Differentiable End-to-End Sensor Array Optimization for Electronic Nose via Local Self-Attention

Electronic noses (e-noses) based on cross-selective gas sensors often suffer from severe redundancy, which degrades odor analysis accuracy and impedes practical deployment. Here, we introduce a fully deep-learning-based sensor array optimization (SAO) framework that learns an optimal subset of sensor array through an end-to-end differentiable manner. Unlike conventional step-by-step workflows, which separate feature extraction, feature selection, sensor selection, and model training, our method jointly optimizes the sensor selection strategy, feature extractor, and task-specific network, enabling global collaborative optimization. Core to the framework is a local self-attention mechanism that infers sensor-wise importance while preserving the temporal structure of individual sensor responses, together with a tanhshrink-based gating function that drives redundant channels exactly to zero without requiring explicit sparsity regularization. Experiments on a 13-channel polyaniline sensor array exposed to six volatile organic compounds show that the proposed method achieves an accuracy of 91.18 ± 0.55% using only four sensors, reducing the number of sensors by nearly 70% while outperforming conventional approaches. These results demonstrate that the differentiable SAO can improve both efficiency and accuracy in practical e-nose applications.