Distributed dynamic load identification method and system based on Transform architecture

Abstract

The invention discloses a distributed dynamic load identification method and a distributed dynamic load identification system based on a Transform framework. The method comprises the following steps: constructing a response-excitation data set in a mode of combining simulation and test, calculating a corresponding load, forming a response matrix # imgabs0 #, performing orthogonal polynomial order selection according to the highest order required by orthogonal polynomial decomposition of a load sample in the data set, and calculating an orthogonal polynomial value matrix L; performing model training on a Transform neural network architecture model according to the response matrix # imgabs1 # and the orthogonal polynomial value matrix L, wherein the model mainly comprises an embedded layer, an encoder and a decoder; and inputting a response signal by using the trained network model to obtain a corresponding distributed load. According to the method, the mapping from the limited response signal to the distributed dynamic load is directly constructed, the problem that a traditional distributed dynamic load method is not qualitative in the inversion process is effectively solved, and the method has good precision in identification of various distributed dynamic loads.

Description

Technical Field

[0001] The present invention belongs to the field of dynamic load identification, and specifically relates to a distributed dynamic load identification method and system based on a Transformer architecture. Background Art

[0002] Structural dynamics has three elements: excitation, system and response. Given the excitation and system, the structural response is a direct problem in dynamics; given the excitation and response, the system is a first-class inverse problem in dynamics; dynamic load identification belongs to the second-class inverse problem in dynamics, which is a "source" problem of solving the excitation given the response and system. In actual engineering, the dynamic loads on the structure are crucial in the design, optimization and health monitoring of the structure. However, due to the limitation of the force position or measurement technology, it is very difficult to determine the load acting on the structure by direct measurement. Relatively speaking, the measurement of the response is much simpler. In response to this engineering demand, dynamic load identification technology came into being. The dynamic load identification method originated in the 1970s and was first used in the field of military aviation. In the following decades, a large number of dynamic load identification methods have been proposed and widely used in various fields.

[0003] Traditional dynamic load identification methods are divided into two categories: frequency domain method and time domain method. The frequency domain method is a method of load identification in the frequency domain through Fourier transform. This method started early and has matured. It is suitable for the identification of long-term or stable loads. The time domain method establishes a mapping relationship between response and load in the time domain to identify the time domain history of load action. It has obvious advantages in short-term situations such as impact loads. The coefficient matrix condition number of the two methods is too large, and it is inevitable to encounter ill-posed problems in the inversion process. At the same time, the frequency domain method is inaccurate in identification near the natural frequency, and the time domain method has problems such as cumulative errors. The data-driven method is a series of new methods represented by machine learning proposed in recent years. Its advantage is that it can directly construct a mapping relationship from response to excitation, avoiding problems such as ill-posedness in the inversion process. The current data-driven and neural network dynamic load identification methods are mostly used for concentrated dynamic load identification, and there are few studies on data-driven methods for distributed load identification. Summary of the invention

[0004] Purpose of the invention: In view of the technical difficulties of neural network distributed dynamic load identification, a distributed dynamic load identification method and system based on Transformer architecture is proposed, which improves the anti-noise ability of dynamic load identification and has a very high stability and accuracy on all types of loads.

[0005] Technical solution: To achieve the above purpose, the present invention adopts the following technical solution:

[0006] A distributed dynamic load identification method based on Transformer architecture includes the following steps:

[0007] Step 1: Construct a response-stimulus data set by combining simulation and experiment to obtain the response matrix

[0008] Step 2: Select the orthogonal polynomial order according to the highest order required for orthogonal polynomial decomposition of load samples in the data set, and calculate the orthogonal polynomial value matrix L based on the Legendre orthogonal polynomial decomposition principle;

[0009] Step 3: First, train the Transformer neural network architecture model based on the simulation data set, and then train the model again based on the test data set to obtain a trained model; the model mainly includes three parts: embedding layer, encoder and decoder. In each round of calculation, the response matrix in the data set Entering the embedding layer, the response information dimension is transformed to be consistent with the model dimension, and the position information is embedded; the output of the embedding layer is then input into the multi-head attention sublayer of the encoder, and the signal is converted into a high-dimensional mapping through the multi-head attention mechanism. The output of the multi-head attention sublayer is input into the feedforward neural network after residual connection and layer normalization to capture local features; the features pass through a decoder with a structure basically the same as that of the encoder, and the output of the decoder is transformed in dimension by a linear layer and multiplied by the orthogonal polynomial value matrix L to obtain the distributed load prediction value;

[0010] Step 4: Use the trained network model to input the response signal and obtain the corresponding distributed load.

