Method for recognizing impact position of net structure based on bionic principle

Abstract

The application provides a mesh structure impact position recognition method based on a bionic principle, relates to the field of structure health monitoring, and specifically comprises the following steps: pasting a piezoelectric sensor at the center of a mesh antenna structure to form a sensor array, and acquiring an impact response signal; labeling the impact point distance and angle of the collected impact response signal, and establishing an "impact position-response signal" database; constructing a multi-kernel support vector machine regression MK-SVR model; optimizing the MK-SVR model parameters by using a particle swarm optimization algorithm; and predicting the impact source position by using the optimized MK-SVR model obtained in step S4. The technical scheme of the application overcomes the problem in the prior art that the impact position of the mesh antenna cannot be accurately recognized.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring, and specifically to a method for identifying the impact location of a mesh structure based on biomimetic principles. Background Technology

[0002] Within the keratinized tissue of a spider's limbs lies an elastic sensory organ. Each organ contains numerous tiny gaps, all surrounded by a thin membrane connecting the terminals of sensory cells. This microscopic tissue is extremely sensitive, capable of detecting even the slightest stimuli. It deforms, causing the membrane to move and rapidly transmit information to the sensory cells. The spider then promptly determines the frequency, intensity, and direction of the web's vibrations, thereby pinpointing the location and size of its prey.

[0003] The fully deployed mesh antenna of a satellite resembles a spider web in its external structure and is widely used in deep space exploration, remote sensing, and navigation. For example, the BeiDou-3 geostationary orbit satellite is equipped with two frame-type deployable mesh antennas. With the increasing amount of space debris, satellite antennas are inevitably subject to impacts from space debris during on-orbit operation, posing certain safety hazards. To ensure the reliability of satellite operation in orbit, real-time monitoring and location of potential impact events on the mesh antenna structure are essential. However, existing impact monitoring technologies are all designed for continuous structures, and there is a lack of available methods for impact identification of mesh antenna structures. Therefore, the biomimetic principle of spider hunting can be used to identify the location of impacts on the mesh antenna.

[0004] Therefore, there is a need for a biomimetic mesh structure impact location identification method that can accurately identify the impact location of a mesh antenna. Summary of the Invention

[0005] The main objective of this invention is to provide a biomimetic method for identifying the impact location of a mesh structure, thereby solving the problem that existing technologies cannot accurately identify the impact location of mesh antennas.

[0006] To achieve the above objectives, this invention provides a biomimetic method for identifying the impact location of a mesh structure, specifically comprising the following steps:

[0007] S1. A piezoelectric sensor is attached to the center of the mesh antenna structure to form a sensor array and obtain the impact response signal.

[0008] S2, mark the distance and angle of the impact point on the collected impact response signal, and establish a "impact location-response signal" database.

[0009] S3, construct a multi-core support vector machine regression MK-SVR model.

[0010] S4. The particle swarm optimization algorithm is used to optimize the parameters of the MK-SVR model.

[0011] S5. Using the optimized MK-SVR model obtained in step S4, the location of the impact source is predicted.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S1.1, the sensor is attached to the innermost layer of the mesh antenna structure to form a piezoelectric sensor network, and the rectangular sensors are arranged symmetrically and covered by the mesh antenna structure.

[0014] S1.2 Connect the sensor to the multi-channel data acquisition system and fix the mesh antenna structure.

[0015] S1.3, use a small ball to fall freely from a predetermined height and impact different positions of the model; after each impact experiment, record the signal output by the sensor; repeat each experiment multiple times and take the average value as the final result.

[0016] S1.4 performs preliminary preprocessing on the data, including noise reduction and normalization.

[0017] Furthermore, step S3 specifically includes:

[0018] Based on combined kernel function Building the MK-SVR model :

[0019] ;

[0020] ;

[0021] in, For Lagrange multipliers; For the bias term of the MK-SVR model, These are weight parameters; Take 1 and 2, ; ; For coefficient scaling factor, The bias constant is Let the order be the order of the polynomial. , This represents any two sets of input features, i.e., input time-domain features. The total number of training samples, This is the new input sample vector to be predicted.

