No.11（2016）9.Output estimation for multi-dimensional nonlinear sensors using support vector regression

Chisato Murakami，Yasuaki Kaneda

  Sensor principle modeling and stochastic modeling are known mathematical approaches used in estimating outputs of a sensor. These models can express sensor characteristics well. However, a modeling precision of the sensor principle model is subject to constraints depending on the number of observable physical quantities. On the other hand, the stochastic approach allows models to be constructed with only input and output information in an assembled sensor without facing such constraints. In this study, we investigated accuracies of output estimations using a stochastic model. Support vector regression (SVR), which is a powerful tool used for nonlinear regression, was used as the stochastic model. A multi-dimensional nonlinear displacement sensor with four degrees of freedom was used as a case study for the sensor output estimation.
 Average full-scale errors in the output estimation were 2.42% and 4.59% using SVR and the sensor principle model, respectively. In evaluating the full-scale errors, we verified that SVR is a better technique for estimating the output of a multi-dimensional nonlinear sensor.

Keywords
Capacitive sensors, Displacement measurement, Parameter estimation, Support vector regression, Multi-dimensional nonlinear sensor

Back to Contents