A wind turbine condition monitoring method based on distributed optical fiber sensing

Abstract

The present application relates to the technical field of wind turbines, and discloses a method for monitoring the state of a wind turbine based on distributed optical fiber sensing. The method includes: collecting historical blade vibration signals and historical optical cable vibration signals to form a historical data set; optimizing the historical data set to obtain an optimized data set, and constructing a vibration transfer function based on the optimized data; capturing the Rayleigh scattering signal from the wind turbine in the optical cable; obtaining the real-time optical cable vibration signal of the optical cable at the wind turbine based on the phase change of the Rayleigh scattering signal; converting the real-time optical cable vibration signal into a real-time blade frequency domain signal based on the vibration transfer function, and restoring the real-time blade frequency domain signal to a real-time blade vibration signal based on the inverse Fourier transform; identifying the real-time blade vibration signal based on a machine learning algorithm to obtain the real-time operating condition of the blade. The present invention can monitor the working condition of the wind turbine without arranging sensors on the blades, thereby reducing the monitoring cost.

Description

Technical Field

[0001] The present application relates to the technical field of wind turbines, and in particular to a method for monitoring the state of a wind turbine based on distributed optical fiber sensing. Background Art

[0002] Wind turbine safety monitoring is the most important component of wind farm operation and maintenance. Due to the harsh operating environment at sea, conventional methods such as manual inspections are difficult to implement.

[0003] In order to conveniently monitor the safety of wind turbines, many remote methods have been adopted. For example, the Chinese patent application with publication number CN107829885A discloses a wind turbine blade vibration monitoring and system that takes environmental parameter correction into account. This method arranges dual-axis acceleration sensors at the corresponding parts of the wind turbine blades and the wind turbine hub to measure the temperature data and vibration data at the wind turbine blades and the hub. The WindBVM data collector obtains the wind turbine SCADA data from the wind turbine main control cabinet to obtain the wind speed and pitch angle environmental parameters; the WindBVM data collector transmits the vibration data, temperature data and SCADA working condition data through internal and external networks, and finally combines the collected blade vibration data and wind turbine environmental parameters to accurately identify the health status of the blades. For example, a Chinese patent application document with publication number CN118008721A discloses a wind turbine fault diagnosis and early warning method, device, equipment and medium. The method obtains physical characteristics that can be detected by the diagnosis object and operating condition information associated with the physical characteristics from a sensor or SCADA system or CMS system, and obtains the main component characteristics and auxiliary characteristics and the target operating condition information associated with them after corresponding processing. Then, the feature data and operating condition information obtained by combining the user-defined single feature threshold are fused and analyzed, and an alarm conclusion is output, thereby realizing fault diagnosis of the wind turbine.

[0004] However, the sampling frequency of SCADA is low (usually 10 minutes apart, less than 0.002 Hz), and it is not sensitive to fault characteristics. At the same time, the monitoring object is mostly temperature. The drastic change of temperature often means that the wind turbine components have been damaged, which cannot achieve the purpose of early warning of faults. Although CMS continuously monitors and measures the vibration, strain, load and other data of the wind turbine structure at a higher frequency (usually more than 50 Hz), it requires the use of accelerometers, acoustic emission technology and fiber grating sensor technology. Numerous sensors are inevitably accompanied by complex line connections, which poses severe challenges in power supply, wiring and safety protection. Moreover, it is only sensitive to parameter changes at the installation location, and the operating environment at sea is extremely harsh, with a large amount of environmental noise and electromagnetic interference, which greatly affects the performance of the sensor. Therefore, a more accurate and convenient method for remotely monitoring the status of wind turbines is needed. Summary of the invention

[0005] To achieve the above objectives, the present application provides a wind turbine condition monitoring method based on distributed optical fiber sensing.