[0022] Furthermore, step S4 specifically includes the following steps:

[0023] S4.1, Define the joint vector to be optimized. Including: penalty factors Insensitive region parameters Coefficient scaling factor nuclear parameters and the weights of each sub-kernel function :

[0024] ;

[0025] in, For the first The weights of each sub-kernel function.

[0026] S4.2, using the Particle Swarm Optimization (PSO) algorithm, obtains the optimized parameter vector. , Includes: Optimal penalty factor Optimal insensitive region parameters Optimal coefficient scaling factor Optimal kernel parameters and the weights of the optimal sub-kernel functions .

[0027] S4.3, with Construct the optimized MK-SVR model.

[0028] Furthermore, step S4.2 specifically includes the following steps:

[0029] S4.2.1, Set the parameter to be optimized: penalty factor Insensitive region parameters nuclear parameters and the weights of each sub-kernel function The upper and lower limits of the search.

[0030] S4.2.2, each set of parameters to be optimized is considered as a position vector of a particle in the search space. Within the search space, randomly generate the initial positions and velocities of S particles. Initialize the individual historical best position of each particle. Set the global optimal position for the initial position of the particle. This is the position with the highest fitness among all current particles.

[0031] Furthermore, step S4.2 also includes the following steps:

[0032] S4.2.3, for each combination of position parameters of a particle, construct the corresponding MK-SVR model; fit the MK-SVR model on the training set, calculate the positioning error evaluation index on the validation set, and set the root mean square error (RMSE). ) is the evaluation indicator:

[0033] );

[0034] in, The number of support vectors; These are the actual impact location coordinates. To predict the coordinates of the impact location.

[0035] The optimized parameter vector minimizes the root mean square error function, i.e.:

[0036] ;

[0037] in, The value of the independent variable is the minimum value.

[0038] S4.2.4 If the current particle's fitness is better than its individual best value, then update the individual's historical best position. If the current particle's fitness is better than the global optimum, then update the global optimum position. .

[0039] S4.2.5, Calculate the velocity of the next generation of particles using the velocity update formula:

[0040] ;

[0041] in, Inertial weight; For the first The velocity of the particles These are individual learning factors and social learning factors, respectively. A random number between [0, 1]; For the first The individual's historical optimal position. for The Each dimension component; This represents the parameter dimension index, and the particle position is adjusted using the position update formula:

[0042] ;

[0043] in, For the first The position of the substitute particle.

[0044] If the particle's position is outside the parameter range, its position is clipped to the boundary value, and its velocity is reversed.

[0045] Furthermore, step S4.2 also includes the following steps:

[0046] S4.2.6, determine whether any of the following termination conditions are met: the maximum number of iterations has been reached. The improvement in global optimal fitness is below the threshold. Continue for k generations; verify that the error reaches the preset target.

[0047] If the termination condition is not met, return to S4.2.3; otherwise, proceed to S4.2.7.

[0048] S4.2.7, determine the globally optimal position corresponding As the final parameter configuration for the MK-SVR model, the MK-SVR model is retrained and validated on both the training and validation sets.

[0049] Further, step S4.3 includes the following steps:

[0050] S4.3.1, with Construct a combined kernel function with optimized parameters :

[0051] ;

[0052] in, Indicates the first Sub-kernel functions, Values ​​1 and 2; ; ; The bias constant is Let the order be the order of the polynomial. This is the parameter vector after optimization for the corresponding sub-kernel.

[0053] S4.3.2, the optimization objective of the MK-SVR model is:

[0054] ;

[0055] in, The constraints are: ; This represents the total number of training samples.

[0056] S4.3.3, in the validation set and test set, for Predict output polar coordinates And convert polar coordinates to rectangular coordinates ( , ), calculate the two-dimensional Euclidean error :

[0057] .

[0058] And calculate the root mean square error. With maximum positioning error :

[0059] ;

[0060] ;

[0061] If the test set Less than the preset threshold and If the engineering requirements are met, the training is considered complete; otherwise, return to step S4.2.1 for re-optimization.

