CN110631792A - Model update method for seismic hybrid test based on convolutional neural network - Google Patents

Abstract

The invention discloses a method for updating an anti-seismic hybrid test model based on a convolutional neural network, using the system input variables of the test substructure and the observed value of the restoring force as a sample set to train the convolutional neural network online, and then obtain values that are more in line with actual conditions Substructure resilience prediction model, which replaces the resilience model of the same or similar part of the numerical substructure as the test substructure. Therefore, the selection error of the model is avoided, the prediction accuracy of the restoring force is significantly improved, and the result of the mixed test is more in line with the real situation. This method removes the pooling layer in the convolutional neural network, improves computational efficiency, and maintains good data feature extraction capabilities and the ability to resist noise. The prediction accuracy of the restoring force of the numerical substructure in the hybrid test is improved, the generalization ability and the ability to resist noise of the seismic hybrid test model update method based on the intelligent algorithm are significantly improved, and the modeling and analysis results of the numerical substructure in the hybrid test are more accurate. for precision.

Description

Model update method for seismic hybrid test based on convolutional neural network

technical field

The invention relates to a test method for evaluating the anti-seismic performance of a structure in the field of civil engineering, in particular to a method for updating an anti-seismic hybrid test model based on a convolutional neural network.

Background technique

There are mainly three kinds of structural seismic test methods commonly used in the field of civil engineering: pseudo-static test, shaking table test and pseudo-dynamic test. Pseudo-static test is to carry out low-cycle and cyclic loading on the specimen according to a certain load control or displacement control, so that the specimen is elastically stressed until it is destroyed, thereby obtaining the nonlinear constitutive model of the structure or structural components. Pseudo-static test is the most widely used because of its simple and stable technology, but its disadvantage is that it cannot consider the influence of seismic waves on structures. The earthquake simulation shaking table test can truly reproduce the earthquake action, but because of the limited size and bearing capacity of the earthquake simulation shaking table, often only scaled model tests can be used. For the model test of tall structures, because the scale is too small, it cannot reflect all the characteristics of the real structure under the action of earthquake. Pseudodynamic test is a kind of online test, which simulates and reproduces the earthquake process through computer-controlled loading, and calculates the dynamic response loading restoring force and displacement according to the numerical integration algorithm. The advantage is that the restoring force characteristics of the structure in the pseudodynamic test method are no longer derived from the mathematical model, but are directly measured from the test structure, avoiding the numerical error caused by the assumed restoring force model, and can be applied to large-scale model tests , and the gradual failure process of the structure can be observed during the test.

The substructure hybrid test is developed on the basis of the traditional pseudodynamic test method. For some large and complex structures, the substructure technology divides the structure into experimental substructure and numerical substructure, and the parts that are vulnerable to damage or have complex nonlinear restoring force characteristics are used as experimental substructures for physical loading, and the rest are used as numerical substructures. Numerical simulations are carried out in the computer, and the two parts are unified in the equation of motion of the structure. The advantage of substructure technology is that it is beneficial to carry out large-scale engineering structure experiments, and reduces the cost and scale of test equipment.

When performing mixed tests on large complex structures. When the overall structure becomes nonlinear, it is impossible to carry out physical loading tests on all key parts. At this time, some key components or parts must be modeled and analyzed in the numerical substructure. However, there are still large model errors in the current substructure hybrid experiment: on the one hand, the numerical substructure model is too simplified to describe the nonlinear characteristics of the real structure; on the other hand, it is due to the numerical substructure model parameters. Uncertainty, such as using assumed numerical model parameters to describe components that cannot be tested and may enter nonlinearity in large and complex structures, when this numerical model accounts for a large proportion, it will reduce the accuracy of the overall hybrid experiment, making the test The results cannot truly reflect the seismic performance and seismic response of the structure. In order to solve the problem of large errors in the numerical substructure model in the substructure hybrid test, many scholars have begun to study the model update method of the numerical substructure in the seismic hybrid test. The model update is divided into updating the selected initial restoring force model parameters and not selecting the initial restoration. There are two kinds of force models and direct prediction models. The intelligent algorithm can be used without presupposing the restoring force model of the structure, and the restoring force model that is more in line with the actual situation can be obtained by inputting the sample data such as the displacement and restoring force of the test substructure into the intelligent algorithm network for training. Furthermore, it can be used to predict the restoring force of the numerical substructure online, and the trained restoring force model can be updated in real time according to the actual situation, and then the next step of the mixed test loading process is carried out.