[0006] In order to achieve the above-mentioned object of the invention, the present invention proposes a method for monitoring the condition of a wind turbine based on distributed optical fiber sensing, comprising:

[0007] Determine a sampling frequency, and synchronously collect historical blade vibration signals and historical optical cable vibration signals in a wind turbine operating environment based on the sampling frequency to form a historical data set;

[0008] Optimizing the historical data set based on historical offshore wind data to obtain an optimized data set, performing data analysis on the optimized data, and constructing a vibration transfer function from a blade vibration signal to an optical cable vibration signal;

[0009] Determine a signal return time according to the distance between the laser emission point and the wind turbine, and capture the Rayleigh scattering signal from the wind turbine in the optical cable based on the signal return time;

[0010] Based on the phase change of the Rayleigh scattering signal, a real-time optical cable vibration signal of the optical cable at the wind turbine is obtained;

[0011] Converting the real-time optical cable vibration signal into a real-time blade frequency domain signal based on the vibration transfer function, and restoring the real-time blade frequency domain signal into a real-time blade vibration signal based on an inverse Fourier transform;

[0012] The real-time blade vibration signal is identified based on a machine learning algorithm to obtain the real-time operating condition of the blade.

[0013] Furthermore, constructing the vibration transfer function comprises the following steps:

[0014] Perform cross-correlation analysis on the historical blade vibration signal and the historical optical cable vibration signal to obtain a delay time, perform time synchronization on the historical blade vibration signal and the historical optical cable vibration signal based on the delay time to obtain a synchronization signal, perform Fourier transform on the synchronization signal to obtain a frequency domain signal, perform spectrum characteristic analysis on the frequency domain signal to obtain system characteristics, the system characteristics include amplitude-frequency characteristics and phase-frequency characteristics, construct an initial function based on the system characteristics, optimize parameters in the initial function based on the least squares method, and obtain the vibration transfer function.

[0015] Furthermore, obtaining the real-time optical cable vibration signal comprises the following steps:

[0016] A laser pulse is injected into the optical cable, and a Rayleigh scattering signal returned by the optical cable is received. The Rayleigh scattering signal is demodulated to obtain a phase change angle, a dynamic strain value is calculated based on the phase change angle, low-frequency drift and high-frequency interference in the dynamic strain value are removed based on a bandpass filter to obtain a filtered signal, and the filtered signal is decomposed by wavelet transform to obtain the real-time optical cable vibration signal.

[0017] Furthermore, identifying the real-time blade vibration signal includes the following steps:

[0018] The machine learning algorithm includes multiple machine learning models, determines the hyperparameters of each machine learning model, constructs the genetic code of the individual based on the hyperparameters, and in the process of using the optimized data set to train each machine learning model, iteratively optimizes the individual based on the genetic algorithm to obtain the optimal individual with the goal of maximizing fitness, determines the hyperparameters of the machine learning model based on the genetic code in the optimal individual, defines a judgment model after training is completed, inputs the real-time blade vibration signal into the judgment model, and obtains multiple recognition results, the recognition results include the probability that the blade belongs to each real-time blade vibration signal, and the recognition results are synthesized based on a voting algorithm to obtain the real-time operating condition.

[0019] Further, training the machine learning model comprises the following steps:

[0020] The time domain features of the historical blade vibration signals in the optimized data set are extracted, and the time domain features are subjected to dimensionality reduction processing based on the PCA algorithm. A Pareto chart is constructed based on the dimensionality reduction results. Based on the Pareto chart, the principal components with a cumulative contribution greater than 90% are selected as the dimensionality reduction results. The machine learning model is trained using a minibatch method based on the dimensionality reduction results to obtain the judgment model.

[0021] Furthermore, synthesizing the recognition results in a voting algorithm includes the following steps:

[0022] Based on the validation set, a first accuracy rate and a second accuracy rate of each judgment model are obtained, wherein the first accuracy rate is an overall judgment accuracy rate, and the second accuracy rate is a judgment accuracy rate for each real-time operating condition, and an initial weight is assigned to each judgment model based on the first accuracy rate;

[0023] After inputting the dimensionality reduction result of the real-time blade vibration signal into the judgment model, each of the judgment models outputs a first probability that the blade is in various real-time operating conditions based on the softmax function, adjusts the first probability to a second probability based on the second accuracy rate, calculates a third probability that the blade is in each real-time operating condition based on the second probability and the initial weight, normalizes the third probability to obtain the judgment probability of each real-time operating condition, and takes the real-time operating condition corresponding to the maximum judgment probability as the final judgment result.

