Shock location method for composite structures based on wavelet packet Tsallis entropy and LLE

Abstract

The present invention discloses a composite material structure impact location method based on wavelet packet Tsallis entropy and LLE. The method comprises the following steps: using an EFPI optical fiber sensor network to monitor impact response signals at different positions on a structure, extracting a wavelet packet Tsallis entropy spectrum of the impact response signal, and using a local linear embedding method to reduce the dimension of the wavelet packet Tsallis entropy spectrum; utilizing the relationship between the wavelet packet Tsallis entropy spectrum of the impact point response signal after the dimension reduction and the impact point position, and using a GA-CNN-BiLSTM-Attention model to realize impact position identification; the present invention analyzes the wavelet packet Tsallis entropy spectrum of the impact response signal monitored by the EFPI sensor on the composite material structure, uses the LLE technology to extract feature vectors from the wavelet packet Tsallis entropy spectrum, and uses the GA-CNN-BiLSTM-Attention neural network model to accurately identify the impact load position of the composite material structure. The method has the characteristics of not requiring a high-speed data acquisition module, having a small amount of signal feature data, being widely applicable, and having good generalization ability of the identification model.

Description

Technical Field

[0001] The present invention relates to the technical field of composite material structure impact positioning, in particular to a composite material structure impact positioning method based on wavelet packet Tsallis entropy and LLE. Background Art

[0002] Composite materials, due to their lightweight, high strength, and corrosion resistance, are widely used in aerospace, automotive, bridge, and other structural applications. However, they are highly sensitive to shock loads and are highly susceptible to impact during manufacturing, transportation, service, and recycling. These subtle damages are difficult to detect using traditional maintenance procedures, threatening structural safety and even leading to catastrophic consequences. To assess damage, prevent accidents, and improve the efficiency of composite material damage detection, the development of composite structural health monitoring technologies with impact location capabilities is crucial.

[0003] Current research on impact load location methods based on the propagation of Lamb waves caused by impact on structural surfaces typically employs methods based on signal arrival time and angle of arrival using sensor arrays arranged in triangles, quadrilaterals, lines, or rings. While these methods have achieved good results in identifying impact locations, they do not guarantee accurate location identification. However, Lamb wave propagation velocities vary in different directions across different materials and structural shapes, making location accuracy dependent on the algorithm used to determine signal arrival time and placing extremely high demands on the sampling rate of the sensor response signal. Furthermore, methods based on signal similarity comparison are sensitive to noise in the impact response signal and exhibit limited anti-interference capabilities.

[0004] To address the requirements for impact monitoring and location on composite structures and the time-frequency characteristics of composite structure impact response signals, a new method was developed that requires no prior knowledge, is adaptable to data acquisition modules with low sampling frequencies, and exhibits a certain degree of noise immunity. To this end, this paper proposes a composite structure impact location method based on wavelet packet Tsallis entropy and LLE. Summary of the Invention

[0005] The purpose of the present invention is to solve the technical problems existing in the prior art and provide a composite material structure impact location method based on wavelet packet Tsallis entropy and LLE.

[0006] To achieve the above object, the present invention provides a technical solution: a composite material structure impact location method based on wavelet packet Tsallis entropy and LLE, the method comprising the following steps:

[0007] Step 1: Use EFPI sensors to monitor the impact response signals on the composite structure, and build a composite impact monitoring system based on EFPI optical fiber sensors to collect the impact point response signals;

[0008] Step 2: Calculate the wavelet packet Tsallis entropy spectrum of the impact point response signal collected in step 1;

[0009] Step 3: Use the LLE method to perform feature dimension reduction on the wavelet packet Tsallis entropy spectrum of the impact point response signal in step 2 to achieve feature extraction of the impact point response signal;

[0010] Step 4: Use the GA-CNN-BiLSTM-Attention neural network model to identify the impact load position on the composite structure.

[0011] Preferably, the impact monitoring system in step 1 includes: a broadband light source, a 1×4 coupler, a first optical circulator, a second optical circulator, a third optical circulator, a fourth optical circulator, a first EFPI sensor, a second EFPI sensor, a third EFPI sensor, a fourth EFPI sensor, a fifth optical circulator, a sixth optical circulator, a seventh optical circulator, an eighth optical circulator, a first FBG sensor, a second FBG sensor, a third FBG sensor, a fourth FBG sensor, a first photodetector, a second photodetector, a third photodetector, a fourth photodetector, a data acquisition module, and a composite material structure;

[0012] The light output by the broadband light source is split into four light paths through a 1×4 coupler. The first light path passes through a first optical circulator and reaches a first EFPI sensor. The light reflected from the first EFPI sensor then passes through a fifth optical circulator and reaches a first FBG sensor. The reflected light from the first FBG sensor is received by a first photodetector.

[0013] The second light path passes through the second optical circulator to reach the second EFPI sensor. The light reflected from the second EFPI sensor passes through the sixth optical circulator to reach the second FBG sensor. The reflected light from the second FBG sensor is received by the second photodetector.

[0014] The third light path passes through the third optical circulator to the third EFPI sensor. The light reflected by the third EFPI sensor passes through the seventh optical circulator to the third FBG sensor. The reflected light of the third FBG sensor is received by the third photodetector.

[0015] The fourth light path passes through a fourth optical circulator to reach a fourth EFPI sensor. The light reflected by the fourth EFPI sensor then passes through an eighth optical circulator to reach a fourth FBG sensor. The reflected light from the fourth FBG sensor is received by a fourth photodetector. Subsequently, a data acquisition module is used to collect impulse response signals output by the first photodetector, the second photodetector, the third photodetector, and the fourth photodetector.

