基于改进YoloV4的电网变压器油液渗漏检测方法

Oil Leakage Detection Method for Power Grid Transformers Based on Improved Yolov4

摘要:

及时发现电网变压器油液渗漏问题对于电网的安全与稳定运行尤为重要；传统电网变压器油液渗漏检测主要依赖于人工定期巡检，但人工巡检无法实现全天候监测，具有滞后性；当前主流目标检测模型应用于电网变压器油液渗漏检测时，存在检测速度较慢、准确率低和鲁棒性较差等问题，无法满足实际应用；为此提出一种改进YoloV4的变压器油液渗漏检测方法；首先，通过引入Mobile-ViT作为模型的骨干结构，利用卷积和Transformer结构有效提取目标的局部和全局信息特征，降低计算量；其次，提出多尺度特征融合层，旨在实现局部和全局信息的多尺度特征融合，增强上下文语义表达，用以更好地实现电网变压器油液渗漏检测；实验结果表明，该方法在电网变压器油液渗漏数据集上检测精度达到了95.3%，检测速度达到了50.6帧/秒；相较于原生YoloV4方法检测精度提高了2.6%，检测速度提升了2.6帧/秒；经实际应用，该方法部署在边缘设备上推理速度也达到了43帧/秒，满足了实际工程的需求。

Abstract:

It is particularly important for the safe and stable operation of power grid to find the oil leakage problem of power grid transformer in time. Traditional power transformer oil leakage detection mainly depends on manual regular detection, but manual detection cannot achieve all-weather monitoring and has lag. When the current mainstream object detection model is applied to the oil leakage detection of power grid transformers, there are some problems such as slow detection speed, low accuracy and poor robustness. It cannot meet the practical application. An improved You Only Look Once Version 4 (YoloV4) transformer oil leakage detection method is proposed. Firstly, by introducing Mobile Vision Transformer (Mobile-ViT) as the backbone structure of the model, the local and global information features of the object are effectively extracted by convolution and Transformer structure, which reduces the computation. Secondly, a multi-scale feature fusion layer is proposed, which aims to realize the multi-scale feature fusion of local and global information and enhance the context semantic expression, so as to better realize the oil leakage detection of power grid transformers. The experimental results show that the detection accuracy of this method on the power grid transformer oil leakage data set reaches 95.3%, and the detection speed reaches 50.6 frames per second; Compared with the native YoloV4 method, the detection accuracy is improved by 2.6%, and the detection speed is improved by 2.6 frames per second. After practical application, the reasoning speed of this method deployed on edge devices also reaches 43 frames per second, which meets the needs of practical engineering.