Roadbed compaction system parameter identification method based on RK-PINN network

Abstract

The invention discloses a roadbed compaction system parameter identification method based on an RK-PINN network, and belongs to the technical field of high-speed railway roadbed continuous compaction detection. The method is used for solving the problem of real-time and accurate identification of the parameters of the roadbed compaction system under the vibration compaction working condition in the prior art. The method comprises the following steps: constructing a two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system; constructing a physical information neural network (RK-PINN) with an embedded Runge-Kutta method; according to a collected vibration acceleration signal of the vibratory roller, an RK-PINN network is combined to be applied to parameter identification of a two-degree-of-freedom nonlinear dynamic model of a vibratory roller-roadbed system, and the identified parameters can reach a high-precision standard and can more truly reflect the compaction state of a soil body; and a basis is provided for establishment of continuous compaction indexes of the high-speed railway roadbed. According to the roadbed compaction system parameter identification method based on the RK-PINN network provided by the invention, the parameter change condition can be identified in real time with high precision so as to reflect the real condition of roadbed compaction.

Description

Technical Field

[0001] The invention relates to the field of roadbed compaction detection, and in particular to a method for identifying parameters of a roadbed compaction system based on a RK-PINN network. Background Art

[0002] In high-speed railway construction projects, the stability of the roadbed has a direct impact on the quality of railway construction. In order to ensure that the roadbed has sufficient strength and stability, roadbed compaction operations are essential and necessary. Its goal is to ensure that the roadbed and pavement structure materials have sufficient density. The evaluation index of the continuous compaction quality of high-speed railway roadbed is compaction. Whether the compaction is qualified is an important criterion for judging the quality of the project. An effective compaction detection method is essential for implementing compaction monitoring.

[0003] Continuous compaction detection of roadbed is a complex detection technology. The continuous compaction parameter model mechanism is complex and its characterization is affected by many factors, such as the parameters of the vibratory roller, the type, thickness, and moisture content of the filler. Traditional roadbed compaction detection methods mostly use sampling detection methods such as sand filling method and ring knife method. It is impossible to measure and evaluate the compaction status in real time during the compaction process. Only limited sample materials can be used for testing after the compaction is completed. Its representativeness is poor and it is difficult to reflect the compaction status of every point on the road. It is easy to cause the phenomenon of missing "weak points". Once missed, it may be fatal to the operation of high-speed railways. How to continuously and accurately detect the compaction degree of the roadbed in real time has also become a key and difficult point.

[0004] With the rapid development of computer and artificial intelligence technology, the agent model represented by deep learning methods has shown a very broad application prospect in the engineering field, providing new ideas and ways to solve complex problems in engineering. And it has been applied in many engineering fields. In recent years, the introduction of physical information neural network (PINN) has provided new ideas and ways to realize the real-time identification of parameters of continuous vibration compaction system of roadbed, and has been applied to the problem of continuous compaction degree detection of roadbed. Physical information neural network (PINN) is usually composed of a deep neural network. Its core idea is to embed the residual form of differential equations and partial differential equations into the loss function of deep neural network through automatic differentiation technology. When training the network model, it is necessary not only to minimize the data residual loss, but also to minimize the physical equation loss to ensure that the prediction results conform to the laws of physics, so that it can be used to solve physical problems related to differential equations and partial differential equations. Physical information neural network can not only solve various related physical equations such as differential equations and partial differential equations, but also invert the unknown parameters in the equations. Although the physical information neural network solves the problem of the inability to measure compaction parameters and evaluate compaction status in real time in traditional roadbed continuous compaction degree detection methods by inverting the compaction parameters in the roadbed compaction dynamics equation, the roadbed compaction process is an extremely complex process, and the dynamic equations that characterize the roadbed compaction process are very complex. When solving the inverse problem of complex differential equations, the physical information neural network faces limitations that cannot be ignored, such as high computational cost, poor convergence and stability, and difficulty in applying to high-order derivative equations. This makes it still challenging for the physical information neural network to identify the roadbed compaction system parameters in real time.

