Ghazala Ansari, Rupali Singh, Sachin Kumar, Ravi Kumar
The increased demand for reliable gas sensing in factory systems has prompted the development of smart sensory systems capable of overcoming errors and signal distortion caused by environmental changes. This paper introduces drift-resistant hybrid feature learning to classify mixed industrial gases, specifically the ethylene-carbon monoxide (CH-CO) and ethylene-methane (CH-CH) mixtures with the help of metal-oxide semiconductor (MOX) sensor arrays. The suggested framework combines both the statistical descriptors, which include the steady-state signal behavior, and the deep temporal-spatial embedding, which is trained by a one-dimensional CNN-GRU architecture. To generate these complimentary representations, a single feature space is produced, which is more interpretable and resilient. Optuna was used to conduct Bayesian search on the models hyperparameters and Extra Trees and LightGBM ensembles were used to conduct classification. The results of the experiments carried out on 180 time-series recordings with labels proved that the fused model attained average accuracies of more than 90% and best fold accuracies of 99% and surpassed a number of advanced machine learning and deep learning baselines. Synthetic baseline drift (0.8-1.4) was also used to test long-term reliability and the framework retained accuracies greater than 96% at all levels of drift. Shapley additive explanations (SHAP) also shed some light by determining kurtosis, skewness, and mid-frequency temporal response as the main discriminative sensor channels. The findings verify that the suggested framework provides a computationally efficient, interpretable and drift-resilient solution that can be used in real-time in an industrial context. In addition to gas analysis, the hybrid approach demonstrates a high potential of more general applications related to non-stationary multivariate sensor data, that is, environmental monitoring, biomedical diagnostics, and process control systems.