Cyclic Temperature Modulation for Lightweight Gas-Mixture Classification via Feature Construction and Response-Phase Decomposition

The recognition of gas mixtures typically demands extensive training data or models with strong representation power, which leads to challenges such as large model size, slow training speed, and the vulnerability to noise. To address these issues, this paper proposes a novel feature construction (FC) via phase-wise aggregation aimed at constructing a gas-mixture recognition model with a simple structure. By leveraging dynamic temperature modulation and feature construction, features highly correlated with gas concentration are effectively built. A cyclic temperature modulation mode is developed for FC, enabling the decomposition of dynamic response signals into transient and steady-state response phases based on the temperature change curve. Both phases are visually analyzed, and the mapping relationship between static features and dynamic responses is thoroughly analyzed. In this study, we formulate the mixture recognition as a 48-class classification task over discretized acetone–toluene concentration combinations. By selecting the average value of the steady-state response phase as discriminative feature, the random forest algorithm achieves a recognition accuracy of 97.92% for gas mixtures, along with rapid training time and enhanced noise robustness. This cost-effective, noise robustness, and lightweight identification model holds considerable promise for application in the domain of portable device integration.