A damage identification method for composite propeller blades of large UAV propulsion system

Abstract

The present invention relates to a damage identification method for composite propeller blades of a large unmanned aerial vehicle propulsion system, comprising: establishing multiple finite element material models, setting multiple sensor measurement points on each finite element material model, and constructing a transmissibility function between the sensor measurement points to obtain a marginal probability density function of each transmissibility function; obtaining a damage characteristic vector and a damage index corresponding to each finite element material model based on the marginal probability density functions of the multiple finite element material models; obtaining a damage judgment threshold based on the damage index, and obtaining a damage identification model based on the damage characteristic vector and a damage position in a damage model; obtaining a transmissibility function to be measured between the sensor measurement points on the blade to be measured, and obtaining a damage index to be measured and a damage characteristic vector to be measured based on the transmissibility function to be measured; judging whether the damage index to be measured exceeds the damage judgment threshold, and if so, obtaining the damage position based on the damage identification model.

Description

Technical Field

[0001] The present invention relates to the technical field of structural damage detection, and in particular to a method for identifying damage to composite propeller blades of a large unmanned aerial vehicle propulsion system. Background Art

[0002] Composite structures such as carbon fiber reinforced polymer (CFRP) are widely used in aerospace due to their advantages such as low density, high specific strength, and corrosion resistance. As a key component of the propulsion system, the integrity and reliability of composite propeller blades are crucial to the service safety of large unmanned aerial vehicles (UAVs). However, minor defects such as inclusions, bubbles, and cracks generated during the manufacturing process can expand under the action of alternating aerodynamic loads, causing damage to composite propeller blades. In addition, blades can also be damaged by impacts from external objects such as debris. Therefore, identifying early damage to composite propeller blades is crucial to ensuring the safe operation of UAVs.

[0003] Damage identification methods based on structural dynamic characteristics have been widely studied. The principle is that damage reduces the local stiffness of the structure, which will cause changes in structural dynamic characteristics such as natural frequency, modal vibration shape and frequency response function. Therefore, structural damage can be effectively detected by identifying changes in dynamic characteristics. However, there are the following problems in the current damage identification methods based on structural dynamic characteristics: (1) Modal parameters such as natural frequency are not sensitive to weak damage and are difficult to use to identify early damage; (2) Modal vibration shapes usually require very dense measurement points to identify high-order modes, and require the support of external equipment such as laser vibrometers. The operation is relatively complicated and cannot be used for in-situ detection; (3) The frequency response function requires an accurate excitation signal, but in actual applications, the structure is usually excited by the environment and cannot provide an accurate excitation signal; (4) Most current damage indicators do not take into account uncertainties such as load differences, environmental noise and calculation errors, resulting in poor robustness of the damage indicators; (5) A single damage indicator can usually achieve damage detection, but it is difficult to achieve damage location. The multi-measurement point multi-feature fusion method is expected to solve this problem. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying damage to composite propeller blades of a large unmanned aerial vehicle propulsion system.

[0005] To achieve the above-mentioned object, the present invention provides a method for identifying damage to composite propeller blades of a large UAV propulsion system, comprising:

[0006] S1. Establishing multiple finite element material models of a composite propeller blade, wherein the multiple finite element material models include: a lossless model and multiple lossy models;

[0007] S2. Setting a plurality of sensor measurement points on each of the finite element material models, constructing a transmissibility function between the sensor measurement points based on the transmissibility functions, and obtaining a marginal probability density function for each of the transmissibility functions, wherein the marginal probability density functions include: a first marginal probability density function corresponding to the lossless model and a second marginal probability density function corresponding to the lossy model;

[0008] S3. Based on the marginal probability density functions of the plurality of finite element material models, obtaining a damage characteristic vector and a damage index corresponding to each of the finite element material models; wherein the damage characteristic vector includes: a first damage characteristic vector corresponding to the lossless model, and a second damage characteristic vector corresponding to the lossy model; and the damage index includes: a first damage index corresponding to the lossless model, and a second damage index corresponding to the lossy model;

[0009] S4. obtaining a damage judgment threshold for determining whether the composite propeller blade is damaged based on the damage indicator, and obtaining a damage identification model for identifying the damage location based on the damage feature vector and the damage location in the damage model;

[0010] S5. Obtaining a measured transmissibility function between sensor measuring points on the blade to be measured, and obtaining a measured damage index and a measured damage feature vector based on the measured transmissibility function;

[0011] S6. Determine whether the damage index to be measured exceeds the damage judgment threshold. If not, determine that the blade to be measured is a healthy blade. If so, obtain the damage position on the blade to be measured based on the damage recognition model and the damage feature vector to be measured.

[0012] According to one aspect of the present invention, in step S1, the step of establishing multiple finite element material models of composite propeller blades includes:

[0013] S11. Establishing a finite element model of the composite propeller blade, and obtaining a plurality of initial models based on the finite element model; wherein the composite propeller blade comprises: a core material and skins disposed on opposite sides of the core material;

[0014] S12. Obtain multiple damage samples in a random generation manner, and append the damage samples to the initial model to obtain multiple lossy models, and use the remaining one of the initial models as a lossless model.

