positive frequency components by -90°. For discrete signals, this is achieved using the FFT. The analytic signal is a complex time function. For the envelope amplitude to be equal to the analytic signal The model. For example, if ,but Analyzing signals envelope (Constant). The actual voiceprint signal is an amplitude-modulated and frequency-modulated wave, the envelope of which varies with time, and the peak value represents the strongest response to the event.
[0081] Since zero-phase filtering removes group delay, the peak time position of the envelope A(t) directly reflects the occurrence time of the main acoustic events in this frequency band.
[0082] Detect the maximum value of A(t) on the time axis, and record the time coordinate of this point relative to the start time of the sub-segment, denoted as . Since the start time of the sub-segment is the reference segmentation point. Given the time coordinates, the time offset of the envelope peak of this frequency band relative to the reference segment point is: Similarly, for and Calculate the envelope peak offset separately and .
[0083] It should be noted that, theoretically, the arrival time of acoustic signatures at different radial positions of the same blade to the microphone varies slightly (the blade tip is closer to the tower, resulting in a shorter sound propagation path). Therefore, the offsets in the three frequency bands are usually different. This difference is precisely the information used for subsequent position correction.
[0084] As can be seen, the embodiment of the present invention employs Hilbert envelope peak detection, which can stably extract the moment when the energy is most concentrated within each frequency band, avoiding peak jitter caused by noise. Simultaneously, defining the offset as absolute time relative to the reference segment point gives subsequent correction operations a clear physical meaning.
[0085] S3.4 Calculate the average offset and correct the reference segment point.
[0086] For each blade j, the offsets corresponding to the three frequency bands are arithmetically averaged to obtain the average offset. The above average values, combining the acoustic propagation characteristics at the three radial positions of the blade root, middle, and tip, yield a correction value that represents the overall phase of the blade passage event. Since different defect types affect the offset distribution in different frequency bands, averaging can provide a robust global correction value even when the defect type is unknown, avoiding extreme offsets caused by local defect interference in a single frequency band.
[0087] Furthermore, the benchmark segmentation points Time coordinates plus After correction, the segment points are obtained. :
[0088] It should be noted that, Originally, this corresponded to the local minimum point of the short-time energy envelope (the energy trough after the blade passes through the tower), while the peak points of the envelope in each frequency band corresponded to the strongest acoustic response at different radial positions. After taking the average offset, The position was corrected to the core acoustic energy concentration area of the blade passage event, rather than the trough area. This corrected segment point serves as the central reference for subsequent steady-state extraction, ensuring that the extracted segments contain the richest blade state information. Experiments show that the time deviation between the corrected segment point and the blade azimuth angle of 0° (directly facing the tower) is less than 0.02× .
[0089] Perform the above operation on all j=1,2,3 to obtain the corrected set of segment points. , , Sort in ascending order by time.
[0090] For each frequency band, the effectiveness of envelope peak detection must simultaneously meet the following two conditions: The ratio of the maximum envelope value to the statistical average of the envelope amplitude A(t) within the entire sub-segment is ≥1.5 (an empirical threshold, adjustable between 1.3 and 2.0); 0.02× before and after the peak point. Within the neighborhood of the peak value, the envelope value decreases by no more than 20% of the peak value, and there are no other local maxima within 10% of the peak value.
[0091] In this embodiment, the statistical averages include, but are not limited to, the arithmetic mean, the weighted average, and the median; when no weighting coefficient is specifically specified, the arithmetic mean is used by default.
[0092] If a frequency band does not meet any of the above conditions, the offset of that frequency band is deemed invalid and discarded. If there are fewer than two valid frequency bands among the three frequency bands, the correction is deemed to have failed, and the process returns to step S1 for re-acquisition. If there are two valid frequency bands, the offsets of these two frequency bands are used to calculate the statistical average; if there are three, all three are used.
[0093] It should be noted that, in this embodiment of the invention, the reference segment points are corrected by multi-band average offset, eliminating systematic errors caused by differences in sound propagation paths or filter group delays. This makes the corrected segment points more accurately correspond to the true phase position of each blade in the acoustic signature. Compared to directly using the original segment points, the stable period segments extracted from the corrected segment points have higher temporal consistency, thereby improving the reliability of subsequent three-blade difference comparisons.
