A method, device, medium and product for reconstructing and locating a time history of an impact load

Abstract

The application discloses a shock load time history reconstruction and positioning method, device, medium and product, relates to the field of shock load identification, and comprises the following steps: acquiring an acceleration signal matrix and a dynamic graph structure of a shock load acting object; according to the acceleration signal matrix and the dynamic graph structure, a time history reconstruction model is used to obtain a reconstructed shock load time history of the shock load acting object; the time history reconstruction model is obtained by training a graph neural network model by using a first training data set and a dynamic graph structure of a shock load acting object for training; according to an encoding distance matrix of the acceleration signal matrix, a shock positioning model is used to determine the shock position of the shock load acting object; and the shock positioning model is obtained by training a fully-connected neural network by using a second training data set. The application can accurately identify the shock load time history and the shock position when the load acting position is unknown.

Description

Technical Field

[0001] This application relates to the field of impact load identification, and in particular to a method, device, medium, and product for reconstructing and locating the time history of impact loads. Background Technology

[0002] Dynamic load identification is a major concern in aerospace structural design and has received widespread attention in recent years. The need for quantification and visualization of dynamic loads on aerospace structures is increasing for structural health monitoring, vibration control, and vibration transmission path analysis. Reconstructing dynamic loads based on the structure's dynamic characteristics and measured responses (such as acceleration, velocity, displacement, and strain) is a challenging inverse problem. Extensive research has been conducted on reconstructing the time history of loads when the impact load location is assumed to be known. Many scholars have handled the severe ill-posedness of the inverse problem relatively accurately using a series of frequency and time domain methods. However, in engineering practice and other real-world problems, load identification is more practical but also more challenging when the impact location is unknown. It is difficult to solve due to two characteristics. First, it is an inverse problem. Second, the transmission relationship between the input (external impact force) and the output (structural response) is also unknown, further complicating the problem. In this case, many studies have set assumptions and used prior knowledge to constrain the problem, making it solvable numerically. By assuming candidate impact locations, the time history of the impact load at each candidate location is considered as part of the total load action at all candidate locations. However, when the impact load is not applied completely to the candidate location, the transfer function between the impact location and the response location is completely uncertain, which challenges the effectiveness of the above method.

[0003] With the rapid development of artificial intelligence in recent years, Artificial Neural Networks (ANNs) have become an effective method for solving complex mapping relationships. As a data-driven approach, ANNs do not rely on mechanical models and only require supervised and / or unsupervised learning of data. Therefore, ANNs have great potential in load recognition applications, especially when the aforementioned model-driven methods show limitations. However, since purely data-driven methods operate like a black box, ANNs require a large amount of data to learn the response of the target structure. Although current research on simple structures has demonstrated that ANNs can achieve good accuracy in reconstructing load time histories using only the temporal information of the structure's dynamic characteristics, providing additional spatial information would certainly help reduce sample size and achieve higher accuracy and reliability on complex structures such as real aircraft structures.

[0004] Recently, Graph Neural Networks (GNNs) have gained increasing attention due to the powerful expressive power of graphs. A graph is a data structure that models a set of objects (nodes) and their relationships (edges). These characteristics also make GNNs less reliant on data-driven methods and more reliant on model-driven methods, where the model is the graph. GNNs can be used to represent a wide range of systems across various fields, including social sciences (such as social networks), natural sciences (such as physical systems and structural health monitoring), knowledge graphs, and many other research areas. Generally, GNN models learn from graph edge and hidden layer representations, which encode the local graph structure and node features. Compared to ANNs, a unique advantage of GNNs is that they consider the connections between nodes, providing additional topological information to the neural network to help achieve the learning objectives. However, due to the generality of GNNs, this topological information is constructed in an abstract way, leading to an inability to accurately identify the time history and impact location of impact loads when the load application location is unknown. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, medium, and product for reconstructing and locating the time history of impact loads, so as to accurately identify the time history and impact location of impact loads when the location of the load application is unknown.

[0006] To achieve the above objectives, this application provides the following solution:

[0007] Firstly, this application provides a method for reconstructing and locating the time history of impact loads, including:

[0008] The acceleration signal matrix and dynamic graph structure of the impact-loaded object are obtained. The dynamic graph structure includes a dynamic weighted metric matrix and a dynamic admittance matrix. The dynamic admittance matrix is ​​determined based on the dynamic weighted adjacency matrix and the dynamic weighted metric matrix. The dynamic weighted adjacency matrix is ​​determined based on the adjacency matrix and the dynamic weight matrix. The adjacency matrix represents the undirected graph structure of the impact-loaded object. The undirected graph structure consists of N nodes and N×(N-1) edges. N acceleration sensors are installed on the impact-loaded object. One node corresponds to one acceleration sensor. The dynamic weight matrix is ​​determined based on the Euclidean distance between the features of each node. The node features are the acceleration signals collected by the acceleration sensors. The dynamic weighted metric matrix is ​​determined based on the dynamic weighted adjacency matrix.

