A sea wind turbine group blade monitoring method and system based on distributed acoustic sensing

Abstract

The application discloses a kind of sea fan group blade monitoring method and system based on distributed acoustic sensing, which comprises the following steps: obtaining the original space-time data of fan based on distributed acoustic sensing (DAS) system and carrying out data preprocessing, to obtain the two-dimensional time-frequency spectrum image of source fan DAS data;The two-dimensional time-frequency spectrum image of source fan DAS data is identified by state classification network, and the current operating condition of source fan is obtained;According to the current operating condition of source fan, blade speed monitoring and high-order modal identification are carried out, to realize the monitoring of sea fan group blade.The present application can capture the weak high-frequency vibration characteristics of fan blades, and realize early warning of fan blade structure health.The present application, as a kind of sea fan group blade monitoring method and system based on distributed acoustic sensing, can be widely applied to the field of fan blade monitoring technology.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade monitoring technology, and in particular to a method and system for monitoring the blades of offshore wind turbine groups based on distributed acoustic sensing. Background Technology

[0002] Wind power generation, especially offshore wind power, offers significant advantages over onshore wind power, including abundant wind resources, more stable wind speeds, higher power generation, and no land occupation. However, these advantages come with more severe operation and maintenance challenges. Offshore wind turbines operate in highly corrosive environments characterized by high temperatures, high salt spray, and high humidity, and are subjected to continuous and complex loads such as typhoons, random wind loads, and wave impacts. This results in a much higher equipment failure rate than onshore wind turbines, with shorter maintenance windows and extremely high maintenance costs. Statistics show that the operation and maintenance costs of offshore wind power can account for 20%-30% of the total life-cycle cost of a project. Within the entire wind turbine generator set, the blades are the core component for capturing wind energy, and their aerodynamic performance and structural health directly determine the generator's power generation efficiency and operational safety. At the same time, the blades are also among the most complex, largest, and most stress-affected components. With the trend towards larger wind turbines, blade lengths have exceeded 100 meters, dramatically increasing the risks they face.

[0003] However, existing technologies for monitoring wind turbine blades include regular manual / drone inspections, point sensor-based monitoring, and SCADA data-based condition monitoring. However, regular manual / drone inspections are "post-incident" checks, unable to capture the transient process and evolution of damage, resulting in poor early warning capabilities. Furthermore, they rely on specialized technicians and expensive equipment (such as high-powered telescopes and drones), and the detection window is severely limited by weather conditions (wind speed, sea state), leading to high overall costs. Point sensor-based monitoring exposes sensors and signal transmission lines to harsh marine and electromagnetic environments, making them susceptible to salt spray corrosion, lightning strikes (EMI), and electromagnetic interference, leading to data distortion and even system failure. Furthermore, deploying numerous sensors and cables on the blades not only increases blade weight and lightning strike risk but also makes installation and subsequent maintenance extremely difficult. SCADA data-based condition monitoring primarily collects macroscopic operating parameters such as wind speed, power, and rotational speed. Anomalies in these data are analyzed to indirectly infer potential blade problems. This method is insensitive to early, minor blade damage, has low diagnostic accuracy, and suffers from high false alarm and false negative rates. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a method and system for monitoring the blades of offshore wind turbine clusters based on distributed acoustic sensing, which can capture the weak high-frequency vibration characteristics of wind turbine blades and achieve early warning of the structural health of wind turbine blades.

[0005] The first technical solution adopted in this invention is: a method for monitoring the blades of offshore wind turbine fleets based on distributed acoustic sensing, comprising the following steps: The raw spatiotemporal data of the wind turbine is acquired based on the distributed acoustic sensing (DAS) system and the data is preprocessed to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. The operating conditions of the source-end wind turbine are identified by using a state classification network to analyze the two-dimensional time-spectrum image of the DAS data. Based on the current operating conditions of the source wind turbine, blade speed monitoring and high-order mode identification are performed to achieve monitoring of the blades of the offshore wind turbine group.

[0006] Furthermore, the step of acquiring the raw spatiotemporal data of the wind turbine based on the distributed acoustic sensing (DAS) system and performing data preprocessing to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data specifically includes: Raw spatiotemporal data of the wind turbine are acquired through a distributed acoustic sensing (DAS) system. Determine the sensing channel position of the source-end fan in the DAS sensing optical cable to obtain the DAS data of the source-end fan; Based on the DAS data of the source-end wind turbine, the corresponding one-dimensional time series vibration signal is extracted and short-time Fourier transform is performed to obtain a two-dimensional time spectrum image of the source-end wind turbine DAS data.

[0007] Furthermore, the step of identifying the operating condition of the source-end wind turbine by analyzing the two-dimensional time-spectrum image of the source-end wind turbine DAS data through a state classification network to obtain the current operating condition of the source-end wind turbine specifically includes: A state classification network is constructed, which includes an initial convolutional layer, several residual attention modules, a global average pooling layer, and a softmax output layer. The several residual attention modules include a two-dimensional convolutional layer, an attention mechanism module, and residual connections. The two-dimensional time-spectrum image of the source-end wind turbine DAS data is input into the state classification network; Based on the initial convolutional layer of the state classification network, feature extraction is performed on the two-dimensional time-spectrum image of the source-end wind turbine DAS data to obtain shallow features of the source-end wind turbine DAS data. Based on the residual attention module of the state classification network, deep feature extraction and downsampling are performed on the shallow features of the source-end wind turbine DAS data to obtain the deep features of the source-end wind turbine DAS data. Based on the global average pooling layer of the state classification network, the deep features of the source-end wind turbine DAS data are transformed to obtain a one-dimensional vector of the source-end wind turbine DAS data. The Softmax output layer based on the state classification network performs probabilistic identification on the one-dimensional vector of the source-end wind turbine DAS data and outputs the current operating condition of the source-end wind turbine.

