Generalized Linear Model-Based Process Control of Multivariate Measurements

点击翻译按钮获取中文摘要

The objectives of this research are to develop model-based control algorithms for multiple variables that operate when the measurements are a mixture of normally and non-normally distributed variables. Situations such as this arise frequently in industry, notably in the semiconductor industry, where the manufacturing and quality databases contain variables such as counts of defects or particle contaminates and electrical parameters, in addition to yield and other performance characteristics. These situations are also frequently encountered in the chemical and process industries. The research may also extend this algorithm to cascade processes; that is, processes with sequential manufacturing steps and significant added value processing occurring at each step. Processes in semiconductor manufacturing and chemical and process industries often add value sequentially over a number of process steps. Several process and product characteristics might be measured at each step. In this type of process, a shift in an upstream subset of variables can propagate into downstream subset variables although nothing is wrong with the process at later stages. A strategy that links the steps to improve process control will be incorporated into this algorithm. If successful, the results of this research will lead to better use of the data to improve the control of processes with many measurements of various types. As automated data acquisition systems become more prevalent, more processes will require the type of analysis proposed. In addition, the focus on cascade processes will improve the ability to detect and isolate process problems through a segmented design that considers the deviance from expected results at each processing step. This approach should providebetter control of these processes, and greater knowledge of the interrelationships between process steps. This integration of different measurement types in cascade processes should lead to new methodology forthe regression-adjustment of nonstandard data and contribute to the research in statistical deviancies. Furthermore, it is expected that the algorithms developed here will become benchmarks for other researchers to use for subsequent extensions to particular cascade processes.