Recurrent neural networks (RNN) have been used to interpret data in situations wherein our knowledge of the active physics is incomplete. The currency of these methods is the data that are generated by a physical system. Unfortunately, if we are uncertain about the physics of the system, we also do not know the level of uncertainty in the data that we use to represent it.
Typically, data provided to an RNN is provided by measurements of system state information, e.g., data that define speed, position, accelerations, configurations of system elements (like the flaps and elevators on an airplane) etc. But recently, data are being collected that indicate the state of the materials themselves that are used to construct the system. Material state may include the defect state of the materials such as the crack density and patterns in composite material in structural elements (obtained from health monitoring data).
In this paper, we address the question of teaching a control system (e.g., for testing equipment, aircraft control systems, health monitoring systems, etc.) to recognize composite material degradation during service and to adjust applied loads and fields as part of a control scheme to avoid failure of the material during service.
Topics will include defining a proper cost function for the above objectives, formulation of a ‘failure hypothesis’ as a regression function, and the quantification of uncertainty when the physics of the situation is not completely defined. Examples of machine learning techniques for a uniaxial fatigue loading of composite coupons with a circular hole are presented. Example models are random forest regression algorithms and artificial neural networks for linear regression.
