融合 BiFPN 和注意力机制的电力设备异常检测算法

An Anomaly Detection Algorithm for Power Equipment Integrating BiFPN and Attention Mechanism

摘要:

为提高电力设备异常检测的精度，提出以融合双向加权特征金字塔网络（BiFPN)和三重注意力（TA）机制改进Yolov5s的电力设备异常检测方法。首先，融合BiFPN是在特征融合结构中加入跨尺度连接线以保留更多深层的语义信息,可以有效促进目标的分类识别和位置精确定位；然后，加入采用三分支结构的注意力机制能够更好地提取空间交互注意力和通道空间交互注意力，抑制无用的特征信息；最后,通过采用Soft NMS来取代传统的NMS算法可以有效减少目标的遗漏，并提升检测的准确性。实验数据显示，改进后的YOLOv5s网络模型相较于原始YOLOv5s模型,精确率从88.3%提升至90%，召回率从89%提升至93%，mAP@0.5值从88.7%提升至92.8%，有效地提高了检测精度。

Abstract:

In order to improve the accuracy of anomaly detection of power equipment, an improved Yolov5s anomaly detection method for power equipment by combining bidirectional weighted feature pyramid network(BiFPN) and triplet attention (TA) mechanism is proposed. In the integrated BiFPN, the cross scale connection lines are added into the feature fusion structure to retain more deep semantic information, which can effectively promote the classification and accurate location of the target. The addition of three branch structure of TA mechanism can better extract the spatial interactive attention and channel spatial interactive attention, and suppress the useless feature information. Finally, using Soft NMS to replace the traditional NMS algorithm can effectively reduce the omission of the target and improve the accuracy of detection. Experimental data show that compared with the original YOLOv5s model, the accuracy rate of the improved YOLOv5s network model has been increased from 88.3% to 90%, the recall rate has been increased from 89% to 93%, and the mAP@0.5 value has been increased from 88.7% to 92.8%, which effectively improves the detection accuracy.