Investigation of a neural network methodology to predict transient performance in FMS.

摘要

Scope and method of study. Most rapid analytical evaluative models for Flexible Manufacturing Systems (FMSs) are based on the steady-state performance. There is a practical need to develop robust, easy to construct, and transportable transient-state evaluative models for FMSs. This study proposes an ANN based metamodeling framework that can capture various post disruption system behaviors of FMS. The proposed ANN based metamodeling scheme consists of a hierarchical taxonomy of multiple ANNs. Each set of ANNs collectively represents a different part of the underlying system modeling domain. The taxonomical arrangement of multiple ANNs overcomes shortcomings often found in single ANN based metamodeling schemes. These shortcomings are generally related to the limited knowledge acquisition capability of these schemes. The study uses an Extend based discrete simulation model that is built after an experimental FMS with a limited disruption trigger and handling capabilities. The simulation model is used to study various post-disruption behaviors by a given FMS and to study the feasibility of the proposed modeling scheme as a viable means to provide "lookahead" capability for a low level controller.; Findings and conclusions. The proposed ANN based metamodeling approach using multiple ANNs, in a taxonomically organized modeling structure, is an efficient way to capture multiple target performance index observation processes with a similar overall post-disruption behavior pattern. Despite its accuracy issues, this methodology was proven especially effective when it has to deal with noisy time series such as TIS at observation under a data rich environment. The study is to prove that the proposed methodology could be a viable means to model transient system behaviors. As long as individual observation processes of the selected performance index can keep their variances smaller among themselves, the accuracy of the overall model would be acceptable. This non-parametric performance modeling technique using hierarchically organized multiple ANNs, is worth further investigation.