Author

Zongyu Geng

Date of Graduation

2015

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Industrial and Managements Systems Engineering

Committee Chair

Feng Yang

Committee Co-Chair

Wafik Iskander

Committee Member

Majid Jaridi

Committee Member

Natalia A Schmid

Committee Member

Nianqiang Wu

Committee Member

Feng Yang

Abstract

Traditional sensor calibration is restricted to mathematically relating the steady-state sensor responses to the target analyte concentrations to realize environment monitoring. However, commonly-used chemical sensors usually require a relatively long time, on the order of minutes, to reach steady-state operation, and exhibit nonlinear drifting behaviors. To achieve real-time monitoring of rapidly-changing environments while accommodating drifting behaviors, this work develops statistical methods for both forward descriptive calibration and inverse dynamic calibration of sensor arrays.;Forward calibration is performed based on experimental data. In this work, multivariate Gaussian processes (GPs) were adapted to obtain the forward calibration model, which quantifies the sensor response as a function of the analyte concentration, the drifting variables, and the sensors' exposure time. The multivariate GP method synergistically models all calibration data collected under a range of drifting conditions, and seeks to produce the calibration model of highest quality with the given experimental data. The forward calibration model is a descriptive model, relating sensors' time-dependent responses to a static environment specified by several variables, hence it is not able to assist in real-time monitoring of rapidly-changing environments.;To achieve real-time monitoring of analyte concentrations while fully utilizing the efficiency of forward calibration rendered by multivariate GPs, an inverse calibration method was developed. This inverse model takes the form of a transfer function regression, infers the time-varying analyte concentrations from the dynamic sensor responses, and thus can be coupled with sensors for real-time monitoring. The inverse transfer function model is estimated from the pseudo-calibration data generated by the forward multivariate GP model, which captures the sensors' dynamic and drifting behaviors as reflected in the real experimental data.;Simulated sensor arrays have been developed from real sensor data, and were used to demonstrate the calibration methods developed in this work.

Share

COinS