A method, device and system for monitoring icing conditions on wind turbine blades

Abstract

The application discloses a kind of for wind turbine blade icing state monitoring method, device and system, belong to wind turbine blade monitoring technical field.This method is based on acoustic emission technology, through arranging acoustic emission sensor in wind turbine blade inner wall, the acoustic emission signal of blade in the process of operation is collected, signal is handled using symplectic geometry modal decomposition method, and state recognition is carried out using deep learning model.This method overcomes the limitations of traditional icing monitoring method, such as response lag, susceptible to environmental interference, influence aerodynamic performance, etc., can realize the accurate identification and timely warning of blade icing state.Experimental verification shows that the identification accuracy of this method for the three states of normal operation, icing and rain of wind turbine blade is more than 99.5%, and has good engineering application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine blade icing state monitoring, in particular to a wind turbine blade icing state monitoring method, device and system. BACKGROUND

[0002] Wind power generation, as a clean and renewable energy source, plays an important role in energy structure transformation. With the expansion of wind farms to highland, mountainous and oceanic areas, the problem of wind turbine blade icing is becoming increasingly prominent, and is no longer limited to traditional cold regions. Blade icing can change the aerodynamic shape of the blade, severely reducing power generation efficiency; at the same time, due to uneven icing, it can increase the unbalanced load, leading to accelerated fatigue of the wind turbine components, and in severe cases, it can cause the wind turbine to collapse; the falling ice blocks can endanger the safety of surrounding personnel and property. According to statistics, blade icing causes billions of dollars in economic losses to the global wind power industry every year, and has become a key problem that needs to be solved in wind power operation and maintenance.

[0003] Currently, wind turbine blade icing monitoring mainly uses visual observation method, power curve method, vibration analysis method, ultrasonic attenuation method and capacitive sensor method. The visual observation method directly observes the icing condition of the blade surface by artificial or camera, although it can achieve high recognition accuracy combined with deep learning technology, but it is easily affected by factors such as light and weather, especially at night or in poor visibility, it is difficult to effectively monitor; the power curve method is based on the power change before and after the blade icing and combines SCADA data to make judgments, although it can obtain more reliable identification results through optimization algorithm, but it is easily disturbed by wind conditions, and may be misjudged due to the increase of surface roughness when initial icing occurs; the vibration analysis method identifies icing state by measuring vibration signal changes, which has high sensitivity, but is easily affected by wind conditions and unit operation, and the sensor arranged on the blade surface will affect the aerodynamic performance; the ultrasonic attenuation method uses sound wave propagation energy change to judge the icing state, which can effectively detect icing conditions at different positions, but acoustic devices need to be arranged at both ends of the monitoring position, which also affects the blade performance; the capacitive sensor method can monitor multiple parameters such as the existence, type and thickness of icing, but it is easily affected by temperature and humidity, and has problems such as poor long-term reliability and increased risk of lightning strikes. These methods have different degrees of response lag, poor anti-interference ability or affect the performance of the blade in practical application.

[0004] In summary, the existing wind turbine blade icing monitoring methods can reflect the icing condition to some extent, but generally have problems such as icing identification response lag, low accuracy, and easy environmental interference. At the same time, the reliability of the sensor cannot be guaranteed for long-term use, and the arrangement method often affects the aerodynamic performance of the blade, which seriously restricts the actual application effect of the icing monitoring system in the wind farm. SUMMARY

[0005] The present application aims to overcome the above-mentioned shortcomings of the prior art, and provides a method, device and system for monitoring the icing state of a wind turbine blade, to solve the problems of the prior art in monitoring the icing of a wind turbine blade, such as the identification of a response lag, low accuracy and susceptibility to environmental interference during the monitoring of icing.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] A method for monitoring the icing state of a wind turbine blade, comprising the following steps:

[0008] S1, collecting an acoustic emission signal; the acoustic emission signal is generated during the operation of the blade, and an acoustic emission sensor collects the acoustic emission signal of the fan blade;

[0009] S2, performing a symplectic geometric modal decomposition on the acoustic emission signal to obtain a main symplectic geometric component;

[0010] S3, reconstructing the acoustic emission signal through the main symplectic geometric component;

[0011] S4, judging the blade state based on the reconstructed acoustic emission signal through a deep neural network; the blade state includes a normal operation state, an icing state and a raining state.

