Hybrid virtual sensor based on RBFN or SVR compared for an embedded application

Abstract

In numerous practical applications the interesting measurands are not explicitly available by existing sensors or unaffordable due to high cost of explicit sensor principle. Virtual sensors, as one particular means to design intelligent integrated sensory systems (I2S2), offer an solution to this problem, by merging various sources of information to generate the desired measurand for the given environmental stimuli. In this paper, radial-basis-function-networks (RBFN) and support-vector-regression (SVR) are compared for knock-detection in combustion engines with regard to ease of learning, generalization, and resource-efficiency. Additionally, the notion of a hybrid virtual sensor (HVS) is introduced here for invariance and complexity reasons. In our experiments, real-world engine data has been applied for method comparison and recommendations for parameter settings. SVR shows better generalization results than RBFN for the criteria correlation coefficient and absolute mean error are applied. In future work, we will integrate HVS concept in our emerging tool for automated I2S2 design. © 2011 Springer-Verlag.