[0016] Preferably, the specific steps of arranging sensors of the impact monitoring system and obtaining the impact point response signal in step 1 are:

[0017] Step 1-1: Fix the composite structural part on a workbench, construct a square impact monitoring area on its surface, and divide it into uniform grids to obtain N grid nodes;

[0018] Step 1-2: Select the lower left corner of the square area as the coordinate origin (0mm, 0mm), horizontally to the right as the positive direction of the X axis, and vertically upward as the positive direction of the Y axis to establish a rectangular coordinate system;

[0019] Step 1-3: Use epoxy resin adhesive to adhere the first, second, third, and fourth EFPI sensors to the four vertices of the square monitoring area on the surface of the composite structural component. Connect the EFPI sensor optical path via fiber optic patch cables to build an impact monitoring system.

[0020] Step 1-4: Use an impact hammer with a fixed energy to apply n impact loads of the same energy to each grid node in the impact monitoring area on the surface of the composite component, and record the original response signal of each EFPI sensor under each impact load.

[0021] Preferably, the specific calculation method of the wavelet packet Tsallis entropy spectrum of the impact point response signal in step 2 is:

[0022] Step 2-1: remove the DC component of the collected original impulse response signal to obtain the time series of the impulse response signal after removing the baseline;

[0023] Step 2-2: Select appropriate wavelet basis function to perform m-layer wavelet packet decomposition on the impulse response signal after removing DC component, and get 2 m The decomposition coefficients of nodes are obtained, and then the wavelet packet decomposition coefficients are reconstructed to obtain the reconstructed signal sequence X of the i-th node under the m-layer wavelet packet decomposition. m,i ={x m,i (k), k=1,2,...,M}, where the length of the reconstructed signal sequence is M;

[0024] Step 2-3: The wavelet packet Tsallis entropy calculation formula of the reconstructed signal of this node is as follows:

[0025]

[0026] Where q is a non-extensive parameter;

[0027] Step 2-4: Calculate the wavelet packet Tsallis entropy of each node in the m-layer wavelet packet decomposition of the baseline-removed shock response signal of each EFPI sensor after the impact load, and merge them into the wavelet packet Tsallis entropy spectrum feature column vector:

[0028] Tsae={Tsa1;Tsa2;Tsa3;Tsa4}

[0029] where the column vector They respectively represent the wavelet packet Tsallis entropy of all nodes in the m-layer wavelet packet decomposition of the baseline-removed impact response signals of the first EFPI sensor, the second EFPI sensor, the third EFPI sensor, and the fourth EFPI sensor after the impact load.

[0030] Preferably, the specific steps of feature dimensionality reduction in step 3 are as follows:

[0031] Step 3-1: Summarize the wavelet packet Tsallis entropy spectra of all N×n sample impact points and h unknown impact points to be located, and obtain a 4×2 m =2 m+2 The wavelet packet Tsallis entropy characteristic matrix T with N rows × n + h columns is:

[0032] T=[Tsae 1，1 ,Tsae 1，2 ,...,Tsae 1，n ,Tsae 2，1 ,Tsae 2，2 ,...,Tsae 2，n ,...,Tsae N，1 ,Tsae N，2 ,...,Tsae N，n ,Tsae 0，1 ,Tsae 0，2 ,...,Tsae 0，h ]

[0033] =[t1,t2,...,t i ,...,t N×n+h ]

[0034] Among them Tsae a,b It represents the wavelet packet Tsallis entropy spectrum characteristic column vector of the bth sample impact load applied to the ath grid node, Tsae 0,s The wavelet packet Tsallis entropy spectrum characteristic column vector representing the sth unknown impact load to be located;

[0035] Step 3-2: Calculate the Euclidean distance between any two impact point wavelet packet Tsallis entropy spectrum eigenvectors, and obtain the wavelet packet Tsallis entropy spectrum eigenvector t from the impact point i The Tsallis entropy spectrum eigenvectors of the wavelet packet of the nearest K impact points are used as the neighborhood;

[0036] Step 3-3: Calculate the reconstruction weight matrix W by minimizing the reconstruction error function Get the reconstruction matrix W:

[0037]

[0038] Where: w ij t i With t j The weight coefficients between If t j Not t i The neighboring point of ij is 0; therefore:

[0039]

[0040]

[0041] Where Z i =(t i -t j )(t i -t j ) T , l K is a K-dimensional all-one vector. The weight coefficient can be obtained by the above two formulas;

[0042]

[0043] Step 3-4: By minimizing the loss function Get the low-dimensional embedding result Y. The loss function and constraints are as follows:

[0044]

[0045] Where: I is the unit matrix, M = (IW) T (IW), y i ,y j x i ,x j Image in low-dimensional space;

[0046] Step 3-5: Sort the eigenvalues ​​of the matrix M from small to large, and select the eigenvectors corresponding to the K smallest non-zero eigenvalues ​​of the matrix M to form the resulting matrix U after the dimensionality reduction of the initial eigenvector matrix T. U is a matrix with K rows and N × n + h columns.

