Inspired by the mammalian olfactory system, electronic nose (e-nose) technology has been widely applied in environmental monitoring, food safety, and medical diagnostics. However, mixed gas identification still faces two challenges. Firstly, the gas sensor array exhibits significant cross-sensitivity and low selectivity, resulting in highly overlapped responses among different gases, with redundant features masking key discriminative information. This leads to a sparse and undefined decision space. Secondly, in practical applications, e-nose systems often suffer from limited training samples, making existing methods unstable under low-resource conditions. To solve these problems, this paper proposes MVSP-SACFormer, a mixed gas identification framework tailored for low-resource situations. In the feature extraction phase, we design a Multi-View Subspace Projection Encoding (MVSP) mechanism, which maps raw sensor responses into multiple low-dimensional subspaces via independent random observation matrices, generating compressed features that are structurally diverse and information-complementary. This enhances feature expressiveness under low resource. In the modeling phase, we construct a Spectral-Aware Convolutional Transformer (SACFormer) that fuses spectral, dynamic temporal, and spatial structural features to deeply model the key patterns within complex mixed gas responses. Experiments on two public mixed gas datasets demonstrate that MVSP-SACFormer achieves outstanding recognition performance even when using only 40 % of the training data: achieving 98.03 % classification accuracy on a methane–ethylene mixture dataset, and 97.06 % on an ethylene–carbon monoxide mixture dataset. These results validate that the proposed method serves as a universal pattern recognition framework for e-nose systems under low-resource conditions.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号