An impact load identification method and system, an electronic device, and a storage medium

Abstract

The application discloses an impact load identification method and system, electronic equipment and a storage medium, and relates to the field of dynamic load identification. The method comprises the following steps: acquiring a plurality of acceleration response signals of a to-be-measured plate structure subjected to an impact; determining the load of the to-be-measured plate structure by using a load identification model according to the plurality of acceleration response signals; wherein the load identification model is determined by training a graph neural network by using a training data set; and the training data set comprises a load time history of each impact position of a training plate structure and each knock and a plurality of acceleration response signals of each impact position of the training plate structure and each knock. The application improves the accuracy of impact load identification.

Description

Technical Field

[0001] This invention relates to the field of dynamic load identification, and in particular to an impact load identification method, system, electronic device, and storage medium. Background Technology

[0002] Dynamic load identification is a key issue in aerospace structural design and has received widespread attention in recent years. For example, quantifying and visualizing dynamic loads on aerospace structures is essential for purposes such as structural health monitoring, vibration control, and vibration transmission path analysis. On the one hand, acquiring dynamic loads can guide structural design and improve the structure, thereby reducing the impact of dynamic loads on structural performance. On the other hand, acquiring dynamic loads can help subsequent designs achieve precise control and effective vibration suppression. However, the actual impact loads (dynamic loads) experienced by a structure are often difficult to measure directly, while the dynamic response of a structure under external excitation is relatively easy to measure. Therefore, how to obtain the load time history based on the structural dynamic response information and necessary load inversion methods is a critical problem that urgently needs to be solved in modern engineering.

[0003] Load identification technology falls under the category of inverse structural dynamics problems. It's a technique that uses the dynamic characteristics of a structural system and measured response signals to invert external excitations. In recent decades, load identification technology has continuously developed, laying a solid foundation for load identification in hypersonic vehicle structures. Existing load identification methods mainly include model-driven methods such as frequency domain methods and time domain methods. These methods have their own distinct advantages and disadvantages, but all require prior knowledge of the mapping relationship between external loads and structural responses. However, in practical engineering, it is often difficult to obtain the impact location of unknown loads, making dynamic load identification even more challenging.

[0004] In recent years, the rapid development of artificial intelligence has made artificial neural networks (ANNs) an effective method for solving complex mapping relationships. As a data-driven method, ANNs do not rely on mechanical models and only require supervised and / or unsupervised learning of data. Therefore, ANNs have great potential in dynamic load identification applications, especially when model-driven methods show limitations. This research directly constructs an inverse model between vibration response and excitation, avoiding the process of solving for model parameters. The time history of impact force identification using ANNs under unknown location or model parameters demonstrates the potential application of deep learning in impact force identification. However, due to the "black box" nature of purely data-driven methods, ANNs require a large number of data samples for learning; insufficient training data may lead to low recognition accuracy in ANN models. Recently, research on graph neural networks (GNNs) or using machine learning to analyze graphs has received increasing attention due to their powerful expressive capabilities. A graph is a data structure used to model a set of objects (nodes) and their relationships (edges). These characteristics make GNNs no longer purely data-driven methods, but more like model-driven methods, where the model is the graph itself. Typically, GNN models learn from the edges and hidden layer representations of a graph, which encode both the graph structure and the features of the nodes. Compared to ANNs, GNNs have the unique advantage of considering the connections between nodes, providing additional topological information to the neural network and thus achieving its learning objectives.

[0005] Furthermore, incorporating physical constraints into deep learning models is also a key factor in improving the interpretability of neural network models. Summary of the Invention

[0006] The purpose of this invention is to provide an impact load identification method, system, electronic device, and storage medium to improve the accuracy of impact load identification.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] An impact load identification method, comprising:

[0009] Acquire multiple acceleration response signals of the impacted test plate structure;

[0010] Based on multiple acceleration response signals, the load on the test plate structure is determined using a load identification model; wherein, the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact.

[0011] Optionally, the graph neural network is trained using a training dataset, specifically including:

[0012] Construct a training dataset and a validation dataset; the validation dataset includes the load-time history of each impact position of the validation plate structure and each impact position of the validation plate structure.

[0013] The training plate structure and the accelerometer are constructed as a graph structure, and a learnable admittance matrix is ​​preset to build a graph neural network.

