Long flexible blade real-time monitoring and fault early warning method

Abstract

The application relates to the technical field of blade monitoring, in particular to a long flexible blade real-time monitoring and fault early warning method, which comprises the following steps: collecting vibration signals of a long flexible blade in a running state, wherein the vibration signals are coupled by multiple factors including aerodynamic load, gravity and centrifugal force; constructing an over-determined equation set containing multi-source excitation characteristics based on a blade dynamics model; decomposing the vibration signals by using a sparse decomposition algorithm to obtain signal decomposition results; solving the over-determined equation set based on the signal decomposition results to obtain modal parameters of each coupled vibration source; and determining a state monitoring result of the long flexible blade according to the modal parameters of each coupled vibration source.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blade monitoring, in particular to a long flexible blade real-time monitoring and fault early warning method. BACKGROUND

[0002] In industrial applications, long flexible blades are widely used in many fields such as wind turbines and aero-engines. The state of the long flexible blade directly affects the operating efficiency, safety and reliability of the entire device. Therefore, accurately monitoring the state of the long flexible blade is of great significance to ensure the stable operation of the device and timely maintenance.

[0003] The traditional monitoring method relies on manual inspection or single-point sensors, which has defects such as poor real-time performance and limited coverage. Therefore, it is urgent to improve. SUMMARY

[0004] Therefore, it is necessary to provide a long flexible blade real-time monitoring and fault early warning method capable of optimizing the sorting process in view of the above technical problems.

[0005] In a first aspect, the present application provides a long flexible blade real-time monitoring and fault early warning method applied to an edge server, which comprises:

[0006] Collecting vibration signals of the long flexible blade in a running state, the vibration signals being subjected to the coupling action of multiple factors such as aerodynamic load, gravity and centrifugal force;

[0007] Based on a blade dynamics model, an overdetermined equation set containing multi-source excitation characteristics is constructed;

[0008] A sparse decomposition algorithm is used to decompose the vibration signals to obtain signal decomposition results;

[0009] Based on the signal decomposition results, the overdetermined equation set is solved to obtain modal parameters of each coupled vibration source;

[0010] According to the modal parameters of each coupled vibration source, a state monitoring result of the long flexible blade is determined.

[0011] In one of the embodiments, the sparse decomposition algorithm is an orthogonal matching pursuit algorithm, and in the iteration process of the orthogonal matching pursuit algorithm, dictionary atoms related to the inherent modes of the blade are preferentially matched.

[0012] In one of the embodiments, the dictionary atoms of the orthogonal matching pursuit algorithm are generated in the following way:

[0013] Combining the Campbell diagram of the long flexible blade to preselect the resonance frequency band;

[0014] Based on the modal confidence factor, an atomic basis function matching the physical structure of the blade is screened.

[0015] In one of the embodiments, the blade dynamics model comprises:

[0016] a forced vibration term caused by aerodynamic load;

[0017] a static offset correction term caused by gravity;

[0018] a stiffness enhancement term caused by centrifugal force.

[0019] In one of the embodiments, the calculation of the static offset correction term comprises:

[0020] adjusting the direction of gravity decomposition according to the installation angle of the long and flexible blade, and introducing a material nonlinear coefficient to compensate for the change of geometric stiffness under large deformation.

[0021] In one of the embodiments, the construction of the overdetermined equation set specifically comprises:

[0022] obtaining prior information of blade modal shape through finite element simulation or experimental calibration;

[0023] embedding the prior information as a constraint condition into the equation set to obtain the overdetermined equation set.

[0024] In one of the embodiments, the prior information acquisition method comprises:

[0025] calibrating the first three order modal shapes of the blade through a laser vibration meter;

[0026] updating the coefficient matrix of the overdetermined equation set in real time by using working condition parameters.

[0027] In one of the embodiments, the vibration signal acquisition specifically comprises:

[0028] arranging three-axis acceleration sensors at the leading edge position, the trailing edge position and the mid-blade position of the long and flexible blade respectively to collect the vibration signals of the long and flexible blade;

[0029] Among them, the arrangement of three-axis acceleration sensors at the leading edge position, the trailing edge position and the mid-blade position respectively adopts an anti-aliasing filter and a synchronous sampling technology to make the phase consistency of multi-channel signals.

[0030] In one of the embodiments, according to the modal parameters of each coupled vibration source, the state monitoring result of the long and flexible blade is determined, comprising:

[0031] using a state recognition model locally deployed by an edge server to analyze the current signal characteristics of the long and flexible blade to obtain the current state corresponding to the long and flexible blade;

[0032] determining the load state analysis result of the long and flexible blade according to the current state;

[0033] determining the vibration state analysis result of the long and flexible blade according to the modal parameters of each coupled vibration source.

[0034] According to the load state analysis result and the vibration state analysis result, the state monitoring result of the long flexible blade is determined;

[0035] The state recognition model is trained according to historical signal features of the long flexible blade and historical states corresponding to the historical signal features.

[0036] The historical signal features include time-frequency domain features of the blade load signal and time domain features of the vibration signal, and aerodynamic change features of the long flexible blade in the historical state.

[0037] The state recognition model is a deep learning model including physical embedding, and a loss function of the state recognition model is constructed according to an aeroelastic equation.

[0038] The long flexible blade real-time monitoring and fault early warning method can collect vibration signals of the long flexible blade under the action of multiple factors such as aerodynamic load, gravity and centrifugal force in the running state, can comprehensively reflect the stress and vibration of the blade in the actual working environment, can provide accurate data basis for subsequent analysis, and is helpful for accurately capturing the vibration characteristics and potential fault information of the blade. The super-determined equation set containing multiple source excitation features is constructed based on the blade dynamics model, and the influence of multiple excitation factors on the vibration of the blade is comprehensively considered. This modeling method can more accurately describe the vibration behavior of the blade, convert complex physical phenomena into mathematical models, provide a theoretical framework for solving the modal parameters of each coupled vibration source, and improve the accuracy and reliability of the analysis.

[0039] The vibration signal is decomposed by using the sparse decomposition algorithm, the complex vibration signal can be decomposed into different components, the useful feature information can be extracted, and the noise and redundant information can be removed. This helps to more clearly understand the composition structure of the vibration signal, provides purer data for subsequent solving of the super-determined equation set and analysis of the modal parameters of each coupled vibration source, and improves the accuracy and efficiency of parameter solving. The modal parameters of each coupled vibration source are obtained by solving the super-determined equation set based on the signal decomposition result, the specific contribution of each excitation factor to the vibration of the blade can be determined, including the frequency, amplitude, phase and other key information of the vibration. These modal parameters are important basis for evaluating the state of the blade, and are helpful for deeply understanding the vibration characteristics and potential failure mechanism of the blade.

[0040] The state monitoring result of the long flexible blade is determined according to the modal parameters of each coupled vibration source, the working state of the blade can be comprehensively and accurately evaluated, and potential fault hidden dangers can be found in time. Through the analysis of the modal parameters, it can be judged whether the blade has abnormal vibration, fatigue damage and other problems, scientific basis is provided for the maintenance and repair of the blade, the reliability and service life of the blade are improved, the maintenance cost is reduced, and the safe and stable operation of the whole system is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application or the related art. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other related drawings can also be obtained from these drawings without creative labor.

[0042] Figure 1 A flowchart of a long flexible blade real-time monitoring and fault early warning method in an embodiment;

[0043] Figure 2 A structural block diagram of a long flexible blade state monitoring device in an embodiment. DETAILED DESCRIPTION

[0044] In order to make the objects, technical solutions and advantages of the present application more clear, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0045] In an exemplary embodiment, a long flexible blade real-time monitoring and fault early warning method is provided, which is applied to an edge server, such as Figure 1 As shown in the figure, the method comprises:

[0046] S101, collecting vibration signals of the long flexible blade in a running state.

[0047] Among them, the vibration signals are coupled by multiple factors of aerodynamic load, gravity and centrifugal force.

[0048] It can be understood that the state of the long flexible blade in actual operation will be reflected through the vibration signals. The vibration signals contain dynamic information of the blade under the action of various forces (aerodynamic load, gravity and centrifugal force, etc.), and through the collection of vibration signals, the basic data for analyzing the state of the blade can be obtained, which is the premise for subsequent fault early warning and state monitoring.

[0049] Principle: A suitable sensor (such as an acceleration sensor) is installed at a proper position of the long flexible blade, the sensor can convert the mechanical vibration of the blade into electrical signals, the characteristics (such as amplitude, frequency, etc.) of these electrical signals correspond to the vibration condition of the blade, so as to realize the collection of vibration signals.

[0050] Implementation process: According to the structure and operating characteristics of long flexible blades, determine the installation position of the sensor to ensure that the signal reflecting the overall vibration of the blade can be accurately collected. Install and debug the sensor to ensure stable operation and convert the vibration of the blade into measurable electrical signals. Use appropriate data acquisition equipment to collect the electrical signals output by the sensor at a certain sampling frequency and store them in a format that can be used for subsequent analysis.

[0051] S102, based on the blade dynamics model, construct an overdetermined equation set containing multi-source excitation characteristics.

[0052] It can be understood that the vibration of the blade is the result of the joint action of multiple excitation factors (aerodynamic load, gravity, centrifugal force, etc.). By constructing an overdetermined equation set containing multi-source excitation characteristics, the vibration phenomenon of the blade can be described by a mathematical model, providing a theoretical basis for subsequent analysis of the influence of each excitation factor on the vibration of the blade, so as to more accurately determine the state of the blade.

[0053] Implementation principle: Based on the basic theory of blade dynamics, consider the action mode and mutual relationship of each excitation factor on the vibration of the blade, and deduce and establish the vibration equation of the blade. Due to the existence of multiple excitation sources and unknown parameters, an overdetermined equation set is formed, and mathematical tools are used to describe the complex dynamic characteristics of the blade vibration.

[0054] Implementation process: In-depth study of the structure, material properties and operating environment of the blade, establish an accurate blade dynamics model, consider the contribution of each excitation factor to the vibration of the blade. According to the dynamics model, each excitation factor is converted into a mathematical expression, and combined with the vibration equation of the blade, an overdetermined equation set containing multiple unknown parameters is constructed. Verify the reasonableness of the constructed equation set to ensure that it can accurately reflect the vibration characteristics of the blade under multi-source excitation.

[0055] S103, decompose the vibration signal using sparse decomposition algorithm to obtain signal decomposition result.

