Abstract:Identifying the underlying dynamics of physical systems can be challenging when only provided with observational data. In this work, we consider systems that can be modelled as first-order ordinary differential equations. By assuming a certain pseudo-Hamiltonian formulation, we are able to learn the analytic terms of internal dynamics even if the model is trained on data where the system is affected by unknown damping and external disturbances. In cases where it is difficult to find analytic terms for the disturbances, a hybrid model that uses a neural network to learn these can still accurately identify the dynamics of the system as if under ideal conditions. This makes the models applicable in some situations where other system identification models fail. Furthermore, we propose to use a fourth-order symmetric integration scheme in the loss function and avoid actual integration in the training, and demonstrate on varied examples how this leads to increased performance on noisy data.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to identify the underlying dynamics of a physical system when only observational data are provided. Specifically, the author focuses on systems that can be modeled as first - order ordinary differential equations. By assuming a certain pseudo - Hamiltonian form, the paper proposes a method that can learn the analytical terms of the internal dynamics of the system even when the training data are affected by unknown damping and external perturbations. Moreover, when it is difficult to find the analytical expression of the external perturbation, the paper also proposes a hybrid model that uses a neural network to learn these perturbations, so that it can accurately identify the dynamics of the system as in ideal conditions. This makes the proposed model more advantageous than other system identification models in some cases.
The main contributions of the paper include:
1. **Pseudo - Hamiltonian system identification**: By assuming the pseudo - Hamiltonian form of the system, a model that can simultaneously learn conservation laws, damping, and external forces is proposed. If the external forces are difficult to express in an analytical form, these external forces can be learned through a neural network model.
2. **Use of symmetric numerical integrators**: The paper suggests using symmetric numerical integrators when training the model and shows the superior performance of this method in dealing with noisy data.
### Specific problems and solutions
- **Problem**: How to accurately identify the dynamics of a physical system from observational data in the presence of external perturbations and noise?
- **Solutions**:
- **Pseudo - Hamiltonian form**: By assuming the pseudo - Hamiltonian form of the system, the internal dynamics and external perturbations are separated.
- **Hybrid model**: For external perturbations that are difficult to express analytically, a neural network is used for learning.
- **Symmetric numerical integrator**: A symmetric numerical integration scheme is used in the loss function instead of the actual integration to reduce the impact of noise.
### Mathematical formulas
- **General form of the pseudo - Hamiltonian system**:
\[
\dot{x}=(S(x)-R(x))\nabla H(x)+F(x, t)
\]
where:
- \(S(x)\) is a skew - symmetric matrix that defines the energy - conserving part of the system.
- \(R(x)\) is a positive - semi - definite matrix that defines the dissipation on each coordinate.
- \(H(x)\) is the Hamiltonian of the system.
- \(F(x, t)\) is the external force.
- **Loss function**:
\[
L = \left\|\frac{x_{n + 1}-x_n}{\Delta t}-\Psi_{\Delta t}(\hat{g}_\theta, x_n, x_{n + 1}, t_n)\right\|^2_2+\lambda_H\|\theta_H\|_1+\lambda_F\|\theta_F\|_1+\lambda_R\|\theta_R\|_1
\]
where:
- \(\Psi_{\Delta t}\) is the selected numerical integrator.
- \(\lambda_H, \lambda_F, \lambda_R\) are regularization parameters.
- \(\theta_H, \theta_F, \theta_R\) are the trainable parameters of the Hamiltonian, external force, and damping matrix respectively.
### Experimental verification
The paper verifies the effectiveness of the proposed method through multiple experiments, including the performance on noise - free and noisy datasets, and compares it with existing system identification methods (such as SINDy and SSINN). The results show that the proposed pseudo - Hamiltonian system identification model has better performance in dealing with complex and realistic systems.