(223b) Anomaly Detection Algorithms to Utilize Non-Injective Gas Sensor Arrays

Gas sensor arrays provide an information-rich response pattern to discriminate among many gas compositions. However, the mapping of gas compositions to sensor array response vectors is often non-injective (not one-to-one), preventing their application for quantitative gas sensing tasks. Here, we show non-injective gas sensor arrays may still be useful in an anomaly detection setting.
We demonstrate computationally by (i) considering a sensor array composed of two gravimetric sensors based on zeolitic imidazolate frameworks; (ii) generating a synthetic (gas composition, sensor array response) data set pertaining to fruit ripening rooms, where CO₂, C₂H₄, and H₂O vary in concentration with anomalies in CO₂ and C₂H₄ concentrations; (iii) training and evaluating a one-class support vector machine as the anomaly detector. We show how the anomaly detection performance diminishes with the magnitude of the variance of (i) the background humidity levels and (ii) measurement noise.