A method and apparatus for dynamic load identification using physically guided convolutional long short-term memory networks

Abstract

This invention discloses a dynamic load identification method and apparatus based on a physically guided convolutional long short-term memory network, relating to the field of data processing. The method includes: constructing and training a dynamic load identification model based on a convolutional long short-term memory network to obtain a trained dynamic load identification model; the loss functions used in the training process of the dynamic load identification model include a data-driven loss function and a physically guided loss function. The physically guided loss function is constructed based on the dynamic equations transformed from the motion equations of a linear system under dynamic action, and uses an overall transfer matrix; acquiring the time-domain signals of acceleration and / or strain responses of each test point collected within each time window when the dynamic load to be identified is applied to one or more points on the structure, and inputting these signals into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points. This invention solves the problem of the current load identification method's strong dependence on data.

Description

Technical Field

[0001] This invention relates to the field of data processing, and specifically to a method and apparatus for dynamic load identification using a physically guided convolutional long short-term memory network. Background Technology

[0002] Load information is fundamental to structural analysis and the reliability and safety of engineering structures, and the need to accurately identify external dynamic loads is increasingly prominent. In practical engineering, the dynamic loads acting on structures are often difficult to measure directly; therefore, identifying dynamic loads through monitoring the structural dynamic response is an effective approach. Most load identification methods rely on accurate structural models; however, due to limitations in experimental and modeling technologies, it is difficult to establish accurate models for complex structures, making it challenging to accurately identify dynamic loads using these physical methods. Therefore, how to accurately identify the time history of loads acting on structures based on data-driven approaches is a research topic of significant importance, and developing effective dynamic load identification techniques using only measured partial structural response data has engineering application value.

[0003] Many physical methods for structural load identification based on the time or frequency domain have been developed, including Kalman filtering methods for unknown inputs. However, load identification using physical methods is an inverse problem, which still presents challenges. With the rapid development of machine learning / deep learning technologies, learning the intrinsic correlation between structural response and structural load based on neural networks to avoid the difficulties of inverse identification is an important data-driven method and development trend, and has received increasing attention and research in recent years.

[0004] Recent research has also focused on identifying the time histories of general unknown dynamic loads using deep learning. For example, Altun et al. studied a method for identifying sinusoidal forces on structural members using BP neural networks and radial basis function neural networks based on strain response data. Yang et al. investigated a method for identifying sinusoidal, impulse, and random excitations based on a bidirectional long short-term memory network. Yang et al. further investigated a method for identifying dynamic load time histories based on deep convolutional neural networks. To achieve load magnitude identification, Weng Shun et al. constructed a two-branch convolutional neural network with both classification and regression capabilities, effectively and synchronously identifying the magnitude and location of harmonic and impulse loads using structural acceleration response. Wang Lei et al. proposed a load identification method based on a time-delay neural network, verifying its superior accuracy compared to the general BP neural network load identification method through numerical examples of dynamic load loading and impact load loading—two representative loading conditions. However, in the above methods, the load forms used for identification and testing are the same as those trained on the deep learning network. This requires assuming the type of unknown load is known, resulting in poor network generalization ability and limited application. Therefore, further research is needed on methods for identifying arbitrary unknown loads based on deep learning.

[0005] Furthermore, most deep learning-based load identification research involves the application of single networks, necessitating the integration of spatial and temporal aspects. Some researchers have explored combining CNNs (Convolutional Neural Networks) and RNNs (Recurrent Neural Networks) to capture spatial features of data first, followed by temporal features. However, this layered network model, employing various feature extraction methods, increases model complexity, prolongs processing time, and sometimes even leads to the loss of crucial original features. When CNNs are applied before LSTM (Long Short-Term Memory) networks, the limited kernel range may overlook basic temporal features, potentially impairing the ability to capture long-term temporal dependencies. Convolutional Long Short-Term Memory (ConvLSTM) networks, proposed by Shi et al., can simultaneously capture both temporal and spatial features of data to accurately predict local rainfall intensity. Subsequently, ConvLSTM models have been widely applied in various fields, including weather forecasting and short-term power load forecasting. However, to date, ConvLSTM models have seen limited application in structural health monitoring. For example, Zhang et al. established a ConvLSTM network spatiotemporal learning framework to predict the strain response of structures with strong temporal and spatial dependencies, successfully predicting the strain response of a concrete bridge over three years. Dizaji et al. proposed a deep learning-based hybrid attention-ConvLSTM architecture, which can extract modal frequencies from limited data and transfer knowledge from publicly available datasets to more complex structures through transfer learning to estimate resonant frequencies. Therefore, ConvLSTM can effectively capture both the spatial and temporal dependencies of field sensing measurements, but there is currently no research applying ConvLSTM to identify unknown loads on structures.

[0006] On the other hand, load identification methods driven solely by data require a large number of data samples to train neural networks. Physical guidance can be added to solve this problem. Therefore, it is urgent to propose a physical guidance-based deep learning method to identify load time histories, so as to reduce the current load identification problem's dependence on data.

[0007] In summary, the main problems with existing load time history identification technologies include:

[0008] (1) The types of loads in actual engineering projects are often different from those in training. Existing research has not taken into account the situation where the types of loads in training are different from those in testing. Furthermore, many techniques are limited to the identification of impact loads.

[0009] (2) Most load identification models based on neural networks are applications of a single network to consider time-series effects, which cannot well reflect the influence of multiple loads and spatial characteristics of the structure's response.

[0010] (3) The implementation of existing data-driven deep learning load identification technology still relies on a large number of data samples. Therefore, this technology innovatively uses a physical-guided neural network deep learning convolutional long short-term memory network (ConvLSTM) to simultaneously capture the temporal and spatial features of the data, thereby achieving accurate identification of unknown load time histories based on a small number of data samples, optimizing the model training process, and maintaining a high accuracy in identifying multiple unknown loads. Summary of the Invention

[0011] The purpose of this application is to propose a dynamic load identification method and device based on a physically guided convolutional long short-term memory network to address the aforementioned technical problems.