[0014] Based on the sliding window, multiple acceleration response signals of each impact position and each impact on the training plate structure are input into the current graph neural network, outputting the first load time history and determining the first total loss function value;

[0015] Based on the first total loss function value, adjust the parameters of the current graph neural network to obtain the trained graph neural network;

[0016] Based on a sliding window, multiple acceleration response signals from each impact position of the verification plate structure at each impact are input into the trained graph neural network, which outputs a second load time history and determines a second total loss function value.

[0017] Determine whether the value of the second total loss function is greater than n times the value of the first total loss function;

[0018] If so, adjust the next learning rate to a preset multiple of the current learning rate, and use the trained graph neural network as the current graph neural network. Return to the step of "based on the sliding window, input multiple acceleration response signals of each impact position of the training plate structure for each impact into the current graph neural network, and output the first load time history" until the preset number of training times is reached.

[0019] If not, the trained graph neural network will be used as the load recognition model.

[0020] Optionally, a training dataset is constructed, specifically including:

[0021] Multiple acceleration sensors are arranged on the training plate structure; the acceleration sensors are connected to a laser vibration meter.

[0022] A training dataset is constructed by repeatedly striking each impact point of a training plate structure with a force hammer, recording the acceleration response signal and load time history; the force hammer is equipped with a force sensor, which is connected to a laser vibration meter; the training dataset includes multiple sets of training data; the load time history and multiple acceleration response signals recorded after striking one impact point constitute one set of training data.

[0023] Optionally, based on a sliding window, multiple acceleration response signals from each impact position of the training plate structure for each strike are input into the current graph neural network, outputting a first load time history and determining a first total loss function value, specifically including:

[0024] Based on the sliding window, multiple acceleration response signals of each impact position and each impact on the training plate structure are input into the current graph neural network, and the first load time history is output.

[0025] Based on the first load time history and the load time history of each impact position and each impact on the training plate structure, the first mean square criterion loss function value for each window length is determined.

[0026] The first total loss function value is determined based on the first mean square criterion loss function value for each window length.

[0027] Optionally, based on a sliding window, multiple acceleration response signals from each impact point of the verification plate structure at each impact are input into the trained graph neural network to output a second load time history and determine a second total loss function value, specifically including:

[0028] Based on a sliding window, multiple acceleration response signals from each impact position of the verification plate structure at each impact are input into the trained graph neural network, which outputs a second load time history.

[0029] Based on the second load time history and the load time history of each impact position and each impact on the verification plate structure, the second mean square criterion loss function value for each window length is determined.

[0030] The second total loss function value is determined based on the second mean square criterion loss function value for each window length.

[0031] Optionally, the acceleration response signal is acquired by an acceleration sensor; multiple acceleration sensors are disposed on the test board structure.

[0032] An impact load identification system, applying the above-mentioned impact load identification method, includes:

[0033] The data acquisition module is used to acquire multiple acceleration response signals of the impacted test plate structure;

[0034] The load identification module is used to determine the load on the test plate structure based on multiple acceleration response signals using a load identification model; wherein the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact.

[0035] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described impact load identification method.

[0036] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described impact load identification method.

[0037] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] The impact load identification method, system, electronic device, and storage medium of this invention acquire multiple acceleration response signals of the impacted plate structure under test. Based on these multiple acceleration response signals, a load identification model is used to determine the load on the plate structure under test. The load identification model is determined by training a graph neural network using a training dataset. The training dataset includes the load time history of each impact at each impact location on the training plate structure and multiple acceleration response signals of each impact at each impact location on the training plate structure. This invention can identify impact loads when the load application is unknown by learning the temporal and spatial information of the structure's dynamic response. Simultaneously, it uses the vibration equation of the plate structure as a constraint, making it more applicable than traditional numerical methods. Compared to existing neural network methods, it has higher accuracy and interpretability with the same amount of training samples. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 Flowchart of the impact load identification method provided by the present invention;

[0041] Figure 2 This is a schematic diagram of the graph structure applied to graph neural networks for plate structures constructed according to the present invention;

[0042] Figure 3 This is a schematic diagram of the neural network training process of the present invention;

[0043] Figure 4 This is a schematic diagram of the two-dimensional plate structure signal acquisition system used in this invention;

[0044] Figure 5This is a comparison curve of different error evaluation indicators between the identified load and the actual applied load in this invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0046] The purpose of this invention is to provide an impact load identification method, system, electronic device, and storage medium to improve the accuracy of impact load identification.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 1 As shown, the impact load identification method provided by the present invention includes:

[0050] Step 101: Acquire multiple acceleration response signals from the impacted test board structure. The acceleration response signals are acquired through acceleration sensors; multiple acceleration sensors are disposed on the test board structure.