[0062] Further, step S5 includes the following steps:

[0063] S5.1, validate set data Input the MK-SVR model obtained in step S4:

[0064] ;

[0065] in, Indicates The constructed MK-SVR model outputs the predicted position coordinates of the impact source in the mesh antenna structure coordinate system. .

[0066] S5.2 Calculate the root mean square error (RMSE) and confidence score; if the predicted location is outside the set safety boundary, it indicates an error has been triggered.

[0067] The present invention has the following beneficial effects:

[0068] This invention utilizes machine learning methods for impact location identification, without needing to consider information such as the excitation duration and peak value of the impact response signal. It can learn the deep internal features by mapping the data in a high dimension. Attached Figure Description

[0069] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0070] Figure 1 The diagram shows the mesh antenna structure and sensor layout of the present invention.

[0071] Figure 2 A schematic diagram of a typical impulse response signal for channel 1 is shown.

[0072] Figure 3 A schematic diagram of a typical impulse response signal for channel 2 is shown.

[0073] Figure 4 A schematic diagram of a typical impulse response signal for channel 3 is shown.

[0074] Figure 5 A schematic diagram of a typical impulse response signal for channel 4 is shown.

[0075] Figure 6 A schematic diagram of a typical impulse response signal for channel 5 is shown.

[0076] Figure 7 A schematic diagram of a typical impulse response signal for channel 6 is shown.

[0077] Figure 8 A physical diagram of the mesh structure and piezoelectric sensor array is shown.

[0078] Figure 9 It shows Figure 8 A magnified view of the details at point A.

[0079] Figure 10 The error distribution diagram for the predicted angle and distance is shown.

[0080] Figure 11 A comparison diagram showing the actual impact location and the predicted impact location is presented. Detailed Implementation

[0081] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] like Figure 1 The method for identifying the impact location of a mesh structure based on biomimetic principles, as shown, specifically includes the following steps:

[0083] S1. A piezoelectric sensor is attached to the center of the mesh antenna structure to form a sensor array, which acquires the impact response signal. Impacts are sequentially applied to different locations on the mesh antenna structure, and the impact response signals received by the sensors are collected.

[0084] S2, mark the distance and angle of the impact point on the collected impact response signal, and establish a "impact location-response signal" database.

[0085] like Figures 2-7This is a schematic diagram of a typical impact response signal. As can be seen, when the impact response reaches the sensor, the signal waveform experiences a sharp spike. A threshold much larger than the noise signal but smaller than the first peak of the signal is set. The program then finds the first time point in each sensor signal that reaches this threshold; this time point is the arrival time of the impact response signal, i.e., the time-domain characteristic.

[0086] By setting a loop in the signal processing program, signal data from multiple sensors can be processed simultaneously, and the temporal characteristics of the signals at different locations on the impact mesh structure can be more clearly reflected. The collected impact response signals are labeled with the distance and angle of the impact point, establishing an "impact location-response signal" database.

[0087] S3. Construct a multi-kernel support vector machine regression (MK-SVR) model. The MK-SVR algorithm is used to establish a nonlinear mapping relationship between input features and the output target, enabling accurate prediction of the impact source location. By integrating the advantages of multiple kernel functions, the multi-kernel support vector regression model effectively improves the model's learning ability and generalization performance, making it suitable for handling the mapping problem of nonlinear impact signals in complex structures.

[0088] S4. The particle swarm optimization algorithm is used to optimize the parameters of the MK-SVR model.

[0089] S5. Using the optimized MK-SVR model obtained in step S4, the location of the impact source is predicted.

[0090] Inspired by spiders sensing prey through their webs, a dense sensor array is formed by attaching piezoelectric sensors to the center of the web structure to sense structural vibration signals caused by impacts. Combined with machine learning methods, this enables accurate identification of the impact location.

