Wavelet transform and fuzzy ARTMAP-based pattern recognition for fast gas identi®cation using a micro-hotplate gas sensor E. Llobet a,* , J. Brezmes a , R. Ionescu a , X. Vilanova a , S. Al-Khalifa b , J.W. Gardner b , N. Ba Ãrsan c , X. Correig a a Department of Electronic Engineering, Universitat Rovira i Virgili, Avda. Paõ Èsos Catalans 26, Campus Sescelades, 43007 Tarragona, Spain b School of Engineering, Warwick University, Coventry CV4 7AL, UK c Institute of Physical Chemistry, Tu Èbingen University, Auf der Morgenstelle, 8 Tu Èbingen, Germany Abstract It is shown that a single thermally-modulated tin oxide-based resistive microsensor can discriminate between two different pollutant gases COandNO 2 )andtheirmixtures.Themethodemploysanovelfeature-extractionandpatternclassi®cationmethod,whichisbasedona1-D discrete wavelet transform and a Fuzzy adaptive resonant theory map ARTMAP) neural network. The wavelet technique is more effective than FFT in terms of data compression and is highly tolerant to the presence of additive noise and drift in the sensor responses. Furthermore, Fuzzy ARTMAP networks lead to a 100% success rate in gas recognition in just two training epochs, which is signi®cantly lower than the number of epochs required to train the back-propagation network. # 2002 Elsevier Science B.V. All rights reserved. Keywords: Discrete wavelet transform; Fuzzy ARTMAP; Micro-hotplate gas sensor; Tin oxide; Additive white noise; Sensor drift 1. Introduction SnO 2 sensors are inexpensive and highly sensitive to a broad spectrum of gases, including atmospheric pollutants such as CO, NO 2 and H 2 S [1±3]. However, well know disadvantages are their lack of selectivity and drift [4], which explain why these sensors are mainly used in low- cost alarm-level gas monitors for domestic and industrial applications [5]. A strategy to enhance selectivity consists of modulating the sensor's working temperature. When the sensor operating temperature is modulated, the kinetics of adsorption and reaction that occur at the sensor surface in the presence of atmospheric oxygen and other reducing or oxidising species are altered. This leads to sensor response patterns that are characteristic of the species present in the gas mixture [6±8]. Many examples of this approach can be found in the literature where the gas sensors employed can be conventional e.g. TGS type [7]) or micromachined [9±12]. In most of these studies, the fast Fourier transform FFT)isusedtoextractimportantfeaturesfromtheacsensor signal together with a back-propagation neural network for predictive classi®cation. In this work we show that the use of a novel feature- extraction technique like the discrete wavelet transform DWT) that replaces FFT coupled to a Fuzzy adaptive resonant theory map ARTMAP) neural network leads to better results. The performance of this technique in the presence of arti®cially generated noise and drift has also been studied. The DWT-based method was found to be highly tolerant to noise and sensor drift. 2. Experimental 2.1. Sensor The device comprises an inert micro-hotplate substrate with a Pt heating resistor sandwiched between two thin silicon nitride layers and a pair of gold electrodes on top. The membrane thickness was about 0.6 mm and the active hot) area of 250 mm  500 mm.Thegassensitivelayerwas formed by depositing a Pd-doped SnO 2 paste onto the electrode area. Finally, on-chip annealing was performed at 450 8C in air see [13] for further details). The micro- hotplate structure provides millisecond thermal response times and milliwatt power consumption required for hand- held units. Sensors and Actuators B 83 2002) 238±244 * Corresponding author. Tel.: 34-977-559623; fax: 34-977-559605. E-mail address: ellobet@etse.urv.es E. Llobet). 0925-4005/02/$ ± see front matter # 2002 Elsevier Science B.V. All rights reserved. PII:S0925-400501)01047-4