Structural impact load identification method and system, medium and equipment
By constructing a structural impact load recognition method based on a learning-based activation function network, the accuracy and interpretability of damage monitoring of aviation structures under impact loads is solved, and efficient and stable impact load recognition and damage assessment are achieved.
Patent Information
- Application Number
- CN202510261686.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-06
- Publication Date
- 2025-07-08
AI Technical Summary
In the prior art, it is difficult to effectively monitor and evaluate damage under the action of impact loads of aviation structures, and the excessive number of neural network models leads to low interpretability, and insufficient recognition accuracy when sensor layout is random.
The structural impact load recognition method based on the learnable activation function network is adopted, and the response signal is collected through the acceleration sensor or strain gauge, and a network model consisting of a flattening layer, a fully connected layer and a learnable activation function layer is constructed. The mean square variance and L2 norm regular terms are used for training, and the strategy is stopped in advance to prevent overfitting, so as to achieve accurate identification of impact loads.
It improves the accuracy and stability of impact load reconstruction, has good generalization performance and noise immunity, the model is simple and easy to visualize, and is suitable for identifying scenarios with uncertainty-point layout of sensors, reducing calculation costs.
Smart Images
Figure CN120277353A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of structural health monitoring, and particularly relates to a method, system, medium and device for identifying structural impact loads based on a learnable activation function network. Background Art
[0002] During the service process of an aircraft structure, it is inevitable to be affected by impact loads, such as impacts caused by bird strikes, falling of maintenance tools, etc. Under the action of impact loads, invisible damages may occur inside the structure, affecting the mechanical properties and reliability of the structure, and even leading to the failure and destruction of the structure. Therefore, monitoring the impact load of the structure and timely detecting and evaluating the impact damage are important means to ensure the safe and reliable operation of the structure. However, directly measuring the impact load received by the structure using force sensors has limitations such as limited measurement positions and high costs.
[0003] The load identification using deep learning methods has the advantages of high accuracy, low computational cost, and insensitivity to the structure. However, most current research on neural networks focuses on fields such as image processing and mechanical fault diagnosis, and there is still a lack of neural networks for force identification; at the same time, due to the excessive number of layers, most neural networks belong to "black box models" and have low interpretability.
[0004] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention
[0005] The present invention provides a method, system, medium and device for identifying structural impact loads based on a learnable activation function network, for monitoring the impact load of the structure and timely detecting and evaluating the impact damage.
[0006] A method for identifying structural impact loads based on a learnable activation function network includes:
[0007] Step S1, collecting impact load force signals and response signals of a monitoring object and making a data set, where the data set includes impact load force signals and response signals;
[0008] Step S2, constructing a structural impact load identification model based on a learnable activation function network;
[0009] Step S3, constructing a learnable activation function layer in the learnable activation function network;
[0010] Step S4, establishing an iterative rule for training the learnable activation function network based on the learnable activation function layer;
[0011] Step S5: Use the data set and the iteration rule to train the structural impact load identification model based on the learnable activation function network until an optimal network model is obtained;
[0012] Step S6: Collect the response signal of the monitoring object under the impact load to be identified, and use the optimal network model to identify the impact load and output the impact load to be identified.
[0013] In the described method for identifying structural impact loads based on a learnable activation function network, in step S1, the response signal sequence of the monitoring object is collected through an acceleration sensor or a strain gauge as the network input feature, and the impact load time history is collected through a force sensor as the sample label.
[0014] In the described method for identifying structural impact loads based on a learnable activation function network, in step S2, the impact load identification model of the learnable activation function network is composed of a flattening layer, a first fully connected layer, a learnable activation function layer, and a second fully connected layer connected in series in sequence, and is expressed as:
[0015]
[0016]
[0017] That is:
[0018] ,
[0019] Where represents the input signal the element in the i-th row and j-th column; represents the element in the -th column in the flattened input vector; represents the output vector, represents the element in the -th column in the output vector; is the activation function matrix of the learnable activation function layer; , are the fully connected weight matrices without activation. The flattening layer converts the multi-channel response signal into a vector convenient for calculation; the matrix reorganizes the time information contained in the response signal and maps it to the activation layer; each activation function in the activation layer is responsible for one time step, parsing and extracting the information; the matrix decodes the information into the amplitude of the impact load at the corresponding historical moment.
