Design of model-based monitoring systems for machine monitoring.

Abstract

Model-Based Monitoring (MBM) systems have been thought to hold much promise for improving manufacturing productivity and quality. However, the performance of these systems usually fall far short of expectations and they are often "turned off". The main reason for the failure of MBM systems is that they are not properly designed. A proper design requires a thorough understanding of the actual manufacturing machine and process. In addition, it requires an understanding of how to select the best set among the available sensors and estimation algorithms for a machine monitoring problem. There is no design methodology which accounts for these factors. This study investigates qualitatively and quantitatively how these factors affect the performance of MBM systems and develops methodologies for improved MBM systems design. The significant factors are identified by analyzing the monitoring model, measurement set, estimation algorithm, and implementation. The central idea of a qualitative design methodology is the recognition of the interdependent relationships between the monitoring model used to represent the process and the estimation technique selected for estimating the system quantities. It is demonstrated that a simpler monitoring model based on a "smart" choice of the system boundary can lead to the solution of very difficult nonlinear process monitoring problems. The quantitative investigation is conducted for one MBM method, i.e. the state observer technique. The key idea behind the quantitative methodology is the development of performance index. This index accounts for quantitatively the factors affecting the transient and steady-state observer performance. This index is a very effective tool not only to tell a priori how the conditioning of a particular state observer might be, but also to supervise the process of designing the observer-based machine monitors. By minimizing the index, a design methodology for the state observer is developed for the machine monitoring applications, in contrast to using classical pole placement techniques. It is illustrated with the examples that observer designed by the proposed method performs much better than observer based on traditional design techniques.