An aero-engine defect detection method via sequential matching and historical experience-guided transfer

Borescope imaging of aero-engines often suffers from specular highlights, oil contamination, low-contrast micro-defects or elongated defects, and probe perturbations, which induce cross-layer reassignment and unstable convergence in end-to-end detectors. We propose an end-to-end aero-engine defect detection framework based on sequential matching and history-guided transfer. Building on Align-DETR, the framework designs three complementary mechanisms spanning features, assignment, and optimization: (1) adaptive history-guided cross-attention leverages historical regions and quality priors to stabilize cross-scale information aggregation and target alignment; (2) adaptive context-guided bias injects a hysteresis term into the linear assignment stage, raising the reassignment threshold and suppressing rematching; (3) matching consistency loss anchors intermediate layers to the final-layer distribution via depth-weighted KL-divergence alignment, enforcing cross-layer consistency. These designs render the originally implicit inter-layer dependencies explicit and differentiable, reducing assignment jitter and gradient conflicts, and promoting monotonic box refinement with stable convergence. Extensive experiments on three aero-engine defect datasets demonstrate consistent gains in detection precision and recall over strong baselines and mainstream competitors. On AE-SD6, our method improves mAP@50:95 over Align-DETR by 5.0%, markedly increases early-stage matching consistency, reduces the number of iterations required to reach a given loss threshold by up to 59%, and converges to a lower and smoother loss plateau. The results verify robustness and generalization under challenging conditions and provide a scalable, engineering-ready solution for intelligent aero-engine defect inspection.