High-Accuracy Concentration Prediction for Gas Mixtures Based on SF-MBNet

Accurate detection of multi-component gas mixtures is crucial for applications in environmental monitoring, industrial safety, and medical diagnostics. Electronic nose (E-nose) technology has garnered widespread attention for its ability to detect multi-component gases non-invasively. However, current approaches face challenges in capturing the inherent spatial dependencies and complex temporal dynamics embedded in sensor response data. These limitations significantly impact the accuracy of concentration predictions. To overcome these challenges, we propose a novel framework named Sparse Fusion Multi-Branch Network (SF-MBNet). Its architecture comprises three parallel branches, including a multi-scale convolutional module for capturing local temporal patterns at different granularities, a gated recurrent unit (GRU) module for modeling long-range dependencies, and a Dual Temporal-Channel Attention (DTCA) module for adaptively prioritizing critical nodes across temporal and channel dimensions. To effectively integrate these heterogeneous features, we design an Adaptive Channel-Sparse Fusion (ACSF) module that dynamically emphasizes complementary information while suppressing redundant representations across branches. It is achieved through an interactive sparsification mechanism that optimizes channel-wise attention based on inter-branch similarity, thereby significantly improving feature fusion performance. Experimental validation on the UCI Gas Sensor Array dataset under dynamic gas mixture conditions confirms the high predictive accuracy of SF-MBNet. The model achieves a determination coefficient (R²) of 0.9828, along with a mean absolute error (MAE) of 0.0158 and a root mean square error (RMSE) of 0.0375. The model demonstrates superior performance in both MAE and RMSE compared to state-of-the-art methods.