[0015] According to one aspect of the present invention, in step S2, the step of setting a plurality of sensor measurement points on each of the finite element material models and constructing a transmissibility function between the sensor measurement points includes:

[0016] S21. Meshing the skin using four-node linear quadrilateral elements and meshing the core using eight-node linear tetrahedral elements;

[0017] S22. Arranging the sensor measuring points on the finite element material model based on preset rules and quantities, wherein the arrangement of the sensor measuring points of each finite element material model is consistent;

[0018] S23. Applying a random vibration signal to the fixed end of the finite element material model and collecting a corresponding strain response signal based on the sensor measuring point;

[0019] S24. Calculate the transferability function based on the strain response signal of each sensor measuring point.

[0020] According to one aspect of the present invention, the transferability function is expressed as:

[0021]

[0022] in, Indicates that at frequency ω k The transfer rate function between sensor point m and sensor point n, ω k =k×1 / f s , k represents the kth frequency point, f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, The spectrum of the strain response signal of the sensor measuring point m obtained by fast Fourier transform is represented by The frequency spectrum of the strain response signal at sensor point n obtained by fast Fourier transform.

[0023] According to one aspect of the present invention, in step S2, the step of obtaining the marginal probability density function of each of the transmissibility functions includes:

[0024] S25. Obtain the covariance matrix of the fast Fourier transform coefficients of the strain response signals of the two sensor measuring points, which is expressed as:

[0025]

[0026] in, represents the covariance matrix of the strain response signal of sensor measuring point m and the strain response signal of sensor measuring point n, and They are and The variance of express and The multiple correlation coefficient is expressed as express and The real part correlation coefficient of express and The imaginary correlation coefficient of for conjugation of;

[0027] S26. Obtaining the probability density function of the strain response signals of the two sensor measuring points based on the covariance matrix and the transmissibility function, which is expressed as:

[0028]

[0029] in, represents the probability density function of the strain response signal of the sensor measuring point m and the strain response signal of the sensor measuring point n, Express about A process variable, and The superscript T represents the transpose operation. express At frequency ω k is a complex random variable at , and ω k =k×1 / f s , f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, express At frequency ω k A real random variable at , express At frequency ω k An imaginary random variable at ;

[0030] S27. Obtain the marginal probability density function of the sensor measurement point based on the probability density function, which is expressed as:

[0031]

[0032] in, The strain response signal between the sensor measuring point m and the sensor measuring point n at the frequency ω k The real part marginal probability density function at , The strain response signal between the sensor measuring point m and the sensor measuring point n at the frequency ω k The imaginary marginal probability density function at .

[0033] According to one aspect of the present invention, in step S3, the step of obtaining the damage characteristic vector and damage index corresponding to each of the finite element material models based on the marginal probability density functions of the plurality of finite element material models includes:

[0034] S31. For each of the finite element material models, respectively, obtaining strain response signals of sensor measuring points of multiple groups of the finite element material models based on the applied random vibration signal;

[0035] S32. Acquire multiple sets of strain response signals corresponding to the non-destructive model, and obtain multiple sets of the first marginal probability density functions corresponding to sensor point pairs on the non-destructive model using step S2; wherein the sensor point pairs are constructed based on the sensor points on the non-destructive model;

[0036] S33. Select any two groups of the first marginal probability density functions to obtain first JSD divergences of each sensor point pair in the lossless model at all frequency points, average the first JSD divergences of all sensor point pairs to obtain a first damage eigenvector of the lossless model, and further average the multidimensional data contained in the first damage eigenvector to obtain a first damage index of the lossless model;

[0037] S34. Obtain multiple sets of strain response signals corresponding to the lossy model, and obtain multiple sets of the second marginal probability density functions corresponding to the sensor point pairs on the lossy model using step S2; wherein the sensor point pairs are constructed based on the sensor points on the lossy model;

[0038] The second marginal probability density function of the sensor point pair on the lossy model and the first marginal probability density function of the sensor point pair on the lossless model are used to obtain a second JSD divergence of each sensor point pair in the lossy model at all frequency points. The second JSD divergences of all the sensor point pairs are averaged to obtain a second damage feature vector of the lossy model. The multidimensional data contained in the second damage feature vector is further averaged to obtain a second damage indicator of the lossy model.

[0039] According to one aspect of the present invention, the JSD divergence is expressed as:

[0040]

[0041] Where KLD(P||Q) represents the KLD divergence between the probability density function P(x) and the probability density function Q(x), JSD(P||Q) represents the JSD divergence between the probability density function P(x) and the probability density function Q(x), and P(x) and Q(x) represent the two marginal probability density functions involved in the JSD divergence calculation.

[0042] According to one aspect of the present invention, in step S4, in the step of obtaining a damage judgment threshold for judging whether the composite propeller blade is damaged based on the damage indicator, the upper limit of the 99% confidence interval of multiple first damage indicators is used as the damage judgment threshold.

[0043] According to one aspect of the present invention, in step S4, in the step of acquiring a damage recognition model for identifying the damage location based on the damage feature vector and the damage location in the lossy model, the damage recognition model is obtained by training an attention bidirectional temporal convolutional network;

[0044] The attention bidirectional temporal convolutional network includes: an input layer, a forward TCN network and a reverse TCN network connected to the input layer, a fully connected layer connected to the forward TCN network and the reverse TCN network, a Sortmax activation layer connected to the forward connection layer, and an output layer connected to the Sortmax activation layer;

[0045] There are two residual blocks in the forward TCN network and the reverse TCN network respectively.