[0008] Furthermore, the step of monitoring blade speed and identifying higher-order modes based on the current operating conditions of the source-end wind turbine to achieve monitoring of the blades of the offshore wind turbine group specifically includes: Based on the current operating conditions of the source-end wind turbine, the current operating conditions of the source-end wind turbine are selected as a one-dimensional time series signal of the wind turbine power generation state; A bandpass filter is applied to the one-dimensional time series signal under wind turbine power generation conditions to obtain the filtered one-dimensional time series signal. The filtered one-dimensional time series signal is segmented using a period segmentation algorithm to obtain a single-period wind turbine blade signal; The blade speed of the wind turbine is obtained by monitoring the blade speed of a single-cycle wind turbine blade signal through signal processing algorithms. High-order mode recognition is performed on the single-cycle wind turbine blade signal to obtain the wind turbine blade health status recognition result; By combining the blade rotation speed of the wind turbine with the identification results of the blade health status, the blades of the offshore wind turbine group can be monitored.

[0009] Furthermore, the step of monitoring the blade speed of the wind turbine blades using a signal processing algorithm in a single cycle to obtain the blade speed specifically includes: Autocorrelation calculation is performed on the single-cycle wind turbine blade signal to determine the first peak value other than zero point, and the time corresponding to the peak value is set as the approximate cycle of blade rotation. The target segment of the filtered one-dimensional time series signal is extracted as a template and subjected to sliding cross-correlation calculation with the approximate period of blade rotation to locate the starting point of each period and obtain the blade speed of the wind turbine.

[0010] Furthermore, the step of performing high-order mode recognition on the single-cycle wind turbine blade signal to obtain the wind turbine blade health status recognition result specifically includes: Using a preset multiple of the single-cycle wind turbine blade signal at rated speed as the trigger value, wind turbine blade signals higher than the trigger value are subjected to high-order mode identification and downsampling to obtain the accumulated signal; The power spectrum of the wind turbine blades is obtained by performing a high-resolution short-time Fourier transform on the accumulated signal. A baseline spectrum of healthy blades is set and compared with the power spectrum of wind turbine blades to determine the health status identification result of wind turbine blades.

[0011] The second technical solution adopted in this invention is: a wind turbine blade monitoring system based on distributed acoustic sensing, comprising: The first module is used to acquire the raw spatiotemporal data of the wind turbine based on the distributed acoustic sensing (DAS) system and perform data preprocessing to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. The second module is used to identify the operating conditions of the source-end wind turbine by using a state classification network to analyze the two-dimensional time-spectrum image of the source-end wind turbine's DAS data, thereby obtaining the current operating conditions of the source-end wind turbine. The third module is used to monitor blade speed and identify higher-order modes based on the current operating conditions of the source wind turbine, thereby enabling monitoring of the blades of the offshore wind turbine group.

[0012] The beneficial effects of the method and system of this invention are as follows: This invention acquires the raw spatiotemporal data of the wind turbine based on a distributed acoustic sensing (DAS) system and performs data preprocessing to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. Since the DAS fiber optic sensor is a passive medium, it is naturally immune to electromagnetic interference and lightning strikes, ensuring the long-term stability and reliability of the monitoring data. Furthermore, the two-dimensional time-spectrum image of the source-end wind turbine DAS data is used to identify the wind turbine's operating condition through a state classification network to obtain the current operating condition of the source-end wind turbine. Finally, based on the current operating condition of the source-end wind turbine, blade speed monitoring and high-order mode identification are performed to realize the monitoring of the blades of the offshore wind turbine group. It can capture weak high-frequency vibration characteristics that are excited by early damage and cannot be identified by SCADA data, thereby achieving a true early warning of structural health. Attached Figure Description

[0013] Figure 1 This is a flowchart of the steps of a method for monitoring the blades of a wind turbine fleet based on distributed acoustic sensing according to the present invention. Figure 2 This is a structural block diagram of a wind turbine blade monitoring system based on distributed acoustic sensing according to the present invention. Figure 3 This is a topology diagram of the accompanying optical cable provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the wind turbine accompanying optical cable provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of machine-collected DAS data waterfall provided in a specific embodiment of the present invention; Figure 6 This is a schematic diagram of offshore wind turbine blade monitoring provided in a specific embodiment of the present invention; Figure 7 This is a time-frequency diagram illustrating the basic operating conditions of a wind turbine as provided in a specific embodiment of the present invention; Figure 8 This is a schematic diagram of the state classification network provided in a specific embodiment of the present invention; Figure 9This is a visual schematic diagram of the frequency band signal transmitted by the blades according to a specific embodiment of the present invention; Figure 10 This is a schematic diagram of an example of blade periodic segmentation provided in a specific embodiment of the present invention; Figure 11 This is a schematic diagram of a high-order mode observation process during a typhoon, provided in a specific embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0015] First, the technical terms used in the embodiments of this invention will be explained: 1) Accompanying Fiber Optic Cable: When laying the submarine main cable connecting the wind farm, one or more such optical cables are usually laid along the way for communication. The submarine cable will pass through the inside of the optical fiber.

