CN111241658A - Beam bridge moving load identification method based on LSTM neural network model - Google Patents

Abstract

The invention discloses a method for identifying a moving load of a beam bridge based on a long-time and short-time memory neural network model. And finally, inputting the bridge response measured by the experiment as test data into the trained LSTM neural network model, outputting the identified moving load speed and size history, and comparing the moving load speed and size history with an actual result. The method overcomes the defect that the moving speed and the load size of the load cannot be identified simultaneously in the traditional moving load identification, realizes multi-parameter high-efficiency and high-precision identification of the moving load of the bridge, and has better precision and applicability when being used for identifying the moving load of the beam bridge.

Description

Identification method of moving load of beam bridge based on LSTM neural network model

technical field

The invention relates to the cross field of bridge moving load identification and computer science, in particular to a beam bridge moving load identification method based on a long short-term memory (LSTM) neural network model.

Background technique

With the rapid development of the bridge field, bridge safety assessment has been paid more and more attention. Many bridges, especially aging bridges, are prone to safety hazards due to increased traffic flow and severe overloading. In order to ensure the safety of bridges and improve the service capacity of bridges, the accurate identification of moving loads has attracted more and more researchers' attention.

At present, the moving vehicle load on the bridge deck is mainly identified by the dynamic response signal of the bridge under the excitation of the vehicle load obtained by the sensor. Most of the moving load identification methods are finally transformed into the solution of linear equations, but due to the ill-posedness of the equations and other reasons, the identification accuracy of each identification method is often insufficient, and the identification effect is not satisfactory.

At present, with the rise of artificial neural field and the continuous deepening of research, many practical problems that are difficult to solve have been successfully solved in the field of civil engineering. Among them, long short-term memory artificial neural network (LSTM), as a time recurrent neural network, is suitable for processing and predicting important events with very long interval and delay in time series. The data information of moving loads happens to have strong temporal correlation, so LSTM neural network has a wide application prospect in the field of moving load identification.

SUMMARY OF THE INVENTION

In view of the deficiencies of the prior art, the present invention provides a method for recognizing moving loads of a beam bridge based on an LSTM neural network model. The LSTM neural network is used to train the load data information. After the training is completed, the neural network can input the bridge response obtained by the input. The signal simultaneously identifies the unknown vehicle moving load speed and size, the identification process is rapid, and the identification accuracy is high.

The object of the present invention is achieved through the following technical solutions:

A method for identifying moving loads of beam bridges based on LSTM neural network model, characterized in that the specific steps of the method for identifying moving loads of beam bridges are as follows:

S1: Build an LSTM multi-layer neural network, which consists of a BLSTM layer, two LSTM layers, and a time-distributed fully connected layer.

S2: Data preparation. By establishing a simplified physical and mechanical model of the bridge, the vibration equation of the beam is solved in the time domain through dynamic response analysis, and a system equation of the form BP=V is obtained, where B is the known coefficient matrix, V is the dynamic response vector of the bridge, P is the load vector of the moving vehicle; according to the system equation, through numerical simulation, the speed, size and dynamic response data of the bridge are obtained for several groups of moving loads, and the sample library is constructed.

(3) Model training. Based on the LSTM multi-layer neural network established in step (1), the multi-layer LSTM neural network model is trained through the sample library.

(4) Model verification. The experimentally measured bridge response is input into the trained LSTM neural network model as a validation set, and the identified moving load speed and size are output.

Further, its input bridge response can be: displacement, strain, and acceleration of the bridge.

Further, the system equation in S2 is constructed by the following method:

S2.1: First establish a simplified physical and mechanical model of the bridge, take the length of the bridge as L, the moving vehicle load p(t) on the bridge deck, move along the bridge deck at a uniform speed v, the linear density of the bridge is ρA, and the flexural stiffness of the bridge is EI, the motion equation of the beam is obtained as shown in equation (1):

where w(x, t) represents the displacement of the beam at time t, at the x position, and δ(t) represents the Dirac function.

