Damping device optimization method and system based on improved physical information neural network

By using a method of improving physical information neural network, combined with deep learning technology and hard boundary constraints, the shock absorption device parameters of multi-layer steel frame structure are optimized, and the problems of low efficiency, insufficient accuracy and insufficient adaptability in the existing technology are solved, and the problem of low efficiency, insufficient accuracy and insufficient adaptability in the existing technology is realized. Efficient and economical shock absorption effect and improve the building's earthquake resistance.

CN119047035APending Publication Date: 2024-11-29CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411093961.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing method of optimization of parameters of shock absorbers has problems such as low efficiency, insufficient accuracy, high calculation cost and insufficient adaptability, which is difficult to meet the needs of rapid design and construction, especially in the improvement of earthquake resistance in multi-layer steel frame structures.

Method used

Using a method based on improved physical information neural network, a hard boundary constraint physical information neural network model is established through the combination of deep learning technology and physical laws, and the total loss function is optimized by the augmented Lagrangian algorithm to obtain the maximum shock absorption device. Excellent parameters.

Benefits of technology

Improves computing efficiency, reduces R&D costs, and significantly reduces the calculation amount of parameters, ensures high accuracy and reliability of optimization results, and demonstrates excellent generalization ability and adaptability, suitable for complex multi-degree-of-freedom structures. system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047035A_ABST
    Figure CN119047035A_ABST
Patent Text Reader

Abstract

The invention discloses a damping device optimization method and system based on an improved physical information neural network, and the method comprises the steps: building a physical system model of a damping device, and obtaining a motion equation of the physical system model; determining to-be-optimized parameters and a parameter range of the damping device, and establishing an objective function of the physical system model; according to an objective function and constraint conditions of the physical system model, adding hard boundary constraint conditions and then establishing a physical information neural network model based on hard boundary constraint; establishing a total loss function of the physical information neural network model based on the hard boundary constraint according to the motion equation of the physical system model; and optimizing the total loss function by adopting an augmented Lagrange algorithm to obtain an optimal parameter of the damping device when the total loss function is minimum. The obtained parameters play a significant role in response to the multi-layer steel frame, the calculation amount is greatly reduced, and high precision and reliability of an optimization result are ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of structural engineering, and specifically designs a shock absorbing device optimization method and system based on an improved physical information neural network. Background Art

[0002] With the acceleration of urbanization and the increase in high-rise buildings, the safety and stability of building structures have become an important issue in the engineering field. Especially in areas where natural disasters such as earthquakes occur frequently, multi-story steel frame structures are prone to large vibrations under external excitation due to their unique structural characteristics, which not only affects the service life of the building, but also poses a serious threat to the safety of people's lives and property.

[0003] To meet this challenge, engineers and researchers have developed a variety of vibration reduction devices, including but not limited to tuned mass dampers (TMDs), tuned liquid dampers (TLDs), and tuned liquid column dampers (TLCDs). These devices effectively reduce the vibration response of buildings by absorbing and dissipating structural vibration energy. However, the existing parameter optimization methods for vibration reduction devices have many shortcomings, which are mainly manifested in the following aspects:

[0004] Efficiency issue: Traditional parameter optimization methods such as genetic algorithms and particle swarm algorithms usually require a large number of iterative calculations to obtain a satisfactory solution, which is often difficult to meet the needs of rapid design and construction in actual engineering applications;

[0005] Accuracy issues: Due to the lack of in-depth consideration of the physical characteristics of the structure, traditional parameter optimization methods may not accurately capture the true response of the structure during the parameter optimization process, resulting in a deviation between the optimization results and the performance in actual applications;

[0006] Computational cost: When performing numerical simulation and parameter optimization, traditional parameter optimization methods often require a lot of computing resources, which not only increases the cost of research and development, but also limits the widespread application of optimization algorithms in practical engineering.

[0007] Adaptability problem: Traditional parameter optimization methods often lack sufficient adaptability and flexibility when dealing with optimization problems under different types of structures and multiple incentive conditions.

[0008] In summary, there is a need for a parameter optimization method for a multi-layer steel frame structure shock absorbing device that can effectively overcome the above-mentioned deficiencies. Summary of the invention

[0009] The present invention provides a shock absorbing device optimization method and system based on an improved physical information neural network, which can effectively overcome the shortcomings of traditional shock absorbing device parameter optimization methods in terms of efficiency, accuracy, resource consumption and adaptability through the close combination of deep learning technology and physical laws.

