Implementation and Optimization of a Data-Driven Soft Sensor in a Welding Process

Soft sensors have been a popular tool for years when it comes to recording certain parameters that are difficult to access. Therefore, in this paper, a basic data-driven soft sensor is implemented in a welding process in order to be able to replace chosen hardware sensors in the future. The focus of this work, however, is on the data pre-processing, which offers numerous opportunities to improve the performance of the soft sensor itself. In the case of process data, as in the example of the welding machine, the clustering of data plays a particularly important role. During this investigation, it was possible to achieve significant improvements in terms of accuracy with an algorithm for clustering the data that was rewritten from scratch. Consequently, the data prepared by PCA is processed with this cluster algorithm and then used for a linear regression model. Furthermore, after the clustering algorithm has been developed and tested in Octave it gets rewritten in Python which is more application-oriented. The calculation time as well as the accuracy of the result are altogether superior compared to alternative approaches.