Feature-Level Fusion by Multi-Objective Binary Particle Swarm Based Unbiased Feature Selection for Optimized Sensor System Design

The performance of recognition systems in intelligent sensor technology can often be improved by using the combined information of several different measurement results, i.e., signal processing and feature computation, of single-sensor and/or multi-sensor information. However, the large dimensional data caused by feature-level fusion imposes an accuracy problem on classification tasks, due to some irrelevant and/or redundant features. For this reason, feature selection as an important dimensionality reduction method is applied according to certain assessment measures. The standard use of a single assessment, e.g., overlap measure, can be unsatisfactory in performance and potentially suffer from saturation effect of the chosen measure. In this paper, we investigated an aggregation approach, i.e., a multi-objective approach using binary particle swarm optimization and leave-one-out method for unbiased feature selection (LOOFS) to find an optimal feature subset in feature-level fusion. The approach was applied to several benchmark problems of increasing complexity and for the most difficult problem the LOOFS approach proved to be superior. Additionally, we investigated the possibility of LOOFS, as an assessment, to obtain the statistical information of the selection-stability. The developed criterion indicates the aptness of feature selection for the regarded task and data and allows improved closed-loop optimization for optimized sensor system design