Search

Measurement systems produce information to support decision making. The added value of measurements is determined by how much the provided information improves decision making as uncertainty about decision consequences is reduced. We discuss through an example the value generation by measurement information in the context statistical decision theory (SDT). This example, a grade change decision, outlines the principle for determining added value in decision making. We discuss separately two value generation mechanisms: the direct value generation by reduced uncertainty on consequences and the indirect value generation through accumulating knowledge about the consequence model parameters. The first mechanism is solvable in analytic form and this allows comparing formulations of SDT in the analysis. We outline the analysis of the indirect value generation and show that even with this simple decision task the analysis is rather complex. We propose that the added-value analysis is the key to optimal design of measurement information systems.

We shall present approaches to optimize the design of a measurement system and to schedule dynamically a versatile measurement resource. The analysis is based on expressing the system management task as a dynamic programming problem in which the system state is partially observable. We shall review the well-known linear-quadratic-Gaussian case, and discuss and give example solutions of discrete state systems with approximate dynamic programming methods to solve such problems in practice. Design problem is then studied by assessing the long-term performance of control and how it depends on properties of measurements. A simple design example is presented for a discrete state system.

Dynamic systems, such as industrial production processes, are measured indirectly with fast on-line measurements. Occasionally, more direct and accurate measurements are made in laboratory to validate the on-line measurements. We discuss the optimal dynamic estimation of the parameters of characteristic curves under the assumption that their validity deteriorates through a stochastic process, describable through a stochastic differential equation. We provide the general formalism and an easy-to-implement algorithm for Gaussian distributions of measurement errors and random-walk deterioration of parameter validity. Furthermore, we analyze the dynamic estimation of the measurand given the earlier information about it deteriorates through (another) stochastic differential equation, and show the effect of uncertainty in the parameters of characteristic curve.