An Improved Data Representation for Smoking Detection with Wearable Respiration Sensors

摘要

This paper proposes a peak-based representation for smoking detection from RIP data as a more robust alternative to the existing segmentation-based representation. The rational for preferring this approach in the presence of noise is that there is much less uncertainty about the location of the peaks in the RIP data than there is about the location of segmentation boundaries. To implement this approach, we first run a peak detection algorithm on the RIP waveform data to extract the peaks (maximum inhalation) and valleys (maximum exhalation) from the input time series. We then extract feature vectors analogous to those used in the segmentation case [1], but defined using relative locations and amplitudes of peaks and valleys. We associate puff (positive) and non-puff (negative) labels with peaks, and learn a classification model to predict the peak labels given the associated feature vectors.