[0051] Step 102: Based on the multiple acceleration response signals, determine the load on the test plate structure using a load identification model; wherein, the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact.

[0052] In practical applications, the vibration dynamics equation for plate structures is shown in equation (1):

[0053]

[0054] Where w is the displacement vibration in the z direction, D is the bending stiffness, h is the thickness of the plate, ρ is the density of the plate, F(t) is the time history of the load applied to the plate in the z direction, and p(x,y) is the position function of the load application.

[0055] By simplification, the vibration equation of this two-dimensional plate structure can be expressed as:

[0056]

[0057] Where K = ρh, acceleration variable

[0058] Therefore, in general, when identifying loads on a known plate structure, the structural parameters of the plate structure, namely D and K in formula (2), can first be measured, and the acceleration response a(t) of the plate vibration can be obtained through the structural vibration response. It is assumed that a neural network model can fit the values ​​in formula (2). The external load time history F(t) of the plate structure can be obtained by training the roar.

[0059] The graph structure in GNN is a data structure that models a set of objects (nodes) and their relationships (edges). These characteristics allow GNN to use less data-driven methods and more model-driven methods, where the model is the graph. Compared with ANNs, the unique advantage of GNN is that it considers the connections between nodes that provide additional topological information to help the graph neural network achieve its learning objectives. Therefore, based on formula (2), this invention designs a graph structure for the GNN model of the vibration equation of a two-dimensional plate structure, as follows: Figure 2 As shown. In Figure 2 There are N nodes in total, where each of nodes 1 to N-1 corresponds to an accelerometer sensor, and its node signal is the acceleration response; node N is the variable to be determined. Since the initial node signals are unknown, zero signals are used instead. No preset values ​​are used in this model. The edge relationships between nodes and acceleration signal nodes are determined using a learnable admittance matrix. Alternatively, through the GNN itself on the admittance matrix Training and learning The edge relationships between nodes and acceleration signal nodes.

[0060] Therefore, a three-layer GNN model is designed, and the input of the entire model is a signal matrix consisting of the acceleration response vector and the zero vector of the window length. As shown in formula (3):

[0061]

[0062] Among them, A1, A2, ..., A N-1 This represents the acceleration response signal from node 1 to node N-1. The window length H, as a hyperparameter of this model, serves to segment the data before inputting it into the model, i.e., through t1-t... H Accelerometer signals at time t are used to predict t HThe payload value at each time step. Inputting all training sample data into the neural network model at once may cause two problems: 1. The hardware memory cannot handle it; 2. Training and fitting all training samples at once may result in excessively long training times and slow convergence speeds. Therefore, setting a hyperparameter for the segmented input window length H is necessary. However, there is currently no specific criterion to determine the size of the window length H, and the window length H only serves to segment the input; changes in its size have no significant impact on the recognition results.

[0063] The first layer of this GNN model is as follows:

[0064]

[0065] Where σ(•) is the activation function, The admittance matrix is ​​learnable. and Let be a learnable weight matrix, where Hid is the number of hidden neurons. This is the output signal of the fitting layer (first layer).

[0066] The second layer will output the signal from the first layer. The input is as follows:

[0067]

[0068] in, and The weight matrix is ​​learnable. This is the output of the output layer (second layer). Rows 1 to N-1 represent the values ​​corresponding to each acceleration signal at that given moment. value.

[0069] The third layer is composed of rows 1 to N-1 of Y2. And the acceleration values ​​extracted from A for each moment constitute As input, A a Y represents the actual acceleration signal from N-1 accelerometers at this moment. a This represents the current position of each of the N-1 accelerometer sensors. The value is calculated using the vibration equation of the plate, i.e., formula (2), as shown below:

[0070] F = Mean(S·D·Y) a +S·K·A a (6)

[0071] Where D is the known bending stiffness of the plate, and K = ρh is also known. Since the load position parameter p(x,y) in formula (2) is unknown, it is calculated based on the unit of each quantity. The area S of the plate is selected to be added to the network for training, which is also known. Mean(·) represents the final load value obtained by averaging the values ​​from the accelerometer. The payload value F output by the GNN network is compared with the control sample at the corresponding time point, and the loss function is calculated before backpropagation is performed.