[0016] 2) Distributed Acoustic Sensing (DAS) System: This system can turn an entire ordinary optical cable (i.e., "accompanying optical cable") into a super-sensitive microphone or vibration sensor from beginning to end. It uses the phase change of back Rayleigh scattered light modulated by vibration to reflect the magnitude of the vibration.

[0017] 3) Spectrogram: It can "draw" a sound or vibration signal. The horizontal axis of this graph is time, the vertical axis is pitch (frequency), and the color intensity represents the volume.

[0018] 4) Autocorrelation function: used to find recurring patterns in a signal.

[0019] 5) Template Matching: This is a "finding the clues" search method. In this embodiment of the invention, a standard waveform (template) is first defined, and then this template is used to compare segments of long data to find the most similar parts.

[0020] 6) Higher-Order Modes: "Mode" can be understood as the inherent vibrational "posture" of an object. "Higher-order" refers to vibrational modes with higher frequencies and less common occurrences.

[0021] Reference Figure 1This invention provides a method for monitoring the blades of offshore wind turbine fleets based on distributed acoustic sensing, the method comprising the following steps: S100. Based on the distributed acoustic sensing (DAS) system, the original spatiotemporal data of the wind turbine is acquired and preprocessed to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. Specifically, the raw spatiotemporal data of the wind turbine is acquired through a distributed acoustic sensing (DAS) system; the sensing channel position of the source-end wind turbine in the DAS sensing optical cable is determined to obtain the DAS data of the source-end wind turbine; based on the DAS data of the source-end wind turbine, the corresponding one-dimensional time-series vibration signal is extracted and a short-time Fourier transform is performed to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data.

[0022] In this embodiment, a distributed acoustic sensing (DAS) system is used to collect spatiotemporal sensing data via the fiber optic cable connecting the wind turbine. In other words, DAS technology is used to transform the ordinary communication fiber optic cable laid on the wind turbine foundation into a long-distance, high-density vibration sensor array.

[0023] Based on the fiber optic cable topology map, the source-end wind turbine for vibration data can be located. The data collected by the DAS system can stretch for tens of kilometers, covering numerous facilities. It is essential to first determine "which data segment corresponds to which wind turbine," such as... Figure 3 As shown, DAS technology collected all vibrations within a ten-kilometer range along the wind turbine's accompanying optical cable. The vast majority of these vibrations are located beneath the seabed. To determine the exact location of the wind turbine's interior, a topological map of the optical cable is needed. This embodiment... Figure 3 The signals at locations around 8.8km and 9.6km in the topology are what this embodiment requires.

[0024] In addition, such as Figure 4 As shown, the wind turbine-accompanying optical cable reaches the wind turbine, the suction cylinder structure climbs to the top surface of the foundation, and then the optical cable goes straight into the tower and comes out from the other side.

[0025] like Figure 5 The image shown is a waterfall plot of the accompanying optical cable monitored by DAS for one minute. A DAS data waterfall plot, also known as a time-space plot, is a method of displaying massive amounts of raw DAS data in the form of a two-dimensional image. It can simultaneously display information in three dimensions: time, space, and vibration intensity. We have determined the location of the wind turbine, and we can index a vibration time series in the DAS data that shows the change in vibration intensity over time. Then, by performing time-frequency transformation on the vibration time series, we can obtain the dataset we use for the wind turbine operating condition classification network.

[0026] Therefore, as Figure 6As shown, firstly, raw spatiotemporal data is received from the DAS demodulator. This data is a two-dimensional matrix, with the Y-axis representing the fiber optic channel (space) and the X-axis representing time. The color values ​​represent vibration intensity. Based on the fiber optic cable topology diagram, the fiber optic channel with the strongest and most stable vibration signal from the target wind turbine foundation is located from the waterfall diagram. Then, the one-dimensional time-series vibration signal on this selected channel is further extracted. One-dimensional time-series signal Converted into a two-dimensional time-spectral image, it is used as input to the convolutional model. Perform a short-time Fourier transform (STFT) to generate the time-frequency spectrum, the expression of which is: ; In the above formula, This represents the calculated time-frequency spectrum matrix. Indicates frequency, This represents the one-dimensional time-series vibration signal from the source-end fan's DAS. Indicated by time The window function centered on This represents the integral variable.

[0027] Based on previous calculations, to obtain an image size close to 512x128, we set NFFT = 1024 and stride HopLength = 228.

