A multitask learning architecture for gas mixture species recognition and concentration inversion based on multiscale transformer

Mixed gas detection holds significant importance in environmental monitoring, industrial safety, and air quality assessment. However, accurate identification of gas species and concentration inversion remain challenging due to the complexity of gas mixtures and the unique characteristics of spectral data. This paper proposes a Multi-scale Attention Convolution Module (MSAtt-MT) integrated with a multi-task learning strategy that can simultaneously handle the two tasks of gas species identification and concentration inversion. A spectral dataset containing target gases NH₃, CH₄, and N₂O were generated using the HITRAN database and experimentally measured data. Experimental results demonstrate 96.71% accuracy in gas species identification with a 50% reduction in concentration prediction error compared to traditional methods. The model exhibits strong generalization capabilities, providing an effective solution for practical gas detection applications.