1. Proposal of a parameter identification method for singledegree-of-freedom nonlinear systems using neural networks
- Author
-
Daiki TAJIRI, Kosuke NAKAJIMA, Masaki IKEDA, Shozo KAWAMURA, and Masami MATSUBARA
- Subjects
nonlinear vibration system ,identification ,neural network ,time domain ,partial differentiation ,Mechanical engineering and machinery ,TJ1-1570 - Abstract
This study investigated a time-domain method that uses a neural network (NN) to identify three linear parameters such as mass, viscous damping coefficient and spring constant, as well as nonlinear forces of a nonlinear vibration system. In this method, the measured excitation force, acceleration, velocity, and displacement are input to the NN, which learns the force equilibrium between the external force, inertial force, damping force, and nonlinear restoring force, thus identifying the characteristics of the vibration system in an explicit manner. The proposed NN consists of two subnetworks: the linear subnetwork and the nonlinear subnetwork. The nonlinear subnetwork is called a global NN and cooperates with a local NN that extracts linear parameters. The features of the proposed method are: I) linear parameters are intentionally extracted based on the equation of motion, and II) guidelines for setting hyperparameters can be obtained from the behavior of the mean squared error (MSE). In this study, numerical simulations for parameter identification were performed to validate the proposed identification method. The vibration systems investigated were those governed by the Duffing and Van der Pol equations, as well as those in which the restoring force is represented by a sine function. The data obtained by numerically solving the equations of motion were considered as experimental data, and the linear parameters and nonlinear forces were identified to confirm the validity and applicability of the proposed method.
- Published
- 2024
- Full Text
- View/download PDF