[0009] Based on the acceleration signal matrix and the dynamic graph structure, a time history reconstruction model is used to obtain the reconstructed impact load time history of the impact load-bearing object; wherein, the time history reconstruction model is obtained by training a graph neural network model using a first training dataset and the dynamic graph structure of the impact load-bearing object used for training; the first training dataset includes the acceleration signal matrix of the impact load-bearing object used for training and the corresponding real impact load time history; the graph neural network model includes a graph interaction layer, a reconstruction layer and a graph fusion layer connected in sequence;

[0010] Based on the encoded distance matrix of the acceleration signal matrix, the impact location of the object subjected to the impact load is determined using an impact location model. The encoded distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix. The reconstructed impact load time history of each node is determined using the reconstruction layer of the time history reconstruction model. The impact location model is obtained by training a fully connected neural network using a second training dataset. The second training dataset includes the encoded distance matrix of the acceleration signal matrix of the object subjected to the impact load and the corresponding true impact location. The fully connected neural network includes a first fully connected layer and a second fully connected layer connected sequentially.

[0011] Optionally, the graph neural network model is trained using the first training dataset and the dynamic graph structure of the training object subjected to the impact load, specifically including:

[0012] Obtain the training dataset, validation dataset, and dynamic graph structure of the object subjected to the validation impact load; the validation dataset includes the acceleration signal matrix of the object subjected to the validation impact load and the corresponding real impact load time history.

[0013] The acceleration signal matrix and dynamic graph structure of the training impact load object are input into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object;

[0014] Based on the reconstructed impact load time history of the training impact load object and the corresponding real impact load time history, the loss function value of the training dataset is determined using the mean squared error loss function.

[0015] The model parameters of the current graph neural network model are adjusted by backpropagation based on the loss function values ​​of the training dataset to obtain the trained graph neural network model.

[0016] The acceleration signal matrix and dynamic graph structure of the object subjected to the verification impact load are input into the currently trained graph neural network model to obtain the reconstructed impact load time history of the object subjected to the verification impact load.

[0017] Based on the reconstructed impact load time history and the corresponding real impact load time history of the object subjected to the impact load for verification, the loss function value of the verification dataset is determined using the mean squared error loss function.

[0018] Determine if the maximum number of iterations has been reached;

[0019] If so, then end the training and use the currently trained graph neural network model as the time history reconstruction model;

[0020] If not, adjust the current learning rate based on the loss function values ​​of the training dataset and the validation dataset, and use the trained graph neural network model as the current graph neural network model, returning "Input the acceleration signal matrix and dynamic graph structure of the training impact load object into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object".

[0021] Optionally, the acceleration signal matrix and dynamic graph structure of the training impact load object are input into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object, specifically including:

[0022] Based on the acceleration signal matrix and dynamic graph structure of the training impact load object, the encoding of each node is determined using the graph interaction layer;

[0023] Based on the encoding of each node, the reconstruction layer is used to determine the reconstruction impact load time history of each node;

[0024] Based on the reconstructed impact load time history of each node, the reconstructed impact load time history of the training impact load object is determined using the graph fusion layer.

[0025] Optionally, based on the reconstructed impact load time history of the training impact load object and the corresponding real impact load time history, the loss function value of the training dataset is determined using the mean squared error loss function, specifically including:

[0026] Using formula Determine the loss function value for the training dataset; where, The time history of the reconstructed impact load is used to train the impact load object; F is the actual impact load time history; pr is the Pearson correlation coefficient; the Pearson correlation coefficient is determined based on the encoding distance of the acceleration signal matrix of the training impact load object and the physical distance between the impact position and the acceleration sensor.

[0027] Optionally, the current learning rate is adjusted based on the loss function values ​​of the training dataset and the validation dataset, specifically including:

[0028] Determine whether the loss function value of the validation dataset is greater than the loss function value of the training dataset by a set multiple;

[0029] If so, adjust the current learning rate.

[0030] Optionally, the encoding distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix, specifically including:

[0031] Using formula Determine the coding distance for each node; where δ is the coding distance; and FG is the reconstruction impact load time history of the node. It is a dynamically weighted metric matrix; =Hadamard product; Mean(·) is the averaging function of operations on each row of the matrix; Diag(·) is the transformation function that transforms a diagonal matrix into a column vector;

[0032] The coding distances corresponding to N nodes constitute a coding distance matrix.

[0033] Optionally, based on the encoded distance matrix of the acceleration signal matrix, the impact location of the object subjected to the impact load is determined using an impact location model, specifically including:

[0034] Using formula Determine the impact location of the object subjected to the impact load; wherein, The impact position of the object subjected to the impact load; σ(·) is the activation function; δ is the encoding distance; T is the transpose; Here is the weight matrix of the first fully connected layer; is the weight matrix of the second fully connected layer; Hid is the number of neurons in the hidden layer.