[0020] In the described method for identifying structural impact loads based on a learnable activation function network, in step S3, the learnable activation function layer is mathematically expressed as , that is:
[0021]
[0022] ,
[0023] Among them, represents the input vector of the learnable activation function layer; represents the output vector of the learnable activation function layer; is a one-dimensional B-spline interpolation function, representing the activation function matrix The element in the i-th row and j-th column; P is the control vertex parameter of the B-spline curve; is the basis function of the B-spline function.
[0024] In the described method for identifying structural impact loads based on a learnable activation function network, in step S4, the iterative rule for training the learnable activation function network in the i-th iteration is expressed as:
[0025] ,
[0026] ,
[0027] ,
[0028] where is the iteration step size; represents the actual load history, represents the model prediction history; represents the output of the first fully connected layer; represents the output of the learnable activation function layer; represents The element in the i-th row of the main diagonal; is the control vertex of the B-spline curve; is the basis function of the B-spline function.
[0029] In the described method for identifying structural impact loads based on a learnable activation function network, step S5 includes:
[0030] Step S5.1, establishing the mean square error MSE as the loss function to describe the difference between the true impact load and the reconstructed impact load for network training, expressed as:
[0031] ,
[0032] where the weight matrix 、 and the control vertex parameter P of the spline function matrix are the learnable parameters in the network;
[0033] Step S5.2: The network performs network training in an iterative loop based on the iterative rules of the learnable activation function network to calculate the optimal parameters.
[0034] Step S5.3: Set the L2-norm regularization term during the iterative loop to prevent the model performance from overly relying on the parameters of specific nodes due to gradient descent.
[0035] Step S5.4: If the iterative loop calculation takes more than 3600 seconds or the number of iterations exceeds 2000 times, terminate the iteration in advance.
[0036] In steps S5.1 to S5.4, an early stopping strategy is adopted. When the loss function value of the network on the validation set does not decrease for several consecutive iterations, it is regarded as model convergence, and the training process of the optimal network model is terminated to prevent overfitting of the optimal network model. The optimal network model has the best parameters and is used for subsequent impact load identification.
[0037] In the preferred embodiment of the method for identifying structural impact loads based on a learnable activation function network, in step S6, the trained network model is tested for impact load identification using the response of the monitoring object under the impact load. The input of the network is the response signal of the monitoring object structure under the impact load to be identified, and the output is the impact load to be identified.
[0038] The identification system of the method includes:
[0039] The first unit is used to collect and make a data set for the monitoring object; the data set includes impact load force signals and response signals.
[0040] The second unit is used to construct an impact load identification model based on a neural network with a learnable activation function.
[0041] The third unit is used to build a learnable activation function layer.
[0042] The fourth unit is used to construct a learnable activation function network based on the learnable activation function layer.
[0043] The fifth unit is used to establish iterative rules for the training of the learnable activation function network.
[0044] The sixth unit is used to train the neural network using the data set and the iterative rules until an optimal network model is obtained.
[0045] The seventh unit is used to collect the response signal of the monitoring object under the impact load to be identified and identify the impact load using the optimal network model, and output the impact load to be identified.
[0046] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the method described above.
[0047] An electronic device, the electronic device includes:
[0048] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,
[0049] When the processor executes the program, the method is implemented.
[0050] Compared with the prior art, the present invention has the following advantages: Compared with the convolutional neural network algorithm and the long short-term memory neural network algorithm, the present invention can improve the accuracy and stability of impact load reconstruction, and exhibit good generalization performance and noise resistance; it can accurately reconstruct the impact force under the condition of randomly arranged sensor layouts, achieving a more accurate impact load identification effect and reflecting excellent computing speed; the neural network provided by the present invention has a simpler structure compared with other models, with the fewest number of model layers and one-dimensional scale of parameters, and the parameters are flat and easy to visualize; the composite relationship is concise, the information flow is highly readable, and it has better interpretability. Description of the Drawings
[0051] By reading the following detailed description of the preferred specific embodiments, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The accompanying drawings in the specification are only for the purpose of showing the preferred embodiments and are not considered as a limitation of the present invention. Obviously, the following described drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.
