Optimisation of System Configuration Using Machine Learning as a Surrogate Model

Abstract

Complex mechanical systems provide a degree of reliability through redundancy. That is, they have duplicate systems, either similar or dissimilar, performing the same function. These duplicate systems will vary in their dynamic properties and may have different vibration transmission paths due to their connections within the overall system. Therefore, the use of redundancy gives rise to different vibrational behaviour of the overall system depending on the machines selected to deliver the intended system output. In some circumstances, it is necessary to minimise vibration levels by selecting a particular system configuration. In large complex systems consisting of multiple sources, this is not a trivial task, and virtually impossible to investigate all possible machinery combinations leading to the lowest vibration levels, whilst delivering the desired output. To address this, a machine learning model has been trained to provide predictions of machinery vibration levels. Data are obtained using an experimental vibration rig, emulating a complex, multi-source mechanical system. The machine learning model is subsequently utilised within a genetic algorithm optimisation routine in order to obtain the system configuration producing the lowest vibration levels at a number of observer locations for a specified system output.