[0012] Further improvements of the present application are as follows:

[0013] Preferably, in S1, the acoustic emission signal is a stress change generated by the icing on the fan blade during the accumulation of the ice layer, the stress on the ice layer and the ice melting process.

[0014] Preferably, the main symplectic geometric component is obtained through analysis of a test signal, and the analysis process is as follows:

[0015] Performing a symplectic geometric modal decomposition on the acoustic emission signal obtained by the test to obtain a plurality of symplectic geometric components, and establishing a scatter plot, a signal superposition plot and a phase space plot based on the symplectic geometric components and the original acoustic emission signal; judging the correlation between each symplectic geometric component and the original acoustic emission signal through the scatter plot, the signal superposition plot and the phase space plot to obtain the main symplectic geometric component;

[0016] Reconstructing the acoustic emission signal through the main symplectic geometric component.

[0017] Preferably, in the step, the correlation between each symplectic geometric component and the original acoustic emission signal is judged through the scatter plot, and if the correlation is greater than a set threshold one, the symplectic geometric component is the main symplectic geometric component.

[0018] Preferably, in the step, the retention rate of each symplectic geometric component on the amplitude feature of the original signal is judged through a signal superposition diagram, and if the retention rate is greater than a set threshold two, the symplectic geometric component is a main symplectic geometric component.

[0019] Preferably, in the step, the retention rate of each symplectic geometric component on the dynamics of the original signal is judged through a phase space diagram, and if the retention rate is greater than a set threshold three, the symplectic geometric component is a main symplectic geometric component.

[0020] Preferably, in S4, the deep neural network model can be any one of a convolutional neural network, a long short-term memory network, a gated recurrent unit or a Transformer neural network.

[0021] Preferably, after the deep neural network model is trained through training set data, the deep neural network model is tested through test set data; the training set data and the test set data are derived from wind turbine blade tests.

[0022] A device for monitoring the icing state of a wind turbine blade, comprising:

[0023] A collection unit is configured to collect an acoustic emission signal; the acoustic emission signal is generated during the operation of the blade, and the acoustic emission sensor is configured to collect the acoustic emission signal of the fan blade;

[0024] A decomposition unit is configured to perform symplectic geometric modal decomposition on the acoustic emission signal to obtain a main symplectic geometric component;

[0025] A reconstruction unit is configured to reconstruct the acoustic emission signal through the main symplectic geometric component;

[0026] A judgment unit is configured to judge the blade state through a deep neural network based on the reconstructed acoustic emission signal; the blade state includes a normal operation state, an icing state and a raining state.

[0027] A wind turbine blade icing state monitoring system for implementing the above monitoring method, comprising an acoustic emission sensor and an information processor; the acoustic emission sensor is arranged in the fan blade, and the acoustic emission sensor and the information processor are electrically connected.

[0028] Compared with the prior art, the present application has the following beneficial effects:

[0029] The application discloses a kind of for wind turbine blade icing state monitoring method, which is based on acoustic emission technology, by arranging acoustic emission sensor in wind turbine blade inner wall, the acoustic emission signal of blade in the process of operation is collected, signal is handled using symplectic geometry modal decomposition method, and state recognition is carried out using deep learning model.The method overcomes the limitations of traditional icing monitoring method, such as response lag, susceptible to environmental interference, influence on aerodynamic performance, etc., and can realize accurate identification and timely warning of blade icing state.Experimental verification shows that the identification accuracy of the method for wind turbine blade normal operation, icing and rain is more than 99.5%, and has good engineering application value.