[0005] There are currently few research results by domestic scholars on the application of deep learning in continuous roadbed compaction detection. How to further improve the application effect of physical information neural networks in the field of roadbed compaction engineering, better and more practically apply deep learning technology to the field of roadbed compaction engineering, and accurately identify the compaction parameters of the roadbed continuous vibration compaction system in real time to reflect the actual situation of roadbed compaction is of great significance for evaluating the compaction quality of high-speed railway roadbed and realizing process control. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for identifying parameters of a roadbed compaction system based on an RK-PINN network, which can realize real-time and accurate identification of parameters of a roadbed continuous vibration compaction system under vibration compaction conditions.

[0007] A first aspect of the present invention provides a dynamic modeling method for a roadbed continuous vibration compaction system, comprising the following steps:

[0008] S1. Construct a two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system;

[0009] Furthermore, the two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system is a model obtained by introducing a nonlinear term characterizing the dry friction generated by the relative motion between the vibratory wheel and the roadbed soil on the basis of the traditional two-degree-of-freedom linear dynamic model of the vibratory roller-roadbed system that can reflect the roadbed compaction process.

[0010] Furthermore, the two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system has a dynamic equation expressed as follows:

[0011]

[0012] Where m1, m2 and m3 represent the mass of the frame, the mass of the vibrating wheel and the mass of the vibrating roadbed soil, respectively; x1 and x2 represent the vertical displacement of the roller frame and the vibrating wheel, respectively; F0 represents the amplitude of the exciting force of the vibrating wheel; ω represents the working angular frequency of the vibrating wheel; k1 and c1 are the parameters representing the stiffness and damping of the vibration roller frame shock absorber, respectively; k2 and c2 are the parameters representing the stiffness and damping of the roadbed soil, respectively. It is a nonlinear term that characterizes the dry friction generated by the relative motion between the vibrating wheel and the roadbed soil.

[0013] A second aspect of the present invention provides a physical information neural network (RK-PINN) with an embedded Runge-Kutta method, comprising:

[0014] S2, constructing a physical information neural network embedded with Runge-Kutta method (RK-PINN);

[0015] Furthermore, the physical information neural network with embedded Runge-Kutta method (RK-PINN) is a new type of neural network that combines the idea of ​​physical information neural network and improves on the structure of traditional physical information neural network. By changing the input and output of the traditional network, the Runge-Kutta algorithm module is embedded to replace the automatic differentiation calculation and physical information constraint process in the traditional network, and the loss function is improved. It is specially used for solving the inverse problem of differential equations with high efficiency and high precision.

[0016] The third aspect of the present invention provides a method for identifying parameters of a roadbed compaction system based on an RK-PINN network, comprising the following steps:

[0017] S3, collecting actual vibration acceleration signals, and inputting the collected vibration acceleration signals into the network for training;

[0018] Furthermore, based on the constructed physical information neural network with embedded Runge-Kutta method (RK-PINN), the network inputs the actual collected roller vibration acceleration signal to achieve the reverse solution of the two-degree-of-freedom nonlinear equation of roadbed compaction. That is, the unknown parameters in the equation are inverted based on the known actual real data.

[0019] S4. Identify compaction system parameters during the actual compaction process.

[0020] Furthermore, the actual collected vibration acceleration signal of the road roller is input into the network for training. After the training is completed, the network outputs the unknown parameters of the equation to be solved in the nonlinear equation, and the parameter results output by the network are the compaction system parameters in the actual compaction process to be identified.

[0021] The present invention provides a method for identifying parameters of a roadbed compaction system based on an RK-PINN network. Compared with the prior art, the present invention has the following effects:

[0022] 1. This application can more realistically and accurately reflect the actual roadbed compaction process and the changes in compaction parameters during the roadbed compaction process by constructing a two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system.

[0023] 2. This application applies deep learning to the technical field of continuous compaction detection of roadbed. The physical information neural network (RK-PINN) with embedded Runge-Kutta method not only perfectly solves the automatic differentiation ill-conditioned problem of traditional physical information neural networks when solving inverse problems of differential equations, but also greatly reduces the network loss space. Not only does it greatly improve the accuracy of the inversion parameter results of the neural network, it also achieves a significant improvement in the magnitude of the solution speed, and also improves the stability of the network. The network meets the requirements of novelty and is creative.