[0010] In a first aspect, the present invention proposes a method for optimizing a shock absorbing device based on an improved physical information neural network, comprising the following steps:

[0011] Establish a physical system model of the shock absorbing device and obtain the motion equation of the physical system model;

[0012] Determine the parameters and parameter ranges to be optimized for the shock absorbing device and establish the objective function of the physical system model;

[0013] According to the objective function and constraint conditions of the physical system model, a physical information neural network model based on hard boundary constraints is established after adding hard boundary constraints;

[0014] According to the motion equation of the physical system model, the total loss function of the physical information neural network model based on hard boundary constraints is established;

[0015] The augmented Lagrangian algorithm is used to optimize the total loss function and obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value.

[0016] Furthermore, the parameters to be optimized of the shock absorbing device include one or more of the head loss coefficient, the total length of the liquid in the U-shaped tube, the mass of the U-shaped tube, the stiffness of the shock absorbing device, the damping of the shock absorbing device, and the cross-sectional area of ​​the U-shaped tube.

[0017] Furthermore, the physical information neural network model based on hard boundary constraints includes a first fully connected neural network module, a second fully connected neural network module, a hard boundary constraint module, an automatic differentiation module, a residual network module, and a weight update module. The inputs of the first fully connected neural network module and the second fully connected neural network module are both time variables. The output of the first fully connected neural network module is used as the input of the hard boundary constraint module, and the output of the hard boundary constraint module is used as the input of the automatic differentiation module. The output of the automatic differentiation module and the optimal parameters of the shock absorbing device output by the second fully connected neural network module are input into the residual network module together. The output of the residual network module is back-propagated to the first fully connected neural network module and the second fully connected neural network module after passing through the weight update module.

[0018] Furthermore, the first fully connected neural network module and the second fully connected neural network module each include an input layer, a plurality of hidden layers and a plurality of output layers; the input layer, the hidden layer and the output layer are connected in sequence; the input variable of the input layer is a time variable, and the output data of the output layer are respectively the displacement velocity and the optimal parameters of the shock absorbing device under the time variable.

[0019] Furthermore, the constraints of the physical information neural network model based on hard boundary constraints include boundary condition constraints and inequality constraints.

[0020] Furthermore, the boundary condition is at least one of a Dirichlet boundary condition, a Neumann boundary condition, and a Robin boundary condition.

[0021] Furthermore, the specific steps of optimizing the total loss function by using the augmented Lagrangian algorithm to obtain the optimal parameters of the shock absorbing device when the total loss function takes the minimum value are as follows:

[0022] Select a deep network neural model and give a penalty coefficient and and the penalty factor β F and β h The initial value and threshold of

[0023] Input the initial values ​​of the motion equations of the physical system model of the shock absorber;

[0024] Set the initial number of iterations to 0, set the maximum number of iterations K and the threshold to be reached;

[0025] Send the data into the deep network neural model to start the initial training until the loss function of the deep network neural model converges;

[0026] If the accuracy requirement is not met or the number of iterations is less than the maximum number of iterations, set the number of iterations k = k + 1, and back propagate to update the weights and biases of the deep network neural model;

[0027] The penalty coefficient and augmented Lagrangian operator are updated, and the learning rate and optimizer of the deep network neural model are adjusted and the training is continued to find the minimum value of the total loss function corresponding to the optimal parameters of the shock absorber.

[0028] Furthermore, the expression of the total loss function is:

[0029]

[0030] Among them, J is the objective function, Loss1 is the partial differential equation loss, and loss2 is the inequality loss.

[0031] Furthermore, the objective function is expressed as:

[0032]

[0033] in, They are the displacement of the main structure of the building predicted by the deep network neural model and the average displacement of the main structure of the building; The deep network neural model predicts the displacement of the legal person's shock-absorbing device and the average displacement of the shock-absorbing device respectively; They are respectively the liquid level difference of the shock absorber and the average liquid level difference of the shock absorber predicted by the deep network neural model.

[0034] In a second aspect, the present invention proposes a shock absorbing device optimization system based on an improved physical information neural network, comprising:

[0035] A physical system model building module is used to build a physical system model of the shock absorbing device and obtain a motion equation of the physical system model;

[0036] An objective function building module is used to determine the parameters and parameter ranges to be optimized for the shock absorbing device and to establish the objective function of the physical system model;

[0037] A physical information neural network model building module is used to build a physical information neural network model based on hard boundary constraints after adding hard boundary constraints according to the objective function and constraints of the physical system model;

[0038] A loss function establishment module, used to establish a total loss function of a physical information neural network model based on hard boundary constraints according to the motion equation of the physical system model;

[0039] The deep learning module is used to optimize the total loss function using the augmented Lagrangian algorithm to obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value.