[0035] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the impact load time history reconstruction and positioning method described above.

[0036] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the impact load time history reconstruction and positioning method described above.

[0037] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the impact load time history reconstruction and positioning method described above.

[0038] According to the specific embodiments provided in this application, this application has the following technical effects:

[0039] This application provides a method, device, medium, and product for reconstructing and locating the time history of impact loads, acquiring the acceleration signal matrix and dynamic graph structure of the impact load-bearing object. The dynamic graph structure includes a dynamic weighted metric matrix and a dynamic admittance matrix. The dynamic admittance matrix is ​​determined based on the dynamic weighted adjacency matrix and the dynamic weighted metric matrix. The dynamic weighted adjacency matrix is ​​determined based on the adjacency matrix and the dynamic weight matrix. The adjacency matrix represents the undirected graph structure of the impact load-bearing object. The undirected graph structure consists of nodes and edges. An acceleration sensor is installed on the impact load-bearing object; one node corresponds to one acceleration sensor. The dynamic weight matrix is ​​determined based on the Euclidean distance between the features of each node. The dynamic weighted metric matrix is ​​determined based on the dynamic weighted adjacency matrix. Based on the acceleration signal matrix and dynamic graph structure, a time history reconstruction model is used to obtain the reconstructed impact load time history of the impact load-bearing object. The time history reconstruction model is obtained by training a graph neural network model using a first training dataset and the dynamic graph structure of the impact load-bearing object. Based on the encoded distance matrix of the acceleration signal matrix, an impact location model is used to determine the impact location of the impact load-bearing object. The encoded distance matrix is ​​determined based on the reconstructed impact load time history and dynamic weighted metric matrix of each node. The reconstructed impact load time history of each node is determined using the reconstruction layer of the time history reconstruction model. The impact location model is obtained by training a fully connected neural network using a second training dataset. This application can accurately identify the impact load time history and impact location when the load application location is unknown. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating an impact load time history reconstruction and positioning method provided in an embodiment of this application;

[0042] Figure 2This is a flowchart illustrating the impact load time history reconstruction and impact location method for plate structures under unknown load application locations as described in this application.

[0043] Figure 3 This is a schematic diagram of the diagram structure used in this application;

[0044] Figure 4 This is a diagram of the helicopter web structure used in this application;

[0045] Figure 5 This is a layout diagram of the sensor acquisition system used in this application;

[0046] Figure 6 This is a comparison diagram of the time history reconstruction of the identified load and the actual applied load, and a schematic diagram of the error evaluation index.

[0047] Figure 7 This is a comparison diagram of the impact positioning results of the identified load and the actual applied load in this application;

[0048] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] This application provides a graph neural network model capable of identifying the time history and impact location of impact loads when the load application is unknown by learning the temporal and spatial information of the structural dynamic response, and applies it to real aircraft structures. This method has stronger applicability than traditional numerical methods and existing neural network methods.

[0052] In one exemplary embodiment, such as Figure 1 As shown, a method for reconstructing and locating the time history of impact loads is provided, including the following steps:

[0053] S1: Obtain the acceleration signal matrix and dynamic graph structure of the object subjected to the impact load; the dynamic graph structure includes a dynamic weighted metric matrix and a dynamic admittance matrix; the dynamic admittance matrix is ​​determined based on the dynamic weighted adjacency matrix and the dynamic weighted metric matrix; the dynamic weighted adjacency matrix is ​​determined based on the adjacency matrix and the dynamic weight matrix; the adjacency matrix is ​​used to represent the undirected graph structure of the object subjected to the impact load; the undirected graph structure consists of N nodes and N×(N-1) edges; N acceleration sensors are set on the object subjected to the impact load; one node corresponds to one acceleration sensor; the dynamic weight matrix is ​​determined based on the Euclidean distance between the features of each node; the node features are the acceleration signals collected by the acceleration sensors; the dynamic weighted metric matrix is ​​determined based on the dynamic weighted adjacency matrix.

[0054] S2: Based on the acceleration signal matrix and the dynamic graph structure, the reconstructed impact load time history of the impact load object is obtained using a time history reconstruction model; wherein, the time history reconstruction model is obtained by training a graph neural network model using a first training dataset and the dynamic graph structure of the impact load object used for training; the first training dataset includes the acceleration signal matrix of the impact load object used for training and the corresponding real impact load time history; the graph neural network model includes a graph interaction layer, a reconstruction layer and a graph fusion layer connected in sequence.

[0055] S3: Based on the encoded distance matrix of the acceleration signal matrix, the impact position of the impact load object is determined using the impact location model; the encoded distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix; the reconstructed impact load time history of each node is determined using the reconstruction layer of the time history reconstruction model; wherein, the impact location model is obtained by training a fully connected neural network using a second training dataset; the second training dataset includes the encoded distance matrix of the acceleration signal matrix of the impact load object used for training and the corresponding true impact position; the fully connected neural network includes a first fully connected layer and a second fully connected layer connected sequentially.