Contents of the invention

Purpose of the invention: The purpose of the present invention is to provide a method for updating the seismic hybrid test model based on convolutional neural network, which can improve the accuracy of the numerical substructure restoring force model in the seismic hybrid test, and improve the prediction ability and anti-interference ability of the model for the restoring force .

Technical scheme: in order to achieve this goal, the present invention adopts following technical scheme:

The method for updating the anti-seismic hybrid test model based on the convolutional neural network of the present invention comprises the following steps:

a. Divide the overall structure into experimental substructure and numerical substructure, establish the motion differential equation of the overall structure according to the number of degrees of freedom of the structure and structural parameters, and solve the target displacement d _{E, i} of the experimental substructure in the i-th step of the hybrid test , transfer the obtained target displacement d _{E, i} to the physical loading system, and the actuator pushes the test substructure to reach the target displacement d _{E, i} , and obtains the restoring force R _{E, i} of the test substructure through the sensor in the actuator;

b. Using the total system input variable {d E, ij ,..., d E, i } of the test substructure and the total system input variables {d _{E, ij} ,..., d _{E, i} } of the j step before the i step of the test substructure and the total j+1 steps including the i step Resilience observations { _{RE, ij} , ..., RE _{, i} } are used as the training sample set of the i-th convolutional neural network {(d _{E, i-} j, _{RE, ij} ), ..., (d _{E, i} , R _{E, i} )}; where d _{E, i} represent the system input variable of the i-th test substructure, and R _{E, i} represent the observed value of the recovery force of the i-th test sub-structure;

c. Training samples {(d _{E, ij} , R _{E, ij} ),..., (d _{E, i} , R _{E, i} )} are first processed through the input layer, then enter the convolution layer, and are convoluted by the convolution kernel The features of the data are processed, and then enter the activation layer. The activation layer contains the activation function to help the training samples express complex features. Finally, the processed features enter the fully connected layer, and the prediction model is obtained after training the convolutional neural network.

将混合试验第i步数值子结构的系统输入变量z_i输入预测模型得到第i步数值子结构的恢复力R_N，i，并将R_N，i反馈给数值积分算法；这样就完成了第i步的混合试验，然后循环步骤a-d直到地震动输入完毕。d. Using the prediction model obtained in step c

Input the system input variable z _i of the numerical substructure of the i-th step of the hybrid test into the prediction model to obtain the restoring force R _N,i of the numerical substructure of the i-th step, and feed back R _N,i to the numerical integration algorithm; thus the completion of the first step The mixing test of step i, and then loop steps ad until the input of earthquake motion is completed.

Furthermore, the actuation signal of the loading equipment for loading the test substructure comes from the integral solution of the differential equation of motion for each step of the structure, and the loading method of the test substructure is displacement-controlled loading.

Further, in step b, the convolutional neural network includes an input layer, a convolutional layer, an activation layer and a fully connected layer, and the pooling layer is removed.

Further, in the step c, the processing of the training samples by the input layer is normalization, that is, the data is changed into a range of 0 to 1; the training samples processed by the input layer are then processed by the convolution kernel in the convolution layer , the convolution kernel sweeps the input samples regularly according to a certain step size, and performs point-by-point multiplication of the data points in the receptive field, and then adds and sums the obtained results, and finally obtains the sample characteristics of the training data, and according to the input The parameters of the convolution kernel need to be adjusted according to the number and dimension of the training samples. The obtained sample features then enter the activation layer, and the activation layer performs nonlinear mapping on the output results of the convolution layer; the final processed training samples enter the The fully connected layer trains the neural network; all neurons in the fully connected layer have weight connections, and the fully connected layer is located at the end of the improved convolutional neural network.

