The health state of the blade directly affects the operating performance and the safety of aero-engines. Therefore, blade damage detection plays a more important role in aero-engine maintenance. Existing intelligent aero-engine blade damage detection methods basically rely on deep learning techniques. However, with the deepening of neural networks, the low-level features critical for localization and the high-level features critical for classification cannot be extracted simultaneously . To address this problem, an adaptive feature enhancement network is proposed, which adaptively enhances the information of low-level features and high-level features. Specifically, a feature enhancement module is embedded in the backbone network of the benchmark model (YOLOv5), which improves the feature extraction capability of the backbone network, the generalization capability and the training speed. In addition, a feature reconstruction fusion module is distributed in the overall structure of YOLOv5 model, which prevents the effective features extracted by the backbone network from being weakened in the subsequent networks. Wise intersection over union loss is utilized as bounding box regression loss to reduce the harmful gradient generated by low quality anchors, further improving the overall performance of the detection model. A simulated blade damage dataset is constructed to validate the proposed method, which obtains a performance improvement of 1.9% for the mean average precision (mAP) and 9.1% for the mean average precision at IoU 0.75 (mAP@75) compared with the baseline model.

aero-engine blade, damage detection, feature enhancement, object detection