[0056] It can be understood that the collected vibration signal is a complex mixed signal containing vibration information of multiple frequency components and different excitation sources. Through the sparse decomposition algorithm, the mixed signal can be decomposed into simpler components with specific characteristics, so that the contribution of each coupled vibration source can be more clearly analyzed, and the modal parameters of each vibration source can be accurately obtained.

[0057] Implementation principle: Sparse decomposition algorithm is based on the sparse representation theory of signal, assuming that the vibration signal can be represented by a small number of atoms in an over-complete dictionary. Through optimization algorithm, find the most suitable atom combination to decompose the original vibration signal into multiple sparse components, each component corresponds to a specific vibration mode or the contribution of an excitation source.

[0058] Implementation process: Select a suitable overcomplete dictionary that can better represent the various characteristics of the blade vibration signal. For example, dictionaries based on Fourier transform, wavelet transform, etc. can be used. Use sparse decomposition algorithms (such as matching pursuit algorithm, etc.) to search for the most matching atomic combination in the overcomplete dictionary with the original vibration signal, and gradually decompose the signal into multiple sparse components. Screen and process the decomposed components to remove noise and other interference factors to obtain accurate signal decomposition results.

[0059] S104, based on the signal decomposition results, solve the overdetermined equation set to obtain the modal parameters of each coupled vibration source.

[0060] It can be understood that the contribution information of each vibration source is obtained through signal decomposition, and the overdetermined equation set describes the relationship between each excitation factor and the blade vibration. By combining the two, solving the overdetermined equation set can obtain the modal parameters (such as frequency, amplitude, phase, etc.) of each coupled vibration source, which can quantitatively describe the influence of each vibration source on the blade vibration and provide key data for blade condition monitoring.

[0061] Implementation principle: Substitute the characteristic information of each component obtained from the signal decomposition results into the overdetermined equation set, and use numerical calculation methods (such as least squares method, etc.) to solve the equation set, so that the solution of the equation set can satisfy the vibration characteristics obtained by signal decomposition, thereby determining the modal parameters of each coupled vibration source.

[0062] Implementation process: Organize and analyze the signal decomposition results, and extract the characteristic information related to the unknown parameters in the overdetermined equation set. Select the appropriate numerical calculation method, substitute the extracted characteristic information into the overdetermined equation set for solving. During the solving process, iterative calculation may be required to gradually approach the accurate solution. Verify and evaluate the modal parameters of each coupled vibration source obtained by solving to ensure their accuracy and reliability.

[0063] S105, according to the modal parameters of each coupled vibration source, determine the condition monitoring results of the long and flexible blade.

[0064] It can be understood that the modal parameters of each coupled vibration source reflect the vibration state of the blade under the action of different excitation factors. By analyzing these parameters, it can be determined whether the blade is in normal operating state, whether there is a potential fault or abnormal situation. According to the modal parameters to determine the condition monitoring results, the problems of the blade can be found in time, and the basis for fault warning and maintenance decision is provided.

[0065] Implementation principle: Establish the standard range or reference value of the modal parameters of each coupled vibration source under the normal operating state of the blade in advance. Compare the currently calculated modal parameters with the standard range, and determine the state of the blade according to the deviation degree and change trend of the parameters. For example, if the frequency or amplitude of a certain vibration source exceeds the normal range, it may indicate that the blade has a corresponding fault or abnormality.

[0066] Implementation process: Establish and maintain a database of modal parameters of each coupled vibration source under the normal operating state of the blade, including the standard range or reference value. Compare the calculated modal parameters of each coupled vibration source with the standard values in the database. According to the comparison results, determine the state monitoring results of the blade according to the predetermined rules and algorithms, such as normal, warning, fault, etc., and output the corresponding result information.

[0067] In an exemplary embodiment, the sparse decomposition algorithm is the orthogonal matching pursuit algorithm, and the dictionary atoms related to the blade natural modes are preferentially matched in the iteration process of the orthogonal matching pursuit algorithm.

[0068] It can be understood that in the analysis of long and flexible blade vibration signals, the natural modes of the blade contain key information about its structure and dynamics. Preferentially matching dictionary atoms related to the blade natural modes can more accurately extract components closely related to the blade's own characteristics from complex vibration signals. This helps to more accurately analyze the vibration state of the blade and identify potential faults or abnormal conditions, because the faults of the blade are often reflected in the changes of its natural modes.

[0069] Implementation principle: The orthogonal matching pursuit algorithm (OMP) is an iterative algorithm, whose core idea is to select the atom that best matches the current residual from the over-complete dictionary in each iteration, add it to the selected atom set, and update the residual until the stopping condition is met. In this method, the dictionary atoms related to the blade natural modes are preferentially selected, based on the understanding that the natural modes of the blade determine its vibration response at a certain frequency. By preferentially matching these atoms, the characteristic vibration information of the blade can be more effectively captured, thereby better analyzing its state.

[0070] Implementation process: Determine the blade inherent modal information: Through theoretical analysis (such as finite element analysis, etc.) or experimental testing (such as modal testing, etc.), obtain the inherent modal information of the long flexible blade, including natural frequency, mode shape, etc. Construct an over-complete dictionary: Select appropriate atomic basis functions to construct an over-complete dictionary, which should be able to cover various vibration modes that may occur in the blade. Iterative matching: In each iteration of the orthogonal matching pursuit algorithm, evaluate the atoms in the dictionary, and preferentially select atoms related to the blade inherent modal. Specifically, by calculating the similarity of the atom and the known inherent modal information (such as natural frequency), the atom with the highest similarity is selected to join the selected atom set. Update the residual: After the selected atom is added to the selected atom set, update the residual and continue the next iteration until the stopping condition is met (such as the residual is less than a certain threshold or the preset number of iterations is reached).

[0071] The dictionary atoms of the orthogonal matching pursuit algorithm are generated in the following way: combining the Campbell diagram of the long flexible blade to pre-select the resonance frequency band, and selecting the atomic basis function that matches the physical structure of the blade based on the modal confidence factor.

[0072] Generation method of dictionary atoms of orthogonal matching pursuit algorithm

[0073] Pre-select the resonance frequency band in combination with the blade Campbell diagram

[0074] It can be understood that the Campbell diagram of the blade shows the distribution of the resonance frequency of the blade at different rotating speeds. During the operation of the long flexible blade, resonance is a key problem, which may cause the vibration of the blade to intensify, and even cause failure. By pre-selecting the resonance frequency band in combination with the Campbell diagram, the generation range of the dictionary atoms can be focused on the frequency area where resonance may occur, improving the algorithm's ability to capture resonance-related vibration information, thus more effectively monitoring the operation state of the blade.

[0075] Implementation principle: The Campbell diagram is drawn based on the dynamics characteristics and operating conditions of the blade, which reflects the relationship between the natural frequency and the excitation frequency of the blade at different rotating speeds. When the excitation frequency is close to the natural frequency of the blade, resonance occurs. Therefore, by analyzing the Campbell diagram, the frequency range that may resonate under different operating conditions, i.e. the resonance frequency band, can be determined. When generating dictionary atoms, preferentially considering the frequency components within these resonance frequency bands can more accurately capture the vibration characteristics of the blade in the resonance state.

[0076] Implementation process: Draw the Campbell chart of the blade: Use the dynamic model of the blade and the operating parameters to draw the Campbell chart of the blade through numerical calculation or experimental test, etc. Determine the resonance frequency band: According to the actual operating speed range of the blade, find out the frequency area that may appear resonance in the Campbell chart, and determine the resonance frequency band. Screen the frequency range of the dictionary atom: When generating the dictionary atom, limit the frequency range of the atom within the pre-selected resonance frequency band, and ensure that the dictionary atom can effectively represent the vibration information of the blade in the resonance state.

[0077] Screening of atom basis functions matching the physical structure of the blade based on modal assurance criterion

[0078] It can be understood that the physical structure of the blade determines its vibration characteristics, and different physical structures will produce different vibration modes. The modal assurance criterion (MAC) is an index for evaluating the similarity between two modal vectors. By screening atom basis functions matching the physical structure of the blade based on the modal assurance criterion, it can be ensured that the atoms in the dictionary can accurately reflect the actual vibration modes of the blade, improve the accuracy and reliability of signal decomposition, and thus more effectively monitor the state of the blade.

[0079] Implementation principle: The modal assurance criterion evaluates the similarity of two modal vectors by calculating their correlation. In this method, the atom basis function is regarded as a kind of modal vector, and by calculating the modal assurance criterion between the atom basis function and the actual modal vector of the blade, the atom basis function with high matching degree to the physical structure of the blade is screened out. This can ensure that the atoms in the dictionary can better represent the actual vibration of the blade, thereby improving the performance of the orthogonal matching pursuit algorithm.

[0080] Implementation process: Obtain the actual modal vector of the blade: Through experimental test (such as modal test, etc.) or theoretical analysis (such as finite element analysis, etc.), obtain the actual modal vector of the long and flexible blade, which reflects the displacement distribution of the blade under different vibration modes.

[0081] Calculate the modal assurance criterion: For each candidate atom basis function, calculate its modal assurance criterion with the actual modal vector of the blade. The calculation of the modal assurance criterion is usually based on the inner product and the modulus of the vector.

[0082] Screen the atom basis function: According to the calculated modal assurance criterion, set a threshold value, and screen out the atom basis functions with a modal assurance criterion greater than the threshold value, and take these atom basis functions as the final dictionary atoms. This can ensure that the atoms in the dictionary can accurately reflect the physical structure and vibration characteristics of the blade.

[0083] In an exemplary embodiment, the blade dynamics model includes: an aerodynamic load-induced forced vibration term, a gravity-induced static deflection correction term, and a centrifugal force-induced stiffness enhancement term.

[0084] The calculation of the static deflection correction term includes dynamically adjusting the direction of gravity decomposition according to the installation angle of the long and flexible blade, and introducing a material nonlinearity coefficient to compensate for the change in geometric stiffness under large deformation.

[0085] It can be understood that the blade dynamics model takes into account the effects of aerodynamic load, gravity and centrifugal force on the vibration of the long and flexible blade, which is analyzed as follows:

[0086] Aerodynamic load-induced forced vibration term

[0087] It can be understood that the aerodynamic load is one of the main external excitation sources of the long and flexible blade during operation. The airflow acting on the blade will produce periodic forces, which will cause the blade to produce forced vibration. Considering the forced vibration term caused by aerodynamic load, the vibration of the blade in actual operation can be more accurately described, which is crucial for analyzing the dynamic response of the blade and monitoring its state.

