Event-Camera-Based Defect Detection of High-Speed Rotating Propellers

Defect detection in rotating propellers is critical because small defects can cause performance degradation, vibration, noise, and safety risks. Conventional frame-based cameras suffer from motion blur, whereas high-speed cameras often require costly equipment and controlled conditions. This study investigates an event-camera-based sensing framework that converts asynchronous luminance changes from high-speed rotating propellers into positive-negative (PN) composite event accumulation images and classifies normal and defective propellers using a convolutional neural network (CNN). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizes class-discriminative regions and extracts candidate defect regions, while normalized cross-correlation generates event accumulation images at the same rotational phase for defect progression monitoring. Experiments with 65 mm three-blade propellers with approximately 1 mm artificial blade defects rotating at approximately 862 rpm showed that, in an exploratory accumulation-time analysis, 900 μs produced an image-level accuracy of 0.93; integrating images over one full rotation yielded no misclassifications under the present dataset. An additional experiment at 2610 rpm tracked defect progression using a quantitative indicator based on distance-transform pixel counts. These results support the feasibility of event-based sensing for defect detection, cropping candidate defect-related regions, and tracking an image-domain indicator associated with defect progression in high-speed rotating propellers under the tested conditions.