Mølgaard, Lasse L.Buus, Ole T.Larsen, JanBabamoradi, HamidThygesen, Ida L.Laustsen, MilanMunk, Jens KristianDossi, EleftheriaO'Keeffe, CarolineLässig, LinaTatlow, SolSandström, LarsJakobsen, Mogens H.2017-05-232017-05-232017-05Molgaard LL, Buus OT, Larsen J, et al., (2017) Improved detection of chemical substances from colorimetric sensor data using probabilistic machine learning. In: Chemical, Biological, Radiological, Nuclear, and Explosives (CBRNE) Sensing XVIII, SPIE Defense + Security, 2017, 9-13 April 2017, Anaheim, California, USAhttps://doi.org/10.1117/12.2262468http://dspace.lib.cranfield.ac.uk/handle/1826/11918We present a data-driven machine learning approach to detect drug- and explosives-precursors using colorimetric sensor technology for air-sampling. The sensing technology has been developed in the context of the CRIM-TRACK project. At present a fullyintegrated portable prototype for air sampling with disposable sensing chips and automated data acquisition has been developed. The prototype allows for fast, user-friendly sampling, which has made it possible to produce large datasets of colorimetric data for different target analytes in laboratory and simulated real-world application scenarios. To make use of the highly multi-variate data produced from the colorimetric chip a number of machine learning techniques are employed to provide reliable classification of target analytes from confounders found in the air streams. We demonstrate that a data-driven machine learning method using dimensionality reduction in combination with a probabilistic classifier makes it possible to produce informative features and a high detection rate of analytes. Furthermore, the probabilistic machine learning approach provides a means of automatically identifying unreliable measurements that could produce false predictions. The robustness of the colorimetric sensor has been evaluated in a series of experiments focusing on the amphetamine pre-cursor phenylacetone as well as the improvised explosives pre-cursor hydrogen peroxide. The analysis demonstrates that the system is able to detect analytes in clean air and mixed with substances that occur naturally in real-world sampling scenarios. The technology under development in CRIM-TRACK has the potential as an effective tool to control traf- ficking of illegal drugs, explosive detection, or in other law enforcement applications.Attribution-NonCommercial 4.0 Internationalhttp://creativecommons.org/licenses/by-nc/4.0/Conference paper