[0040] Beneficial effects:

[0041] 1. The present invention closely combines deep learning technology with physical laws and uses a hard boundary constraint physical information neural network algorithm to obtain the optimal parameters of TLCD more quickly, thereby improving calculation efficiency and reducing R&D costs. The obtained parameters have a significant effect on the response of multi-story steel frames, greatly reducing the amount of calculation and ensuring the high accuracy and reliability of the optimization results.

[0042] 2. The present invention demonstrates excellent generalization and adaptability, and is not only applicable to single-degree-of-freedom systems, but can also be extended to complex multi-degree-of-freedom structural systems. Through the verification of the shaking table test, the present invention has proven its innovation and practicality in the field of seismic optimization of multi-story steel frame structures, and provides an efficient and economical solution for improving the seismic performance of building structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the method of the present invention;

[0044] Figure 2 It is the schematic diagram of the traditional physical information neural network;

[0045] Figure 3 It is a schematic diagram of a hard boundary constrained physical information neural network;

[0046] Figure 4 It is the flow chart of augmented Lagrangian algorithm;

[0047] Figure 5 It is a principle block diagram of the system of the present invention;

[0048] Figure 6 It is a principle block diagram of the computer terminal of the present invention;

[0049] Figure 7 is a schematic diagram of a multi-layer steel frame-TLCD system in the present invention;

[0050] Figure 8 It is the PDE loss function training result diagram of the hard boundary constraint physical neural information network;

[0051] Fig. 9 It is the total loss function training result diagram of the hard boundary constraint physical neural information network;

[0052] Fig.10 It is a schematic diagram of the structural time-history response analysis under seismic wave load;

[0053] Fig.11 It is a schematic diagram of the displacement response of the main structure without control measures, TMD-multi-story steel frame, and TLCD-multi-story steel frame under different simple harmonic loads;

[0054] Fig.12 It is a schematic diagram of the acceleration time history of the main structure, TMD-multi-story steel frame, and TLCD-multi-story steel frame without taking control measures under different simple harmonic loads. DETAILED DESCRIPTION

[0055] The present invention is further described below in conjunction with embodiments and drawings.

[0056] Embodiment 1:

[0057] like Figure 1 As shown, this embodiment proposes a shock absorbing device optimization method based on an improved physical information neural network, and the specific steps are as follows:

[0058] Step 1: Establish a physical system model of the shock absorbing device and obtain the motion equation of the physical system model. Specifically:

[0059] According to the method of physical information neural network model, a domain is established. Physical system model;

[0060] The PDE equation for the model is:

[0061]

[0062] Establish the boundary conditions (or initial conditions) of the model:

[0063]

[0064] In formula (1) and (2):

[0065] is the operator of the partial differential control equation, which contains N partial differential equations {F1, F2, F3, ... F N}; Operators for arbitrary initial or boundary conditions of physical problems; is the boundary of the domain Ω; u(x) is the solution of the partial differential equation, Determined by the parameter γ(x). The solution u(x) of the partial differential equation and the unknown parameter γ(x) in the equation (also called the parameter to be optimized) must satisfy equations (1) and (2).

[0066] An objective function related to the solution u(x) of the partial differential equation and the parameter γ(x) to be optimized is defined to find the solution u(x) of the partial differential equation when the parameter γ(x) is optimal.

[0067] Step 2: Determine the parameters and parameter ranges to be optimized for the shock absorbing device and establish the objective function of the physical system model;

[0068] The present invention only considers inequality constraints, and uses the constraint method of adding equations (1) and (2) to the loss function to minimize the total loss to deal with inequality constraints, so that the unconstrained problem can be transformed into a constrained optimization problem. Then the objective function can be expressed as:

[0069]

[0070] In formula (3) Satisfied with:

[0071]

[0072] The three equations in formula (4) represent partial differential equations (PDEs), boundary conditions (or initial conditions) and inequality constraints, respectively. From formula (1), it can be seen that there may be multiple local optimal solutions for the parameter γ(x) to be optimized in the optimization problem, so attention should be paid to the overfitting problem of the model.

[0073] Step 3: According to the objective function and constraint conditions of the physical system model, a physical information neural network model based on hard boundary constraints is established after adding hard boundary constraints;

[0074] The structure of the traditional physical information neural network is as follows Figure 2As shown in the figure, it is composed of an input layer, a hidden layer, and an output layer. If the number of neurons in the traditional neural network model is large enough, nonlinear functions can be approximated with arbitrary precision. This is based on the universal approximation theorem of neural networks, which uses neural networks to obtain approximate solutions to PDEs based on a large amount of training data. This allows the physical information neural network to embed the physical control equations into the loss function.

