Differentiation between Measurement Uncertainty and Model Error in the Description of Measurement Processes

Research Project "RENUMBER" to Start Soon

 

Measurement processes are always subject to uncertainty so that the statements derived from the measurement data are also uncertain. If the measurement uncertainty in a product inspection is too high and the characteristic under consideration is close to the specification limits, the decision as to whether the characteristic is within or outside the specification is fraught with risk. Thus, correctly produced parts can be mistakenly rejected, or defective parts can be mistakenly released.

Since the risk of incorrect decisions cannot be determined and thus controlled without knowledge of the measurement uncertainty when collecting measurement data, measurement values without measurement uncertainty information are worthless. To determine the measurement uncertainty, the so-called model of the measurement is needed. With this model, the measurement uncertainty is determined based on the natural variations of the input variables. Systematic deviations within the measurement process are taken into account when determining the measurement uncertainty, provided that they cannot be eliminated. Deviations between reality and model in the sense of a model error, however, are often not taken into account when determining the measurement uncertainty. As a result, the specified measurement uncertainty may be overestimated and costs may be incurred due to the incorrect specification of the manufacturing tolerance range.

This problem is to be solved in the future. In the research project "RENUMBER" at the Laboratory for Machine Tools and Production Engineering (WZL) of RWTH Aachen University, a method will be developed over the next two years that addresses the differentiation between measurement uncertainty and model error. The procedure is based on a Bayesian approach and is divided into four steps: the selection of the a priori distribution of the model and error parameters, the model building using machine learning considering the a priori distributions, the differentiation between model error and uncertainty using the Bayesian approach and the validation of the calculated information.

To make the procedure tangible for practical application, the individual steps are combined into an overall procedure and implemented in a freely accessible programming language. The procedure is validated using several example measurement processes to ensure applicability in an industrial context.

Funding Notice:
The research project "RENUMBER" is funded by the German Research Foundation (DFG) (GZ: SCHM 1856/122-1). The responsibility for the content of this publication lies with the authors.