[0012] In a first aspect, the present invention provides a dynamic load identification method based on a physically guided convolutional long short-term memory network, comprising the following steps:

[0013] A dynamic load identification model based on a convolutional long short-term memory network is constructed and trained to obtain a trained dynamic load identification model. The loss functions used in the training process of the dynamic load identification model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equation transformed from the motion equation of a linear system under dynamic action, and the global transfer matrix is ​​used in the physical guidance loss function.

[0014] The time-domain signals of acceleration and / or strain response of each test point collected in each time window are obtained when the dynamic load to be identified is applied to one or more points on the structure, and input into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

[0015] As a preferred method, the process of constructing the dynamic equations is as follows:

[0016] The equations of motion for a linear system subjected to dynamic forces are as follows:

[0017] ;

[0018] in, These are the mass matrix, damping matrix, and stiffness matrix of the structure, respectively. Let be the acceleration vector, velocity vector, and displacement vector of the linear system at time t. This represents the dynamic load vector of the linear system at time t. For dynamic load vectors The influence matrix;

[0019] Rewrite the equations of motion as state-space formulas:

[0020] ;

[0021] in, Let be the state matrix of the linear system at time t, composed of the velocity vector and displacement vector, and its expression is: , This represents the first derivative of the state matrix with respect to time. The system matrix is ​​expressed as follows: ; The control matrix is ​​expressed as follows: , It is the identity matrix;

[0022] make Then the equations of motion are transformed to obtain the transformation formula:

[0023] ;

[0024] in, This is the measured acceleration and / or strain response vector. The strain observation matrix is ​​expressed as follows: ;

[0025] Divide the continuous time t into several sampling time points t. n Discretize the state-space formula and the transformation equation as shown in the following equation:

[0026] ;

[0027] ;

[0028] in, The discretized state matrix is ​​expressed as follows: , and These are continuous state matrices. At the (n+1)th time sampling point t n+1 and the nth time sampling point t n The value of is determined by the (n+1)th time sampling point t in the linear system. n+1 and the nth time sampling point t n Composed of velocity vector and displacement vector, , The total number of time sampling points. for and The time interval between; The discretized control matrix is ​​expressed as follows: ; For a linear system at the nth time sampling point t n The dynamic load vector; For the nth time sampling point t n Discretized measured acceleration and / or strain response vectors;

[0029] And the integrated formula is obtained by sorting it out:

[0030] ;

[0031] in, For a linear system at the nth time sampling point t n The transfer matrix is ​​expressed as follows: ,and , ; Let be the transfer matrix of the linear system at the 0th time sampling point t0. Discretized state matrix The k-1 powers satisfy the recurrence relation: ; For the linear system at the nth time sampling point t n-k The dynamic load vector;

[0032] Rewrite the integrated formula in matrix form:

[0033] ;

[0034] And further written as:

[0035] ;

[0036] in, The overall transfer matrix of dynamic load and acceleration and / or strain response is expressed as follows: ; The total vector of acceleration and / or strain response measured at multiple points across the entire time domain is expressed as follows: ; The total vector of dynamic loads across the entire time domain and multiple points of application is expressed as follows: .

[0037] As a preferred method, the construction process of the overall transfer matrix is ​​as follows:

[0038] Dynamic loads are applied sequentially to m points on the structure using a hammer impact method. Sensors are placed at p measurement points to record the time-domain signals of the dynamic loads and the time-domain signals of the acceleration and / or strain responses, respectively, as follows: and And construct the training data for the dynamic load identification model, wherein, Let l be the time-domain signal of the dynamic load applied at the l-th point of action at time t, where l = 1, 2, ..., m; Let j be the time-domain signal of acceleration and / or strain response collected at the j-th measurement point when the l-th point of impact is subjected to an impact at time t, where j = 1, 2, ..., p;

[0039] Performing a Fast Fourier Transform (FFT) on the time-domain signal of the dynamic load and the time-domain signals of acceleration and / or strain response converts the time domain to the frequency domain, yielding the frequency-domain signals of the dynamic load and the acceleration and / or strain response, respectively expressed as follows: and ,in, The frequency domain signal of the dynamic load applied at the l-th point of action. The frequency domain signal of the acceleration and / or strain response acquired at the j-th measurement point when a dynamic load is applied at the l-th point of application;

[0040] The frequency response function is calculated based on the frequency domain signal of the dynamic load and the frequency domain signal of the acceleration and / or strain response, as shown in the following equation:

[0041] ;

[0042] in, Let be the frequency response function of the j-th measurement point when a dynamic load is applied to the l-th point of application;

[0043] By performing an inverse fast Fourier transform on the frequency response function, the time-domain signals of the acceleration and / or strain response under a unit pulse load are obtained, as shown in the following equation:

[0044] ;

[0045] in, This represents the inverse fast Fourier transform. This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application;

[0046] Discretize the time-domain signals of acceleration and / or strain response under a unit pulse load over time. , where t1, t2, ... , t N-1 For discrete time steps, This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application; therefore, the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application is:

[0047] ;

[0048] in, For time intervals, This represents the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application;

[0049] The overall transfer matrix consists of the transfer matrices of all m action points and p measurement points. Therefore, the expression for the overall transfer matrix is:

[0050] .

[0051] As a preferred option, the expression for the physical guidance loss function is:

[0052] ;

[0053] in, For physical guidance loss function, The predicted dynamic loads corresponding to one or more points of action on the structure are identified by the dynamic load identification model. It is an L2 norm.

[0054] As a preferred option, the total loss function used during the training process of the dynamic load identification model is:

[0055] ;

[0056] in, For the total loss function, The data-driven loss function has α and β as the first and second weighting coefficients, respectively, and their expressions are as follows:

[0057] ;

[0058] .

[0059] As a preferred embodiment, the dynamic load identification model includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes a first-layer convolutional short-term memory network, a batch normalization layer, and a K-layer convolutional short-term memory network connected in sequence. Each layer of the convolutional short-term memory network contains s convolutional short-term memory units connected in sequence. The K-layer convolutional short-term memory network outputs the predicted dynamic load corresponding to one or more application points.