[0072] The process of training a graph neural network (GNN) is as follows: Figure 3 As shown, it specifically includes:

[0073] S1: Construct training and validation datasets; the validation dataset includes the load-time history of each impact position and each impact on the validation plate structure and multiple acceleration response signals of each impact position and each impact on the validation plate structure.

[0074] As an optional implementation, constructing a training dataset specifically includes:

[0075] Multiple acceleration sensors are arranged on the training plate structure; the acceleration sensors are connected to a laser vibration meter.

[0076] A training dataset is constructed by repeatedly striking each impact point of a training plate structure with a force hammer, recording the acceleration response signal and load time history; the force hammer is equipped with a force sensor, which is connected to a laser vibration meter; the training dataset includes multiple sets of training data; the load time history and multiple acceleration response signals recorded after striking one impact point constitute one set of training data.

[0077] In practical applications, accelerometers and acquisition systems are arranged on the training plate structure (training plate structure), such as... Figure 4As shown in the diagram, N-1 accelerometers are used, forming an envelope that covers the plate structure. These accelerometers are connected to an external laser vibrometer to collect the acceleration time history (acceleration response signal) of the plate structure during vibration. A force sensor on the hammer is also connected to the external laser vibrometer to collect the dynamic load time history of each impact. Mc impact locations are uniformly marked on the plate as impact points. Approximately 15% of these Mc impact locations are selected as training set locations, and the validation set locations are the same as the training set locations. One set of load time histories and N-1 sets of acceleration responses on the plate structure are recorded for each impact at each impact location, forming one set of data. M1 impacts are performed at each impact location in the training set to form training set samples, M2 impacts are performed at each impact location in the validation set to form validation set samples, and M3 impacts are performed at each impact location in the test set to form test set samples. The ratio of M1:M2:M3 is approximately 15:1:1.

[0078] S2: Construct a graph structure by combining the training plate structure and the accelerometer, and pre-set a learnable admittance matrix to build a graph neural network.

[0079] In practical applications, the plate structure and sensor acquisition system are constructed as a graph structure. And a learnable admittance matrix is ​​preset.

[0080] S3: Based on the sliding window, input multiple acceleration response signals of each impact position and each impact of the training plate structure into the current graph neural network, and output the first load time history.

[0081] S4: Based on the first load time history and the load time history of each impact position and each impact on the training plate structure, determine the first mean square criterion loss function value for each window length.

[0082] S5: Determine the first total loss function value based on the first mean square criterion loss function value for each window length.

[0083] S6: Adjust the parameters of the current graph neural network according to the first total loss function value to obtain the trained graph neural network.

[0084] S7: Based on the sliding window, input multiple acceleration response signals of each impact position of the verification plate structure for each impact to the trained graph neural network, and output the second load time history.

[0085] S8: Based on the second load time history and the load time history of each impact position of the verification plate structure for each impact, determine the second mean square criterion loss function value for each window length.

[0086] S9: Determine the second total loss function value based on the second mean square criterion loss function value for each window length.

[0087] S10: Determine whether the value of the second total loss function is greater than n times the value of the first total loss function.

[0088] S11: If so, adjust the next learning rate to a preset multiple of the current learning rate, and use the trained graph neural network as the current graph neural network. Return to the step of "based on the sliding window, input multiple acceleration response signals of each impact position of the training plate structure for each impact to the current graph neural network, and output the first load time history" until the preset number of training times is reached.

[0089] S12: If not, then the trained graph neural network will be used as the load recognition model.