[0047] Preferably, the specific steps of using the GA-CNN-BiLSTM-Attention neural network model to identify the impact load position on the composite material structure in step 4 are:

[0048] Step 4-1: Transpose the result matrix U extracted in step 3-5, and take the Tsallis entropy spectrum of the wavelet packet of the sample impact point response signal after dimensionality reduction of the first N×n rows and the corresponding impact point coordinate position as the sample data set;

[0049] Step 4-2: Input the sample dataset data into the GA-CNN-BiLSTM-Attention neural network model and use the genetic algorithm to optimize the number and size of convolution kernels in the CNN-BiLSTM-Attention model;

[0050] Step 4-3: Input the optimal parameters obtained by genetic algorithm optimization and the wavelet packet Tsallis entropy spectrum characteristics of the response signal of the unknown impact point to be located after LLE dimensionality reduction into the CNN-BiLSTM-Attention neural network model to realize the location of the impact load on the composite structure.

[0051] Preferably, the specific process steps of optimizing the CNN-BiLSTM-Attention neural network model by the genetic algorithm in step 4-2 are as follows:

[0052] (1) Initialize the parameters of the genetic algorithm-optimized CNN-BiLSTM-Attention neural network, randomly divide the sample data set into a training set and a test set, and use the average positioning error of the test set as the fitness function of the genetic algorithm. The formula for the average positioning error is as follows:

[0053]

[0054] in Represents the average distance error between the predicted value and the true value, x j Indicates the actual horizontal coordinate value, x' j Indicates the predicted horizontal coordinate value, y j Indicates the actual vertical coordinate value, y' j Represents the predicted ordinate value, and n represents the number of test sets;

[0055] (2) Selection operation: Introduce an elite selection strategy to retain the best individuals and directly enter the next generation to replace individuals with poor fitness;

[0056] (3) Crossover operation: Using the real number crossover method, the gene values ​​of the parent individuals are linearly combined to generate new offspring individuals, thereby retaining the excellent characteristics;

[0057] (4) Mutation operation: using the exchange mutation method, two gene positions in an individual are randomly selected and their values ​​are exchanged to generate a new individual;

[0058] (5) Repeat operations (2) to (4) until the preset number of iterations is reached, and select the optimal individual parameter value as the parameters for the number and size of convolution kernels in the CNN-BiLSTM-Attention neural network model.

[0059] Preferably, the specific process steps of the GA-CNN-BiLSTM-Attention neural network model in step 4-3 for locating the impact load on the composite material structure are as follows:

[0060] (1) Set the input and output dimensions of the GA-CNN-BiLSTM-Attention neural network;

[0061] (2) The wavelet packet Tsallis entropy spectrum feature of the response signal of the unknown impact point to be located obtained after LLE dimensionality reduction is input into the sequence folding layer of the trained CNN-BiLSTM-Attention neural network; the sequence folding layer receives the data and folds it into a small batch format to improve computational efficiency and enable the model to more effectively learn the temporal features within the sequence;

[0062] (3) The convolution layer compresses the input data and extracts features through sliding window operations;

[0063] (4) The sequence refolding layer is responsible for unfolding the folded sequence data processed by the convolution layer, thereby ensuring that the subsequent BiLSTM layer can correctly receive the data and effectively capture the long-term dependencies in the time series;

[0064] (5) The expanded sequence data is passed to the BiLSTM layer for bidirectional time series feature learning and capturing context information;

[0065] (6) The Attention layer automatically weights time information, highlights important information, and improves data processing efficiency;

[0066] (7) Finally, the impact point position coordinate prediction results are output through the Dropout layer and the fully connected layer.

[0067] Beneficial effects of the present invention:

[0068] The present invention analyzes the wavelet packet Tsallis entropy spectrum of the impact response signal monitored by the EFPI sensor on the composite material structure, adopts the LLE technology to extract feature vectors from the wavelet packet Tsallis entropy spectrum, and uses the GA-CNN-BiLSTM-Attention neural network model to accurately identify the impact load position of the composite material structure. It has the characteristics of not requiring a high-speed data acquisition module, small amount of signal feature data, wide applicability, and good generalization ability of the identification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The drawings described herein are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0070] Figure 1 This is a schematic diagram of the impact location system for composite material structures based on EFPI optical fiber sensors in the present invention;

[0071] Figure 2 It is the overall flow chart of the present invention;

[0072] Figure 3 This is a diagram of the impact positioning results of the present invention.

[0073] Figure annotation:

[0074] 1- broadband light source, 2- 1×4 coupler, 3- first optical circulator, 4- second optical circulator, 5- third optical circulator, 6- fourth optical circulator, 7- first EFPI sensor, 8- second EFPI sensor, 9- third EFPI sensor, 10- fourth EFPI sensor, 11- fifth optical circulator, 12- sixth optical circulator, 13- seventh optical circulator, 14- eighth optical circulator, 15- first FBG sensor, 16- second FBG sensor, 17- third FBG sensor, 18- fourth FBG sensor, 19- first photodetector, 20- second photodetector, 21- third photodetector, 22- fourth photodetector, 23- data acquisition module, 24- composite material structure. DETAILED DESCRIPTION

[0075] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it should not be understood as a limitation on the scope of protection of the present invention.

[0076] In the description of the present invention, if there is a description of first and second, it is only for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0077] The present invention provides a method for locating impacts in composite structures based on wavelet packet Tsallis entropy and LLE. This method uses an EFPI fiber optic sensor network to monitor impact response signals at different locations on the structure, extracts the wavelet packet Tsallis entropy spectrum of the impact response signal, and uses a local linear embedding method to reduce the dimension of the wavelet packet Tsallis entropy spectrum. The relationship between the wavelet packet Tsallis entropy spectrum of the impact point response signal after dimensionality reduction and the impact point location is then used to identify the impact location using a GA-CNN-BiLSTM-Attention model. This method is suitable for quantitatively estimating the time-frequency domain characteristics of the impact signal and can be applied to the impact location of four-side clamped composite structures in the aerospace field. This method does not require extensive prior knowledge and can be applied to data acquisition modules with lower sampling frequencies to identify the location of impact loads on composite structures.