[0030] The application also discloses a kind of for wind turbine blade icing state monitoring system, which comprises acoustic emission sensor arranged on the inner wall of wind turbine blade, and signal processor connected with acoustic emission sensor electric signal, the data collected or recognized by acoustic emission sensor, combined with signal processing and deep learning technology, realize accurate identification of blade icing state, improve the safety and economic benefit of wind farm operation.The system arranges acoustic emission sensor in the inside of wind turbine blade, based on the rapid response characteristics of acoustic emission signal, can find early icing in time, and can avoid the influence on blade aerodynamic performance, effectively protect acoustic emission sensor, avoid the influence of equipment caused by severe environment;The system has simple structure, low maintenance cost, and is easy to implement in engineering. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a schematic diagram for acoustic emission signal monitoring of wind turbine blade icing;

[0032] Wherein: 1, wind turbine blade;2, acoustic emission sensor;3, acoustic emission signal;4, icing layer;F1, wind load;F2, blade support force;F3, centrifugal force;F4, adhesion;F5, gravity;

[0033] Figure 2 It is a flow chart of the monitoring method of the application;

[0034] Figure 3 It is a flow chart of the monitoring method in specific embodiment of the application;

[0035] Figure 4 It is a recognition effect diagram of deep neural network model;

[0036] Figure 5 It is an icing state SGC1 component and original signal comparison analysis diagram;

[0037] Wherein, (a) is a scatter plot;(b) is signal superposition diagram;(c) is phase space diagram;

[0038] Figure 6Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0039] Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0040] Figure 7 Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0041] Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0042] Figure 8 Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0043] Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0044] Figure 9 Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0045] Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0046] Figure 10 Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state;

[0047] Fig. 4 is a comparative analysis diagram of SGC2 component and original signal in the icing state. DETAILED DESCRIPTION

[0048] The application will be further described below in conjunction with the accompanying drawings:

[0049] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application; the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance; in addition, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection; it can be directly connected, or indirectly connected through an intermediate medium; it can be the communication between two elements inside. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0050] Referring to Figure 1 The first aspect of the present application discloses an icing state monitoring system for wind turbine blade, comprising acoustic emission sensor 2 and information processor, wherein the acoustic emission sensor 2 is arranged inside the outer side wall of the wind turbine blade 1. The outer side wall here refers to the wall surface of the wind turbine blade 1 far from the rotating shaft. Arranging the acoustic emission sensor 2 on the outer side wall of the wind turbine blade 1 can avoid affecting the aerodynamic performance of the wind turbine blade 1, and arranging the acoustic emission sensor 2 inside the wind turbine blade 1 can avoid the influence of external environment on the sensor and improve the service life. The information processor can be a computer or other peripheral industrial equipment, and the information processor and the acoustic emission sensor 2 are electrically connected to analyze the signals returned by the acoustic emission sensor 2.

[0051] For the wind turbine blade 1, the parameters that can affect the propagation characteristics of acoustic emission signals are the interface propagation parameters, including the ice layer acoustic impedance, the blade material acoustic impedance and the transmission coefficient. Specifically, the ice layer acoustic impedance is about 3.5×10 6 kg / (m²·s); the blade material acoustic impedance is about 6.0×10 6 kg / (m²·s), and the transmission coefficient is close to 1. The above interface propagation parameters can ensure the effective transmission of signals; and the wind turbine blade itself has the characteristics of low damping, low frequency scattering effect and small energy attenuation. The above interface propagation parameters and the material characteristics of the wind turbine blade itself can enable the acoustic emission sensor 2 to sensitively collect the changes of the icing state on the wind turbine blade.

[0052] Once the wind turbine blade 1 has an ice layer 4, the ice layer 4 will generate acoustic emission signal sources 3 due to various situations during operation, which are then received by the acoustic emission sensor 2 to realize icing state monitoring of the wind turbine blade 1.

[0053] The acoustic emission signal sources 3 described above include the following situations:

[0054] (1) Ice layer accumulation process, specifically including stress waves generated when supercooled water droplets hit the blade, structural rearrangement during ice crystal formation and growth, and stress changes at the ice layer and blade interface.

