Multi-task learning with attention for efficient gas classification and concentration prediction in sensor arrays

Abstract As an effective method of olfactory simulation and gas identification, electronic nose (E-nose) sensing systems are extensively employed in industrial production sectors, playing a critical role in addressing substantial emissions of toxic and hazardous gases released during manufacturing processes. Compared with traditional chemical analysis instruments, E-nose has quickly occupied the market by virtue of its high performance-price ratio, high recognition rate and simple portability. However, given the heterogeneous composition and intricate characteristics of gas data samples, how to accurately classify and predict gas concentration is still a certain challenge. This study presents a novel framework named multi-task learning-LACNet for gas detection applications. The model combines long short-term memory networks with a residual-based 1D convolutional neural network architecture, enabling the concurrent determination of gas species and quantification of their concentrations. An attention-enhanced module specifically designed enables autonomous learning of critical features relevant to detection objectives, significantly improving classification accuracy and concentration regression performance. Simulation validation on two public datasets demonstrates the framework’s superior capabilities: achieving 98.90% classification accuracy with 99% R 2 value for concentration prediction in the first dataset, while attaining 99.39% classification precision and 98.76% determination coefficient in the second dataset, outperforming conventional detection approaches. This study provides an efficient and reliable solution for intelligent gas sensing technology with substantial engineering application value.