PTADA: An unsupervised domain-adversarial regression algorithm for sensor drift in mixed gas scenarios

Abstract

Compared to single-gas detection, regression prediction for mixed gas is more challenging, due to the feature diversity among different gas components and distribution discrepancy between source and target domains that caused by sensor drift effect. To address these challenges, we propose an unsupervised adversarial domain adaptation neural network called Progressive dual-stream Temporal network with Attention for Domain Adaptation (PTADA), offering an effective drift compensation solution for mixed gas regression tasks through adversarial learning. PTADA employs a multi-scale modeling approach to progressively extract local and global dependencies from temporal gas data, constructing a hierarchical temporal dependency structure and enhancing feature representation. Additionally, PTADA employs Multi-Kernel Maximum Mean Discrepancy (MK-MMD) for regularization to boost domain adaptation. Furthermore, to address the issue of sample imbalance in the concentration data, we designed a hybrid compressed focal loss function (H-Focal Loss) that combines linear error and logarithmic error. We evaluated PTADA on a self-collected H₂S–SO₂ mixed gas drift dataset, focusing on the novel and challenging task of regression-based drift compensation for mixed gases. The experimental results show that PTADA achieved R² values of 0.924 and 0.937 for H₂S and SO₂, respectively, outperforming traditional models and other drift compensation methods (such as Informer, SWD, and ADDA), which highlights its significant advantages in regression tasks under mixed gas drift conditions. Furthermore, noise resistance and sensor failure experiments verified the robustness of PTADA, and additional tests on public dataset and other mixed gas dataset demonstrated its strong generalization capability.