Further, in the step c, the activation function in the activation layer is selected as two:

or

处理得到的y_i(z_i)作为恢复力R_N，i。Further, in the step d, the restoring force R _N,i of the numerical substructure is the system input variable z _i of the numerical substructure, and after the prediction model

The obtained y _i ( _zi ) is treated as the restoring force R _N,i .

Furthermore, the neural network has no requirement on the size of the input training samples, can adopt various training sample sizes, and can adjust the weights in the convolution kernel as needed to adapt to different data types.

Beneficial effects: compared with the prior art, the present invention has the following significant advantages:

(1) The convolutional neural network can well meet the needs of the hybrid test model update. Using the trained convolutional neural network model to predict the resilience of the current step of the numerical substructure is more in line with the specific situation, and has a certain influence on the noise of the training samples. Stronger anti-interference ability; this method improves the prediction accuracy of the numerical substructure restoring force in the update of the substructure hybrid test model, and significantly improves the accuracy of the hybrid test;

(2) The neural network can adapt well to changes in the dimension of the input training samples, and can use a variety of training sample sizes, and will not cause sudden changes in the number of neurons in the neural network due to changes in the size and dimension of the input samples. sudden increase; and the weight in the convolution kernel can be adjusted as needed to meet the needs of different data types;

(3) The convolutional neural network involved in the present invention does not need to compress the features extracted by the convolutional layer, so the pooling layer is removed to improve the calculation efficiency and speed, and is more suitable for nonlinear fitting.

Description of drawings

Fig. 1 adopts the model update mixing test flowchart of this method in the specific embodiment of the present invention;

Fig. 2 is the algorithm flowchart of numerical substructure model updating method in the specific embodiment of the present invention;

3 is a schematic diagram of a convolutional neural network in a specific embodiment of the present invention;

Fig. 4 is a schematic structural diagram of a convolutional neural network in a specific embodiment of the present invention;

5 is a schematic diagram of convolution kernel processing training samples in a specific embodiment of the present invention;

Fig. 6 is a schematic diagram of a mixing test using a shaking table of the present invention.

Detailed ways

The technical solution of the present invention will be further introduced below in combination with specific implementation methods and accompanying drawings. Fig. 1 is the model updating hybrid test flow chart adopting this method in the specific embodiment of the present invention, and this specific embodiment discloses a kind of anti-seismic hybrid test model updating method based on convolutional neural network, as shown in Fig. 2, comprises the following steps :

a: According to the number of degrees of freedom and structural parameters of the structure, the differential equation of motion of the overall structure is established. The target displacement d _E,i of the test substructure and the target displacement d _N,i of the numerical substructure in the i-th step of the hybrid test are solved by using the numerical integration algorithm. The obtained target displacement d _{E, i} is converted into an electrical signal, which is transmitted to the controller and converted into the actuating displacement of the hydraulic servo actuator, so as to realize the physical loading on the test substructure. The test substructure is pushed by the actuator to reach the target displacement d _E,i , and the restoring force R _E,i of the test substructure is obtained through the sensor in the actuator.

b: Use the total system input variable {d E, ij ,..., d E, i } of the test substructure and the total system input variables {d _{E, ij} ,..., d _{E, i} } of the test substructure of the j step before the i-th step and the total j+1 steps including the i-th step of the test substructure Resilience observations { _{RE, ij} , ..., RE _{, i} } are used as the training sample set of the i-th convolutional neural network {(d _{E, ij} , _{RE, ij} ), ..., (d _{E, i} , R _{E, i} )}, where d _{E, i} represent the system input variables of the i-th trial substructure. _{RE, i} represents the observed value of the restoring force of the test substructure in the i-th step.