[0055] (2) Ice layer stress changes, specifically including periodic stress changes caused by blade rotation, ice layer deformation caused by wind load, and crack formation under centrifugal force.

[0056] (3) Ice melting process, specifically including thermal stress caused by temperature gradient, structural changes caused by ice layer strength reduction, and energy release during ice layer shedding process.

[0057] It should be understood that the number and position of the acoustic emission sensor 2 can be adjusted according to actual conditions.

[0058] In some embodiments of the present application, the acoustic emission sensor 2 is preferably arranged on the inner wall of the blade tip at about 1 / 3 of the length of the blade, which is the area where icing is most severe and melts last, and is subjected to a large centrifugal force, and the ice-acoustic emission signal is more significant. By monitoring the icing state at this position, the icing state of the fan blade 1 can be quickly and accurately obtained.

[0059] In some embodiments of the present application, the acoustic emission sensor 2 is a narrow-band acoustic emission sensor with a frequency range of 10KHz-400KHz, including a data acquisition unit with a sampling rate of 2MHz, and a preamplifier with a gain of 40dB.

[0060] In the above, after the data acquisition unit receives the signal generated by the acoustic emission signal source 3, the collected signal is converted and amplified in the preamplifier, and the signal is output to the signal processor, which is a hardware device for monitoring the icing state of the blade of the wind turbine.

[0061] The monitoring principle of the present application is shown in Figure 1 During operation of the fan blade, the ice layer is subjected to multiple actions of centrifugal force (F3), wind load (F1), gravity (F5) and adhesion force (F4), etc., generating acoustic emission signals. Since the acoustic impedance of ice and the blade material (such as glass fiber / epoxy resin composite material) is close, and the acoustic transmission coefficient is close to 1, the acoustic emission signal can be effectively transmitted to the acoustic emission sensor 2 arranged inside the fan blade 2.

[0062] Referring to Figure 2 The second aspect of the present application discloses a method for monitoring the icing state of the blade of the wind turbine, comprising the following steps:

[0063] S1, collecting acoustic emission signals; the acoustic emission signals 3 are generated during operation of the blade, and the acoustic emission sensor 2 collects the acoustic emission signals of the fan blade 1;

[0064] S2, performing SGMD processing on the collected acoustic emission signals, decomposing the original signals into multiple SGCs, and selecting the SGCs containing the main characteristic information of the signals as the main SGCs;

[0065] S3, reconstructing the acoustic emission signals through the main SGCs;

[0066] S4, judging the blade state through a deep neural network based on the reconstructed acoustic emission signals; the blade state includes normal operation state, icing state and raining state.

[0067] In some embodiments of the present application, in S2, the original signal is decomposed into n SGCs (SGC1-SGC n) and a residual term.

[0068] In some embodiments of the present application, the main symplectic geometric component is obtained by analyzing the experimental data, and the analysis process is as follows:

[0069] (1) The acoustic emission signal of the experiment is subjected to symplectic geometric modal decomposition to obtain a plurality of symplectic geometric components, each symplectic geometric component is used to reconstruct the acoustic emission signal to obtain a symplectic geometric component, and a scatter plot, a signal superposition plot and a phase space plot are established based on the symplectic geometric component and the original acoustic emission signal; the correlation between each symplectic geometric component and the original acoustic emission signal is determined through the scatter plot, the signal superposition plot and the phase space plot, and the main symplectic geometric component is obtained.

[0070] (2) The acoustic emission signal is reconstructed based on the main symplectic geometric component.

[0071] Specifically, in step (1), the scatter plot is used to evaluate the correlation between the symplectic geometric component and the acoustic emission signal; the specific process is that the correlation between each symplectic geometric component and the original acoustic emission signal is determined through the scatter plot, and if the correlation is greater than a set threshold one, the symplectic geometric component is the main symplectic geometric component.