[0088] Principle: Based on the principle of aerodynamics, the interaction between airflow and blade surface produces a pressure difference, thereby forming aerodynamic load. The size and direction of this load are related to factors such as airflow speed, blade shape, angle of attack, etc. Through aerodynamic theory and relevant experimental data, a relationship model between aerodynamic load and blade vibration is established, expressing the aerodynamic load as a function related to the motion state of the blade, thereby obtaining the forced vibration term.

[0089] Implementation process: First, determine the flow characteristics of the airflow, such as speed, pressure distribution, etc., according to the geometry of the blade and the operating environment. Then, use computational fluid dynamics (CFD) methods or empirical formula-based calculation methods to calculate the aerodynamic load acting on the blade. Finally, convert the aerodynamic load into the forced vibration term in the blade vibration equation, usually expressed as a function related to the displacement, velocity, etc. of the blade.

[0090] Gravity-induced static deflection correction term

[0091] It can be understood that gravity will cause the long and flexible blade to produce static deflection, affecting the balance position and vibration characteristics of the blade. Especially in the case of large deformation of the blade, the effect of gravity cannot be simply regarded as a constant force, and the change in the direction of decomposition and the change in geometric stiffness need to be considered. Therefore, setting the gravity-induced static deflection correction term helps to more accurately describe the mechanical behavior of the blade under the action of gravity and improve the accuracy of the model.

[0092] Implementation principle: The size and direction of gravity are fixed, but during the motion of the blade, due to the dynamic change of its installation angle, the decomposition direction of gravity in the blade coordinate system will change. At the same time, when the blade deforms greatly, its geometry will change, causing the stiffness of the material to change, so a material nonlinear coefficient needs to be introduced to compensate for this change in geometric stiffness. By considering these factors, the static deflection caused by gravity is corrected to make it more consistent with the actual stress of the blade.

[0093] Implementation process: For the adjustment of the decomposition direction of gravity, according to the relationship between the installation angle of the blade and time, the gravity vector is decomposed in the local coordinate system of the blade to obtain the components of gravity in each direction at different times. For the compensation of geometric stiffness change, the nonlinear coefficient of the material is determined through experiment or theoretical analysis, which reflects the stiffness change characteristics of the material under large deformation. Then, the gravity component and the material nonlinear coefficient are combined to obtain the static deflection correction term caused by gravity, which is used to correct the gravity term in the blade dynamics model.

[0094] Stiffness enhancement term caused by centrifugal force

[0095] It can be understood that the long and flexible blade will be subjected to centrifugal force during rotation, which will cause the blade to stretch and deform, thereby increasing its stiffness. This stiffness enhancement effect will affect the vibration frequency and modal characteristics of the blade, and must be considered when analyzing the dynamics of the blade. Setting the stiffness enhancement term caused by centrifugal force can more accurately describe the mechanical properties of the blade in the rotating state, and provide a more reliable theoretical basis for the state monitoring and fault warning of the blade.

[0096] Implementation principle: According to the calculation formula of centrifugal force, centrifugal force is related to the rotation speed of the blade, mass distribution and distance to the rotation axis. When the blade is subjected to centrifugal force, stress will be generated inside the blade, which will cause the material of the blade to stretch and deform, thereby increasing the stiffness of the blade. By establishing a relationship model between centrifugal force and blade stiffness change, the stiffness enhancement effect caused by centrifugal force is expressed as a function related to the rotation speed of the blade, geometric parameters, etc., thereby obtaining the stiffness enhancement term.

[0097] Implementation process: First, determine the mass distribution and geometric parameters of the blade, as well as the position of the rotation axis. Then, according to the rotation speed, calculate the centrifugal force acting on each point of the blade. Next, using the theory of material mechanics and elasticity, analyze the deformation and stress distribution of the blade under the action of centrifugal force, and establish the relationship between centrifugal force and blade stiffness change. Finally, this relationship is converted into a stiffness enhancement term in the blade dynamics model, which is usually expressed in the form of being proportional to the square of the rotation speed, and is combined with the stiffness matrix of the blade to correct the dynamics equation of the blade.

[0098] Analysis of steps in static offset correction term calculation: dynamically adjust the direction of gravity decomposition according to the installation angle of long flexible blades.

[0099] It can be understood that the installation angle of long flexible blades will dynamically change with the movement of the blades and changes in the external environment during operation. Gravity is a constant force in a certain direction, but its decomposition direction in the local coordinate system of the blade will change with the change of the installation angle. If the dynamic change of the installation angle is not considered, the actual effect of gravity on the blade at different positions and attitudes cannot be accurately described, which will affect the calculation accuracy of the static offset of the blade. Therefore, it is necessary to dynamically adjust the direction of gravity decomposition according to the installation angle, which can more truly reflect the mechanical behavior of the blade under the action of gravity.

[0100] Principle of implementation: use trigonometric functions to decompose the gravity vector in the coordinate system with the blade as the reference. The installation angle determines the angle between the gravity vector and each coordinate axis of the blade coordinate system, and through these angles, the components of gravity in each direction of the blade coordinate system can be calculated. With the change of the installation angle, these angles will also change accordingly, so as to realize the dynamic adjustment of the direction of gravity decomposition.

[0101] Implementation process: real-time monitoring or calculating the change rule of the installation angle of long flexible blades with time through kinematic model. Then, according to the specific value of the installation angle, use the trigonometric function formula to calculate the components of gravity in each direction of the blade coordinate system. For example, in the two-dimensional case, if the installation angle is (theta) and the gravity is G, then the gravity components in the tangential and normal directions of the blade are (Gsintheta) and (Gcostheta), respectively. In the three-dimensional case, more complex coordinate transformation and trigonometric function calculation are needed to determine the components of gravity in three coordinate axis directions to realize the dynamic adjustment of the direction of gravity decomposition.

[0102] Introduce material nonlinear coefficient to compensate for the change of geometric stiffness under large deformation:

[0103] It can be understood that when the long flexible blade deforms greatly, its geometric shape will change significantly, and the mechanical properties of the material will also exhibit nonlinear characteristics, resulting in that the stiffness of the blade is no longer a constant, but is related to the deformation degree. The traditional linear model cannot accurately describe the change of stiffness under this condition, which will cause large deviation between the calculation results and the actual situation. The introduction of material nonlinear coefficient can consider the nonlinear behavior of the material under large deformation, compensate for the change of geometric stiffness, and thus improve the description accuracy of the model for the mechanical behavior of the blade under large deformation.

[0104] Implementation principle: The material nonlinear coefficient is obtained through experimental research or theoretical analysis, which reflects the stiffness variation law of the material under different deformation degrees. During the large deformation process of the blade, the stiffness of the blade is corrected using the material nonlinear coefficient according to the deformation size. Generally, the material nonlinear coefficient is a function related to deformation, and the coefficient gradually decreases with the increase of deformation, to reflect the characteristic that the stiffness decreases with the increase of deformation. By multiplying the material nonlinear coefficient with the elastic stiffness matrix of the blade, the stiffness matrix considering the change of geometric stiffness can be obtained, so as to more accurately describe the mechanical behavior of the blade under large deformation.

[0105] Implementation process: First, the relationship function between the material nonlinear coefficient and the deformation is determined through experimental test or theoretical analysis. The mechanical properties of the blade material under different deformation conditions can be tested by tensile test, bending test, etc., to obtain the stress-strain curve of the material, and then the expression of the material nonlinear coefficient is fitted. Then, when calculating the static deflection of the blade, the corresponding deformation is calculated according to the current deformation state of the blade, and the material nonlinear coefficient under the current deformation is obtained by substituting the expression of the material nonlinear coefficient. Finally, the coefficient is multiplied with the initial elastic stiffness matrix of the blade to obtain the corrected stiffness matrix, which is used to calculate the static deflection correction term caused by gravity, to consider the influence of the change of geometric stiffness under large deformation on the static deflection of the blade.

[0106] In an exemplary embodiment, the construction of the overdetermined equation set specifically includes: obtaining prior information of blade modal shapes through finite element simulation or experimental calibration, and embedding the prior information as a constraint condition into the equation set to obtain the overdetermined equation set.

[0107] Wherein, the acquisition method of prior information includes: calibrating the first three order modal shapes of the blade by laser vibration meter, and updating the coefficient matrix of the overdetermined equation set in real time by using working condition parameters.

[0108] Analysis of overdetermined equation set construction steps: obtain prior information of blade modal shapes through finite element simulation or experimental calibration.

[0109] It can be understood that the blade modal shape reflects the vibration form of the blade under different vibration frequencies, which is an important embodiment of the dynamic characteristics of the blade. Obtaining these prior information can provide key data support for subsequent construction of overdetermined equation set. Finite element simulation can simulate the mechanical behavior of the blade by computer, and predict the modal shape of the blade from the theoretical level; while experimental calibration obtains the real modal shape data of the blade through actual measurement. The two methods complement each other, ensure that the obtained prior information is comprehensive and accurate, and help to improve the reliability of the overdetermined equation set, and then more accurately analyze the vibration state of the blade.

[0110] Implementation principle: Finite element simulation: Based on the finite element theory, the blade is discretized into multiple finite-sized elements, and the mechanical model of the entire blade is established by analyzing the mechanics of each element. In the model, factors such as the material properties, geometric shape, and boundary conditions of the blade are considered, and numerical calculation methods are used to solve the vibration equation to obtain the modal shape of the blade.

[0111] Experimental calibration: Use specific measurement equipment, such as a laser vibration meter, to test the vibration of the actual blade. The laser vibration meter obtains the vibration information of each point on the blade surface by emitting a laser beam and measuring the frequency change of the reflected light. By arranging multiple measurement points on the blade surface, vibration data at different positions are collected, and through data processing and analysis, the modal shape of the blade is determined.

[0112] Implementation process: Finite element simulation: Establish a three-dimensional geometric model of the blade to accurately describe its shape and size. Define the material properties of the blade, including the elastic modulus, Poisson's ratio, etc. Divide the blade into a suitable number and size of elements through meshing. Set the boundary conditions to simulate the installation and constraints of the blade in actual work. Choose appropriate solvers and calculation parameters, run the finite element analysis software, solve the vibration equation, and obtain the modal shape data of the blade.