[0060] Secondly, the present invention provides a dynamic load identification device based on a physically guided convolutional long short-term memory network, comprising:

[0061] The model building module is configured to build and train a dynamic load recognition model based on a convolutional long short-term memory network to obtain a trained dynamic load recognition model. The loss functions used in the training process of the dynamic load recognition model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equations transformed from the motion equations of a linear system subjected to dynamic action, and the overall transfer matrix is ​​used in the physical guidance loss function.

[0062] The prediction module is configured to acquire the time-domain signals of the acceleration and / or strain responses of each test point collected in each time window when the dynamic load to be identified is applied to one or more points on the structure, and input them into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

[0063] Thirdly, the present invention provides an electronic device including one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0064] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0065] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method as described in any of the implementations in the first aspect.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) The physical-guided convolutional long short-term memory network dynamic load identification method mentioned in this invention is based on deep learning and can learn from structural response and load data, learning the mapping relationship between acceleration and / or strain response and dynamic load in the structure. Combined with the learning ability of convolutional long short-term memory network on time series, the load time history of the structure can be identified. Based on the transfer matrix between response and dynamic load, even under external excitation with different load types, it can identify any unknown external dynamic load well.

[0068] (2) The dynamic load identification model mentioned in the physical-guided convolutional long short-term memory network dynamic load identification method mentioned in this invention adopts a convolutional long short-term memory network, which can process spatiotemporal data more effectively and has excellent spatiotemporal feature capture capability. It solves the limitation of previous technologies that cannot capture the temporal and spatial features of structural response and multiple dynamic load data at the same time. Since the proposed method does not require structural physical modeling, it can be applied to the identification of multiple dynamic loads on complex structures.

[0069] (3) The physical-guided convolutional long short-term memory network dynamic load identification method mentioned in this invention introduces the physical relationship between dynamic load and structural response as a physical guide, and combines the errors of dynamic load and structural response to form a data-physical hybrid loss function. The physical-guided convolutional long short-term memory network reduces the dependence on a large amount of data, optimizes the model training process, and thus improves the accuracy of the model in identifying load time histories; it does not require the construction of a structural physical model, but obtains the overall transfer matrix by using the measured structural response as training data through experiments, establishes the physical relationship between excitation and structural response, and effectively realizes the identification of load time histories. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0071] Figure 1 This is a flowchart illustrating the dynamic load identification method using a physically guided convolutional long short-term memory network, as an embodiment of this application.

[0072] Figure 2 This is a schematic diagram of the structure and training process of the dynamic load recognition model of the physical guided convolutional long short-term memory network dynamic load recognition method according to an embodiment of this application;

[0073] Figure 3 This is a schematic diagram of the load application location and sensor arrangement location of the dynamic load identification method using a physically guided convolutional long short-term memory network, which is one embodiment of this application.

[0074] Figure 4 This is a comparison diagram of the white noise F1 value and the true value of the dynamic load identification method of the physically guided convolutional long short-term memory network in one embodiment of this application.

[0075] Figure 5 A comparison diagram of the identified value and the true value of white noise F2 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application;

[0076] Figure 6 This is a comparison diagram of the identified value and the true value of the pulse load F1 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0077] Figure 7This is a comparison diagram of the identified value and the true value of the pulse load F2 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0078] Figure 8 This is a comparison diagram of the identified value and the true value of sinusoidal load F1 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0079] Figure 9 This is a comparison diagram of the identified value and the true value of sinusoidal load F2 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0080] Figure 10 This is a comparison diagram of the white noise load F1 and the true value of the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0081] Figure 11 This is a comparison diagram of the white noise load F2 and the true value of the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0082] Figure 12 This is a comparison diagram of the pulse load F1 of the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application, showing the identification value and the true value of the load under different noise levels.

[0083] Figure 13 This is a comparison diagram of the pulse load F2 and the true value of the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application, under different noise levels.

[0084] Figure 14 This is a comparison diagram of the identified and true values ​​of sinusoidal load F1 under different noise levels in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0085] Figure 15 This is a comparison diagram of the identified and true values ​​of sinusoidal load F2 under different noise levels in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0086] Figure 16 A comparison diagram of the identified value and the true value of stationary random load F1 in the dynamic load identification method of the physically guided convolutional long short-term memory network of one embodiment of this application.

[0087] Figure 17This is a comparison diagram of the identified value and the true value of the stationary random load F2 in the dynamic load identification method of the physically guided convolutional long short-term memory network according to one embodiment of this application.

[0088] Figure 18 This is a schematic diagram of a dynamic load identification device for a physically guided convolutional long short-term memory network, as an embodiment of this application.

[0089] Figure 19 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0091] Figure 1 This application illustrates an embodiment of a dynamic load identification method based on a physically guided convolutional long short-term memory network, comprising the following steps:

[0092] S1. Construct and train a dynamic load identification model based on a convolutional long short-term memory network to obtain a trained dynamic load identification model. The loss functions used in the training process of the dynamic load identification model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equation transformed from the motion equation of a linear system subjected to dynamic action.

[0093] In a specific embodiment, the dynamic load identification model includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes a first-layer convolutional short-term memory network, a batch normalization layer, and a K-layer convolutional short-term memory network connected in sequence. Each layer of the convolutional short-term memory network contains s convolutional short-term memory units connected in sequence. The K-layer convolutional short-term memory network outputs the predicted dynamic load corresponding to one or more application points.

[0094] Specifically, in the embodiments of this application, a dynamic load identification model is first constructed, such as... Figure 2As shown, the dynamic load identification model consists of an input layer, a hidden layer, and an output layer. The input layer receives the time-domain signals of the acceleration and / or strain response of each test point within each time window when a dynamic load is applied to one or more points on the structure. The time-domain signals of the acceleration and / or strain response of each test point within each time window can be obtained by delaying the acquired raw acceleration data or strain response data. The hidden layer contains a K+1-layer convolutional short-term memory (LSTM) network and a batch normalization layer. The time-domain signals of acceleration and / or strain responses collected at each test point within each time window are input to the first-layer LSTM network to obtain its output features. These output features are then input to the batch normalization layer, and so on. Finally, the output features of the (K-1)th layer LSTM network are input to the Kth layer LSTM network to obtain the predicted dynamic loads for one or more application points. Each layer of the Convolutional Long Short-Term Memory (ConvLSTM) network consists of several sequentially connected Convolutional Long Short-Term Memory (ConvLSTM) cells.