[0046] According to one aspect of the present invention, a Squeeze and Excitation attention module is inserted into the residual block.

[0047] According to one solution of the present invention, the present invention can perform in-situ detection on composite propeller blades of a large UAV propulsion system and realize the location and identification of damage, which helps to ensure the operational safety of the large UAV.

[0048] According to one solution of the present invention, the transmissibility function used in the present invention is sensitive to damage, and its calculation process is based only on the response signal without the need for an excitation signal, and is therefore suitable for in-situ detection of structures during service.

[0049] According to a solution of the present invention, a new damage index and damage feature vector are proposed based on the uncertainty model of the transmissibility function and the JSD divergence, which can handle uncertainty problems caused by environmental noise, measurement noise, and calculation errors, thereby improving the robustness of damage identification.

[0050] According to one solution of the present invention, the attention bidirectional temporal convolutional network adopted in the present invention can assign dynamic weights to different features, so that the model pays more attention to important features, thereby improving the accuracy of damage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a block diagram schematically illustrating steps of a composite propeller blade damage identification method according to one embodiment of the present invention;

[0052] Figure 2 is a flow chart schematically illustrating a method for identifying damage to a composite propeller blade according to one embodiment of the present invention;

[0053] Figure 3 is a diagram schematically showing a finite element material model of a composite propeller blade according to an embodiment of the present invention;

[0054] Figure 4 is a diagram schematically showing a blade health damage index threshold value according to one embodiment of the present invention;

[0055] Figure 5 is a diagram schematically showing damage indicators of different damage degrees at a certain damage position of a blade according to one embodiment of the present invention;

[0056] Figure 6 Schematically illustrates a structure of an attention bidirectional temporal convolutional network according to an embodiment of the present invention;

[0057] Figure 7 is a diagram schematically showing a test result of a damage identification model according to an embodiment of the present invention;

[0058] Figure 8 is a diagram schematically showing damage indicators of different blades to be tested according to an embodiment of the present invention;

[0059] Figure 9 is a diagram schematically showing a damage characteristic vector of a damaged blade to be tested according to an embodiment of the present invention;

[0060] Figure 10 is a diagram schematically showing a damage characteristic vector of another damaged blade to be tested according to an embodiment of the present invention;

[0061] Figure 11 Schematically shows the damage position identification results of two damaged blades to be tested according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0063] When describing the embodiments of the present invention, the orientation or positional relationship expressed by the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside" and "outside" are based on the orientation or positional relationship shown in the relevant drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the above terms should not be understood as limiting the present invention.

[0064] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described one by one here, but the embodiments of the present invention are not limited to the following embodiments.

[0065] Combine Figure 1 and Figure 2 As shown, according to one embodiment of the present invention, a composite propeller blade damage identification method for a large UAV propulsion system of the present invention includes:

[0066] S1. Establishing multiple finite element material models of a composite propeller blade, wherein the multiple finite element material models include: a lossless model and multiple lossy models;

[0067] S2. Setting a plurality of sensor measurement points on each finite element material model, constructing a transmissibility function between the sensor measurement points, and obtaining a marginal probability density function for each transmissibility function, wherein the marginal probability density function includes: a first marginal probability density function corresponding to the lossless model and a second marginal probability density function corresponding to the lossy model;

[0068] S3. Based on the marginal probability density functions of the multiple finite element material models, obtain a damage characteristic vector and a damage index corresponding to each finite element material model; wherein the damage characteristic vector includes: a first damage characteristic vector corresponding to the lossless model, and a second damage characteristic vector corresponding to the lossy model; and the damage index includes: a first damage index corresponding to the lossless model, and a second damage index corresponding to the lossy model;

[0069] S4. Obtaining a damage judgment threshold for determining whether the composite propeller blade is damaged based on the damage index, and obtaining a damage identification model for identifying the damage location based on the damage feature vector and the damage location in the damage model;

[0070] S5. Obtaining the measured transmissibility function between the sensor measuring points on the blade to be measured, and obtaining the measured damage index and the measured damage feature vector based on the measured transmissibility function;

[0071] S6. Determine whether the damage index to be measured exceeds the damage judgment threshold. If not, determine that the blade to be measured is a healthy blade. If so, obtain the damage position on the blade to be measured based on the damage recognition model and the damage feature vector to be measured.

[0072] According to one embodiment of the present invention, in step S1, the step of establishing multiple finite element material models of composite propeller blades includes:

[0073] S11. Establish a finite element model for a composite propeller blade, and obtain multiple initial models based on the finite element model; wherein the finite element model can adopt an ABAQUS finite element model. In this embodiment, the composite propeller blade includes: a core material and a skin provided on opposite sides of the core material; in this embodiment, the core material is a foam structure, for example, polymethacrylimide (PMI) foam. In this embodiment, the skin is made of a CFRP laminate, and the laminate is composed of two layers of CFRP woven cloth and multiple layers of CFRP unidirectional prepreg, wherein the thickness of the skin gradually decreases from the root (i.e., the fixed end of the composite propeller blade) to the tip. Therefore, the blade is divided into five regions along the length direction of the composite propeller blade, as shown in Table 1, wherein the thickness of each layer of CFRP woven cloth is 0.22 mm, and the thickness of each layer of CFRP unidirectional prepreg is 0.1 mm. The ply directions of the CFRP unidirectional prepreg are mainly 0°, 30°, and -30°.