[0113] Experimental calibration: Prepare experimental equipment, including a laser vibration meter, a signal acquisition system, etc., and debug and calibrate them. Install the blade on the experimental table to simulate its actual working state, set appropriate excitation methods to make the blade vibrate, and evenly arrange multiple measurement points on the blade surface and mark them with reflective materials to improve the measurement accuracy of the laser vibration meter. Start the laser vibration meter, collect vibration data at each measurement point, and record the vibration response of the blade at different frequencies. Process and analyze the collected data, and extract the modal shape of the blade through specific algorithms such as modal parameter identification algorithms.

[0114] (1) Embed the prior information as a constraint condition into the equation set to obtain an overdetermined equation set:

[0115] It can be understood that simply constructing an equation set to describe the vibration of the blade may have multiple solutions, resulting in an inaccurate determination of the vibration state of the blade. The prior information of the blade modal shape contains the inherent characteristics of the blade vibration, which is embedded as a constraint condition into the equation set, limiting the range of solutions and making the equation set have a unique solution or a solution more consistent with the actual situation. The overdetermined equation set obtained in this way can more accurately reflect the vibration relationship of the blade under multi-source excitation, providing a reliable mathematical model for subsequent solving of the modal parameters of each coupled vibration source.

[0116] Implementation principle: When constructing the blade vibration equation set, the modal shape information obtained through finite element simulation or experimental calibration is converted into mathematical constraint conditions. These constraint conditions can be the value relationship of blade displacement, velocity, or acceleration at specific locations and frequencies. Combine these constraint conditions with the original equation set to form an over-determined equation set, and solve the equation set using mathematical optimization methods to make the solution not only satisfy the vibration relationship described by the original equation set, but also comply with the constraints imposed by the modal shape prior information.

[0117] Implementation process: Analyze the obtained blade modal shape prior information and determine the key features that can be used to construct constraint conditions, such as vibration amplitude, phase at specific locations, etc.

[0118] Convert these features into mathematical expressions, for example, if the vibration amplitude at a certain location under a certain modal is known as A, then the constraint equation can be constructed: (u(x,y,z,t) = A) (where u is the displacement function, (x,y,z) is the position coordinate, and t is the time). Combine these constraint equations with the original equation set that describes the blade vibration to form an over-determined equation set. Select an appropriate numerical solution method, such as the least squares method, to solve the over-determined equation set and obtain the solution that satisfies the constraint conditions, i.e., the modal parameters of each coupled vibration source.

[0119] (2) Real-time update of the coefficient matrix of the over-determined equation set using operating parameters:

[0120] It can be understood that in the actual running process of a long and flexible blade, operating parameters such as rotational speed, load, and environmental temperature will change continuously. Changes in these parameters will affect the aerodynamic load, centrifugal force, and other excitation factors experienced by the blade, thereby changing the dynamics of the blade. Therefore, the coefficient matrix of the over-determined equation set should also change. By using operating parameters to update the coefficient matrix in real time, the over-determined equation set can always accurately describe the vibration relationship of the blade under the current operating conditions, improving the real-time and accuracy of blade vibration state monitoring and fault warning.

[0121] Implementation principle: Establish a mathematical relationship model between operating parameters and the coefficients of the over-determined equation set. This model is based on blade dynamics theory and experimental data, and analyzes the influence of different operating parameters on the excitation force experienced by the blade and the dynamics parameters such as stiffness and damping. These influences are converted into the change rules of each element in the coefficient matrix. When the operating parameters change, according to the established mathematical relationship model, the element values in the coefficient matrix are calculated and updated in real time, so that the over-determined equation set can adapt to the current operating conditions of the blade.

[0122] Implementation process: Collect a large amount of blade operation data under different working conditions, including working condition parameters (rotation speed, load, ambient temperature, etc.) and corresponding vibration response data. Based on the blade dynamics theory, a mathematical model is established between the working condition parameters and the coefficient matrix of the overdetermined equation system. Regression analysis, neural network, etc. can be used to train and fit the collected data to determine the parameters of the model. During the operation of the blade, real-time monitoring of the changes in the working condition parameters is carried out, and the rotation speed, load, ambient temperature, etc. are collected by sensors. The real-time collected working condition parameters are input into the established mathematical model to calculate the updated values of each element in the coefficient matrix of the overdetermined equation system. According to the calculated updated values, the coefficient matrix of the overdetermined equation system is updated in real time to ensure that the equation system can accurately reflect the vibration characteristics of the blade under the current working condition.

[0123] In an exemplary embodiment, the collection of the vibration signal specifically includes: arranging three-axis acceleration sensors at the leading edge position, the trailing edge position and the mid-blade position of the long and flexible blade respectively, and collecting the vibration signal of the long and flexible blade.

[0124] Among them, arranging three-axis acceleration sensors at the leading edge position, the trailing edge position and the mid-blade position respectively uses anti-aliasing filter and synchronous sampling technology to make the phase consistency of multi-channel signals.

[0125] (1) Arranging three-axis acceleration sensors at the leading edge position, the trailing edge position and the mid-blade position of the long and flexible blade respectively.

[0126] It can be understood that the vibration characteristics of the long and flexible blade at different positions are different. The vibration at the leading edge position directly bears the impact of the airflow, and its vibration reflects the influence of the aerodynamic load; the vibration at the trailing edge position is related to the factors such as blade wake; the vibration at the mid-blade position comprehensively reflects the bending, torsion and other deformation conditions of the whole blade. Arranging three-axis acceleration sensors at the three key positions can comprehensively obtain the vibration information of the blade in different directions (x, y, z axis directions), reflect the running state of the blade from multiple dimensions, and provide rich data support for subsequent accurate analysis of the dynamics characteristics and fault diagnosis of the blade.

[0127] Implementation principle: The three-axis acceleration sensor can sense the acceleration changes in three perpendicular directions at its position. When the blade vibrates, the sensor moves with the blade, and the internal sensitive element will produce corresponding electrical signal changes according to the acceleration. By measuring the amplitude and frequency of these electrical signal changes, the acceleration information of the blade in each direction can be converted to describe the vibration condition of the blade at this position.

[0128] Implementation process: Sensor selection: According to the working environment of long flexible blades (such as temperature, humidity, vibration intensity range, etc.) and the measurement accuracy requirements, select the appropriate type of three-axis acceleration sensor to ensure its sufficient sensitivity and frequency response range. Position determination: Use the design drawings and actual installation of the blade to accurately determine the leading edge, trailing edge and mid-blade positions. For the leading edge position, generally select a suitable point on the contour line at the front end of the blade; the trailing edge position is determined on the contour line at the tail end of the blade; the mid-blade position can be determined by measuring the length of the blade and taking the midpoint position. Installation and fixation: Use appropriate installation methods, such as using bolts or adhesives, to firmly install the three-axis acceleration sensor at the predetermined position. When installing, make sure that the coordinate axis direction of the sensor is consistent with the reference coordinate axis direction of the blade, so as to accurately measure the acceleration in each direction.

[0129] (2) The leading edge position, trailing edge position and mid-blade position are respectively arranged with three-axis acceleration sensors Anti-aliasing filter and synchronous sampling technology are used to make the phase consistency of multi-channel signals.

[0130] It can be understood that during the vibration signal collection process, if not handled, the actual signal may contain high-frequency noise and other components, which will produce aliasing phenomenon during sampling, resulting in distortion of the sampled signal and inability to accurately reflect the characteristics of the original signal. Anti-aliasing filter is used to filter out high-frequency noise and ensure the authenticity of the sampled signal. Synchronous sampling technology is crucial for multi-channel signal collection, because there is a phase relationship between the vibration signals at different positions of the long flexible blade. Only by ensuring synchronous sampling of each channel signal can the phase difference and mutual relationship between signals be accurately analyzed, and the overall vibration state of the blade be accurately evaluated. If the channels are not sampled synchronously, phase errors will be introduced, affecting the accuracy of subsequent analysis of blade vibration characteristics and fault diagnosis.

[0131] Implementation principle: Anti-aliasing filter: Anti-aliasing filter is a low-pass filter whose cutoff frequency is set below half of the sampling frequency (Nyquist frequency). Its working principle is to allow low-frequency signals to pass through while attenuating high-frequency signals. When the vibration signal containing high-frequency noise passes through the anti-aliasing filter, the high-frequency noise component is greatly weakened, and only the low-frequency signal that truly reflects the blade vibration can pass through, thereby avoiding the aliasing phenomenon in the sampling process.

[0132] Synchronous sampling technique: Synchronous sampling technique makes the sampling time of each channel strictly consistent through some synchronization mechanism. The common implementation methods are hardware synchronization and software synchronization. Hardware synchronization usually uses a special synchronization clock circuit to provide a unified clock signal for the sampling devices of each channel, ensuring that they sample at the same time. Software synchronization adjusts the time of each channel before data collection through algorithms and program control, so that they remain time synchronized during the collection process. In this way, the consistency of multi-channel signals in time is ensured, so that the phase relationship between signals can be accurately reflected.

[0133] Implementation process: Anti-aliasing filter setting: Filter selection: According to the frequency range of the vibration signal and the sampling frequency, select the appropriate type and parameter of the anti-aliasing filter, such as Butterworth low-pass filter, etc. The cutoff frequency of the filter should be reasonably determined according to the Nyquist criterion to ensure that high-frequency noise can be effectively filtered out without affecting the useful signal. Filter connection: Connect the anti-aliasing filter between the signal output end of each three-axis acceleration sensor and the sampling device to ensure that the signal is filtered before entering the sampling device.

[0134] Synchronous sampling technique implementation: Hardware synchronization: Synchronous clock circuit construction: Design and build a high-precision synchronous clock circuit that can generate stable and accurate clock signals. Clock signal distribution: Distribute the synchronous clock signal to each channel's sampling device (such as a data acquisition card) through a special signal transmission line to ensure that each sampling device can receive the same clock signal, thus achieving synchronous sampling.

[0135] Software synchronization: Time calibration algorithm design: Write a time calibration algorithm to calibrate the sampling devices of each channel before data collection. The algorithm can send a specific calibration signal, measure the time delay of the signal in each channel, and then adjust the sampling start time of each channel according to the delay time.

[0136] Synchronous sampling control program writing: Develop a synchronous sampling control program to accurately control the sampling time of each channel's sampling device during data collection according to the calibrated time, ensuring the synchronous collection of multi-channel signals. During the collection process, the sampling state of each channel needs to be monitored in real time, and any abnormalities need to be adjusted and corrected in a timely manner.