[0028] S200. The operating conditions of the source-end wind turbine are identified by the two-dimensional time-spectrum image of the DAS data of the source-end wind turbine through the state classification network, so as to obtain the current operating conditions of the source-end wind turbine. Specifically, a state classification network is constructed, comprising an initial convolutional layer, several residual attention modules, a global average pooling layer, and a Softmax output layer. The residual attention modules include a two-dimensional convolutional layer, an attention mechanism module, and residual connections. A two-dimensional time-spectrum image of the source-end wind turbine DAS data is input into the state classification network. Based on the initial convolutional layer of the state classification network, feature extraction is performed on the two-dimensional time-spectrum image of the source-end wind turbine DAS data to obtain shallow features. Based on the residual attention modules of the state classification network, deep feature extraction and downsampling processing are performed on the shallow features of the source-end wind turbine DAS data to obtain deep features. Based on the global average pooling layer of the state classification network, the deep features of the source-end wind turbine DAS data are transformed to obtain a one-dimensional vector of the source-end wind turbine DAS data. Based on the Softmax output layer of the state classification network, probability recognition is performed on the one-dimensional vector of the source-end wind turbine DAS data to output the current operating condition of the source-end wind turbine.

[0029] In this embodiment, the DAS data of the located wind turbine is input into a state classification network to identify its operating condition. First, the located wind turbine channel data is converted into a time spectrum and input into a pre-trained deep learning "state classification network". This network automatically identifies and outputs the current operating condition of the wind turbine, such as: start-up state, power generation state, shutdown state, maintenance state, etc.

[0030] First, it should be noted that the wind turbine status classification network used in this embodiment of the invention uses DAS data collected over a long period of time from wind turbines under different real operating conditions. By synchronizing and cross-validating with the operating logs recorded by the wind farm SCADA system, each segment of DAS data (and its generated time spectrum) is accurately labeled with its operating condition (such as startup, power generation, shutdown, etc.), thereby constructing a high-quality training set containing tens of thousands of labeled samples.

[0031] A 512x128 time-spectral image is input into a pre-trained 2D-CNN + CBAM + residual connection classification network. The network architecture first extracts shallow features through initial convolutional layers, then stacks multiple residual attention modules. Each module contains a convolutional layer, an attention mechanism module, and residual connections for deep feature extraction and downsampling. When the dimensions are mismatched, the residual connections use 1x1 convolutions with stride for projection. A global average pooling layer then converts the feature map into a one-dimensional vector. Finally, a softmax output layer outputs the final classification probability. The final output shows one of the five operating states of the wind turbine: [Power Generation State], [Start-up State], [Shutdown State], [Standby State], and [Maintenance State].

[0032] like Figure 7 As shown, specifically: 1) Startup: During the startup phase, multiple significant frequency clusters can be observed gradually rising from low frequencies. These frequency clusters originate from the acceleration processes of higher-order rotating components in the wind turbine drivetrain (such as gearboxes and generators). When these frequency clusters finally stabilize at their respective rated frequencies, it indicates that the wind turbine has entered normal power generation.

[0033] 2) Generating state: When the wind turbine is running stably, the high-order frequency clusters originating from the transmission chain tend to stabilize. At this time, in the low-frequency band of 10Hz to 40Hz, a periodic pulse signal with concentrated energy can be clearly observed. This is the blade-tower sweep signal that this method focuses on, and its periodicity directly reflects the rotational speed of the blade.

[0034] 3) Shutdown state: This process is the opposite of the start-up state, and is characterized by all rotation-related frequency clusters gradually decreasing from the steady state until they disappear.

[0035] 4) Standby mode: In this state, all frequency characteristics related to the rotation of the blades and drive train have disappeared from the spectrum. Only the 50Hz power frequency signal and its harmonics are clearly visible in the figure. This proves that the fan has stopped rotating, but its control system is still powered on and ready to operate.

[0036] 5) Maintenance Status: When the wind turbine enters a periodic inspection or maintenance phase, all its power supply systems are disconnected. Therefore, it exhibits no specific frequency characteristics in the time spectrum, with only weak background noise present.

[0037] More specifically, such as Figure 8 As shown, the wind turbine status classification network is a deep convolutional neural network. Its core purpose is to receive a 512x128 wind turbine time spectrum (i.e., (Height=512, Width=128, Channels=1)) and automatically classify it into one of five predefined operating conditions: start-up, power generation, shutdown, standby, or maintenance.

[0038] The network structure is mainly divided into two parts: 1) Feature extraction backbone: It consists of 5 cascaded, powerful residual downsampling modules used to extract increasingly abstract features from the time spectrum layer by layer.

[0039] 2) Classifier Head: Responsible for receiving the final features extracted by the backbone network and mapping them to the probabilities of the five operating conditions.

[0040] The feature extraction backbone is the core of the network, consisting of five consecutive residual modules (Conv2D-16 to Conv2D-256). The numbers (16, 32, 64, 128, 256) represent the number of channels (Filters) in the output feature map of each module.

[0041] The internal structure of the module (taking Conv2D-16 as an example in this embodiment) is as follows: Figure 8 As shown, each large module in the diagram (such as Conv2D-16) is designed as a "residual downsampling module". It contains two paths: 1) Main Processing Path: This is the path for main feature extraction and size reduction. The input data will pass through the following paths in sequence: 3x3 convolutional layer (3*3 2DConv): Extracts spatial features from the input feature map.

[0042] Batch normalization (BatchNorm): Normalizes the data distribution, accelerates model convergence, and prevents overfitting.

[0043] CBAM Module: This is an attention mechanism module. It "intelligently" analyzes feature maps, enhancing useful channels and spatial regions while suppressing irrelevant noise, making the network more focused on key features.