[0094] Furthermore, to prevent over-correction from causing a disorder in the segmentation order, in the output... Perform order verification before assembling: Check Whether this condition is met, and whether the time interval between two adjacent points is greater than the preset minimum interval (set to 0.1× in this embodiment). If the sequence condition is not met, it indicates that the correction offset is too large, which may be caused by abnormal noise or segmentation errors. In this case, discard the current data and return to step S1. The above operations can ensure the physical rationality of the segment points after correction, avoid overlap or out-of-bounds errors in subsequent stable period segment extraction due to outliers, and enhance the fault tolerance of the method.
[0095] S4: Extract the acoustic signature segments of each stage of the current blade based on the corrected segment points, and compare them with the historical acoustic signature segments of the other two blades of the same wind turbine. Determine the defect type based on the distribution pattern of the acoustic signature differences among the three blades.
[0096] S4.1 Extract the stable period segment of the current blade to be detected based on the corrected segment points.
[0097] The corrected set of segment points is obtained from step S3. , , These correspond to the time points after phase correction for the three blades during the second rotation cycle. One of the three blades is randomly selected as the current blade to be detected; in this embodiment, the corrected segment point is used. The following explanation will be based on the corresponding blade (denoted as blade A).
[0098] by Using a fixed time interval as the center point, extend a fixed duration to both sides of the time axis to obtain the stationary period segments. The extension duration is set to the measured rotation period. Multiplied by a preset scaling factor , The value range is 0.04 to 0.08. In this embodiment... Taking 0.05, that is, the one-sided propagation time is 0.05 × The total segment length is 0.1× The basis for choosing this duration is that after the blade passes directly in front of the tower, the acoustic signature signal will exhibit relatively stable characteristics within a short and stable window (approximately 0.1 times the period). Within this window, the blade attitude is basically fixed, and the background noise changes little, making it suitable for amplitude comparison.
[0099] The start time of the expanded fragment is The end time is From the original voiceprint sequence The voiceprint data within this time range is extracted and denoted as... Boundary checks are required during the interception process: if the starting time is earlier than the start time of the second rotation cycle (i.e., ... ) or the end time is later than the end of the second rotation cycle (i.e. If the data is found to be out of bounds, it is processed by mirror symmetry or by extracting the largest usable segment within the effective range. If the effective length is less than 80% of the total segment length, the current data is discarded and the process returns to step S1 to collect data again.
[0100] It should be noted that this invention extends symmetrically around the corrected segment point, which can extract a stable acoustic pattern before and after the blade passes through the event. This preserves the main features of the event while avoiding the mixing of large-scale background noise, providing a high signal-to-noise ratio input for subsequent cross-blade comparison.
[0101] S4.2 Retrieve historical stable period fragments of two other blades from the historical voiceprint database.
[0102] Wind turbines are typically equipped with three blades, denoted as blade A, blade B, and blade C. In this embodiment, the blade to be detected is blade A, and the other two blades are blade B and blade C.
[0103] A historical voiceprint database is pre-built and maintained as follows: After each successful inspection (i.e., after steps S1 to S3 without triggering re-collection), the stable period segments corresponding to the three corrected segment points obtained in this collection are stored in the database (extracted in the same way as in S4.1). Each stored record includes: turbine number, blade identifier (B or C), collection timestamp, and cumulative rotation cycle count (the total number of rotations accumulated since the unit was put into operation, which can be obtained from the measured rotation cycle). The original voiceprint data was obtained by estimating the start time of the data collection, as well as the raw voiceprint data of the stable period segment.
[0104] During this inspection, historical stable period segments of blades B and C were queried from the database. The query criteria were: the segment closest to the current acquisition time, and the absolute value of the difference between the cumulative number of rotation cycles and the cumulative number of rotation cycles acquired in this acquisition does not exceed a preset threshold (set to 3 rotations in this embodiment). The basis for setting this threshold is that a difference of ≤3 rotation cycles means that the rotation phase of the blades was close during the two acquisitions, and the changes in operating conditions such as wind speed and blade pitch angle were small, making the acoustic waveform baselines comparable.