[0095] Deep learning technology involves building a deep neural network with multiple hidden layers. It requires a large amount of data to train the network model, allowing it to learn complex and useful information from the data, thereby improving the accuracy of regression prediction or classification. Long Short-Term Memory (LSTM) neural networks are an improved algorithm based on Recurrent Neural Networks (RNNs). LSTM units are introduced into the hidden layers of RNNs to replace the original hidden layer units, effectively solving the gradient vanishing problem of RNNs. The innovation of LSTM neural networks compared to RNNs lies in the addition of input, forget, and output gates to obtain a variable self-recurrent weight. When the model parameters are fixed, the integral scale can be dynamically changed at different times, thus avoiding the shortcomings of RNNs.

[0096] ConvLSTM networks are an improvement on traditional RNNs. Their core essence remains the same as LSTM, using the output of the previous layer as the input of the next. The difference lies in the addition of convolutional operations, which not only allows for the extraction of temporal relationships but also enables the extraction of spatial features, much like convolutional layers.

[0097] ConvLSTM networks are an improvement on traditional Long Short-Term Memory (LSTM) networks, replacing the fully connected layers in traditional LSTMs with convolutional layers to better capture local spatial features and temporal dynamic changes in data with spatiotemporal structure. Its core mechanism retains the gating structure of LSTMs, including forget gates, input gates, cell states, and output gates; however, all transformations between inputs and forget gates, input gates, and cell states are implemented through convolutional operations. ConvLSTM networks combine the advantages of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs), enabling more efficient handling of the mapping relationships between multiple acceleration and / or strain responses and multiple dynamic loads. CNNs excel at capturing spatial features, while LSTMs excel at processing time-series data. By combining these two types of networks, ConvLSTM networks can more effectively process spatiotemporal data, exhibiting superior spatiotemporal feature capture capabilities.

[0098] Based on the concept of physical guidance, this application proposes a ConvLSTM network structure. Within this ConvLSTM network framework, LSTM layers are used to extract temporal correlation information from the response sequences of partial structures, while convolutional layers are used to extract spatial features. This achieves effective encoding of the structural partial response data features, enabling the network to more comprehensively understand the dynamic changes and load patterns of the structure throughout the entire time series, thereby improving the accuracy and stability of dynamic load identification.

[0099] In a specific embodiment, the process of constructing the dynamic equations is as follows:

[0100] The equations of motion for a linear system subjected to dynamic forces are as follows:

[0101] ;

[0102] in, These are the mass matrix, damping matrix, and stiffness matrix of the structure, respectively. Let be the acceleration vector, velocity vector, and displacement vector of the linear system at time t. This represents the dynamic load vector of the linear system at time t. For dynamic load vectors The influence matrix;

[0103] Rewrite the equations of motion as state-space formulas:

[0104] ;

[0105] in, Let be the state matrix of the linear system at time t, composed of the velocity vector and displacement vector, and its expression is: , This represents the first derivative of the state matrix with respect to time. The system matrix is ​​expressed as follows: ; The control matrix is ​​expressed as follows: , It is the identity matrix;

[0106] make Then the equations of motion are transformed to obtain the transformation formula:

[0107] ;

[0108] in, The measured acceleration and / or strain response vectors, The strain observation matrix is ​​expressed as follows: ;

[0109] Divide the continuous time t into several sampling time points t. n Discretize the state-space formula and the transformation equation as shown in the following equation:

[0110] ;

[0111] ;

[0112] in, The discretized state matrix is ​​expressed as follows: , and These are continuous state matrices. At the (n+1)th time sampling point t n+1 and the nth time sampling point t n The value of is determined by the (n+1)th time sampling point t in the linear system. n+1 and the nth time sampling point t n Composed of velocity vector and displacement vector, , The total number of time sampling points. for and The time interval between; The discretized control matrix is ​​expressed as follows: ; For a linear system at the nth time sampling point t n The dynamic load vector; For the nth time sampling point t n Discretized measured acceleration and / or strain response vectors;

[0113] And the integrated formula is obtained by sorting it out:

[0114] ;

[0115] in, For a linear system at the nth time sampling point t n The transfer matrix is ​​expressed as follows: ,and , ; Let be the transfer matrix of the linear system at the 0th time sampling point t0. Discretized state matrix The k-1 powers satisfy the recurrence relation: ; For the linear system at the nth time sampling point t n-k The dynamic load vector;

[0116] Rewrite the integrated formula in matrix form:

[0117] ;

[0118] And further written as:

[0119] ;

[0120] in, The overall transfer matrix of dynamic load and acceleration and / or strain response is expressed as follows: ; The total vector of acceleration and / or strain response measured at multiple points across the entire time domain is expressed as follows: ; The total vector of dynamic loads across the entire time domain and multiple points of application is expressed as follows: .

[0121] Specifically, embodiments of this application introduce the physical relationship between dynamic load and acceleration and / or strain response as a physical guide into the dynamic load identification model. The errors of dynamic load and acceleration and / or strain response are combined to form a data-physics hybrid loss function. This data-physics hybrid loss function serves as the total loss function used during the training of the dynamic load identification model, and it consists of a data-driven loss function and a physical guide loss function. Introducing a physical guide loss function into the total loss function reduces the dependence on large amounts of data, optimizes the training process of the dynamic load identification model, and thus improves the accuracy of the dynamic load identification model in recognizing the time history of dynamic loads.

[0122] First, we discuss the physical mechanism of time history identification techniques for unknown loads. Specifically, this involves: constructing the equations of motion for a dynamically driven linear system, rewriting them as state-space formulas, transforming the equations to obtain transformation formulas, and further discretizing the time in both the state-space and transformation formulas to construct the overall transfer matrix. The overall transfer matrix, also known as the Markov parameter, represents the response of the discrete linear system to a unit impulse. The Markov parameter represents the inherent system response characteristics of a given linear system and is uniquely determined for each linear system. Therefore, the overall transfer matrix can be obtained by constructing a training dataset using measured acceleration and / or strain responses and known dynamic loads, thereby achieving dynamic load time history identification.