[0011] Furthermore, the embedding layer output of the model is expressed as:

[0012]

[0013] in, represents the acceleration response signal of m sensors with a total sampling number of n, O E represents the output of the embedding layer, W d ×m , P d×n Represent the weight matrix and the position encoding matrix respectively, d represents the model dimension, and the encoding method of the position encoding matrix is ​​as follows:

[0014]

[0015] pos represents the position of the element in the column vector, and the encoding of the odd and even positions is generated by the cosine function and the sine function respectively.

[0016] Furthermore, the multi-head attention mechanism transforms the signal into a high-dimensional mapping, expressed as:

[0017]

[0018] In the formula Represents the result of the i-th self-attention calculation. The data structure is a d×n matrix. A represents the final output of the multi-head attention sublayer, heads represents the number of attention heads, and flatten represents spliced ​​together along the feature dimension, Q i , K i 、V i They represent the query vector, key vector, and value vector of the attention mechanism, respectively, and are obtained by multiplying the output of the previous layer by three learnable weight matrices:

[0019]

[0020] Where W Q ,W K ,W V ∈R n×d is the weight matrix available for learning.

[0021] Furthermore, the decoder has an additional masked multi-head attention layer compared to the encoder, and the calculation result of the i-th masked attention is expressed as:

[0022]

[0023] Among them, M masked is a mask matrix, the upper triangular elements of which are negative infinity and the rest are 0;

[0024] The output of the masked attention layer is used as the input of the next attention layer to calculate the V matrix. The input of the Q, K matrix is ​​provided by the output of the encoder. The output expression of the next attention layer of the masked attention layer is:

[0025]

[0026] in:

[0027] Q encoder =W Q O encoder

[0028] K encoder =W K O encoder

[0029]

[0030] O encoder represents the encoder output, represents the output of the masked attention layer, W Q ,W K ,W VSame as in the encoder, is the weight matrix available for learning.

[0031] Furthermore, during the model training process, the model parameters are updated according to the back-propagation method.

[0032] Furthermore, a response-stimulus data set is constructed by combining simulation and experiment, including:

[0033] The simulation data is calculated through the modal superposition method. First, a modal test is performed on the target structure used in the test to obtain the modal parameters. Simulation calculations are performed based on the modal parameters to obtain the response of the target structure under different distributed loads to form a response-excitation data set. The test data is obtained by applying different distributed loads on the target structure and collecting acceleration response signals at multiple points to form a response-excitation data set.

[0034] A distributed dynamic load identification system based on Transformer architecture, comprising:

[0035] The data set construction module is used to construct the response-stimulus data set by combining simulation and experiment to obtain the response matrix

[0036] An orthogonal polynomial decomposition module is used to select the orthogonal polynomial order according to the highest order required for the orthogonal polynomial decomposition of the load samples in the data set, and calculate the orthogonal polynomial value matrix L based on the Legendre orthogonal polynomial decomposition principle;

[0037] The model training module is used to first train the Transformer neural network architecture model based on the simulation data set, and then train the model again based on the test data set to obtain a trained model; the model mainly includes three parts: the embedding layer, the encoder and the decoder. In each round of calculation, the response matrix in the data set Entering the embedding layer, the response information dimension is transformed to be consistent with the model dimension, and the position information is embedded; the output of the embedding layer is then input into the multi-head attention sublayer of the encoder, and the signal is converted into a high-dimensional mapping through the multi-head attention mechanism. The output of the multi-head attention sublayer is input into the feedforward neural network after residual connection and layer normalization to capture local features; the features pass through a decoder with a structure basically the same as that of the encoder, and the output of the decoder is transformed in dimension by a linear layer and multiplied by the orthogonal polynomial value matrix L to obtain the distributed load prediction value;

[0038] The load identification module is used to use the trained network model, input the response signal, and obtain the corresponding distributed load.

[0039] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the distributed dynamic load identification method based on the Transformer architecture as described above are implemented.

