Multilevel Discrete Wavelet Decomposition-Assisted Lightweight Multi-Order Gated Aggregation Network for Gas Concentration Retrieval in WMS

This paper presents a lightweight multigated aggregation network assisted by multilevel discrete wavelet decomposition (MDWD-LiteMogaNet) for gas detection in wavelength modulation spectroscopy (WMS). By integration of the wavelet transform for data filtering and feature extraction, MDWD-LiteMogaNet significantly reduces data volume and enhances computational efficiency. The multigated feature extraction and fusion mechanism ensures comprehensive feature representation, while the gating mechanism optimizes feature selection for efficient utilization. Experimental results demonstrate that multilevel discrete wavelet decomposition effectively reduces computational complexity while maintaining high retrieval accuracy, achieving a reduction in multiply-accumulate operations (MACs) to just 0.003 GFLOPs. Compared with MogaNet, MDWD-LiteMogaNet achieves higher retrieval accuracy while significantly reducing resource consumption, making it more suitable for deployment on lightweight devices. Long-term testing indicates that sensors using MDWD-LiteMogaNet exhibit stable performance, with minimal impact from noise on detection results. This study presents an innovative method for deep learning in gas detection, highlighting the potential of MDWD LiteMogaNet in complex feature extraction and efficient computation.