Anomaly Detection and Feature Inference in Assembly Quality Based on Auto Encoders

Aero-engine assembly encounters challenges in detecting subtle anomalies with scarce fault data and deriving optimal tolerances from complex performance metrics. Traditional methods suffer from overfitting, subjective thresholds, and disconnects between diagnostics and tolerances. This study proposes an intelligent framework integrating Bagging-Stacked Auto Encoders (BSAE) featuring: (1) Ensemble learning for robust feature extraction, reducing reconstruction errors and overfitting; (2) A self-calibrating dynamic threshold in latent space for unsupervised high-accuracy anomaly detection; (3) A bidirectional decoder translating critical performance parameters (e.g., thrust) inversely into manufacturable tolerance specifications, linking diagnostics with control. Validated on engine assembly lines, the framework reliably detects concealed defects (e.g., blade misalignment) while optimizing tolerance intervals to improve first-pass yield. It integrates AI-driven anomaly detection with closed-loop quality optimization, advancing intelligent quality assurance for high-precision aerospace systems from inspection to process optimization.