[0091] Specifically, step S1 includes the following steps:

[0092] S1.1, the sensor is attached to the innermost layer of the mesh antenna structure to form a piezoelectric sensor network, such as... Figure 1 As shown, rectangular piezoelectric sensors (PZTs) are symmetrically arranged and covered with a mesh antenna structure to ensure comprehensive capture of vibration signals generated by impacts. The geometric dimensions of the mesh antenna structure can be determined according to actual needs. The sensor in this invention uses a rectangular piezoelectric element, which has excellent directionality. Because the sensor is prone to inconsistencies in response due to factors such as manufacturing processes and bonding methods, calibration or compensation of the sensor is necessary before testing to obtain reliable monitoring results.

[0093] S1.2 Connect the sensor to the multi-channel data acquisition system and fix the mesh antenna structure.

[0094] S1.3, a small ball of a specific mass is dropped freely from a predetermined height to impact different positions of the model; after each impact experiment, the signal output by the sensor is recorded; to ensure the reliability of the experimental data, each group of experiments is repeated multiple times, and the average value is taken as the final result.

[0095] S1.4 Perform preliminary preprocessing on the data, including noise reduction and normalization, to improve the accuracy of subsequent analysis.

[0096] Specifically, step S3 is as follows:

[0097] Based on combined kernel function Building the MK-SVR model :

[0098] ;

[0099] ;

[0100] in, For Lagrange multipliers; For the bias term of the MK-SVR model, These are weight parameters; Take 1 and 2, ; ; For coefficient scaling factor, The bias constant is Let be the order of the polynomial. The larger the value, the higher the mapping dimension, and the stronger the learning ability, based on empirical values. Take 2; , This represents any two sets of input features, i.e., input time-domain features. The total number of training samples, This is the new input sample vector to be predicted.

[0101] Kernel functions are mainly divided into local kernel functions and global kernel functions. Local kernel functions are characterized by strong learning ability but weak generalization performance. Global kernel functions, on the other hand, are characterized by strong generalization performance but weak learning ability. Therefore, combining support vector machines based on these two types of kernel functions into a new prediction model can leverage their strengths and mitigate their weaknesses, resulting in a prediction model with strong learning ability and good generalization performance. Linear kernel functions are characterized by few parameters, simple structure, and fast running speed; in this invention, they are used to construct the model by combining local kernel functions. To integrate the modeling capabilities of different kernel functions in different feature spaces, this invention performs a weighted combination of multiple selected kernel functions to form a multi-kernel function. .

[0102] To achieve two-dimensional polar coordinates of the impact source Independent location prediction was achieved by establishing two independent regression channels, corresponding to the radius respectively. and angle The coordinate prediction ensures that the final output is the two-dimensional polar coordinates of the impact source location. : The regression channel uses input features as independent variables. The coordinates are the dependent variables; The regression channel also uses input features as independent variables. The coordinates are the dependent variables.

[0103] The experimental data in the database were divided into training, testing, and validation sets according to a predetermined ratio. Each data point contains a feature vector (representing the arrival time of each sensor) and a target vector (the true coordinates of the impact source). The training set is used for support vector calculation and parameter fitting of the model, the validation set is used to monitor model performance during training to prevent overfitting, and the test set is only used for final performance evaluation to ensure the objectivity of the results. The multi-core SVR model is trained using the training set data, employing standardized input and cross-validation to improve training accuracy. After training, the results from multiple models obtained through cross-validation are arithmetically averaged and saved along with the kernel function weight parameters to support real-time prediction in subsequent impact positioning systems.

[0104] Specifically, this invention introduces a parameter optimization mechanism. By uniformly optimizing the kernel parameters, kernel weights, and hyperparameters of the multi-kernel support vector regression model, the impact positioning accuracy and model generalization ability are further improved, enhancing the robustness of the model in practical applications. Step S4 specifically includes the following steps:

[0105] S4.1, Define the joint vector to be optimized. Including: penalty factors Insensitive region parameters Coefficient scaling factor nuclear parameters and the weights of each sub-kernel function :

[0106] ;

[0107] in, For the first The weights of each sub-kernel function.