c: Training samples {(d _{E, ij} , _{RE, ij} ),..., (d _{E, i} , _{RE, i} )} are first processed through the input layer, then enter the convolution layer, and are convoluted by the convolution kernel Process the features of the obtained data. Then enter the activation layer, which contains the activation function to help express complex features. Finally, the processed features enter the fully connected layer, and the prediction model is obtained after training

将混合试验第i步数值子结构的系统输入变量z_i输入预测模型得到第i步数值子结构的恢复力R_N，i，并将R_N，i反馈给数值积分算法。这样就完成了第i步的混合试验，然后循环步骤a-d直到地震动输入完毕。d: The prediction model obtained by using step c

Input the system input variable z _i of the numerical substructure of the ith step of the hybrid test into the prediction model to obtain the restoring force R _N,i of the numerical substructure of the ith step, and feed back R _N,i to the numerical integration algorithm. In this way, the mixing test of the i-th step is completed, and then the steps ad are circulated until the input of the earthquake motion is completed.

In step a, the actuation signal of the loading equipment for loading the test substructure comes from the integral solution of the differential equation of motion for each step of the structure, and the loading method of the test substructure is displacement-controlled loading. In step b, the convolutional neural network is based on the standard convolutional neural network. Since the features extracted by the convolutional layer do not need to be compressed, the pooling layer is removed to improve computational efficiency and is more suitable for nonlinear fitting. Happening.

The schematic diagram of the convolutional neural network involved in step c is shown in Figure 3, and the schematic diagram of the specific convolutional neural network structure is shown in Figure 4, which mainly consists of an input layer, a convolutional layer, an activation layer and a fully connected layer. The data obtained from the experiment forms a training sample input into the convolutional neural network to form a matrix. The training samples enter the convolutional layer. The convolution kernel will scan the training samples according to certain rules to form a receptive field. Then the data points in the receptive field will be convoluted, as shown in Figure 5, a', b', c', d', e', f', g', h', i' are in the convolution kernel The weights of , α, β, a', β' are the output training samples after convolution kernel processing.

α=1×a+2×b+3×c+4×δ+5×e+6×f+7×g+8×h+9×i

β=4×a+5×b+6×c+7×d+8×e+9×f+10×g+11×h+12×i

The size of the convolution kernel can be determined according to the actual situation. When the convolution process of the current step is completed, the convolution sum will slide forward with an artificially limited step size, and the next convolution process will be performed. Obtain training samples that expose data features layer by layer, and these data features are integrated, as shown in Figure 5. R _{E, i-1} · Δd _{E, i} is the energy consumption of the i-th step structure, E _i-1 represents the cumulative energy consumption of the i-1 step structure, E _i-1 = E _i-2 +|d _{E, i-1} · R _{E, i-1} |. Then enter the activation layer. Considering the fitting ability of introducing nonlinear functions, the activation functions selected in the activation layer are the following two:

收敛速度快，求梯度简单；(1) RELU function:

The convergence speed is fast, and the gradient is simple;

(2) tanh function:

In the calculation process, the RELU function is preferred, and if the effect is not ideal, the tanh function is selected.

The activation layer can help samples better display nonlinear characteristics through the mapping of activation functions. Subsequently, the training data finally reaches the fully connected layer, and all neurons between each layer of the fully connected layer have weight connections, which are located at the end of the convolutional neural network. Train the convolutional neural network to obtain a more realistic resilience model

为通过既有恢复力模型得出的恢复力，

为通过基于卷积神经网络的模型更新模块预测得到的恢复力。R_E，k为试验子结构恢复力，

为数值积分计算得到的输入给控制器的试验子结构位移命令，F_k为外部激励，此处为地震动作用，一般为

其中m为结构质量矩阵，为地震动加速度，k，(k＝1，2，…)为混合试验中每一步数值积分的步数。The method will be described in detail below in conjunction with specific examples. As shown in FIG. 6 , it is a shaking table mixing test using the convolutional neural network model update method disclosed in the present invention. In the shaking table mixing test, there will be large errors in the modeling and analysis of the numerical substructure due to model errors. Computational models of similar structures or components in the numerical substructure can be updated using test data from the experimental substructure. However, since the data generated during the shaking table hybrid test process contains more noise, to meet the needs of model prediction and noise resistance at the same time, the hybrid test model update method based on the convolutional neural network proposed by the present invention can be used. In Fig. 6, M _i , K _i , and C _i (i=1, 2, 3) are structural parameters, which are respectively mass, stiffness and damping. M _N , C _N are the mass and damping matrices of the numerical substructure, respectively. a _k is acceleration and v _k is velocity.