[0044] Max Pooling: Performs downsampling (usually 2x2), halving the height and width of the feature map (e.g., 512x128 -> 256x64). This serves two purposes: first, to expand the "receptive field" of subsequent convolutional layers, and second, to reduce computational cost.

[0045] 2) Shortcut Connection: This is the essence of ResNet, which allows information to propagate "across layers" and solves the problem of training deep networks.

[0046] 1x1 Convolution (1*1 2DConv): The input data passes through a 1x1 convolutional layer in parallel. Because the main path involves MaxPooling (size halving) and a change in the number of channels (e.g., from 1 to 16), the 1x1 convolution on this shortcut path must have a stride of 2 and 16 filters to force its output dimension to be exactly the same as the output dimension of the main path.

[0047] 3) Module Output (Residual Connection) Finally, the output of the main path and the output of the shortcut connection are added element by element.

[0048] The formula can be expressed as: ; in, It is the input of the module.

[0049] It is the main processing path. This is the quick connection path for the projector. It is the final output of the entire module.

[0050] Data Dimension Changes (Example): Input: (512, 128, 1); Conv2D-16 Output: (256, 64, 16); Conv2D-32 Output: (128, 32, 32); Conv2D-64 Output: (64, 16, 64); Conv2D-128 Output: (32, 8, 128); Conv2D-256 Output: (16, 4, 256); Classifier Head 1) Global Average Pooling (Green Bar) operation: It "flattens" the feature map (16, 4, 256) into a one-dimensional feature vector of length 256. Output dimension (256).

[0051] 2) Operation of the first fully connected layer (FC Layer 1 / Hidden Layer): The feature vector of (256) is fed into the first fully connected layer. This layer acts as a "hidden layer" and is responsible for performing deep nonlinear combinations of the features. The number of neurons is 128, and the activation function uses ReLU to increase nonlinearity, with an output dimension of (128).

[0052] The activation function formula is: ; 3) Operation of the second fully connected layer (FC Layer 2 / Output Layer): The (128) output vector of the first FC layer (after Dropout) is fed into the second fully connected layer.

[0053] The number of neurons is 5, which is the number of categories. The activation function does not use ReLU or other activation functions because its output (called logits) will be directly fed into the Softmax layer, with an output dimension of (5).

[0054] The activation function formula is: ; 4) Softmax layer (softmax - Light Grey Bar) operation: This step remains unchanged. It receives 5 logits from the second FC layer and converts them into probability distributions corresponding to the five operating conditions, using the following formula: ; Output: A probability distribution containing 5 values, such as [0.05, 0.90, 0.02, 0.02, 0.01], indicating that the model has 90% confidence that the current time spectrum belongs to the power generation state.

[0055] S300 performs blade speed monitoring and high-order mode identification based on the current operating conditions of the source wind turbine, thereby enabling monitoring of the blades of the offshore wind turbine group.

[0056] S310. Based on the current operating conditions of the source-end wind turbine, select the current operating conditions of the source-end wind turbine as a one-dimensional time series signal under the wind turbine power generation state; S320. Perform bandpass filtering on the one-dimensional time series signal under wind turbine power generation conditions to obtain the filtered one-dimensional time series signal. In this embodiment, when the wind turbine is generating electricity, the blade passing signal is extracted. The algorithm will only continue to execute if the identification result indicates that the wind turbine is generating electricity. Under the generating electricity condition, the blades will periodically rotate and pass over the tower, generating a very distinctive periodic vibration signal, namely the "blade passing signal" or "tower-skimming signal".

[0057] like Figure 9 As shown, the key frequency band of the blade-tower sweep signal is concentrated between 10Hz and 40Hz. Therefore, when the system identifies that the wind turbine is in power generation mode, this 10~40Hz bandpass filter is immediately applied to extract the original signal.

[0058] The filtered signal waterfall plot clearly shows the characteristic signal generated by the periodic passage of the three blades. All subsequent speed calculations and higher-order mode information extractions are based on this filtered signal.

[0059] Specifically, low-frequency signals related to blade rotation are separated from the original signal, and a one-dimensional time series signal of the selected channel is obtained. ,right Bandpass filtering is performed to remove high-frequency noise and power frequency components, retaining only the frequency band containing the blade passage frequency (usually within the range of 10 to 40 Hz).

[0060] S330. The filtered one-dimensional time series signal is segmented by a period segmentation algorithm to obtain a single-period wind turbine blade signal. Specifically, such as Figure 10 As shown, the signal after bandpass filtering from 10-40Hz is first processed. Perform autocorrelation analysis of the envelope, through To determine the overall machine rotation cycle (1P cycle), such as Figure 10 The 6.25s shown is an example. Subsequently, in this... Within the period, template matching (cross-correlation) is applied. Precise positioning is achieved by three blades ( , , Each of the three independent pulse events generated during tower sweeps will be used to determine the consecutive pulses based on their timestamps. The signal stream is cut and categorized into the corresponding blades ( This allows for independent segmentation of the signal for each blade and enables the calculation of the instantaneous rotational speed. .