[0075] The weights and biases of each neuron in the network are continuously updated through the gradient descent method, minimizing the loss function to obtain an approximate solution to the PDE. The optimization problem is to define the loss term of the PDE in the form of a residual, and add it to the loss function in the form of a penalty term to transform the unconstrained problem into a constrained problem.

[0076] Since the physical information neural network has multiple hidden layers, the input is distributed to the neurons in the hidden layer through linear transformation, and then the solution of PDE is calculated through nonlinear transformation of activation function. Neurons are also connected in a fully connected way. The output of the next layer of neurons This process is called forward propagation and can be expressed as:

[0077]

[0078] In formula (5), are the weights and biases from the neurons in the l-1th layer to the neurons in the lth layer, σ is the activation function, and in this embodiment, the hyperbolic tangent function (tanh) is selected as the activation function.

[0079] According to the partial differential equations and boundary conditions (or initial conditions) defined in equations (1) and (2), PINN is used to approximate the solution u(x) of the partial differential equation, and the mean square error can be constructed as the loss function, which is expressed as:

[0080]

[0081] Where: N u , are the number of known data nodes and the number of training nodes in the domain, respectively; are the training point sets of partial differential equations in the space-time domain; are the data point sets for the boundary or initial conditions of the equations, respectively;

[0082] Based on this, it can be concluded that the loss function of the physical information neural network can be expressed as:

[0083]

[0084] In this embodiment, a uniform sampling method is used. In the residual network, the neural network needs to calculate the partial derivative of the output with respect to the input, which is achieved through automatic differentiation technology. Automatic differentiation technology is automatically solved by the back propagation algorithm library (AD) in Pytorch, and the principle of the back propagation algorithm is to use the chain derivation rule to calculate the partial derivative of the loss function for each weight (or bias), and then update the weight (or bias) layer by layer in reverse according to the network model, and let the loss function converge by continuously adjusting the parameters in the network model.

[0085] This embodiment adds conditions for hard constraints on the basis of PINN to form a physical neural information network model based on hard boundary constraints. The structure of the physical information neural network model based on hard boundary constraints is as follows: Figure 3 As shown, it includes a first fully connected neural network module, a second fully connected neural network module, a hard boundary constraint module, an automatic differentiation module, a residual network module, and a weight update module. The inputs of the first fully connected neural network module and the second fully connected neural network module are both time variables. The output of the first fully connected neural network module is used as the input of the hard boundary constraint module, and the output of the hard boundary constraint module is used as the input of the automatic differentiation module. The output of the automatic differentiation module and the optimal parameters of the shock absorbing device output by the second fully connected neural network module are input into the residual network module together. The output of the residual network module is back-propagated to the first fully connected neural network module and the second fully connected neural network module after passing through the weight update module.

[0086] Furthermore, the first fully connected neural network module and the second fully connected neural network module each include an input layer, a plurality of hidden layers and a plurality of output layers; the input layer, the hidden layer and the output layer are connected in sequence; the input variable of the input layer is a time variable, and the output data of the output layer are respectively the displacement velocity and the optimal parameters of the shock absorbing device under the time variable.

[0087] The constraint conditions of the physical information neural network model based on hard boundary constraints include boundary condition constraints and inequality constraints. The boundary condition is at least one of the Dirichlet boundary condition, the Neumann boundary condition, and the Robin boundary condition. In this embodiment, the Dirichlet boundary condition is used. Assume a solution u i (1≤i≤n) on the Dirichlet frontier:

[0088] u i (x, t) = g0(x, t), x∈Γ D (9)

[0089] In the formula, Belongs to the subset of Dirichlet boundary. In order to make the prediction value of the network satisfy the Dirichlet boundary, a function g(x,t) that is meaningful in the entire domain is constructed, and its function satisfies the special solution of the initial conditions and boundary conditions. At the same time, in order to facilitate the analysis and calculation of the function g0(x,t), D arrive The function g(x, t) in the domain has the same definition as the function g0(x, t), which can be expressed as:

[0090]

[0091] In formula (9), N(x; θ u ) is the original network output, A function is a length factor and satisfies the following two conditions:

[0092]

[0093] In formula (11), the function g(x, t) and the function The network's adjustable parameters (the network's weights and biases) are not included. The function g(x,t) satisfies the boundary (or initial) conditions of the physical equations, while the function It is 0 at the boundary (or initial) condition and increases at points outside the boundary. The constructed function is added to the original network output to form a new network output through network training. In this embodiment, only the physical control equation of the Dirichlet boundary is targeted, and its approximate solution can be defined as:

[0094]

[0095] Step 4: Establish a total loss function of the physical information neural network model based on hard boundary constraints according to the motion equation of the physical system model;

[0096] In this embodiment, the Lagrangian term is augmented to approximate the constraint conditions of the objective function, and the minimized total loss function can be expressed as:

[0097]

[0098] In formula (2), is the penalty coefficient of the kth iteration of the inequality constraint, β h >1; is the penalty coefficient of the kth iteration of PDEs, It is an indicator function, the result is 1 if the condition is true, otherwise it is 0; is the augmented Lagrangian operator of the kth iteration of the augmented Lagrangian method.