[0137] In an exemplary embodiment, the state monitoring result of the long flexible blade is determined according to the modal parameters of each coupled vibration source, including:

[0138] The state recognition model deployed locally on the edge server is used to analyze the current signal characteristics of the long flexible blade, and the current state corresponding to the long flexible blade is obtained; according to the current state, the load state analysis result of the long flexible blade is determined; according to the modal parameters of each coupled vibration source, the vibration state analysis result of the long flexible blade is determined; according to the load state analysis result and the vibration state analysis result, the state monitoring result of the long flexible blade is determined.

[0139] It can be understood that the vibration signal characteristics of the long flexible blade are the external manifestations of its operating state, but these characteristics are often complex and difficult to directly interpret. The state recognition model can extract valuable information from complex signal characteristics and convert it into an intuitive state description after being trained with a large amount of data. Deploying the model locally on the edge server can reduce data transmission delay, realize real-time analysis, quickly respond to changes in blade state, and timely detect potential problems.

[0140] Principle: The state recognition model is usually constructed based on machine learning or deep learning algorithms, such as neural networks, support vector machines, etc. In the training phase, the model learns the mapping relationship between vibration signal characteristics and corresponding states of a large number of known states, establishing a feature-state association model. In actual application, the current signal characteristics of the long flexible blade are input into the trained model, and the model analyzes and judges the input characteristics according to the learned mapping relationship, and outputs the current state of the blade, such as normal, abnormal, etc.

[0141] Implementation process: Data collection and preprocessing: Collect vibration signal data of the long flexible blade under different operating states, and perform preprocessing operations such as cleaning and filtering on these data to extract key features that can reflect the state of the blade, such as frequency, amplitude, etc. Model training: Select appropriate machine learning or deep learning algorithms and use preprocessed data to train the model. By continuously adjusting the parameters of the model, the model can accurately map signal characteristics to corresponding states. Model deployment: Deploy the trained state recognition model on the edge server to ensure that the model can run normally. Real-time analysis: Input the current signal characteristics of the long flexible blade into the state recognition model deployed on the edge server, and the model analyzes and outputs the current state of the blade. According to the current state, the load state analysis result of the long flexible blade is determined.

[0142] It can be understood that the current state of the long flexible blade is closely related to the load it bears. Different load states will cause different vibration responses of the blade, which will affect its state. By analyzing the current state of the blade, its load state can be inferred, and whether the blade bears abnormal load can be understood, providing an important basis for subsequent fault diagnosis and maintenance decisions.

[0143] Implementation principle: Based on a large amount of experimental data and theoretical analysis, the corresponding relationship between the blade state and the load state is established. For example, when the blade is in an abnormal state, it may be due to excessive aerodynamic load, centrifugal force, etc. By judging the current state and combining the corresponding relationship established in advance, the load state of the blade can be analyzed, such as excessive load, uneven load distribution, etc.

[0144] Implementation process: Establishing the corresponding relationship: Through experiments and theoretical research, collect the state data of the blade under different load states, analyze the correlation between them, and establish the corresponding relationship table or model between the blade state and the load state. State judgment: According to the current state of the long and flexible blade obtained in the previous step, find the corresponding load state in the corresponding relationship. Result output: Output the load state analysis result of the long and flexible blade, such as "load normal" "load too large" etc.

[0145] It can be understood that the modal parameters (such as frequency, amplitude, phase, etc.) of each coupled vibration source directly reflect the vibration characteristics of the long and flexible blade. By analyzing these modal parameters, the vibration state of the blade can be deeply understood, and whether the blade has abnormal vibration such as resonance, flutter, etc. can be judged. Abnormal vibration may cause fatigue damage to the blade, affecting its service life and safety, so accurate analysis of the vibration state is crucial to ensure the normal operation of the blade.

[0146] Implementation principle: Under normal operating conditions, the modal parameters of each coupled vibration source of the long and flexible blade have certain ranges and characteristics. When the blade fails or is abnormal, these modal parameters will change. By comparing the current modal parameters with the parameter range under normal conditions, analyzing the parameter change trend and characteristics, and judging whether the vibration state of the blade is normal. For example, if the frequency of a certain vibration source approaches the natural frequency of the blade, resonance may occur, which needs special attention.

[0147] Implementation process: Determine the normal parameter range: Through a large number of experiments and operating data, determine the range and characteristics of the modal parameters of each coupled vibration source of the long and flexible blade under normal operating conditions. Parameter comparison: Compare the calculated modal parameters of each coupled vibration source with the normal parameter range, and analyze the parameter differences. Vibration state judgment: According to the results of parameter comparison, judge the vibration state of the blade. If the parameters are within the normal range, it is considered that the vibration state is normal; if the parameters exceed the normal range or show abnormal change trend, it is judged that the blade has abnormal vibration. Result output: Output the vibration state analysis result of the long and flexible blade, such as "vibration normal" "resonance risk exists" etc.

[0148] It can be understood that the load state and the vibration state are two important aspects of the long flexible blade operating state, and they influence and correlate with each other. Analyzing the load state or the vibration state alone may not comprehensively and accurately evaluate the overall state of the blade. Comprehensive consideration of the analysis results of the two aspects can more comprehensively and accurately judge the operating state of the blade, discover potential fault risks in time, and provide a scientific basis for the maintenance and management of the blade.

[0149] Implementation principle: According to the pre-prepared evaluation rules, the load state analysis result and the vibration state analysis result are comprehensively considered. For example, if the load state is normal and the vibration state is also normal, it is considered that the overall state of the blade is good; if the load state is abnormal or the vibration state is abnormal, or both are abnormal, further analysis of the severity and possible influence of the abnormality is needed to determine the final state monitoring result of the blade, such as "normal operation", "need attention", "exist fault risk", etc.

[0150] Implementation process: Formulate evaluation rules: According to the design requirements, operation experience and fault cases of the long flexible blade, formulate comprehensive evaluation rules to clearly define the corresponding state monitoring results under different load state and vibration state combinations. Result synthesis: The load state analysis result and the vibration state analysis result obtained in the foregoing are substituted into the evaluation rules for comprehensive judgment. Output final result: Output the state monitoring result of the long flexible blade, and give corresponding suggestions according to the result, such as continue to observe, perform maintenance inspection, etc.

[0151] Among them, the state recognition model is trained according to the historical signal characteristics of the long flexible blade and the historical state corresponding to the historical signal characteristics.

[0152] It can be understood that the load borne by the blade is different under different states, and the load signal contains important information of the working state of the blade. Time-frequency domain feature analysis can reflect the characteristics of the signal in time and frequency. Through time-frequency domain analysis of the blade load signal, such as using wavelet transform method, the signal can be decomposed into different time and frequency scales, and features such as energy distribution of different frequency components and peak frequency can be extracted. These features can reflect the stress of the blade under different states, such as under high wind speed state, the blade may bear greater load, and the energy of certain frequency components of the load signal will increase significantly.

[0153] It can be understood that the long flexible blade will generate vibration during operation, and the time domain characteristics of the vibration signal directly reflect the vibration state of the blade. The time domain characteristics include the mean, variance, peak value, kurtosis and the like of the vibration signal. Different states will cause different vibration modes of the blade, for example, when the blade is in failure or abnormal state, the mean, variance and other characteristics of the vibration signal will change obviously. By analyzing the time domain characteristics of the vibration signal, the abnormal vibration condition of the blade can be found in time, so as to judge the state of the blade.

[0154] Optionally, the selected historical signal features are taken as input, and the corresponding historical state is taken as output, and a machine learning model (such as a neural network, a decision tree, etc.) is trained. During the training process, the model will continuously adjust its own parameters, so that the mapping relationship between the input historical signal features and the output historical state is as accurate as possible. Through a large amount of historical data, the model can learn the rules and patterns of the long flexible blade signal features under different states, thereby having the ability to identify unknown states.

[0155] For example, the abnormal state of the long flexible blade includes a pollution state and an icing state, and the pollution state mainly refers to the accumulation of dust, salt or oil stains and other foreign matters on the surface of the blade, which increases the surface roughness. This will damage the aerodynamic shape of the blade, cause airflow separation and increase turbulence, thereby increasing the additional resistance and reducing the power generation efficiency. The time-frequency domain characteristics change, including: low-frequency vibration enhancement (0-20Hz): the increase of surface roughness causes airflow separation, forming low-frequency vortex shedding, and causing periodic low-frequency vibration. The frequency band energy distribution changes, including: in the 1-10Hz frequency band, the energy rises significantly, reflecting the enhancement of vortex-induced vibration (VIV). High-frequency noise suppression, including: pollution may mask part of the high-frequency vibration signal (such as small changes in the natural frequency of the blade). The measured data of a certain offshore wind farm shows that the energy of the blade root acceleration signal in the 5-15Hz frequency band under the pollution state is increased by about 40% compared with the normal state, and the time domain waveform presents periodic fluctuations.

[0156] For example, the icing state of a long flexible blade can change the blade's aerodynamic shape (e.g., the ice layer attachment causes the leading edge thickness to increase), while increasing the blade mass. Uneven ice layer distribution can also cause dynamic imbalance, triggering severe vibration and even structural damage. Time-frequency domain feature changes include: natural frequency shift: the increase in ice layer mass changes the blade's stiffness-mass ratio, causing the natural frequency to shift to a lower frequency (e.g., the normal state natural frequency is 20 Hz, and after icing, it may drop to 18 Hz). High-frequency transient impact (>100 Hz): transient impact signals generated when the ice layer falls off, appearing as time-domain spikes and high-frequency energy surges. Multiple harmonic peaks appear: uneven ice layer distribution can excite high-order modal vibrations, appearing as multiple discrete peaks (e.g., 20 Hz, 40 Hz, 60 Hz) in the frequency spectrum. In some icing states, the blade vibration signal has a significant resonance peak near 18 Hz, accompanied by transient impact signals in the 100-150 Hz frequency band, which are highly correlated with ice layer shedding events.

[0157] The historical state includes an actual state and / or a simulation state, and the historical signal feature corresponding to the historical state includes an actual signal feature and / or a simulation signal feature.

[0158] The simulation signal feature is a response of a simulation model of the long flexible blade in a simulation state.

[0159] Further, a state recognition model locally deployed on the edge server is used to analyze the current signal feature of the long flexible blade to obtain a current state corresponding to the long flexible blade, including: using an online transfer learning module to dynamically calibrate the current signal feature of the long flexible blade collected by the sensor to obtain a calibrated signal feature; using the state recognition model locally deployed on the edge server to analyze the calibrated signal feature to obtain the current state corresponding to the long flexible blade.

