Chen Junjie, Pan Siqi, Chan Yanping, Ni Yuedong, Ye Donghua
State Grid Zhangzhou Power Supply Company, Zhangzhou, 363000, Fujian, China.
Sci Rep. 2024 Apr 23;14(1):9362. doi: 10.1038/s41598-024-60126-2.
Artificial intelligence has demonstrated notable advancements in the realm of visual inspection and defect detection in substations. Nevertheless, practical application presents challenges, with issues arising from the dynamic shooting environment and limited dataset resulting in suboptimal defect identification accuracy and instability. To address these concerns, a pioneering approach based on hybrid pruning YOLOv5 and multiscale data augmentation is proposed for enhancing defect detection in substations. Initially, an enhanced multiscale data augmentation method is proposed. The improved multiscale data augmentation mitigates the impact of the time-varying shooting environment on recognition accuracy and enhances defect detection precision. Subsequently, YOLOv5 is employed for training and detecting defects within multi-scale image data. To alleviate the potential destabilizing effects of YOLOv5's large-scale parameters on model stability, a new model pruning method is implemented. This method strategically prunes parameters to bolster the model's defect identification accuracy. The efficacy of the proposed methodology is evaluated through testing on substation defect images, confirming its effectiveness in enhancing defect detection capabilities.
人工智能在变电站视觉检测和缺陷检测领域已取得显著进展。然而，实际应用面临挑战，动态拍摄环境和有限数据集引发的问题导致缺陷识别准确率欠佳且不稳定。为解决这些问题，提出一种基于混合剪枝YOLOv5和多尺度数据增强的开创性方法，以提高变电站的缺陷检测能力。首先，提出一种增强型多尺度数据增强方法。改进后的多尺度数据增强减轻了时变拍摄环境对识别准确率的影响，提高了缺陷检测精度。随后，使用YOLOv5对多尺度图像数据中的缺陷进行训练和检测。为减轻YOLOv5大规模参数对模型稳定性的潜在不稳定影响，实施了一种新的模型剪枝方法。该方法通过策略性地剪枝参数来提高模型的缺陷识别准确率。通过对变电站缺陷图像进行测试，评估了所提方法的有效性，证实其在增强缺陷检测能力方面的有效性。
