Author:

Keywords:

Design of Experiments, EVOP, Simplex

Abstract:

In industry, guaranteeing good product quality is essential and requires continuous monitoring and control of the production process. When the quality is not optimal, the process settings need to be changed. Under laboratory conditions this would be performed using an experimental design during which well-chosen combinations of the process settings are imposed so that the optimal settings can be defined. However, this approach is not feasible when searching for optimal settings in the case of a full-scale process since it typically requires exploring the extreme regions of the process where the probability of producing unsaleable product is very high. In order to overcome the drawbacks of classical experimentation, methods for improving full scale processes are investigated in this dissertation. Special attention will be paid to methods which are easy to implement and which are applicable to processes which involve a large number of factors that possibly interact, since this is the situation which is found in contemporary processes. In the first part of the thesis, a literature survey of methods that are suitable for Online Sequential Process Improvement was conducted. This literature study resulted in two potential candidates for online experimentation, being the Evolutionary Operation (EVOP) and Basic Simplex method. Extensions to the EVOP methodology were developed to allow for an automated implementation along with a novel way to deal with the borders of the experimental domain. Furthermore, a steepest ascent search was combined with the EVOP method. In conjunction with these extensions, a Matlab software package was created that allows the easy execution of these methods for process improvement in practice. The final part of the methodology research was focussed on presenting a framework for selecting the starting point for these methods should no prior be established by process experience or offline experimentation. Space-filling designs combined with Gaussian Process modelling was shown to achieve a good initial starting point, using a small number of measurements. In the second part of the dissertation simulation studies were performed to thoroughly investigate the applicability of the methods on contemporary processes. The EVOP and Simplex methods were compared and this study showed that Simplex is the preferred choice when dealing with deterministic or low-noise systems. The EVOP method proved to be attractive to improve processes characterized by the presence of a substantial amount of noise and/or a dimensionality above three factors. Therefore, the feasibility of a more efficient design -such as a fractional factorial of at least resolution III- was investigated. It was concluded that applying such minimalistic designs offers a significant improvement in the total number of measurements required to reach the optimum and opens up possibilities for the use of EVOP in processes in which fast decisions are required. In the final simulation chapter, first results about the statistical power that is required for an efficient improvement was researched. It was shown that the optimal statistical power increases when the dimensionality increases. However, the exact choice of the power is not too critical and shows a broad, almost flat valley for high dimensions. This allows for much flexibility in the base design for EVOP, depending on the process under study. For processes with a low sampling rate or with a non-stationary behaviour a low power is recommended, whereas a higher power is advised when a high sample rate is possible or the process is stationary. In the third part, the methodology was validated on a practical case study. Using the developed method, the energy-efficiency of a badminton robot was improved. The problem under study was a minimization of consumed energy subject to a time constraint. A novel approach was presented in which this problem was treated as a multi-objective problem which was then transformed to a single-objective criterion using desirability functions. The energy consumption was reduced by 5% compared to the current implemented energy-efficient solution. Furthermore -by applying more stringent time constraints- the precision of the system could be improved to the maximum precision possible, but with a reduction in energy consumption of 52% compared to the current maximum precision implementation. In conclusion, this work presents the extensions necessary for the Online Sequential Process Improvement methodology to deal with contemporary processes and shows the potential of using the methodology under practical conditions. Furthermore, a software package was developed that allows for the fast execution of the improvement methods, which was shown on a practical case study.