[0099] According to the Lagrange method, the gradient of the function must be perpendicular to the constraint surface, so we can get and The expression is:

[0100]

[0101] Step 5: Use the augmented Lagrangian algorithm to optimize the total loss function and obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value;

[0102] The augmented Lagrangian algorithm is also called the multiplier method. In the augmented Lagrangian algorithm, not only a penalty term is used, but also an augmented Lagrangian multiplier is added. Its function is also to serve as a penalty term. In each iteration, the augmented Lagrangian multiplier method will simultaneously update the penalty coefficient and Lagrange multipliers And the penalty term is used to approximate the constraints of the objective function and minimize the total loss function.

[0103] From equation (13), it can be seen that the augmented Lagrangian optimization problem is to optimize the solution of the equation and the optimal parameters in the partial differential equation by training the network, minimizing the objective function, the loss term of PDEs, the inequality loss term, and the total loss of the augmented Lagrangian term. The steps of the augmented Lagrangian method in hPINN are: first, the penalty coefficient is given and penalty factor β h , β F and the augmented Lagrange multiplier The initial value of , and then randomly initialize to start training the network to get the total loss term, and then update the network parameters and penalty coefficients through iterations and penalty factor β h , β F and the augmented Lagrange multiplier The total loss gradually converges, and the specific process is as follows Figure 4 As shown, the specific steps are as follows:

[0104] Step 5.1: Select a deep network neural model and set a penalty coefficient and and the penalty factor β F and β h The initial value and threshold of

[0105] Step 5.2, input the initial value of the motion equation of the physical system model of the shock absorbing device;

[0106] Step 5.3, set the initial number of iterations to 0, set the maximum number of iterations K and the threshold to be reached;

[0107] Step 5.4, send the data to the deep network neural model to start the initial training until the loss function of the deep network neural model converges;

[0108] Step 5.5: If the accuracy requirement is not met or the number of iterations is less than the maximum number of iterations, set the number of iterations k = k + 1, and back propagate to update the weights and biases of the deep network neural model;

[0109] Step 5.6, update the penalty coefficient and augmented Lagrangian operator, adjust the learning rate and optimizer of the deep network neural model and continue training to find the minimum value of the total loss function corresponding to the optimal parameters of the shock absorber. That is, when the total loss function reaches the minimum value, the optimal parameters of the shock absorber under the current time parameters are obtained.

[0110] The pseudo code of the augmented Lagrangian method is shown in the following table:

[0111]

[0112] Embodiment 2:

[0113] like Figure 5 As shown, this embodiment proposes a shock absorbing device optimization system based on an improved physical information neural network for implementing the method described in Embodiment 1, comprising:

[0114] A physical system model building module is used to build a physical system model of the shock absorbing device and obtain a motion equation of the physical system model;

[0115] An objective function building module is used to determine the parameters and parameter ranges to be optimized for the shock absorbing device and to establish the objective function of the physical system model;

[0116] A physical information neural network model building module is used to build a physical information neural network model based on hard boundary constraints after adding hard boundary constraints according to the objective function and constraints of the physical system model;

[0117] A loss function establishment module, used to establish a total loss function of a physical information neural network model based on hard boundary constraints according to the motion equation of the physical system model;

[0118] The deep learning module is used to optimize the total loss function using the augmented Lagrangian algorithm to obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value.

[0119] like Figure 6As shown, this embodiment also proposes a computer terminal based on the shock absorbing device optimization method based on the improved physical information neural network described in Example 1, the computer terminal includes a processor, a memory and a communication interface; the memory and the communication interface are coupled to the processor, and the memory is used to store computer program instructions; wherein, when the processor executes the computer program instructions, the computer terminal implements the shock absorbing device optimization method as described above.

[0120] An embodiment of the present application also provides a computer-readable storage medium, which stores computer program instructions. When the computer program instructions are executed by a processor, the above-mentioned shock absorbing device optimization method based on the improved physical information neural network is implemented.