[0074] Table 1

[0075] area 1 2 3 4 5 Distance to the root 65mm 115mm 190mm 265mm 336mm CFRP woven fabric layers 2 2 2 2 2 Number of CFRP unidirectional prepreg layers 21 13 9 7 3 Total thickness 2.54mm 1.74mm 1.34mm 1.14mm 0.74mm

[0076] Furthermore, the performance parameters of polymethacrylimide (PMI) foam used as the core material, CFRP woven fabric, and CFRP unidirectional prepreg are shown in Table 2.

[0077] Table 2

[0078]

[0079] S12. A plurality of damage samples are obtained by random generation, and the damage samples are appended to the initial model to obtain a plurality of lossy models, and the remaining initial model is used as a lossless model. In this embodiment, a plurality of random damage samples are generated by a Python-based script file, wherein the damage sample is a square damage sample, and its size can be set to 10 mm × 10 mm. The damage position is represented by the x-coordinate and y-coordinate of the Cartesian coordinate system, that is, the center coordinate of the damage in the damage sample, such as Figure 3 As shown. In this embodiment, the damage parameters in the damage sample are generated randomly, wherein the degree of damage is represented by the degree of decrease in the local Young's modulus within the square damage range. In this embodiment, the relevant parameters of the randomly generated damage sample are imported into the ABAQUS finite element model, and the parameters are saved to generate a damage model. The remaining initial model does not need to set the parameters of the damage sample and can be used as a non-destructive model. In this embodiment, one non-destructive model in a healthy state and 360 damage models in different damage states can be generated.

[0080] According to one embodiment of the present invention, in step S2, the step of setting a plurality of sensor measurement points on each finite element material model and constructing a transmissibility function between the sensor measurement points includes:

[0081] S21. Use four-node linear quadrilateral elements to mesh the skin, and use eight-node linear tetrahedral elements to mesh the core. In this embodiment, the mesh size of the four-node linear quadrilateral element (S4R) and the eight-node linear tetrahedral element (C3D4) are both 1 mm.

[0082] S22. Arrange sensor measuring points on the finite element material model based on preset rules and quantities, wherein the arrangement of the sensor measuring points of each finite element material model is consistent; in this embodiment, the number of sensor measuring points can be set to 16, wherein the principle of arranging the sensor measuring points is to obtain as many signals as possible while ensuring that the sensors on the actual blade do not interfere with each other, so the spacing between each sensor measuring point is approximately 30 mm.

[0083] S23. Apply a random vibration signal to the fixed end of the finite element material model, and collect the corresponding strain response signal based on the sensor measuring point; in this embodiment, the fixed end of the finite element material model is fixed (for example, a length of 40 mm from the root is used as a fixed position), and then the Lanczos eigenvalue extraction method is used to obtain the natural frequency and modal vibration shape. A zero-mean, unit-variance normally distributed random vibration signal is applied to the fixed end along the blade radial direction (Z axis), with a bandwidth of 0 to 2500 Hz and an amplitude of 5g; in this embodiment, the transient modal dynamics module in the ABAQUS simulation platform is used to perform dynamic analysis on the finite element material model, and the strain response signal of the sensor measuring point is collected, wherein the time increment of the dynamic analysis is set to 0.0002s, the sampling frequency is 5000Hz, and a group of samples is 5s long. 500 groups of samples are generated for the healthy state and 200 groups of samples are generated for each damage state.

[0084] S24. Calculate the transferability function based on the strain response signal of each sensor measuring point. In this embodiment, the transferability function (TF) between any two sensor measuring points is calculated based on the strain response signal of each sensor measuring point. There are 16×(16-1) / 2, or 120 transferability functions (TF). However, some transferability functions (TF) (such as TF 1,10 , i.e., the transferability function between measurement point 1 and measurement point 10) is too long and overlaps with multiple other paths. Therefore, in order to reduce the dimensionality of the data, only the transferability function between each sensor measurement point and its adjacent measurement points is calculated, for a total of 36 transferability functions (TFs), see Figure 3 shown.

[0085] In this embodiment, the strain response signals measured at the two sensor measurement points m and n are y(m) and y(n), respectively, and the spectrum Y is obtained by fast Fourier transform (FFT). m and Y n , then the frequency between the two points is ω k =k×1 / f s (k is an integer, representing the kth frequency point, f s is the sampling frequency of the strain response signal at the sensor measuring point) is expressed as:

[0086]

[0087] in, Indicates that at frequency ω k The transfer rate function between sensor point m and sensor point n, ω k =k×1 / f s , k represents the kth frequency point, f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, The spectrum of the strain response signal of the sensor measuring point m obtained by fast Fourier transform is represented by The frequency spectrum of the strain response signal at sensor point n obtained by fast Fourier transform.

