Reinforcement-learning-based Identification of the System for the Purpose of Structural Change Detection

摘要

Reinforcement learning provides an edge in many problems of mechanical engineering. That is because its ability to solve problems on-the-fly based on interactions with initially unknown environment. In this paper the reinforcement learning approach to identification, adaptive control and structural change detection is presented. The approach consists of training feedforward artificial neural networks in a task of adaptive control of an unknown and time-variable system. The nets are used to build system model and propose optimal control prescription to actively suppress vibration of a selected system element. The difference between expected and obtained quality of outcome is used as a damage indicator - in a novelty-detection-based framework. A series of numerical experiments for a 5DOF system is used to evaluate method's performance. It is shown that the method is able to operate for both impulse and white noise external excitation and can work despite random and significant changes of system parameters.