[0108] S4.2 addresses the high-dimensionality and non-convex nature of the parameter combination space by employing the Particle Swarm Optimization (PSO) algorithm. This method can be used to search for equal hyperparameter combinations of multi-kernel support vector machines (MK-SVR), using the PSO algorithm to obtain the optimized parameter vector. , Includes: Optimal penalty factor Optimal insensitive region parameters Optimal coefficient scaling factor Optimal kernel parameters and the weights of the optimal sub-kernel functions .

[0109] S4.3, with Construct the optimized MK-SVR model.

[0110] Specifically, step S4.2 includes the following steps:

[0111] S4.2.1, Set the parameter to be optimized: penalty factor Insensitive region parameters nuclear parameters and the weights of each sub-kernel function The upper and lower bounds of the search are set to ensure that the parameter range covers the possible optimal solutions.

[0112] S4.2.2, each set of parameters to be optimized is considered as a position vector of a particle in the search space. Within the search space, randomly generate the initial positions and velocities of S particles. To ensure uniform distribution, initialize the individual historical best position for each particle. Set the global optimal position for the initial position of the particle. This is the position with the highest fitness among all current particles.

[0113] Specifically, step S4.2 also includes the following steps:

[0114] S4.2.3, for each combination of position parameters of a particle, construct the corresponding MK-SVR model; fit the MK-SVR model on the training set, and calculate the positioning error evaluation index on the validation set. This invention sets the root mean square error (RMSE) to quantitatively evaluate the performance of parameter combinations. ) is the evaluation indicator:

[0115] );

[0116] in, This refers to the number of support vectors, which is the number of samples that are effective in the final model. These are the actual impact location coordinates. To predict the coordinates of the impact location.

[0117] The optimized parameter vector minimizes the root mean square error function, i.e.:

[0118] ;

[0119] in, The value of the independent variable is the minimum value.

[0120] The reciprocal of RMSE is used as the fitness value of the particle; the larger the fitness value, the better the parameter combination.

[0121] S4.2.4 If the current particle's fitness is better than its individual best value, then update the individual's historical best position. If the current particle's fitness is better than the global optimum, then update the global optimum position. ;

[0122] S4.2.5, Calculate the velocity of the next generation of particles using the velocity update formula:

[0123] ;

[0124] in, The inertial weights are used to control the search range and convergence speed. For the first The velocity of the particles These are individual learning factors and social learning factors, respectively. A random number between [0, 1]; For the first The individual's historical optimal position. for The Each dimension component; This represents the parameter dimension index, and the particle position is adjusted using the position update formula:

[0125] ;

[0126] in, For the first The position of the substitute particle.

[0127] If the particle's position is outside the parameter range, its position is clipped to the boundary value, and its velocity is reversed.

[0128] Specifically, step S4.2 also includes the following steps:

[0129] S4.2.6, determine whether any of the following termination conditions are met: the maximum number of iterations has been reached. The improvement in global optimal fitness is below the threshold. Continue for k generations; verify that the error reaches the preset target.

[0130] If the termination condition is not met, return to S4.2.3; otherwise, proceed to S4.2.7.

[0131] S4.2.7, determine the globally optimal position corresponding As the final parameter configuration for the MK-SVR model, the MK-SVR model was retrained and validated on the training and validation sets to confirm the final performance.

[0132] Specifically, step S4.3 includes the following steps:

[0133] S4.3.1, with Construct a combined kernel function with optimized parameters :

[0134] ;

[0135] in, Indicates the first Sub-kernel functions, Values ​​1 and 2; ; ; The bias constant is Let the order be the order of the polynomial. This is the parameter vector after optimization for the corresponding sub-kernel.

[0136] S4.3.2, In the training set, solve for the Lagrange multipliers of the MK-SVR model using a quadratic programming algorithm. and bias terms This yields the set of support vectors. The optimization objective of the MK-SVR model is:

[0137] ;

[0138] in, The constraints are: ; This represents the total number of training samples.