[0056] As an optional implementation, the encoding distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix, specifically including:

[0057] Using formula Determine the coding distance for each node; where δ is the coding distance; and FG is the reconstruction impact load time history of the node. It is a dynamically weighted metric matrix; =Hadamar product; Mean(·) is the average function of operations on each row of the matrix; Diag(·) is the transformation function that transforms a diagonal matrix into a column vector.

[0058] The coding distances corresponding to N nodes constitute a coding distance matrix.

[0059] As an optional implementation, S3 specifically includes:

[0060] Using formula Determine the impact location of the object subjected to the impact load; wherein, The impact position of the object subjected to the impact load; σ(·) is the activation function; δ is the encoding distance; T is the transpose; Here is the weight matrix of the first fully connected layer; is the weight matrix of the second fully connected layer; Hid is the number of neurons in the hidden layer.

[0061] Inspired by GNNs, this application develops a novel graph neural network model for practical applications. Driven by a graph structure and constrained by real physical information, this model can learn from the temporal and spatial information of the structural dynamic response, simultaneously identifying the time history and impact location of the impact load. Finally, an application verification based on graph neural network-based impact load reconstruction and localization is conducted using a typical aircraft structure, such as a helicopter web structure.

[0062] For an impact load identification system for a helicopter web structure, N acceleration sensors S i (i = 1, 2, ..., N) are arranged on the helicopter web structure. Position, used to measure the time-domain acceleration signal. (S) i The measured data are in vector form S i It is indicated that vector S i The length of the vector is L, named the sensor signal vector. The load is a single-point impact force, and the impact position is P. F (x, y) are random within the identification region. The reconstructed impact load time history is represented by a vector. The measured data of the applied load are represented as a vector F, named the true impact load time history, and used as a comparison sample.

[0063] For the helicopter web structure, the load identification control equation in the time domain is as follows:

[0064]

[0065] Where S(t) is the time history of the sensor signal, P F (x, y) represents the impact location, P s(x, y) represents the sensor position, and F(t) represents the actual impact load time history. To reconstruct the time history of the impact load, ε represents the acceptable error between the two. Ω s,F It is the transfer function between the sensor signal and the impact load signal.

[0066] Accelerometers are abstracted as response nodes, and all response nodes are connected pairwise by undirected edges, ensuring that each accelerometer is connected to all other accelerometers. The time-domain acceleration signal S acquired by each accelerometer is then... i As a feature of each node, this undirected graph structure has N nodes and N×(N-1) edges. This undirected graph structure is represented by an adjacency matrix. Represented as:

[0067]

[0068] Among them, the adjacency matrix It is a square matrix with all diagonal elements being 0 and all other elements being 1. It's worth noting that the adjacency matrix... The element values ​​in the table only represent the abstract topological connections between nodes. However, the characteristic differences between nodes, i.e., the temporal characteristic differences between accelerometers, are closely related to the edges between nodes. These differences are due to the different arrangements of the accelerometers, and therefore can be quantified using similarity criteria.

[0069] Therefore, the feature S of each pair of nodes is calculated. i and S j The Euclidean distance M between them i，j As shown below:

[0070] M i，j =||S i -S j ||2 (4)

[0071] Among them, M i，j Let M be the Euclidean distance between the features of node i and node j, where i, j = 1, 2, ..., N, and M is the Euclidean distance between them. i，j The larger the number of nodes, the worse the similarity between their features.

[0072] Therefore, the following Euclidean distance matrix can be obtained.

[0073]

[0074] in, It is a symmetric matrix with zero elements on its diagonal. The differences in its elements can be considered as weights for the effectiveness of information arising from the different positions of the accelerometers. It is called a dynamic weight matrix.

[0075] By adjacency matrix and dynamic weight matrix Element-wise multiplication allows the effective differences in information between node features to be used as a concrete representation of edges, replacing the abstract topological relationships of matrices. This yields a dynamically weighted adjacency matrix. As shown below:

[0076]

[0077] in, It is the Hadamard product. Then, a dynamically weighted metric matrix can be obtained. As shown below:

[0078]

[0079] Among them, the dynamic weighted metric matrix The element represents the sum of the information validity weights between the node and other nodes in the graph.

[0080] With dynamic weighted adjacency matrix and dynamic weighted metric matrix A matrix called the dynamic admittance matrix can be calculated. The matrix is ​​shown below:

[0081]

[0082] Therefore, based on the above theory, the designed three-layer graph neural network model is shown below. First, the node features of each node are used to form the input matrix. As shown below:

[0083]

[0084] The first layer is the graph interaction layer, where the input matrix is... and dynamic graph structures (including dynamic weighted metric matrices) and dynamic admittance matrix Together, they serve as input. Each node's features exchange and propagate information through a graph structure. The first layer can be represented as follows:

[0085]

[0086] Where σ(·) is the activation function, typically the ReLU function. and It is a learnable weight matrix, where Hid is the number of hidden neurons. This is the output signal of the graph interaction layer, i.e., the encoding of each node. The first term... Effective information extraction for each node's features, the second term Δ·S·W 12 This represents the exchange and dissemination of information throughout the entire graph structure.