[0088] According to one embodiment of the present invention, in step S2, the step of obtaining the marginal probability density function of each of the transmissibility functions includes:

[0089] S25. Obtain the covariance matrix of the fast Fourier transform coefficients of the strain response signals of the two sensor measuring points; in this embodiment, the fast Fourier transform (FFT) coefficients of the structural response under random excitation approximately obey a complex Gaussian distribution with a mean of zero, and the covariance matrix of the two fast Fourier transform (FFT) coefficients is expressed as:

[0090]

[0091] in, represents the covariance matrix of the strain response signal of sensor measuring point m and the strain response signal of sensor measuring point n, and They are and The variance of express and The multiple correlation coefficient is expressed as express and The real part correlation coefficient of express and The imaginary correlation coefficient of for conjugation of;

[0092] S26. Obtain the probability density function of the strain response signal of the two sensor measuring points based on the covariance matrix and the transmissibility function; in this embodiment, the transmissibility function (TF) vector obeys a circularly symmetric complex Gaussian ratio distribution, and the corresponding probability density function is expressed as:

[0093]

[0094] in, represents the probability density function of the strain response signal of the sensor measuring point m and the strain response signal of the sensor measuring point n, Express about A process variable, and The superscript T represents the transpose operation. express At frequency ωk is a complex random variable at , and ω k =k×1 / f s , f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, express At frequency ω k A real random variable at , express At frequency ω k An imaginary random variable at ;

[0095] S27. Obtaining the marginal probability density function of the sensor measurement point based on the probability density function; In this embodiment, the complex random variable can be written as but The marginal probability density functions (PDFs) of the real and imaginary parts are:

[0096]

[0097] in, The strain response signal between the sensor measuring point m and the sensor measuring point n at the frequency ω k The real part marginal probability density function at , The strain response signal between the sensor measuring point m and the sensor measuring point n at the frequency ω k The imaginary marginal probability density function at .

[0098] In this embodiment, when a single transmissibility function (TF) is used as a baseline or test sample, uncertainty arises. Therefore, the marginal probability density functions of the real and imaginary parts of the transmissibility function (TF) are used as damage characteristics, and the difference between the marginal probability density function of the test sample and the marginal probability density function of the baseline is constructed as a damage indicator.

[0099] According to one embodiment of the present invention, in step S3, the step of obtaining the damage characteristic vector and damage index corresponding to each finite element material model based on the marginal probability density function of the multiple finite element material models includes:

[0100] S31. For each finite element material model, multiple groups of strain response signals of the sensor measuring points of the finite element material model are obtained based on the applied random vibration signal; in this embodiment, as mentioned above, multiple groups of strain response signals can be obtained by simulating the lossy model and the lossless model respectively, for example, 500 groups are generated corresponding to the lossless model, and 200 groups are generated corresponding to each lossless model.

[0101] S32. Obtain multiple sets of strain response signals corresponding to the non-destructive model, and obtain multiple sets of first marginal probability density functions corresponding to the sensor point pairs on the non-destructive model using step S2; wherein the sensor point pairs are constructed based on the sensor points on the non-destructive model; specifically, randomly select two groups of samples (200 samples each) from the 500 groups of healthy samples corresponding to the non-destructive model, one group as baseline data and the other group as test data, obtain the corresponding strain response signals, and obtain the first marginal probability density functions corresponding to all the sensor point pairs on the non-destructive model using step S2;

[0102] S33. Select any two groups of first marginal probability density functions to obtain the first JSD divergence of each sensor point pair in the lossless model at all frequency points, and average the first JSD divergences of all sensor point pairs to obtain the first damage characteristic vector of the lossless model, and then average the multidimensional data contained in the first damage characteristic vector to obtain the first damage index of the lossless model; in this embodiment, in the step of selecting any two groups of first marginal probability density functions to obtain the first JSD divergence of each sensor point pair in the lossless model at all frequency points, 36 transferability functions (TFs) of each state of the lossless model are calculated, and the 36 transferability functions (TFs) of any two states are selected to calculate multiple first JSD divergences of each sensor point pair in the lossless model, wherein the first JSD divergence (i.e., Jensen-Shannon divergence) is used to evaluate the difference between the marginal probability density functions.

[0103] Furthermore, all the first JSD divergences obtained are averaged to obtain a first damage feature vector of the lossless model, and the multidimensional data included in the first damage feature vector are averaged again to obtain a first damage index of the lossless model.

[0104] In this embodiment, 200 groups of samples in a healthy state, each containing 36 transmissibility functions, each consisting of 512 frequency points, can generate a probability density function for each transmissibility function at each frequency point. Then, 200 groups of samples in another state are selected, and a probability density function for each transmissibility function at each frequency point is similarly generated. Thus, a first JSD divergence can be calculated based on the two probability density functions to obtain the JSD index between the probability density functions of the corresponding frequency points of the transmissibility functions in the two states. This generates 36*512 first JSD divergences. The first JSD divergences of the 512 frequency points in each transmissibility function are then averaged to obtain a 36*1 damage feature vector. All features in the damage feature vector are then averaged to obtain a single first damage feature vector.

