Pulse repetition interval (PRI) modulation type recognition of radar signals is a critical component for signal sorting and threat assessment in electronic warfare systems, as its accuracy directly affects battlefield electromagnetic situational awareness and the effectiveness of countermeasure strategies. To address the limitations of traditional methods—namely, strong reliance on expert knowledge and insufficient feature extraction capacity of single models—this paper proposes a deep neural network architecture based on multi-dimensional feature space-time fusion. The approach constructs a cascade of convolutional neural networks (CNN) and long short-term memory (LSTM) networks to capture the intrinsic correlations between the spatiotemporal characteristics of radar pulses and the temporal evolution of modulation patterns. A sequential attention mechanism is incorporated to weight and fuse the extracted temporal features, emphasizing key moments of PRI variation and enabling accurate identification of PRI modulation types. Experiments are conducted on a dataset comprising five typical modulation types: Fixed, jitter, group change, stagger, and slip. Results show that, under ideal interference-free conditions, the model achieves an overall recognition accuracy of 99.40%. Even under 70% pulse loss interference, recognition accuracy remains 70.93%, and under equal-strength false pulse interference, it reaches 96.13%. The proposed model significantly outperforms mainstream comparison models, including CNN, CNN-LSTM, GRU-Attention, and SE-NET, demonstrating enhanced accuracy and robustness in complex electromagnetic environments. Furthermore, the model requires only 1.54 ms for single-sample inference and has 653 814 parameters, offering strong real-time capability and potential for lightweight deployment. The proposed method provides an effective technical solution for real-time, accurate recognition of radar PRI modulation types in complex electromagnetic environments, with both theoretical significance and practical engineering applicability.