[0040] The present invention also provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the distributed dynamic load identification method based on the Transformer architecture as described above are implemented.

[0041] Beneficial effects: The distributed dynamic load identification method based on the Transformer architecture proposed in the present invention realizes the identification of distributed dynamic loads based on Legendre orthogonal polynomials and Transformer neural networks, directly constructs the mapping from finite response signals to distributed dynamic loads, and effectively overcomes the ill-posed problem of traditional distributed dynamic load methods in the inversion process. The algorithm has good accuracy in the identification of sinusoidal, multi-frequency and impact distributed dynamic loads, and has excellent anti-noise ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the overall architecture diagram of the Transformer neural network;

[0043] Figure 2 It is a flow chart of the distributed load identification method based on Transformer architecture;

[0044] Figure 3 It is a simply supported beam model under distributed load;

[0045] Figure 4 This is a schematic diagram of the downward trend of the loss function;

[0046] Figure 5 is the identification result of the quadratic distributed sinusoidal load over time;

[0047] Figure 6 is the identification result of the quadratic distributed sinusoidal load with position;

[0048] Figure 7 It is the identification result of the secondary distribution multi-frequency superposition load over time;

[0049] Figure 8 It is the identification result of the secondary distribution multi-frequency superposition load with position;

[0050] Fig. 9 It is the identification result of the secondary distributed impact load over time;

[0051] Fig.10 It is the identification result of the secondary distributed impact load with position. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0053] The present invention proposes a distributed dynamic load identification method based on Legendre orthogonal polynomial decomposition and Transformer architecture. The distributed dynamic load identification target is converted into orthogonal polynomial coefficients through Legendre orthogonal polynomials, and a response-orthogonal polynomial value matrix pair is constructed. The final output of the network is multiplied by the orthogonal polynomial value matrix and the target load as an error, and back propagation training is performed to obtain a distributed load identification model. The specific technical solution includes:

[0054] 1. Distributed load reconstruction method based on Legendre orthogonal polynomials

[0055]

[0056] In formula (1), w(x N ,t s ) indicates that the Nth measurement point is at t = t s The response value at time, Indicates HT mjk At t = t s The value of HT mjk Calculated from formula (2), J and K represent the highest order of the first and second dimensions of the two-dimensional orthogonal polynomial, respectively, and W m (x N ) represents the value of the mth-order vibration mode function at the Nth measurement point. In formula (1), a jk represents the coefficients of the jk-th order two-dimensional orthogonal polynomial, M m represents the mth order principal mass. In formula (2), h m represents the m-th order impulse response function, L j represents the jth order Legendre orthogonal polynomial, l represents the beam length, and τ is the integral variable in the convolution calculation. The above formula can be used to obtain the mapping relationship from the distributed load to the response. The traditional dynamic load identification method inverts the above formula to obtain the distributed load. The present invention directly constructs the mapping relationship from the response to the orthogonal polynomial coefficient from the data-driven perspective.

[0057] For the load function f(x,t), decomposing it using orthogonal polynomials yields:

[0058]

[0059] L and A in formula (3) are called orthogonal polynomial value matrix and orthogonal polynomial coefficient matrix respectively.

[0060] 2. Transformer neural network architecture construction

[0061] The Transformer architecture is good at processing sequence information and capturing long-distance dependencies through the attention mechanism. Its main structure is divided into three parts: embedding layer, encoder, and decoder. The overall architecture of the Transformer neural network is as follows: Figure 1 As shown. First, the response information dimension is transformed to be consistent with the model dimension through the embedding layer, and the position information is embedded:

[0062]

[0063] In formula (4), Represents the acceleration response signal of m sensors with a total sampling number of n as the model input, O E represents the output of the embedding layer, W d×m , P d×n They represent the weight matrix and the position encoding matrix respectively. For position encoding, pos represents the position of an element in the column vector, d represents the model dimension, sin and cos are used alternately to represent the position, and the encodings of odd and even positions are generated by cosine and sine functions respectively. In this way, different position encoding vectors can be obtained through linear transformation.