[0105] The query results are denoted as and If there are no records in the database that meet the conditions (e.g., during the first inspection), this step cannot be executed. The system will issue a prompt and end the current inspection, and will continue to inspect after accumulating sufficient historical data.
[0106] As can be seen, this invention effectively eliminates the influence of different operating conditions (wind speed, rotational speed, load) on acoustic signature amplitude by constraining the selection of historical data through the cumulative number of rotational cycles. This ensures that the comparison between blades across time is based on similar operating conditions, thereby improving the accuracy of defect determination. At the same time, the use of the rule of closest time ensures the freshness of historical data and avoids baseline drift caused by long-term blade aging.
[0107] S4.3: Calculate the root mean square difference between the three-blade acoustic waveform segments.
[0108] Among them, the three voiceprint segments are: the stable period segment of the current blade A. (Length is N sampling points), a stable period fragment of historical blade B Stable period fragment of historical blade C The original sampling rates of the three segments should be the same. If the sampling rates of historical segments in the database are inconsistent with the sampling rates of the currently acquired segments, a linear interpolation resampling method is used to unify the historical segments to the current sampling rate (in this embodiment, the current sampling rate is 25.6 kHz).
[0109] Furthermore, calculate the root mean square difference of the three groups: (1) Root mean square difference between current leaf A and historical leaf B for:
[0110] in, for The Amplitude at each sampling point, for The resampled first Amplitude of each sampling point.
[0111] (2) Similarly, the root mean square difference between the current leaf A and the historical leaf C can be calculated. The root mean square difference between historical leaf B and historical leaf C is .
[0112] Before calculation, each segment undergoes normalization preprocessing to eliminate overall energy differences caused by wind speed, distance, etc.: subtract the DC bias within the segment. Then, energy normalization is performed: the effective value (RMS) of the fragment is calculated to obtain the normalized fragment. If the RMS is less than the preset noise threshold (set to 0.01 times the ADC full scale in this embodiment), the segment is determined to be invalid, the current data is discarded, and the process returns to step S1.
[0113] This normalization method ensures These indicators primarily reflect differences in waveform shape rather than differences in gain.
[0114] After normalization, The value reflects only the difference in waveform shape rather than the difference in absolute energy, which is robust to changes in acoustic energy under different wind speeds.
[0115] It should be noted that the root mean square difference (RMSD) is a simple and effective measure of waveform similarity, capable of comprehensively reflecting the amplitude difference between two segments in the time domain. By simultaneously calculating the RMSD of three pairwise combinations, complete pairwise comparison information is provided for subsequent defect determination.
[0116] S4.4: Determine whether there is a defect in the current blade based on the differential distribution pattern of the three blade acoustic signatures.
[0117] This embodiment employs an outlier detection strategy: if blade A has a defect, the difference between its acoustic signature segment and those of normal blades B and C should be significantly greater than the difference between normal blades B and C themselves. Therefore, the determination criteria are as follows: like( > ) and( > If the condition is met, then blade A is determined to have a defect; otherwise, blade A is determined to be normal.
[0118] The physical basis of this judgment logic is that, under normal operating conditions, the acoustic signatures of the three blades should be basically consistent. , , The three values are similar and all are relatively small; when a blade is damaged (such as cracks, pinholes, or corrosion), the sound pattern of that blade will change significantly, causing the root mean square difference between it and the other two normal blades to increase, while the root mean square difference between the other two normal blades remains at a low level.
[0119] To prevent missed detections when three blades have similar damage simultaneously, this embodiment adds a degradation-level fault-tolerant judgment to the above-mentioned judgment logic. When condition ( > ) and( > If this condition is not met, and the system initially determines that the current blade is normal, then the system will further... , , The data was compared with a preset absolute defect threshold. The absolute defect threshold was calibrated as follows: at least 20 successful inspections of acoustic signature segments, manually confirmed as defect-free, were collected during the initial operation of the wind turbine. The data under all historical normal conditions were then calculated. The values are used, and the 95th percentile of these values is taken as the absolute defect threshold. If... , , If all three values exceed the absolute defect threshold, it is determined that there may be a synchronization anomaly in the three blades. In this case, instead of outputting a conclusion that the current blade is normal, a multi-blade synchronization anomaly alarm is triggered, the judgment result of the current blade under inspection is changed to a suspected defect, and it is recommended that maintenance personnel use alternative inspection methods such as drones or manual visual inspection for verification. This fault tolerance mechanism is only activated as a fallback measure when the historical data of the reference blade may also no longer be healthy, and does not affect the main judgment logic under normal operating conditions.