[0052] In the drawings:
[0053] Figure 1 Figure (a) is a flowchart of a method for identifying structural impact loads based on a learnable activation function network in an embodiment of the present invention; Figure 1 Figure (b) is the overall structure of a learnable activation function network impact load identification model in an embodiment of the present invention;
[0054] Figure 2 is a schematic diagram of identifying impact loads on a composite material plate structure in an embodiment of the present invention;
[0055] Figure 3 Figure 3 is a schematic diagram of the impact event applied to the composite material plate and the strain response signals of the composite material plate structure at four response measurement points in an embodiment of the present invention; wherein, Figure 3In (a) is the applied impact load, Figure 3 In (b) are the strain response signals of measuring points S1 to S4;
[0056] Figure 4 is a schematic diagram of the impact load identification result of the composite material plate in the embodiment of the present invention; wherein, Figure 4 In (a) is the time history reconstruction result of the impact load learnable activation function network, Figure 4 In (b) is the error distribution of the time history reconstruction of the impact load learnable activation function network;
[0057] Figure 5 is a schematic diagram of the identification result of the impact load at the non-training sample position of the composite material plate in the example of the present invention; wherein, Figure 5 In (a) is the time history reconstruction result of the impact load learnable activation function network at the non-training sample position, Figure 5 In (b) is the error distribution of the time history reconstruction of the impact load learnable activation function network at the non-training sample position;
[0058] Figure 6 is a flowchart of the structural impact load identification method based on the learnable activation function network in an embodiment of the present invention.
[0059] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments. Detailed implementation manners
[0060] The specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be completely conveyed to those skilled in the art.
[0061] It should be noted that certain terms are used in the description and claims to refer to particular components. Those skilled in the art should understand that a technician may use different terms to refer to the same component. The description and claims of this specification do not distinguish components by the difference in terms, but by the difference in functions of the components. As used throughout the description and claims, the terms "comprising" or "including" are open-ended terms and should be interpreted as "including but not limited to". The following description is of the preferred embodiments for implementing the present invention, but the description is for the purpose of the general principles of the specification and is not intended to limit the scope of the present invention. The scope of protection of the present invention shall be defined by the appended claims.
[0062] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not limit the embodiments of the present invention.
[0063] As Figures 1 to 6 shown, the structural impact load identification method based on the learnable activation function network includes the following steps:
[0064] Step S1, collect the impact load force signal and the response signal of the monitoring object and make a data set, where the data set includes the impact load force signal and the response signal;
[0065] Step S2, construct a structural impact load identification model based on the learnable activation function network;
[0066] Step S3, construct the learnable activation function layer in the learnable activation function network;
[0067] Step S4, establish an iterative rule for training the learnable activation function network based on the learnable activation function layer;
[0068] Step S5, use the data set and the iterative rule to train the structural impact load identification model based on the learnable activation function network until an optimal network model is obtained;
[0069] Step S6, collect the response signal of the monitoring object under the impact load to be identified, and use the optimal network model to identify the impact load and output the impact load to be identified.
[0070] In the preferred embodiment of the structural impact load identification method based on the learnable activation function network, in step S1, the response signal sequence of the monitoring object is collected by an acceleration sensor or a strain gauge as the network input feature, and the impact load time history is collected by a force sensor as the sample label.
[0071] In the preferred embodiment of the structural impact load identification method based on the learnable activation function network, in step S2, the impact load identification model of the learnable activation function network is composed of a flattening layer, a first fully connected layer, a learnable activation function layer, and a second fully connected layer connected in series in sequence, expressed as:
[0072]
[0073]
[0074] That is:
[0075] ,
[0076] where represents the element of the input signal in the i-th row and j-th column; represents the element in the -th column of the flattened input vector; represents the output vector, and represents the element in the th column of the output vector; is the activation function matrix of the learnable activation function layer; , is the fully connected weight matrix without activation. The flattening layer converts the multi-channel response signals into vectors for easy calculation; the matrix reorganizes the time information contained in the response signals and maps it to the activation layer; each activation function in the activation layer is responsible for one time step, parsing and extracting the information; the matrix decodes the information into the amplitude of the impact load at the corresponding historical moment.
[0077] In the preferred embodiment of the method for identifying structural impact loads based on a learnable activation function network, in step S3, the learnable activation function layer is mathematically represented as , that is:
[0078]
[0079] ,
[0080] where represents the input vector of the learnable activation function layer; represents the output vector of the learnable activation function layer; is a one-dimensional B-spline interpolation function, representing the element in the th row and th column of the activation function matrix; P is the B-spline curve control vertex parameter;
[0081] In the preferred embodiment of the method for identifying structural impact loads based on a learnable activation function network, in step S4, the iterative rule for training the learnable activation function network in the
[0082] th iteration is expressed as:
[0083] ,
[0084] ,
[0085] where is the iteration step size; represents the actual load history, represents the model prediction history; represents the output of the first fully connected layer; represents the output of the learnable activation function layer; represents The element in the \(i\)-th row of the main diagonal; is the control vertex of the B-spline curve; is the basis function of the B-spline function.