[0105] S34. Obtain multiple sets of strain response signals corresponding to the lossy model, and use step S2 to obtain multiple sets of second marginal probability density functions corresponding to the sensor measurement point pairs on the lossy model; wherein the sensor measurement point pairs are constructed based on the sensor measurement points on the lossy model;

[0106] The second marginal probability density function of the sensor point pair on the lossy model and the first marginal probability density function of the sensor point pair on the lossless model are used to obtain the second JSD divergence of each sensor point pair in the lossy model at all frequency points, and the second JSD divergence of all sensor point pairs is averaged to obtain the second damage characteristic vector of the lossy model, and the multidimensional data contained in the second damage characteristic vector is averaged to obtain the second damage index of the lossy model; in this embodiment, the second marginal probability density function of the sensor point pair on the lossy model and the first marginal probability density function of the sensor point pair on the lossless model are used to obtain the second damage index of the lossy model. Density function, in the step of obtaining the second JSD divergence of each sensor measurement point pair in the lossy model at all frequency points, 36 transferability functions (TFs) of the lossless model are calculated, as well as 36 transferability functions (TFs) of each state of the lossless model. The 36 transferability functions (TFs) of any state of the lossless model and the 36 transferability functions (TFs) of the lossy model are selected to calculate multiple second JSD divergences between each sensor measurement point pair in the lossy model and the lossless model, wherein the second JSD divergence (i.e., Jensen-Shannon divergence) is used to evaluate the difference between the marginal probability density functions.

[0107] Furthermore, all obtained second JSD divergences are averaged to obtain a second damage feature vector for the lossy model, and the multidimensional data contained in the second damage feature vector is averaged again to obtain a second damage index for the lossy model. In this embodiment, the above calculation steps are implemented based on 200 sets of samples from the lossy model and 200 sets of samples from the lossless model.

[0108] According to one embodiment of the present invention, the first JSD divergence and the second JSD divergence can be obtained using a unified JSD divergence calculation formula, and the JSD divergence is expressed as:

[0109]

[0110] Where KLD(P||Q) represents the KLD divergence between the probability density function P(x) and the probability density function Q(x), JSD(P||Q) represents the JSD divergence between the probability density function P(x) and the probability density function Q(x), and P(x) and Q(x) represent the two marginal probability density functions involved in the JSD divergence calculation.

[0111] According to one embodiment of the present invention, in step S4, in the step of obtaining a damage judgment threshold for determining whether a composite propeller blade is damaged based on the damage index, the upper limit of the 99% confidence interval of the plurality of first damage indexes is used as the damage judgment threshold. In this embodiment, in order to obtain the threshold of the damage index, two groups of samples (200 samples each) are randomly selected from 500 healthy samples, one group of which is used as baseline data and the other group as test data. The random sampling and calculation are repeated 50 times to generate 50 damage indicators, such as Figure 4 As shown, the upper limit of the 99% confidence interval of DI is used as the threshold, as shown by the red dotted line. Figure 5 The damage index at a certain damage position of the blade has different damage degrees. When the damage degree is greater than or equal to 20%, the damage index exceeds the threshold, and the damage index increases with the increase of the damage degree, thereby realizing damage detection.

[0112] like Figure 1 As shown, according to one embodiment of the present invention, in step S4, in the step of obtaining a damage identification model for identifying the damage location based on the damage feature vector and the damage location in the damage model, the damage feature vector of the finite element material model in each damage state is calculated, a set of damage feature vectors containing different damages is obtained, and the damage feature vector set is used to train the damage identification model for identifying the damage location. In this model, the input is the damage feature vector of the finite element material model in each damage state, and the output is the damage location of each damage state, that is, the X-axis coordinate and the Y-axis coordinate.

[0113] like Figure 6 As shown, in this embodiment, the damage recognition model is obtained by training an attention bidirectional temporal convolutional network; wherein the attention bidirectional temporal convolutional network includes: an input layer, a forward TCN network and a reverse TCN network connected to the input layer, a fully connected layer connected to the forward TCN network and the reverse TCN network, a Sortmax activation layer connected to the forward connection layer, and an output layer connected to the Sortmax activation layer; in this embodiment, two residual blocks are respectively set in the forward TCN network and the reverse TCN network, wherein the convolution kernel size is 3, the number of convolution kernels in each layer is 64, and the expansion factor is set to d=2 (n-1) , where n represents the nth convolutional layer. In this embodiment, a Squeeze and Excitation attention module is inserted into each residual block to assign weights to input features, allowing the model to learn more important features, thereby improving model accuracy. In this embodiment, the batch size is set to 32, the learning rate is 0.001, the Adam optimizer is applied, and the mean squared error (MSE) minimization is used as the loss function.

[0114] In this embodiment, among the 360 ​​damage models with different damage states, 70% are used for training the model, 20% are used for verifying the model, and the remaining 10% are used for testing the model. The model test results are shown in FIG. Figure 7 As shown, Figure 7 (a) is the prediction of the X-axis coordinate of the damage in the test sample by the damage identification model. Figure 7 (b) is the prediction of the Y-axis coordinate of the damage in the test sample by the damage identification model. Figure 7 (c) is the comprehensive prediction result of the damage identification model on the damage location in the test sample.

[0115] According to one embodiment of the present invention, in step S5, in the step of obtaining the measured transmissibility function between sensor measuring points on the blade to be measured and obtaining the measured damage index and the measured damage characteristic vector based on the measured transmissibility function, multiple strain gauges are arranged on the blade to be measured, wherein the arrangement positions of the strain gauges are consistent with the positions of the sensor measuring points in the aforementioned step. The blade to be measured is fixed to a vibration table, and the vibration table is controlled by a controller to generate an experimental vibration excitation signal of random vibration, thereby acquiring a corresponding experimental strain response signal based on the strain gauges on the blade to be measured.