[0024] 3. This application has innovatively improved the physical information neural network and constructed a physical information neural network (RK-PINN) with an embedded Runge-Kutta method. It combines the powerful data fitting ability of the neural network with the powerful high-precision and high-computational efficiency of the Runge-Kutta algorithm in equation iteration, and combines the advantages of the neural network in inverting differential equations with the speed and accuracy advantages of the Runge-Kutta algorithm in forward solving differential equations. It can input the vibration acceleration signal collected in real time through the network, and the network directly outputs the compaction system parameter results to be identified, thereby realizing the parameter inversion of the nonlinear dynamic equation. The identified parameter results can accurately match the real compaction parameters of the continuous vibration compaction system of the roadbed, and can truly reflect the compaction state of the roadbed soil during the vibration compaction condition, solving the problem that the existing roadbed continuous compaction degree detection technology cannot accurately identify the compaction parameters of the roadbed continuous vibration compaction system in real time, and has the practicality to be used in actual roadbed compaction engineering operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process structure of the roadbed compaction system parameter identification method based on the RK-PINN network;

[0026] Figure 2 It is the structural diagram of the two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system;

[0027] Figure 3 It is the structural framework diagram of the physical information neural network (RK-PINN) embedded with Runge-Kutta method; DETAILED DESCRIPTION

[0028] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0029] In order to better highlight the advantages of the present invention, the research scheme of the present invention will be described in detail below in conjunction with the accompanying drawings. Figure 1 As shown in the figure, a method for identifying parameters of roadbed compaction system based on RK-PINN network is presented. The main steps are described in detail as follows:

[0030] S1: Construct a two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system;

[0031] Considering that during the roadbed compaction operation, a certain amount of dry friction will be generated due to the contact and relative movement between the vibrating wheel and the roadbed soil, and the soil damping includes Coulomb dry friction damping, which makes the soil stiffness show a certain degree of nonlinear change. Based on the traditional two-degree-of-freedom linear dynamic model of the vibrating roller-roadbed system, a nonlinear term characterizing the dry friction generated by the contact and relative movement between the vibrating wheel and the roadbed soil is introduced to construct a two-degree-of-freedom nonlinear dynamic model. The structure of the constructed two-degree-of-freedom nonlinear dynamic model of the vibrating roller-roadbed system is as follows: Figure 2 shown.

[0032] Through Figure 2 The nonlinear dynamic model structure stress analysis shown in the figure establishes the two-degree-of-freedom nonlinear dynamic equation of roadbed compaction, which is expressed as follows:

[0033]

[0034] Where m1, m2 and m3 represent the mass of the frame, the mass of the vibrating wheel and the mass of the vibrating roadbed soil, respectively; x1 and x2 represent the vertical displacement of the roller frame and the vibrating wheel, respectively; F0 represents the amplitude of the exciting force of the vibrating wheel; ω represents the working angular frequency of the vibrating wheel; k1 and c1 are the parameters representing the stiffness and damping of the vibration roller frame shock absorber, respectively; k2 and c2 are the parameters representing the stiffness and damping of the roadbed soil, respectively. It is a nonlinear term that characterizes the dry friction generated by the relative motion between the vibrating wheel and the roadbed soil.

[0035] Among them, m1, m2, m3, F0, ω, are known parameters, x1 and x2 are the parameters actually collected on site during the roadbed compaction operation, and k1, c1, k2 and c2 are the parameters of the roadbed continuous vibration compaction system to be identified during the roadbed compaction process.

[0036] S2: Constructing a physical information neural network embedded with Runge-Kutta method (RK-PINN);

[0037] The physical information neural network with embedded Runge-Kutta method (RK-PINN) is improved on the basis of the traditional physical neural network structure, changes the input and output of the network structure, and embeds the Runge-Kutta algorithm module to replace the automatic differentiation calculation process in the traditional network and the physical information constraint process in the loss space. It is a new type of neural network for solving the real-time and accurate inversion problem of complex differential equation parameters.