[0087] The second layer is the reconstruction layer. It encodes each node output from the first layer. Reconstructed into a single load feature. Similar to the first layer, the second layer can be represented as follows:

[0088]

[0089] in, and It is a learnable weight matrix. It's worth noting that the output of the second layer... It should be the time history of the reconstructed impact load at each node. Therefore, by... and dynamic weighted metric matrix A latent variable can be calculated, and a direct linear relationship can be established between this latent variable and the different distances from the sensor at the moment of load impact. This latent variable is named the coded distance δ. N×1 , means as follows:

[0090]

[0091] Here, Mean(·) is the averaging function applied to each row of the matrix, and Diag(·) is the transformation that converts a diagonal matrix into a column vector. To assign a coding distance matrix δ... N×1 In physical terms, constraints are applied to guide the model in extracting the physical distance matrix ρ between the impact location and the sensor location. N×1 The relevant underlying characteristics. The Pearson correlation coefficient is used to measure δ. N×1 and ρ N×1 The linear correlation between them is expressed as:

[0092]

[0093] Where pr is The covariance of ρ and their standard deviation and The ratio of the product of , so the result is always between -1 and 1.

[0094] The third layer is the graph fusion layer, which uses the output of the second layer. As input, the time histories of reconstructed impact loads from each node are fused. Finally, the time-domain impact load characteristics are output. The third layer can be represented in the following form:

[0095]

[0096] Among them, the trainable weight matrix Node features used for fusing and decoding graphs.

[0097] The graph neural network model uses a modified mean squared error loss function as shown below:

[0098]

[0099] The |pr| term is added as a denominator to the reconstructed impact load time history. In the mean squared error loss function between the loss function and the control sample F, the encoded distance, which has a poor correlation with physical distance, is penalized. Finally, the loss function LossGNN is backpropagated to update the weights of the graph neural network model.

[0100] After the graph neural network is trained, a two-layer fully connected neural network is set up as shown below. First, the encoding distance δ between the training dataset samples and the validation dataset samples is extracted. N×1 The samples used as input to the training and validation sets of the fully connected network ultimately output the location of the impact load. As shown below:

[0101]

[0102] in, The impact position of the output. and These are the weight matrices for the first and second layers of the two-layer fully connected neural network model, respectively.

[0103] The two-layer fully connected neural network uses the mean squared error loss function as shown below:

[0104]

[0105] Finally, the loss function LossBP is backpropagated to update the weights of the fully connected neural network model.

[0106] As an optional implementation, the graph neural network model is trained using a first training dataset and the dynamic graph structure of the training impact load object, specifically including:

[0107] Step 1: Obtain the training dataset, the validation dataset, and the dynamic graph structure of the object subjected to the validation impact load; the validation dataset includes the acceleration signal matrix of the object subjected to the validation impact load and the corresponding real impact load time history.

[0108] Step 2: Input the acceleration signal matrix and dynamic graph structure of the training impact load object into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object.

[0109] As an optional implementation, step 2 specifically includes:

[0110] Step 21: Based on the acceleration signal matrix and dynamic graph structure of the training impact load object, determine the encoding of each node using the graph interaction layer.

[0111] Step 22: Based on the encoding of each node, use the reconstruction layer to determine the reconstruction impact load time history of each node.

[0112] Step 23: Based on the reconstructed impact load time history of each node, use the graph fusion layer to determine the reconstructed impact load time history of the training impact load object.

[0113] Step 3: Based on the reconstructed impact load time history of the training impact load object and the corresponding real impact load time history, determine the loss function value of the training dataset using the mean squared error loss function.

[0114] As an optional implementation, step 3 specifically includes:

[0115] Using formula Determine the loss function value for the training dataset; where, The time history of the reconstructed impact load is used to train the impact load object; F is the actual impact load time history; pr is the Pearson correlation coefficient; the Pearson correlation coefficient is determined based on the encoding distance of the acceleration signal matrix of the training impact load object and the physical distance between the impact position and the acceleration sensor.

[0116] Step 4: Adjust the model parameters of the current graph neural network model by backpropagation based on the loss function values ​​of the training dataset to obtain the trained graph neural network model.

[0117] Step 5: Input the acceleration signal matrix and dynamic graph structure of the object subjected to the verification impact load into the currently trained graph neural network model to obtain the reconstructed impact load time history of the object subjected to the verification impact load.

[0118] Step 6: Based on the reconstructed impact load time history and the corresponding real impact load time history of the object subjected to the impact load for verification, determine the loss function value of the verification dataset using the mean squared error loss function.

[0119] Step 7: Determine if the maximum number of iterations has been reached.

