Bingzhi Wang

In acoustic sensing, spectrum-based representations such as Channel Impulse Response (CIR) and Doppler Frequency Shift (DFS) are important as they transform reflected acoustic signals into structured spectra for motion analysis. Learning effective representations from these spectra, however, remains challenging because they preserve intrinsic temporal–physical structure and often exhibit structural sparsity, where informative motion-related responses are localized amid widespread weak background responses. To address these challenges, we propose SURE, a fusion-based self-supervised predictive learning framework for acoustic sensing. Specifically, SURE employs a dense predictive path, which performs multi-scale masked reconstruction to preserve global temporal–physical structure, and a sparse structural path, which is explicitly designed to handle structural sparsity by leveraging sample-adaptive support inference and submanifold sparse convolution to identify informative regions and preserve localized motion-related structure. The two paths are coupled through reconstruction bottleneck fusion during pretraining, and a multi-scale fuse extractor further integrates multi-level dense features with sparse-guided readout to form the final acoustic descriptor. Experiments on one self-collected CIR-based dataset and one public DFS-based dataset show that SURE learns discriminative representations from unlabeled data, generalizes across different spectrum forms, maintains robustness under noisy conditions, and outperforms representative supervised and self-supervised baselines under a unified evaluation protocol.