A Lightweight YOLOv8s Algorithm for Ceiling Fan Blade Defect Detection With Optimized Pruning and Knowledge Distillation

While industrial ceiling fans play a vital role in maintaining efficient airflow and system stability within factory settings, surface defects on their blades that arise over time can disrupt their functionality and create safety hazards. Detecting these surface defects is crucial; however, accurate and rapid detection typically involves complex machine vision algorithms, such as You Only Look Once (YOLO) networks, that require considerable computing resources, which contradicts the industry’s preference for simpler algorithms that can be deployed using low-cost computing power. We propose a lightweight algorithm for detecting surface defects on ceiling fan blades using YOLOv8s, incorporating optimizations and enhancements in model pruning and knowledge distillation. Using our extensively built dataset, the training process yields a YOLOv8s network capable of reliably identifying the surface defects, functioning as the teacher network for knowledge distillation. The dynamic self-tuning strategy for sparsity factors, developed for sparse training, effectively determines the regularization intensity that reflects network channel contributions, providing more appropriate network pruning results to form the lightweight student network. The proposed hybrid loss function promotes sufficient knowledge transfer of intermediate and output features from the teacher to the student network, thus preserving defect identification capabilities that might otherwise degrade through knowledge distillation. Experimental results confirm that our algorithm achieves an mAP of 0.976 and operates at 91.4 FPS, while reducing the number of parameters by 76.6% and floating-point operations by 85.2%. These results demonstrate superior accuracy, speed, and deployment efficiency over conventional methods and highlight our algorithm’s considerable potential for surface defect detection applications beyond ceiling fan blades. This work is anticipated to present a new solution for developing and applying lightweight YOLO-based networks.