The evaluation of practical skills in aerospace manufacturing, maintenance,
repair, and overhaul (MRO) conventionally relies on examinations or observations
from instructors. However, these methods deliver slow feedback, reduce the
efficiency of skill acquisition and require substantial instructional workload.
To overcome these challenges, this study proposes an edge-intelligent real-time
evaluation system for skill training in the field of aerospace manufacturing.
Specifically, the system comprises five sensing modalities, which are then
processed through a hierarchical three-layer structure. Then, the neural
architecture search (NAS) is introduced to identify the optimal network
structure and the quantization-aware training (QAT) approach is employed to
improve the inference speed. In addition, temporal modeling is used to realize
real-time inference and multi-level evaluation. Moreover, a conflict-aware
adaptive fusion algorithm is introduced to ensure robust scoring under harsh
workshop conditions and deliver interpretable feedback. Finally, federated
learning and the differential privacy mechanism are used for cross-institutional
model improvement without raw data exposure. Empirical results on turbofan
engine MRO assembly, airframe structural component CNC manufacturing, and
aviation maintenance human factors training show that the system achieves low
latency, high expert correlation, improved learning efficiency and strong
robustness, which offers a scalable and privacy-aware solution for
competency-based vocational training in the aerospace manufacturing field.