Zero-Shot Wireless Blockage Detection via Foundation Models

Wireless blockage detection (WBD) is crucial for ensuring stable communication services, yet traditional methods and deep learning (DL)-based approaches face limitations that conventional techniques struggle to distinguish line-of-sight (LoS) blockage from complex channel dynamics, while DL models rely heavily on large-scale labeled wireless data, which are costly and difficult to annotate. To address these challenges, this paper proposes FM-WBD (Foundation Model-based Wireless Blockage Detection), a zero-shot WBD framework leveraging the cross-modal reasoning and generalization capabilities of foundation models (FMs). We design a custom wireless encoder to extract semantic features from received power measurements, and employ contrastive learning to align wireless features with image features from pretrained FMs. Zero-shot detection is achieved by measuring the similarity between aligned wireless features and text prompt embeddings representing blocked/unblocked states. Additionally, a few-shot enhancement strategy is introduced to refine detection accuracy using a small number of labeled samples. Evaluations on the DeepSense 6G dataset demonstrate that FM-WBD achieves competitive performance with fully supervised baselines, and exhibits superior generalizability to unseen scenarios. Although its zero-shot performance is inferior to that of few-shot learning, the zero-shot capability enables label-free WBD, which is of significant practical and research value in scenarios where labeled data are unavailable or prohibitively costly to obtain.