Applications of the Kalman Filter to Chemical Sensors for Downstream Machine Learning

• Published in: IEEE sensors journal 2018-07, Vol.18 (13), p.5455-5463
• Main Authors: Weiss, Matthew, Wiederoder, Michael S., Paffenroth, Randy C., Nallon, Eric C., Bright, Collin J., Schnee, Vincent P., McGraw, Shannon, Polcha, Michael, Uzarski, Joshua R.
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Chemical and biological sensors; Chemical sensors; Chemicals; Classification; Coated electrodes; Kalman filters; Machine learning; Organic chemistry; Organophosphates; Power consumption; Preprocessing; Remote sensors; Sensor arrays; Sensor phenomena and characterization; Sensors; Signal processing; Time series analysis
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2018.2836183

Chemical sensors play an important role in a variety of civilian and military domains. In these contexts, the ability to accurately and quickly identify chemical agents is of utmost importance. In practice, constraints on physical footprint, power consumption, ease of use, and time required for accurate detection often restrict the utility of sensors, particularly in remote and isolated regions. One solution to address this problem is the engineering of advanced signal processing techniques, which decrease the time required for accurate detection. This allows software to facilitate the construction of hardware that meet stringent power and concept of operations guidelines. In this paper, we propose the Kalman filter as a preprocessing technique applicable to chemical sensor time series data for downstream machine learning. Using data collected from a sensor array of multiple unique polymer-graphene nanoplatelet coated electrodes, we show accurate and early detection of both organophosphates and interferents is improved when the Kalman filter is used as a preprocessing technique. In particular, within two seconds of analyte exposure to the sensor array, classification using Kalman filtered first derivative estimates achieve an error of less than 10%. By comparison, the non-Kalman filtered data set has a classification error rate above 40% within this time. An advantage of our approach is classification does not depend on a set parameter, such as maximum resistance change, or a pre-determined exposure time, and which allows rapid classification immediately after analyte introduction.