[0064]

[0065] The output of the embedding layer is then fed into the first partial multi-head attention sublayer of the encoder, which converts the signal into a high-dimensional mapping through the multi-head attention mechanism:

[0066]

[0067] In formula (6), Represents the calculation result of the i-th self-attention. The data structure is a d×n matrix. heads represents the number of attention heads. flatten represents Spliced ​​together along the feature dimension, the explanation of the three matrices Q, K, and V is shown in formula (7), which is obtained by multiplying the three learnable weight matrices with the output of the previous layer:

[0068]

[0069] W in formula (7) Q ,W K ,W V ∈R n×d is the weight matrix available for learning.

[0070] O ARepresents the final output of the multi-head attention sub-layer, which is obtained by concatenating the total heads attention results into a (d*heads)×n matrix and performing full connection. The principle is to calculate possible attentions multiple times and select features that are more suitable for the model through the linear layer.

[0071] The output of the multi-head attention sublayer will then be input into the feedforward neural network after residual connection and layer normalization. The attention mechanism can extract long-distance dependencies, and the feedforward neural network is used to capture local features. The feedforward neural network consists of two fully connected layers and a nonlinear activation function. The present invention selects the Leaky ReLU activation function. The output mathematical expression of the feedforward layer is:

[0072]

[0073] In formula (8), They represent the weights and biases of the first and second layers of the fully connected neural network of the feedforward neural network. The output of the feedforward neural network will also be subjected to residual connection and layer normalization as the output of a layer of encoder.

[0074] The encoder is composed of the above sub-layers. The output results of multiple encoders stacked together will be input into the decoder as Q, K of the second multi-head attention sub-layer of the decoder. The decoder contains roughly the same content as the encoder, except that the decoder adds a masked multi-head attention layer. During the training process, Transformer does not perform step budgeting like RNN, but calculates in parallel and calculates global timing information at the same time. In order to prevent subsequent timing information from affecting the results of the previous moment, the masked attention mechanism is introduced, which makes the attention mechanism only capture the historical information before and at the current moment during training. The specific implementation is as follows:

[0075]

[0076] Among them, M masked is a mask matrix whose upper triangular elements are negative infinity:

[0077]

[0078] The output of the masked attention layer will be used as the input of the next attention layer to calculate the V matrix. The input of the Q and K matrices is provided by the output of the encoder. Therefore, the output expression of the next attention layer of the masked attention layer is:

[0079]

[0080] in:

[0081]

[0082] In formula (11), O encoder represents the encoder output, represents the output of the masked attention layer, W Q ,W K ,W V Same as in the encoder, is the weight matrix available for learning.

[0083] The output of the decoder is transformed by the linear layer and then multiplied by the orthogonal polynomial value matrix (i.e., L in equation (3)) to obtain the distributed load prediction value.

[0084] 3. Transformer neural network training method

[0085] The premise of neural network training is high-quality labeled data. The present invention constructs a data set by combining simulation and experiment to obtain response-stimulus data with sufficient sample size, and divides the training and validation sets in a ratio of 9:1. The mean square error loss function (MSE) is selected as the loss function, and the Adam optimizer is used for neural network training. Unlike traditional neural network data labeling, the sample data for Transformer neural network training needs to be shifted right, that is, the response at time t corresponds to the stimulus at time t-1.

[0086] The principle of the distributed load identification method of the present invention is based on Legendre orthogonal polynomial fitting, so firstly, it is necessary to select a suitable orthogonal polynomial order for the load characteristics to be identified, and the selected orthogonal polynomial order needs to be greater than or equal to the maximum order required for orthogonal polynomial decomposition of all distributed loads. After selecting a sufficiently large orthogonal polynomial order, the value matrix of the orthogonal polynomial in the identification interval is calculated to determine whether the final output result of the model is accurate.

[0087] According to the above invention concept, referring to Figure 2 The specific implementation steps of the method of the present invention are:

[0088] Step 1: Construct a response-stimulus dataset by combining simulation and experiment.

[0089] The simulation data is obtained by modal superposition method. First, the modal test is performed on the target structure used in the test (a simply supported beam in the embodiment) to obtain the modal parameters. The simulation calculation is performed based on the modal parameters to obtain the response of the target structure under different distributed loads to form a response-excitation data set. The test data is obtained by applying different distributed loads on the target structure and collecting acceleration response signals at multiple points, thereby forming a response-excitation data set.