[0090] In practical applications, the training set samples are shuffled and then processed from t1 to t2. H Start by taking t2-t in sequence H+1 ,...,t n -t n-1+H The acceleration sensor signal at time t and the zero signal are combined into a signal matrix according to formula (3). The input is fed into the GNN model above for training, and the output payload value is... The dynamic load t collected from the training set samples n-1+H The values ​​at each time point are compared, and the mean squared loss function (MSE) corresponding to the training set samples for each window length is calculated. After sequentially taking windows and summing the loss functions of all training set samples, the final value Loss1 is output, followed by backpropagation. After training and backpropagation of the training set samples are complete, the same operation is performed on the validation set samples, and the loss function Loss2 is calculated. Validation set samples do not participate in the backpropagation of the GNN model. If Loss2 > 1.1 × Loss1, the learning rate is adjusted to 0.7 times the learning rate of the previous step to prevent overfitting of the GNN model. This process completes one training iteration of the GNN model. Approximately 200 training iterations are performed on the GNN model, observing the loss function in each training iteration and adjusting the initial learning rate of the GNN model until the loss function curve converges.

[0091] After the GNN model is trained, the hyperparameters and weights are fixed and remain unchanged. The test set samples are then windowed and input into the trained graph neural network in the manner described above. After each windowing, the dynamic load value at the corresponding time step is input. After windowing is completed, a complete dynamic load vector f is formed. recon This refers to the identified test set dynamic load time history.

[0092] Two criteria for evaluating recognition performance are defined as follows: First, the description of the actual applied load vector f. load Graph neural network reconstructs the load vector f recon The overall relative error between them RE = Meanwhile, since it is an impact load identification, the peak value of the impact load is also a very important indicator. It is the actual applied load vector f load Reconstructing the load vector f using GNN recon The peak error between the two is calculated. For each set of sample data in the test set, the recognition error is calculated and the PE and RE of all sample data are averaged. Recognition is considered successful when the averaged recognition error PE < 10% and RE < 20%. The graph neural network reconstruction payload vector f for each set of sample data is then output. recon .

[0093] Four accelerometers and a data acquisition system were deployed on the plate structure to be trained. The load-time history and acceleration response of the plate structure at 14 different impact locations were recorded as training, validation, and test sets. The training set consisted of 16 impacts at the 14 different impact locations, the validation set consisted of one impact at each of the 14 different impact locations, and the test set consisted of one impact at each of the 49 different impact locations. The plate structure and the sensor data acquisition system were constructed as a graph structure. Acceleration signals from four sensors are used as signal features for acceleration nodes. The test set samples are input into the three-layer graph neural network for training using a segmented windowing method. The hyperparameters of the neural network are adjusted until the mean squared error (MSE) loss function converges. Validation set samples are used in the training to prevent overfitting. The test set samples are then input into the trained graph neural network, which outputs the identified dynamic load time history of the test set. The average error of all identification results on the test set is calculated to be RE = 19.48% and PE = 8.23%, meeting the error criteria. The identification result closest to the average value is selected. Figure 5 As shown, (a) is a graph of the overall error between the identified load and the actual applied load, and (b) is a graph of the peak error between the identified load and the actual applied load.

[0094] As can be seen from the results of the above embodiments, the present invention can identify the dynamic load time history at different unknown locations on the plate structure, and the magnitude and regularity are basically consistent with the applied load.

[0095] This invention uses the structural acceleration response acquired by an accelerometer as the input to a neural network. The load time history of a hammer excitation is used as the output of the neural network. The edges connecting the load nodes and response nodes represent complex and uncertain mapping relationships that the neural network needs to learn. A "forward problem" approach is adopted to solve the inverse problem of load identification, mitigating the non-convergence issues caused by inverting the transfer function matrix in traditional numerical methods.

[0096] This invention encodes the influence of impact loads at different locations on the sensor response into a learnable admittance matrix by calibrating and sampling the loads at different locations. By using a graph structure as a model to abstractly represent the relative distance relationship between load and response, which in turn affects the output of hidden layer neurons, and by incorporating the physical dynamics equations of plate structure vibration, the interpretability of the GNN model is further improved. Ultimately, it achieves the identification of the time history of impact loads acting on uncalibrated locations, with a significant improvement in accuracy compared to ANNs methods.

[0097] This invention requires relatively little training hardware and can guarantee that the recognition accuracy meets the requirements, making it suitable for widespread use in engineering applications.

[0098] Example 2

[0099] In order to execute the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, an impact load identification system is provided below, including:

[0100] The data acquisition module is used to acquire multiple acceleration response signals of the impacted test plate structure.