[0024] Furthermore, optimizing the historical data set includes the following steps:

[0025] A transition data set is obtained after outlier detection and missing value processing are performed on the historical data set. Each data in the transition data set has a data label, and the data label is the blade operating condition corresponding to the data. Based on the trained hidden Markov model, the offshore wind data is used as an observation sequence, and the blade operating condition is used as hidden data. The blade operating condition at each time point in the transition data set is predicted, and the high confidence parameters are used as prediction results. The prediction results and the data labels at the same time point are compared to optimize the historical data set.

[0026] Further, the sampling frequency is determined based on the Nyquist sampling theorem.

[0027] Furthermore, the machine learning model includes a random forest model, a BP neural network model, and a SVM model.

[0028] Furthermore, the real-time operating conditions of the blades include starting conditions, stopping conditions, stalling conditions, speed change conditions and abnormal start-stop conditions.

[0029] Beneficial effects: The present invention collects the blade vibration signals of the wind turbine through sensors, and establishes a historical data set in combination with the optical cable vibration signals, and then optimizes the historical data set through historical offshore wind data, thereby correcting abnormal data and further improving the reliability and accuracy of the data. After the vibration transfer function is established through the historical data set, the actual monitored optical cable vibration signal is converted into a blade vibration signal through the wind power accompanying optical cable, combined with distributed optical fiber sensing technology and vibration transfer function, and the blade vibration signal is analyzed by using a machine learning algorithm, thereby achieving an accurate estimation of the blade working condition. Compared with traditional methods, the present invention does not deploy sensors on the wind turbine, but monitors the blade working condition through the wind power accompanying optical cable, which reduces the monitoring cost and avoids the influence of environmental noise and electromagnetic interference on data collection and transmission.

[0030] The present invention can measure the axial strain along the optical cable by sending laser pulses and detecting the phase change of Rayleigh backscattering generated by impurities in the optical cable. Therefore, the existing wind turbine companion optical cable can be used to monitor the wind turbine vibration signal in real time, monitor the operating status and vibration of the wind turbine, understand its vibration characteristics under different wind speeds and sea conditions, improve the safety of wind turbines, and reduce the operation and maintenance costs of wind turbine farms. In addition, distributed optical fiber sensing technology can effectively make up for the shortcomings of existing technologies by virtue of its significant characteristics such as passivity, ability to achieve long-distance monitoring, and large response bandwidth. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0032] Figure 1 A flowchart of the steps of a wind turbine condition monitoring method based on distributed optical fiber sensing for this application;

[0033] Figure 2 This is a schematic diagram of the laying of the wind power accompanying optical cable for this application;

[0034] Figure 3 Schematic diagram of PCA dimensionality reduction results for this application

[0035] Figure 4 This is a schematic diagram of the voting process for this application. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0037] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0038] First of all, it should be noted that the present invention does not directly place vibration sensors on the blades, but uses existing wind power accompanying optical cables for monitoring, that is, while protecting the safety of the optical cables, it also monitors the working conditions of the wind turbine blades. The working conditions monitored in this embodiment mainly include blade start, stop, rotation, stall, speed change, and abnormal start and stop. In addition, since the optical cable and the wind turbine supply cable pass in parallel, the condition of the wind turbine supply cable can also be monitored. The specific implementation method is introduced below.

[0039] like Figure 1 As shown, a wind turbine condition monitoring method based on distributed optical fiber sensing comprises:

[0040] S1: Determine a sampling frequency, and synchronously collect historical blade vibration signals and historical optical cable vibration signals in a wind turbine operating environment based on the sampling frequency to form a historical data set.

[0041] In this embodiment, the sampling frequency is determined based on the Nyquist sampling theorem.

[0042] According to the Nyquist sampling theorem, the selected sampling frequency should be greater than twice the target maximum frequency. Assuming that the maximum frequency of wind turbine blade vibration is 15 Hz, the sampling frequency is set to 30 Hz. Then, according to the sampling frequency, a vibration sensor is set on the wind turbine to collect more accurate historical blade vibration signals, and the optical cable vibration signal at the point where the optical cable passes through the wind turbine platform is calculated through the Rayleigh scattering signal of the optical cable. In particular, the above-mentioned acquisition of historical blade vibration signals through vibration sensors is only for the purpose of establishing a vibration transfer function in the early stage. When the blade working condition is actually determined later, the corresponding blade vibration signal is only estimated through the optical cable vibration signal, and will not be obtained by the sensor.