[0072] Specifically, in step (1), the signal superposition plot is used to evaluate and analyze the retention rate of each symplectic geometric component to the amplitude characteristics of the original signal, and if the retention rate is greater than a set threshold two, the symplectic geometric component is the main symplectic geometric component.

[0073] Specifically, in step (1), the phase space plot is used to evaluate and analyze the retention rate of each symplectic geometric component to the dynamic characteristics of the original signal, so as to verify the dynamic characteristics retention. If the retention rate is greater than a set threshold three, the symplectic geometric component is the main symplectic geometric component.

[0074] It should be noted that the set threshold one, the set threshold two and the set threshold three can be adjusted according to the actual accuracy requirement of the method in the application process. The higher the accuracy, the higher the requirement for the number of each set threshold, and the higher the repetition rate of the symplectic geometric component and the original acoustic emission signal.

[0075] It should be noted that when the main symplectic geometric component is determined through the scatter plot, the signal superposition plot and the phase space plot, if the three determination results are different, the main symplectic geometric component needs to be determined according to the specific application scene and research purpose. In some cases, if the similarity of the signal is the most important, the scatter plot is more critical; if the amplitude information of the signal is important for analysis, the signal superposition plot is more important; if the dynamic characteristics of the signal are the main research object, the phase space plot is more critical. Therefore, it needs to be determined according to the actual situation.

[0076] In some embodiments of the present application, a deep learning model is used to classify and identify the reconstructed signal to determine the blade state. In this step, the deep learning model can use a neural network model such as a convolutional neural network (CNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), or a Transformer neural network. It should be understood that the deep learning model is not limited to the above four models and can be adjusted according to actual conditions.

[0077] In some embodiments of the present application, the training set and test set of the deep learning model are constructed by laboratory data.

[0078] A third aspect of the present application discloses an icing state monitoring device for a wind turbine blade, comprising:

[0079] The acquisition unit is used to acquire an acoustic emission signal. The acoustic emission signal is generated during the operation of the blade, and the acoustic emission sensor acquires the acoustic emission signal of the fan blade.

[0080] The decomposition unit is used to perform symplectic geometric modal decomposition on the acoustic emission signal to obtain a main symplectic geometric component.

[0081] The reconstruction unit is used to reconstruct the acoustic emission signal by the main symplectic geometric component.

[0082] The judgment unit is used to determine the blade state by a deep neural network based on the reconstructed acoustic emission signal. The blade state includes a normal operation state, an icing state, and a raining state.

[0083] The following will be further described in conjunction with a specific embodiment.

[0084] Embodiment

[0085] A small fan blade test bench is constructed in the laboratory to simulate the operating state of the fan blade. The motor speed is controlled by a frequency converter controller. The three fan blades of the test bench are 1.6 m long glass fiber / epoxy composite blades. The speed of the adjustable speed motor is 0-280 revolutions / minute. An acoustic emission sensor is arranged in each simulated fan blade to acquire the corresponding acoustic emission signal.

[0086] The experimental conditions of the small fan blade test bench are as follows: the fan blade speed is 0-140 revolutions / minute with 8 gears; the environment includes windless and windy states; and the blade state includes normal, icing, and raining states.

[0087] A series of training sets and test sets are obtained by the above-mentioned test bench to train and construct a deep neural network model. The input of the training set and the test set is the acoustic emission signal, and the output is the blade state.

[0088] Four deep neural network models, CNN model, LSTM model, GRU model and Transformer model, are built. The parameters of each model are as follows:

[0089] (1) CNN model

[0090] - Input layer: 128-dimensional time series

[0091] - Convolutional layer: 3 feature extraction blocks

[0092] - Pooling layer: average pooling with step 2

[0093] - Fully connected layer: 3-class output

[0094] (2) LSTM model

[0095] - Input layer: same as CNN

[0096] - LSTM layer: 128 hidden units

[0097] - Fully connected layer: 3-class output

[0098] (3) GRU model

[0099] - Input layer: same as CNN

[0100] - GRU layer: 512 hidden units

[0101] - Fully connected layer: 3-class output

[0102] (4) Transformer model

[0103] - Input layer: same as CNN

[0104] - Encoder: 6 layers of attention mechanism

[0105] - Decoder: position encoding

[0106] - Output layer: 3-class classification.