[0101] The load identification module is used to determine the load on the test plate structure based on multiple acceleration response signals using a load identification model; wherein the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact.

[0102] Example 3

[0103] The present invention provides an electronic device, including: a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the impact load identification method of Embodiment 1.

[0104] Example 4

[0105] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the impact load identification method of Embodiment 1.

[0106] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0107] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying impact loads, characterized in that, include: Acquire multiple acceleration response signals of the impacted test plate structure; Based on multiple acceleration response signals, the load on the test plate structure is determined using a load identification model; wherein, the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact. Training a graph neural network using a training dataset specifically includes: Construct a training dataset and a validation dataset; the validation dataset includes the load-time history of each impact position of the validation plate structure and multiple acceleration response signals of each impact position of the validation plate structure. The training plate structure and the accelerometer are constructed as a graph structure, and a learnable admittance matrix is ​​preset to build a graph neural network. Based on the sliding window, multiple acceleration response signals of each impact position and each impact of the training plate structure are input into the current graph neural network, outputting the first load time history and determining the first total loss function value; Based on the first total loss function value, adjust the parameters of the current graph neural network to obtain the trained graph neural network; Based on a sliding window, multiple acceleration response signals from each impact position of the verification plate structure at each impact are input into the trained graph neural network, which outputs a second load time history and determines a second total loss function value. Determine whether the value of the second total loss function is greater than n times the value of the first total loss function; If so, adjust the next learning rate to a preset multiple of the current learning rate, and use the trained graph neural network as the current graph neural network. Return to the step of "based on the sliding window, input multiple acceleration response signals of each impact position of the training plate structure for each impact into the current graph neural network, and output the first load time history" until the preset number of training times is reached. If not, the trained graph neural network will be used as the load recognition model.

2. The impact load identification method according to claim 1, characterized in that, Constructing the training dataset specifically includes: Multiple acceleration sensors are arranged on the training plate structure; the acceleration sensors are connected to a laser vibration meter. A training dataset is constructed by repeatedly striking each impact point of a training plate structure with a force hammer, recording the acceleration response signal and load time history; the force hammer is equipped with a force sensor, which is connected to a laser vibration meter; the training dataset includes multiple sets of training data; the load time history and multiple acceleration response signals recorded after striking one impact point constitute one set of training data.

3. The impact load identification method according to claim 1, characterized in that, Based on a sliding window, multiple acceleration response signals from each impact point on the training plate structure are input into the current graph neural network, outputting the first load time history and determining the first total loss function value, specifically including: Based on the sliding window, multiple acceleration response signals of each impact position and each impact on the training plate structure are input into the current graph neural network, and the first load time history is output. Based on the first load time history and the load time history of each impact position and each impact on the training plate structure, the first mean square criterion loss function value for each window length is determined. The first total loss function value is determined based on the first mean square criterion loss function value for each window length.

4. The impact load identification method according to claim 1, characterized in that, Based on a sliding window, multiple acceleration response signals from each impact point on the verification plate structure are input into the trained graph neural network to output a second load time history and determine a second total loss function value, specifically including: Based on a sliding window, multiple acceleration response signals from each impact position of the verification plate structure at each impact are input into the trained graph neural network, which outputs a second load time history. Based on the second load time history and the load time history of each impact position and each impact on the verification plate structure, the second mean square criterion loss function value for each window length is determined. The second total loss function value is determined based on the second mean square criterion loss function value for each window length.

5. The impact load identification method according to claim 1, characterized in that, The acceleration response signal is acquired by an acceleration sensor; multiple acceleration sensors are disposed on the test board structure.

6. An impact load identification system, characterized in that, The impact load identification system is used to implement the impact load identification method according to any one of claims 1-5, and the impact load identification system includes: The data acquisition module is used to acquire multiple acceleration response signals of the impacted test plate structure; The load identification module is used to determine the load on the test plate structure based on multiple acceleration response signals using a load identification model; wherein the load identification model is determined by training a graph neural network using a training dataset; the training dataset includes the load time history of each impact position of the training plate structure for each impact and multiple acceleration response signals of each impact position of the training plate structure for each impact.

7. An electronic device, characterized in that, include: A memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to cause the electronic device to perform the impact load identification method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the impact load identification method according to any one of claims 1-5.

Images