[0090] Step 2: Select the orthogonal polynomial order according to the highest order required for orthogonal polynomial decomposition of the load samples in the data set, and calculate the orthogonal polynomial value matrix.

[0091] Step 3: Build the Transformer neural network architecture model.

[0092] Step 4: Divide the simulation data set, train the model, calculate the gradient, and update the model parameters according to the back propagation method.

[0093] Step 5: Based on the model trained with simulation data, test sample data is trained and the model is fine-tuned. This allows a distributed dynamic load identification network that conforms to the real physical model to be trained and greatly reduces the demand for test data sample size.

[0094] In the present invention, the simulation data and the test data are trained twice, considering that although the simulation data and the test data are similar, there are errors and the data features are not exactly the same. The amount of simulation data is large, and the training of the simulation data first can more effectively construct the dynamic mapping relationship and avoid the overfitting phenomenon due to the small amount of data. The model trained based on the simulation data can be considered to be similar to the real physical model, but because of the difference between the simulation and test data, it is not exactly the same. Therefore, fine-tuning using test data is a more reasonable and efficient way.

[0095] Step 6: Use the trained network model and input the response signal to obtain the corresponding distributed load.

[0096] The following is a specific example to illustrate the specific implementation method of the present invention and verify the performance of the method. Taking the identification of distributed dynamic load on a simply supported beam as an example, the simply supported beam model is as follows: Figure 3 The parameters of the simply supported beam are shown in Table 1:

[0097] Table 1 Geometric parameters and material properties of simply supported beams

[0098]

[0099] exist Figure 3 A distributed load is applied to the simply supported beam model shown in FIG. 1 , and 10 acceleration sensors are evenly arranged on the beam. The sensor positions are shown in Table 2.

[0100] Table 2 Sensor layout

[0101]

[0102] The specific method is as follows:

[0103] Step 1: First, construct the incentive data set in the following form. The incentive mathematical expression is:

[0104] F(x,t)=f1(x)f2(t) (13)

[0105] in

[0106] f1(x)=ax 2 -bx+c,f2(x)=sin(2πft) (14)

[0107] a, b, c are randomly selected within a certain range, f is the frequency, the value range is [30, 80], the interval is 5Hz, the load under the same frequency is guaranteed to have different spatial distribution, and a total of 50 samples are generated. For multi-frequency and impact loads, the spatial distribution function f1(x) is generated in the same way. For multi-frequency superposition loads, f2(x) = sin(2πp1t) + sin(2πp2t), p1, p2 are randomly selected frequencies in the interval, and for impact loads, the time history function f2(x) is replaced by a half-sine wave, changing the pulse bandwidth and impact amplitude to generate 50 different samples.

[0108] Step 2: Taking the Bernoulli-Euler continuous beam with simple supports on both sides as the analysis object, a one-dimensional beam dynamic model is established. Based on the modal coordinate transformation method, the analytical solution of the vibration response under the distributed dynamic load obtained in step 1 is derived to form a response-excitation data set. In order to enhance the model's anti-noise ability, different degrees of noise of 20dB, 15dB, and 10dB are added to the data, and the training set and validation set are divided in a ratio of 9:1.

[0109] Step 3: Determine the order of Legendre orthogonal polynomial. Here, the load characteristic spatial distribution is quadratic (highest), the highest frequency in the time domain is 80 Hz, and the load of 1 second is identified. The order of the two-dimensional Legendre orthogonal polynomial is 3×160, and the Legendre orthogonal polynomial value matrix (L ij ):

[0110]

[0111] The domain of Legendre orthogonal polynomial is [0,1], so we first need to normalize the domain of the function. x, t represent the original space and time sampling points, max, min represent the maximum and minimum values, x mapped ,t mapped Represents the normalized data, L ij represents the ijth order of the two-dimensional Legendre orthogonal polynomial, L i ,L j They represent the i-th and j-th orders of one-dimensional Legendre orthogonal polynomials respectively.

[0112] Step 4: Build the Transformer neural network model. Here, the model dimension is set to 20, the number of heads is set to 20, the maximum number of iterations is set to 500, the data set is randomly shuffled, and the training is performed in batch training mode with a batch size of 5. The loss function changes with the training generation as follows: Figure 4 As shown. Figure 4 It can be seen that the model converges quickly, and the decline of the training set and the validation set is basically the same, without overfitting.

