An Improved Lightweight YOLOv8 using Focaler Training for Fault Recognition of Smart Meters

The recognition of visual faults in smart meters based on computer vision and deep learning methods can effectively improve the accuracy and efficiency of fault detection, ensuring the stable operation of the power system. Existing object detection technologies based on lightweight YOLO algorithms seldom consider the issue of data class imbalance, especially in fault detection where some faults occur infrequently, leading to classification results biased towards the majority class. This paper proposes an imbalanced object detection method based on lightweight YOLOv8 for visual fault recognition in smart meters, improving detection accuracy while reducing model parameters through an improved imbalanced IoU loss and neural network pruning. The proposed method first introduces a Focaler-IoU imbalanced loss, which focuses on different difficulty levels of regression samples by combining multiple IoU calculation methods and dynamically adjusts sample weights to optimize regression performance. For model lightweighting, this paper adopts a sparse momentum pruning method to selectively optimize model parameters by choosing the most effective connection weights under a specified sparsity level. Experiments qualitatively and quantitatively verify that the proposed method achieves higher accuracy and fewer parameters compared to existing YOLO algorithms in visual fault detection of smart meters.