An Automatic Feature Extraction Method for Gas Sensors Based on Color-Enhanced Phase Space

Abstract

Aiming to improve the effectiveness and the identity of features extracted from gas sensor responses, a novel automatic feature extraction method is proposed and studied. A simple color-enhanced phase-space approach is proposed to convert the dynamic gas sensor signals into images, which emphasizes the internal features of phase space. A lightweight neural network, i.e., MobileNetV2, is adopted to automatically extract the features and classify the odors. The method has been embedded into a lab system to classify the freshness of yellow peaches, and the final freshness classification accuracy reaches 98.58%, which is more than 20% improvement of average classification accuracy than the traditional time domain or frequency domain feature extraction and recognition methods. Compared to the original phase space, more than 10% improvement in average classification accuracy has also been obtained.