Drift-Insensitive Features for Learning Artificial Olfaction in E-Nose System

• Published in: IEEE sensors journal 2018-09, Vol.18 (17), p.7173-7182
• Main Authors: Ur Rehman, Atiq, Bermak, Amine
• Format: Article
• Language: English
• Subjects: Calibration; Classification; Classification algorithms; Computer memory; Cosine similarity; Detection; Drift; electronic nose; Electronic noses; Feature extraction; Gas detectors; Gas sampling; Gas sensors; Gases; Heuristic methods; Machine learning; particle swarm optimization; Performance degradation; Sensor arrays; sensor drift; Sensor systems; Sensors; Uniqueness
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2018.2853674

Domain features and independence maximization are proposed recently for learning the domain-invariant subspace to handle drift in gas sensors. The proposed domain features were the acquisition time and a unique device label for the collected gas samples. In real-time applications of gas sensing, a sample is usually collected using a multi-sensor sensing approach, so a unique device label is not possible in that case, which results in performance degradation. Similarly, semisupervised learning algorithms are proposed to handle drift for gas sensing applications, but getting data from the target domain for the calibration of the system is not always possible. To address these problems, this paper proposes a novel approach to handle the drift in gas sensors, with the following merits: 1) a new classification system based on cosine similarity is developed and features are exploited using a metaheuristic; the outcome is drift-insensitive features that are capable of handling drift in gas sensors; 2) the proposed system is robust against the drift without requiring any re-calibration, domain transformation, or data from target domain; 3) the classification system is an integration of two classifiers; this enables the system to outperform other baseline methods; and 4) only median values of drift-insensitive features are used for learning, so the system requires very few memory cells for storage. The proposed system is validated against a large-scale data set of 13910 samples from six gases, with 36 months' drift and has demonstrated 86.01% classification accuracy, which is 2.76% improvement, when compared with other state-of-the-art methods.