Aeroengine Blades Damage Detection and Measurement Based on Multimodality Fusion Learning

Aeroengines often work in harsh conditions such as high load, high-speed rotation, and strong corrosion, under these conditions, the aeroengine blades are easily damaged by the impact of foreign objects, which seriously affects the aeroengine performance and flight safety. Therefore, it is very necessary to carry out the research on aeroengine blades damage detection and measurement. In this article, a multimodality intelligent damage detection method based on visual image and depth map and an automatic damage measurement method for aeroengine blades are proposed. An aeroengine blade damage visual-depth multimodality dataset (ABDM dataset) is constructed. The dataset contains four common types of engine blade damage, namely, nick, tear, bent, and chamfer. According to the different fusion stages, three fusion networks are designed: visual-depth data level fusion network (VDFNet-data), visual-depth feature level fusion network (VDFNet-feature), and visual-depth decision level fusion network (VDFNet-decision). Among them, VDFNet-feature has the best damage detection performance, with its mean average precision (mAP) of 85.60% and inference speed of 37.48 frames/s (fps). In the backbone, multibranch concatenation block (Multi-Concat-Block), parallel down sampling block (Parallel-Down-Block), and cross-stage partial spatial pyramid pooling (CSPSPP) block are designed to solve the challenges of damage intelligent detection caused by dim detection environment light, a large change in damage size and small size of some damage. In addition, a stacked symmetrical network (SSNet) is designed to extract the damage feature points, and then damage size is calculated according to the spatial coordinates of feature points on the depth map. The percentage of correct keys (PCKs) and size error (SE) of the measurement method proposed in this article are 93.28% and 0.12 mm, respectively.