[0120] Step 8: If yes, end the training and use the currently trained graph neural network model as the time history reconstruction model.

[0121] Step 9: If not, adjust the current learning rate according to the loss function value of the training dataset and the loss function value of the validation dataset, and use the trained graph neural network model as the current graph neural network model, and return "Input the acceleration signal matrix and dynamic graph structure of the training impact load object into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object".

[0122] As an optional implementation, adjusting the current learning rate based on the loss function values ​​of the training dataset and the validation dataset specifically includes:

[0123] Determine whether the loss function value of the validation dataset is greater than the loss function value of the training dataset by a set multiple.

[0124] If so, adjust the current learning rate.

[0125] In practical applications, after constructing the graph neural network model and the fully connected neural network, as follows: Figure 2 As shown, the specific training steps are as follows:

[0126] Step 1: Deploy sensors and a data acquisition system on the helicopter's web structure. The sensors consist of N accelerometers connected to an external laser vibrometer to collect the acceleration time history of the web structure during vibration. Force sensors on a hammer are also connected to the external laser vibrometer to collect the impact load time history for each impact. The entire web structure is divided into Q sub-regions. At any location within each sub-region, the hammer strikes Q1, Q2, and Q3 times respectively, forming the training, validation, and test datasets, where Q1:Q2:Q3 is approximately 15:1:1. Simultaneously, record the actual distance ρ between each impact location and the N accelerometers. N×1 .

[0127] Step 2: Construct the sensor system on the web structure as a graph structure, such as... Figure 3 As shown, the adjacency matrix is ​​calculated. The sensor acceleration signal is used as a feature of the graph structure nodes.

[0128] Step 3: Calculate the dynamic weight matrix based on the node features between each node. Based on dynamic weight matrix and adjacency matrix The dynamic weighted adjacency matrix is ​​calculated. Dynamic weighted metric matrix and dynamic admittance matrix

[0129] Step 4: After shuffling the training dataset, assemble the acceleration signal matrix according to Eq. The input is fed into the above graph neural network model for training, and the output is the reconstructed impact load time history. Compared with the actual impact load time history collected from the training set samples A comparison is performed, and the loss function LossGNNTr for the graph neural network is calculated based on the Eq. and the Pearson correlation coefficient pr corresponding to the sample group, followed by backpropagation. After training and backpropagation on the training dataset are complete, the same operation is performed on the validation dataset, calculating the loss function LossGNNVa. The validation dataset is not involved in the backpropagation of the graph neural network model. If LossGNNVa > 1.1 × LossGNNTr, the learning rate is adjusted to 0.7 times the learning rate of the previous step to prevent overfitting. This process is repeated to complete one training iteration of the graph neural network model. Approximately 200 training iterations are performed, and the loss function is observed in each iteration. The initial learning rate of the graph neural network model is adjusted until the loss function curve converges.

[0130] Step 5: After the graph neural network model is trained, fix the hyperparameters and weights so that they do not change. Input the training dataset and validation dataset into the trained graph neural network in the above manner, and output the encoding distances corresponding to the training dataset and validation dataset.

[0131] Step 6: After the graph neural network model is trained, fix the hyperparameters and weights so that they do not change. Input the test dataset into the trained graph neural network in the above manner, and output the reconstructed test set impact load time history. The distance between the test set and the encoded distance.

[0132] Step 7: After shuffling the encoded distances corresponding to the training dataset, input them sequentially into the fully connected neural network described above for training. The output is the load impact location. Compared with the actual load impact locations collected in the training dataset For comparison, the loss function LossBPTr of the fully connected neural network is calculated according to Eq., and then backpropagation is performed. After training and backpropagation of the encoding distance corresponding to the training dataset are completed, the same operation is performed on the encoding distance corresponding to the validation dataset, and the loss function LossBPVa is calculated. The encoding distance corresponding to the validation dataset is not included in the backpropagation of the fully connected neural network. If LossBPVa > 1.1 × LossBPTr, the learning rate is adjusted to 0.7 times the learning rate of the previous step to prevent overfitting of the fully connected neural network. This process is repeated to complete one training iteration of the fully connected neural network. Approximately 1000 training iterations are performed, and the loss function is observed in each training iteration. The initial learning rate of the fully connected neural network is adjusted until the loss function curve converges.

[0133] Step 8: After the fully connected neural network model has been trained, fix the hyperparameters and weights so that they do not change. Input the encoded distance corresponding to the test dataset into the trained fully connected neural network, and output the impact location of the test set load.

[0134] Step 9: To uniformly evaluate the generalization and reliability of the graph neural network model, two evaluation criteria are proposed to assess the accuracy of the impact load time history reconstruction and the accuracy of impact location, respectively: For the impact load time history reconstruction, due to the high and concentrated energy characteristics of impact loads, the peak error standard PE commonly used in structural health monitoring is used to evaluate the true impact load vector F and the reconstructed impact load time history. The peak difference, of which For load impact localization, the actual load impact position (x, y) and the reconstructed load impact position are used. The distance difference DE between them is used for evaluation, where For each set of sample data in the test dataset, identify and calculate the error. Average the PE and DE of all sample data. When the average recognition error PE < 10% and DE < 10cm, the recognition is considered successful.

