A new multi-scale neural approach

Dynamical systems describe the evolution of natural phenomena over time and space through mathematical frameworks, often using differential equations. Accurate predictions in these systems are important for a variety of applications, but traditional methods face challenges due to their rigidity and complex dynamic behavior.

Existing models often oversimplify these systems, introducing biases and errors.

Purely data-driven approaches and physics-based machine learning methods have emerged to address these issues. However, there is still a need for more effective methods to accurately learn and predict the behavior of dynamic systems under rigorous conditions. These challenges have necessitated intensive research.

The new study, conducted by researchers at Qingdao University in China, introduces a multi-scale differential algebraic neural network (MDANN) method. International Journal of Mechanical System Dynamics March 20, 2024,This research aims to improve the learning of dynamical systems,characterized by particularly stringent conditions.

MDANN methods incorporate Lagrangian mechanics and multi-scale information to improve the accuracy and efficiency of predictions in complex systems.

The MDANN method has two main components: a Lagrangian mechanics module and a multi-scale module. The Lagrangian mechanics module simplifies the learning process by embedding the system in Cartesian coordinates, using a differential algebraic equation formalism, and explicitly imposing constraints through Lagrangian multipliers.

The multi-scale module effectively transforms high-frequency components into low-frequency components through radial scaling, allowing the system to learn sub-processes at different speeds.Experimental validation on a coupled pendulum system, a double pendulum system, and a scissor-type deployable mast system demonstrates the excellent performance of the MDANN method.

This approach resulted in a mean squared error (MSE) of 3.214e-2 for position and 2.590e-3 for energy for the coupled pendulum system. For the double pendulum system, the MSE for position was 9.638e-02 and the MSE for energy was 5.091e-01. Additionally, we optimized the control forces of the scissor-type deployable mast system to achieve uniform motion.

These results highlight the ability of MDANN to manage stringent system complexity and significantly improve prediction accuracy and computational efficiency.

“The development of the MDANN method is a major advancement in the field of dynamic systems learning. By integrating Lagrangian mechanics and multi-scale information, we have addressed a long-standing challenge associated with rigid systems. This approach not only improves prediction accuracy but also provides a practical solution for complex engineering applications,” said Professor Jieyu Ding, one of the principal investigators.

The application of the MDANN method will transform fields that rely on precise dynamic system modeling, such as aerospace and robotics. The method's ability to improve control strategies and ensure operational safety is ushering in a new era of efficiency and reliability in system prediction and optimization.

Provided by Qingdao University