Efficient Sensor Selection for Individualized Prediction Based on Biosignals

Soft sensors combine a hardware component with an intelligent algorithmic processing of the raw sensor signals. While individualization of software components according to a person’s specific needs is comparably cheap, individualization of the sensor hardware itself is usually impossible in mass production. At the same time, the number of raw sensors should be minimum to reduce production costs. In this contribution, we propose to model this challenge as a feature selection problem, which optimizes a feature set simultaneously with respect to a family of functions corresponding to individualized post-processing of sensor signals. This concept is integrated into a number of different classical feature selection schemes, and evaluated in the context of the placement of pressure sensors as part of a shoe insole. It turns out that feature selection respecting the class of functions is superior to both placement based on anatomical considerations and classical feature selection methods.