[0160] It can be understood that the actual state refers to various working states experienced by the long flexible blade in a real running environment, such as different wind speeds, wind directions, temperatures, humidities, and other environmental conditions, as well as running parameters such as the running speed and load of the blade. The simulation state is a virtual working state generated by computer simulation technology according to the design parameters, material properties, and various possible external conditions of the long flexible blade, and is used to study the behavior of the blade under different assumed conditions.

[0161] The historical signal feature corresponds to a signal feature of a historical state, including an actual signal feature and / or a simulation signal feature. The actual signal feature is various physical quantity data of the long flexible blade collected by the sensor in an actual state, such as vibration acceleration, strain, displacement, etc. The simulation signal feature is response data generated by a simulation model of the long flexible blade in a simulation state. These responses are calculated through complex physical models and mathematical algorithms, and can reflect the dynamic behavior of the blade in the simulation state.

[0162] It can be understood that the sensor may be affected by environmental changes, self-aging and other factors during long-term use, resulting in deviation of its measurement results. The online transfer learning module can correct the measurement error of the sensor in real time by dynamically calibrating the current signal characteristics of the long and flexible blade collected by the sensor, and improve the accuracy and reliability of the signal.

[0163] In the online transfer learning module, the recursive least squares method is used to update the calibration coefficients of the sensor.

[0164] In the online transfer learning module, the recursive least squares method is used to update the calibration coefficients of the sensor. The recursive least squares method is an adaptive filtering algorithm that can recursively calculate the optimal calibration coefficients based on the observation data at the current time and the estimated value at the previous time. This method has the characteristics of high computational efficiency and fast convergence speed, and can adjust the calibration coefficients of the sensor in real time to ensure the accuracy of the calibrated signal characteristics.

[0165] Optionally, the core steps of the sensor dynamic calibration based on the recursive least squares method (RLS) include: initialization: setting the initial calibration coefficient w(0) (usually a zero vector); initializing the error covariance matrix P(0) (a large diagonal matrix, such as σ 2 I; set the forgetting factor λ (0 < λ ≤ 1, control the influence of historical data).

[0166] Data acquisition: collect the sensor raw signal x(k) (such as vibration acceleration, strain) and obtain the reference signal y(k) (high-precision measurement value or simulation theoretical value)

[0167] Dynamic update: gain matrix calculation: K(k) = P(k-1)x(k) / [λ+x(k)^T P(k-1)x(k)];

[0168] Calibration coefficient update: w(k) = w(k-1) + K(k)[y(k)-x(k)^T w(k-1)];

[0169] Error covariance update: P(k) = [P(k-1)-K(k)x(k)^T P(k-1)] / λ.

[0170] Real-time calibration: correct the signal with the updated coefficient: y_cal(k) = x(k)^T w(k)

[0171] Loop iteration: continuously collect new data and repeat the above process to dynamically optimize the calibration parameters.

[0172] In this embodiment, the calibration coefficient is dynamically adjusted according to the real-time residual error to compensate for the sensor error, the weight of new and old data is balanced by λ to adapt to state changes, offline retraining is not required, and real-time updating is directly performed on the edge device.

[0173] S102, determine the fault warning strategy according to the current state.

[0174] Optionally, in S102, according to the current state of the long flexible blade obtained in S101, the fault warning strategy corresponding to the state is determined from the pre-set mapping relationship between state and fault warning strategy. This mapping relationship is usually established based on a large amount of historical data, expert experience and in-depth analysis of the running characteristics of the long flexible blade, which can ensure that the most suitable warning strategy is selected under different states.

[0175] Consider the severity and trend of the state: different states may represent different degrees of problem severity or potential failure risk. For example, some states may only be minor abnormalities that require only simple warning prompts; while other states may indicate that a serious failure is about to occur, requiring immediate emergency measures. Therefore, S102 needs to consider the severity of the state and its change trend to determine the corresponding level of fault warning strategy.

[0176] Combine real-time data with preset rules: when determining the fault warning strategy, not only the current state obtained in S101, but also other related data such as environmental parameters, running time, etc. and some pre-set rules and thresholds will be combined. In this way, the warning strategy can be more accurate and flexible, better responding to various complex situations.

[0177] For example, assume that the long flexible blade has three states: normal state, minor abnormal state and serious abnormal state.

[0178] Normal state: if S101 analysis shows that the long flexible blade is in normal state, the fault warning strategy determined in S102 may be to issue no warning signal, continue to monitor and record normally.

[0179] Minor abnormal state: when S101 identifies that the long flexible blade is in a minor abnormal state, S102 may determine a fault warning strategy to issue a level one warning signal, reminding the staff to observe the operation of the blade and suggesting a routine inspection and maintenance in the near future.

[0180] Serious abnormal state: if S101 judges that the long flexible blade is in a serious abnormal state, S102 will determine to issue a level two warning signal, immediately notify the maintenance personnel to the scene for emergency inspection and treatment, and may also automatically start some backup devices or take other emergency measures to prevent the failure from further expanding.

[0181] In this embodiment, the state recognition model is deployed on an edge server instead of a cloud or other remote server. The edge server is close to the data source and can respond quickly, reducing the delay of data transmission. This is particularly important for the state recognition of long and flexible blades, because the state of the blade can change very quickly, and timely and accurate state recognition can provide strong support for subsequent control and maintenance.

[0182] The signal features of icing, contamination, and other states may overlap in the time-frequency domain (e.g., similar low-frequency vibrations). In order to accurately recognize the state, in an exemplary embodiment, the historical signal features include time-frequency domain features of the blade load signal and time domain features of the vibration signal, as well as aerodynamic change features of the long and flexible blade in the historical state. The state recognition model is a deep learning model with physical embedding, and the loss function of the state recognition model is constructed according to the aeroelastic equation.

[0183] Further, the aerodynamic change features include at least one of the following: the icing thickness distribution outside the blade, the contamination surface roughness, the micro-cracks inside the blade, and the surface temperature distribution and microstructure changes of the blade.

[0184] It can be understood that the microstructure changes may specifically represent the following changes:

[0185] Crystal structure changes: The blade is usually made of composite materials or metal materials, and changes in the internal crystal structure of the blade will affect the performance of the blade. For example, during long-term use, due to factors such as external force and temperature changes, the crystal structure inside the metal blade may dislocate, slip, and other phenomena, causing the crystal arrangement to change. Such microstructure changes will affect the strength and stiffness of the blade, and indirectly affect the aerodynamic performance of the blade, because the structural stability of the blade will affect its vibration characteristics and shape retention ability in the air flow.

[0186] Fiber orientation changes: For blades made of fiber-reinforced composite materials, the orientation of the fibers is designed and distributed during manufacturing. However, during use, the orientation of the fibers may change due to factors such as fatigue load and impact. For example, some fibers may twist and dislocate, no longer maintaining the original ideal arrangement direction. This will affect the overall mechanical properties of the composite material, causing the deformation of the blade to change when subjected to aerodynamic loads, thereby affecting the aerodynamic performance.

[0187] Porosity variation: The pores present inside the material are part of the microstructure. During the manufacturing process or during service, the porosity can change. For example, due to manufacturing defects or long-term environmental erosion, fatigue damage, etc., the pores inside the blade material can increase, enlarge, or connect. The change in porosity will affect the density and mechanical properties of the material, and then affect the vibration and deformation of the blade in the airflow, affecting the aerodynamic performance.

[0188] Interfacial bonding variation: In composite blades, the interfacial bonding state between fibers and matrix is an important aspect of the microstructure. After long-term use, phenomena such as debonding and crack propagation may occur at the interface, causing the bonding force between the fiber and the matrix to decrease. This change in interfacial bonding will reduce the overall performance of the composite material, affecting the mechanical response of the blade under aerodynamic load, and then changing its aerodynamic performance.

[0189] It can be understood that the aerodynamic performance of the blade is of great importance to its state. Different abnormal states (such as icing, contamination, presence of micro-cracks, etc.) will significantly change the aerodynamic shape and surface characteristics of the blade, and then affect its aerodynamic performance. For example, icing will form an ice layer on the surface of the blade, changing the aerodynamic shape of the blade, increasing air resistance, and reducing lift; contamination will increase the roughness of the blade surface, affecting the flow characteristics of the airflow; micro-cracks may cause changes in the local aerodynamic performance of the blade. Therefore, considering these aerodynamic change characteristics can more comprehensively reflect the state of the blade.

[0190] Traditional deep learning models are mainly based on data-driven and lack consideration of physical laws. In the long flexible blade condition monitoring, the aeroelastic equation describes the mechanical response and deformation of the blade under the action of airflow, which is an important physical basis for blade dynamics. Embedding physics into deep learning models means incorporating aeroelastic equations into the structure or loss function of the model. The advantage of this is that the model can follow the physical laws while learning the data features, improving the generalization ability and interpretability of the model. For example, when encountering a state that has not appeared in the training data but conforms to the physical law, the model can more accurately identify it.

[0191] The loss function is used to measure the difference between the model's predicted results and the true values, and the model's parameters are optimized by minimizing the loss function. Building a loss function based on the aeroelastic equation is to add a constraint term related to the aeroelastic equation to the loss function. This allows the model to not only make the predicted results as close as possible to the true state label during the training process, but also to satisfy the physical relationship described by the aeroelastic equation. For example, if the model's predicted blade response deviates significantly from the result calculated by the aeroelastic equation, the value of the loss function will increase, prompting the model to adjust its parameters so that the predicted results better conform to the physical law.

[0192] Optionally, sensors are used to collect blade load signals and vibration signals, and various measurement methods are used to obtain the aerodynamic change characteristics of the blade, such as using laser scanning to measure the ice thickness distribution, using a roughness meter to measure the contaminated surface roughness, using non-destructive testing technology to detect internal micro-cracks in the blade, using an infrared thermal imager to measure the blade surface temperature distribution, etc.

[0193] Time-frequency domain feature extraction of blade load signals: methods such as wavelet transform and short-time Fourier transform are used to analyze the time-frequency domain of blade load signals, and features such as energy and peak frequency of different frequency components are extracted.

[0194] Time domain feature extraction of vibration signals: calculate the mean, variance, peak value, kurtosis and other time domain features of the vibration signal.

[0195] Aerodynamic change feature processing: quantize and encode the collected aerodynamic change features (such as ice thickness distribution, contaminated surface roughness, etc.) so that they can be used as inputs to the model.