[0123] It should be noted that the embodiments mentioned in this application... It is defined using a general response vector derived from the equations of motion of a linear system subjected to dynamic forces. It can describe both acceleration and strain responses. The above reasoning uses acceleration response as an example for derivation; it can also be derived using strain response. The specific derivation process will not be elaborated here. Therefore... It can also be viewed as a continuously measured acceleration and / or strain response vector.

[0124] In a specific embodiment, the process of constructing the overall transfer matrix is ​​as follows:

[0125] Dynamic loads are applied sequentially to m points on the structure using a hammer impact method. Sensors are placed at p measurement points to record the time-domain signals of the dynamic loads and the time-domain signals of the acceleration and / or strain responses, respectively, as follows: and ,in, Let l be the time-domain signal of the dynamic load applied at the l-th point of action at time t, where l = 1, 2, ..., m; Let j be the time-domain signal of acceleration and / or strain response collected at the j-th measurement point when the l-th point of impact is subjected to an impact at time t, where j = 1, 2, ..., p;

[0126] Performing a Fast Fourier Transform (FFT) on the time-domain signal of the dynamic load and the time-domain signals of acceleration and / or strain response converts the time domain to the frequency domain, yielding the frequency-domain signals of the dynamic load and the acceleration and / or strain response, respectively expressed as follows: and ,in, The frequency domain signal of the dynamic load applied at the l-th point of action. The frequency domain signal of the acceleration and / or strain response acquired at the j-th measurement point when a dynamic load is applied at the l-th point of application;

[0127] The frequency response function is calculated based on the frequency domain signal of the dynamic load and the frequency domain signal of the acceleration and / or strain response, as shown in the following equation:

[0128] ;

[0129] in, Let be the frequency response function of the j-th measurement point when a dynamic load is applied to the l-th point of application;

[0130] By performing an inverse fast Fourier transform on the frequency response function, the time-domain signals of the acceleration and / or strain response under a unit pulse load are obtained, as shown in the following equation:

[0131] ;

[0132] in, This represents the inverse fast Fourier transform. This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application;

[0133] Discretize the time-domain signals of acceleration and / or strain response under a unit pulse load over time. , where t1, t2, ... , t N-1 For discrete time steps, This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application; therefore, the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application is:

[0134] ;

[0135] in, For time intervals, This represents the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application;

[0136] The overall transfer matrix consists of the transfer matrices of all m action points and p measurement points. Therefore, the expression for the overall transfer matrix is:

[0137] .

[0138] Specifically, construct the overall transfer matrix. The steps are as follows: A dynamic load is applied experimentally as an excitation, and time-domain signals are collected using sensors to construct training data. Specifically, a hammer is used to impact different locations on the structure, and the time-domain signals collected by the force sensor, acceleration sensor, and / or strain sensor are recorded—that is, the time-domain signals of the dynamic load and the time-domain signals of the acceleration and / or strain response. These time-domain signals are converted to frequency-domain signals using Fast Fourier Transform (FFT), and the frequency response function (FRF) is estimated. Subsequently, the FRF is converted to the time-domain impulse response function (IRF) using IFFT, and the transfer matrix is ​​constructed using the IRF. Here, the impulse response function is also the time-domain signal of the acceleration and / or strain response at the measurement point when a unit impulse load is applied at the point of application. After discretizing the IRF over time, the transfer matrix at the j-th measurement point when the dynamic load is applied at the l-th point of application is constructed by combining the time intervals, and the overall transfer matrix is ​​further constructed. .

[0139] Impact loads are applied sequentially to m load points on the structure using the hammer impact method. Sensors are placed at p measurement points to record the time-domain signals acquired by the sensors: Time-domain signal of dynamic load: Where l = 1, 2, ..., m, corresponding to the dynamic load at the l-th point of application; the time-domain signals of acceleration and / or strain response: Where j = 1, 2, ..., p, corresponding to the acceleration response at the j-th measurement point and the l-th point of impact. A fast Fourier transform is performed on the time-domain signal to convert it to the frequency domain, obtaining the frequency domain signal of the dynamic load. and frequency domain signals of acceleration and / or strain response .

[0140] The frequency response function is defined as the ratio of the frequency domain signal of the acceleration and / or strain response to the frequency domain signal of the dynamic load, i.e., the FRF at the j-th measurement point under excitation at the k-th application point. Performing an inverse fast Fourier transform on the FRF yields the impulse response function (IRF) in the time domain. Discretizing the IRF over time allows us to construct the transfer matrix at the j-th measurement point when the dynamic load is applied at the l-th application point. With m application points and p measurement points, the overall transfer matrix... The complete form of the dimension is .

[0141] Based on the above analysis, embodiments of this application propose a method for identifying multiple unknown load time histories using a physically guided ConvLSTM network to construct a mapping relationship between multiple responses and multiple excitations, thereby identifying multiple unknown load time histories applied to the structure.

[0142] In a specific embodiment, the expression for the physical guidance loss function is:

[0143] ;

[0144] in, For physical guidance loss function, The predicted dynamic loads corresponding to one or more points of action on the structure are identified by the dynamic load identification model. It is an L2 norm.

[0145] In a specific embodiment, the total loss function used by the dynamic load identification model during training is:

[0146] ;

[0147] in, For the total loss function, The data-driven loss function has α and β as the first and second weighting coefficients, respectively, and their expressions are as follows:

[0148] ;

[0149] .

[0150] Specifically, in the embodiments of this application, the time-domain signals of acceleration and / or strain response in the training data are used as input data for the dynamic load identification model, and the corresponding predicted dynamic load is output. During the training process, the dynamic load identification model learns the mapping relationship between partial responses on the structure and dynamic loads.

[0151] The choice of loss function is crucial in training ConvLSTM networks, as it guides the tuning of model parameters to minimize error. A well-chosen loss function not only contributes to convergence and model generalization but also aligns the model's learning objective with the desired goal. In the embodiments of this application, a loss function called the data-physical hybrid loss function is proposed. This data-physical hybrid loss function includes not only a data-driven loss function reflecting data error but also a physical-guided loss function reflecting consistency with the physical mechanism.