[0086] In the preferred embodiment of the method for identifying structural impact loads based on a learnable activation function network, the step S5 includes:
[0087] Step S5.1, establish the mean squared error (MSE) as the loss function to describe the difference between the true impact load and the reconstructed impact load for network training, expressed as:
[0088] ,
[0089] where the weight matrix , and the spline curve control vertex parameter \(P\) of the spline function matrix are the learnable parameters in the network,
[0090] Step S5.2, the network trains based on the iterative rule of the learnable activation function network and trains the network in an iterative loop manner to calculate the optimal parameters;
[0091] Step S5.3, set the L2-norm regularization term during the iterative loop process to prevent the model performance from overly relying on the parameters of specific nodes due to gradient descent;
[0092] Step S5.4, if the iterative loop calculation takes more than 3600 seconds or the number of iterations is higher than 2000 times, terminate the iteration in advance;
[0093] In the steps S5.1 to S5.4, an early stopping strategy is adopted. When the loss function value of the network on the validation set does not decrease continuously for multiple iterations, it is regarded as model convergence, and the training process of the optimal network model is terminated to prevent overfitting of the optimal network model. The optimal network model has the best parameters and is used for subsequent impact load identification.
[0094] In the preferred embodiment of the method for identifying structural impact loads based on a learnable activation function network, in the step S6, the trained network model is used to perform impact load identification tests using the response of the monitored object under impact loads. The input of the network is the response signal of the monitored object structure under the impact load to be identified, and the output is the impact load to be identified.
[0095] The identification system of the method includes:
[0096] The first unit is used to collect and make a dataset for the monitored object; the dataset includes impact load force signals and response signals;
[0097] The second unit, which is used to construct an impact load identification model based on a neural network with a learnable activation function;
[0098] The third unit, which is used to construct a learnable activation function layer;
[0099] The fourth unit, which is used to construct a learnable activation function network based on the learnable activation function layer;
[0100] The fifth unit, which is used to establish an iterative rule for training the learnable activation function network;
[0101] The sixth unit, which is used to train the neural network using the data set and the iterative rule until an optimal network model is obtained;
[0102] The seventh unit, which is used to collect the response signal of the monitoring object under the impact load to be identified, and use the optimal network model to identify the impact load and output the impact load to be identified.
[0103] A computer storage medium, the storage medium includes computer instructions, which when running on a computer, cause the computer to execute the method described above.
[0104] An electronic device, the electronic device includes:
[0105] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,
[0106] When the processor executes the program, the method described above is implemented.
[0107] In one embodiment, referring to Figure 1 (a) in, the present invention provides a structural impact load identification method and system based on a learnable activation function network. The method collects and creates a database containing impact load force signals and response signals on the monitored object, constructs a learning network based on the learnable activation function of the impact load, and performs model training and model testing, including the following steps:
[0108] S1. Collect and create a data set for the monitored object;
[0109] In step S1, the impact load force signal at the impact monitoring point, the transfer function S between the impact monitoring point and the response measurement point, and the vibration response signal x of the measurement point are obtained by performing a hammering test on the monitored structure, and a data set is constructed.
[0110] S2. Construct a learnable activation function network impact load identification model;
[0111] The model is composed of a flattening layer, a first fully-connected layer, a learnable activation function layer (LANlayer), and a second fully-connected layer connected in series in sequence. As shown in Figure 1 (b) of
[0112]
[0113]
[0114] That is:
[0115] ,
[0116] Where represents the input signal the element in the i-th row and j-th column; represents the element in the -th column in the flattened input vector; represents the output vector, represents the element in the -th column in the output vector; is the activation function matrix of the learnable activation function layer; 、 are the fully-connected weight matrices without activation. The flattening layer converts the multi-channel response signal into a vector convenient for calculation; the matrix reorganizes the time information contained in the response signal and maps it to the activation layer; each activation function in the activation layer is responsible for one time step, parsing and extracting the information; represents the matrix decodes the information into the amplitude of the impact load corresponding to the historical moment.
[0117] S3. Construct the learnable activation function layer in the network;
[0118] The learnable activation function network layer LANlayer is mathematically expressed as That is:
[0119]
[0120]
[0121] Where, represents the input vector of LANlayer; represents the output vector of LANlayer; is a one-dimensional B-spline interpolation function, representing the activation function matrix the element in the i-th row and j-th column; P is the control vertex parameter of the B-spline curve; is the basis function of the B-spline function.