[0043] S2: Optimizing the historical data set based on historical offshore wind data to obtain an optimized data set, performing data analysis on the optimized data, and constructing a vibration transfer function from a blade vibration signal to an optical cable vibration signal.

[0044] The historical data set includes blade vibration signals. In the future, it is necessary to train the machine learning model based on the blade vibration signals and construct the vibration transfer function. To this end, it is necessary to determine the correctness of the blade vibration signal data label to ensure the accuracy of subsequent training and transfer function construction. The optimized historical data set is defined as the optimized data set. The vibration transfer function from the optical cable vibration signal to the blade vibration signal is constructed based on the optimized data set. The vibration transfer function represents the relationship between the blade vibration signal and the optical cable vibration signal. For example, when the blade rotates, the blade will vibrate, and the vibration is transmitted to the platform through the main structure, thereby causing the vibration of the optical cable. The specific correction method and the construction method of the transfer function will be introduced later.

[0045] S3: determining a signal return time according to the distance between the laser emission point and the wind turbine, and capturing a Rayleigh scattering signal originating from the wind turbine in the optical cable based on the signal return time.

[0046] S4: Based on the phase change of the Rayleigh scattering signal, a real-time optical cable vibration signal of the optical cable at the wind turbine is obtained.

[0047] Specifically, the ratio of distance to the speed of light can be used as the signal return time. After determining the return time, the Rayleigh scattering signals originating from the location of the wind turbine can be screened out based on the return time. The Rayleigh scattering signal is then phase-demodulated to obtain the corresponding phase change, and then the real-time optical cable vibration signal at the wind turbine is determined based on the phase change.

[0048] S5: Converting the real-time optical cable vibration signal into a real-time blade frequency domain signal based on the vibration transfer function, and restoring the real-time blade frequency domain signal into a real-time blade vibration signal based on inverse Fourier transform.

[0049] S6: Identify the real-time blade vibration signal based on a machine learning algorithm to obtain the real-time operating condition of the blade.

[0050] In this embodiment, the operating conditions of the blades include a starting condition, a stopping condition, a stalling condition, a speed changing condition, and an abnormal starting and stopping condition.

[0051] Through the vibration transfer function constructed previously, the optical cable frequency domain signal can be converted into a blade frequency domain signal. For machine learning, it is easier to judge the working condition of the blade based on the time domain characteristics. Therefore, the obtained blade frequency domain characteristics are restored to the real-time blade vibration signal, and then the time domain features are extracted from the blade vibration signal. The time domain features are input into the machine learning algorithm to identify the blade working condition.

[0052] The present invention collects the blade vibration signals of the wind turbine through sensors, and establishes a historical data set in combination with the optical cable vibration signals, and then optimizes the historical data set through historical offshore wind data, thereby correcting abnormal data and further improving the reliability and accuracy of the data. After the vibration transfer function is established through the historical data set, the actual monitored optical cable vibration signal is converted into a blade vibration signal through the wind power accompanying optical cable, combined with distributed optical fiber sensing technology and vibration transfer function, and the blade vibration signal is analyzed by using a machine learning algorithm, thereby achieving an accurate estimation of the blade working condition. Compared with traditional methods, the present invention does not deploy sensors on the wind turbine, but monitors the blade working condition through the wind power accompanying optical cable, which reduces the monitoring cost and avoids the influence of environmental noise and electromagnetic interference on data collection and transmission.

[0053] It is particularly noteworthy that the present invention can monitor the working condition of the wind turbine without arranging sensors on the blades, thereby reducing the monitoring cost.

[0054] In this embodiment, constructing the vibration transfer function includes the following steps:

[0055] Perform cross-correlation analysis on the historical blade vibration signal and the historical optical cable vibration signal to obtain a delay time, perform time synchronization on the historical blade vibration signal and the historical optical cable vibration signal based on the delay time to obtain a synchronization signal, perform Fourier transform on the synchronization signal to obtain a frequency domain signal, perform spectrum characteristic analysis on the frequency domain signal to obtain system characteristics, the system characteristics include amplitude-frequency characteristics and phase-frequency characteristics, construct an initial function based on the system characteristics, optimize parameters in the initial function based on the least squares method, and obtain the vibration transfer function.