[0152] The embodiments of this application introduce an overall transfer matrix. A physical guidance loss function is constructed, and a data-driven loss function is constructed by combining the difference between the predicted dynamic load and the actual collected dynamic load. The physical guidance loss function and the data-driven loss function are weighted to obtain the total loss function used in the training of the dynamic load identification model. The prediction results of the dynamic load identification model and the structural dynamic equilibrium relationship are incorporated into a unified error measurement system. The total loss function includes two parts: data error and physical error. Data error measures the difference between the predicted dynamic load output by the dynamic load identification model and the actual dynamic load; physical error measures the consistency between the predicted dynamic load and the dynamic equation. This physical guidance mechanism improves the interpretability of the model, reduces its dependence on large-scale training data, and effectively reduces identification errors. The final total loss function is composed of both types of errors, and the parameters of the dynamic load identification model are continuously updated through backpropagation to obtain the trained dynamic load identification model.

[0153] S2, acquire the time-domain signals of acceleration and / or strain response of each test point collected in each time window when the dynamic load to be identified is applied to one or more points on the structure, and input them into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

[0154] Specifically, the trained dynamic load identification model is deployed. When the dynamic load to be identified is applied to one or more points on the structure, the time-domain signals of acceleration and / or strain response of each test point are collected in each time window and input into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

[0155] To illustrate the technical effects of the embodiments of this application, the effectiveness of the method proposed in the embodiments of this application is demonstrated through the analysis of numerical simulation of a simply supported beam with 10 degrees of freedom under two loads.

[0156] Take a simply supported beam under multiple concentrated loads as an example. The beam length is 20m, and the Young's modulus of elasticity is... mass density cross-sectional area Moment of inertia The simply supported beam was divided into 10 elements, each 2m in length, and the acceleration response was observed at the lower surface of each node. With a sampling frequency of 200Hz and a duration of 7.5s, the dimension of a single observed acceleration response is... .

[0157] Apply concentrated loads at nodes 4 and 5, such as Figure 3As shown in Table 1, the specific forms of the test load are illustrated. The time-domain signals of partial acceleration and strain responses were used as input to the trained dynamic load identification model to test its learning ability. The number of responses needed to be greater than the number of loads; therefore, the acceleration responses at the lower surfaces of nodes 1, 3, 4, 5, 6, and 8, and the strain responses at the lower surfaces of nodes 4 and 5 were observed.

[0158] Table 1 Specific forms of test loads

[0159] serial number Load type load describe 1 White noise <![CDATA[F1、F2]]> The amplitude distribution follows a Gaussian distribution, and the power spectral density is constant. 2 pulse <![CDATA[F1、F2]]> The amplitude, variance, mean, and direction of action follow a normal distribution. 3 sine <![CDATA[F1、F2]]> <![CDATA[F1=1.6sin(18πt); F2=1.2sin(18πt)]]>

[0160] During the training phase, concentrated loads of white noise type are applied to the structure, and internal random numbers are changed to generate a large number of different white noise loads, thereby generating different structural responses. The time step is set to 10 steps. Since the acceleration and strain response dimension generated by one white noise load is 1×1500, through recursive processing, the acceleration and strain response generated by one white noise load can be processed into 1491 training samples. In the embodiments of this application, only the response data generated by 30 white noise loads are processed, which generates 44730 training samples. During the testing phase, to verify the effectiveness of the method proposed in the embodiments of this application, concentrated loads of white noise, impulse, and sinusoidal types are applied to the structure. The time-domain signals of the structural acceleration and strain response under each type of load are collected as input to the trained dynamic load identification model, and the value of the first predicted dynamic load is used as the dynamic load identification result for the current time step.

[0161] Figure 4-9 The comparison between the identification results and actual values ​​of the test loads is shown. For white noise loads and sinusoidal loads, only the identification results for 2-4 seconds are shown, and for pulse loads, only the identification results near the pulse loading are shown.

[0162] The R² index (also known as the coefficient of determination) is used to assess how well a model fits. It represents the proportion of variability in the dependent variable (target variable) explained by the independent variables (characteristic variables). R² values ​​range from 0 to 1; the closer to 1, the better the model explains the variability in the dependent variable.

[0163] The average R² values ​​for white noise, impulse load, and sinusoidal load are 0.955, 0.845, and 0.993, respectively. The identification error for white noise load is within an acceptable range. Although the identification accuracy for impulse load is lower, it performs well in identifying the amplitude of impulse excitation. This indicates that the trained dynamic load identification model proposed in the embodiments of this application can effectively identify the time history of unknown loads applied to simply supported beams.

[0164] Since the data acquired by the sensor can be affected by observation noise, which can impact the recognition results, in order to discuss whether the method proposed in the embodiments of this application is robust to noise, the embodiments of this application added different levels of Gaussian white noise to the acceleration and strain responses observed in all test sets. Figure 10 - 15 shows the load recognition performance of white noise load, pulse load, and sinusoidal load under 1% and 2% noise influence, respectively, during the test. To more intuitively demonstrate the recognition performance, only the recognition performance within 2-4 seconds is shown for white noise load and sinusoidal load, and only the recognition results near the pulse excitation loading point are shown for pulse load. Table 2 shows the performance evaluation index of the method proposed in the embodiments of this application under different load forms and different levels of noise.

[0165] Table 2. Evaluation of load identification effectiveness under different noise levels (average R²)

[0166] No noise 2% noise White noise load 0.9549 0.9454 Pulse load 0.8445 0.8289 sinusoidal load 0.9926 0.9888

[0167] Depend on Figure 10-15 The load identification diagram and identification effect evaluation index in Table 2 intuitively demonstrate that the method proposed in this application can still identify the time history data of unknown loads well, even under the influence of noise, exhibiting good robustness and noise resistance. The lower index for pulse loads is due to the inclusion of noise in the response at load zero, resulting in a larger error in load identification in this region; however, the accuracy of amplitude identification is high. Based on the principle of the transfer matrix, this application proposes a time history identification technique for multiple unknown dynamic loads using structural partial response and a physically guided convolutional long short-term memory network, thus overcoming the limitation of existing load identification methods that are only applicable to impact loads. Furthermore, this method can effectively identify the time history of unknown external excitations even under external excitations different from the training load type. Therefore, by introducing the physically guided loss into the ConvLSTM network, high-precision load time history identification can be achieved with only a small number of samples. Numerical simulations to identify unknown multi-load time histories verified the effectiveness of the method proposed in this application, providing crucial technical support for the accurate identification of load time histories in structures.