[0122] S4. Establish an iterative rule for training the learnable activation function network based on the learnable activation function layer;
[0123] The weight matrix in the network model described in S2 、 and the spline function matrix The parameter P of is a learnable parameter in the network model. The mean square error is used as the loss function to describe the difference between the true impact load and the reconstructed impact load. The network model performs gradient descent iteration through the backpropagation algorithm, and the i-th iteration is expressed as:
[0124]
[0125]
[0126]
[0127] where is the iteration step size; y represents the actual load history, and ŷ represents the model prediction history; represents the output of the first fully connected layer; represents the output of the LANlayer layer; represents the element in the i-th row of the main diagonal; is the control vertex of the B-spline curve; is the B-spline function basis function.
[0128] S5. Use the dataset and the iterative rule to train the algorithm network until an optimal network model is obtained, including the following steps:
[0129] S5.1: Establish the mean square error MSE as the loss function to describe the difference between the true impact load and the reconstructed impact load for network training, expressed as:
[0130]
[0131] where the weight matrix 、 and the spline curve control vertex parameter P of the spline function matrix are learnable parameters in the network.
[0132] Step S5.2, the network trains the network in an iterative loop based on the iterative rule for training the learnable activation function network to calculate the optimal parameters;
[0133] Step S5.3, set the L2 norm regularization term during the iterative loop to prevent the model performance from overly depending on the parameters of specific nodes;
[0134] Step S5.4, if the time-consuming of the iterative loop calculation is higher than 3600 seconds, or the number of iterations is higher than 2000 times, then terminate the iteration in advance;
[0135] In the above steps S5.1 to S5.4, an early stopping strategy is adopted. When the loss function value of the network on the validation set does not decrease for several consecutive iterations, it is regarded as the model convergence, and the training process of the optimal network model is terminated to prevent the optimal network model from overfitting. The optimal network model has the best parameters and is used for subsequent impact load identification.
[0136] S6. Collect the response signal of the monitoring object under the impact load to be identified, and use the optimal network model to identify the impact load and output the impact load to be identified;
[0137] Use the response of the monitoring object under the impact load to conduct an impact load identification test on the trained network model. The input of the network is the response signal of the monitoring object structure under the impact load to be identified, and the output is the impact load to be identified.
[0138] Compared with the traditional method of solving the impact load identification model through iterative algorithms, the above embodiments are applicable to a wider range of situations and can accurately reconstruct the time history of the impact load when the sensor layout is not fixed (randomly placed). The learning network constructed by the learnable activation function layer has a low depth, significantly reducing the disadvantages of the fully connected layer. At the same time, the fully connected layer considers all inputs, enabling the LANlayer layer to extract all the time information in the input data, reorganize and map it to the activation layer, thereby improving the identification accuracy of the impact force. In addition, the parameters of the LANlayer layer only include two fully connected matrices and a set of spline functions, which are easier to visualize than other network structures. Since this network can have the fewest model layers and one-dimensional scale of parameters while maintaining the model performance, the trained LANlayer model has higher interpretability in network structure design than traditional artificial neural networks for other types of networks.
[0139] In one embodiment, a composite plate model with four sides fixed is used as the research object to identify the impact load. As Figure 2 shown, the degrees of freedom constraints of the four sides of this structure are zero. The size of this composite plate is 500×600×6mm, and the ply situation is . Taking this composite plate model as the research object, the process of identifying its impact load is as follows:
[0140] 1. As shown in Figure 2, establish the coordinate system shown on the composite material plate. At coordinates R1(70, 540), R2(150, 540), R3(350, 540), R4(430, 540), R5(70, 60), R6(150, 60), R7(350, 60), and R8(430, 60), collect the strain response signals of the composite material plate in the X-axis direction. Take the data at R1, R4, R5, and R8 as the effective response signals. As Figure 2 Set 154 impact monitoring points, and apply 10 different impact loads at each impact monitoring point to obtain a total of 1540 training samples as the model training data; as Figure 2 Set 122 impact monitoring points, and apply 10 different impact loads at each impact monitoring point to obtain a total of 1220 training samples as the model generalization ability evaluation data. The sampling frequency when obtaining the samples is 10240 Hz, which satisfies the Nyquist sampling theorem; the sampling time is 25 ms, and the corresponding data length is 256 to collect the complete time history of the impact load; the combined force vector length is 256, and the input combined strain signal vector length is 256×4 = 1024.