[0120] If the condition is deemed normal, the inspection ends, and the stable period segment of current blade A is stored as new historical data in the database. Before storage, the queue length of the current blade is checked: if it has not reached the configured maximum length, it is inserted directly; if it has reached the maximum length, the scores of all existing and new records are calculated according to the aforementioned weighted elimination mechanism, and the record with the lowest score is removed before inserting the new record. This process ensures that a high-quality blade health baseline is always maintained in the historical database.
[0121] Each blade independently maintains its own historical record queue. To adapt to the storage and computing resource constraints of large-scale wind farm deployments, the database management adopts the following optimization strategies: the queue length is dynamically configured based on the available storage capacity of the local edge computing node, or set by maintenance personnel through a configuration file, supporting a range of 5 to 30 times; when the queue is full and a new normal segment needs to be inserted, a weighted scoring elimination strategy based on data quality is adopted. The score of each record is based on three dimensions: the average wind speed change rate during the inspection period (reflecting the stability of the operating condition, the lower the change rate, the higher the score), the signal-to-noise ratio of the acoustic signature segment (the higher the signal-to-noise ratio, the higher the score), and the average waveform similarity between the segment and other segments in the queue (retaining typical samples and eliminating outliers). The record with the lowest score is eliminated each time; to reduce storage overhead, the database supports two storage modes—original waveform mode (default) and feature-level storage mode. In feature-level storage mode, the original acoustic signature waveform is not saved, but the following feature vectors of the stable period segment are saved: the envelope peak amplitude and root mean square value of the envelope for each frequency band (low frequency, mid frequency, high frequency), and the first 12 dimensions of the Mel frequency cepstral coefficients. Feature-level storage can compress the amount of data recorded in a single instance to less than 5% of the original waveform, while still meeting the calculation requirements for subsequent root mean square difference comparison and defect location. The mode selection is specified by the system configuration file, with the default factory setting being the original waveform mode.
[0122] When blade A is determined to be normal, only the historical queue of blade A is updated; the historical queues of blades B and C remain unchanged; when blade A is determined to have a defect, its historical queue is not updated to prevent defect data from contaminating the health baseline. Simultaneously, the defect data is separately tagged and stored in a defect record library for subsequent trend analysis.
[0123] At any given time, when blade B or C is selected as the currently inspected blade, the corresponding historical acoustic signature segments retrieved are all from their respective health history queues (i.e., the most recently determined record of normal operation). If a blade has never been determined to be normal (e.g., a new unit), it cannot be used as a reference, and the system will prompt that sufficient health data needs to be accumulated.
[0124] If a defect is determined to exist, proceed to step S4.5 to locate the defect type.
[0125] As can be seen from the above operations, the present invention does not rely on absolute thresholds, but uses the relative differences between the three blades of the same unit to determine defects, effectively offsetting common-mode interference such as environmental noise, wind speed changes, and sensor gain drift, and has self-calibration characteristics, which significantly improves the robustness and accuracy of detection.
[0126] S4.5 Defect type location: Defect location is mapped based on the root mean square difference of the frequency band.
[0127] When a defect is determined to exist in leaf A, the specific type of defect is further analyzed (defect near the leaf root, defect near the leaf middle, or defect near the leaf tip). For this purpose, three stationary phase segments ( , , Perform the same multi-band bandpass filtering process as in step S3.2, that is, extract sub-segments of the low frequency band (50-200Hz), mid frequency band (200-500Hz), and high frequency band (500-1000Hz).
[0128] Furthermore, for each frequency band ( = low, mid, high), calculate the following two root mean square differences: The root mean square difference between current blade A and historical blade B in this frequency band ; The root mean square difference between current blade A and historical blade C in this frequency band .