is the restoring force derived from the existing restoring force model,

is the resilience predicted by the convolutional neural network-based model update module. R _{E, k} is the restoring force of the test substructure,

is the displacement command of the test substructure input to the controller obtained by numerical integration calculation, _Fk is the external excitation, here is the earthquake action, generally

where m is the structural mass matrix, is the ground motion acceleration, k, (k=1, 2, ...) is the number of steps of numerical integration for each step in the hybrid test.

The specific implementation steps are as follows:

1. Extract the part of the overall structure that is difficult to model and analyze, and the analysis results are inaccurate as the test substructure. Treat the rest as numeric substructures. Carry out modeling analysis on the numerical substructure, and conduct real physical loading tests on the test substructure;

2. Establish the motion differential equation of the overall structure. At the beginning of the test, the velocity, displacement and acceleration of the structure are all zero, and the displacement and acceleration of the shaking table are also equal to 0;

并将其传递给计算机，反馈给数值积分算法；3. Assuming k=1, the restoring force _RE,1 of the test substructure =0, and calculate the displacement, velocity and acceleration of the numerical substructure under the joint action of the seismic inertial force and _RE,1 and will signal Measuring the restoring force of the test substructure at the end of the first step as drive commands are passed to the actuator and shaker

And pass it to the computer, feed back to the numerical integration algorithm;

的共同作用下，数值子结构与试验子结构交界处的结构动态响应位移，速度和加速度

同时采集振动台的位移u_k，并将

作为驱动命令传递给作动器。4. Calculate the earthquake motion and the restoring force of the test substructure in step k (k=2, 3, 4...)

The structural dynamic response displacement, velocity and acceleration at the interface between the numerical substructure and the experimental substructure under the joint action of

At the same time, the displacement u _k of the shaking table is collected, and

Passed to the actuator as a drive command.

于试验子结构，同时输入第i步的地震加速度到振动台，由作动器内的传感器测量得到第k步结束时的试验子结构实际的位移

和试验子结构的反力

反馈给计算机，形成训练样本集

5. Actuator application

For the test substructure, input the seismic acceleration of step i to the shaking table at the same time, and measure the actual displacement of the test substructure at the end of step k by the sensor in the actuator

and the reaction force of the test substructure

Feedback to the computer to form a training sample set

首先由输入层进行处理之后，进入卷积层，由卷积核卷积处理得到数据的特征。随后进入激活层，激活层包含激励函数以表达复杂特征，最终经过处理的特征进入全连接层，对卷积神经网络进行训练，最终得到预测模型 6. Training samples

First, after processing by the input layer, it enters the convolution layer, and the convolution kernel is used to obtain the characteristics of the data. Then enter the activation layer, the activation layer contains the activation function to express complex features, and finally the processed features enter the fully connected layer, train the convolutional neural network, and finally get the prediction model

将数值子结构中与试验子结构相似或相同部分的第k步系统输入变量

输入预测模型得到第k步数值子结构中该部分的恢复力预测值

将得到的

反馈给数值积分算法。数值子结构中的其余部分采用既有的恢复力模型，将数值子结构计算得到的位移

输入恢复力模型得到恢复力

进行数值子结构计算。这样就完成了第k步的混合试验，然后循环步骤1-7直到地震动输入完毕，最终就能求解出整体结构的动态响应。7. Utilize the resulting predictive model

Enter the kth step system input variables of the similar or identical part of the numerical substructure and the experimental substructure

Input the prediction model to get the predicted value of restoring force of this part in the k-th step numerical substructure

will get

Feedback to the numerical integration algorithm. The rest of the numerical substructure adopts the existing restoring force model, and the displacement calculated by the numerical substructure

Input the restoring force model to get the restoring force

Perform numerical substructure calculations. In this way, the mixing test of the kth step is completed, and then steps 1-7 are repeated until the input of the earthquake motion is completed, and finally the dynamic response of the overall structure can be solved.

