MDFE-Net: A Meta-Learning Driven Dual-Branch Feature Extraction Network for E-Nose Sensor Drift Adaptation

Gas sensors are prone to ageing, environmental fluctuations, and device-to-device variability during long-term deployment. The resulting sensor drift progressively erodes the recognition performance of Electronic nose (E-nose) systems. Here, a Meta-learning Driven Dual-branch Feature Extraction E-nose Drift Adaptation Network (MDFE-Net) was developed for few-shot drift compensation. A dual-branch feature extraction module (DBFE) was constructed to encode temporal dynamics and cross-sensor spatial response patterns, enabling discriminative feature-level modelling of drifted samples. Model-Agnostic Meta-Learning (MAML) was incorporated to learn task-shared priors and to enable rapid adaptation to unseen drift conditions with only a few labelled samples. An adaptive triplet loss and a dynamically reweighted cross-entropy loss were used to tighten intra-class clusters and enlarge inter-class margins. MDFE-Net achieves 95.53% accuracy under long-term drift and 95.71% accuracy under short-term drift on the Gas Sensor Array Drift Dataset. It maintains 97.89% accuracy under cross-device evaluation on the Twin Gas Sensor Arrays benchmark. Together, MDFE-Net couples deep metric learning with meta-learning to mitigate long-term nonlinear drift and improve cross-device generalisation when calibration labels are scarce.