IMPROVEMENT OF SURFACE-ACOUSTIC-WAVE GAS SENSOR RESPONSE-TIME USING NEURAL-NETWORK PATTERN-RECOGNITION

Surface acoustic wave (SAW) gas-sensor signal processing may allow fir st, detection of gases, secondly, their identification and thirdly, if possible, their quantification. For a few years now, pattern-recognit ion techniques using artificial neural networks have been applied to s ensor arrays with promising results. Nevertheless, data sets needed fo r these techniques are always built with well-established and stable s ensor responses. Sometimes, the SAW gas-sensor response times are quit e long due to kinetic factors concerning the gas adsorption. Moreover, in some applications, such as military or safety, the gases involved are extremely dangerous or toxic. The detection speed is hence an esse ntial parameter. Thus, we are developing a neural-network-based signal -processing system that aims to allow dynamic gas detection: the senso r steady-state response does not need to be reached to ascertain the p resence of gas. This signal-processing system includes three steps. Fi rst, the gas-sensor output is pre-processed in order to extract some c haracteristic parameters. These then constitute the input pattern of a three-layer neural network trained with the back-propagation learning rule. Finally, its output is post-processed to decide whether or not to turn an alarm on. Up to now, we have only used the results given by one sensor consisting of a dual SAW delay-line oscillator for NOx sen sing. We propose a pre- and a post-algorithm. Results are presented an d discussed. In particular, we emphasize that some attention has to be given to the constitution of the data set and to the definition of th e neural-network performance.