[0061] S340. The blade speed of the wind turbine is obtained by monitoring the blade speed of the single-cycle wind turbine blade signal through a signal processing algorithm. Specifically, autocorrelation is performed on the single-cycle wind turbine blade signal to determine the first peak value excluding zero. The time corresponding to this peak value is set as the coarse cycle of blade rotation. The target segment of the filtered one-dimensional time series signal is extracted as a template and subjected to sliding cross-correlation calculation with the coarse cycle of blade rotation to locate the starting point of each cycle and obtain the blade speed of the wind turbine.

[0062] In this embodiment, a period segmentation algorithm is applied to transform the continuous signal into a single-cycle blade signal. Signal processing algorithms (such as autocorrelation analysis or template matching) are used to accurately locate the start and end points of each rotation cycle in the continuous signal. Then, this long signal is "cut" into a series of "single-cycle blade signal" data segments to extract the rotational speed information.

[0063] Perform autocorrelation on the filtered signal to find the first peak value excluding zero. The time corresponding to this peak value is the approximate period of the blade rotation. ,by For reference, a typical "glide tower" pulse waveform is extracted from the signal as a template. Through sliding cross-correlation calculations, the starting point of each cycle is precisely located, thus obtaining a more accurate cycle. Output blade speed (RPM = 60 / ).

[0064] S350: Perform high-order mode recognition on the single-cycle wind turbine blade signal to obtain the health status recognition result of the wind turbine blade; Specifically, a preset multiple of the single-cycle wind turbine blade signal at rated speed is used as the trigger value. Wind turbine blade signals higher than the trigger value are subjected to high-order mode identification and downsampling to obtain an accumulated signal. The accumulated signal is subjected to a high-resolution short-time Fourier transform to obtain the power spectrum of the wind turbine blade. A healthy blade reference spectrum is set and compared with the power spectrum of the wind turbine blade to determine the health status identification result of the wind turbine blade.

[0065] In this embodiment, a time-domain alarm threshold for the blade signal is set, and long-term signals are accumulated to identify higher-order modes. A preset time-domain characteristic threshold is used to filter valid signal periods. In this embodiment, the threshold is set to 1.5 times the maximum amplitude of the signal passing through the blade at rated speed. Signals exceeding the threshold are accumulated over a long period (10 minutes), and the time-spectrum of the blade is printed to determine the higher-order harmonic information triggered by the blade under abnormal conditions (such as typhoons).

[0066] High-precision spectral analysis is performed on the accumulated signal to diagnose the health status of the blades. The reference spectrum is established by collecting long-term accumulated signals from blades that have undergone professional maintenance and been confirmed to be healthy, under various typical wind speeds, and then performing spectral analysis.

[0067] Threshold judgment: The trigger value is 1.5 times the maximum amplitude of the single-channel signal at rated speed of the wind turbine. Signals exceeding this baseline are used for higher-order mode identification. Signals accumulated over a long period (10 minutes) are downsampled to 10Hz. .

[0068] High-resolution time-spectrum transform: for accumulated signals Perform high-resolution short-time Fourier transform.

[0069] Modal identification: The calculated power spectrum is compared with the pre-stored "healthy leaf baseline spectrum". The focus is on analyzing whether there is a significant increase in the energy (peak height) of a known high-order modal frequency.

[0070] Output: Final diagnostic conclusion, such as "Leaf health status is normal" or "Warning: Abnormal high-order mode was found at 2.28, and the leaf is suspected of having early structural damage."

[0071] like Figure 11 As shown, under typhoon conditions, the spectrum clearly reveals four key frequency bands that are excited: 1) is the wave signal frequency band located at the bottom of the spectrum, and its energy comes from the disturbance of gravity waves on the ocean surface; 2) The blade passes through the fundamental frequency (1P) at 0.37Hz, which is the blade-tower-grazing signal of interest to us; 3) It is a second harmonic (2P) of 0.74 Hz, which is the second harmonic of the signal passing through the blades and is significantly excited during the typhoon process; 4) is a 2.28Hz sixth harmonic (6P), which is also a harmonic of the signal passing through the blades. Its excitation is also attributed to the extreme conditions of the typhoon.

[0072] S360 combines the blade speed of the wind turbine with the identification results of the health status of the wind turbine blades to realize the monitoring of the blades of the offshore wind turbine group.

[0073] In summary, the purpose of this invention is to achieve full-time online monitoring and early warning: establishing a 24 / 7 automated monitoring system that transforms "periodic checkups" into "real-time health monitoring," enabling early warnings at the initial stage of damage. By utilizing the "tower-grabbing effect," the overall blade condition is indirectly assessed. The invention does not aim for complete physical coverage of the blade itself, but rather for a precise analysis of a highly sensitive and regular signal at the wind turbine base—the "blade-tower-grabbing effect"—to indirectly but holistically assess the structural health of the blade. Each time a blade passes over the tower, its stiffness, mass distribution, and aerodynamic characteristics are transmitted to the foundation in the form of a unique vibrational "acoustic pattern." The purpose of this invention is to capture and analyze subtle changes in this "acoustic pattern," thereby achieving a comprehensive diagnosis of the blade's overall dynamic characteristics without physical contact. Utilizing fiber optic sensors deployed at the stationary wind turbine base, with their inherent immunity to electromagnetic interference and lightning strikes, ensures the monitoring system can provide highly reliable data stably over a long period, completely avoiding the challenge of installing any electronic components on the rotating blades, which face the highest risk of lightning strikes. Achieving low-cost, high-efficiency monitoring of wind turbine clusters: Leveraging the long-distance monitoring advantages of DAS technology and reusing existing accompanying communication optical cables in the wind farm, a single system can monitor the entire wind farm. This significantly avoids the high-risk, high-cost fiber optic laying, signal transmission, and subsequent maintenance work on the rotating blades, thereby greatly reducing the system's deployment difficulty and total lifecycle cost.

