无人机图像配电导线断股检测的深度学习方法

Deep Learning Method of Distribution Broken Strands Intelligent Detection via Unmanned Aerial Vehicle Images

配电导线的断股易导致断线事故,给用电需要和用电安全带来了极大的负面影响。传统人工检视方法费时费力,而基于无人机(unmanned aerial vehicle,UAV)获取导线影像,利用全卷积网络(fully convolutional network,FCN)深度学习的方法进行配电导线断股快速准确检测可以事半功倍。首先,顾及断股图像空间低占比特点,利用卷积层代替池化层以减少细节损失,实现改进FCN网络架构构建;其次,通过图像变换方法增强原始数据集,提升网络泛化能力;然后,选用BReLU(bilateral rectified linear unit)激活函数,弥补常用激活函数存在的梯度弥散缺陷,提高准确识别率;最后以某电力配电导线巡检项目为例,训练、优化网络,并对UAV获取的图像进行检测,取得了93%的正确率。同时,对比分析了该方法与传统方法的检测效果,结果表明了在配电导线断股智能识别中该方法的鲁棒性和准确性显著占优。

The strand breaking of the transmission lines is likely to cause wire breakage accidents,which has a great negative impact on electricity demand and safety.The traditional manual inspection method is time-consuming and laborious.Therefore,this paper proposes a method uses UAV(unmanned aerial vehicle)to obtain wire images,and using FCN(fully convolutional network)deep learning method to quickly and accurately identify the broken strands of the distribution wire.Firstly,we take into account the low bit-occupancy of the broken strand image space,and use the convolutional layer instead of the pooling layer to reduce the loss of details.Secondly,the original data set is enhanced by the image transformation method to improve the network generalization ability;then,the BReLU(Bilateral Rectified Linear Unit)activation function is used to make up for the defect of gradient dispersion and improve the accuracy of recognition.Finally,the network of an electric power distribution conductor inspection project is trained and optimized as an example,and the images acquired by the UAV are detected,and a 93%correct detection rate is obtained.At the same time,this paper comparatively analyzes the detection effect of this method and the traditional method.The results show that the robustness and accuracy of this method is significantly superior in the intelligent identification of distribution conductor broken strands.