Method for detecting the external acoustic signature of a fan blade

Abstract

The application provides an external acoustic print detection method for a fan blade, relates to a data processing technology, converts an operation acoustic print signal into a time-frequency feature map, performs clustering analysis on the time-frequency feature map to obtain a clustering label sequence, processes the clustering label sequence to obtain an effective candidate segment, performs scale normalization and time domain padding processing on the effective candidate segment to generate a to-be-detected acoustic segment and decompose the to-be-detected acoustic segment into complementary sub-frequency bands, inputs the complementary sub-frequency bands into a deep feature extraction network to obtain an embedding vector of each complementary sub-frequency band and a confidence score of a known defect category, groups the embedding vectors according to a device operation cycle to obtain multiple detection groups, calculates the consistency of the intragroup features of the detection groups, if the intragroup consistency meets a preset condition, calculates the intergroup similarity of the detection groups to obtain an outlier group, compares the embedding vector of the outlier group with an abnormal class library to obtain a recognition result.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method for detecting the external acoustic signature of wind turbine blades. Background Technology

[0002] As a crucial component of clean energy, the safety and stability of wind turbine operation are paramount. Wind turbine blades, the core components for capturing wind energy, are constantly exposed to harsh natural environments and are highly susceptible to external damage such as lightning strikes, corrosion, and cracks. Acoustic signature monitoring technology, due to its non-contact and real-time advantages, is gradually becoming an important means of early warning for wind turbine blade failures. By collecting and analyzing the aerodynamic noise generated by blade rotation, structural anomalies on the blade surface can be detected in a timely manner, which is of great significance for reducing operation and maintenance costs and preventing catastrophic accidents.

[0003] However, existing acoustic signature detection technologies for wind turbine blades still face significant challenges in practical applications. Traditional monitoring methods often rely on statistical thresholds of sound pressure level or energy in specific frequency bands. However, in complex and variable wind environments, strong background noise and non-stationary wind speed changes often mask weak early fault signals, resulting in low detection sensitivity. Although subsequent deep learning-based supervised fault diagnosis methods have improved the recognition rate to some extent, they heavily depend on massive amounts of detailed fault samples for model training. In practice, wind turbine fault samples are scarce, and fault modes are diverse and difficult to cover. Furthermore, due to fluctuations in operating cycles and instantaneous gusts, the collected acoustic signature signals often suffer from temporal misalignment and inconsistent operating conditions, leading to poor model recognition capabilities under different wind fields or operating conditions, making it difficult to accurately distinguish between environmental interference and actual faults.

[0004] Therefore, how to extract effective acoustic segments in complex and dynamic noise environments, and how to achieve relatively stable identification and classification of abnormal voiceprints by utilizing the periodic consistency characteristics of the signal itself in the absence of a large number of fault samples, has become an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a method for detecting external acoustic signatures of wind turbine blades, which can effectively extract acoustic segments in complex and dynamic noise environments, and achieve relatively stable identification and classification of abnormal acoustic signatures by utilizing the periodic consistency characteristics of the signal itself in the absence of a large number of fault samples.

[0006] A first aspect of the present invention provides a method for detecting the external acoustic signature of wind turbine blades, comprising:

[0007] The voiceprint signal is converted into a time-frequency feature map, and cluster analysis is performed on the time-frequency feature map to obtain a cluster label sequence. The cluster label sequence is then processed to obtain valid candidate segments.

[0008] The effective candidate segments are scaled and time-domain padded to generate the acoustic segments to be tested and decomposed into complementary sub-bands. The complementary sub-bands are then input into a deep feature extraction network to obtain the embedding vectors of each complementary sub-band and the confidence scores of the known defect categories.

[0009] The embedded vectors are grouped according to the device operating cycle to obtain multiple detection groups. The consistency of the features within the detection group is calculated. If the consistency within the group meets the preset conditions, the inter-group similarity of the detection group is calculated to obtain the outlier group. The embedded vectors of the outlier group are compared with the anomaly class library to obtain the recognition result.

[0010] Optionally, in one possible implementation of the first aspect, processing the clustering label sequence to obtain valid candidate fragments includes:

[0011] Identify bright regions in the cluster label sequence and lock the original label positions of the bright regions as a protection set;

[0012] The clustering label sequence is smoothed to generate a smoothed label sequence;

[0013] The original tag positions in the protection set are backfilled into the corresponding positions in the smoothed tag sequence to obtain the integrated tag sequence;

[0014] The integrated tag sequence is traversed, and segments with a duration lower than a preset threshold are removed to obtain valid candidate segments.

[0015] Optionally, in one possible implementation of the first aspect, the step of performing scale normalization and temporal padding on the valid candidate segments to generate the acoustic segment to be tested includes:

[0016] The time length of the valid candidate segment is obtained. If the time length is greater than the preset standard length, the segment is truncated based on the center. If the time length is less than the preset standard length, the segment is padded at the boundary to obtain a normalized segment.

[0017] When the time center interval between adjacent normalized segments is greater than the preset threshold of the device operating cycle, a time positioning point is determined between the two adjacent normalized segments with the device operating cycle as the step size, and supplementary data is generated at the time positioning point.

[0018] The normalized segment and the padded data are combined in time sequence to generate the acoustic segment to be tested.

[0019] Optionally, in one possible implementation of the first aspect, calculating the consistency of the intra-group features of the detection group, and if the intra-group consistency meets a preset condition, calculating the inter-group similarity of the detection group to obtain the outlier group, includes:

[0020] The mean cosine similarity among all embedded vectors within the same detection group is used as the intra-group consistency index.

[0021] If the consistency index within the group is lower than the preset consistency threshold, the corresponding detection group is deleted.

[0022] If the consistency index within the group meets the consistency threshold, then the corresponding detection group is used as the comparison group, and the comparison group is statistically analyzed to obtain a candidate set.

[0023] Obtain the centroids of all comparison groups in the candidate set, and calculate the distance between the mean of the intra-group features of each comparison group and the centroid as the inter-group separation degree.

[0024] When the inter-group separation is determined to be greater than a preset outlier threshold, the corresponding comparison group is designated as the outlier group.

[0025] Optionally, in one possible implementation of the first aspect, it also includes:

[0026] When the identification result is determined to be a damage result, the corresponding wind turbine blade is regarded as the damaged blade, and a damage model corresponding to the damaged blade is constructed.

[0027] The damaged blades are customized based on the damaged area in the damage model and the limited length of the carrier vehicle.

