Active classification with arrays of tunable chemical sensors

a r t i c l e i n f o This paper presents Posterior-Weighted Active Search (PWAS), an active-sensing algorithm for classification of volatile compounds with arrays of tunable chemical sensors. The algorithm combines concepts from feature subset selection and sequential Bayesian filtering to optimize the sensor array tunings on-the-fly based on information from previous measurements. Namely, the algorithm maintains an estimate of the posterior probability associated with each chemical class, and updates it sequentially upon arrival of each new sensor observations. The updated posteriors are then used to bias the selection of the next sensor tunings towards the most likely classes, in this way reducing the number of measurements required for discrimination. We characterized PWAS on a database of infrared absorption spectra with 250 analytes, and then validated it experimentally on an array of metal-oxide sensors. Our results show that PWAS outperforms passive-sensing approaches based on sequential forward selection, both in terms of classification performance and robustness to noise in sensor measurements. Chemical sensors are generally used as first-order devices, where one measures the sensor's response at a fixed setting, e.g., absorption of an optical sensor at a specific wavelength, or conductivity of a solid-state sensor at a specific operating temperature [1]. In many cases, additional information can be extracted by modulating some internal property of the sensor. As an example, measuring the conductivity of a metal-oxide chemical sensor at different temperatures can provide a wealth of discriminatory information [2]. However, this additional information comes at a cost, such as sensing times or power consumption. For this reason, feature subset selection (FSS) techniques are commonly used to identify a subset of the most informative sensor configurations. Over the past decade, a handful of investigators in the chemical sensor community have explored active sensing as an alternative to FSS [3–7]. In contrast with FSS, where the sensor configurations are optimized off-line, active sensing adapts the sensor configurations in real-time based on information obtained from previous measurements. In previous work [6,7], we showed that active sensing can achieve higher classification performance than FSS with fewer measurements and provides a trade-off between sensing costs and classification performance. Unfortunately, these active-sensing methods were developed for individual sensors, and do not scale up to sensor arrays. First, the number of operating configurations for a sensor array grows exponentially with the size of the array; given an N-sensor array with D configurations per sensor, …