Response surface methodology (RSM) is a collection of statistical design and numerical optimization techniques used to optimize processes and product designs. The original work in this area dates from the 1950s and has been widely used, especially in the chemical and process industries. The last 15 years have seen the widespread application of RSM and many new developments. In this review paper we focus on RSM activities since 1989. We discuss current areas of research and mention some areas for future research.

In most statistical process control (SPC) applications, it is assumed that the quality of a process or product can be adequately represented by the distribution of a univariate quality characteristic or by the general multivariate distribution of a vector consisting of several correlated quality characteristics. In many practical situations, however, the quality of a process or product is better characterized and summarized by a relationship between a response variable and one or more explanatory variables. Thus, at each sampling stage, one observes a collection of data points that can be represented by a curve (or profile). In some calibration applications, the profile can be represented adequately by a simple straight-line model, while in other applications, more complicated models are needed. In this expository paper, we discuss some of the general issues involved in using control charts to monitor such process- and product-quality profiles and review the SPC literature on the topic. We relate this application to functional data analysis and review applications involving linear profiles, nonlinear profiles, and the use of splines and wavelets. We strongly encourage research in profile monitoring and provide some research ideas.

The purpose of this paper is to provide an overview and our perspective of recent research and applications of statistical process monitoring. The focus is on work done over the past decade or so. We review briefly a number of important areas, including health-related monitoring, spatiotemporal surveillance, profile monitoring, use of autocorrelated data, the effect of estimation error, and high-dimensional monitoring, among others. We briefly discuss the choice of performance metrics. We provide references and offer some directions for further research.

We compare the performance of two phase II monitoring schemes for linear profiles, one based on the classical calibration method monitoring the deviations from the regression line (referred to as the NIST method) and the second based on individually monitoring the parameters of the linear profile (referred to as the KMW method). The comparison criterion is average run length performance under different sustained shifts in the intercept, slope and error standard deviation of the linear calibration line. A simulation study shows that the NIST method performs poorly compared to the combined control charting scheme of the KMW method.

In many industrial experiments, time and/or cost constraints often force certain factors in a designed experiment to be much harder to change than others. An appropriate approach to such an experiment restricts the randomization, which leads to a split-plot structure. This paper first establishes how one can modify the common central-composite design to efficiently accommodate a split-plot structure. The proposed designs allow pure-error estimates of the two variance components. We next discuss the conditions on the design that make ordinary least squares and weighted least squares estimates equivalent. These conditions are easy to obtain in practice. An important consequence is that people can use standard experimental design software to analyze second-order response surfaces. We give examples of central composite and Box-Behnken designs that meet the conditions. We conclude with an example.

Computer-generated designs are useful for situations where standard factorial, fractional factorial or response surface designs cannot be easily employed. Alphabetically-optimal designs are the most widely used type of computer-generated designs, and of these, the D-optimal (or D-efficient) class of designs is extremely popular. D-optimal designs are usually constructed by algorithms that sequentially add and delete points from a potential design by using a candidate set of points spaced over the region of interest. We present a technique to generate D-efficient designs using genetic algorithms (GA). This approach eliminates the need to explicitly consider a candidate set of experimental points and it can be used in highly constrained regions while maintaining a level of performance comparable to more traditional design construction techniques.