18 December 2025 in IEEE Transactions on Instrumentation and Measurement
Artificial olfactory systems play an increasingly vital role in environmental monitoring, industrial safety, and medical diagnostics. However, their deployment in real-world scenarios is often constrained by fragmented model design and limited adaptability to complex gas mixtures. To address these challenges, we propose a unified modeling approach that supports both gas component identification and concentration estimation. First, we extend the Gas Component Strategy (GCS) into a modular learning paradigm that isolates individual gas components during training, enabling scalable generalization to unseen mixtures with minimal retraining. Second, we introduce the Signal Graph Framework (SGF), a unified model architecture that transforms sensor responses into dynamic signal graphs. SGF provides a shared architectural backbone for both classification and regression. Task-specific models are trained independently but retain structural consistency, which simplifies code maintenance, facilitates hardware deployment, and streamlines system integration. Experiments on a custom dataset containing binary and ternary gas mixtures demonstrate over 99% identification accuracy and up to 35% reductions in estimation error compared with state-of-the-art baselines. Beyond laboratory settings, GCS and SGF mitigate class imbalance and error entanglement, while sustaining accuracy on unseen combinations. Together, they provide a scalable and precise foundation for multifunctional artificial olfaction in dynamic, resource-constrained environments.