[0129] The calculation method is the same as S4.3, but normalization and root mean square error are calculated independently for each frequency band. The segment length of each frequency band may vary slightly due to the group delay of the filter, and the number of sampling points must be strictly aligned after resampling.
[0130] For each frequency band The larger of the two root mean square differences is taken as the final difference value for that frequency band.
[0131] Then compare the differences among the three frequency bands: , , Find the frequency band corresponding to the maximum value and denote it as the target frequency band.
[0132] Based on the preset correlation between the target frequency band and the radial position of the blade, the defect type is output: If the target frequency band is a low frequency band (50-200Hz), it is determined that there is a defect near the blade root. Specific types include, but are not limited to, root cracks, loose bolts, or structural fatigue.
[0133] If the target frequency band is the mid-frequency band (200-500Hz), it is determined that there is a defect near the middle of the blade, and the specific types include but are not limited to central sand holes, delamination, or interlayer damage.
[0134] If the target frequency band is a high-frequency band (500-1000Hz), it is determined that there is a defect near the tip of the current blade, and the specific types include but are not limited to leading edge corrosion, trailing edge cracking or lightning damage.
[0135] If the difference values of two or three frequency bands are equal or the difference value is less than the preset relative tolerance (for example, the maximum difference value does not exceed 1.1 times the second difference value), it is determined to be a composite defect or the frequency band distinction is unclear. In this case, the output cannot locate the defect type and manual review is recommended.
[0136] It should be noted that the main defects at different radial locations affect the aeroacoustic characteristics of the blade in different frequency bands (root defects mainly modulate the low-frequency band, while blade tip defects mainly affect the high-frequency band). By comparing the differences in multiple frequency bands, defects can be mapped to specific spatial regions, providing maintenance personnel with precise maintenance guidance, avoiding blind inspections or complete blade replacement, and significantly reducing maintenance costs.
[0137] Finally, after completing the defect determination (normal or defect type location), the inspection results are output to the monitoring system or maintenance terminal. The output includes: inspection time, turbine number, current blade identifier (blade A), and determination conclusion (normal / defect type). If the defect is determined, the characteristic frequency band, difference value, and suggested radial location for inspection are additionally output.
[0138] At the same time, the stable period fragment of the current leaf A collected this time will be... The corresponding collection timestamp and cumulative rotation cycle count are stored in the historical voiceprint database.
[0139] In this embodiment, the historical segments of blades B and C are used only for comparison and their historical data are not updated due to the current inspection (unless B or C becomes the current blade to be inspected). The advantage of this design is that it avoids contaminating the historical normal baseline due to current blade defects, thus ensuring the reliability of the historical reference segments.
[0140] On the other hand, a wind turbine blade intelligent inspection system based on blade acoustic pattern segmentation comparison and defect radial position mapping is also provided. This system further includes: a dual-cycle independent segmentation module, which collects blade acoustic patterns from two consecutive complete rotation cycles and performs segmentation operations independently to obtain a first-cycle segmentation point set and a second-cycle segmentation point set; a cycle consistency verification module, which performs cycle consistency verification on the first and second-cycle segmentation point sets. If the time interval between the two sets of segmentation points matches one rotation cycle, the second-cycle segmentation point set is used as the reference segmentation point; a frequency band-radial position correction module, which divides the acoustic pattern segment corresponding to the reference segmentation point into multiple frequency band sets, each frequency band set is associated with a different radial position of the blade, and performs position correction on the reference segmentation point based on the radial position to obtain the corrected segmentation point; and a three-blade asynchronous mutual calibration and defect mapping module, which extracts acoustic pattern segments of each stage of the current blade based on the corrected segmentation point and compares them with the historical corresponding stage acoustic pattern segments of the other two blades of the same wind turbine, determining the defect type based on the distribution pattern of the three-blade acoustic pattern differences.
[0141] The system also includes one or more processors and memory.
[0142] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.
[0143] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software, the software including instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, including the flow of the intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping of the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.
[0144] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.
[0145] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.
[0146] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if necessary, the program can be implemented in assembly or machine language.
