A Multi-Modal Fusion Network of Visible-Ultraviolet Features for Detecting Aero-Engine Blade Microcracks

Detecting microcracks in aero-engine blades is paramount for ensuring flight safety, yet conventional inspection methods exhibit significant limitations in precision, efficiency, and applicability. This paper proposes a multi-modal fusion detection network for blade microcracks-VUFNet (visible and ultraviolet multi-modal fusion detection network)-which accurately identifies blade microcracks by integrating visible and ultraviolet image features. First, the blade surface undergoes chromic acid anodisation to enhance microcrack features. Subsequently, visible and ultraviolet light illumination is applied to construct a microcrack dataset. VUFNet employs dual backbone networks to extract visible and ultraviolet features. It incorporates CSPSTR (cross-stage partial bottleneck with swin transformer) to optimise feature extraction capability, and VUFM (visible-ultraviolet multi-modal adaptive fusion module) to deeply fuse visible and ultraviolet features. Finally, MFConv (multi-branch fusion convolution) enhances detection accuracy in microcrack target regions. On the held-out test subset acquired from a single compressor-blade type under a fixed laboratory imaging protocol, VUFNet achieved mAP_0.5 and mAP_0.5:0.95 values of 96.8% and 85.5%, respectively, with a detection speed of 52.7 fps. These results demonstrate improved within-domain detection performance over the evaluated baselines. Ablation studies further validate the synergistic optimisation effect of VUFM, CSPSTR, and MFConv on model performance.