Cross-domain specific emitter identification (SEI) is vital for RF systems management but faces challenges from temporal signal variability, hardware state drifts, and cross-domain feature distribution mismatches. We propose a Multi-scale Adaptive Attention Block (MAAB) architecture addressing these issues through coordinated multi-temporal feature learning, combining a Multi-scale Dilated Convolution Module (MDCM) that captures hierarchical patterns via progressive dilation rates and a Dynamic Adaptive Attention Module (DAAM) enabling channel-wise feature recalibration. The MDCM-DAAM synergy facilitates robust cross-scale feature alignment while maintaining hardware fluctuation resilience. Our framework integrates MK-MMD distribution matching with adversarial domain adaptation, creating a dual optimization mechanism that simultaneously minimizes inter-domain discrepancies and enhances discriminative feature transfer. Comprehensive evaluations using real-world WiFi datasets confirm the system’s superior performance, demonstrating 12.7% accuracy improvement over conventional methods in cross-scenario tests. The proposed solution achieves 94.2% average identification accuracy across diverse operational conditions, particularly excelling in handling temporal window variations and long-term hardware degradation effects, proving its practical viability for real-world SEI deployment under non-stationary signal environments.