<p>Bearing fault diagnosis is critical for ensuring operational reliability and airworthiness. However, significant challenges remain in diagnosing bearing faults with limited samples under unknown variable speed. To address these dual constraints, this paper proposes a novel dual-branch ensemble learning framework for bearing fault diagnosis based on Ensemble Feature Selection and Wasserstein Distance (EFS-WD). The primary novelty lies in an adaptive diagnostic architecture that integrates a dual-branch feature selection strategy to systematically satisfy distinct diagnostic requirements. First, a comprehensive pool of physically interpretable diagnostic indicators is extracted across multiple domains from multi-sensor signals. Subsequently, an accuracy-driven ensemble branch identifies highly discriminative fault-related features, while a parallel branch leverages the Wasserstein distance to evaluate and select speed-invariant features. By adaptively integrating a tree-based ensemble classifier and a distribution-matching approach based on the Wasserstein distance, the framework achieves robust and accurate fault identification under both known and unknown speed scenarios, even with extremely limited samples. Case studies on an aero-engine bearing dataset and an aeronautical bearing dataset demonstrate the effectiveness of the proposed method.</p>