Aiming at the critical issue of interlayer debonding damage susceptibility in reusable launch vehicle thermal protection structures under complex multi-physics coupling environments, a non-destructive testing method integrating ultrasonic guided waves with domain-adaptive transfer learning was proposed. Four typical bonding types were designed in thermal protection tile specimens, enabling efficient full-coverage inspection of bonded areas through a bidirectional orthogonal scanning strategy coupled with an ultrasonic excitation-reception mechanism. To solve the problem of signal drift caused by individual differences of specimens, an adaptive phase alignment method based on peak proportion threshold is proposed, and an appropriate window length is selected to realize the retention of key features of debonding damage while suppressing the interference of redundant data. A Domain-Adaptive Transfer Learning (DATL) was further proposed to align cross-specimen damage feature distributions. Experimental results demonstrate that in cross-specimen testing scenarios, the DATL model exhibits only a 3.9% accuracy decline, with inter-domain distribution discrepancy reduced from 0.31 to 0.10. With target domain data below 40%, DATL achieves 85% accuracy, outperforming CNN by 19.4%. The methodology mitigates reliance on damage patterns and specimen consistency, effectively reducing false alarms and missed detections in debonding testing for in-service thermal protection systems, which provides a practical solution for rapid non-destructive evaluation and structural health monitoring of reusable launch vehicle.