• Agi, Kamil
• Garzon, Fernando H.
• Halley, Sleight

Abstract

<jats:p>Methane emissions from natural gas pipeline infrastructure contribute to over 2 billion dollars in losses in the US and 1.6 x 10<jats:sup>8</jats:sup> metric tons of CO<jats:sub>2</jats:sub>-equivalent per year.[1-2] Widely deployable sensors are needed to quickly identify and repair leaks so that their environmental and economic impact can be minimized. Mixed potential electrochemical sensors (MPES) are a low cost, highly robust system which can be deployed in the field. Mixture identification and quantification using sensor signals produced by MPES devices is challenging due to cross interference effects and complicated by the presence of interferent methane sources including wetlands and agricultural resources.</jats:p><jats:p>We have developed a machine learning system with artificial neural networks using MPES devices as inputs and either concentration or mixture label as outputs. We have previously demonstrated using this approach to automatically decode sensor signals associated with automotive emissions gas mixtures [3-4]. Artificial neural networks created using the Google Tensorflow library for the Python programming language were trained on MPES sensor signals to identify and quantify mixtures as natural gas, CH<jats:sub>4</jats:sub> only, and a CH<jats:sub>4</jats:sub>+NH<jats:sub>3</jats:sub> mix as a simulant for bovine emissions. The results we have collected showed &gt;99% test accuracy in identification of these gas mixtures as shown in the confusion matrix in Figure 1 in the concentration range of 1000-4000 PPM CH<jats:sub>4</jats:sub> and 75-255 PPM NH<jats:sub>3</jats:sub>. Optimization of the ANN architecture to minimize processing time and facilitate deployment of machine learning onto portable hardware will also be presented. We will also compare the performance of ANNs to alternate ML techniques.</jats:p><jats:p>This work was supported by the US Department of Energy under award DE-FOA-0031864.</jats:p><jats:p>[1] A.J. Marchese, T. L. Vaugh, D. J. Zimmerle, D. M. Martinez, L. L. Williams, A. L. Robinson, A. L. Mitchell, R. Subramanian, D. S. Tkacik, J. R. Roscioli, S. C. Herndon. <jats:italic>Science</jats:italic>. 7204 (2018).</jats:p><jats:p>[2] Kort, E. A.; Frankenberg, C. Costigan, K.R.; Lindenmaier, R.; Dubey, M. K; Wunch, D. Geophys. Res. Lett, 2014, 41, 19, 6898</jats:p><jats:p>[3] L. Tsui, A. Benavidez, P. Palanisamy, L. Evans, F. Garzon, Electrochim. Acta 283 (2018) 141–148.</jats:p><jats:p>[4] L. Tsui, A. Benavidez, P. Palanisamy, L. Evans, F. Garzon, Sens. Act. B: Chem. 249 (2017), 673-684</jats:p><jats:p>Figure 1. Confusion matrix for (a) natural gas simulant, (b) CH4, (c) NH3, and (d) CH4+NH3. Mixtures (b) and (d) serve as surrogates for wetlands and agricultural bovine emissions respectively.</jats:p><jats:p><jats:inline-formula><jats:inline-graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="1315fig1.jpg" xlink:type="simple" /></jats:inline-formula></jats:p><jats:p>Figure 1</jats:p><jats:p />

Topics

• impedance spectroscopy
• machine learning