[0139] S4.3.3, in the validation set and test set, for Predict output polar coordinates And convert polar coordinates to rectangular coordinates ( , ), calculate the two-dimensional Euclidean error :

[0140] ;

[0141] And calculate the root mean square error. With maximum positioning error :

[0142] ;

[0143] ;

[0144] If the test set Less than the preset threshold and If the engineering requirements are met, the training is considered complete; otherwise, return to step S4.2.1 for re-optimization.

[0145] Specifically, the optimal multi-kernel support vector machine regression (MK-SVR) model, trained and validated in step S4, is applied to the localization of actual impact events in a mesh structure, achieving a real-time closed loop from data acquisition to localization result output. Step S5 includes the following steps:

[0146] S5.1, validate set data Input the MK-SVR model obtained in step S4:

[0147] ;

[0148] in, Indicates The constructed MK-SVR model outputs the predicted position coordinates of the impact source in the mesh antenna structure coordinate system. .

[0149] S5.2 Calculate the root mean square error (RMSE) and confidence score; if the predicted location is outside the set safety boundary, it indicates an error has been triggered, and event information is provided for subsequent analysis.

[0150] To verify the effectiveness of the method of the present invention, such as Figure 8 and Figure 9 As shown, impact experiments were conducted on an aluminum mesh structure with an outer radius of 0.25 m. Six rectangular piezoelectric sensors were arranged at 60-degree intervals along the innermost ring. Impact experiments were performed at different locations on the mesh structure, and impact response signals were acquired in real time, resulting in a total of 1368 impact response signals.

[0151] The invention was developed by optimizing a multi-kernel support vector machine using parameters. For R and Independent MK-SVR model channels were constructed, using 70% of the database training samples as training data, 15% as validation data, and 15% as test data. The samples were randomly divided into six subsets, each containing 228 data points. Different models were trained on each subset, and cross-validation was used. The arithmetic mean of the results from the six models was taken as the final prediction output. The impact locations predicted by the trained models were compared with the actual impact locations to obtain the angle and distance error distributions, as shown below. Figure 10 As shown in the error distribution map obtained from 1368 sets of data, 96% of the distance error is concentrated within 0.015m; 98.6% of the angle error is concentrated within 5 degrees, indicating ideal prediction results.

[0152] Analysis of the training results shows that the trained model can not only recognize the original samples and use the training samples as input to test the trained model, but also accurately predict the location of the impact.

[0153] Some impact location identification results are as follows Figure 11 As shown in the figure. The results indicate that by using parameter-optimized multi-kernel support vector machines, the relationship between the impact response signal and the impact location can be learned without analyzing signal features, and this method has broad application prospects.