[0038] The RK-PINN network is mainly composed of a feedforward network structure and a Runge-Kutta algorithm module. Figure 3 As shown in the figure, its characteristics are that, on the one hand, the Runge-Kutta algorithm module is embedded in the network structure to replace the automatic differentiation calculation process in the traditional physical information neural network and the optimization process of the physical information constraint in the loss function, thus reducing the loss space of the network. On the other hand, by changing the input and output of the traditional network and solving the differential equations by the Runge-Kutta algorithm instead of the process of fitting the positive solution of the differential equations by the feedforward network in the traditional physical information neural network, it no longer relies on the neural network to perform global fitting of the solution. The above characteristics not only reduce the consumption of computing resources, but also ensure the high accuracy, stability and convergence of the network.

[0039] The feedforward network structure of the RK-PINN network is the same as that of the traditional physical information network, both of which use a fully connected neural network. However, unlike the traditional physical information neural network, the input and output of the RK-PINN network have changed. The network no longer uses the time coordinate t as input, but the actual collected acceleration signal. As the input of the network; the network output becomes the compaction system parameters k1, c1, k2 and c2 to be inverted, rather than the forward solutions x1 and x2 of the differential equation. The system parameters k1, c1, k2 and c2 are no longer trainable parameters trained and optimized together with the network parameters, but are directly used as the output parameters of the network. The fully connected neural network consists of three parts: the input layer, the hidden layer and the output layer. Each layer includes neurons that are unconnected within the layer and fully connected between layers. Among them, the number of layers of the network and the number of neurons in each layer can be set according to the complexity of the actual problem.

[0040] The Runge-Kutta algorithm module in the RK-PINN network is composed of a fourth-order Runge-Kutta numerical solution method. The network outputs the compaction system parameters, which are passed to the Runge-Kutta module. The Runge-Kutta algorithm module uses the fourth-order Runge-Kutta method to forward solve the nonlinear dynamic equation under the current network output parameters based on the parameter prediction values ​​output by the network, combined with the known initial conditions and step size of the nonlinear dynamic equation, and obtains the positive solution of the nonlinear dynamic equation under the compaction parameters output by the current network. This process replaces the automatic differentiation calculation and physical information constraint process in the traditional physical information network. The traditional physical information neural network embeds the physical equation into the network loss function through automatic differentiation technology, and realizes the fusion of physical information and neural network by minimizing the loss function. However, this process not only consumes a lot of computing resources, but also greatly increases the complexity of the loss function space, resulting in slow network training speed, poor convergence and stability, and poor practicality in the technical field of continuous compaction detection of roadbed. The physical information in the RK-PINN network is reflected in the embedding of physical equations into the fourth-order Runge-Kutta algorithm. The advantages of the Runge-Kutta method in numerically solving differential equations are fast and accurate. Based on the current output parameters of the network, the dynamic equations are iteratively forward solved. This not only perfectly solves the automatic differentiation ill-posed problem in traditional networks, but also can potentially add physical information to the loss function and realize the forced constraint of physical information into the network, which not only ensures the effective integration of physical information into the network, but also reduces the loss function space.

[0041] The loss function of the RK-PINN network is composed of data loss. The network outputs compaction system parameters, which are passed to the Runge-Kutta module. The fourth-order Runge-Kutta algorithm in the Runge-Kutta module iteratively solves the physical equations based on the parameters output by the network, and generates the numerical solution of the equation under the compaction parameters output by the current network. The data residual is formed by combining the numerical solution of the equation solved by the fourth-order Runge-Kutta algorithm with the input of the network (actually collected real data), which constitutes the loss function of the network. In this way, the solution results of the Runge-Kutta algorithm are associated with real data, which provides a basis for the training and optimization of the network.

[0042] S3, collecting actual vibration acceleration signals, and inputting the collected vibration acceleration signals into the network for training;

[0043] S4. Identify compaction system parameters during the actual compaction process.