Claims (7)

1. A method for updating an earthquake-resistant hybrid test model based on a convolutional neural network is characterized by comprising the following steps: the method comprises the following steps:

a. dividing the integral structure into a test substructure and a numerical substructure, establishing a motion differential equation of the integral structure according to the number of degrees of freedom and structural parameters of the structure, and solving a target displacement d of the test substructure at the ith step of the hybrid test_E，iThe obtained target displacement d_E，iTransferring to a physical loading system, and pushing the test substructure to reach a target displacement d by an actuator_E，iObtaining the restoring force R of the test substructure by a sensor in the actuator_E，i；

b. Using the total system input variable { d } of the test sub-structure of j steps before the ith step and j +1 steps including the ith_E，i-j，…，d_E，i) And observed values of restoring force { R } of the test substructures_E，i-j，…，R_E，iUsing the training sample set of the ith convolutional neural network as (d)_E，i-j，R_E，i-j)，…，(d_E，i，R_E，i) }; wherein d is_E，iSystem input variable, R, representing the test substructure of step i_E，iRepresenting the observed value of the restoring force of the test substructure in the ith step;

c. training sample { (d)_E，i-j，R_E，i-j)，…，(d_E，i，R_E，i) Firstly, processing the data through an input layer, entering a convolutional layer, obtaining the characteristics of the data through convolutional processing of a convolutional kernel, then entering an activation layer, wherein the activation layer comprises an excitation function for helping a training sample to express complex characteristics, finally, entering a full connection layer through the processed characteristics, and obtaining a prediction model through training a convolutional neural network

d. Using the prediction model obtained in step c

System input variable z of i-th step numerical substructure of mixed test_iInputting a prediction model to obtain the restoring force R of the ith step numerical substructure_N，iAnd R is_N，iFeeding back to a numerical integration algorithm; thus, the mixing test of the step i is completed, and then the steps a-d are circulated until the seismic input is finished.

2. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 1, wherein: the loading device for loading the test substructure has an actuating signal from integral solution of a motion differential equation of each step of the structure, and the loading mode of the test substructure is a displacement control loading mode.

3. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 1, wherein: in step b, the convolutional neural network comprises an input layer, a convolutional layer, an activation layer and a full-link layer, and the pooling layer is removed.

4. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 1, wherein: in the step c, the input layer normalizes the training samples, namely, the data are all changed into the range from 0 to 1; the training samples processed by the input layer are processed by convolution kernels in the convolution layer, the convolution kernels regularly sweep the input samples according to a certain step length, data points in a receptive field are multiplied point by point, obtained results are added and summed, finally, the sample characteristics of the training data are obtained, parameters of the convolution kernels are adjusted according to actual needs including the number and dimensionality of the input training samples, the obtained sample characteristics enter the activation layer, and the activation layer performs nonlinear mapping on the output results of the convolution layer; the finally processed training sample enters a full connection layer to carry out the training of a neural network; all neurons of the full connection layer are connected in a weight mode, and the full connection layer is located at the tail portion of the improved convolutional neural network.

5. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 4, wherein: in step c, the activation functions in the activation layer are selected as two:

or

6. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 1, wherein: in said step d, the restoring force R of the numerical substructure_N，iIs a system input variable z of a numerical substructure_iBy means of a prediction model

Processing the obtained y_i(z_i) As restoring force R_N，i。

7. The convolutional neural network-based seismic hybrid test model updating method as claimed in claim 1, wherein: the neural network has no requirement on the size of an input training sample, can adopt various training sample sizes, can adjust the weight in the convolution kernel according to the requirement, and is suitable for different data types.

Images