Gas Identification Driven by Dynamic Response Fragments and Multi-Task Learning

Abstract

Abstract The electronic nose system has attracted much attention due to its wide application in industrial detection, atmospheric monitoring and noninvasive disease diagnosis. However, the traditional methods are limited in accuracy and generalization due to the lack of selectivity and reliability of metal oxide semiconductor (MOS) sensors, as well as the synchronization requirements of gas qualitative and quantitative in applications. To address these challenges, this work introduces multitask learning based temporal convolutional bidirectional attention network (MTL-TCBANet), a novel model for gas recognition and concentration prediction. The proposed method harnesses MOS response dynamics by extracting peak-centered fragments with rising and recovery features and employs a temporal convolutional multitask framework with bidirectional cross-attention and dynamic weighting to jointly optimize recognition and regression. Experiments show that MTL-TCBANet attains 0.96 accuracy in gas recognition and an R2 of 0.93 in concentration prediction, surpassing seven traditional baselines and two single-task deep models. Remarkably, it preserves full-sample recognition accuracy with only 50% of the response sequence, demonstrating strong generalization to varying sample lengths. Ablation studies further confirm the essential roles of the multitask temporal convolution, cross-attention, and dynamic weighting modules in performance gains. To validate the generalization capability, three mixed-gas data sets with increasing complexity were introduced. The proposed method achieves consistently superior performance across different data distributions, with accuracy up to 0.97 and R2 up to 0.95. Even under turbulent conditions, the model maintains stable performance, demonstrating strong robustness and cross-data set generalization ability. This study offers an enhanced input processing mechanism for gas identification and concentration prediction, while also presenting a novel approach for multitask model architecture design.