Abstract— Artificial olfaction is an emerging application field for machine learning practitioners. In this work, we propose a holistic approach to pattern classification in electronic noses applications. Specifically, we show how classification results based on a complete measurement cycle can be combined with an assessment provided by real time classifiers acting on the single instantaneous measurement sample. A running classification confidence measure allows for obtaining fast and reliable outcomes. A safety critical scenario has been selected for the testing of the proposed pattern analysis strategy involving the identification and discrimination of surface contaminants on composite panels in pre-bonding Non Destructive Tests (NDT) during lightweight aircraft assembly. A reject option has been introduced to refuse low classification confidence panels improving both FP and FN rates. Results show how this strategy can efficiently exploit two different views of the electronic nose olfactive fingerprinting process that are currently seen as alternative. Index Terms— Artificial olfaction; Electronic noses; Reject Option; Classifiers Combination; Non Destructive Test. I. INTRODUCTION Electronic noses are intelligent chemical multi-sensors devices that have been applied to several fields where chemical mixture detection, identification and quantification are concerned. Based on array of chemical sensors, they provide for multivariate measurements that can represent an electronical counterpart of biological olfactive fingerprints [1]. Several attempts have been made to develop pattern recognition systems able to extract semantic contents from e- nose response patterns for classification or regression applications. Many of these have then been applied in commercial solutions in different applicative scenarios [2]. A significant amount of e-nose pattern recognition systems relies on a fixed set of descriptive features that are computed by using steady state response of the e-nose’s chemical sensors. These are collected within a measurement procedure that typically encompass 1. a baseline response acquisition phase, 2. a response acquisition phase during which the sensors are exposed to the relevant chemicals mixture, 3. a flushing phase in which the sensors are exposed to filtered air again. The authors are with ENEA, Italian Agency for New Technologies, energy and Sustainable Economic Development, P.le E.Fermi, 1, 80055 Portici (NA) (e-mail: maria.salvato@enea.it; saverio.devito@enea.it). During the last decades several descriptors have been devised to enhance discriminative power of e-nose basic response patterns [3][21][23]. Many of them have shown that transient phase holds significant information that could be exploited for enhancing e-nose overall performances ([4],[5],[23]). The introduction of Ion Mobility Spectrometers (IMS) in artificial olfaction and the application of temperature modulation procedures to Metal Oxide chemical sensors (MOS) as well as voltage scanning in voltammetric sensors have led to the development of descriptive features and PARC systems capable to deal with spectra-like responses ([6], [7]). As opposed to the use of fixed measurement procedures, during the last decade, continuous measurement techniques with open sampling have been proposed. These are particularly interesting for on-field operating systems usually requiring a fast and continuous evaluation of the olfactive patterns. Moreover, the need for portable, low cost, battery operated devices has increased the adoption of continuous operating e-noses since the classic approach requires a significant amount of economically and energetically costly devices such as filters, carrier gas reservoirs and pumps [8]. As emerging from multiple reviews, currently, the two approaches are seen as alternatives in literature. In contrast, in this work we propose their combined use as complementary “views” of the same chemical sensing process [3][22]. In our view, on the hypothesis that they individually provide sufficient and independent information, their combined use may leads to increased performance. In facts, features extracted by using the overall response to a fixed measurement procedure give a concise description of the sensor response trajectory in the multivariate space [5]. However, discriminatory power is well present in each instantaneous response sample due to the peculiar ratio among the single sensors response levels [24][25]. Moreover, the use of instantaneous responses may allow for the user to obtain rapid assessment during the measurement procedure execution. The full execution of the procedure may be avoided if the assessment of the stream of successive samples produce consistent classification estimations. In this paper, we propose an holistic approach to electronic nose response pattern analysis built up by combination of multiple pattern recognition systems that are able to combine the two above mentioned “views” to produce enhanced performance indexes [9-10]. Finally, a classification confidence measure is also provided and used to reject low confidence samples requiring a more careful inspection [11]. A demanding scenario as Non Destructive Test (NDT) for aerospace industry has been selected as an application field for this proposal. In this specific scenario, the e-nose is required to identify and classify surface contaminations on composite MARIA SALVATO, SAVERIO DE VITO, ELENA ESPOSITO, ETTORE MASSERA, MARA MIGLIETTA, GRAZIA FATTORUSO, GIROLAMO DI FRANCIA An holistic approach to e-nose response patterns analysis. An application to non-destructive tests This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/JSEN.2015.2513818 Copyright (c) 2016 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.