[0147] In any case, the language can be either compiled or interpreted.
[0148] Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit.
[0149] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.
[0150] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.
[0151] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.
[0152] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.
[0153] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.
[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for intelligent inspection of wind turbine blades based on segmented comparison of blade acoustic signatures and radial location mapping of defects, characterized in that: include: The blade acoustic signatures for two consecutive complete rotation cycles are collected, and segmentation operations are performed independently to obtain the first cycle segmentation point set and the second cycle segmentation point set. Perform a period consistency check on the first and second period segment point sets. If the time interval between the two sets of segment points matches one rotation cycle, then the second period segment point set is used as the reference segment point. The acoustic signature segment corresponding to the reference segment point is divided into multiple frequency band sets, and each frequency band set is associated with a different radial position of the blade. The reference segment point is then corrected based on the radial position to obtain the corrected segment point. Based on the corrected segment points, the acoustic signature segments of each stage of the current blade are extracted and compared with the historical acoustic signature segments of the other two blades of the same wind turbine. The defect type is determined based on the distribution pattern of the acoustic signature differences among the three blades.
2. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 1, characterized in that: The separate and independent execution of segmented operations includes: The low-frequency impact sound generated each time the blade passes directly in front of the tower is used as the initial trigger signal. By detecting two consecutive impact sounds and using the condition that the number of wind turbine blades is known, two adjacent trigger signals belonging to the same blade are identified, and the time interval between the two triggers is determined as the measured rotation period. The acoustic data collection begins with the first trigger signal. After the second trigger signal of the same blade is detected and the measured rotation cycle is obtained, acoustic data for the duration of one measured rotation cycle is collected to obtain the original acoustic sequence covering two complete measured rotation cycles. The voiceprints within the first rotation cycle of the original voiceprint sequence are segmented by detecting local minima of the short-time energy envelope to obtain the first cycle segmentation point set; The number of segment points in the first cycle segment point set is equal to the number of blades in the wind turbine; For the voiceprint in the second rotation cycle of the original voiceprint sequence, the independent repeating segmentation method is used to obtain the second cycle segmentation point set, and the number of segmentation points in the second cycle segmentation point set is the same as that in the first cycle segmentation point set.
3. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 2, characterized in that: The periodic consistency check includes: A time window is predicted for each segment point in the first cycle segment point set using the measured rotation period. The center of the time window is the time coordinate of the first cycle segment point plus a measured rotation period, and the window width is a preset ratio multiplied by the measured rotation period. Within the corresponding time window, search for the segment point in the second cycle segment point set that is closest in time as the pairing point, so that the segment points in the first cycle and the second cycle are paired one by one by the blade through events. Calculate the time difference between the second cycle segment point and the first cycle segment point in each pair. If the deviation of each time difference relative to the measured rotation cycle does not exceed the preset deviation tolerance, it is determined that the cycles of the two sets of segment points are consistent; otherwise, discard the current data and re-acquire it. The second set of periodic segment points is used as the reference set of segment points.
4. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 1, characterized in that: The position correction of the reference segment points to obtain the corrected segment points includes: Starting from each segment point in the set of reference segment points, extract a first set duration of voiceprint segments from the original voiceprint sequence to obtain the same number of voiceprint segments as the number of blades. The first set duration is less than the measured rotation period. Each acoustic signature segment is passed through multiple bandpass filters with preset frequency bands. Each frequency band is pre-associated with a radial position on the blade, corresponding to low frequency, mid frequency, and high frequency from the blade root to the blade tip, respectively. For each frequency band filtered segment, extract the time point corresponding to the maximum envelope value, and calculate the time offset between the time point and the corresponding reference segment point; For each acoustic signature segment, calculate the statistical average of the time offset corresponding to all frequency bands, and add the statistical average to the time coordinate of the corresponding reference segment point to obtain the corrected segment point of each blade. The corrected segmentation points of all blades constitute the corrected segmentation point set.
5. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 4, characterized in that: The step of extracting acoustic signature segments of each stage of the current blade and comparing them with the historical acoustic signature segments of the corresponding stages of two other blades of the same wind turbine includes: Select one of the three blades of the wind turbine as the current blade to be tested. Based on the segment point corresponding to the blade to be tested in the corrected segment point set, extend the segment point to both sides for a second set time to obtain the current stable period segment. The second set duration is less than half of the measured rotation cycle; Based on the segment points corresponding to the other two blades in the corrected segment point set, retrieve the historical stable period segments of these two blades from the pre-constructed historical voiceprint database respectively. The historical stable period segment is the stable period segment extracted in the same way in the most recent successful inspection before this acquisition, and the difference between the cumulative number of rotation cycles corresponding to the historical stable period segment and the cumulative number of rotation cycles in this acquisition does not exceed the preset number of revolutions threshold. The historical voiceprint database maintains an independent queue for each blade. The queue length is dynamically configured based on local storage resources or set by the user. When the queue is full, an elimination strategy based on the signal-to-noise ratio of voiceprint segments and the stability of operating conditions is adopted to prioritize the retention of healthy samples with high confidence.
6. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 5, characterized in that: The method of determining the defect type based on the distribution pattern of the acoustic signature differences of the three blades includes: Calculate the root mean square difference between the current stable period segment and the first historical stable period segment, calculate the root mean square difference between the current stable period segment and the second historical stable period segment, and calculate the root mean square difference between two historical stable period segments. If the root mean square difference between the current stable period segment and the two historical stable period segments is greater than the root mean square difference between the two historical stable period segments, then the current leaf is determined to have a defect; otherwise, the current leaf is determined to be normal. When the current blade is determined to be normal, the root mean square difference between each blade is compared with the preset absolute defect threshold. If the root mean square difference of all three is greater than the absolute defect threshold, a multi-blade synchronous abnormal alarm is triggered, and the determination result of the current blade to be inspected is changed to a suspected defect. At the same time, it is recommended to use backup inspection methods for verification. The preset absolute defect threshold is determined by the root mean square difference statistical quantile of the acoustic waveform segments during the historical healthy period of the wind turbine.
7. The intelligent inspection method for wind turbine blades based on blade acoustic signature segmentation comparison and defect radial position mapping as described in claim 6, characterized in that: This also includes identifying the defect type when a defect is determined to exist: Calculate the root mean square difference between the current stable period segment and each historical stable period segment in each preset frequency band. For each frequency band, take the larger of the two root mean square differences as the difference value of the frequency band. Compare the differences between the frequency bands, determine the frequency band with the largest difference as the target frequency band, and output the defect type based on the radial position corresponding to the target frequency band: If the defect is in the low-frequency band, it is determined to be a defect near the leaf root; if it is in the mid-frequency band, it is determined to be a defect near the leaf middle; and if it is in the high-frequency band, it is determined to be a defect near the leaf tip.
8. A wind turbine blade intelligent inspection system based on blade acoustic signature segment comparison and defect radial position mapping, based on the wind turbine blade intelligent inspection method based on blade acoustic signature segment comparison and defect radial position mapping as described in any one of claims 1 to 7, characterized in that: Also includes: The dual-cycle independent segmentation module collects the blade acoustic signatures for two consecutive complete rotation cycles, and performs segmentation operations independently to obtain the first cycle segmentation point set and the second cycle segmentation point set. The period consistency verification module performs period consistency verification on the first and second period segment point sets. If the time interval between the two sets of segment points matches a rotation cycle, the second period segment point set is used as the reference segment point. The frequency band-radial position correction module divides the acoustic pattern segment corresponding to the reference segment point into multiple frequency band sets. Each frequency band set is associated with a different radial position of the blade. The reference segment point is corrected according to the radial position to obtain the corrected segment point. The three-blade asynchronous cross-calibration and defect mapping module extracts acoustic signature segments of each stage of the current blade based on the corrected segment points, and compares them with the historical corresponding acoustic signature segments of the other two blades of the same wind turbine. The defect type is determined based on the distribution pattern of the acoustic signature differences of the three blades.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent inspection method for wind turbine blades based on blade acoustic pattern segmentation comparison and defect radial position mapping as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent inspection method for wind turbine blades based on blade acoustic pattern segmentation comparison and defect radial position mapping as described in any one of claims 1 to 7.