[0056] The following example illustrates the above steps. First, the blade and the optical cable are regarded as a whole system, the optical cable vibration signal is the input signal of the system, and the blade vibration signal is the output signal of the system. Based on this, the cross-correlation analysis is used to calculate the delay time between the two signals. The delay time between the two signals is determined by determining the peak positions of the two signals. For example, the historical optical cable vibration signal is delayed by 1s compared to the historical blade vibration signal. According to the calculated delay time, the historical optical cable vibration signal is shifted forward by 1s to complete the time alignment and synchronization of the two signals. Then, the synchronization signal is Fourier transformed to obtain the frequency domain signal. The amplitude-frequency characteristics and phase-frequency characteristics of the entire system are calculated based on the frequency domain signal. These two characteristics respectively describe the gain and phase changes of the system for different frequency components.

[0057] The initial function is constructed based on the system characteristics, and the specific form of the initial function is determined according to the classical function form of the multi-degree-of-freedom system, for example , Where N is the number of characteristic patterns, is the system gain factor of the i-th mode. When calculating, the average value of the low-frequency band of the amplitude-frequency characteristic can be used as the initial value of the gain factor. f represents the frequency component of the entire system, which is obtained through Fourier transform. is the cutoff frequency of the i-th characteristic mode, specifically the starting position of the high-frequency attenuation in the amplitude-frequency characteristic, j is the complex part, which is used to represent the complex characteristics of the frequency domain transfer function, is the delay time of the i-th characteristic mode; when determining the number of characteristic modes, we can start from 1 and gradually fit until the fitting error is minimized, thereby determining the final value of i. Finally, after optimizing the fitting of historical data, the transfer function obtained is In subsequent calculations, the frequency signal of the optical cable vibration signal is multiplied by the transfer function, and the result is inverse Fourier transformed to obtain the blade vibration signal.

[0058] In this embodiment, obtaining the real-time optical cable vibration signal includes the following steps:

[0059] A laser pulse is injected into the optical cable, and a Rayleigh scattering signal returned by the optical cable is received. The Rayleigh scattering signal is demodulated to obtain a phase change angle, a dynamic strain value is calculated based on the phase change angle, low-frequency drift and high-frequency interference in the dynamic strain value are removed based on a bandpass filter to obtain a filtered signal, and the filtered signal is decomposed by wavelet transform to obtain the real-time optical cable vibration signal.

[0060] like Figure 2 As shown in the figure, the optical cable returns to the ground after passing through the wind turbine platform. When monitoring, a pulsed laser with a wavelength of 1550 nm and a pulse width of 10-50 ns is first injected into the optical cable through a fiber coupler. The tiny structural inhomogeneities in the optical fiber will cause Rayleigh scattering. After the scattered signal returns over time, a high-speed photodetector is used to receive the Rayleigh scattering signal, and the detected Rayleigh scattering signal is collected through a high-speed sampling card. By comparing the Rayleigh scattering signals of two consecutive pulses, the phase change is extracted, which is calculated specifically with reference to the first formula, which is: ,in, represents the phase change of the Rayleigh scattering signal at time t, is the Rayleigh scattered complex signal corresponding to the first laser pulse received at time t, represents the complex conjugate of the Rayleigh scattering complex signal corresponding to the second laser pulse received at time t, Im represents the imaginary part of the complex number, and Re represents the real part of the complex number.

[0061] The dynamic strain value is calculated based on the second formula, which is: ,in, represents the dynamic strain value, is the laser wavelength, n is the fiber refractive index, and L is the calibration length, whose value is calculated by the sampling frequency and the width of the laser pulse. The dynamic strain value includes normal vibration signals, low-frequency drift signals, and high-frequency interference signals. Therefore, a bandpass filter is needed to remove the low-frequency drift signals and high-frequency interference signals to obtain the filtered signal. Then, through wavelet transform, the signal is decomposed into multiple frequency components, from which signals in a specific frequency range, such as 10 ~ 50 Hz, are screened out to obtain the optical cable vibration signal.