[0107] The above models are trained, taking the Transformer model as an example. The architecture includes an input layer, an encoder layer, and a classification layer. The input layer first normalizes the acquired data and slices the signal into 128-point length signals. The encoder layer processes the data through multi-head self-attention mechanism, feedforward neural network, and normalization layer, encodes the data, and passes the encoded results to the classification layer. The classification layer classifies the encoded results through a fully connected layer and a Softmax activation function to obtain the final classification result. During the training process of the Transformer model, the batch size in the training parameters is 64, the learning rate is 0.001, the optimizer is Adam, and the loss function is cross-entropy.

[0108] Referring to Figure 4 The visual comparison chart of the identification effects of the four deep learning models specifically includes the confusion matrix and t-SNE feature visualization results of each model. Among them:

[0109] The confusion matrix shows the identification results of each model on the three states of normal operation, icing state and rain state. The CNN model has an identification accuracy of 100% on the normal state and icing state, and only 1.4% of the rain state samples are misjudged as normal state; the identification accuracy of the LSTM model and the GRU model is more than 99.6%, and the identification of the icing state is particularly accurate; the Transformer model performs best, and the identification accuracy of the three states is more than 99.9%.

[0110] The t-SNE feature visualization chart shows the feature extraction and classification effect of the four models on the three state samples. Among them, the blue points represent the normal operation state, the green points represent the icing state, and the orange points represent the rain state. From the visualization results, it can be seen that all models have effectively separated the three state samples and formed a clear clustering structure. In particular, the samples of the icing state (green points) maintain a clear boundary interval from the other two states, which verifies that the monitoring method based on acoustic emission can reliably identify the icing state of the blade. The feature distribution of the Transformer model is the most compact, and the class boundary is the clearest, which corresponds to its optimal identification accuracy.

[0111] The four neural network models are verified by experimental data, and the average identification accuracy of the four deep learning models is more than 99.5%; among them, the Transformer model performs best, with an accuracy of 99.9%, and the standard deviation of the accuracy is less than 0.001, showing good stability; and each round of training only takes 0.09 seconds, with high computing efficiency. Through the comparative analysis of the confusion matrix and t-SNE visualization, the excellent performance of each deep learning model in the fan blade state identification task is intuitively displayed, verifying the effectiveness and reliability of the method.

[0112] The above neural network model is applied to a specific icing monitoring system, which includes an acoustic emission sensor and a signal processor; the acoustic emission sensor is a narrow frequency sensor with a size of Φ19x15mm, which is fixedly installed in the inside of the fan blade through a magnetic clamp, and the acoustic coupling is ensured through a special coupling agent during signal transmission. The acoustic emission sensor is provided with a data acquisition unit, the dynamic range is 70dB, the sampling accuracy is 16 bits, and the storage capacity is 64G.

[0113] The acoustic emission sensor is electrically connected with the signal transmission line and the information processor, the signal transmission line is a 75Ω impedance coaxial cable, a shielding layer is arranged in the coaxial cable to suppress electromagnetic interference, and the coaxial cable is connected with the acoustic emission sensor through a waterproof joint.

[0114] The wind turbine blade icing state monitoring method through the above device comprises the following steps:

[0115] S1, collect acoustic emission signals, and after signal collection, the signals need to be sequentially subjected to signal noise reduction and data normalization;

[0116] S2, performing singular geometric modal decomposition (SGMD) processing on the collected acoustic emission signals.

[0117] Through the singular geometric modal decomposition, the rain and icing state acoustic emission original signals are decomposed into six singular geometric components (SGC1-SGC6) and a residual term.

