GDAM-DSTCN-LSTM Hybrid Model: Gaussian Dependency-Attention Guided Dual-Scale Mixed Gas Classification

Aiming at the problems of traditional electronic nose in the recognition of complex gas mixtures, such as lack of long-range dependence modeling, difficulty in capturing dynamic time dependence and limitation in multi-scale feature extraction, an end-to-end deep learning model, Gaussian Dependent Attention Dual-scale Temporal Convolutional Long Short-Term Memory Network (GDAM-DSTCN-LSTM), is proposed in this paper. The model uses Gaussian Dependent Attention Module (GDAM) to explicitly model the feature-space dependence between time steps and enhance the feature representation, uses Dual-Scale Temporal Convolutional Network (DSTCN) to extract multi-scale local temporal features in parallel, and combines Long Short-Term Memory Network (LSTM) to capture the long-term evolution of sensor data. Experimental results on UCI electronic nose dataset show that GDAM-DSTCNLSTM achieves $98.48 \%$ classification accuracy in mixed gas classification task. It is significantly better than traditional machine learning methods (such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Extreme Learning Machine (ELM)) and a variety of single or combined deep learning models (such as Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM), Temporal Convolutional Network (TCN), TCN-LSTM). Ablation experiments verify the effectiveness of GDAM and DSTCN modules, and the cosine annealing scheduling strategy effectively improves the stability and convergence rate of model training. In summary, GDAM-DSTCN-LSTM provides an efficient, stable and scalable end-to-end solution for high-precision identification of complex gas mixtures, which has a good prospect for industrial and environmental monitoring applications.