[0028] Optionally, in one possible implementation of the first aspect, the customized processing of the damaged blade based on the damaged area in the damage model and the defined length of the transport vehicle includes:

[0029] When the damage region in the damage model is determined to be a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut. The blade to be cut is then cut based on a limited length to obtain segmented sections.

[0030] When the damage region in the damage model is determined to be an axial crack, the cutting position corresponding to the axial crack is determined based on a limited length, and the damage model is cut according to the cutting position to obtain segmented segments.

[0031] When the damage region in the damage model is determined to be a blocky crack, the cutting position corresponding to the blocky crack is determined based on a limited length, and the damage model is cut according to the cutting position to obtain segmented segments.

[0032] Optionally, in one possible implementation of the first aspect, when determining that the damage region in the damage model is a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut, including:

[0033] When the damage region in the damage model is determined to be a circumferential crack, the circumferential crack length is obtained.

[0034] When the length of the circumferential crack is less than or equal to a preset fixed length, the center position of the circumferential crack is obtained as the fixed position;

[0035] When the length of the circumferential crack is greater than the preset fixed length, a wrapping region corresponding to the circumferential crack is constructed, and the number of divided regions is obtained by rounding up the ratio of the circumferential crack length to the preset fixed length.

[0036] The package area is evenly divided based on the number of division areas to obtain multiple fixed areas. Each fixed area is then divided into three equal parts to obtain an upper area, a middle area, and a lower area for each fixed area.

[0037] The region with the most crack pixels in the upper, middle, and lower regions of each fixed region is selected as the fixed position of each fixed region;

[0038] The circumferential crack is fixed at the fixed position to obtain the blade to be cut.

[0039] Optionally, in one possible implementation of the first aspect, when determining that the damage region in the damage model is an axial crack, determining the cutting position corresponding to the axial crack based on a limited length includes:

[0040] When the damage region in the damage model is determined to be an axial crack, the axial crack length of the axial crack, the first endpoint of the axial crack that is closest to the blade root, and the second endpoint that is closest to the blade tip are obtained.

[0041] When the axial crack length is less than or equal to the predetermined length, the remaining length is obtained based on the difference between the predetermined length and the axial crack length. The remaining length is then halved to obtain the extension length.

[0042] Starting from the first endpoint and the second endpoint respectively, the extension is extended to both sides based on the extension length to obtain the cutting position corresponding to the axial crack;

[0043] When the axial crack length is greater than the limit length, the number of cut segments is obtained by rounding up the ratio of the axial crack length to the limit length.

[0044] The selection area for axial cracks is determined based on the number of cut segments;

[0045] The location of the smallest crack width in the selected region of the axial crack is taken as the cutting position of the axial crack.

[0046] Optionally, in one possible implementation of the first aspect, the step of extending the line outwards from the first endpoint and the second endpoint respectively, based on the extension length, to obtain the cutting position corresponding to the axial crack, includes:

[0047] Based on the aforementioned extension length, starting from the first endpoint, the line is extended along a circumferential direction pointing towards the blade root and perpendicular to the wind turbine blade to obtain the first line segment;

[0048] Based on the aforementioned extension length, starting from the second endpoint, the second line segment is obtained by extending along the direction pointing towards the blade tip and perpendicular to the circumferential direction of the fan.

[0049] Obtain the remaining construction endpoints in the first and second line segments, and construct circumferential cutting lines perpendicular to the first and second line segments respectively as the cutting positions of the axial cracks based on the construction endpoints;

[0050] The process of determining the selectable region for axial cracks based on the number of cut segments includes:

[0051] Based on the ratio of the axial crack length to the number of cut segments, the average length is obtained. Taking the first or second endpoint as the starting point, the center point is determined sequentially in the axial crack based on the average length.

[0052] The safe length is obtained by taking half of the difference between the limited length and the average length, and a defined perpendicular line is constructed that passes through the center point and is perpendicular to the circumferential direction of the wind turbine blade.

[0053] Starting from the center point, determine the division points along the defined vertical line and in both directions away from the center point based on the safe length;

[0054] Using the defined dividing points as a reference, construct circumferential dividing lines perpendicular to the defined vertical lines, and obtain the area between the circumferential dividing lines as the area to be selected.

[0055] Optionally, in one possible implementation of the first aspect, when determining that the damage region in the damage model is a blocky crack, determining the cutting position corresponding to the blocky crack based on a defined length includes:

[0056] When the damage region in the damage model is determined to be a blocky crack, the maximum axial distance of the blocky crack in the axial direction of the wind turbine blade is obtained.

[0057] When the maximum axial distance is less than or equal to the defined length, the block extension distance is obtained by taking half of the difference between the defined length and the maximum axial distance.

[0058] Starting from the two extreme points of the blocky crack in the axial direction of the wind turbine blade, the cutting position corresponding to the blocky crack is obtained by extending the blocky extension distance to both sides.

[0059] When the maximum axial distance is greater than the limited length, the number of cutting blocks is obtained by rounding up the ratio of the maximum axial distance to the limited length.

[0060] The selection area for block-shaped cracks is determined based on the number of cutting blocks;

[0061] Obtain the fragment area of ​​the blocky fragment in the selected area of ​​the blocky crack, and select the center point of the blocky fragment corresponding to the largest fragment area as the blocky cutting point;

[0062] Using the block-shaped cutting points as a reference, a circumferential dividing line of the block-shaped crack is constructed as the cutting position of the block-shaped crack.

[0063] A second aspect of the present invention provides an electronic device comprising: a memory, a processor, and a computer program, the computer program being stored in the memory, and the processor executing the computer program to perform the methods described in the first aspect of the present invention and various possible methods related to the first aspect.

[0064] A third aspect of the present invention provides a storage medium storing a computer program, which, when executed by a processor, is used to implement the first aspect of the present invention and various methods possibly involved in the first aspect.

[0065] The beneficial effects of this invention are as follows:

[0066] 1. This invention converts acoustic signature signals into time-frequency feature maps and performs cluster analysis, enabling adaptive learning of background noise patterns in the current environment. This allows for automatic segmentation of effective candidate segments even in complex and variable wind conditions, avoiding the problem of fixed thresholds failing to adapt to environmental changes. Simultaneously, this invention employs complementary sub-band decomposition, extracting features from both low-frequency structural and high-frequency surface perspectives. This prevents high-energy low-frequency wind noise from masking weak high-frequency fault features, covering damage modes with different physical mechanisms. By calculating intra-group consistency and inter-group similarity of detection groups, it effectively distinguishes between real fault signals with periodic patterns and random environmental noise, significantly reducing the false alarm rate and achieving good identification of abnormal acoustic signatures even in the absence of massive fault samples.