S2.2: According to the modal superposition method, w(x, t) is expressed as:

S2.3: Solve η _n (t) in the time domain by the modal superposition method, and obtain the beam displacement w(x, t) as

得到应变，并把应变写成离散形式S2.4: According to the relationship between the displacement and the strain of the beam

get the strain and write the strain in discrete form

S2.5: Obtain the system equation of BP=V from discrete form

BP=V (5)

The beneficial effects of the present invention are:

The invention overcomes the defect that the load moving speed and the load size cannot be simultaneously identified in the traditional moving load identification, and realizes the multi-parameter efficient and high-precision identification of the bridge moving load. The proposed method has better accuracy for the identification of the beam bridge moving load. and applicability. The present invention is expected to promote the development of the data-driven civil engineering structure operation state monitoring technology.

Description of drawings

Fig. 1 is the architecture diagram of multilayer LSTM neural network in concrete step (1) of the present invention;

Fig. 2 is the schematic diagram of the single-layer LSTM model in the concrete step (1) of the present invention;

Fig. 3 is the device diagram of concrete step (4) experimental verification of the present invention;

Fig. 4 is the dynamic response signal diagram of concrete step (4) experimental verification of the present invention;

Fig. 5 is the effect diagram of moving load identification; Fig. 5a is the effect diagram of moving load speed identification, and Fig. 5b is the effect diagram of moving load size identification.

Detailed ways

The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become clearer.

The invention provides a method for recognizing moving loads of a beam bridge based on a long short-term memory (LSTM) neural network model, and the specific steps are as follows:

(1) Build an LSTM multi-layer neural network model, as shown in Figure 1, which consists of a BLSTM layer 3, two LSTM layers 2 and a time-distributed fully connected layer 1. The schematic diagram of an LSTM layer model is shown in the figure 2 shown. In the figure, 1 represents the hyperbolic tangent function, 2 represents the logical sigmoid function, 3 represents the point state operation, and the forward propagation of the LSTM unit can be expressed as

Equation (1) can be simplified as

表示忘记门的输出，

和

分别表示的是输入门和输出门的输出。

是lstm单元在t时刻，ith隐藏层的状态。⊙表示两个向量的点乘。lstm表示lstm单元的正向传播函数。where t and l represent time and the hidden layer, respectively. σ and tanh are the logistic sigmoid activation function and the hyperbolic tangent activation function.

represents the output of the forget gate,

and

represent the output of the input gate and output gate, respectively.

is the state of the ith hidden layer of the lstm unit at time t. ⊙ represents the dot product of two vectors. lstm represents the forward propagation function of the lstm unit.

The BLSTM layer is the most critical part of moving load inversion, it will receive forward and reverse dynamic responses. The use of BLSTM layers takes into account the physics behind the structural dynamics, as the vibrational response is contributed by the long-term excitation force.

In this example, the entire multi-layer LSTM model has 1000 units per layer, and each LSTM unit has 128 hidden neurons.

(2) Data preparation. Firstly, a simplified physical and mechanical model of the bridge is established, taking the length of the bridge as L, the moving vehicle load p(t) on the bridge deck, and moving along the bridge deck at a uniform speed v, the linear density of the bridge is ρA, and the flexural stiffness of the bridge is EI

The equation of motion of the beam is obtained as shown in equation (3):

where w(x, t) represents the displacement of the beam at time t, at the x position, and δ(t) represents the Dirac function.

According to the modal superposition method, w(x,t) can be expressed as:

By solving η _n (t) in the time domain by the modal superposition method, the beam displacement w(x, t) is obtained as

可以得到应变，并写成离散形式。According to the relationship between the displacement and the strain of the beam

Strain can be obtained and written in discrete form.