[0074] The embodiments of the present invention differ from the prior art in the following technical features: 1) Operating condition classification method based on DAS time spectrum: The one-dimensional time series data of DAS is transformed into a two-dimensional time spectrum, and then the spectrum is input into a deep learning network (CNN) as a two-dimensional array to realize the automatic classification of five operating conditions of wind turbines (start-up, power generation, shutdown, standby and maintenance).

[0075] 2) WindTurbine-DAS-SFNet Network Architecture: The backbone consists of five consecutive "residual downsampling modules," supplemented by a classification head composed of global average pooling and two fully connected layers (FC). The residual downsampling modules sequentially include a 3*3 2DConv, BatchNorm, CBAM, and MaxPooling. A 1*1 2DConv with stride=2 is used for projection to match the dimensions after downsampling the main path, and finally, the two paths are added element-wise.

[0076] 3) Hierarchical analysis logic of "identify first, then diagnose": First, the wind turbine is confirmed to be in power generation state through operating condition identification, and then the subsequent signal extraction, period segmentation and modal diagnosis are activated (triggered).

[0077] 4) Individual blade signal segmentation method: The rotation period (T1P) of the entire machine is determined by autocorrelation analysis, and then template matching is used within this period to accurately locate and separate each individual blade. , , The method of signal pulses.

[0078] 5) Threshold-triggered long-term accumulation diagnostic method: For triggering: set a specific time domain threshold (e.g., 1.5 times the maximum amplitude of the rated speed); For accumulation: only accumulate the signal for a long time (e.g., 10 minutes) when the signal is higher than the threshold; For identification: perform high-resolution time spectrum analysis on the accumulated signal and diagnose the damage by identifying abnormal high-order harmonics (e.g., 6P octave at 2.28 Hz).

[0079] 6) Data link and physical basis for non-invasive monitoring: The embodiments of the present invention utilize the accompanying optical cable of the wind turbine foundation (non-rotating component) as a sensor to capture the indirect vibration signal of "blade sweeping the tower effect" to realize the whole set of technical ideas for health assessment of rotating blades.

[0080] Therefore, the embodiments of the present invention have the following advantages compared with the prior art: 1) Compared to solutions relying on electronic "point sensors" such as strain gauges, the embodiments of this invention offer inherent safety and high reliability. Offshore wind farms are subject to strong electromagnetic interference (EMI) and lightning strike risks, which can easily cause traditional electronic sensors to fail or distort data. The DAS fiber optic sensor utilized in the embodiments of this invention is a passive medium, naturally immune to EMI and lightning strikes, ensuring long-term stable and reliable monitoring data. Furthermore, by monitoring the stationary wind turbine foundation, the high-risk and complex construction and maintenance work associated with installing sensors on rotating blades is completely avoided.

[0081] 2) Compared to indirect analysis methods based on SCADA data, this invention provides higher diagnostic accuracy and early warning capabilities. SCADA data is macroscopic and coarse, and insensitive to early, minor structural damage. This invention, however, employs an innovative two-stage algorithm. First, it uses a CNN state classification network to accurately pinpoint the "power generation condition," and then performs high-resolution high-order modal analysis on the "blade-tower-grabbing signal" under this condition. This method can capture weak high-frequency vibration characteristics (such as sixth-order harmonics) excited by early damage that cannot be identified by SCADA data, thereby achieving true early warning of structural health.

[0082] Reference Figure 2 A wind turbine blade monitoring system based on distributed acoustic sensing, comprising: The first module 201 is used to acquire the original spatiotemporal data of the wind turbine based on the distributed acoustic sensing DAS system and perform data preprocessing to obtain a two-dimensional time-spectrum image of the source wind turbine DAS data. The second module 202 is used to identify the operating conditions of the wind turbine by using a state classification network to analyze the two-dimensional time-spectrum image of the source-end wind turbine DAS data, and to obtain the current operating conditions of the source-end wind turbine. The third module 203 is used to monitor the blade speed and identify higher-order modes based on the current operating conditions of the source wind turbine, so as to realize the monitoring of the blades of the offshore wind turbine group.