[0116] In this embodiment, a special clamping tool is used to hold the blade to be tested, wherein the root of the blade to be tested is fixed and the fixed length is 40 mm. In this embodiment, the experimental vibration excitation signal applied to the blade to be tested is the same as the vibration excitation signal in the above steps. In this embodiment, the blade to be tested can be set to three, one of which is a non-destructive blade to be tested and two are damaged blades to be tested, wherein the structure of the blade to be tested is consistent with the structure of the above-mentioned finite element material model. In this embodiment, the damaged blade to be tested uses a double layer of 0.03 mm thick polytetrafluoroethylene film embedded between two layers to cause local debonding of the composite material during the curing process to simulate delamination damage. Specifically, the damage area of ​​the two damaged blades to be tested is 10 mm × 10 mm. As in the above steps, the delamination damage is distributed in two different positions, namely (1) X-axis coordinate is 100 mm, Y-axis coordinate is -15 mm; (2) X-axis coordinate is 170 mm, Y-axis coordinate is 5 mm.

[0117] In this way, the strain response signals of one healthy blade and two damaged blades can be collected, with a sampling frequency of 5000 Hz and a sampling time of 1000 s.

[0118] In this embodiment, the transferability function of the blade to be tested is calculated based on the experimental strain response signals of each measuring point of the blade to be tested, and the damage index DI of the healthy blade to be tested and the two damaged blades to be tested are calculated, such as Figure 8As shown in the figure, the damage index of the healthy blade to be tested is less than the threshold and is correctly judged to be in a healthy state. The damage indexes of the damaged blades to be tested at two different damage locations are both greater than the threshold and are correctly judged to be in a damaged state.

[0119] According to one embodiment of the present invention, in step S6, in the step of obtaining the damage position on the blade to be measured based on the damage recognition model and the damage feature vector to be measured, the damage features (DF) of the damaged blade to be measured at two different damage positions are calculated, respectively as follows: Figure 9 and Figure 10 As shown. The damage characteristics of the damaged blades at two different damage locations are brought into the damage identification model to obtain the predicted X-axis coordinates and Y-axis coordinates of the damage as shown in Figure 11 As shown, the prediction errors of the X-axis coordinate and Y-axis coordinate of the damage in the first damaged blade to be tested are both less than 5 mm, and the prediction errors of the X-axis coordinate and Y-axis coordinate of the damage in the second damaged blade to be tested are 8.6 mm and 5.8 mm, respectively, which are both within the acceptable range in engineering.

[0120] The above contents are merely examples of specific solutions of the present invention. For devices and structures not described in detail, it should be understood that they can be implemented by adopting general devices and methods available in the art.

[0121] The above description is merely one embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for identifying damage to composite propeller blades in a large UAV propulsion system, characterized in that: include: S1. Establishing multiple finite element material models of a composite propeller blade, wherein the multiple finite element material models include: a lossless model and multiple lossy models; S2. Setting a plurality of sensor measurement points on each of the finite element material models, constructing a transmissibility function between the sensor measurement points based on the transmissibility functions, and obtaining a marginal probability density function for each of the transmissibility functions, wherein the marginal probability density functions include: a first marginal probability density function corresponding to the lossless model and a second marginal probability density function corresponding to the lossy model; S3. Based on the marginal probability density functions of the plurality of finite element material models, obtaining a damage characteristic vector and a damage index corresponding to each of the finite element material models; wherein the damage characteristic vector includes: a first damage characteristic vector corresponding to the lossless model, and a second damage characteristic vector corresponding to the lossy model; and the damage index includes: a first damage index corresponding to the lossless model, and a second damage index corresponding to the lossy model; S4. obtaining a damage judgment threshold for determining whether the composite propeller blade is damaged based on the damage indicator, and obtaining a damage identification model for identifying the damage location based on the damage feature vector and the damage location in the damage model; S5. Obtaining a measured transmissibility function between sensor measuring points on the blade to be measured, and obtaining a measured damage index and a measured damage feature vector based on the measured transmissibility function; S6. Determine whether the damage index to be measured exceeds the damage judgment threshold. If not, determine that the blade to be measured is a healthy blade. If so, obtain the damage position on the blade to be measured based on the damage recognition model and the damage feature vector to be measured.

2. The composite propeller blade damage identification method according to claim 1, characterized in that: In step S1, the step of establishing multiple finite element material models of composite propeller blades includes: S11. Establishing a finite element model of the composite propeller blade, and obtaining a plurality of initial models based on the finite element model; wherein the composite propeller blade comprises: a core material and skins disposed on opposite sides of the core material; S12. Extract one of the initial models as a lossless model, and obtain multiple damaged samples in a random generation manner, and append the damaged samples to the remaining initial models to obtain multiple lossy models.