[0067] 2. This invention addresses the problem of deformed and easily broken old blades in complex environments such as mountainous wind farms, which prevents the use of standard equipment for overall disassembly and transportation. After identifying the damage, a damage model is constructed and customized processing is performed. By targeting the blades based on the damaged area and the limited length of the transport vehicle, the risk of blade slippage or breakage caused by forced transportation is avoided, ensuring operational safety. Attached Figure Description

[0068] Figure 1 A flowchart illustrating the external acoustic signature detection method for wind turbine blades provided by this invention;

[0069] Figure 2 A schematic diagram showing the central position provided by the present invention;

[0070] Figure 3 This is a schematic diagram of the packaging area provided by the present invention;

[0071] Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided by the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0074] It should be understood that in the various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0075] It should be understood that in this invention, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0076] It should be understood that in this invention, "multiple" refers to two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "Contains A, B, and C", "Contains A, B, and C" means that all three A, B, and C are contained; "Contains A, B, or C" means that one of A, B, and C is contained; "Contains A, B, and / or C" means that any one, two, or three of A, B, and C are contained.

[0077] It should be understood that in this invention, "B corresponding to A", "B corresponding to A", "A and B correspond", or "B and A correspond" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Matching A and B is defined as a similarity between A and B that is greater than or equal to a preset threshold.

[0078] Depending on the context, "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection."

[0079] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0080] This invention provides a method for detecting the external acoustic signature of wind turbine blades, such as... Figure 1 As shown, steps S1-S3 are included:

[0081] S1, convert the running voiceprint signal into a time-frequency feature map, perform cluster analysis on the time-frequency feature map to obtain a cluster label sequence, process the cluster label sequence to obtain effective candidate segments.

[0082] It's important to note that wind turbines operate in extremely harsh environments, and the raw audio signals collected are often mixed with significant background noise, such as continuous wind noise, distant thunder, and birdsong. Current technologies often directly threshold the raw waveform, which struggles to distinguish between noise and the sound of blades sweeping across the surface, leading to inaccurate extraction of effective segments. Therefore, the server first converts the acoustic signature of the wind turbine into a readily identifiable time-frequency feature map, using a graphical representation of energy distribution to characterize the sound. More importantly, given the highly variable nature of wind farm environments, pre-setting fixed thresholds is impractical. Through clustering, the system adaptively learns the background noise patterns of the current environment, automatically segmenting continuous sound signals into multiple segments for easier extraction of the actual blade sound.

[0083] Among them, the operating acoustic signal refers to the raw continuous audio data collected by the sound pickup device during wind turbine operation. For example, it could be a sound sensing unit deployed at the bottom of the tower to collect acoustic signals during operation. The time-frequency characteristic map refers to a two-dimensional spectrum obtained through short-time Fourier transform or Mel-frequency cepstral coefficient transform, where the horizontal axis represents time, the vertical axis represents frequency, and the pixel brightness represents energy intensity.

[0084] Understandably, the system first performs frame-by-frame windowing processing on the acquired acoustic signature signal and transforms it to the frequency domain, generating a time-frequency feature map that reflects energy changes over time. Then, the system treats each column of this map, i.e., the spectrum at each time point, as a sample point and inputs it into a clustering algorithm. The algorithm automatically divides these sample points into different clusters based on the similarity of energy distribution. Finally, the system outputs the cluster ID of each time point in chronological order, forming a continuous sequence of numbers, i.e., the clustering label sequence. The clustering analysis used here is existing technology and will not be elaborated upon here; for example, it could be done using K-Means clustering.

[0085] In some embodiments, step S1 (processing the clustering label sequence to obtain valid candidate fragments) includes:

[0086] S11, Identify bright regions in the cluster label sequence and lock the original label positions of the bright regions as a protection set.

[0087] It should be noted that early fault signals of wind turbine blades, such as crack opening sounds and bolt loosening impact sounds, often have the characteristics of short duration and high energy. That is, the duration is extremely short, but the energy burst in the spectrum is obvious. When processing cluster labels, existing technologies often filter indiscriminately in order to obtain a smooth waveform. This approach is very likely to misclassify these short fault pulses as noise and smooth them out, leading to missed detections. Therefore, the system identifies key areas based on energy characteristics. These high-energy areas, i.e., bright areas, usually contain the most significant tower sweeping sounds or abnormal impact sounds. By locking their coordinate positions in advance for protection, that is, analyzing energy characteristics to identify key signal segments with physical energy significantly higher than the background, and locking them into the protection set, i.e., bright areas.

[0088] In this context, "bright" regions can be time segments that are marked as bright by the clustering algorithm and whose actual energy values ​​exceed the dynamic background threshold. The dynamic background threshold is obtained by extracting the average and standard deviation of the energy data of all time frames marked as dark (i.e., background noise) from the time-frequency feature map. This is existing technology and will not be elaborated here. For example, 1 represents the bright class and 0 represents the dark class. The dynamic background threshold is calculated based on the average and standard deviation of the dark class, and is generally the sum of the average and standard deviation by a fixed multiple. The original label position refers to the time index of the bright region in the clustering label sequence, which can be understood as the time segment position corresponding to 1.

[0089] It is easy to understand that the server will select the bright positions in the cluster label sequence corresponding to the voiceprint for protection. Here, protection means backing up the signal at that position, which will be restored to the smoothed sequence later.

[0090] S12, the clustering label sequence is smoothed to generate a smoothed label sequence.

[0091] It is understandable that isolated noise points can be removed using a sliding window majority voting algorithm to obtain a smooth label sequence, which is the existing smoothing method.

[0092] S13, the original tag positions in the protection set are backfilled into the corresponding positions in the smoothed tag sequence to obtain the integrated tag sequence.

[0093] It should be noted that while simple smoothing removes noise, it often also smooths out extremely important short-term fault signals to zero. Therefore, the system will backfill the original tag positions in the protection set into the corresponding positions in the smoothed tag sequence to obtain the integrated tag sequence.

[0094] S14, traverse the integrated tag sequence, remove segments with a duration lower than a preset threshold, and obtain valid candidate segments.

[0095] It should be noted that after the above processing, there may still be a very small number of fragments that are neither bright nor very short in the sequence. Blade sweeping sound or meaningful fault sound has a certain duration, and very short fragments are meaningless in the frequency domain.