[0118] In order to verify the SGMD decomposition effect and determine the optimal reconstruction scheme, the acoustic emission signals in the icing state are taken as the research object, and the scatter diagram, signal superposition diagram and phase space Figure Three The characteristic relationship between each SGC component and the original signal is analyzed. Figures 6-8 is the demonstration process of the SGC1+SGC2 reconstruction scheme finally determined for the icing signal, Figure 9 is the performance of the SGC1+SGC2 reconstruction scheme finally determined for the icing signal, Figure 10 is the performance of the SGC1+SGC2 reconstruction scheme finally determined for the rain signal.

[0119] Specifically, firstly, the correlation of each SGC component reconstructed signal and the original signal is evaluated by scatter plot, the results show that SGC1 and SGC2 components have high correlation with the original signal (correlation coefficient R>0.90), and the correlation of each subsequent component starting from SGC3 is significantly reduced (correlation coefficient R<0.45); secondly, the degree of maintaining the amplitude characteristics of the reconstructed signal to the original signal is analyzed by signal superposition diagram, it is found that SGC1 and SGC2 components can maintain more than 85% of the amplitude characteristics of the original signal, and the amplitude maintenance rate of each subsequent component starting from SGC3 is less than 60%; finally, the phase space diagram is used to verify the degree of maintaining the dynamic characteristics of the reconstructed signal to the original signal, the results show that the phase space trajectory structure of SGC1 and SGC2 components is highly similar to the original signal (structure similarity>0.80), and the phase space trajectory of each subsequent component starting from SGC3 is significantly deformed. Based on the above analysis results, the application selects SGC1+SGC2 as the optimal reconstruction scheme, which not only realizes more than 95% correlation coefficient and more than 90% amplitude maintenance rate, but also shows good applicability under different working conditions, which can effectively reduce the influence of noise.

[0120] Specific analysis, the analysis results of SGC1 component ( Figure 5 ) show significant correlation characteristics: the scatter plot shows a highly consistent distribution pattern with the reference line (correlation coefficient R=0.92), the signal superposition diagram shows that the reconstructed signal effectively retains the main oscillation characteristics of the original signal (amplitude maintenance rate>85%), and the phase space reconstruction confirms the maintenance of the basic dynamic characteristics. Similarly, SGC2 component ( Figure 6 ) also shows excellent signal maintenance capability, which is particularly evident in phase space reconstruction, its trajectory structure is highly similar to the original signal (structure similarity index>0.80). However, starting from SGC3 ( Figure 7 ), the characteristic performance of the reconstructed signal is significantly degraded. The scatter plot shows that the deviation of the data points from the reference line increases (the correlation coefficient decreases to 0.45), the matching accuracy in the signal superposition diagram decreases (the amplitude maintenance rate is less than 60%), and the phase space trajectory structure changes significantly. This degradation is more prominent in SGC4 component ( Figure 8 ), the reconstructed signal is difficult to reflect the main characteristics of the original signal (correlation coefficient<0.30), and the phase space trajectory presents essential morphological differences.

[0121] Based on the above comprehensive analysis, the embodiment selects the combination of SGC1 and SGC2 components for signal reconstruction. As shown in Figure 9 and Figure 10As shown, the SGC1+SGC2 reconstruction scheme achieved the best performance on all evaluation indexes: the data points in the scatter plot were closely distributed around the reference line (correlation coefficient R=0.95), the signal superposition plot showed excellent waveform matching (amplitude retention rate >90%), and the phase space reconstruction verified the good preservation of the kinetic characteristics, including the shape and structural characteristics of the trajectory (structural similarity index >0.85).

[0122] Based on the principal symplectic geometric component, after superposition and reconstruction of the principal symplectic geometric component, more than 90% of the amplitude characteristics of the original signal can be maintained, the correlation coefficient is more than 0.95, and the influence of noise is effectively reduced.

[0123] The technical scheme of the present application successfully solves the many problems existing in the traditional monitoring method, realizes accurate identification and timely warning of the icing state of the fan blade, and has important engineering application value.

[0124] Through the above detailed description, the technical scheme of the present application is fully demonstrated, which not only solves the limitations of the traditional monitoring method, but also has good engineering practical value. Experimental results show that the method has high reliability, timely response and easy implementation, and provides an innovative technical solution for wind turbine blade icing monitoring.