[0154] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying the impact location of a mesh structure based on biomimetic principles, characterized in that, Specifically, the steps include the following: S1. A piezoelectric sensor is attached to the center of the mesh antenna structure to form a sensor array and obtain the impact response signal. S2, mark the distance and angle of the impact point on the collected impact response signals, and establish a "impact location - response signal" database; S3, Construct a multi-core support vector machine regression MK-SVR model; S4. The particle swarm optimization algorithm is used to optimize the parameters of the MK-SVR model. S5. Using the optimized MK-SVR model obtained in step S4, predict the location of the impact source. Step S1 specifically includes the following steps: S1.1, The sensor is attached to the innermost layer of the mesh antenna structure to form a piezoelectric sensor network. Rectangular sensors are arranged symmetrically and covered with the mesh antenna structure. S1.2, Connect the sensor to the multi-channel data acquisition system and fix the mesh antenna structure; S1.3, use a small ball to fall freely from a predetermined height and impact different positions of the model; after each impact experiment, record the signal output by the sensor; repeat each experiment multiple times and take the average value as the final result; S1.4, Perform preliminary preprocessing on the data, including noise reduction and normalization; Step S3 is as follows: Based on combined kernel function Building the MK-SVR model : ； ； in, For Lagrange multipliers; For the bias term of the MK-SVR model, These are weight parameters; Take 1 and 2, ; ; For coefficient scaling factor, The bias constant is Let the order be the order of the polynomial. , This represents any two sets of input features, i.e., input time-domain features. The total number of training samples, The new input sample vector to be predicted; Step S4 specifically includes the following steps: S4.1, Define the joint vector to be optimized. Including: penalty factors Insensitive region parameters Coefficient scaling factor nuclear parameters and the weights of each sub-kernel function : ； in, For the first The weights of each sub-kernel function; S4.2, using the Particle Swarm Optimization (PSO) algorithm, obtains the optimized parameter vector. , Includes: Optimal penalty factor Optimal insensitive region parameters Optimal coefficient scaling factor Optimal kernel parameters and the weights of the optimal sub-kernel functions ; S4.3, with Construct an optimized MK-SVR model; Step S4.2 specifically includes the following steps: S4.2.1, Set the parameter to be optimized: penalty factor Insensitive region parameters nuclear parameters and the weights of each sub-kernel function The upper and lower limits of the search; S4.2.2, each set of parameters to be optimized is considered as a position vector of a particle in the search space. Within the search space, randomly generate the initial positions and velocities of S particles. Initialize the individual historical best position of each particle. Set the global optimal position for the initial position of the particle. This is the position with the highest fitness among all current particles; Step S4.2 also includes the following steps: S4.2.3, for each combination of position parameters of a particle, construct the corresponding MK-SVR model; fit the MK-SVR model on the training set, calculate the positioning error evaluation index on the validation set, and set the root mean square error (RMSE). ) is the evaluation indicator: )； in, The number of support vectors; These are the actual impact location coordinates. To predict the coordinates of the impact location; The optimized parameter vector minimizes the root mean square error function, i.e.: ； in, The value of the independent variable that takes the minimum value; S4.2.4 If the current particle's fitness is better than its individual best value, then update the individual's historical best position. If the current particle's fitness is better than the global optimum, then update the global optimum position. ; S4.2.5, Calculate the velocity of the next generation of particles using the velocity update formula: ； in, Inertial weight; For the first The velocity of the particles These are individual learning factors and social learning factors, respectively. A random number between [0, 1]; For the first The individual's historical optimal position. for The Each dimension component; This represents the parameter dimension index, and the particle position is adjusted using the position update formula: ； in, For the first The position of the substitute particle; If the particle's position is outside the parameter range, its position is clipped to the boundary value, and its velocity is reversed. Step S4.2 also includes the following steps: S4.2.6, determine whether any of the following termination conditions are met: the maximum number of iterations has been reached. The improvement in global optimal fitness is below the threshold. Continue for k generations; verify that the error reaches the preset target. If the termination condition is not met, return to S4.2.3; otherwise, proceed to S4.2.

7. S4.2.7, determine the globally optimal position corresponding As the final parameter configuration for the MK-SVR model, the MK-SVR model is retrained and validated on both the training and validation sets.

2. The method for identifying the impact location of a mesh structure based on biomimetic principles according to claim 1, characterized in that, Step S4.3 includes the following steps: S4.3.1, with Construct a combined kernel function with optimized parameters : ； in, Indicates the first Sub-kernel functions, Values ​​1 and 2; ; ; The bias constant is Let the order be the order of the polynomial. This is the parameter vector optimized for the corresponding sub-kernel; S4.3.2, the optimization objective of the MK-SVR model is: ； in, The constraints are: ; The total number of training samples; S4.3.3, in the validation set and test set, for Predict output polar coordinates And convert polar coordinates to rectangular coordinates ( , ), calculate the two-dimensional Euclidean error : ； And calculate the root mean square error. With maximum positioning error : ； ； If the test set Less than the preset threshold and If the engineering requirements are met, the training is considered complete; otherwise, return to step S4.2.1 for re-optimization.

3. The method for identifying the impact location of a mesh structure based on biomimetic principles according to claim 2, characterized in that, Step S5 includes the following steps: S5.1, validate set data Input the MK-SVR model obtained in step S4: ； in, Indicates The constructed MK-SVR model outputs the predicted position coordinates of the impact source in the mesh antenna structure coordinate system. ; S5.2 Calculate the root mean square error (RMSE) and confidence score; if the predicted location is outside the set safety boundary, it indicates an error has been triggered.

Images