[0096] Therefore, the server will traverse and integrate the tag sequence, remove segments whose duration is less than a preset threshold, and obtain valid candidate segments.

[0097] The preset threshold can be the minimum effective time window set by the fan speed and sampling rate, or it can be a threshold set manually based on the actual situation.

[0098] S2, scale normalization and temporal padding are performed on the effective candidate segments to generate the acoustic segments to be tested and decompose them into complementary sub-bands. The complementary sub-bands are input into the deep feature extraction network to obtain the embedding vectors of each complementary sub-band and the confidence scores of the known defect categories.

[0099] It should be noted that the wind turbine rotation speed changes in real time with the wind speed, resulting in varying durations of the sound period when the blades sweep across the tower. Directly inputting audio signals of varying lengths into the neural network would lead to input dimension mismatch or feature misalignment. Therefore, scale normalization and time-domain completion are necessary to eliminate time axis differences caused by rotation speed fluctuations and ensure that all samples are aligned in the time dimension. Secondly, there are frequency domain differences in fault features, meaning that different types of blade damage have completely different physical acoustic manifestations. For example, internal structural cavities typically induce low-frequency resonance, while surface cracks or corrosion produce high-frequency howling. If full-band signals are mixed and input, high-energy low-frequency wind noise often masks weak high-frequency fault features. Therefore, this invention employs a complementary sub-band decomposition strategy, which allows the model to extract features from two independent channels: a low-frequency structural perspective and a high-frequency surface perspective, thereby comprehensively covering damage modes with different physical mechanisms.

[0100] Among them, complementary sub-bands refer to the division of a full-band signal into multiple complementary frequency domain intervals based on the physical sound generation mechanism, such as low frequency and high frequency.

[0101] Understandably, the processing flow for this step is as follows: First, the system receives valid candidate segments and scales or pads them on the time axis according to the preset standard input length to generate a uniformly formatted acoustic segment to be tested. Second, the system uses a filter bank to decompose the segment in the frequency domain, generating at least two sets of complementary sub-frequency bands, such as a low-frequency spectrum and a high-frequency spectrum. Finally, these sub-frequency bands are input into a pre-trained deep feature extraction network, which performs forward propagation calculations. For each sub-frequency band, two types of results are output in parallel: an embedding vector representing acoustic texture features for subsequent outlier analysis and a confidence score representing the specific fault probability for direct alarm.

[0102] It's worth noting that the system first collects a massive amount of unlabeled normal samples. Instead of directly labeling them as normal, it uses unsupervised clustering algorithms, such as K-Means, to automatically subdivide these normal samples into N different subclasses based on differences in acoustic texture, assigning each subclass a pseudo-label from 1 to N. These N pseudo-labeled normal samples are then merged with a small number of k known anomalous samples with true labels, such as leading-edge erosion and cracks, to jointly train a lightweight CNN network. By forcing the network to distinguish these N and k subtle differences, it learns an extremely powerful feature embedding space, becoming a model capable of recognizing fine acoustic textures. This reduces the reliance on rare anomalous samples, enabling both the determination of known fault classes and the extraction of feature vectors from voiceprint data—essentially a deep feature extraction network.

[0103] In some embodiments, step S2 (performing scale normalization and temporal completion on the valid candidate segments to generate the acoustic segment to be tested) includes:

[0104] S21, obtain the time length of the valid candidate segment. If the time length is greater than the preset standard length, then the segment is truncated based on the center. If the time length is less than the preset standard length, then the segment is padded at the boundary to obtain a normalized segment.

[0105] It should be noted that in actual operation, the rotor speed of a wind turbine fluctuates in real time due to changes in wind speed. This results in varying durations of the sweeping sound produced by the blades sweeping across the tower. Current technologies often employ resampling or linear stretching to unify the data dimensions input to deep learning models. While this unifies the length, it alters the original frequency characteristics of the sound, thus disrupting the acoustic texture of subtle faults such as cracks. Since the energy peak of the blade sweeping sound is typically located at the center of the segment, for long segments, the central core area is preserved while edge redundancy is removed; for short segments, the original data is retained, with only edge padding. This approach unifies the data dimensions while better ensuring that the frequency and phase characteristics of the original sound wave are not distorted.

[0106] The preset standard length refers to the fixed time dimension of the input layer of the deep feature extraction network, which can be preset.

[0107] Understandably, the system first reads the number of sampling points for each valid candidate segment, compares it with the standard length, and if the time length is greater than the preset standard length, the system calculates the length to be cut off. Taking the center of the segment as the origin, the system retains the data to the left and right respectively, which is the data of the preset standard length, and cuts off the excess parts on both sides.

[0108] If the time length is less than the preset standard length, the system calculates the length that needs to be supplemented, generates a zero sequence or noise sequence with the length of the supplement, and splices them evenly at the beginning and end of the segment.

[0109] S22, when the time center interval between adjacent normalized segments is greater than the preset threshold of the device operating cycle, a time positioning point is determined between two adjacent normalized segments with the device operating cycle as the step size, and supplementary data is generated at the time positioning point.

[0110] It should be noted that the signal extraction process may occasionally miss a sweeping event due to gusts of wind or excessive distance from the sound source. For example, one of the three blades might be missing during sweeping. Such omissions are unrecognizable for subsequent time-series analysis. Directly connecting discontinuous segments would disrupt the periodicity of the wind turbine's operation, misleading the model into believing a sudden change in turbine speed. Therefore, this solution utilizes the physical inertia of the wind turbine's operation to repair the data. The time difference between segments is checked. If an excessively large interval is found, such as twice the cycle length, indicating a missed event, the system automatically inserts a dummy data point at the time when the sound should have been recorded, based on the current speed cycle. This ensures the temporal integrity of the input data, satisfying the requirement of three blade sweeping sounds and generating corresponding data at that specific time point.

[0111] Among them, the time center interval refers to the difference between the center time of the previous segment and the center time of the next segment, the equipment operation cycle refers to the theoretical tower sweeping interval calculated based on the current wind turbine speed, and the supplementary data refers to the generated virtual segments, so that the sound of the three blade sweepings is more complete and it is also convenient for subsequent grouping.

[0112] It is easy to understand that, using the equipment operation cycle as the step size, the theoretical center position of the missing segment, i.e. the time positioning point, is calculated. The system generates a standard supplementary data at this time positioning point and inserts it into the sequence to fill the time gap.

