Multi-Scale Temporal-Dependency Learning for Single-Sensor Gas Classification

Single-sensor gas classification remains challenging because metal oxide semiconductor sensors exhibit limited selectivity and different gases can produce similar response patterns. This paper proposes a multi-scale temporal-dependency-learning framework that combines reservoir-based multi-scale dynamic learning (MsDL) with lagged-correlation features. To reduce information leakage caused by window construction, each continuous gas-response sequence is divided chronologically into training, validation, and test blocks before window extraction, and 2000 raw samples are excluded between adjacent subsets. The primary evaluation uses non-overlapping windows and fits all data-dependent preprocessing using the training subset only. On a dataset containing seven volatile organic compounds measured using a single sensor, the proposed MsDL+Corr representation with a random-forest classifier achieves a test accuracy of 92.14% and a macro-F1 score of 92.08%. Across ten random-forest seeds, the accuracy is 92.43% ± 0.37%. The corresponding micro-average one-vs-rest AUC is 0.9948 (99.48%); this threshold-independent ranking metric is distinct from classification accuracy. Feature ablation shows that combining MsDL and correlation features improves the macro-F1 score over either branch alone. These results demonstrate that temporal-dependency representations provide useful discriminative information under a conservative temporally separated evaluation protocol.