LGR-Net: A Lightweight Defect Detection Network Aimed at Elevator Guide Rail Pressure Plates

Abstract

In elevator systems, pressure plates secure guide rails and limit displacement, but defects compromise their performance under stress. Current detection algorithms face challenges in achieving high localization accuracy and computational efficiency when detecting small defects in guide rail pressure plates. To overcome these limitations, this paper proposes a lightweight defect detection network (LGR-Net) for guide rail pressure plates based on the YOLOv8n algorithm. To solve the problem of excessive model parameters in the original algorithm, we enhance the baseline model’s backbone network by incorporating the lightweight MobileNetV3 and optimize the neck network using the Ghost convolution module (GhostConv). To improve the localization accuracy for small defects, we add a high-resolution small object detection layer (P2 layer) and integrate the Convolutional Block Attention Module (CBAM) to construct a four-scale feature fusion network. This study employs various data augmentation methods to construct a custom dataset for guide rail pressure plate defect detection. The experimental results show that LGR-Net outperforms other YOLO-series models in terms of overall performance, achieving optimal results in terms of precision (p = 98.7%), recall (R = 98.9%), mAP (99.4%), and parameter count (2,412,118). LGR-Net achieves low computational complexity and high detection accuracy, providing an efficient and effective solution for defect detection in elevator guide rail pressure plates.