[0121] Next, take the multi-degree-of-freedom TLCD-multi-layer steel frame system as an example to optimize the parameters:

[0122] Step 1: Establish the physical system model of TLCD-multi-layer steel frame system, such as Figure 7 As shown, the motion equations of the physical system model of TLCD-multi-story steel frame are obtained;

[0123] The mass, stiffness, and damping of each layer are represented by M i , K i , C i (i is the number of layers of the structure). Assume that the main structure with n degrees of freedom is subjected to external load P(t) and earthquake load The equation of motion under the action of is in matrix form:

[0124] MX(t)+CX(t)+KX(t)=P(t)-rMx g (t)

[0125] X(0) = X(16)

[0126]

[0127] The above matrix is ​​expanded as follows:

[0128]

[0129] In formula (17): M, C, K are the mass matrix, damping matrix, and stiffness matrix of the TLCD-multi-story steel frame system; X(t) is the acceleration, velocity and displacement response of the TLCD-multi-story steel frame system; X0, is the initial condition of velocity and displacement of TLCD-multi-story steel frame system; P(t) is the external load vector of TLCD-multi-story steel frame system; is the horizontal seismic acceleration; r is the position vector (all components are 1).

[0130] in:

[0131]

[0132] In formula (18): m0, c0, k0 are the uniform mass, damping and stiffness of the main structure.

[0133] The expressions of frequency and damping ratio of TLCD-multi-story steel frame system are as follows:

[0134]

[0135] Step 2: Determine the parameters and parameter ranges to be optimized for the TLCD-multi-story steel frame system, and establish the objective function of the physical system model;

[0136] In the network model, the input of the fully connected network is only time t. In the TLCD structure system, the parameters to be optimized for TLCD include the head loss coefficient η, the total length of the liquid in the U-tube L, and the mass M of the U-tube. C , the stiffness k of TLCD tlcd 、TLCD damping c tlcd , the cross-sectional area A of the U-shaped tube, where it is assumed that the liquid in the TLCD container is water (the density of water is 1000Kg / m3).

[0137] The specific parameters of the research object are determined, and the optimization parameter range of TLCD is determined as: η∈[0,500], L∈[0,45], M C ∈[0, 0.05M1], k tlcd ∈[0,1350], c tlcd ∈[0,5], A∈[0,50].

[0138] Step 3: According to the objective function and constraint conditions of the physical system model, a physical information neural network model based on hard boundary constraints is established after adding hard boundary constraints;

[0139] First, determine the initial constraints

[0140] Then, Pytorch is used to implement the code of the physical information neural network model (hPINN network) based on hard boundary constraints. A "1+5+3" fully connected network is constructed for the TLCD structural system to obtain the peak displacement of the structure. In this network, "1" is the input layer (the input is time t), "5" is the hidden layer (the number of neurons in each hidden layer is 64), and "3" is the output layer (the output includes the peak displacement x(t) of the main structure, the displacement y(t) of TLCD, and the height difference u(t) of the TLCD liquid level, and the main focus is on the value of x(t)). And a "1+5+6" fully connected network is used to optimize the parameters to be optimized of TLCD. In this network, "1" is the input layer (the input is time t), "5" is the hidden layer (the number of neurons in each hidden layer is 64), and "6" is the output layer (the output includes η, L, M C , k tlcd 、c tlcd , A). The activation function selects the tanh function, and the optimizer first selects the Adam algorithm for optimization. The optimizer continuously optimizes the loss function and updates the network parameters to make the solution of the prediction equation more accurate.

[0141] In the optimization problem of the TLCD structural system, the minimization of the peak displacement of the structural system is set as the objective function, the parameters to be optimized of TLCD are taken as variables, and the problems of multiple constraints and inequality constraints are considered at the same time. Then the optimization problem of the structural system can be defined as:

[0142]

[0143] In formula (20), J is the objective function, which can be specifically expressed as:

[0144]

[0145] In formula (21): They are the main structure displacement and average displacement predicted by the network model respectively; is the TLCD displacement and TLCD average displacement predicted by the network model; It is the TLCD liquid level difference and average liquid level difference predicted by the network model.

[0146] Step 4: Establish a total loss function of the physical information neural network model based on hard boundary constraints according to the motion equation of the physical system model;

[0147] First, the loss function of PDE is expressed as:

[0148]

[0149] In the formula, Equations of motion learned by the network is the penalty coefficient of the kth iteration of PDEs,

[0150] Secondly, the inequality loss function is expressed as:

[0151]

[0152] In formula (23), is the penalty coefficient of the kth iteration of the inequality constraint, β h >1; It is an indicator function, the result is 1 if the condition is true, otherwise it is 0; is the augmented Lagrangian operator of the kth iteration of the augmented Lagrangian method.

[0153] Finally, based on the PDE loss function and the inequality loss function, the expression of the total loss function is established as follows:

[0154]

[0155] Step 5: Use the augmented Lagrangian algorithm to optimize the total loss function, that is, send the data into the network model to start the initial training until the network loss function converges, and find the minimum value of the loss function under the optimal parameters, that is,

[0156] The training results of PDE loss function are as follows Figure 8 As shown, the total loss function training result is as follows Fig. 9 shown.

