Microchip Employing Short Period Thermal Modulation for the Detection of H₂S and SO₂ Mixtures

Abstract

Most of the previous works used to improve the performance of single gas sensors by forming the array. However, more discrete sensors would bring higher cost, larger volume, and power consumption. With the technique of micro-electro-mechanical systems (MEMS), a microchip with three sensitive materials (ZIF8-WO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sub> , ZIF8-In <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> O <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sub> , and ZIF8-SnO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ) has been fabricated to detect the mixtures of H <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> S and SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> . By combining the short-period thermal modulation with the periodic sampling, the response curve could be split into several sub-curves, and each sub-curve was regarded as the virtual device operated at a constant temperature. It can reduce power consumption without increasing the physical volume. Based on the fluctuation and distinction of the descriptors, the optimal 24 features were extracted from the initial 144 features. A stacked denoising autoencoder (SDAE) was employed to enhance the generalization ability of Bagging neural networks. Toward the unknown 30 mixtures of H <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> S and SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> , the recognition model under short-period thermal modulation (SPT, 78.67 %) had better performance than that under constant temperature (CT, 62.67 %). SPT (92.63 %) has a smaller non-overlapping area of feature buffers than CT (50.45 %), which might be the physics behind the performance improvement. Certainly, this work has broad application prospects in the future.