Frequency response based permittivity sensors for measuring air contaminants

Master of ScienceDepartment of Biological and Agricultural EngineeringNaiqian ZhangPermittivity, displayed when a dielectric material is exposed to an electric field, is a useful property for measuring impurities in a dielectric medium. These impurities often have a dipole moment different from the pure material, and the dipoles align through polarization and impede electric current. By measuring the resulting impedance in a known geometry, the permittivity can be determined. Four permittivity sensors were utilized to measure contaminants that are associated with biofuels, specifically glycerol, ethanol, and ammonia. These sensors were based around either stainless steel or aluminum plates to ensure durability and reliability. By connecting each of these sensors to a signal generating control box, the gain and phase can be measured at 609 frequencies, from 10 kHz up to 120 MHz. Data from each of the three contaminants were run through a method for detection. Measurements for ambient air and air with the contaminants were compared with a statistical analysis. Glycerol, ethanol, and ammonia each had significantly different measurements in the gain and phase data at a unique set of frequencies. Using a neural network analysis for detection resulted in a 95.8%, 93.9%, and 97.1% success rate for detecting glycerol, ethanol, and ammonia, respectively. For ethanol and ammonia, where multiple concentrations were measured, regression methods were used to relate the frequency response data to the contaminant concentration. Stepwise regression, wavelet transformation followed by stepwise regression, partial least squares regression, and neural network regression were the four methods used to establish these relationships. Several regressions over-fit the data, showing coefficient of determination (R[superscript]2) values of 1.000 for training data, yet very low R[superscript]2 values for validation data. However, the best R[superscript]2 values of all the regressions were 1.000 and 0.996 for the training and validation data, respectively, from measuring ammonia