[0157] A four-story steel frame structure is selected as the analysis object. The specific parameters of the frame structure are listed in Table 1, and its natural frequency is obtained by numerical solution using Etabs software. The relevant data are summarized in Table 2. The structure is dynamically loaded by displacement excitation, and the excitation is applied to the bottom of the structure. The effect of simple harmonic load and seismic wave on the TLCD-multi-story steel frame system is studied. The specific simple harmonic excitation frequency is given in Table 3 according to the 0.2 to 2 multiples of the first-order vibration frequency of the frame structure of 1.302Hz, and the duration of the excitation is 30 seconds.

[0158] Table 1 Quality of four-story steel frame structure

[0159]

[0160] Table 2 Structural characteristics of four-story steel frame structure

[0161]

[0162] Table 3 Simple harmonic load excitation frequency

[0163]

[0164] Considering that TLCD usually has a large size and volume, and the inter-layer space of the multi-layer steel frame is relatively limited, in this embodiment, the TLCD is installed at the top floor of the multi-layer steel frame.

[0165] According to the specific parameters of the research object, the optimization parameter range of TLCD is: η∈[0,500], L∈[0,45], M C ∈[0, 0.05M1], k tlcd ∈[0,1350], c tlcd ∈[0,5], A∈[0,50]. The optimal parameter range of TMD is: M TMD ∈[0, 0.05M1], K TMD ∈[0,1540], C TMD ∈[0, 10].

[0166] Fig.10 It is a structural time-history response analysis under seismic wave load (where a, b, c, and d correspond to the first to fourth floors of a four-story steel frame structure, respectively). The selected seismic wave is the horizontal EI seismic wave with a peak acceleration of 0.1g. As can be seen from the figure, when TMD is installed on a multi-story steel frame, the displacement of each floor can be reduced, but compared with TLCD, the vibration control performance of TMD on the main structure is not ideal. TLCD significantly reduces the displacement response of each floor of the multi-story steel frame. The maximum displacement of the uncontrolled structure of the first floor is 0.01246m, the maximum displacement of the structure equipped with TMD is 0.0057m, and the maximum displacement of the multi-story steel frame equipped with TLCD is 0.00254m. The maximum displacement of the uncontrolled structure of the fourth floor is 0.03562m, the maximum displacement of the structure equipped with TMD is 0.0149m, and the maximum displacement of the multi-story steel frame equipped with TLCD is 0.00541m. Among them, TLCD has the best vibration reduction effect on the fourth floor, and it can be obtained that the vibration reduction rate of TLCD on multi-story steel frames is between 70% and 80%.

[0167] According to the optimal parameters obtained by the method described in this embodiment, under different simple harmonic loads, the displacement responses and acceleration time histories of the main structure without control measures, the TMD-multi-layer steel frame, and the TLCD-multi-layer steel frame are respectively as follows: Fig.11 and 12As shown (where a, b, c, and d correspond to the first to fourth floors of the four-story steel frame structure, respectively). By analyzing and comparing these time history curves, it can be observed that TLCD has the most significant vibration reduction effect on multi-story steel frames. Therefore, the optimal parameters of the vibration reduction device obtained by the method described in this embodiment can effectively suppress structural vibration.

[0168] In summary, the embodiments of the present invention, through the close combination of deep learning technology and physical laws, can more quickly obtain the optimal parameters of TLCD using the hard boundary constraint physical information neural network algorithm, thereby improving calculation efficiency and reducing R&D costs. The obtained parameters have a significant effect on the response of the multi-story steel frame, greatly reducing the amount of calculation and ensuring the high accuracy and reliability of the optimization results. At the same time, the present invention demonstrates excellent generalization ability and adaptability, and is not only applicable to single-degree-of-freedom systems, but can also be extended to complex multi-degree-of-freedom structural systems. Through the verification of the shaking table test, the present invention has proved its innovation and practicality in the field of seismic optimization of multi-story steel frame structures, and provides an efficient and economical solution for improving the seismic performance of building structures.

[0169] Finally, it should be noted that the above description is only a preferred embodiment of the present invention. Under the guidance of the present invention, ordinary technicians in this field can make various similar expressions without violating the purpose and claims of the present invention, and such changes all fall within the scope of protection of the present invention.