[0135] like Figure 4 and Figure 5As shown, seven sensors and a data acquisition system are arranged on the helicopter's web structure. Accelerometers are connected to an external laser vibrometer to collect the acceleration time history of the web structure during vibration. Force sensors on a hammer are connected to the external laser vibrometer to collect the dynamic load time history of each impact. The entire web structure is divided into 84 sub-regions. At any point in each sub-region, the hammer is used to strike the structure 15 times, 1 time, and 1 time, respectively, forming the training set, validation set, and test set samples. The distance between the impact position and the sensor for each sample is recorded. The sensor system on the web structure is constructed as a graph structure, and the adjacency matrix is ​​calculated. The sensor acceleration signal is used as the node feature of the graph structure. The dynamic weight matrix is ​​calculated through the node features between each node. and adjacency matrix The dynamic weighted adjacency matrix is ​​calculated together. Dynamic weighted metric matrix and dynamic admittance matrix After shuffling the training dataset samples, an acceleration signal matrix is ​​constructed sequentially according to Eq. The samples are input into the graph neural network model for training. The hyperparameters of the graph neural network are adjusted until the loss function converges, and validation dataset samples are used in the training to prevent overfitting. After the graph neural network model is trained, the encoded distances corresponding to the training dataset samples and validation dataset samples are output. The encoded distances of the training set are shuffled and then sequentially input into the fully connected neural network model for training. The hyperparameters of the fully connected neural network are adjusted until the loss function converges, and the encoded distances of the validation set are used in the training to prevent overfitting. The test set samples are input into the trained graph neural network, and the outputs are the identified dynamic load time history and the test set encoded distance. The test set encoded distance is input into the trained fully connected neural network, and the outputs are the identified dynamic load impact locations of the test set. The average errors of all identification results on the test set are PE = 9.12% and DE = 6.73cm, which meet the error criteria. The impact load time history identification result closest to the average value is taken as follows: Figure 6 As shown, the impact positioning results of the helicopter web structure are as follows: Figure 7 As shown.

[0136] As can be seen from the results of the above embodiments, this application can identify the dynamic load time history and the location of impact load at different unknown positions on the helicopter web structure, and the magnitude and regularity are basically consistent with the applied load.

[0137] This application uses the structural acceleration response acquired by the sensor and the actual distance between the sensor and the impact location as input to a graph neural network, and the load time history of the hammer excitation as the output of the graph neural network. With the help of physical distance, the graph neural network learns the complex mapping relationship between the impact load at different locations and the sensor through graph structure learning, ultimately realizing the reconstruction of the impact load time history when the impact load location is unknown, which is impossible for traditional model-driven methods.

[0138] This application incorporates the influence of impact loads at different locations on sensor response into a graph neural network by calibrating and sampling the load at different locations. It constructs a latent variable encoding distance and establishes a linear relationship between the encoded distance and the true distance. This achieves high-precision positioning of the impact load through a simple fully connected neural network. Compared with ANNs, it requires fewer sensors and is more interpretable.

[0139] This application has been tested on a real helicopter web structure, requires less training hardware, and can ensure that the recognition accuracy meets the requirements, making it suitable for widespread use in real aerospace structures.

[0140] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for reconstructing and locating the time history of impact loads.

[0141] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method for reconstructing and locating the time history of impact loads.

[0142] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method for reconstructing and locating the time history of impact loads.

[0143] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for reconstructing and locating the time history of impact loads.

[0144] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0147] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, 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 this application.