[0168] Further testing was conducted using random loads, and the results of the random load tests are as follows: Figure 16 and Figure 17 As shown. By Figure 16 and Figure 17It can be seen that the trained dynamic load identification model still exhibits good identification ability under unknown random load conditions. The identification curves for stationary random loads F1 and F2 are highly consistent with the actual loads, and the prediction results accurately follow the load fluctuation trend, maintaining small deviations in peak positions and overall amplitude changes, demonstrating the model's stability and generalization ability under random load conditions. The results show that this method can not only identify deterministic loads but also maintains reliable adaptability and high accuracy under random conditions.

[0169] Further reference Figure 18 As an implementation of the methods shown in the above figures, this application provides an embodiment of a dynamic load identification device using a physically guided convolutional long short-term memory network. This device embodiment is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0170] This application provides a dynamic load identification device based on a physically guided convolutional long short-term memory network, including:

[0171] Model building module 1 is configured to build and train a dynamic load recognition model based on a convolutional long short-term memory network to obtain a trained dynamic load recognition model. The loss functions used in the training process of the dynamic load recognition model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equation transformed from the motion equation of a linear system subjected to dynamic action, and the overall transfer matrix is ​​used in the physical guidance loss function.

[0172] Prediction module 2 is configured to acquire the time-domain signals of acceleration and / or strain response of each test point collected in each time window when the dynamic load to be identified is applied to one or more points on the structure, and input them into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

[0173] Figure 19 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 19 As shown, the electronic device in this embodiment includes a processor 1901 and a memory 1902; wherein the memory 1902 is used to store computer execution instructions; and the processor 1901 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0174] Alternatively, the memory 1902 can be either standalone or integrated with the processor 1901.

[0175] When the memory 1902 is set up independently, the electronic device also includes a bus 1903 for connecting the memory 1902 and the processor 1901.

[0176] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by the processor 1901, implement the above method.

[0177] This invention also provides a computer program product, including a computer program that, when executed by a processor 1901, implements the above-described method.

[0178] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0179] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0180] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit formed by the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0181] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor 1901 to execute certain steps of the methods of the various embodiments of this application.

[0182] It should be understood that the processor 1901 described above can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor, or the processor 1901 can be any conventional processor 1901. The steps of the method disclosed in this invention can be directly manifested as execution by the hardware processor 1901, or execution by a combination of hardware and software modules within the processor 1901.

[0183] The memory 1902 may include high-speed RAM memory, and may also include non-volatile memory NVM, such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk or optical disc, etc.

[0184] Bus 1903 can be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 1903 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 1903 in the accompanying drawings of this application is not limited to only one bus 1903 or one type of bus 1903.

[0185] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0186] An exemplary storage medium is coupled to a processor 1901, enabling the processor 1901 to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor 1901. The processor 1901 and the storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor 1901 and the storage medium can exist as discrete components in an electronic device or a host device.

[0187] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic load identification method based on a physically guided convolutional long short-term memory network, characterized in that, Includes the following steps: A dynamic load recognition model based on a convolutional long short-term memory network is constructed and trained to obtain a trained dynamic load recognition model. The loss functions used in the training process of the dynamic load recognition model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equations transformed from the motion equations of a linear system under dynamic action, and uses a global transfer matrix. The expression of the physical guidance loss function is as follows: ； in, Let be the physical guidance loss function. The predicted dynamic loads corresponding to one or more points of action on the structure are output by the dynamic load identification model; It is the total vector of acceleration and / or strain response measured at multiple measurement points across the entire time domain; It is the L2 norm. The overall transfer matrix is ​​constructed as follows: Dynamic loads are applied sequentially to m points on the structure using a hammer impact method. Sensors are placed at p measurement points to record the time-domain signals of the dynamic loads and the time-domain signals of the acceleration and / or strain responses, respectively, as follows: and And construct the training data for the dynamic load identification model, wherein, Let l be the time-domain signal of the dynamic load applied at the l-th point of action at time t, where l = 1, 2, ..., m; Let j be the time-domain signal of acceleration and / or strain response collected at the j-th measurement point when the l-th point of impact is subjected to an impact at time t, where j = 1, 2, ..., p; Performing a Fast Fourier Transform (FFT) on the time-domain signal of the dynamic load and the time-domain signals of acceleration and / or strain response converts the time domain to the frequency domain, yielding the frequency-domain signals of the dynamic load and the acceleration and / or strain response, respectively expressed as follows: and ,in, The frequency domain signal of the dynamic load applied at the l-th point of action. The frequency domain signal of the acceleration and / or strain response acquired at the j-th measurement point when a dynamic load is applied at the l-th point of application; The frequency response function is calculated based on the frequency domain signal of the dynamic load and the frequency domain signal of the acceleration and / or strain response, as shown in the following equation: ； in, Let be the frequency response function of the j-th measurement point when a dynamic load is applied to the l-th point of application; The frequency response function is subjected to an inverse fast Fourier transform to obtain the time-domain signal of the acceleration and / or strain response under a unit pulse load, as shown in the following equation: ； in, This represents the inverse fast Fourier transform. This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application; Discretize the time-domain signals of the acceleration and / or strain response under the applied unit pulse load over time. , where t0, t1, t2, ... , t N-1 For discrete time steps, This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application at each discrete time point. The total number of time sampling points; therefore, when the dynamic load is applied at the l-th point, the transfer matrix of the j-th measurement point is: ； in, For time intervals, This represents the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application; The overall transfer matrix consists of the transfer matrices of all m action points and p measurement points. Therefore, the expression for the overall transfer matrix is: ； The time-domain signals of acceleration and / or strain response of each test point collected in each time window are obtained when the dynamic load to be identified is applied to one or more points on the structure, and input into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