The system equation of BP=V is obtained from the discrete form

BP=V (7)

According to the system equation, input the speed and size of multiple groups of moving loads, then you can get multiple groups of theoretical bridge dynamic response signal data, and then use the response signal as the input of the LSTM neural network model, and the speed and size of the moving load as the LSTM neural network model. , the data training set can be constructed by numerical simulation.

In this embodiment, in the simplified physical and mechanical model, the length of the girder bridge is 4.5m, ρA=10.512kg/m, EI=14600Nm ² . Based on the above method, 550 groups of dynamic response signals of the bridge under different moving loads are calculated and established. training sample library.

(3) Model training. Based on the neural network model established in step (1), the multi-layer LSTM neural network model is trained through the sample library.

In this embodiment, the mean square error (MSE) is selected as the loss function. The optimizer chooses Adam. The learning rate is 0.005.

In order to verify the accuracy of the identification method of the present invention, the same parameters of the bridge model as in the simulation process of the identification method are used to build a corresponding experimental device, and a trolley is used to simulate the vehicle load acting on the beam. A sensor is used to sense the dynamic response of the bridge, as shown in Figure 3. The 50 groups of dynamic response signals measured in the experiment (as shown in Figure 4) are used as the 50 groups of samples in the verification set, and are input into the trained LSTM neural network model, and the identified moving load speed and size are output, which are compared with the actual ones. The results are compared, as shown in Figure 5. It can be seen from the figure that the moving load speed and size obtained by the identification method of the present invention are very close to the load speed and size of the real trolley. Therefore, it is proved that the identification method of the present invention has high identification accuracy.

Those of ordinary skill in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still understand the Modifications are made to the technical solutions described in the foregoing examples, or equivalent replacements are made to some of the technical features. All modifications and equivalent replacements made within the spirit and principle of the invention shall be included within the protection scope of the invention.

Claims (3)

1. A method for identifying a moving load of a beam bridge based on an LSTM neural network model is characterized by comprising the following specific steps:

s1: and constructing an LSTM multilayer neural network, which consists of a BLSTM layer, two LSTM layers and a time-distributed fully-connected layer.

S2: preparing data: the method comprises the steps of solving a vibration equation of a beam in a time domain through dynamic response analysis by establishing a simplified physical mechanical model of the bridge to obtain a system equation in the form of BP (back propagation) V, wherein B is a known coefficient matrix, V is a dynamic response vector of the bridge, and P is a load vector of a moving vehicle; according to the system equation, through numerical simulation, the speed and the size of a plurality of groups of moving loads and dynamic response data of the bridge are obtained, and a sample library is constructed.

S3: model training: and (3) training a multilayer LSTM neural network model through a sample library based on the LSTM multilayer neural network established in the step (1).

S4: and (3) model verification: and inputting the bridge response measured by the experiment as a verification set into the trained LSTM neural network model, and outputting the recognized moving load speed and magnitude.

2. The LSTM neural network model-based beam bridge moving load identification method of claim 1, which inputs the bridge response: displacement, strain, acceleration of the bridge.

3. The LSTM neural network model-based beam bridge moving load identification method of claim 1, wherein: the system equation in S2 is constructed by the following method:

s2.1: firstly, establishing a simplified physical mechanics model of a bridge, taking the length of the bridge as L, and the load p (t) of a bridge deck moving vehicle, moving along the bridge deck at a uniform speed v, wherein the linear density of the bridge is rho A, the bending rigidity of the bridge is EI, and the motion equation of the obtained beam is shown as the formula (1):

where w (x, t) represents the displacement of the beam at the x position at time t, and δ (t) represents the dirac function.

S2.2: according to modal superposition, w (x, t) is expressed as:

s2.3 solving η in time domain by modal superposition method_n(t) obtaining a beam displacement w (x, t) of

S2.4: according to the relationship of displacement and strain of the beam

Obtaining strain and writing the strain in discrete form

S2.5: system equation for obtaining BP ═ V from discrete form

BP＝V (5)。

Images