Transformer-driven multi-scale contextual alignment for robust E-nose drift compensation

doi:10.1016/j.knosys.2025.114136

Transformer-driven multi-scale contextual alignment for robust E-nose drift compensation

2025-07-19

0

PRE

AI

DOI：10.1016/j.knosys.2025.114136

Original

Original request for help

Save

Abstract

En 中文

Sensor drift poses a significant challenge to the performance and reliability of electronic noses (e-noses) in practical applications. To address this, we propose a Transformer-driven Multi-Scale Contextual Alignment (TMSCA) framework. TMSCA comprises three mutually reinforcing modules: (i) Cross-Domain Prior Attention, encoding source knowledge as a learnable prior to guide alignment even when target classes are missing; (ii) k-nearest-neighbor-constrained Local Maximum Mean Discrepancy, providing fine-grained distribution matching; and (iii) Approximate Low-Rank Canonical Correlation Analysis, preserving cross-domain correlations with sub-quadratic complexity. By synergistically leveraging these components, TMSCA systematically tackles the limitations of prior methods, such as incomplete target data adaptation, local distribution discrepancies, and prohibitive computational costs associated with deep models. Experimental results on two public e-nose datasets, under both long-term and short-term drift scenarios, demonstrate that TMSCA achieves strong and consistent performance in sensor drift compensation, suggesting potential improvements in e-nose accuracy and reliability.

Keywords:

sensor drift

electronic nose

domain adaptation

Transformer

multi-scale alignment

Transformer-driven multi-scale contextual alignment for robust E-nose drift compensation

Abstract

Journal

Organization