Unmanned Aerial Vehicles (UAVs) face significant challenges in rotor fault diagnosis under small-sample and imbalanced data conditions, which severely limit the effectiveness of existing data-driven methods. In practical applications, fault samples are often fewer than 50 per class, with imbalance ratios typically 0.3–0.5, under which conventional deep learning approaches exhibit notable performance degradation. To address these issues, this paper proposes an MSCNN-BiLSTM-Attention framework for UAV rotor fault diagnosis. The method integrates multi-scale convolution to enhance spatial feature extraction from limited vibration data, bidirectional LSTM to capture temporal dependencies, and an attention mechanism combined with Focal Loss to improve minority-class learning. Experiments demonstrate that the proposed framework consistently outperforms conventional methods under small-sample, noisy, and imbalanced conditions, achieving a maximum improvement of up to 9 percentage points, which confirms its robustness and effectiveness in challenging diagnostic scenarios.