A framework for factory-frained virtual sensor models based on censored production data

Authors Sabrina Luftensteiner
Michael Zwick
Editors Sven Hartmann
Josef Küng
Gabriele Kotsis
A Min Tjoa
Ismail Khalil
Title A framework for factory-frained virtual sensor models based on censored production data
Booktitle DEXA 2020: Database and Expert Systems Applications, Part II
Type in proceedings
Publisher Springer
Series Lecture Notes in Computer Science
Volume 12392
ISBN 978-3-030-59002-4
DOI 10.1007/978-3-030-59051-2_1
Month September
Year 2020
Pages 3-16
SCCH ID# 20044

In Industrial Manufacturing, the process variable is often not directly observable using a physical measuring device. For this reason, Virtual Sensors are used, which are surrogate models for a physical sensor trained on previously collected process data. In order to continuously improve such virtual sensor models, it is desirable to make use of data gathered during production, e.g. to adapt the model to client-specific tools not included in the basic training data. In many real-world scenarios, it is only possible to gather data from the production process to a limited extent, as feedback is not always available. One example of such a situation is the production of a workpiece within required quality bounds. In case the finished workpiece meets the quality criteria or else is irreversibly damaged in the process, there is no feedback regarding the model error. Only in the case where the operator is able to specify a correction to reach the desired target value in a second run, the correction error can be used as an approximation for the model error.

To make use of this additional censored data, we developed a framework for preprocessing such data prior to the training of a virtual sensor model. The approach used is independent from the chosen sensor model as well as the method used for correction of the censored data. We tested our approach in combination with three different correction methods for censored data (Tobit regression, constrained convex optimization, OptNet neural network optimization layer) using data gathered in a real-world industrial manufacturing setting. Our results show that including the data can approximate the uncensored data up to a censorship percentage of 60% while at the same time improving the performance of the virtual sensor model up to 30%.