[0078] Reference Figure 1-Figure 3 The present invention provides a composite material structure impact location method based on wavelet packet Tsallis entropy and LLE, the method comprising the following steps:

[0079] Step 1: Use EFPI sensors to monitor the impact response signals on the composite structure, and build a composite impact monitoring system based on EFPI optical fiber sensors to collect the impact point response signals;

[0080] Step 2: Calculate the wavelet packet Tsallis entropy spectrum of the impact point response signal collected in step 1;

[0081] Step 3: Use the LLE method to perform feature dimension reduction on the wavelet packet Tsallis entropy spectrum of the impact point response signal in step 2 to achieve feature extraction of the impact point response signal;

[0082] Step 4: Use the GA-CNN-BiLSTM-Attention neural network model to identify the impact load position on the composite structure.

[0083] Furthermore, the impact monitoring system in step 1 includes: a broadband light source 1, a 1×4 coupler 2, a first optical circulator 3, a second optical circulator 4, a third optical circulator 5, a fourth optical circulator 6, a first EFPI sensor 7, a second EFPI sensor 8, a third EFPI sensor 9, a fourth EFPI sensor 10, a fifth optical circulator 11, a sixth optical circulator 12, a seventh optical circulator 13, an eighth optical circulator 14, a first FBG sensor 15, a second FBG sensor 16, a third FBG sensor 17, a fourth FBG sensor 18, a first photodetector 19, a second photodetector 20, a third photodetector 21, a fourth photodetector 22, a data acquisition module 23, and a composite material structure 24;

[0084] The light output by the broadband light source 1 is split into four light paths by the 1×4 coupler 2. The first light path passes through the first optical circulator 3 and reaches the first EFPI sensor 7. The light reflected by the first EFPI sensor 7 then passes through the fifth optical circulator 11 and reaches the first FBG sensor 15. The reflected light from the first FBG sensor 15 is received by the first photodetector 19.

[0085] The second light path passes through the second optical circulator 4 and reaches the second EFPI sensor 8. The light reflected by the second EFPI sensor 8 then passes through the sixth optical circulator 12 and reaches the second FBG sensor 16. The reflected light from the second FBG sensor 16 is received by the second photodetector 20.

[0086] The third light path passes through the third optical circulator 5 and reaches the third EFPI sensor 9. The light reflected by the third EFPI sensor 9 then passes through the seventh optical circulator 13 and reaches the third FBG sensor 17. The reflected light from the third FBG sensor 17 is received by the third photodetector 21.

[0087] The fourth light path passes through the fourth optical circulator 6 and reaches the fourth EFPI sensor 10. The light reflected by the fourth EFPI sensor 10 then passes through the eighth optical circulator 14 and reaches the fourth FBG sensor 18. The reflected light from the fourth FBG sensor 18 is received by the fourth photodetector 22. Subsequently, the data acquisition module 23 is used to collect the impulse response signals output by the first photodetector 19, the second photodetector 20, the third photodetector 21, and the fourth photodetector 22.

[0088] The present invention builds a composite material impact monitoring system based on EFPI sensors. This sensor monitoring system has a simple structure and high sensitivity. Compared with piezoelectric sensor arrays, it has the characteristics of anti-electromagnetic interference and low cost, and can effectively monitor impact vibration signals.

[0089] The present invention calculates the Tsallis entropy spectrum of the impact point response signal wavelet packet and uses the LLE method to reduce the dimension of the Tsallis entropy spectrum of the impact response signal wavelet packet to achieve impact response signal feature extraction. This method is suitable for non-stationary impact signals and can efficiently and accurately characterize the signal's time-frequency domain characteristics. In the process of calculating the Tsallis entropy of the impact point response signal wavelet packet, the wavelet basis function is usually db4. The size of the impact monitoring area grid division and the dimension of the resulting matrix after the LLE method dimensionality reduction depend on the requirements for positioning accuracy and positioning speed in actual engineering. This calculation method is suitable for processing the impact positioning problem of composite materials structure, with a small amount of data and good robustness.

[0090] The present invention constructs an impact point position identification model on composite material structures through the GA-CNN-BiLSTM-Attention model and optimizes the model parameters through a genetic algorithm. This method is suitable for the identification of non-stationary impact signals and can effectively establish a relationship model between the Tsallis entropy spectrum of the wavelet packet of the impact response signal and the impact point position. The identification model has good generalization ability.

[0091] The present invention is applicable to a data acquisition module with a low sampling frequency and can locate the impact load without a large amount of prior knowledge, thereby enhancing engineering practicability.