2. The dynamic load identification method based on physically guided convolutional long short-term memory network according to claim 1, characterized in that, The process of constructing the dynamic equations is as follows: The equations of motion for a linear system subjected to dynamic forces are as follows: ； in, These are the mass matrix, damping matrix, and stiffness matrix of the structure, respectively. Let be the acceleration vector, velocity vector, and displacement vector of the linear system at time t. This represents the dynamic load vector of the linear system at time t. For dynamic load vectors The influence matrix; The equations of motion are rewritten as state-space formulas: ； in, The state matrix, composed of the velocity vector and displacement vector of the linear system at time t, is expressed as follows: , This represents the first derivative of the state matrix with respect to time. The system matrix is ​​expressed as follows: ; The control matrix is ​​expressed as follows: , It is the identity matrix; make Then the equation of motion is transformed into the transformation formula: ； in, This is the measured acceleration and / or strain response vector. The strain observation matrix is ​​expressed as follows: ; Divide the continuous time t into several sampling time points t. n Discretize the state-space formula and transformation equation as shown below: ； ； in, The discretized state matrix is ​​expressed as follows: , and These are continuous state matrices. At the (n+1)th time sampling point t n+1 and the nth time sampling point t n The value of is determined by the (n+1)th time sampling point t in the linear system. n+1 and the nth time sampling point t n The velocity vector and displacement vector constitute the vector. , for and The time interval between; The discretized control matrix is ​​expressed as follows: ; For the linear system at the nth time sampling point t n The dynamic load vector; For the nth time sampling point t n Discretized measured acceleration and / or strain response vectors; And the integrated formula is obtained by sorting it out: ； in, For the linear system at the nth time sampling point t n The transfer matrix is ​​expressed as follows: ,and , ; Let be the transfer matrix of the linear system at the 0th time sampling point t0. Discretized state matrix The k-1 powers satisfy the recurrence relation: ; For the linear system at the nth time sampling point t n-k The dynamic load vector; The integrated formula is rewritten in matrix form: ； And further written as: ； in, The overall transfer matrix of dynamic load and acceleration and / or strain response is expressed as follows: ; The expression is: ; The total vector of dynamic loads across the entire time domain and multiple points of application is expressed as follows: .

3. The dynamic load identification method based on physically guided convolutional long short-term memory network according to claim 2, characterized in that, The total loss function used during the training process of the dynamic load identification model is: ； in, For the total loss function, The data-driven loss function has α and β as the first and second weighting coefficients, respectively, and their expressions are as follows: ； 。 4. The dynamic load identification method based on physically guided convolutional long short-term memory network according to claim 1, characterized in that, The dynamic load identification model includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes a first-layer convolutional short-term memory network, a batch normalization layer, and a K-layer convolutional short-term memory network connected in sequence. Each layer of the convolutional short-term memory network contains s convolutional short-term memory units connected in sequence. The K-layer convolutional short-term memory network outputs the predicted dynamic load corresponding to one or more application points.

5. A dynamic load identification device based on a physically guided convolutional long short-term memory network, characterized in that, include: The model building module is configured to construct and train a dynamic load recognition model based on a convolutional long short-term memory network, resulting in a trained dynamic load recognition model. The loss functions used during the training of the dynamic load recognition model include a data-driven loss function and a physical guidance loss function. The physical guidance loss function is constructed based on the dynamic equations transformed from the motion equations of a linear system subjected to dynamic forces, and uses a global transfer matrix. The expression of the physical guidance loss function is as follows: ； in, Let be the physical guidance loss function. The predicted dynamic loads corresponding to one or more points of action on the structure are output by the dynamic load identification model; It is the total vector of acceleration and / or strain response measured at multiple measurement points across the entire time domain; It is the L2 norm. The overall transfer matrix is ​​constructed as follows: Dynamic loads are applied sequentially to m points on the structure using a hammer impact method. Sensors are placed at p measurement points to record the time-domain signals of the dynamic loads and the time-domain signals of the acceleration and / or strain responses, respectively, as follows: and And construct the training data for the dynamic load identification model, wherein, Let l be the time-domain signal of the dynamic load applied at the l-th point of action at time t, where l = 1, 2, ..., m; Let j be the time-domain signal of acceleration and / or strain response collected at the j-th measurement point when the l-th point of impact is subjected to an impact at time t, where j = 1, 2, ..., p; Performing a Fast Fourier Transform (FFT) on the time-domain signal of the dynamic load and the time-domain signals of acceleration and / or strain response converts the time domain to the frequency domain, yielding the frequency-domain signals of the dynamic load and the acceleration and / or strain response, respectively expressed as follows: and ,in, The frequency domain signal of the dynamic load applied at the l-th point of action. The frequency domain signal of the acceleration and / or strain response acquired at the j-th measurement point when a dynamic load is applied at the l-th point of application; The frequency response function is calculated based on the frequency domain signal of the dynamic load and the frequency domain signal of the acceleration and / or strain response, as shown in the following equation: ； in, Let be the frequency response function of the j-th measurement point when a dynamic load is applied to the l-th point of application; The frequency response function is subjected to an inverse fast Fourier transform to obtain the time-domain signal of the acceleration and / or strain response under a unit pulse load, as shown in the following equation: ； in, This represents the inverse fast Fourier transform. This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application; Discretize the time-domain signals of the acceleration and / or strain response under the applied unit pulse load over time. , where t0, t1, t2, ... , t N-1 For discrete time steps, This represents the time-domain signal of the acceleration and / or strain response at the j-th measurement point when a unit pulse load is applied at the l-th point of application at each discrete time point. The total number of time sampling points; therefore, when the dynamic load is applied at the l-th point, the transfer matrix of the j-th measurement point is: ； in, For time intervals, This represents the transfer matrix at the j-th measurement point when a dynamic load is applied at the l-th point of application; The overall transfer matrix consists of the transfer matrices of all m action points and p measurement points. Therefore, the expression for the overall transfer matrix is: ； The prediction module is configured to acquire the time-domain signals of the acceleration and / or strain responses of each test point collected in each time window when the dynamic load to be identified is applied to one or more points on the structure, and input them into the trained dynamic load identification model to obtain the predicted dynamic load corresponding to one or more points.

6. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

Images