[0196] Further, a physically embedded deep learning model is constructed, including:

[0197] Model structure design: convolutional neural networks (CNN), recurrent neural networks (RNN) or their combinations can be used as the basic model structure. Modules related to the aeroelastic equation are added to the model, such as introducing constraint conditions based on the aeroelastic equation in certain layers.

[0198] The loss function consists of two parts: one is the traditional classification loss Lclass (such as cross-entropy loss), which measures the difference between the model's predicted state category and the true label; the other is the constraint loss Lphysics based on the aeroelastic equation. For example, calculate the difference between the model's predicted blade response and the theoretical response calculated according to the aeroelastic equation, and use it as a constraint loss term. The total loss function can be represented as: L = Lclass + aLphysics.

[0199] This is a combined loss function, where Lclass represents the classification loss, which is usually used to measure the difference between the model's prediction and the actual label; Lphysics represents the loss based on physical constraints, which here refers specifically to the aeroelastic equation, to ensure that the model's prediction conforms to the physical laws. a is a hyperparameter that adjusts the weight of the two parts of the loss.

[0200] Lclass is the classification loss: this part of the loss measures the model's performance on the classification task, such as cross-entropy loss, which calculates the difference between the model's predicted probability distribution and the true label's probability distribution.

[0201] Lphysics is a constraint loss based on aeroelastic equations: This part of the loss ensures that the model's predictions comply with aeroelastic equations, which describe the interaction between a fluid (such as air) and an elastic structure (such as an airplane wing or a wind turbine blade). By minimizing this part of the loss, the model is encouraged to learn behaviors that comply with physical laws.

[0202] α is a hyperparameter: This parameter allows the model trainer to adjust the relative importance between the classification loss and the physical constraint loss according to the specific task requirements. For example, if the physical constraint is particularly important, a larger value of α may be chosen; if classification accuracy is more critical, a smaller value of α may be chosen.

[0203] In an exemplary embodiment, the state recognition model comprises:

[0204] A multi-modal input layer, including a signal feature branch and an aerodynamic shape branch, the signal feature branch being configured to receive time-frequency domain features of blade load signals and time domain features of vibration signals, and the aerodynamic shape branch being configured to receive aerodynamic variation features.

[0205] A feature fusion module configured to generate multi-dimensional fusion features based on the time-frequency domain features of blade load signals, the time domain features of vibration signals, and the aerodynamic variation features.

[0206] A physical embedding constraint module configured to predict a predicted value of a physical field quantity based on the multi-dimensional fusion features, and to correct the multi-dimensional fusion features based on the predicted value of the physical field quantity and aeroelastic equations, to obtain corrected features.

[0207] A classification output layer configured to determine membership probabilities of the corrected features belonging to each historical state.

[0208] It can be understood that the state information of a long and flexible blade contains multiple aspects, and a single signal feature or aerodynamic shape feature may not be able to accurately reflect its state. Therefore, by using a multi-modal input layer, different types of features are processed separately, which can make full use of the information of each type of feature and improve the state recognition ability of the model.

[0209] Signal feature branch: The time-frequency domain features of blade load signals and the time domain features of vibration signals reflect the mechanical response of the blade during operation. The time-frequency domain features can show the distribution of the load signal in time and frequency, and the time domain features reflect the basic statistical characteristics of the vibration signal. These features can capture the dynamic changes of the blade in different states, such as vibration abnormalities and load fluctuations.

[0210] Aerodynamic shape branch: Aerodynamic variation features are closely related to the aerodynamic performance of the blade. Different states (such as icing, contamination, etc.) can cause changes in the aerodynamic shape of the blade, thereby affecting its aerodynamic performance. By receiving the aerodynamic variation features, the model can analyze the state of the blade from the aerodynamic point of view.

[0211] It can be understood that different types of features each contain partial information about the state of the blade, but there may be limitations in using these features alone. The role of the feature fusion module is to integrate the features of the signal feature branch and the aerodynamic shape branch to generate multi-dimensional fusion features. By fusing features of different modalities, the model can comprehensively consider the mechanical response and aerodynamic performance of the blade, thereby more comprehensively describing the state of the blade. Multi-dimensional fusion features can provide more abundant information, which helps the model discover potential relationships between different features and improve the accuracy and robustness of state recognition.

[0212] It can be understood that based on the multi-dimensional fusion features, the model predicts the predicted values of the physical field quantities. The physical field quantities can be parameters related to the aeroelasticity of the blade, such as stress, strain, displacement, etc. The predicted values of these physical field quantities reflect the model's estimation of the physical state of the blade based on the input features. The aeroelastic equation describes the mechanical response and deformation of the blade under the action of airflow, which is an important physical basis for blade dynamics. By comparing the predicted values of the physical field quantities with the aeroelastic equation, the multi-dimensional fusion features are modified according to the comparison results. The purpose of this is to make the prediction results of the model more consistent with the physical laws, and to avoid the model making predictions that do not conform to the physical reality. By introducing physical constraints, the generalization ability and interpretability of the model are enhanced.

[0213] Optionally, the classification output layer calculates the membership probabilities of the modified features belonging to each historical state. By analyzing and judging the modified features, the model can determine which historical state the current state of the blade is most likely to belong to. The membership probability represents the confidence of the model in classifying the state of the blade. The final membership probability can help users judge the state of the blade, detect abnormal states in a timely manner, and take appropriate measures.

[0214] Further, the predicted values of the physical field quantities include a predicted displacement field and a predicted aerodynamic force; and the physical embedding constraint module includes:

[0215] An aeroelastic residual calculation unit is configured to determine a residual between the predicted values of the physical field quantities and the regular values of the physical field quantities according to the aeroelastic equation.

[0216] A feature modification unit is configured to adjust the feature extraction process based on the gradient of the residual, suppress noise features in the multi-dimensional fusion features that violate the physical laws, and strengthen effective features in the multi-dimensional fusion features that conform to the physical laws, to obtain modified features.

[0217] The aeroelastic equation is constructed according to a mass matrix, a damping matrix and a stiffness matrix, as follows:

[0218]

[0219] M: mass matrix (structural inertia property)

[0220] C: damping matrix (energy dissipation property)

[0221] K: stiffness matrix (elastic recovery property)

[0222] u(t): physical field quantity vector (displacement / velocity / pressure, etc.)

[0223] The residual is defined as:

[0224]

[0225] The residual reflects the deviation of the predicted solution (numerical solution / neural network output) from the physical law (analytical solution / experimental data), and represents the degree of non-equilibrium of the system.

[0226] The physical constraint mechanism of the feature correction unit includes: constructing a Lagrange function: L = Ldata + R(u) 2 ;

[0227] The residual penalty term is introduced, and the gradient is realized through back propagation:

[0228] Feature space optimization includes: noise feature suppression: applying a negative gradient (penalty term) to the feature dimension sensitive to the residual; effective feature enhancement: applying a positive gradient (reward term) to the feature dimension that meets the physical law; realizing orthogonal decomposition of the feature space: z = zvalid + znoise, where zvalid perpendicular to znoise (based on the orthogonality of the residual gradient).

[0229] Optionally, the feature correction unit is also used to use a differentiable physical solver to back-propagate the residual to the feature fusion module to correct the network parameters in the feature extraction process.

[0230] The network parameters in the feature extraction process include the network parameters in the classification output layer, the physical embedding constraint module, the feature fusion module and the multi-modal input layer.

[0231] Further, the feature correction unit is configured with a dynamic feature channel gating mechanism, which is used to: generate a channel weight vector according to the L2 norm of the residual, and perform channel-level weighting on the multi-dimensional fusion features based on the channel weight vector; when the residual is large, the weight of the multi-dimensional fusion features is reduced, and the physical law prediction value is preferred. When the residual is small, the data-driven feature details are retained.

[0232] It can be understood that the main purpose of the feature correction unit is to correct the network parameters in the feature extraction process by combining physical laws and data-driven features. This helps to improve the understanding of the model for physical phenomena and prediction accuracy.

[0233] The network parameters involved in the correction process include the classification output layer, the physical embedding constraint module, the feature fusion module, and the network parameters in the multi-modal input layer.

[0234] It can be understood that in a deep learning model, the feature extraction process is usually composed of multiple network layers, such as the multi-modal input layer, the feature fusion module, the physical embedding constraint module, and the classification output layer. The differentiable physics solver is a tool that can calculate the derivative of the physical equation, which allows the seamless integration of the physical model with the deep learning model.

[0235] After calculating the aeroelastic residual, the residual is backpropagated to each layer of the feature extraction network through the differentiable physics solver. According to the chain rule, the gradient of the residual with respect to the network parameters is calculated, and then these parameters are updated using an optimization algorithm such as stochastic gradient descent. The purpose of this is to let the network learn a feature representation that conforms to the physical laws, reducing the deviation between the prediction results and the physical laws.

[0236] It can be understood that the core idea of the dynamic feature channel gating mechanism is to adaptively adjust the weight of each channel in the multi-dimensional fusion feature according to the size of the residual. The L2 norm of the residual can be used as an indicator to measure the deviation between the prediction results and the physical laws.

[0237] When the residual is large: it indicates that the current data-driven feature has a large deviation from the physical laws, at this time the weight of the multi-dimensional fusion feature is reduced, and the model relies more on the predicted value of the physical laws to ensure that the output of the model conforms to the physical constraints. When the residual is small: it indicates that the data-driven feature is consistent with the physical laws, and the details of the data-driven feature are preserved to make full use of the information in the data for more accurate prediction.

[0238] Residual backpropagation and parameter update, including: assuming that the loss function L is some kind of measure of the residual (such as the sum of the squares of the residuals): Where N is the number of samples.

[0239] According to the A chain rule, calculate the gradient of the loss function L with respect to the network parameters θ: Update the network parameters using an optimization algorithm such as stochastic gradient descent: Where α is the learning rate.

[0240] For the dynamic feature channel gating mechanism, calculate the L2 norm of the residual:

[0241]

[0242] where r i is the i-th element of the residual vector R, and d is the dimension of the residual vector.

[0243] The channel weight vector w can be generated using a Sigmoid function:

[0244] w = σ(-r norm )

[0245] where is the Sigmoid function.

[0246] The multi-dimensional fusion feature F is channel-level weighted:

[0247] F weighted = F ⊙ w.

[0248] where ⊙ denotes element-wise multiplication.

[0249] In an exemplary embodiment, the state recognition model further comprises an online adversarial data augmentation module, configured to: based on the residual, train an adversarial generative network to generate synthetic signal features consistent with the physical law of the current state, and mix the synthetic signal features with the current signal features and input them into the classification output layer, through a consistency regularization loss function, and update the parameters of the generator and the classifier in the adversarial generative network.