[0141] 2. Construct a learnable activation function network impact load identification model:
[0142] The learnable activation function network impact load identification model is composed of a flattening layer, a first fully connected layer, a learnable activation function layer (LANlayer), and a second fully connected layer connected in series in sequence, as shown in Figure 1(b). The mathematical expression is:
[0143]
[0144]
[0145] That is:
[0146] ,
[0147] where represents the input signal the element in the i-th row and j-th column; represents the element in the -th column in the flattened input vector; represents the output vector, represents the element in the -th column in the output vector; is the activation function matrix of the learnable activation function layer; , are the fully connected weight matrices without activation. The flattening layer converts the multi-channel response signals into vectors for easy calculation; the matrix The time information contained in the reconstructed response signal is mapped to the activation layer; each activation function in the activation layer is responsible for one time step, parsing and extracting the information; the representation matrix Decode the information into the amplitude of the impact load at the corresponding historical moment.
[0148] 3. Construct the learnable activation function layer in the network:
[0149] Based on the Kolmogorov-Arnold representation theorem, the learnable activation function network layer LANlayer constructed in the present invention is mathematically expressed as , that is:
[0150]
[0151]
[0152] Among them, represents the input vector of LANlayer; represents the output vector of LANlayer; is a one-dimensional B-spline interpolation function, representing the activation function matrix The element in the i-th row and j-th column is used as the non-linear activation function of the learnable activation function network layer, replacing the constant parameters of the weight matrix of the traditional neural network ; P is the control vertex parameter of the B-spline curve; is the basis function of the B-spline function.
[0153] 4. Establish the iterative rule for training the learnable activation function network.
[0154] Based on the above network, an iterative rule for training the learnable activation function network is established. The weight matrix , and the spline function matrix The parameter P of is the learnable parameter in the network. The mean square error is used as the loss function to describe the difference between the true impact load and the reconstructed impact load. To make the solution reach the minimum parameters , , P, the network performs gradient descent iteration through the backpropagation algorithm, and the i-th iteration is expressed as:
[0155]
[0156]
[0157]
[0158] Among them is the iteration step size; x represents the actual load history, and y represents the model prediction history; represents the output of the fully connected layer 1; represents the output of the LANlayer layer; represents the element in the i-th row of the main diagonal; is the control vertex of the B-spline curve; is the basis of the B-spline function, which is independent of the network parameters. This is the backpropagation iteration algorithm for the learnable activation function.
[0159] 5. Train the learnable activation function network using an existing dataset containing impact load force signals and response signals to obtain an optimal network model, including the following steps:
[0160] S5.1: Establish the mean square error MSE as the loss function to describe the difference between the true impact load and the reconstructed impact load for network training, expressed as:
[0161]
[0162] where the weight matrix , and the spline curve control vertex parameter P of the spline function matrix are the learnable parameters in the network.
[0163] Step S5.2, the network performs network training in an iterative loop manner based on the iterative rule of the learnable activation function network training to calculate the optimal parameters;
[0164] Step S5.3, set the L2 norm regularization term during the iterative loop process to prevent the model performance from being overly dependent on the parameters of specific nodes due to gradient descent;
[0165] Step S5.4, if the iterative loop calculation takes more than 3600 seconds or the number of iterations is higher than 2000 times, abort the iteration in advance;
[0166] In the above steps S5.1 to S5.4, an early stopping strategy is adopted. When the loss function value of the network on the validation set does not decrease continuously for multiple iterations, it is regarded as model convergence, and the training process of the optimal network model is terminated to prevent overfitting of the optimal network model. The optimal network model has the optimal parameters and is used for subsequent impact load identification.
[0167] 6. Identify the impact load through the trained learnable activation function network. The input of the network is the response signal of the monitored object structure under the impact load to be identified, and the output is the impact load to be identified.
[0168] To quantitatively evaluate the performance of the structural impact load identification method based on the learnable activation function network in identifying impact loads, the following relative error and peak relative error of impact load are defined:
[0169]
[0170]
[0171] Among them, and are the true time history of the impact load and the identification result of the impact load respectively, represents the L2 norm of the vector, represents the maximum value element in the vector.
[0172] Figure 3 shows the true time history of the impact load acting on E6 and the strain response signals of the corresponding four response measurement points. Among them, Figure 3 in (a) is the applied impact load, Figure 3 in (b) are the strain response signals of measurement points R1, R4, R5, and R8.
[0173] Figure 4 shows the identification results of the learnable activation function network identification method for the impact load acting on the composite material plate. Among them, Figure 4 in (a) and Figure 4 in (b) are the reconstructed results of the impact time history of the impact load based on the learnable activation function network and the error distribution of the reconstructed impact time history of the impact load based on the learnable activation function network respectively. In the impact samples shown in Figure 4 , the average relative error and average peak relative error of the present invention are 5.81% and 1.90% respectively; the relative error of 98% of the samples is less than 11.47% and the peak relative error is less than 3.75%. The time history of the impact load is identified relatively accurately and stably.