[0062] In this embodiment, identifying the real-time blade vibration signal includes the following steps:

[0063] The machine learning algorithm includes multiple machine learning models, determines the hyperparameters of each machine learning model, constructs the genetic code of the individual based on the hyperparameters, and in the process of using the optimized data set to train each machine learning model, iteratively optimizes the individual based on the genetic algorithm to obtain the optimal individual with the goal of maximizing fitness, determines the hyperparameters of the machine learning model based on the genetic code in the optimal individual, defines a judgment model after training is completed, inputs the real-time blade vibration signal into the judgment model, and obtains multiple recognition results, the recognition results include the probability that the blade belongs to each real-time blade vibration signal, and the recognition results are synthesized based on a voting algorithm to obtain the real-time operating condition.

[0064] The machine learning models include random forest model, BP neural network model and SVM model.

[0065] For the random forest model, its hyperparameters include the number of decision trees, the maximum depth of the decision tree, the minimum number of samples contained in each leaf node, etc. For the BP neural network model, its hyperparameters include the number of hidden layers, the number of neurons in the hidden layer, and the connection weights between neurons. For the SVM model, its hyperparameters include kernel functions, regularization parameters, etc. Later, during the training process, these hyperparameters are optimized by genetic algorithms to speed up the convergence of the machine learning model. After the model training is completed, the real-time blade vibration signal obtained is input into the judgment model so that each judgment model outputs the recognition result. Finally, the optimal recognition result is selected by voting as the implementation operation condition. The specific voting method will be introduced later.

[0066] In this embodiment, training the machine learning model includes the following steps:

[0067] The time domain features of the historical blade vibration signals in the optimized data set are extracted, and the time domain features are subjected to dimensionality reduction processing based on the PCA algorithm. A Pareto chart is constructed based on the dimensionality reduction results. Based on the Pareto chart, the principal components with a cumulative contribution greater than 90% are selected as the dimensionality reduction results. The machine learning model is trained using a minibatch method based on the dimensionality reduction results to obtain the judgment model.

[0068] The time domain features in this embodiment include standard deviation, skewness, kurtosis, peak-to-peak value, average value, maximum value and minimum value. After the time domain features are extracted, they are subjected to dimensionality reduction processing based on the PCA dimensionality reduction algorithm, such as Figure 3As shown in the figure, the principal component with a cumulative contribution of more than 90% is selected as the dimension reduction result and input into the machine learning model for training. During training, the minibatch method is used to speed up the model training. After the real-time blade vibration signal is subsequently generated, the time domain features extracted from the real-time blade vibration signal are directly reduced in dimension by combining the dimension reduction matrix obtained by the previous dimension reduction, and input into the recognition result described in the judgment model.

[0069] This embodiment integrates the recognition results based on the voting algorithm and includes the following steps:

[0070] Based on the validation set, the first accuracy and the second accuracy of each judgment model are obtained, wherein the first accuracy is the overall judgment accuracy, and the second accuracy is the judgment accuracy for each real-time operating condition. An initial weight is assigned to each judgment model based on the first accuracy.

[0071] After inputting the dimensionality reduction result of the real-time blade vibration signal into the judgment model, each of the judgment models outputs a first probability that the blade is in various real-time operating conditions based on the softmax function, adjusts the first probability to a second probability based on the second accuracy rate, calculates a third probability that the blade is in each real-time operating condition based on the second probability and the initial weight, normalizes the third probability to obtain the judgment probability of each real-time operating condition, and takes the real-time operating condition corresponding to the maximum judgment probability as the final judgment result.

[0072] like Figure 4 As shown, the following examples are used to illustrate the above steps. It is assumed that the blade has a startup condition, a stop condition, and an operating condition, which are hereinafter referred to as condition y1, condition y2, and condition y3. There are three machine learning models, which are hereinafter referred to as model M1, model M2, and model M3. During the training process, the first accuracy and the second accuracy of each judgment model are obtained based on the validation set. For example, the overall accuracy of model M1, model M2, and model M3 are 80%, 70%, and 60%, respectively, and the prediction accuracy of model M1 for condition y1 is 0.796. The initial weight of each model is generated based on the third formula. The third formula is ,in, is the initial weight of the nth machine learning model, is the first probability of the nth machine learning model, and m is the total number of machine learning models. After calculation by the third formula, the initial weights of model M1, model M2, and model M3 are determined to be 0.381, 0.333, and 0.286, respectively.