[0092] Example 1

[0093] The present invention provides a composite material structure impact location method based on wavelet packet Tsallis entropy and LLE, and the specific implementation steps are as follows:

[0094] Step 1: Use EFPI sensors to monitor the impact response signals on the composite structure, and build a composite impact monitoring system based on EFPI optical fiber sensors to collect the impact point response signals;

[0095] Step 1-1: Fix the composite structural member 24 on a workbench, select a 400 × 400 mm square area at the center of its surface as the impact monitoring area, and divide the monitoring area into a uniform grid with a spacing of 5 mm, obtaining 17 rows and 17 columns with a total of 289 grid nodes;

[0096] Step 1-2: Select the lower left corner of the square area as the coordinate origin (0mm, 0mm), horizontally to the right as the positive direction of the X axis, and vertically upward as the positive direction of the Y axis to establish a rectangular coordinate system;

[0097] Step 1-3: Using epoxy resin adhesive, adhere the first EFPI sensor 7, the second EFPI sensor 8, the third EFPI sensor 9, and the fourth EFPI sensor 10 to the four vertices of the square monitoring area on the surface of the composite structural member 24. Connect the EFPI sensor optical path via optical fiber jumpers to build an impact monitoring system.

[0098] Step 1-4: Use an impact hammer with a fixed energy of 1 J to apply an impact load of the same energy to each grid node in the impact monitoring area on the surface of the composite component 24 three times, and record the original response signal of each EFPI sensor under each impact load.

[0099] Step 2: Calculate the Tsallis entropy spectrum of the impact point response signal wavelet packet;

[0100] Step 2-1: remove the DC component of the collected original impulse response signal to obtain the time series of the impulse response signal after removing the baseline;

[0101] Step 2-2: Select the Db4 wavelet basis function to perform a 6-layer wavelet packet decomposition on the impulse response signal after removing the DC component, and obtain the decomposition coefficients of 64 nodes. Then reconstruct the wavelet packet decomposition coefficients to obtain the reconstructed signal sequence X of the i-th node under the 6-layer wavelet packet decomposition. 6,i ={x 6,i (k), k=1,2,...,64}

[0102] Step 2-3: The wavelet packet Tsallis entropy calculation formula of the reconstructed signal of this node is as follows:

[0103]

[0104] Wherein: q is a non-extensive parameter, and q=2 is selected in the present invention.

[0105] Step 2-4: Calculate the wavelet packet Tsallis entropy of each node in the 6-layer wavelet packet decomposition of the baseline-removed shock response signal of each EFPI sensor after the impact load, and merge them into the wavelet packet Tsallis entropy spectrum feature column vector:

[0106] Tsae={Tsa1;Tsa2;Tsa3;Tsa4}

[0107] Where the column vector Tsa1={Tsa1 6,1 ;Tsa1 6,2 ;...;Tsa1 6,64}、Tsa2={Tsa2 6,1 ;Tsa2 6,2 ;...;Tsa2 6,64}、Tsa3={Tsa3 6,1 ;Tsa36,2 ;...;Tsa3 6,64}、Tsa4={Tsa4 6,1 ;Tsa4 6,2 ;...;Tsa4 6,64} respectively represent the wavelet packet Tsallis entropy of all nodes in the six-layer wavelet packet decomposition of the baseline-removed impact response signals of the first EFPI sensor 7, the second EFPI sensor 8, the third EFPI sensor 9, and the fourth EFPI sensor 10 after the impact load.

[0108] Step 3: Use the LLE method to perform feature dimension reduction on the Tsallis entropy spectrum of the collected impulse response signal wavelet packet;

[0109] Step 3-1: Summarize the wavelet packet Tsallis entropy spectra of the response signals of all 289*3 867 sample impact points and 50 unknown impact points to be located, and obtain a wavelet packet Tsallis entropy feature matrix T with 256 rows and 917 columns, where Tsae a,b It represents the wavelet packet Tsallis entropy spectrum characteristic column vector of the bth sample impact load applied to the ath grid node, Tsae 0,s It represents the wavelet packet Tsallis entropy spectrum characteristic column vector of the sth unknown impact load to be located. The Tsallis entropy characteristic matrix T is expressed as follows:

[0110] T=[Tsae 1，1 ,Tsae 1，2 ,Tsae 1，3 ,Tsae 2，1 ,Tsae 2，2 ,Tsae 2，3 ,...,Tsae 289，1 ,Tsae 289，2 ,Tsae 289，3 ,Tsae 0，1 ,Tsae 0，2 ,...,Tsae 0，50 ]

[0111] =[t1,t2,...,t i ,...,t 917 ]

[0112] Step 3-2: Calculate the Euclidean distance between any two impact point wavelet packet Tsallis entropy spectrum eigenvectors, and obtain the wavelet packet Tsallis entropy spectrum eigenvector t from the impact point i The eigenvectors of the wavelet packet Tsallis entropy spectrum of the 10 nearest impact points are used as the neighborhood;

[0113] Step 3-3: Calculate the reconstruction weight matrix W by minimizing the reconstruction error function Get the reconstruction matrix W:

[0114]

[0115] Where: w ij t i With t j The weight coefficients between If t j Not t i The neighboring point of ij is 0; therefore

[0116]

[0117]

[0118] Where Z i =(t i -t j )(t i -t j ) T , l K is a 3-dimensional vector of all 1s. The weight coefficient can be obtained by the above two formulas;

[0119]

[0120] Step 3-4: By minimizing the loss function Get the low-dimensional embedding result Y. The loss function and constraints are as follows:

[0121]

[0122] Where: I is the unit matrix, M = (IW) T (IW), y i ,y j x i ,x j Image in low-dimensional space.

[0123] Step 3-5: Sort the eigenvalues ​​of the matrix M from small to large, and select the eigenvectors corresponding to the 10 smallest non-zero eigenvalues ​​of the matrix M to form the resulting matrix U after the dimensionality reduction of the initial eigenvector matrix T. U is a matrix with 10 rows and 917 columns.

