Deep learning-based defects detection of certain aero-engine blades and vanes with DDSC-YOLOv5s.

When performed by a person, aero-engine borescope inspection is easily influenced by individual experience and human factors that can lead to incorrect maintenance decisions, potentially resulting in serious disasters, as well as low efficiency. To address the absolute requirements of flight safety and improve efficiency to decrease maintenance costs, it is imperative to realize the intelligent detection of common aero-engine defects. YOLOv5 enables real-time detection of aero-engine defects with a high degree of accuracy. However, the performance of YOLOv5 is not optimal when detecting the same defects with multiple shapes. In this work, we introduce a deformable convolutional network into the structure of YOLOv5s to optimize its performance, overcome the disadvantage of the poor geometric transformability of convolutional neural networks, and enhance the adaptability of feature maps with large differences in the shape features. We also use a depth-wise separable convolution to improve the efficiency of multichannel convolution in extracting feature information from each channel at the same spatial position while reducing the increased computational effort due to the introduction of deformable convolution networks and use k-means clustering to optimize the size of anchor boxes. In the test results, mAP50 reached 83.8%. The detection accuracy of YOLOv5s for common aero-engine defects was effectively improved with only a 7.9% increase in calculation volume. Compared with the metrics of the original YOLOv5s, mAP@50 was improved by 1.9%, and mAP@50:95 was improved by 1.2%. This study highlights the wide application potential of depth science methods in achieving intelligent detection of aero-engine defects. In addition, this study emphasizes the integration of DDSC-YOLOv5s into borescope platforms for scaled-up engine defect detection, which should also be enhanced in the future.