In non-cooperative application scenarios such as electronic reconnaissance, radar pulse sequences often exhibit alternating working modes over time and are accompanied by high pulse loss rates and complex interference, which pose significant challenges to stable-mode recognition and accurate detection of mode transition segments. To address these challenges, this paper proposes a radar mode recognition method based on multi-granularity feature fusion, which achieves joint recognition of stable working modes and transition segments by modeling the structural characteristics of pulse sequences across multiple temporal scales. The proposed method constructs a multi-branch feature extraction framework composed of a Mamba module, a pulse-aware temporal convolution module, and a local grouped-query attention module. These components are designed to capture long-term temporal dependencies, local pulse structural features, and fine-grained temporal correlation information, respectively. By integrating features extracted at different granularities, the proposed framework enhances its capability to represent complex mode evolution processes. Experimental results conducted on simulated datasets with varying pulse loss and interference pulses demonstrate that the proposed method overall outperforms the compared LSTM, CNN, GRU, TCN, and Transformer-based models in terms of both working mode recognition accuracy and transition segment detection accuracy, and exhibits superior robustness under high pulse loss conditions. These results validate the effectiveness of the multi-granularity feature fusion strategy for radar pulse sequence mode recognition in complex environments.