[0124] Step 4: Use the GA-CNN-BiLSTM-Attention neural network model to identify the impact load location on the composite structure;

[0125] Step 4-1: Transpose the result matrix U extracted in step 3-5, and take the first 867 rows of the sample impact point response signal wavelet packet Tsallis entropy spectrum and the corresponding impact point coordinate position after dimensionality reduction as the sample data set;

[0126] Step 4-2: Input the sample dataset data into the GA-CNN-BiLSTM-Attention neural network model. In this invention, the CNN model consists of a Conv layer, a Batch Norm layer, a ReLU layer, and a Max-pooling layer. The number and size of the convolution kernels in the Conv layer are optimized using a genetic algorithm to obtain the optimal parameters. The BiLSTM layer is composed of three stacked layers, and the parameters of the three layers are exactly the same.

[0127] Step 4-3: Input the optimal parameters obtained by genetic algorithm optimization and the wavelet packet Tsallis entropy spectrum characteristics of the response signal of the unknown impact point to be located after LLE dimensionality reduction into the CNN-BiLSTM-Attention neural network model to realize the impact load positioning on the composite structure.

[0128] The specific process steps of the genetic algorithm optimization CNN-BiLSTM-Attention neural network model in step 4-2 are as follows:

[0129] (1) Initialization parameters are set for the genetic algorithm-optimized CNN-BiLSTM-Attention neural network, where the number of chromosomes is in the range of 10-30, the crossover rate is in the range of 0.8-0.95, the evolutionary generations are in the range of 20-30, and the mutation rate is in the range of 0.001-0.1. The present invention preferably sets the number of chromosomes to 25, the number of genes to 10, the evolutionary generations to 30, the crossover rate to 0.92, and the mutation rate to 0.02. The population of the genetic algorithm is initialized, and 817 rows of sample impact point response signal wavelet packet Tsallis entropy spectra and their corresponding impact point coordinate positions after dimensionality reduction are randomly selected from the sample data set as the training set. The remaining 50 rows of sample impact point response signal wavelet packet Tsallis entropy spectra and their corresponding impact point coordinate positions after dimensionality reduction in the sample data set are used as the test set. The average positioning error of the test set is used as the fitness function of the genetic algorithm. The average positioning error formula is as follows:

[0130]

[0131] in Represents the average distance error between the predicted value and the true value, x j Indicates the actual horizontal coordinate value, x' j Indicates the predicted horizontal coordinate value, y j Indicates the actual horizontal coordinate value, y' jRepresents the predicted horizontal axis value, and n represents the number of test sets.

[0132] (2) Selection operation: Introduce an elite selection strategy to retain the best individuals and directly enter the next generation to replace the individuals with poor fitness;

[0133] (3) Crossover operation: Using the real number crossover method, the gene values ​​of the parent individuals are linearly combined to generate new offspring individuals, thereby retaining the excellent characteristics;

[0134] (4) Mutation operation: using the exchange mutation method, two gene positions in an individual are randomly selected and their values ​​are exchanged to generate a new individual;

[0135] (5) Repeat operations (2) to (4) until the preset number of iterations, 30 generations, is reached, and the best individual is selected as the number and size parameters of the convolution kernel in the CNN-BiLSTM-Attention neural network model;

[0136] The specific processing steps of the GA-CNN-BiLSTM-Attention neural network model in step 4-3 to locate the impact load on the composite structure are as follows:

[0137] (1) Set the input and output dimensions of the GA-CNN-BiLSTM-Attention neural network;

[0138] (2) The wavelet packet Tsallis entropy spectrum features of the response signal of the unknown impact point to be located obtained after LLE dimensionality reduction are input into the sequence folding layer of the trained CNN-BiLSTM-Attention neural network. The sequence folding layer receives data and folds it into a small batch format to improve computational efficiency and enable the model to more effectively learn the temporal features within the sequence;

[0139] (3) Specific process steps of the convolutional layer neural network module:

[0140] ① Convolution layer: The convolution layer compresses the original features of the data and extracts local features. The genetic algorithm optimizes the number of convolution kernels to 8, the size of the convolution kernel to 3×3, and the initial step size of the convolution kernel to 1.

[0141] ②Batch normalization layer, which normalizes the output of the convolutional layer to accelerate the convergence of the model;

[0142] ③ReLU layer, which uses the ReLU activation function to perform nonlinear mapping on the output of the convolutional layer, enabling the neural network to learn nonlinear relationships;

[0143] ④The maximum pooling layer performs dimensionality reduction, reduces the number of parameters, and extracts the main features.

[0144] (4) The sequence refolding layer unfolds the folded sequence data that has been processed by the convolutional layer, thereby ensuring that the subsequent BiLSTM layer can correctly receive the data and effectively capture the long-term dependencies in the time series;

[0145] (5) The expanded sequence data is passed to the BiLSTM layer for bidirectional time series feature learning and capturing context information;

[0146] (6) The Attention layer automatically weights time information, highlights important information, and improves data processing efficiency;

[0147] (7) Finally, the impact point position coordinate prediction results are output through the Dropout layer and the fully connected layer.

[0148] The present invention analyzes the wavelet packet Tsallis entropy spectrum of the impact response signal monitored by the EFPI sensor on the composite material structure, adopts the LLE technology to extract feature vectors from the wavelet packet Tsallis entropy spectrum, and uses the GA-CNN-BiLSTM-Attention neural network model to accurately identify the impact load position of the composite material structure. It has the characteristics of not requiring a high-speed data acquisition module, small amount of signal feature data, wide applicability, and good generalization ability of the identification model.