[0073] The softmax function can normalize multiple values ​​to between 0 and 1, and the total probability sum is 1. Through the softmax function, the first probabilities of the output conditions y1, y2, and y3 of model M1 are 0.7, 0.2, and 0.1 respectively. The second probability is obtained by multiplying the first probability with the corresponding second accuracy. For model M1, the second probability of condition y1 is 0.7*0.796=0.557. Then the third probability of condition y1, condition y2, and condition y3 is calculated by the fourth formula. The fourth formula is ,in, is the third probability of the i-th blade operating condition, is the initial weight of the nth machine learning model, is the second probability of the nth machine learning model for the ith blade operating condition, and G is the sum of the second probabilities of all machine learning models corresponding to various operating conditions.

[0074] Finally, the judgment probability of each real-time operating condition is obtained by normalizing each third probability. After normalization, the judgment probabilities corresponding to conditions y1, y2, and y3 are 0.626, 0.195, and 0.178, respectively. Condition y1 is selected as the final judgment result.

[0075] Optimizing the historical data set includes the following steps:

[0076] A transition data set is obtained after outlier detection and missing value processing are performed on the historical data set. Each data in the transition data set has a data label, and the data label is the blade operating condition corresponding to the data. Based on the trained hidden Markov model, the offshore wind data is used as an observation sequence, and the blade operating condition is used as hidden data. The blade operating condition at each time point in the transition data set is predicted, and the high confidence parameters are used as prediction results. The prediction results and the data labels at the same time point are compared to optimize the historical data set.

[0077] This implementation uses the 3 times standard deviation rule for outlier detection and uses linear interpolation to supplement missing values. In the transition data set, each data has a data label, such as the historical blade vibration signal 1 corresponds to the blade operating condition 1, indicating that the historical blade vibration signal 1 was generated under the blade operating condition 1. However, due to labeling errors and other reasons, the labels in the transition data set are not guaranteed to be completely correct, and wrong labels will cause the machine learning model to obtain poor training results. The size of the offshore wind is correlated with the condition of the blades. For example, when the wind is strong, the blade speed is faster. If the wind turbine is in the startup condition at a time point with strong winds, but the blade condition corresponding to the blade vibration signal is the stop condition at this time, it indicates that the data label is abnormal. To this end, the present invention corrects it based on the following steps.

[0078] First, define the hidden state of the hidden Markov model, which is the blade working condition in this embodiment, so that it is regarded as not directly observable. The wind speed in the offshore wind data is used as the observation sequence, that is, the data that can be directly observed in the hidden Markov model. Then, the hidden Markov model is trained using a completely correct data set. After the training is completed, the offshore wind data at a certain time point in the entire transition data set is input into the hidden Markov model as an observation sequence, so as to obtain the confidence in a certain working condition at the blade, and the confidence greater than 90% is used as the prediction result. If the prediction result is the same as the data label of the historical blade vibration signal and the historical optical cable vibration signal at the corresponding time point, it indicates that the label is correct. If it is not the same, the label is marked as abnormal and a reminder is made to perform manual verification.

[0079] Through this step, the historical performance of each judgment model can be combined to comprehensively select the most reliable identification result as the real-time operating condition output of the blade, thereby further increasing the reliability of system monitoring.

[0080] It should be understood that the various technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for monitoring the condition of a wind turbine based on distributed optical fiber sensing, characterized in that: Determine a sampling frequency, and synchronously collect historical blade vibration signals and historical optical cable vibration signals in a wind turbine operating environment based on the sampling frequency to form a historical data set; Optimizing the historical data set based on historical offshore wind data to obtain an optimized data set, performing data analysis on the optimized data set, and constructing a vibration transfer function from a blade vibration signal to an optical cable vibration signal; Determine a signal return time according to the distance between the laser emission point and the wind turbine, and capture the Rayleigh scattering signal from the wind turbine in the optical cable based on the signal return time; Based on the phase change of the Rayleigh scattering signal, a real-time optical cable vibration signal of the optical cable at the wind turbine is obtained; Converting the real-time optical cable vibration signal into a real-time blade frequency domain signal based on the vibration transfer function, and restoring the real-time blade frequency domain signal into a real-time blade vibration signal based on an inverse Fourier transform; Identifying the real-time blade vibration signal based on a machine learning algorithm to obtain the real-time operating condition of the blade; Wherein, constructing the vibration transfer function comprises the following steps: Perform cross-correlation analysis on the historical blade vibration signal and the historical optical cable vibration signal to obtain a delay time, perform time synchronization on the historical blade vibration signal and the historical optical cable vibration signal based on the delay time to obtain a synchronization signal, perform Fourier transform on the synchronization signal to obtain a blade frequency domain signal, perform spectrum characteristic analysis on the blade frequency domain signal to obtain system characteristics, the system characteristics include amplitude-frequency characteristics and phase-frequency characteristics, construct an initial function based on the system characteristics, optimize parameters in the initial function based on the least squares method, and obtain the vibration transfer function.

