Abstract
An adaptive hybrid learning procedure is proposed for change detection and on-line identification of nonstationary manufacturing processes. Sequential measurements in manufacturing systems and on-line measured signals in speech recognition systems can be well described by nonstationary models (e.g. piece-wise or quasi-stationary models). Multiple models are used for describing a nonstationary process behavior which may randomly switch from one model to another within a set of stationary models. The model orders and parameter values of the members are known initially and they are expanded by on-line model identification. Two novel methods are employed in the design of the proposed hybrid learning procedure: an adaptive segmentation algorithm and an adaptive neural network. The first method is developed for detecting a change in the underlying process model, while the second performs on-line model identification of the changed process. The design of the adaptive hybrid system consists of the following four components: (i) a bank of parallel filters (prediction neural sub-networks) constructed from multiple models, (ii) a model classification neural network, (iii) a sequential model change detection procedure based on an adaptive segmentation algorithm and (iv) the recursive least squares parameter estimation method. The objective of this research is to design an adaptive signal processing system for sequential monitoring of a quasi-stationary process; the primary concern is to detect the time of a change and to perform on-line model adaptation. Two significant contributions are provided: a development of a sensitive change detection procedure and a design of an efficient modular neural network for on-line processing of nonlinear and nonstationary time series. The proposed adaptive hybrid system may be thought of as an extension of the conventional multi-model approach. Here nonlinearity and nonstationarity axe approximated by quasi-stationary piece-wise linear, or piece-wise nonlinear models. Experimental results show that the proposed neural network approach is capable as an on-line learning system with a sensitive detection response in nonstationary environments. A manufacturing process change behavior can be effectively captured by the proposed on-line learning hybrid system.
Wang, Gi-nam (1993). An adaptive hybrid neural network approach for learning nonstationary manufacturing processes. Texas A&M University. Texas A&M University. Libraries. Available electronically from
https : / /hdl .handle .net /1969 .1 /DISSERTATIONS -1483778.