[0174] To characterize the generalization performance of the learnable activation function network, the generalization test data of the impact load at the non-training sample position of the composite material plate is used to evaluate the model performance. Figure 5 shows the identification results of the learnable activation function network identification method for the impact load at the non-training sample position of the composite material plate. Among them, Figure 5 in (a) and Figure 5 in (b) are the reconstructed results of the impact time history of the impact load based on the learnable activation function network and the error distribution of the reconstructed impact time history of the impact load based on the learnable activation function network respectively. In Figure 5In the shown impact samples, the average relative error and average peak relative error of the present invention are 15.80% and 4.95% respectively; the relative error of 98% of the samples is less than 29.28%, and the peak relative error is less than 13.38%. The time history of the impact load is relatively accurately identified, with excellent generalization performance and stability.
[0175] To verify the technical advancement of the learnable activation function network identification method described in the present invention, through systematic comparative experiments, its multi-dimensional performance is evaluated with typical load history reconstruction models (including convolutional neural network algorithms and long short-term memory neural network algorithms).
[0176] Experimental data shows that in the model prediction accuracy test, the relative error and peak relative error of the method of the present invention reach 5.81% and 1.90% respectively, significantly superior to 11.77% and 6.11% of the convolutional neural network algorithm, and 8.81% and 3.09% of the long short-term memory neural network algorithm, showing obvious technical advantages in the accuracy of time-domain feature reconstruction. In terms of computational efficiency, the training of the model of the present invention takes 631 seconds, which has better computing performance than 642 seconds of the convolutional neural network algorithm and 1123 seconds of the long short-term memory neural network algorithm, verifying the computing speed of the algorithm architecture. The test results for the model generalization performance show that the relative error of the method of the present invention under unknown working conditions is 15.80% and the peak relative error is 4.95%, showing obvious technical advantages compared with 20.38% and 8.03% of the convolutional neural network algorithm, and 19.21% and 7.69% of the long short-term memory neural network algorithm, fully demonstrating the generalization performance of the model and its adaptability to complex working conditions. In terms of identification stability, the 98th percentiles of the relative error and peak relative error of the method of the present invention are 29.28% and 13.38% respectively, with a tighter error distribution interval compared with 49.90% and 23.07% of the convolutional neural network algorithm, and 48.46% and 19.69% of the long short-term memory neural network algorithm, showing obvious technical advantages in the stability of time-domain feature reconstruction. In terms of anti-noise performance, in the Gaussian noise interference environment of 25 dB to 10 dB, the average relative error of the method of the present invention is 19.26% and the average peak relative error is 6.56%. Its anti-noise performance is significantly superior to 24.88% and 7.31% of the convolutional neural network algorithm, and 21.21% and 7.14% of the long short-term memory neural network algorithm, maintaining higher identification accuracy in the noise environment.
[0177] The comprehensive experimental data show that, compared with the convolutional neural network algorithm and the long short-term memory neural network algorithm, the present invention can improve the accuracy and stability of impact load reconstruction, and exhibits good generalization performance and noise resistance. The present invention is applicable to identifying structural impact loads and can reconstruct the time history of impact loads. Compared with the convolutional neural network algorithm and the long short-term memory neural network algorithm, the network identification method and system provided by the present invention can obtain solutions with higher accuracy, have a wider range of applicable situations, and exhibit good noise resistance; in terms of structural design, compared with other neural networks, the model has fewer layers and is easier to visualize, the composite relationship between layers is concise, the readability of the information flow is high, and it has better interpretability.
[0178] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all belong to the scope of protection of the present invention.
Claims
1. A method for identifying structural impact loads based on a learnable activation function network, characterized in that, It includes the following steps: Step S1: Collect impact load force signals and response signals of the monitoring object and make a data set, where the data set includes impact load force signals and response signals; Step S2: Construct a structural impact load identification model based on a learnable activation function network; Step S3: Construct a learnable activation function layer in the learnable activation function network; Step S4: Establish an iterative rule for training the learnable activation function network based on the learnable activation function layer; Step S5: Use the data set and the iterative rule to train the structural impact load identification model based on the learnable activation function network until an optimal network model is obtained; Step S6: Collect the response signal of the monitoring object under the impact load to be identified, and use the optimal network model to identify the impact load and output the impact load to be identified.
