Introduction Since the introduction of gas sensor arrays in 1982 [1], they have been tested in numerous fields but still suffer from weak sensitivity and selectivity in gas mixtures [2]. To account for that, the focus lays on new evaluation algorithms (e.g. artificial neural networks with deep learning) following recent advances in computing capabilities [3]. However, often the underlying hardware (i.e. sensors) is the limiting factor. As a result, the often applied “black box” approach correlating sensor signals with chemical perception (e.g., woody taste of wine) holds high risk of bogus correlations as the relevant analyte, responsible for the actual odor, aroma or disease, might not be generating the measured sensor outputs in the first place [2].Here, the effect of orthogonality by selective material design on array performance is investigated experimentally with a basic multivariate linear regression model. This is done by systematically replacing non-selective SnO2-based sensors in an array by distinctly-selective Si:MoO3 and Si:WO3 for ammonia [4] and acetone [5], respectively. The arrays are tested with 60 random combinations of ammonia, acetone and ethanol in the ppb to low ppm range at dry conditions. Next, array size and composition is varied systematically to understand the contribution of each sensor on analyte prediction. Finally, the stability of the extracted regression coefficients is evaluated. Method Flame spray pyrolysis was used to generate sensing materials with distinct selectivities (i.e. Pt:SnO2, Pd:SnO2, Si:SnO2, Si:MoO3 and Si:WoO3). These sensors were combined to arrays and applied to 60 different gas mixtures consisting of randomly combined concentrations of ammonia (250, 500, 800, 1200, 1600 and 2000 ppb), acetone (250, 400, 600, 800, 1200 and 1800 ppb) and ethanol (50, 100, 150, 200, 400 and 600 ppb) corresponding to typical human breath concentrations. In specific, a purely SnO2-based (i.e. Pd:SnO2, Pt:SnO2 and Si:SnO2) array and one with distinctly different sensing materials (i.e. Si:WO3, Si:MoO3 and Si:SnO2) were chosen, denoted as conventional and orthogonal array, respectively. A multivariate linear regression was applied to predict the analyte concentrations, whereas a 6-fold cross-validation was used to separate the data points into calibration and test sets to assess the performance and prevent overfitting. The effect of array size and composition was investigated by evaluating the significance and stability of the regression coefficients [6] and quantifying the accuracy and precision of the arrays. Results and Conclusions Figure 1 shows the average estimation error of the conventional (Pd:SnO2, Pt:SnO2 and Si:SnO2) and orthogonal (Si:WO3, Si:MoO3 and Si:SnO2) arrays in 60 different 3-analyte gas mixtures. The conventional array estimates all analytes with large average errors. For example, for ammonia an average error of 687 ppb is obtained whereas the actual concentrations ranged from 250 to 2000 ppb. Apparently, the influence acetone and ethanol at varying concentrations prevented an accurate ammonia estimation. This is similar when estimating acetone or ethanol with smaller but still significant errors of 177 and 158 ppb, respectively. In contrast, the orthogonal array estimates all analyte concentrations more accurately. In fact, average estimation errors of 123, 45 and 67 ppb are obtained for ammonia, acetone and ethanol, respectively [7].Most likely, the conventional array with the non-specific SnO2 sensors result in a high degree of collinearity. Such arrays feature low discrimination power resulting in weak analyte estimations and strong sensitivity to small changes in background gases, humidity or temperature [8], as demonstrated experimentally here. When exchanging such collinear sensors (e.g, Pt:SnO2 and Pd:SnO2) with distinctly selective ones (Si:WO3 and Si:MoO3) the degree of orthogonality is increased resulting in accurate array estimations.As a result, we provide experimental evidence that sensor array performance depends largely on the orthogonality and thus selectivity of its constituent sensors. Utilizing distinctly selective sensors in arrays improved discrimination power and stability of the regression coefficients, necessary for accurate and precise analyte predictions in gas mixtures. Such orthogonal arrays are needed in breath analysis [9] or human search and rescue [10] due to the low concentrations and high number of interferants.
Search from 250 M+ research papersSearch from 250 M+ research papers
Oct 15, 2023
Mingguang Yu + 6
Open Access