[0113] Step 5: Use the trained model and input the acceleration response to obtain the corresponding distributed load on the beam. The recognition result is as follows: Figures 5 to 10 shown. Figure 5 , 7 , 9 represents the load identification time history at x = 0.4m, Figure 6 , 8 , 10 represents the load identification spatial distribution result at 0.46s.

[0114] Table 3 Recognition results under different noise levels

[0115]

[0116] It can be seen from Table 3 that, under the noise levels of 20dB, 15dB and 10dB, the signal-to-noise ratios of the sinusoidal distributed load identification results are 30.556dB, 28.785dB and 26.016dB respectively, the signal-to-noise ratios of the multi-frequency superposition load identification results are 29.016dB, 27.860dB and 24.132dB respectively, and the signal-to-noise ratios of the impact load identification results are 29.332dB, 27.659dB and 23.458dB respectively. It can be seen that the method of the present invention has good accuracy in the identification of sinusoidal, multi-frequency and impact distributed dynamic loads, and has excellent anti-noise ability.

[0117] Based on the same technical concept as the method embodiment, the present invention also provides a distributed dynamic load identification system based on the Transformer architecture, including:

[0118] The data set construction module is used to construct the response-stimulus data set by combining simulation and experiment to obtain the response matrix

[0119] An orthogonal polynomial decomposition module is used to select the orthogonal polynomial order according to the highest order required for the orthogonal polynomial decomposition of the load samples in the data set, and calculate the orthogonal polynomial value matrix L based on the Legendre orthogonal polynomial decomposition principle;

[0120] The model training module is used to first train the Transformer neural network architecture model based on the simulation data set, and then train the model again based on the test data set to obtain a trained model; the model mainly includes three parts: the embedding layer, the encoder and the decoder. In each round of calculation, the response matrix in the data set Entering the embedding layer, the response information dimension is transformed to be consistent with the model dimension, and the position information is embedded; the output of the embedding layer is then input into the multi-head attention sublayer of the encoder, and the signal is converted into a high-dimensional mapping through the multi-head attention mechanism. The output of the multi-head attention sublayer is input into the feedforward neural network after residual connection and layer normalization to capture local features; the features pass through a decoder with a structure basically the same as that of the encoder, and the output of the decoder is transformed in dimension by a linear layer and multiplied by the orthogonal polynomial value matrix L to obtain the distributed load prediction value;

[0121] The load identification module is used to use the trained network model, input the response signal, and obtain the corresponding distributed load.

[0122] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the distributed dynamic load identification method based on the Transformer architecture as described above are implemented.

[0123] The present invention also provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the distributed dynamic load identification method based on the Transformer architecture as described above are implemented.

[0124] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices, computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0125] The present invention is described with reference to a flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and a combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the functions specified in the process. Figure 1 The steps of a specified function in a process or multiple processes.

Claims

1. A distributed dynamic load identification method based on Transformer architecture, characterized in that: The following steps are involved: Step 1: Construct a response-stimulus data set by combining simulation and experiment to obtain the response matrix Step 2: Select the orthogonal polynomial order according to the highest order required for orthogonal polynomial decomposition of load samples in the data set, and calculate the orthogonal polynomial value matrix L based on the Legendre orthogonal polynomial decomposition principle; Step 3: First, train the Transformer neural network architecture model based on the simulation data set, and then train the model again based on the test data set to obtain a trained model; the model mainly includes three parts: embedding layer, encoder and decoder. In each round of calculation, the response matrix in the data set Entering the embedding layer, the response information dimension is transformed to be consistent with the model dimension, and the position information is embedded; the output of the embedding layer is then input into the multi-head attention sublayer of the encoder, and the signal is converted into a high-dimensional mapping through the multi-head attention mechanism. The output of the multi-head attention sublayer is input into the feedforward neural network after residual connection and layer normalization to capture local features; the features pass through a decoder with a structure basically the same as that of the encoder, and the output of the decoder is transformed in dimension by a linear layer and multiplied by the orthogonal polynomial value matrix L to obtain the distributed load prediction value; Step 4: Use the trained network model to input the response signal and obtain the corresponding distributed load.