Claims

1. A method for reconstructing and locating the time history of impact loads, characterized in that, include: The acceleration signal matrix and dynamic graph structure of the object subjected to the impact load are obtained; the dynamic graph structure includes a dynamic weighted metric matrix and a dynamic admittance matrix. The dynamic admittance matrix is ​​determined based on the dynamic weighted adjacency matrix and the dynamic weighted metric matrix; the dynamic weighted adjacency matrix is ​​determined based on the adjacency matrix and the dynamic weight matrix; the adjacency matrix is ​​used to represent the undirected graph structure of the impact load object; the undirected graph structure consists of N nodes and N×(N-1) edges; N acceleration sensors are set on the impact load object; one node corresponds to one acceleration sensor; the dynamic weight matrix is ​​determined based on the Euclidean distance between the features of each node; the node features are the acceleration signals collected by the acceleration sensors; the dynamic weighted metric matrix is ​​determined based on the dynamic weighted adjacency matrix. Based on the acceleration signal matrix and the dynamic graph structure, a time history reconstruction model is used to obtain the reconstructed impact load time history of the impact load-bearing object; wherein, the time history reconstruction model is obtained by training a graph neural network model using a first training dataset and the dynamic graph structure of the impact load-bearing object used for training; the first training dataset includes the acceleration signal matrix of the impact load-bearing object used for training and the corresponding real impact load time history; the graph neural network model includes a graph interaction layer, a reconstruction layer and a graph fusion layer connected in sequence; Based on the encoded distance matrix of the acceleration signal matrix, the impact location of the object subjected to the impact load is determined using an impact location model. The encoded distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix. The reconstructed impact load time history of each node is determined using the reconstruction layer of the time history reconstruction model. The impact location model is obtained by training a fully connected neural network using a second training dataset. The second training dataset includes the encoded distance matrix of the acceleration signal matrix of the object subjected to the impact load and the corresponding true impact location. The fully connected neural network includes a first fully connected layer and a second fully connected layer connected sequentially. The encoding distance matrix is ​​determined based on the reconstructed impact load time history of each node and the dynamic weighted metric matrix, specifically including: Using formula Determine the coding distance for each node; where δ is the coding distance; and FG is the reconstruction impact load time history of the node. It is a dynamically weighted metric matrix; =Hadamard product; Mean(·) is the averaging function of operations on each row of the matrix; Diag(·) is the transformation function that transforms a diagonal matrix into a column vector; The encoding distances corresponding to N nodes constitute an encoding distance matrix; Based on the encoded distance matrix of the acceleration signal matrix, the impact location of the object subjected to the impact load is determined using the impact positioning model, specifically including: Using formula Determine the impact location of the object subjected to the impact load; wherein, The impact position of the object subjected to the impact load; σ(·) is the activation function; δ is the encoding distance; T is the transpose; Here is the weight matrix of the first fully connected layer; is the weight matrix of the second fully connected layer; Hid is the number of neurons in the hidden layer.

2. The method for reconstructing and locating the time history of impact load according to claim 1, characterized in that, The graph neural network model is trained using the first training dataset and the dynamic graph structure of the training impact load object, specifically including: Obtain the training dataset, validation dataset, and dynamic graph structure of the object subjected to the validation impact load; the validation dataset includes the acceleration signal matrix of the object subjected to the validation impact load and the corresponding real impact load time history. The acceleration signal matrix and dynamic graph structure of the training impact load object are input into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object; Based on the reconstructed impact load time history of the training impact load object and the corresponding real impact load time history, the loss function value of the training dataset is determined using the mean squared error loss function. The model parameters of the current graph neural network model are adjusted by backpropagation based on the loss function values ​​of the training dataset to obtain the trained graph neural network model. The acceleration signal matrix and dynamic graph structure of the object subjected to the verification impact load are input into the currently trained graph neural network model to obtain the reconstructed impact load time history of the object subjected to the verification impact load. Based on the reconstructed impact load time history and the corresponding real impact load time history of the object subjected to the impact load for verification, the loss function value of the verification dataset is determined using the mean squared error loss function. Determine if the maximum number of iterations has been reached; If so, then end the training and use the currently trained graph neural network model as the time history reconstruction model; If not, adjust the current learning rate based on the loss function values ​​of the training dataset and the validation dataset, and use the trained graph neural network model as the current graph neural network model, then return "Input the acceleration signal matrix and dynamic graph structure of the training impact load object into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object".

3. The method for reconstructing and locating the time history of impact load according to claim 2, characterized in that, The acceleration signal matrix and dynamic graph structure of the training impact load object are input into the current graph neural network model to obtain the reconstructed impact load time history of the training impact load object, specifically including: Based on the acceleration signal matrix and dynamic graph structure of the training impact load object, the encoding of each node is determined using the graph interaction layer; Based on the encoding of each node, the reconstruction layer is used to determine the reconstruction impact load time history of each node; Based on the reconstructed impact load time history of each node, the reconstructed impact load time history of the training impact load object is determined using the graph fusion layer.

4. The method for reconstructing and locating the time history of impact load according to claim 2, characterized in that, Based on the reconstructed impact load time history and the corresponding real impact load time history of the object subjected to the training impact load, the loss function value of the training dataset is determined using the mean squared error loss function, specifically including: Using formula Determine the loss function value for the training dataset; where, The time history of the reconstructed impact load is used to train the impact load object; F is the actual impact load time history; pr is the Pearson correlation coefficient; the Pearson correlation coefficient is determined based on the encoding distance of the acceleration signal matrix of the training impact load object and the physical distance between the impact position and the acceleration sensor.

5. The method for reconstructing and locating the time history of impact load according to claim 2, characterized in that, Adjusting the current learning rate based on the loss function values ​​of the training dataset and the validation dataset specifically includes: Determine whether the loss function value of the validation dataset is greater than the loss function value of the training dataset by a set multiple; If so, adjust the current learning rate.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the impact load time history reconstruction and localization method according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the impact load time history reconstruction and localization method as described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the impact load time history reconstruction and localization method as described in any one of claims 1-5.

Images