[0149] Under the premise that no conflict occurs, those skilled in the art may freely combine and superimpose the above-mentioned additional technical features.

[0150] The above descriptions are only preferred embodiments of the present invention. Any technical solution that achieves the purpose of the present invention by substantially the same means shall fall within the scope of protection of the present invention.

Claims

1. A composite material structure impact location method based on wavelet packet Tsallis entropy and LLE, characterized by: The method comprises the following steps: Step 1: Use EFPI sensors to monitor the impact response signals on the composite structure, and build a composite impact monitoring system based on EFPI optical fiber sensors to collect the impact point response signals; Step 2: Calculate the wavelet packet Tsallis entropy spectrum of the impact point response signal collected in step 1; Step 3: Use the LLE method to perform feature dimension reduction on the wavelet packet Tsallis entropy spectrum of the impact point response signal in step 2 to achieve feature extraction of the impact point response signal; Step 4: Use the GA-CNN-BiLSTM-Attention neural network model to identify the impact load location on the composite structure; The impact monitoring system in step 1 includes: a broadband light source (1), a 1×4 coupler (2), a first optical circulator (3), a second optical circulator (4), a third optical circulator (5), a fourth optical circulator (6), a first EFPI sensor (7), a second EFPI sensor (8), a third EFPI sensor (9) and a fourth EFPI sensor (10), a fifth optical circulator (11), a sixth optical circulator (12), a seventh optical circulator (13), an eighth optical circulator (14), a first FBG sensor (15), a second FBG sensor (16), a third FBG sensor (17), a fourth FBG sensor (18), a first photodetector (19), a second photodetector (20), a third photodetector (21) and a fourth photodetector (22), a data acquisition module (23) and a composite material structural member (24); Light output by the broadband light source (1) is divided into four paths of light through a 1×4 coupler (2), the first path of light passes through a first optical circulator (3) and reaches a first EFPI sensor (7), the light reflected from the first EFPI sensor (7) passes through a fifth optical circulator (11) and reaches a first FBG sensor (15), and the reflected light of the first FBG sensor (15) is received by a first photodetector (19); The second path of light passes through the second optical circulator (4) and reaches the second EFPI sensor (8). The light reflected from the second EFPI sensor (8) passes through the sixth optical circulator (12) and reaches the second FBG sensor (16). The reflected light from the second FBG sensor (16) is received by the second photodetector (20). The third light path passes through the third optical circulator (5) and reaches the third EFPI sensor (9). The light reflected from the third EFPI sensor (9) passes through the seventh optical circulator (13) and reaches the third FBG sensor (17). The reflected light from the third FBG sensor (17) is received by the third photodetector (21). The fourth path of light passes through the fourth optical circulator (6) to reach the fourth EFPI sensor (10), and the light reflected from the fourth EFPI sensor (10) passes through the eighth optical circulator (14) to reach the fourth FBG sensor (18), and the reflected light from the fourth FBG sensor (18) is received by the fourth photodetector (22); then, the data acquisition module (23) is used to acquire impulse response signals outputted by the first photodetector (19), the second photodetector (20), the third photodetector (21), and the fourth photodetector (22); The specific steps for arranging the sensors of the impact monitoring system and obtaining the impact point response signal in step 1 are as follows: Step 1-1: Fix the composite structural member (24) on the workbench, construct a square impact monitoring area on its surface, and divide it into uniform grids to obtain grid nodes; Step 1-2: Select the lower left corner of the square area as the coordinate origin (0mm, 0mm), horizontally to the right as the positive direction of the X axis, and vertically upward as the positive direction of the Y axis to establish a rectangular coordinate system; Step 1-3: Using epoxy resin adhesive, affix the first EFPI sensor (7), the second EFPI sensor (8), the third EFPI sensor (9), and the fourth EFPI sensor (10) to the four vertices of the square monitoring area on the surface of the composite structural member (24), and connect the EFPI sensing optical path through optical fiber jumpers to build an impact monitoring system; Step 1-4: Use a fixed energy impact hammer to apply a fixed energy impact to each grid node in the monitoring area of ​​the surface of the composite structural component (24). Apply impact loads of the same energy level for several times and record the original response signals of each EFPI sensor under each impact load.

2. The composite material structure impact location method based on wavelet packet Tsallis entropy and LLE according to claim 1, characterized in that: The specific calculation method of the wavelet packet Tsallis entropy spectrum of the impact point response signal in step 2 is: Step 2-1: remove the DC component of the collected original impulse response signal to obtain the time series of the impulse response signal after removing the baseline; Step 2-2: Select the wavelet basis function to perform the impulse response signal after removing the DC component. Layer wavelet packet decomposition, we get The decomposition coefficients of the nodes are then reconstructed to obtain the wavelet packet decomposition coefficients. Layer wavelet packet decomposition Reconstructed signal sequence of a node , where the length of the reconstructed signal sequence is ; Step 2-3: The wavelet packet Tsallis entropy calculation formula of the reconstructed signal of this node is as follows: ； in is a non-extensive parameter; Step 2-4: Calculate the baseline shock response signal of each EFPI sensor after the shock load The wavelet packet Tsallis entropy of each node under the layer wavelet packet decomposition is merged into the wavelet packet Tsallis entropy spectrum feature column vector: ； where the column vector Respectively represent the baseline-free shock response signals of the first EFPI sensor (7), the second EFPI sensor (8), the third EFPI sensor (9) and the fourth EFPI sensor (10) after the shock load The wavelet packet Tsallis entropy of all nodes under the layer wavelet packet decomposition.