2. The structural impact load identification method based on a learnable activation function network according to claim 1, wherein Preferably, in step S1, the response signal sequence of the monitoring object is collected through an acceleration sensor or a strain gauge as the network input feature, and the impact load time history is collected through a force sensor as the sample label.
3. A method for identifying structural impact loads based on a learnable activation function network according to claim 1, characterized in that, In step S2, the impact load identification model of the learnable activation function network is composed of a flattening layer, a first fully connected layer, a learnable activation function layer, and a second fully connected layer connected in series in sequence, expressed as: ， ， That is: ， wherein represents the input signal the element at the i-th row and j-th column; represents the element at the j-th column in the flattened input vector; represents the output vector, represents the element at the j-th column in the output vector; is the activation function matrix of the learnable activation function layer; 、 are the fully connected weight matrices without activation. The flattening layer converts the multi-channel response signals into vectors for easy calculation; the matrix of the first fully connected layer reorganizes the time information contained in the response signals and maps them to the activation layer ; each activation function in the activation layer is responsible for one time step, parsing and extracting the information; the matrix of the first fully connected layer decodes the information into the amplitude of the impact load at the corresponding historical moment.
4. A method for identifying structural impact loads based on a learnable activation function network according to claim 1, characterized in that, In the said step S3, the mathematical representation of the learnable activation function layer is , that is: ， ， Among them, represents the input vector of the learnable activation function layer; represents the output vector of the learnable activation function layer; is a one-dimensional B-spline interpolation function, representing the activation function matrix the element in the i-th row and j-th column; P is the B-spline curve control vertex parameter; is the B-spline function basis function.
5. A method for identifying structural impact loads based on a learnable activation function network according to claim 3, characterized in that, In step S4, the i-th iteration of the iterative rule for training the learnable activation function network is expressed as: ， ， ， wherein is the iteration step size; represents the actual load history, represents the model prediction history; represents the output vector of the first fully connected layer; represents the output vector of the learnable activation function layer; represents the element of the i-th row of the main diagonal; is the control vertex of the B-spline curve; is the basis function of the B-spline function.
6. A structural impact load identification method based on a learnable activation function network according to claim 5, characterized in that, Step S5 includes: Step S5.1: Establish the mean square error MSE as the loss function to describe the difference between the real impact load and the reconstructed impact load for network training, expressed as: ， Among them, the weight matrix , and the spline function matrix The spline curve control vertex parameter P of is a learnable parameter in the network, Step S5.2: The network is trained in an iterative loop manner based on the iterative rule for training the learnable activation function network to calculate the optimal parameters; Step S5.3: Set the L2 norm regularization term during the iterative loop process to prevent the model performance from being overly dependent on the parameters of specific nodes due to gradient descent; Step S5.4: If the iterative loop calculation takes more than 3600 seconds or the number of iterations is more than 2000 times, the iteration is terminated in advance; In steps S5.1 to S5.4, an early stopping strategy is adopted. When the loss function value of the network on the validation set does not decrease continuously for multiple iterations, it is regarded as model convergence, and the training process of the optimal network model is terminated to prevent overfitting of the optimal network model. The optimal network model has the optimal parameters for subsequent impact load identification.
7. A method for identifying structural impact loads based on a learnable activation function network according to claim 1, characterized in that, In step S6, the trained network model is used to perform impact load identification testing on the response of the monitoring object under the impact load. The input of the network is the response signal of the monitoring object structure under the impact load to be identified, and the output is the impact load to be identified.
8. An identification system for implementing the method according to any one of claims 1-7, characterized in that, It includes: The first unit is used to collect and make a data set for the monitoring object; the data set includes impact load force signals and response signals; The second unit is used to construct an impact load identification model based on a learnable activation function neural network; The third unit is used to construct a learnable activation function layer; The fourth unit is used to construct a learnable activation function network based on the learnable activation function layer; A fifth unit for establishing an iterative rule for training a learnable activation function network; A sixth unit for training the neural network using the data set and the iterative rule until an optimal network model is obtained; A seventh unit for collecting a response signal of a monitoring object under an impact load to be recognized and identifying the impact load using the optimal network model and outputting the impact load to be recognized.
9. A computer storage medium, characterized in that, The storage medium includes computer instructions which, when running on a computer, cause the computer to execute the method according to any one of claims 1-7.
10. An electronic device, characterized in that, The electronic device includes: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein When the processor executes the program, the method according to any one of claims 1-7 is implemented.