Adversarial Domain Adaptation Guided by Farthest Distance for open set electronic nose drift compensation

Abstract

With advancements in modern science and technology, electronic noses (ENs) have gained significant attention for their applications in environmental monitoring, food quality inspection, and medical equipment. ENs mimic biological olfactory systems to classify gases using arrays of sensors and pattern recognition models. However, gas sensor drift poses a major challenge, leading to performance degradation in EN systems. To address this, Domain Adaptation (DA) methods align source domain data with target domain drift data. While traditional DA methods assume identical class compositions in both domains, this is often unrealistic in practice, leading to suboptimal results. Open Set Domain Adaptation (OSDA) methods address unknown classes in the target domain, but they often focus too much on distinguishing unknown classes, neglecting accurate recognition of known classes. To overcome these limitations, we propose the Adversarial Domain Adaptation Guided by Farthest Distance (ADA-FDG), comprising two complementary modules: Farthest Distance Guide (FDG) and Confidence Normalized Adaptive Factor (CNAF). FDG adaptively builds a guide set that lies farthest from the source distribution in feature space, ensuring adversarial alignment learns to the edge region distribution. CNAF assigns a weight to each batch proportional to its classification confidence, preventing unknown-class samples from contaminating the ADA process. By integrating FDG and CNAF in an adversarial training framework, ADA-FDG achieves more precise alignment of source and target distributions while preserving clear separation between known and unknown classes. Extensive experiments on two benchmark datasets demonstrate that ADA-FDG consistently outperforms state-of-the-art closed and open set DA methods, delivering significant improvements in overall, known-class, and unknown-class accuracy.