[0044] The specific implementation of steps S3 and S4 includes: after step S2 determines the RK-PINN network model for real-time and accurate identification of compaction parameters of the roadbed continuous vibration compaction system in the roadbed compaction operation, real-time acquisition of vibration acceleration signals of the vibrating roller in the roadbed compaction operation Input to the network model, the network outputs a set of random compaction parameter values ​​k1, c1, k2 and c2 according to the input acceleration signal. Subsequently, the parameters k1, c1, k2 and c2 are input to the Runge-Kutta module. The fourth-order Runge-Kutta algorithm in the module combines the parameters k1, c1, k2 and c2 to forward solve the two-degree-of-freedom nonlinear dynamic equation of roadbed compaction, and obtain the forward numerical solution of the nonlinear dynamic equation under the parameters k1, c1, k2 and c2. The equation is numerically solved by the fourth-order Runge-Kutta method Input to the network The data residual term is formed and the mean square error is used to form the network's loss function. The network optimizes the loss function through gradient descent, and minimizes the loss function by continuously optimizing and adjusting the network parameters. After the training is completed, the network outputs the true and accurate compaction system parameter value. The true and accurate compaction system parameter value output by the network is the compaction system parameter to be identified in the roadbed compaction project, thereby realizing the real-time and accurate identification of the compaction parameters of the roadbed continuous vibration compaction system.

[0045] The physical information neural network (RK-PINN) with embedded Runge-Kutta method constructed by the present invention combines the powerful data fitting ability of the neural network with the powerful high-precision and high-computational efficiency equation iteration calculation ability of the Runge-Kutta method through innovative improvements on the architecture and loss function of the traditional physical information neural network, and combines the advantage of the neural network in inverting differential equations with the speed and accuracy advantages of the Runge-Kutta algorithm in forward solving differential equations. The network can input the vibration acceleration signal collected in real time, and the network directly outputs the compaction system parameter results to be identified. The identified parameter results can accurately match the real compaction parameters of the roadbed continuous vibration compaction system, and can truly reflect the compaction state of the roadbed soil during the vibration compaction working condition, which solves the problem that the compaction parameters of the roadbed continuous vibration compaction system cannot be accurately identified in real time in the existing roadbed continuous compaction degree detection technology, and has practical engineering significance in actual roadbed compaction engineering operations.

[0046] The above is only an illustrative description of the principles and effects of the present invention, and is not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for identifying parameters of a roadbed compaction system based on an RK-PINN network, characterized in that: The steps include: S1. Construct a two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system; S2, constructing a physical information neural network embedded with Runge-Kutta method (RK-PINN); S3, collecting actual vibration acceleration signals, and inputting the collected vibration acceleration signals into the network for training; S4. Identify compaction system parameters during the actual compaction process.

2. According to claim 1, a method for identifying parameters of a roadbed compaction system based on an RK-PINN network is characterized in that: The two-degree-of-freedom nonlinear dynamic model of the vibratory roller-roadbed system is based on the traditional two-degree-of-freedom linear dynamic model of the vibratory roller-roadbed system that can reflect the roadbed compaction process, and the model is obtained by introducing a nonlinear term that characterizes the dry friction generated by the relative motion between the vibratory wheel and the roadbed soil.

3. The method for identifying parameters of a roadbed compaction system based on an RK-PINN network according to claim 1, characterized in that: The Physics Informed Neural Networks embedded with Runge-Kutta Method (RK-PINN) is a new type of network specifically used for solving inverse problems of differential equations, based on the traditional Physics Informed Neural Networks (PINN), which embeds the Runge-Kutta algorithm module to replace the automatic differentiation calculation process and physical information constraints in the traditional network, and improves the network input and output structure and loss function.

4. The method for identifying parameters of a roadbed compaction system based on an RK-PINN network according to claim 1, characterized in that: The identification of compaction system parameters in the actual compaction process is essentially equivalent to a physical information neural network (RK-PINN) based on an embedded Runge-Kutta method, which inverts the unknown parameters in the two-degree-of-freedom nonlinear dynamic equation of roadbed compaction based on known actual real data.

5. The method for identifying parameters of a roadbed compaction system based on a RK-PINN network according to claim 1, characterized in that: The steps S3 and S4 include: The real-time collected roller vibration acceleration signal (actual real data) is used as the network input, and the unknown parameters of the two-degree-of-freedom nonlinear dynamic equation to be solved are used as the network output. The parameter results of the network output are the compaction system parameters in the actual compaction process to be identified.

Images