Claims

1. A shock absorbing device optimization method based on improved physical information neural network, characterized in that: The steps include: Establish a physical system model of the shock absorbing device and obtain the motion equation of the physical system model; Determine the parameters and parameter ranges to be optimized for the shock absorbing device and establish the objective function of the physical system model; According to the objective function and constraint conditions of the physical system model, a physical information neural network model based on hard boundary constraints is established after adding hard boundary constraints; According to the motion equation of the physical system model, the total loss function of the physical information neural network model based on hard boundary constraints is established; The augmented Lagrangian algorithm is used to optimize the total loss function and obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value.

2. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 1, characterized in that: The parameters to be optimized of the shock absorbing device include one or more of the head loss coefficient, the total length of the liquid in the U-shaped tube, the mass of the U-shaped tube, the stiffness of the shock absorbing device, the damping of the shock absorbing device, and the cross-sectional area of ​​the U-shaped tube.

3. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 1, characterized in that: The physical information neural network model based on hard boundary constraints includes a first fully connected neural network module, a second fully connected neural network module, a hard boundary constraint module, an automatic differentiation module, a residual network module, and a weight update module. The inputs of the first fully connected neural network module and the second fully connected neural network module are both time variables. The output of the first fully connected neural network module is used as the input of the hard boundary constraint module, and the output of the hard boundary constraint module is used as the input of the automatic differentiation module. The output of the automatic differentiation module and the optimal parameters of the shock absorbing device output by the second fully connected neural network module are input into the residual network module together. The output of the residual network module is back-propagated to the first fully connected neural network module and the second fully connected neural network module after passing through the weight update module.

4. The method and system for optimizing a shock absorbing device based on an improved physical information neural network according to claim 3, characterized in that: The first fully connected neural network module and the second fully connected neural network module each include an input layer, a plurality of hidden layers and a plurality of output layers; the input layer, the hidden layer and the output layer are connected in sequence; the input variable of the input layer is a time variable, and the output data of the output layer are respectively the displacement velocity and the optimal parameters of the shock absorbing device under the time variable.

5. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 1, characterized in that: The constraint conditions of the physical information neural network model based on hard boundary constraints include boundary condition constraints and inequality constraints.

6. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 5, characterized in that: The boundary condition is at least one of a Dirichlet boundary condition, a Neumann boundary condition, and a Robin boundary condition.

7. The method and system for optimizing a shock absorbing device based on an improved physical information neural network according to claim 1, characterized in that: The specific steps of optimizing the total loss function by using the augmented Lagrangian algorithm to obtain the optimal parameters of the shock absorbing device when the total loss function takes the minimum value are as follows: Select a deep network neural model and give a penalty coefficient and and the penalty factor β F and β h The initial value and threshold of Input the initial values ​​of the motion equations of the physical system model of the shock absorber; Set the initial number of iterations to 0, set the maximum number of iterations K and the threshold to be reached; Send the data into the deep network neural model to start the initial training until the loss function of the deep network neural model converges; If the accuracy requirement is not met or the number of iterations is less than the maximum number of iterations, set the number of iterations k = k + 1, and back propagate to update the weights and biases of the deep network neural model; The penalty coefficient and augmented Lagrangian operator are updated, and the learning rate and optimizer of the deep network neural model are adjusted and the training is continued to find the minimum value of the total loss function corresponding to the optimal parameters of the shock absorber.

8. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 7, characterized in that: The expression of the total loss function is: L = J + Loss1 + loss2, Among them, J is the objective function, Loss1 is the partial differential equation loss, and loss2 is the inequality loss.

9. The method for optimizing a shock absorbing device based on an improved physical information neural network according to claim 8, characterized in that: The expression of the objective function is: in, They are the displacement of the main structure of the building predicted by the deep network neural model and the average displacement of the main structure of the building; The deep network neural model predicts the displacement of the legal person's shock-absorbing device and the average displacement of the shock-absorbing device respectively; They are respectively the liquid level difference of the shock absorber and the average liquid level difference of the shock absorber predicted by the deep network neural model.

10. A shock absorber optimization system based on an improved physical information neural network, characterized in that: include: A physical system model building module is used to build a physical system model of the shock absorbing device and obtain a motion equation of the physical system model; An objective function building module is used to determine the parameters and parameter ranges to be optimized for the shock absorbing device and to establish the objective function of the physical system model; A physical information neural network model building module is used to build a physical information neural network model based on hard boundary constraints after adding hard boundary constraints according to the objective function and constraints of the physical system model; A loss function establishment module, used to establish a total loss function of a physical information neural network model based on hard boundary constraints according to the motion equation of the physical system model; The deep learning module is used to optimize the total loss function using the augmented Lagrangian algorithm to obtain the optimal parameters of the shock absorbing device when the total loss function reaches the minimum value.

Citation Information

Cited By

  • Seismic absorption and isolation parameter optimization design method based on artificial neural network

    CN120317053A