[0113] S23, combine the normalized segment and the padded data in time sequence to generate the acoustic segment to be tested.

[0114] Understandably, the normalized segments and the imputed data are arranged according to their respective physical time order. That is, the system splices these segments in the channel dimension or the time dimension to obtain a set of acoustic segments to be tested. This set of acoustic segments to be tested contains real acoustic textures and maintains a strict periodic structure, which meets the input requirements of deep feature extraction networks.

[0115] S3, group the embedded vectors according to the device operation cycle to obtain multiple detection groups, calculate the consistency of features within the detection group, if the consistency within the group meets the preset conditions, calculate the inter-group similarity of the detection group to obtain an outlier group, compare the embedded vector of the outlier group with the anomaly library to obtain the recognition result.

[0116] It should be noted that real blade damage, such as cracks, will produce an impact sound every time the impeller sweeps across the tower as it rotates. This means that the abnormal signal should be relatively stable over multiple consecutive rotation cycles. In contrast, environmental noise is usually random and discrete and cannot remain stable over multiple cycles.

[0117] Among them, the detection group refers to a set of embedded vectors aggregated according to the time window. For example, 10 consecutive scanning cycles are divided into 1 group, and the number of groups can be preset according to the actual situation.

[0118] Therefore, this scheme first performs intra-group consistency verification to check if the signal is stable. If it is unstable, for example, intermittent, it is directly judged as environmental noise and removed, thus greatly reducing the false alarm rate. If multiple sets of sounds within a group are inconsistent, it indicates that there is a lot of environmental noise, and the group is invalid. If the consistency threshold is met, the subsequent inter-group similarity verification is performed, and each group is compared to filter out the groups with larger deviations.

[0119] The anomaly database is a pre-built database of anomaly data. High-dimensional feature embedding vectors of identified outliers are used to retrieve historical outliers and compared with candidate vectors in the same frequency domain within the database. For new outliers not previously observed, their feature embedding vectors are added to the database to create new anomaly categories, which are then continuously evaluated. If the similarity to the feature embedding vector of a certain category of historical outliers is sufficiently high, the matching count for that category is accumulated and the time window is updated. Subsequently, when a certain category of feature vectors continuously appears more than a preset number of times and has a similarity greater than a preset value during continuous operation, a corresponding anomaly event is generated and an alarm is recorded. This is the existing comparison method and will not be elaborated upon further.

[0120] In some embodiments, step S3 (calculating the consistency of intra-group features of the detection group; if the intra-group consistency meets a preset condition, calculating the inter-group similarity of the detection group to obtain the outlier group) includes:

[0121] S31, the mean cosine similarity among all embedded vectors within the same detection group is used as the intra-group consistency index.

[0122] S32, if the consistency index within the group is lower than the preset consistency threshold, then the corresponding detection group is deleted.

[0123] S33, if the consistency index within the group meets the consistency threshold, then the corresponding detection group is used as the comparison group, and the comparison group is statistically analyzed to obtain the candidate set.

[0124] It should be noted that wind farm environmental noise is highly random and transient. For example, a sudden strong gust of wind, a bird's call, or a distant clap of thunder can all generate high-energy signals in a short period of time. Existing technologies, relying solely on energy thresholds or single-shot feature matching, are prone to misjudging these sporadic environmental noises as equipment malfunctions. This invention utilizes the physical law of the periodic recurrence of mechanical faults. Real blade damage, such as cracks or loose bolts, will inevitably repeat in each sweeping cycle as the impeller rotates, resulting in relatively stable feature vectors over time. Therefore, this solution incorporates an intra-group consistency screening mechanism. The system checks whether a detection group, for example, containing 10 consecutive cycles of sound, is internally consistent. If the feature vector consistency of these 10 cycles is low, it indicates random noise and is directly deleted; only when the features of these 10 cycles are highly similar are they considered valid sound sources, i.e., stable blade sweeping sound, and retained as a comparison group. The consistency threshold can be a threshold calculated using cosine similarity from historical data. For example, a large number of similarity values ​​can be calculated, and the threshold can be determined by statistically analyzing the probability distribution of these values. Alternatively, it can be preset by humans based on the actual situation.

[0125] S34, obtain the centroids of all comparison groups in the candidate set, and calculate the distance between the mean of the intra-group features and the centroid of each comparison group as the inter-group separation degree.

[0126] S35, when it is determined that the inter-group separation degree is greater than the preset outlier threshold, the corresponding comparison group is regarded as the outlier group.

[0127] It is easy to understand that the previous steps have screened out relatively stable sound sources within each group, with both faulty and non-faulty sounds being relatively stable. Therefore, it is necessary to compare the data between the groups and select the group that deviates significantly from the other data, that is, the group that deviates significantly from the overall structural pattern, as the outlier group.

[0128] If a comparison group, such as group A, is internally stable but its characteristic mean is very far from the global centroid, indicating a large inter-group separation, it suggests a rare stable pattern. In wind turbine operation and maintenance logic, rare and stable patterns usually indicate a fault; conversely, if the mean is close to the centroid, it indicates normal normal operation.

[0129] The centroid refers to the arithmetic mean of all feature vectors of the comparison groups in the candidate set, and the within-group feature mean of the comparison group refers to the arithmetic mean of all embedded vectors contained in the comparison group. The distance between each comparison group and the centroid is calculated. When the distance is greater than the preset outlier threshold, the corresponding comparison group is regarded as the outlier group. The preset outlier threshold can be a threshold calculated from a large amount of historical data or a threshold set by humans in advance based on the actual situation.

[0130] Based on the above embodiments, A1-A2 are also included:

[0131] A1, when the identification result is determined to be a damage result, the corresponding wind turbine blade is taken as the damaged blade, and a damage model corresponding to the damaged blade is constructed.

[0132] It should be noted that in complex environments such as mountain wind farms, damaged blades have often been in operation for over a decade, and their root flange surfaces may have deformed. Furthermore, the damaged areas of the blades are prone to breakage. Current technology typically relies on specialized lifting vehicles for complete disassembly and transportation, but this requires precise alignment between the lifting vehicle's adapter and the blade's root holes. For damaged, old blades, it is impossible to securely lock them onto the lifting vehicle; forced transportation can easily lead to blade slippage or breakage.

[0133] Among them, the damage result is surface damage to the wind turbine blades.

[0134] A2, based on the damaged area in the damage model and the limited length of the transport vehicle, the damaged blade is customized.

