Accurate load monitoring is essential for constructing load spectra and enabling the in-service evolution of airframe digital twins. However, many aircraft lack sufficient strain sensors due to weight and cost constraints, making real-time load tracking challenging. This study presents a real-time load tracking method combining deep learning-based strain prediction with an improved inverse–direct approach to enable load estimation under limited-sensor conditions. A deep neural network incorporating temporal features is trained on flight data to predict sparse strain points using flight parameters as inputs, outperforming traditional regression methods. The improved inverse–direct approach, based on reduced-order aerodynamic modes, further enhances load reconstruction accuracy. The integrated system predicts full-field loads and deformations in real time using only flight parameters, enabling sensor-free structural health monitoring. Experimental validation on a 3.2-meter wingspan UAV demonstrates high accuracy, generalization, and feasibility, providing a practical pathway toward digital twin implementation in existing aircraft fleets.