Optimization of a Data-Driven Soft Sensor for Use Cases with Multiple Dimensions

Supplementing physical sensors with soft sensors is a proven way to support the rapidly advancing digitalization in production halls, especially in cases where it is very complicated to record the desired parameter. For such an implementation to succeed, the soft sensor design must be well optimized for the specific use case. This work focuses on adapting a soft sensor in the best possible way for multiple dimensions. Feature selection therefore represents a large part of this. Hence, it is a matter of optimizing the use of the available dimensions or parameters. To achieve this and to make the result meaningful for a wide range of applications, this work concentrates on a total of three decision criteria: The error, the computation time and the accuracy per time. In contrast, most existing feature selecting algorithms are limited to error minimization only. Furthermore, this error minimization is usually measured on the basis of training data. The approach presented here applies the model created with training data to separated test data, which is much more close to reality and leads to more accurate results in the end.