[0135] In some embodiments, step A2 (customizing the damaged blade based on the damaged area in the damage model and the limited length of the carrier vehicle) includes A21-A23:

[0136] A21, when the damage region in the damage model is determined to be a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut, and the blade to be cut is cut based on a limited length to obtain the segment.

[0137] It should be noted that circumferential cracks, also known as transverse through-cracks, are a relatively dangerous form of structural damage to wind turbine blades. Under the influence of gravity, the blade is highly susceptible to fracture along the circumferential crack. During on-site handling, if blades with circumferential cracks are directly hoisted or cut, vibration and stress changes can easily cause the crack to propagate instantly, resulting in the blade disintegrating in mid-air or being cut off, leading to a serious safety accident.

[0138] Therefore, the fixed position at the circumferential crack is first fixed to obtain the blade to be cut. Then, the circumferential crack that is being reinforced in the blade to be cut is extended to the left and right sides, that is, in the axial direction, until the length is equal to the fixed length, and then cut, so as to avoid secondary damage to the damaged part. The remaining part can be cut freely with a fixed length. Existing cutting methods can be used here as long as the circumferential crack is avoided, which will not be elaborated here.

[0139] In some embodiments, step A21 (when the damage region in the damage model is determined to be a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut, and the blade to be cut is cut based on a limited length to obtain a segment) includes A211-A216:

[0140] A211, when the damage region in the damage model is determined to be a circumferential crack, the circumferential crack length of the circumferential crack is obtained.

[0141] The circumferential crack length is the length of the crack extending along the circumference of the blade cross-section. It can be the length identified by existing technologies, such as extracting the point cloud set corresponding to the crack through a point cloud segmentation algorithm to obtain the length of the circumferential crack, or it can be obtained by connecting the pixels at both ends of the main crack along the main crack. This will not be elaborated here.

[0142] A212, when the length of the circumferential crack is less than or equal to a preset fixed length, the center position of the circumferential crack is obtained as the fixed position.

[0143] It should be noted that short circumferential cracks have not yet damaged the overall keel structure of the blade. If the crack length is within a safe range, complex fixing is not required; simply fixing at the geometric center of the crack is sufficient to prevent it from propagating during subsequent cutting vibrations.

[0144] Therefore, when the length of the circumferential crack is less than or equal to a preset fixed length, the center position of the circumferential crack is taken as the fixed position. For example, if the crack length is 10cm, the center position of the circumferential crack can be taken as 5cm from any end of the crack. See [reference needed]. Figure 2 The center position of the crack can be determined directly using existing technology, and this position can be used as a fixed position.

[0145] A213, when the length of the circumferential crack is greater than the preset fixed length, a wrapping region corresponding to the circumferential crack is constructed, and the number of divided regions is obtained by rounding up the ratio of the circumferential crack length to the preset fixed length.

[0146] It should be noted that when the circumferential crack is too long, for example, when it extends through more than half of the blade's circumference, the blade may break at any time. In this case, single-point fixation is no longer sufficient to withstand the enormous tension, so multi-point distributed reinforcement is adopted.

[0147] Therefore, when the length of the circumferential crack is greater than a preset fixed length, a wrapping region corresponding to the circumferential crack is constructed, see [reference]. Figure 3 That is, by connecting the protruding parts around the crack with line segments to form a region that surrounds the crack, the number of regions is obtained by rounding up the ratio of the circumferential crack length to a preset fixed length.

[0148] The preset fixed length can be a length that is set in advance by humans. If it is greater than this length, there is a risk. Subsequently, the fixed position is determined within the area corresponding to the corresponding length based on the preset fixed length.

[0149] A214, the package area is evenly divided based on the number of division areas to obtain multiple fixed areas, and the fixed areas are divided into three equal parts to obtain the upper area, middle area and lower area of ​​each fixed area.

[0150] It is easy to understand that this solution will evenly divide the packaged area according to the number of division areas to obtain multiple fixed areas. However, the distribution of cracks within the fixed areas is uneven. Therefore, this solution will divide the fixed areas into three equal parts to obtain the upper, middle and lower areas of each fixed area. This division can be done evenly according to the crack length of the transverse crack.

[0151] A215, select the region with the most crack pixels in the upper, middle and lower regions of each fixed region as the fixed position of each fixed region.

[0152] It is easy to understand that a device would be installed in the area with more cracks to fix the area.

[0153] A216, the circumferential crack is fixed based on the fixed position to obtain the blade to be cut.

[0154] A22, when the damage area in the damage model is determined to be an axial crack, the cutting position corresponding to the axial crack is determined based on the limited length, and the damage model is cut according to the cutting position to obtain segmented segments.

[0155] It's easy to understand that when cutting, you should try to avoid cracks. If you can't avoid them, you need to choose a location with a smaller crack width to cut. After cutting off the cracked part, you get a segment. The normal part can then be freely cut according to the specified length.

[0156] In some embodiments, step A22 (determining the cutting position corresponding to the axial crack based on a limited length when determining that the damage region in the damage model is an axial crack) includes A221-A226:

[0157] A221, when the damage region in the damage model is determined to be an axial crack, the axial crack length of the axial crack, the first endpoint of the axial crack that is closest to the blade root, and the second endpoint of the axial crack that is closest to the blade tip are obtained.

[0158] It is easy to understand that the first and second endpoints are the two ends of the axial crack. The axial crack is the crack from the blade root to the blade tip, while the circumferential crack is the crack that surrounds the wind turbine blade.

[0159] A222, when the axial crack length is less than or equal to the predetermined length, the remaining length is obtained based on the difference between the predetermined length and the axial crack length, and the remaining length is halved to obtain the extension length.

[0160] A223, taking the first endpoint and the second endpoint as starting points respectively, extends to both sides based on the extension length to obtain the cutting position corresponding to the axial crack.

[0161] It is easy to understand that by extending the length of the first and second endpoints to both sides, the cutting position corresponding to the axial crack is obtained.

[0162] In some embodiments, step A223 (extending the extension length to both sides based on the first endpoint and the second endpoint respectively to obtain the cutting position corresponding to the axial crack) includes:

[0163] Based on the aforementioned extension length, starting from the first endpoint, the line is extended along a circumferential direction pointing towards the blade root and perpendicular to the wind turbine blade to obtain the first line segment.

[0164] Based on the aforementioned extension length, starting from the second endpoint, the second line segment is obtained by extending along the direction pointing towards the blade tip and perpendicular to the circumferential direction of the fan.