[0250] Optionally, the principle of the online adversarial data augmentation module is to generate synthetic signal features consistent with the physical law of the current state by training an adversarial generative network (GAN). This method uses the generator of the generative adversarial network to create realistic data samples, which are mixed with real data and input into the classification output layer, thereby enhancing the model's recognition ability for the current state. Through a consistency regularization loss function, the model can learn more rich feature representations while maintaining data consistency.

[0251] The specific implementation process is as follows:

[0252] Residual calculation: First, calculate the residual under the current state, i.e., the difference between the model prediction value and the true value. The residual reflects the shortcomings of the model under the current state. Adversarial generative network training: use the residual information to train an adversarial generative network. This network includes a generator and a discriminator. The generator is responsible for generating synthetic signal features consistent with the physical law of the current state, while the discriminator is used to distinguish between generated data and real data.

[0253] Synthetic signal feature generation: the generator generates synthetic signal features according to the rules learned from the training process. These features should be consistent with the statistical characteristics of the real signal features, but can also supplement the information that may be missing in the real data.

[0254] Feature mixing: The generated synthetic signal features are mixed with the current signal features to form an enhanced feature set. This mixing can be a simple concatenation or a weighted average, depending on the model design.

[0255] Consistency regularization loss function: A loss function is designed that considers not only the accuracy of classification but also the consistency of features. By minimizing this loss function, the model can learn the classification task while maintaining the consistency of generated features with real features.

[0256] Parameter update: The parameters of the generator and discriminator are updated through the backpropagation algorithm. The goal of the generator is to generate increasingly realistic features to deceive the discriminator; while the goal of the discriminator is to increasingly accurately distinguish between real features and generated features. This adversarial training process iterates until a balance point is reached.

[0257] Model evaluation and optimization: During the training process, the performance of the model is regularly evaluated, including classification accuracy and feature consistency. Based on the evaluation results, model parameters or training strategies are adjusted to further improve the recognition ability of the model.

[0258] In this embodiment, the online adversarial data augmentation module can effectively enhance the robustness and generalization ability of the state recognition model, so that it can maintain a high recognition accuracy when facing complex and variable states.

[0259] It can be understood that although each step in the flowchart involved in the above embodiments is displayed in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise stated herein, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above embodiments can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with other steps or steps or stages in other steps.

[0260] Based on the same inventive concept, the embodiments of the present application also provide a long flexible blade state monitoring device for implementing the long flexible blade real-time monitoring and fault early warning method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more long flexible blade state monitoring device embodiments provided below can refer to the limitations of the long flexible blade real-time monitoring and fault early warning method described above, and will not be repeated here.

[0261] In one example embodiment, as shown in Figure 2 A long flexible blade condition monitoring device is provided, which comprises:

[0262] A collection module 11 is configured to collect vibration signals of the long flexible blade in a running state, the vibration signals being coupled by multiple factors including aerodynamic load, gravity and centrifugal force;

[0263] An equation construction module 12 is configured to construct an overdetermined equation set containing multi-source excitation characteristics based on a blade dynamics model;

[0264] A signal decomposition module 13 is configured to decompose the vibration signals by using a sparse decomposition algorithm to obtain a signal decomposition result;

[0265] A parameter extraction module 14 is configured to solve the overdetermined equation set based on the signal decomposition result to obtain modal parameters of each coupled vibration source;

[0266] A condition monitoring module 15 is configured to determine a condition monitoring result of the long flexible blade according to the modal parameters of each coupled vibration source.

[0267] The above modules of the long flexible blade condition monitoring device can be realized by software, hardware or a combination thereof. The above modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform operations corresponding to the above modules.

[0268] In one example embodiment, a computer device is provided, which comprises a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps:

[0269] Collect vibration signals of the long flexible blade in a running state, the vibration signals being coupled by multiple factors including aerodynamic load, gravity and centrifugal force;

[0270] Construct an overdetermined equation set containing multi-source excitation characteristics based on a blade dynamics model;

[0271] Decompose the vibration signals by using a sparse decomposition algorithm to obtain a signal decomposition result;

[0272] Solve the overdetermined equation set based on the signal decomposition result to obtain modal parameters of each coupled vibration source;

[0273] Determine a condition monitoring result of the long flexible blade according to the modal parameters of each coupled vibration source.

[0274] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0275] The vibration signal of the long and flexible blade in a running state is collected, and the vibration signal is coupled by multiple factors of aerodynamic load, gravity and centrifugal force;

[0276] An overdetermined equation set containing multi-source excitation characteristics is constructed based on a blade dynamics model;

[0277] A sparse decomposition algorithm is used to decompose the vibration signal to obtain a signal decomposition result;

[0278] Based on the signal decomposition result, the overdetermined equation set is solved to obtain modal parameters of each coupled vibration source;

[0279] According to the modal parameters of each coupled vibration source, a state monitoring result of the long and flexible blade is determined.

[0280] In one embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the following steps:

[0281] The vibration signal of the long and flexible blade in a running state is collected, and the vibration signal is coupled by multiple factors of aerodynamic load, gravity and centrifugal force;

[0282] An overdetermined equation set containing multi-source excitation characteristics is constructed based on a blade dynamics model;

[0283] A sparse decomposition algorithm is used to decompose the vibration signal to obtain a signal decomposition result;

[0284] Based on the signal decomposition result, the overdetermined equation set is solved to obtain modal parameters of each coupled vibration source;

[0285] According to the modal parameters of each coupled vibration source, a state monitoring result of the long and flexible blade is determined.

[0286] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, database or other medium used in each embodiment of the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in each embodiment of the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without limitation. The processor involved in each embodiment of the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without limitation.

[0287] Each technical feature of the above embodiments can be combined arbitrarily. In order to make the description simple, not all possible combinations of each technical feature in the above embodiments are described. However, as long as the combination of these technical features does not exist, it should be considered as the scope of the present application.

[0288] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for real-time monitoring and fault early warning of long, flexible blades, characterized in that, Applied to edge servers, the method includes: Vibration signals of long flexible blades are collected during operation. These vibration signals are affected by aerodynamic loads, gravity, and centrifugal force. An overdetermined set of equations containing multi-source excitation features was constructed based on the blade dynamics model; The vibration signal is decomposed using a sparse decomposition algorithm to obtain the signal decomposition result; Based on the signal decomposition results, the overdetermined equations are solved to obtain the modal parameters of each coupled vibration source; Based on the modal parameters of each coupled vibration source, the state monitoring results of the long flexible blade are determined; The determination of the state monitoring results of the long flexible blade includes: The current signal characteristics of the long flexible blade are analyzed using the state recognition model deployed locally on the edge server to obtain the current state of the long flexible blade. The state recognition model is a deep learning model that includes physical embeddings, and the loss function of the state recognition model is constructed based on the aeroelastic equations; the state recognition model includes: The multimodal input layer includes a signal feature branch and an aerodynamic shape branch. The signal feature branch is used to receive the time-frequency domain characteristics of the blade load signal and the time-domain characteristics of the vibration signal, and the aerodynamic shape branch is used to receive aerodynamic variation characteristics. The feature fusion module is used to generate multi-dimensional fused features based on the time-frequency domain features of the blade load signal, the time-domain features of the vibration signal, and the aerodynamic change features. The physical embedding constraint module is used to predict the predicted values ​​of physical field quantities based on the multidimensional fusion features, and to modify the multidimensional fusion features based on the predicted values ​​of physical field quantities and the aeroelastic equations to obtain modified features. The classification output layer is used to determine the membership probability of the modified feature to each historical state.

2. The method according to claim 1, characterized in that, The sparse decomposition algorithm is an orthogonal matching pursuit algorithm, in which dictionary atoms related to the blade's intrinsic modes are matched first during the iteration process.

3. The method according to claim 2, characterized in that, The dictionary atoms of the orthogonal matching pursuit algorithm are generated in the following manner: Combined with the Campbell diagram preselection of the resonant frequency band of the aforementioned long flexible blade; Atomic basis functions that match the physical structure of the blade were selected based on modal confidence factors.

4. The method according to claim 1, characterized in that, The blade dynamics model includes: Forced vibration term caused by aerodynamic load; Correction term for static offset caused by gravity; The stiffness enhancement term caused by centrifugal force; The calculation of the static offset correction term includes: The direction of gravity decomposition is dynamically adjusted according to the installation angle of the long flexible blade, and a material nonlinear coefficient is introduced to compensate for the change in geometric stiffness under large deformation.

5. The method according to claim 1, characterized in that, The construction of the overdetermined system of equations specifically includes: Prior information on blade mode shapes can be obtained through finite element simulation or experimental calibration. The overdetermined equation system is obtained by embedding prior information as a constraint condition into the equation system.

6. The method according to claim 5, characterized in that, The methods for obtaining the prior information include: The first three mode shapes of the blade were calibrated using a laser vibration meter. The coefficient matrix of the overdetermined equation system is updated in real time using operating parameters.

7. The method according to claim 1, characterized in that, The acquisition of the vibration signal specifically includes: Triaxial accelerometers are arranged at the leading edge, trailing edge, and middle of the long flexible blade to collect vibration signals of the long flexible blade. Among them, triaxial accelerometers are arranged at the leading edge, trailing edge and mid-blade positions, respectively, and anti-aliasing filters and synchronous sampling technology are used to ensure phase consistency of multi-channel signals.

8. The method according to claim 1, characterized in that, The determination of the state monitoring results of the long flexible blade based on the modal parameters of each coupled vibration source also includes: Based on the current state, determine the load state analysis results of the long flexible blade; Based on the modal parameters of each coupled vibration source, the vibration state analysis results of the long flexible blade are determined; Based on the load state analysis results and the vibration state analysis results, the state monitoring results of the long flexible blade are determined; The state recognition model is trained based on the historical signal features of the long flexible blade and the historical states corresponding to the historical signal features. The historical signal characteristics include the time-frequency domain characteristics of the blade load signal and the time-domain characteristics of the vibration signal, as well as the aerodynamic variation characteristics of the long flexible blade under the historical state.

9. The method according to claim 8, characterized in that, The aerodynamic change characteristics include at least one of the following: the distribution of ice thickness on the outside of the blade, the roughness of the contaminated surface, the microcracks inside the blade, and the surface temperature distribution and microstructure changes of the blade.

Images