[0083] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0084] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for monitoring the blades of an offshore wind turbine fleet based on distributed acoustic sensing, characterized in that, Includes the following steps: The raw spatiotemporal data of the wind turbine is acquired based on the distributed acoustic sensing (DAS) system and the data is preprocessed to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. The operating conditions of the source-end wind turbine are identified by using a state classification network to analyze the two-dimensional time-spectrum image of the DAS data. Based on the current operating conditions of the source-end wind turbine, the current operating conditions of the source-end wind turbine are selected as a one-dimensional time series signal of the wind turbine power generation state; A bandpass filter is applied to the one-dimensional time series signal under wind turbine power generation conditions to obtain the filtered one-dimensional time series signal. The filtered one-dimensional time series signal is segmented using a period segmentation algorithm to obtain a single-period wind turbine blade signal; Autocorrelation calculation is performed on the single-cycle wind turbine blade signal to determine the first peak value other than zero point, and the time corresponding to the peak value is set as the approximate cycle of blade rotation. The target segment of the filtered one-dimensional time series signal is extracted as a template and subjected to sliding cross-correlation calculation with the approximate period of blade rotation to locate the starting point of each period and obtain the blade speed of the wind turbine. Using a preset multiple of the single-cycle wind turbine blade signal at rated speed as the trigger value, wind turbine blade signals higher than the trigger value are subjected to high-order mode identification and downsampling to obtain the accumulated signal; The power spectrum of the wind turbine blades is obtained by performing a high-resolution short-time Fourier transform on the accumulated signal. Set a baseline spectrum for healthy blades and compare it with the power spectrum of wind turbine blades to determine the health status identification results of wind turbine blades. By combining the blade rotation speed of the wind turbine with the identification results of the blade health status, the blades of the offshore wind turbine group can be monitored.

2. The method for monitoring the blades of a marine wind turbine fleet based on distributed acoustic sensing according to claim 1, characterized in that, The step of acquiring the raw spatiotemporal data of the wind turbine based on the distributed acoustic sensing (DAS) system and performing data preprocessing to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data specifically includes: Raw spatiotemporal data of the wind turbine are acquired through a distributed acoustic sensing (DAS) system. Determine the sensing channel position of the source-end fan in the DAS sensing optical cable to obtain the DAS data of the source-end fan; Based on the DAS data of the source-end wind turbine, the corresponding one-dimensional time series vibration signal is extracted and short-time Fourier transform is performed to obtain a two-dimensional time spectrum image of the source-end wind turbine DAS data.

3. The method for monitoring the blades of a wind turbine fleet based on distributed acoustic sensing according to claim 2, characterized in that, The step of identifying the operating condition of the source-end wind turbine by using a state classification network to analyze the two-dimensional time-spectrum image of the source-end wind turbine's DAS data to obtain the current operating condition of the source-end wind turbine specifically includes: A state classification network is constructed, which includes an initial convolutional layer, several residual attention modules, a global average pooling layer, and a softmax output layer. The several residual attention modules include a two-dimensional convolutional layer, an attention mechanism module, and residual connections. The two-dimensional time-spectrum image of the source-end wind turbine DAS data is input into the state classification network; Based on the initial convolutional layer of the state classification network, feature extraction is performed on the two-dimensional time-spectrum image of the source-end wind turbine DAS data to obtain shallow features of the source-end wind turbine DAS data. Based on the residual attention module of the state classification network, deep feature extraction and downsampling are performed on the shallow features of the source-end wind turbine DAS data to obtain the deep features of the source-end wind turbine DAS data. Based on the global average pooling layer of the state classification network, the deep features of the source-end wind turbine DAS data are transformed to obtain a one-dimensional vector of the source-end wind turbine DAS data. The Softmax output layer based on the state classification network performs probabilistic identification on the one-dimensional vector of the source-end wind turbine DAS data and outputs the current operating condition of the source-end wind turbine.

4. A monitoring system for the blades of an offshore wind turbine fleet based on distributed acoustic sensing, characterized in that, Includes the following modules: The first module is used to acquire the raw spatiotemporal data of the wind turbine based on the distributed acoustic sensing (DAS) system and perform data preprocessing to obtain a two-dimensional time-spectrum image of the source-end wind turbine DAS data. The second module is used to identify the operating conditions of the source-end wind turbine by using a state classification network to analyze the two-dimensional time-spectrum image of the source-end wind turbine's DAS data, thereby obtaining the current operating conditions of the source-end wind turbine. The third module is used to select the current operating condition of the source-end wind turbine as a one-dimensional time series signal of the wind turbine's power generation state based on the current operating condition of the source-end wind turbine. A bandpass filter is applied to the one-dimensional time series signal under wind turbine power generation conditions to obtain the filtered one-dimensional time series signal. The filtered one-dimensional time series signal is segmented using a period segmentation algorithm to obtain a single-period wind turbine blade signal; Autocorrelation calculation is performed on the single-cycle wind turbine blade signal to determine the first peak value other than zero point, and the time corresponding to the peak value is set as the approximate cycle of blade rotation. The target segment of the filtered one-dimensional time series signal is extracted as a template and subjected to sliding cross-correlation calculation with the approximate period of blade rotation to locate the starting point of each period and obtain the blade speed of the wind turbine. Using a preset multiple of the single-cycle wind turbine blade signal at rated speed as the trigger value, wind turbine blade signals higher than the trigger value are subjected to high-order mode identification and downsampling to obtain the accumulated signal; The power spectrum of the wind turbine blades is obtained by performing a high-resolution short-time Fourier transform on the accumulated signal. Set a baseline spectrum for healthy blades and compare it with the power spectrum of wind turbine blades to determine the health status identification results of wind turbine blades. By combining the blade rotation speed of the wind turbine with the identification results of the blade health status, the blades of the offshore wind turbine group can be monitored.

Images