[0165] Obtain the remaining construction endpoints in the first and second line segments, and construct circumferential cutting lines perpendicular to the first and second line segments respectively as the cutting positions of the axial cracks based on the construction endpoints.

[0166] It is easy to understand that the endpoints other than the first and second endpoints of the first and second line segments are obtained as construction endpoints, and circumferential cutting lines perpendicular to the first and second line segments are constructed based on the construction endpoints as the cutting positions of the axial cracks.

[0167] A224. When the axial crack length is greater than the limit length, the number of cut segments is obtained by rounding up the ratio of the axial crack length to the limit length.

[0168] A225, the area to be selected for axial cracks is determined based on the number of cut segments.

[0169] In some embodiments, step A225 (determining the selectable region of the axial crack based on the number of cut segments) includes:

[0170] Based on the ratio of the axial crack length to the number of cut segments, the average length is obtained. Taking the first or second endpoint as the starting point, the center point is determined sequentially in the axial crack based on the average length.

[0171] It is not difficult to understand that determining multiple points as center points at the crack by taking the equal length and the first or second endpoint as the starting point is actually determining the points sequentially based on the equal length.

[0172] The safe length is obtained by taking half of the difference between the defined length and the average length, and then constructing a defined perpendicular line that passes through the center point and is perpendicular to the circumferential direction of the wind turbine blade.

[0173] Starting from the center point, determine the division points along the defined vertical line and in both directions away from the center point based on the safe length.

[0174] Using the defined dividing points as a reference, construct circumferential dividing lines perpendicular to the defined vertical lines, and obtain the area between the circumferential dividing lines as the area to be selected.

[0175] A226, the location of the smallest crack width in the selected area of ​​the axial crack is obtained as the cutting position of the axial crack.

[0176] It is easy to understand that the location of the smallest crack width in the area to be selected is taken as the cutting position of the axial crack, that is, the smallest crack is selected in the area where cutting is allowed.

[0177] It is worth mentioning that common cracks may have both axial and circumferential directions, meaning the cracks are oblique. Therefore, a combination of circumferential and axial methods can be used: first fix the crack, then select the smaller crack for cutting. This will not be elaborated on here; simply use the appropriate methods for circumferential and axial directions.

[0178] A23, when the damage area in the damage model is determined to be a blocky crack, the cutting position corresponding to the blocky crack is determined based on the limited length, and the damage model is cut according to the cutting position to obtain segmented segments.

[0179] In some embodiments, step A23 (determining the cutting position corresponding to the block crack based on a limited length when the damage region in the damage model is determined to be a block crack) includes A231-A237:

[0180] A231, when the damage region in the damage model is determined to be a blocky crack, the maximum axial distance of the blocky crack in the axial direction of the wind turbine blade is obtained.

[0181] Among them, blocky cracks are cracks in blocky areas, which can be caused by the impact of foreign objects on the surface of the wind turbine blades, resulting in regional shattering. The maximum axial distance can be obtained directly by existing technology or by calculating the distance between the extreme points on both sides.

[0182] A232, when the maximum axial distance is less than or equal to the limited length, the block extension distance is obtained by taking half of the difference between the limited length and the maximum axial distance.

[0183] A233, taking the two extreme points of the blocky crack in the axial direction of the wind turbine blade as the starting point, and extending to both sides based on the blocky extension distance, the cutting position corresponding to the blocky crack is obtained.

[0184] It is not difficult to understand that the principle here is similar to that of steps A222-A223. The subsequent cutting can be extended to both sides to avoid damaging the crack.

[0185] A234, when the maximum axial distance is greater than the limited length, the number of cutting blocks is obtained by rounding up the ratio of the maximum axial distance to the limited length.

[0186] A235, determine the selectable area of ​​the block crack based on the number of cutting blocks.

[0187] It is easy to understand that this is consistent with the principle of step A225, where the area to be selected is obtained at a position that meets the cutting length.

[0188] A236, obtain the fragment area of ​​the blocky fragment in the selected area of ​​the blocky crack, and select the center point of the blocky fragment corresponding to the largest fragment area as the blocky cutting point.

[0189] A237, using the block-shaped cutting point as a reference, construct the circumferential dividing line of the block-shaped crack as the cutting position of the block-shaped crack.

[0190] See Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. The electronic device 40 includes: a processor 41, a memory 42, and a computer program; wherein...

[0191] The memory 42 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0192] The processor 41 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0193] Alternatively, the memory 42 can be either standalone or integrated with the processor 41.

[0194] When the memory 42 is a device independent of the processor 41, the device may further include:

[0195] Bus 43 is used to connect the memory 42 and the processor 41.

[0196] The present invention also provides a readable storage medium storing a computer program, which, when executed by a processor, is used to implement the methods provided in the various embodiments described above.

[0197] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0198] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0199] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the external acoustic signature of wind turbine blades, characterized in that, include: The voiceprint signal is converted into a time-frequency feature map, and cluster analysis is performed on the time-frequency feature map to obtain a cluster label sequence. The cluster label sequence is then processed to obtain valid candidate segments. The effective candidate segments are scaled and time-domain padded to generate the acoustic segments to be tested and decomposed into complementary sub-bands. The complementary sub-bands are then input into a deep feature extraction network to obtain the embedding vectors of each complementary sub-band and the confidence scores of the known defect categories. The embedded vectors are grouped according to the device operating cycle to obtain multiple detection groups. The consistency of the features within the detection group is calculated. If the consistency within the group meets the preset conditions, the inter-group similarity of the detection group is calculated to obtain the outlier group. The embedded vectors of the outlier group are compared with the anomaly class library to obtain the recognition result.

2. The method according to claim 1, characterized in that, The process of processing the clustered label sequence to obtain valid candidate fragments includes: Identify bright regions in the cluster label sequence and lock the original label positions of the bright regions as a protection set; The clustering label sequence is smoothed to generate a smoothed label sequence; The original tag positions in the protection set are backfilled into the corresponding positions in the smoothed tag sequence to obtain the integrated tag sequence; The integrated tag sequence is traversed, and segments with a duration lower than a preset threshold are removed to obtain valid candidate segments.

3. The method according to claim 1, characterized in that, The step of performing scale normalization and temporal completion processing on the valid candidate segments to generate the acoustic segment to be tested includes: The time length of the valid candidate segment is obtained. If the time length is greater than the preset standard length, the segment is truncated based on the center. If the time length is less than the preset standard length, the segment is padded at the boundary to obtain a normalized segment. When the time center interval between adjacent normalized segments is greater than the preset threshold of the device operating cycle, a time positioning point is determined between the two adjacent normalized segments with the device operating cycle as the step size, and supplementary data is generated at the time positioning point. The normalized segment and the padded data are combined in time sequence to generate the acoustic segment to be tested.

