Radar-based respiratory pattern recognition method using stacking-ELCA multi-domain feature fusion

Abstract To address the limitations of traditional radar-based respiratory recognition methods that rely on single-domain features, this study proposes a deep learning framework based on multi-domain feature fusion. First, a ZMTI clutter suppression method is designed in the preprocessing stage, combining MTI inter-frame differencing and zero-crossing analysis, which improves the signal-to-noise ratio by 15% compared to conventional MTI. Second, the Grasshopper Optimization algorithm (GOA) is employed to adaptively optimize the parameters of VMD, and a lightweight ELCA network is constructed with embedded coordinate attention to model the correlations between time–range and frequency domain spectrograms, achieving single-domain classification accuracies of 89.5% and 92%, respectively. Finally, a Stacking-based multi-domain fusion strategy is proposed. This strategy concatenates probability vectors and uses an MLP to model cross-domain relationships, boosting overall accuracy to 97%. The model has only 0.08 m parameters and achieves an inference speed of 135 frames per second, significantly reducing the misclassification rate of slow breathing to 10%, thus offering an efficient solution for medical monitoring applications.