[0125] The above is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring the icing status of wind turbine blades, characterized in that, The wind turbine blade icing status monitoring system includes an acoustic emission sensor (2) and an information processor. The acoustic emission sensor (2) is installed inside the outer wall of the wind turbine blade (1). The outer wall refers to the wall surface of the wind turbine blade (1) that is away from the shaft. The monitoring method includes the following steps: S1, Acquire acoustic emission signal; the acoustic emission signal (3) is generated during the operation of the blade, and the acoustic emission sensor (2) acquires the acoustic emission signal of the wind turbine blade (1); S2, perform symplectic geometric mode decomposition on the acoustic emission signal to obtain the main symplectic geometric components; S3, reconstructing the acoustic emission signal through the main symplectic geometric components; S4. Based on the reconstructed acoustic emission signal, the blade status is determined by a deep neural network; the blade status includes normal operation status, icing status, and rain status.

2. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that, In S1, the acoustic emission signal (3) is the stress change generated by ice covering the wind turbine blades (1) during the ice accumulation, ice stress and melting process.

3. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that, The main symplectic geometric components were obtained through experimental signal analysis. The analysis process is as follows: (1) Perform symplectic geometric mode decomposition on the acoustic emission signal obtained from the experiment to obtain several symplectic geometric components. Based on the symplectic geometric components and the original acoustic emission signal, establish a scatter plot, signal superposition plot and phase space plot; determine the correlation between each symplectic geometric component and the original acoustic emission signal through the scatter plot, signal superposition plot and phase space plot to obtain the main symplectic geometric components. (2) The acoustic emission signal is reconstructed by the symplectic geometric components of the optimal reconstruction scheme.

4. The method for monitoring the icing status of wind turbine blades according to claim 3, characterized in that, In step (1), the correlation between each symplectic geometric component and the original acoustic emission signal is determined by scatter plot. If the correlation is greater than the set threshold, the symplectic geometric component is the main symplectic geometric component.

5. A method for monitoring the icing status of wind turbine blades according to claim 3, characterized in that, In step (1), the retention rate of each symplectic geometric component to the amplitude characteristics of the original signal is determined by the signal overlay diagram. If the retention rate is greater than the set threshold, the symplectic geometric component is the main symplectic geometric component.

6. The method for monitoring the icing status of wind turbine blades according to claim 3, characterized in that, In step (1), the retention rate of each symplectic geometric component to the dynamic characteristics of the original signal is determined by the phase space diagram. If the retention rate is greater than the set threshold three, then the symplectic geometric component is the main symplectic geometric component.

7. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that, In S4, the deep neural network model can be any one of convolutional neural network, long short-term memory network, gated recurrent unit or Transformer neural network.

8. A method for monitoring the icing status of wind turbine blades according to claim 7, characterized in that, The deep neural network model is trained using training set data and then tested using test set data; the training set data and test set data are derived from wind turbine blade experiments.

9. A device for monitoring the icing status of wind turbine blades, characterized in that, The wind turbine blade icing status monitoring system includes an acoustic emission sensor (2) and an information processor. The acoustic emission sensor (2) is installed inside the outer wall of the wind turbine blade (1). The outer wall refers to the wall surface of the wind turbine blade (1) that is away from the shaft. The monitoring device includes: Acquisition unit, used to acquire acoustic emission signals; the acoustic emission signal (3) is generated during the operation of the blade, and the acoustic emission sensor (2) acquires the acoustic emission signal of the wind turbine blade (1); The decomposition unit is used to perform symplectic geometric mode decomposition on the acoustic emission signal to obtain the main symplectic geometric components; Reconstruction unit, used to reconstruct the acoustic emission signal through the main symplectic geometric components; The judgment unit is used to determine the blade state based on the reconstructed acoustic emission signal using a deep neural network; the blade state includes normal operation state, icing state, and rain state.

Images