4. The method according to claim 1, characterized in that, The calculation of the consistency of intra-group features of the detection group, and if the intra-group consistency meets a preset condition, the calculation of inter-group similarity of the detection group to obtain the outlier group, includes: The mean cosine similarity among all embedded vectors within the same detection group is used as the intra-group consistency index. If the consistency index within the group is lower than the preset consistency threshold, the corresponding detection group is deleted. If the consistency index within the group meets the consistency threshold, then the corresponding detection group is used as the comparison group, and the comparison group is statistically analyzed to obtain a candidate set. Obtain the centroids of all comparison groups in the candidate set, and calculate the distance between the mean of the intra-group features of each comparison group and the centroid as the inter-group separation degree. When the inter-group separation is determined to be greater than a preset outlier threshold, the corresponding comparison group is designated as the outlier group.

5. The method according to claim 1, characterized in that, Also includes: When the identification result is determined to be a damage result, the corresponding wind turbine blade is regarded as the damaged blade, and a damage model corresponding to the damaged blade is constructed. The damaged blades are customized based on the damaged area in the damage model and the limited length of the carrier vehicle.

6. The method according to claim 5, characterized in that, The customized processing of the damaged blade based on the damaged area in the damage model and the limited length of the carrier vehicle includes: When the damage region in the damage model is determined to be a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut. The blade to be cut is then cut based on a limited length to obtain segmented sections. When the damage region in the damage model is determined to be an axial crack, the cutting position corresponding to the axial crack is determined based on a limited length, and the damage model is cut according to the cutting position to obtain segmented segments. When the damage region in the damage model is determined to be a blocky crack, the cutting position corresponding to the blocky crack is determined based on a limited length, and the damage model is cut according to the cutting position to obtain segmented segments.

7. The method according to claim 6, characterized in that, When the damage region in the damage model is determined to be a circumferential crack, the fixed position at the circumferential crack is fixed to obtain the blade to be cut, including: When the damage region in the damage model is determined to be a circumferential crack, the circumferential crack length is obtained. When the length of the circumferential crack is less than or equal to a preset fixed length, the center position of the circumferential crack is obtained as the fixed position; When the length of the circumferential crack is greater than the preset fixed length, a wrapping region corresponding to the circumferential crack is constructed, and the number of divided regions is obtained by rounding up the ratio of the circumferential crack length to the preset fixed length. The package area is evenly divided based on the number of division areas to obtain multiple fixed areas. Each fixed area is then divided into three equal parts to obtain an upper area, a middle area, and a lower area for each fixed area. The region with the most crack pixels in the upper, middle, and lower regions of each fixed region is selected as the fixed position of each fixed region; The circumferential crack is fixed at the fixed position to obtain the blade to be cut.

8. The method according to claim 6, characterized in that, When the damage region in the damage model is determined to be an axial crack, the cutting position corresponding to the axial crack is determined based on a limited length, including: When the damage region in the damage model is determined to be an axial crack, the axial crack length of the axial crack, the first endpoint of the axial crack that is closest to the blade root, and the second endpoint that is closest to the blade tip are obtained. When the axial crack length is less than or equal to the predetermined length, the remaining length is obtained based on the difference between the predetermined length and the axial crack length. The remaining length is then halved to obtain the extension length. Starting from the first endpoint and the second endpoint respectively, the extension is extended to both sides based on the extension length to obtain the cutting position corresponding to the axial crack; When the axial crack length is greater than the limit length, the number of cut segments is obtained by rounding up the ratio of the axial crack length to the limit length. The selection area for axial cracks is determined based on the number of cut segments; The location of the smallest crack width in the selected region of the axial crack is taken as the cutting position of the axial crack.

9. The method according to claim 8, characterized in that, The step of extending the length to both sides based on the extension length, starting from the first endpoint and the second endpoint respectively, to obtain the cutting position corresponding to the axial crack includes: Based on the aforementioned extension length, starting from the first endpoint, the line is extended along a circumferential direction pointing towards the blade root and perpendicular to the wind turbine blade to obtain the first line segment; Based on the aforementioned extension length, starting from the second endpoint, the second line segment is obtained by extending along the direction pointing towards the blade tip and perpendicular to the circumferential direction of the fan. Obtain the remaining construction endpoints in the first and second line segments, and construct circumferential cutting lines perpendicular to the first and second line segments respectively as the cutting positions of the axial cracks based on the construction endpoints; The process of determining the selectable region for axial cracks based on the number of cut segments includes: Based on the ratio of the axial crack length to the number of cut segments, the average length is obtained. Taking the first or second endpoint as the starting point, the center point is determined sequentially in the axial crack based on the average length. The safe length is obtained by taking half of the difference between the limited length and the average length, and a defined perpendicular line is constructed that passes through the center point and is perpendicular to the circumferential direction of the wind turbine blade. Starting from the center point, determine the division points along the defined vertical line and in both directions away from the center point based on the safe length; Using the defined dividing points as a reference, construct circumferential dividing lines perpendicular to the defined vertical lines, and obtain the area between the circumferential dividing lines as the area to be selected.

10. The method according to claim 6, characterized in that, When the damage region in the damage model is determined to be a blocky crack, the cutting position corresponding to the blocky crack is determined based on a limited length, including: When the damage region in the damage model is determined to be a blocky crack, the maximum axial distance of the blocky crack in the axial direction of the wind turbine blade is obtained. When the maximum axial distance is less than or equal to the defined length, the block extension distance is obtained by taking half of the difference between the defined length and the maximum axial distance. Starting from the two extreme points of the blocky crack in the axial direction of the wind turbine blade, the cutting position corresponding to the blocky crack is obtained by extending the blocky extension distance to both sides. When the maximum axial distance is greater than the limited length, the number of cutting blocks is obtained by rounding up the ratio of the maximum axial distance to the limited length. The selection area for block-shaped cracks is determined based on the number of cutting blocks; Obtain the fragment area of ​​the blocky fragment in the selected area of ​​the blocky crack, and select the center point of the blocky fragment corresponding to the largest fragment area as the blocky cutting point; Using the block-shaped cutting points as a reference, a circumferential dividing line of the block-shaped crack is constructed as the cutting position of the block-shaped crack.

Images