Micro-crack detection of aero-engine inlet casing struts based on convolutional neural network
Measurement Science and Technology · 21.08.2026 · 37 (33) · 335601
DOI: 10.1088/1361-6501/ae7e1a ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 19.08.2026. Verfasst von Mingwei Xu und 4 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: National High-tech Research and Development Program, Natural Science Foundation of Liaoning Province, National Natural Science Foundation of China.
Abstract (Originalsprache)
Micro-crack detection of aero-engine inlet casing struts is a critical task for ensuring flight safety. However, the extreme scarcity of crack samples in real-world operations leads to severe class imbalance, making high-recall detection exceptionally challenging. Standard class-imbalanced learning techniques often fail in this setting, as they either overfit to the few positives or are overwhelmed by the majority class. To address this largely unexplored problem, we present a specialized deep learning framework tailored for percussion-based micro-crack detection in aero-engine inlet casing struts, specifically designed for such extreme data scarcity. The acoustic response signal is first normalized by the peak value of the impact force signal. Then, the logarithmic short-time Fourier transform spectrogram is computed, so as to preserve the resonant features in crack-sensitive frequency bands. Subsequently, a lightweight two-dimensional CNN model is constructed for micro-crack detection. Finally, multiple recall-oriented training strategies, including data augmentation, weighted sampling, and weighted cross-entropy loss function, are applied to handle the extreme class imbalance where crack samples account for less than 1% of the dataset. Experiments are conducted on a real-world dataset containing 2970 samples (only 5 of which are crack samples) using five-trial stratified 5-fold cross-validation. The proposed method achieves an average recall of 100% with 20.6 false positives per 2970 tests. This achieves a promising balance in safety-critical screening, lowering the workload of subsequent precise inspection by over 99% and outperforming other baseline methods. Meanwhile, Grad-CAM visualizations confirm that the network focuses on physically relevant time–frequency regions.
Förderung
- National High-tech Research and Development Program
- Natural Science Foundation of Liaoning Province
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
- the Project of Fault Diagnosis Technology for Thin-walled Parts
Zitieren
APA
Xu, M., Chen, Z., Yin, F., Yang, M., & Wang, M. (2026). Micro-crack detection of aero-engine inlet casing struts based on convolutional neural network. Measurement Science and Technology, 37(33), 335601. https://doi.org/10.1088/1361-6501/ae7e1a
Vancouver
Xu M, Chen Z, Yin F, Yang M, Wang M. Micro-crack detection of aero-engine inlet casing struts based on convolutional neural network. Measurement Science and Technology. 2026;37(33):335601. doi:10.1088/1361-6501/ae7e1a
BibTeX
@article{Xu2026,
author = {Xu, Mingwei and Chen, Zhe and Yin, Fuliang and Yang, Ming-sui and Wang, Meng},
title = {Micro-crack detection of aero-engine inlet casing struts based on convolutional neural network},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {33},
pages = {335601},
doi = {10.1088/1361-6501/ae7e1a}
}
EUVIMED