3. The composite propeller blade damage identification method according to claim 2, characterized in that: In step S2, a plurality of sensor measuring points are respectively set on each of the finite element material models, and a transmissibility function between the sensor measuring points is constructed based on the sensor measuring points, including: S21. Meshing the skin using four-node linear quadrilateral elements and meshing the core using eight-node linear tetrahedral elements; S22. Arranging the sensor measuring points on the finite element material model based on preset rules and quantities, wherein the arrangement of the sensor measuring points of each finite element material model is consistent; S23. Applying a random vibration signal to the fixed end of the finite element material model and collecting a corresponding strain response signal based on the sensor measuring point; S24. Calculate the transferability function based on the strain response signal of each sensor measuring point.

4. The composite propeller blade damage identification method according to claim 3, characterized in that: The transmissibility function is expressed as: in, Indicates the frequency ω k Sensor measurement points m and sensor points n The transfer rate function between ω k = k ×1 / f s , k Indicates the k frequency points, f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, Indicates sensor measurement point m The spectrum of the strain response signal obtained by fast Fourier transform is Indicates sensor measurement point n The spectrum of the strain response signal obtained by fast Fourier transform.

5. The composite propeller blade damage identification method according to claim 3 or 4, characterized in that: In step S2, the step of obtaining the marginal probability density function of each of the transmissibility functions includes: S25. Obtain the covariance matrix of the fast Fourier transform coefficients of the strain response signals of the two sensor measuring points, which is expressed as: in, Indicates sensor measurement point m The strain response signal and sensor measurement point n The covariance matrix of the strain response signal, and They are and The variance of express and The multiple correlation coefficient is expressed as , express and The real part correlation coefficient of express and The imaginary correlation coefficient of for conjugation of; S26. Obtaining the probability density function of the strain response signals of the two sensor measuring points based on the covariance matrix and the transmissibility function, which is expressed as: in, Indicates sensor measurement point m The strain response signal and sensor measurement point n The probability density function of the strain response signal, Express about A process variable, and , superscript T represents the transpose operation, express In frequency is a complex random variable at , and , , f s Indicates the sampling frequency of the strain response signal at the sensor measuring point, express In frequency A real random variable at , express In frequency An imaginary random variable at ; S27. Obtain the marginal probability density function of the sensor measurement point based on the probability density function, which is expressed as: in, Indicates sensor measurement point m The strain response signal and sensor measurement point n The strain response signal between the frequency The real part marginal probability density function at , Indicates sensor measurement point m The strain response signal and sensor measurement point n The strain response signal between the frequency The imaginary marginal probability density function at .

6. The composite propeller blade damage identification method according to claim 5, characterized in that: In step S3, the step of obtaining the damage characteristic vector and damage index corresponding to each of the finite element material models based on the marginal probability density functions of the plurality of finite element material models includes: S31. For each of the finite element material models, respectively, obtaining strain response signals of sensor measuring points of multiple groups of the finite element material models based on the applied random vibration signal; S32. Acquire multiple sets of strain response signals corresponding to the non-destructive model, and obtain multiple sets of the first marginal probability density functions corresponding to sensor point pairs on the non-destructive model using step S2; wherein the sensor point pairs are constructed based on the sensor points on the non-destructive model; S33. Select any two groups of the first marginal probability density functions to obtain first JSD divergences of each sensor point pair in the lossless model at all frequency points, average the first JSD divergences of all sensor point pairs to obtain a first damage eigenvector of the lossless model, and further average the multidimensional data contained in the first damage eigenvector to obtain a first damage index of the lossless model; S34. Obtain multiple sets of strain response signals corresponding to the lossy model, and obtain multiple sets of the second marginal probability density functions corresponding to the sensor point pairs on the lossy model using step S2; wherein the sensor point pairs are constructed based on the sensor points on the lossy model; The second marginal probability density function of the sensor point pair on the lossy model and the first marginal probability density function of the sensor point pair on the lossless model are used to obtain a second JSD divergence of each sensor point pair in the lossy model at all frequency points. The second JSD divergences of all the sensor point pairs are averaged to obtain a second damage feature vector of the lossy model. The multidimensional data contained in the second damage feature vector is further averaged to obtain a second damage indicator of the lossy model.

7. The composite propeller blade damage identification method according to claim 6, characterized in that: The JSD divergence is expressed as: in, represents the probability density function P ( x ) and the probability density function Q ( x ) KLD Divergence, represents the probability density function P ( x ) and the probability density function Q ( x ) JSD Divergence, and Represent two marginal probability density functions involved in the JSD divergence calculation.

8. The composite propeller blade damage identification method according to claim 6, characterized in that: In step S4, in the step of obtaining a damage judgment threshold for judging whether the composite propeller blade is damaged based on the damage indicator, the upper limit of the 99% confidence interval of multiple first damage indicators is used as the damage judgment threshold.

9. The composite propeller blade damage identification method according to claim 6, characterized in that: In step S4, in the step of acquiring a damage recognition model for identifying the damage location based on the damage feature vector and the damage location in the lossy model, the damage recognition model is obtained by training an attention bidirectional temporal convolutional network; The attention bidirectional temporal convolutional network includes: an input layer, a forward TCN network and a reverse TCN network connected to the input layer, a fully connected layer connected to the forward TCN network and the reverse TCN network, a Softmax activation layer connected to the fully connected layer, and an output layer connected to the Softmax activation layer; There are two residual blocks in the forward TCN network and the reverse TCN network respectively.

10. The composite propeller blade damage identification method according to claim 9, characterized in that: A Squeeze and Excitation attention module is inserted into the residual block.

Images