3. The composite material structure impact location method based on wavelet packet Tsallis entropy and LLE according to claim 2, characterized in that: The specific steps of feature dimensionality reduction in step 3 are as follows: Step 3-1: For all collected Sample impact points and The wavelet packet Tsallis entropy spectrum of the unknown impact point response signal to be located is summarized to obtain a OK The wavelet packet Tsallis entropy characteristic matrix of the column : ； in Indicates the Apply the first The wavelet packet Tsallis entropy spectrum characteristic column vector of the sub-sample impact load, Indicates the The wavelet packet Tsallis entropy spectrum characteristic column vector of the unknown impact load to be located; Step 3-2: Calculate the Euclidean distance between any two impact point wavelet packet Tsallis entropy spectrum feature vectors to obtain the wavelet packet Tsallis entropy spectrum feature vector from the impact point The nearest The Tsallis entropy spectrum eigenvector of the impact point wavelet packet is used as the neighborhood; Step 3-3: Calculate the reconstruction weight matrix , by minimizing the reconstruction error function Get the reconstruction matrix : ； Where: for and The weight coefficients between ,like no The neighboring points of is 0; So we have: ； ； In the formula , for dimensional all-1 vector; the weight coefficient can be obtained by the above two formulas; ； Step 3-4: By minimizing the loss function Get low-dimensional embedding results , the loss function and constraints are as follows: ； Where: is the identity matrix, , They are Image in low-dimensional space; Step 3-5: Matrix Sort the eigenvalues ​​of from small to large and select the matrix of The eigenvectors corresponding to the smallest non-zero eigenvalues ​​form the initial eigenvector matrix The resulting matrix after dimensionality reduction , for OK Column matrix.

4. The composite material structure impact location method based on wavelet packet Tsallis entropy and LLE according to claim 3, characterized in that: The specific steps for using the GA-CNN-BiLSTM-Attention neural network model to identify the impact load position on the composite structure in step 4 are as follows: Step 4-1: Extract the result matrix obtained in steps 3-5 Transpose the front The wavelet packet Tsallis entropy spectrum of the sample impact point response signal after row dimension reduction and its corresponding impact point coordinate position are used as the sample data set; Step 4-2: Input the sample dataset data into the GA-CNN-BiLSTM-Attention neural network model and use the genetic algorithm to optimize the number and size of convolution kernels in the CNN-BiLSTM-Attention model; Step 4-3: Input the optimal parameters obtained by genetic algorithm optimization and the wavelet packet Tsallis entropy spectrum characteristics of the response signal of the unknown impact point to be located after LLE dimensionality reduction into the CNN-BiLSTM-Attention neural network model to realize the location of the impact load on the composite structure.

5. The composite material structure impact location method based on wavelet packet Tsallis entropy and LLE according to claim 4, characterized in that: The specific process steps of the genetic algorithm optimization of the CNN-BiLSTM-Attention neural network model in step 4-2 are as follows: (1) Initialize the parameters of the genetic algorithm-optimized CNN-BiLSTM-Attention neural network, randomly divide the sample data set into a training set and a test set, and use the average positioning error of the test set as the fitness function of the genetic algorithm. The formula for the average positioning error is as follows: ； in It represents the average distance error between the predicted value and the true value. Indicates the actual horizontal coordinate value, Represents the predicted horizontal coordinate value, Indicates the actual vertical coordinate value, Represents the predicted vertical coordinate value, Indicates the number of test sets; (2) Selection operation: Introduce an elite selection strategy to retain the best individuals and directly enter the next generation to replace individuals with poor fitness; (3) Crossover operation: Using the real number crossover method, the gene values ​​of the parent individuals are linearly combined to generate new offspring individuals, thereby retaining the excellent characteristics; (4) Mutation operation: Using the exchange mutation method, two gene positions in an individual are randomly selected and their values ​​are exchanged to generate a new individual; (5) Repeat steps (2) to (4) until the preset number of iterations is reached, and select the optimal individual parameter values ​​as the parameters for the number and size of convolution kernels in the CNN-BiLSTM-Attention neural network model.

6. The composite material structure impact location method based on wavelet packet Tsallis entropy and LLE according to claim 5, characterized in that: The specific process steps of the GA-CNN-BiLSTM-Attention neural network model in step 4-3 to locate the impact load on the composite structure are as follows: (1) Set the input and output dimensions of the GA-CNN-BiLSTM-Attention neural network; (2) The wavelet packet Tsallis entropy spectrum feature of the response signal of the unknown impact point to be located obtained after LLE dimensionality reduction is input into the sequence folding layer of the trained CNN-BiLSTM-Attention neural network; the sequence folding layer receives the data and folds it into a small batch format to improve computational efficiency and enable the model to more effectively learn the temporal features within the sequence; (3) The convolution layer compresses the input data and extracts features through sliding window operations; (4) The sequence refolding layer is responsible for unfolding the folded sequence data processed by the convolution layer, thereby ensuring that the subsequent BiLSTM layer can correctly receive the data and effectively capture the long-term dependencies in the time series; (5) The expanded sequence data is passed to the BiLSTM layer for bidirectional time series feature learning and capturing context information; (6) The Attention layer automatically weights time information, highlights important information, and improves data processing efficiency; (7) Finally, the impact point position coordinate prediction results are output through the Dropout layer and the fully connected layer.

Images