Enabling a transferable calibration model for metal-oxide type electronic noses

摘要

While electronic noses have been around for over 30 years, little effort has been devoted to the development of transferable calibration models, which are models that can be applied to multiple equivalent devices without adjustment. The majority of published results limit itself to data sets gathered with a single device. This lack of insight in transferable models hampers large scale implementations of eNose-based applications as individual calibration of a multitude of devices for a specific application area is generally unrealistic due to the requirement of actual samples to be measured. For simple gases this may be do-able, but in the case of more complex samples such as biological patient material it is logistically impossible. In this paper we show the influence of the deviation of the sensor temperature on the measurement reproducibility and by inference on the transferability of calibration models. We introduce the total inertia (F_2) as a measure for the heterogeneity within the measured data. The total inertia is an objective measure known from linear algebra, where it is used to calculate the correspondence between matrices. We use 5 micro-hotplate metal-oxide sensors from the same wafer, with an inter-sensor heater temperature difference of approximately 15 ℃ in combination with 2 substances, n-butyl-acetate and hexane. This research demonstrates the increase in heterogeneity of the measured response values in relation to a temperature shift. A shift of 15 ℃ at the sensor surface causes an increase of heterogeneity that is 10-15 times higher than the increase in heterogeneity caused by inter-sensor responses to the substances when operated at exactly the same temperature. Some mixtures of substances will be separable by pattern recognition under virtually any condition, and strict temperature control will neither improve nor deteriorate results. However the significant contribution of temperature deviation toward data heterogeneity renders it plausible that optimized temperature control, and by inference lower data heterogeneity, is a prerequisite for transferability of a calibration model. This holds true when applied to metal-oxide sensors and for mixtures containing substances showing a fair degree of similarity.