A lightweight vision-based framework for automated aero-engine blade damage inspection
- 1School of Aviation, Zhejiang Institute of Communications, Hangzhou, Zhejiang, China
- 2School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou, Zhejiang, China
- 3Loongair(Hangzhou) Aviation Maintenance Engineering Co., Ltd., Hangzhou, Zhejiang, China
Abstract
Automated visual inspection of aero-engine blades is central to aviation safety, yet it is challenged by damage that is small and low-contrast against a strongly textured metal surface, and by the need for models that are light enough for edge deployment. This paper presents a lightweight vision-based framework for automated blade damage inspection. Rather than assembling independent detection modules, the framework is organized as a three-stage inspection pipeline: (1) a multi-scale feature-discrimination stage (MSBlock) captures defects across a wide size range at negligible computational cost; (2) a texture-suppression stage (TSA) filters high-frequency machining-background noise through product-based channel–spatial attention, exposing genuine defects; and (3) an efficient-prediction stage (Detect_Efficient) fuses cross-scale features and recalibrates confidence with roughly half the parameters of a standard head. On a self-built engine-blade dataset, the framework improves recall, F1-score, and [email protected] by 4.8, 1.9, and 0.8 percentage points over YOLOv11n (reaching 92.0%, 91.6%, and 94.2%), while cutting parameters and computation by 50.0% and 58.7% and reaching 527 FPS. It also outperforms YOLOv8n, YOLOv9t, and YOLOv10n in recall and F1-score, confirming its suitability for real-time edge-based blade inspection.
Related
- A Three-Year Longitudinal Analysis of Primary Care Quality Improvement Under a Regional Inspection Program in Shanghai: an Evaluation Using the PDSA Cycle as an Analytical Framework
- An explainable vision transformer approach for automated neurological disorder classification from brain MRI scans
- Cytomegalovirus Retinitis Across the Immunosuppression Spectrum: A Contemporary Framework for Risk Stratification, Diagnosis, and Vision-Preserving Therapy
- Quantum-inspired NSGA-III-BERT hybrid decision engine for high-frequency supply chain resilience configuration: design, simulation and engineering framework
- Unifying Vision and Language: Benchmarking End to End Transformer Model Against the ClipCap Framework
BibTeX
@article{ning2026lightweight,
title = {{A lightweight vision-based framework for automated aero-engine blade damage inspection}},
author = {Xiaobo Ning and Hang Zhao and Kang Zeng and Zhengwei Zhang},
journal = {PLOS One},
year = {2026},
month = nov,
doi = {10.1371/journal.pone.0362006},
url = {https://doi.org/10.1371/journal.pone.0362006},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A lightweight vision-based framework for automated aero-engine blade damage inspection
AU - Xiaobo Ning
AU - Hang Zhao
AU - Kang Zeng
AU - Zhengwei Zhang
PY - 2026
DA - 2026/11/05
JO - PLOS One
DO - 10.1371/journal.pone.0362006
UR - https://doi.org/10.1371/journal.pone.0362006
AB - Automated visual inspection of aero-engine blades is central to aviation safety, yet it is challenged by damage that is small and low-contrast against a strongly textured metal surface, and by the need for models that are light enough for edge deployment. This paper presents a lightweight vision-based framework for automated blade damage inspection. Rather than assembling independent detection modules, the framework is organized as a three-stage inspection pipeline: (1) a multi-scale feature-discrimination stage (MSBlock) captures defects across a wide size range at negligible computational cost; (2) a texture-suppression stage (TSA) filters high-frequency machining-background noise through product-based channel–spatial attention, exposing genuine defects; and (3) an efficient-prediction stage (Detect_Efficient) fuses cross-scale features and recalibrates confidence with roughly half the parameters of a standard head. On a self-built engine-blade dataset, the framework improves recall, F1-score, and [email protected] by 4.8, 1.9, and 0.8 percentage points over YOLOv11n (reaching 92.0%, 91.6%, and 94.2%), while cutting parameters and computation by 50.0% and 58.7% and reaching 527 FPS. It also outperforms YOLOv8n, YOLOv9t, and YOLOv10n in recall and F1-score, confirming its suitability for real-time edge-based blade inspection.
SN - 1932-6203
ER -