2. The method according to claim 1, characterized in that: Acquiring the real-time optical cable vibration signal comprises the following steps: A laser pulse is injected into the optical cable, and a Rayleigh scattering signal returned by the optical cable is received. The Rayleigh scattering signal is demodulated to obtain a phase change angle, a dynamic strain value is calculated based on the phase change angle, low-frequency drift and high-frequency interference in the dynamic strain value are removed based on a bandpass filter to obtain a filtered signal, and the filtered signal is decomposed by wavelet transform to obtain the real-time optical cable vibration signal.

3. The method according to claim 1, characterized in that: Identifying the real-time blade vibration signal includes the following steps: The machine learning algorithm includes multiple machine learning models, determines the hyperparameters of each machine learning model, constructs the genetic code of the individual based on the hyperparameters, and in the process of using the optimized data set to train each machine learning model, iteratively optimizes the individual based on the genetic algorithm to obtain the optimal individual with the goal of maximizing fitness, determines the hyperparameters of the machine learning model based on the genetic code in the optimal individual, defines a judgment model after training is completed, inputs the real-time blade vibration signal into the judgment model, and obtains multiple recognition results, the recognition results include the probability that the blade belongs to each real-time blade vibration signal, and the recognition results are synthesized based on a voting algorithm to obtain the real-time operating condition.

4. The method according to claim 3, characterized in that: Training the machine learning model includes the following steps: The time domain features of the historical blade vibration signals in the optimized data set are extracted, and the time domain features are subjected to dimensionality reduction processing based on the PCA algorithm. A Pareto chart is constructed based on the dimensionality reduction results. Based on the Pareto chart, the principal components with a cumulative contribution greater than 90% are selected as the dimensionality reduction results. The machine learning model is trained using the minibatch method and based on the dimensionality reduction results to obtain the judgment model.

5. The method according to claim 4, characterized in that The steps of synthesizing the recognition results based on the voting algorithm include: Based on the validation set, a first accuracy rate and a second accuracy rate of each judgment model are obtained, wherein the first accuracy rate is an overall judgment accuracy rate, and the second accuracy rate is a judgment accuracy rate for each real-time operating condition, and an initial weight is assigned to each judgment model based on the first accuracy rate; After inputting the dimensionality reduction result of the real-time blade vibration signal into the judgment model, each of the judgment models outputs a first probability that the blade is in various real-time operating conditions based on the softmax function, adjusts the first probability to a second probability based on the second accuracy rate, calculates a third probability that the blade is in each real-time operating condition based on the second probability and the initial weight, normalizes the third probability to obtain the judgment probability of each real-time operating condition, and takes the real-time operating condition corresponding to the maximum judgment probability as the final judgment result.

6. The method according to claim 1, characterized in that Optimizing the historical data set includes the following steps: After performing outlier detection and missing value processing on the historical data set, a transition data set is obtained, each data in the transition data set has a data label, and the data label is the blade operating condition corresponding to the data. Based on the trained hidden Markov model, the historical offshore wind data is used as an observation sequence, and the blade operating condition is used as hidden data. The blade operating condition at each time point in the transition data set is predicted, and the high confidence parameters are used as prediction results. The prediction results and the data labels at the same time point are compared to optimize the historical data set.

7. The method according to claim 1, characterized in that The sampling frequency is determined based on the Nyquist sampling theorem.

8. The method according to claim 3, characterized in that The machine learning models include random forest model, BP neural network model and SVM model.

9. The method according to claim 1, characterized in that: The real-time operating conditions of the blades include starting conditions, stopping conditions, stalling conditions, speed change conditions and abnormal start-stop conditions.

Images