2. The method according to claim 1, characterized in that The embedding layer output of the model is represented as: in, represents the acceleration response signal of m sensors with a total sampling number of n, O E represents the output of the embedding layer, W d×m , P d ×n Represent the weight matrix and the position encoding matrix respectively, d represents the model dimension, and the encoding method of the position encoding matrix is ​​as follows: pos represents the position of the element in the column vector, and the encoding of the odd and even positions is generated by the cosine function and the sine function respectively.

3. The method according to claim 1, characterized in that: The multi-head attention mechanism transforms the signal into a high-dimensional map, expressed as: In the formula Represents the result of the i-th self-attention calculation. The data structure is a d×n matrix. A represents the final output of the multi-head attention sublayer, heads represents the number of attention heads, and flatten represents spliced ​​together along the feature dimension, Q i , K i 、V i They represent the query vector, key vector, and value vector of the attention mechanism, respectively, and are obtained by multiplying the output of the previous layer by three learnable weight matrices: Where W Q ,W K ,W V ∈R n×d is the weight matrix available for learning.

4. The method according to claim 1, characterized in that: The feedforward neural network consists of two fully connected layers and a nonlinear activation function, and the output mathematical expression is: in, Respectively represent the weights and biases of the first layer and the second layer of the fully connected neural network of the feedforward neural network, O A represents the final output of the multi-head attention sub-layer, O E represents the embedding layer output, and Leak ReLU(·) represents the LeakyReLU activation function.

5. The method according to claim 3, characterized in that: The decoder has one more masked multi-head attention layer than the encoder, and the calculation result of the i-th masked attention is expressed as: Among them, M masked is a mask matrix, the upper triangular elements of which are negative infinity and the rest are 0; The output of the masked attention layer is used as the input of the next attention layer to calculate the V matrix. The input of the Q, K matrix is ​​provided by the output of the encoder. The output expression of the next attention layer of the masked attention layer is: in: Q encoder =W Q O encoder K encoder =W K O encoder O encoder represents the encoder output, represents the output of the masked attention layer, W Q ,W K ,W V Same as in the encoder, is the weight matrix available for learning.

6. The method according to claim 1, characterized in that During the model training process, the model parameters are updated according to the back propagation method.

7. The method according to claim 1, characterized in that The response-stimulus dataset is constructed by combining simulation and experiment, including: The simulation data is calculated through the modal superposition method. First, a modal test is performed on the target structure used in the test to obtain the modal parameters. Simulation calculations are performed based on the modal parameters to obtain the response of the target structure under different distributed loads to form a response-excitation data set. The test data is obtained by applying different distributed loads on the target structure and collecting acceleration response signals at multiple points to form a response-excitation data set.

8. A distributed dynamic load identification system based on Transformer architecture, characterized in that: include: The data set construction module is used to construct the response-stimulus data set by combining simulation and experiment to obtain the response matrix An orthogonal polynomial decomposition module is used to select the orthogonal polynomial order according to the highest order required for the orthogonal polynomial decomposition of the load samples in the data set, and calculate the orthogonal polynomial value matrix L based on the Legendre orthogonal polynomial decomposition principle; The model training module is used to first train the Transformer neural network architecture model based on the simulation data set, and then train the model again based on the experimental data set to obtain a trained model; the model mainly includes three parts: the embedding layer, the encoder and the decoder. In each round of calculation, the response matrix in the data set Entering the embedding layer, the response information dimension is transformed to be consistent with the model dimension, and the position information is embedded; the output of the embedding layer is then input into the multi-head attention sublayer of the encoder, and the signal is converted into a high-dimensional mapping through the multi-head attention mechanism. The output of the multi-head attention sublayer is input into the feedforward neural network after residual connection and layer normalization to capture local features; the features pass through a decoder with a structure basically the same as that of the encoder, and the output of the decoder is transformed in dimension by a linear layer and multiplied by the orthogonal polynomial value matrix L to obtain the distributed load prediction value; The load identification module is used to use the trained network model, input the response signal, and obtain the corresponding distributed load.

9. A computer device, characterized in that: include: one or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the distributed dynamic load identification method based on the Transformer architecture as described in any one of claims 1-7 are implemented.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the distributed dynamic load identification method based on the Transformer architecture as described in any one of claims 1 to 7 are implemented.

Images