面向输电线路覆冰厚度辨识的多感受野视觉边缘智能识别方法研究

Receptive Field Vision Edge Intelligent Recognition for Ice Thickness Identification of Transmission Line

输电线路覆冰积雪表现出强随机性和不可抗拒性,导致实际输电线路覆冰应急处理极度困难,亟需覆冰监测终端的边缘智能识别能力。为此,以电力视觉边缘智能为基础,提出了一种基于轻量型多感受野特征表达网络的输电线路覆冰厚度终端级辨识方法。该方法首先通过轻量化的卷积神经网络MobileNetV3提取覆图像的特征,然后引入多感受野模块增大模型对覆冰影像的映射区域,从而增强其特征提取能力,其次采用多尺度目标检测网络(single shot multibox detector,SSD)实现覆冰厚度的辨识与监测。最后,采用实际场景下感知到的覆冰影像在计算资源有限的边缘智能装置中进行试验验证。试验结果表明:所提出的边缘智能识别方法能够实现覆冰厚度的终端级识别,并能在极端天气的覆冰场景下保持较高识别精度,避免了覆冰影像的长距离传输,实现了极端天气下覆冰监测的边缘智能自治,具有很强的泛化能力和实际适用价值。

The strong randomness and irresistibility of the transmission lines covered with ice and snow leads to the extreme difficulty in handling such emergency,therefore the intelligent edge recognition ability of icing monitoring terminal is urgently needed.This paper proposes a front-end identification method of transmission line icing grade based on lightweight receptive field feature expression network.Firstly,a lightweight convolutional neural network named MobileNetV3 is utilized to extract feature information of the icing image,and the receptive field block is introduced to enlarge the mapping area of the model to the icing image,thereby enhancing the feature extraction ability of the network.Then the multi-scale target detection network SSD(single shot multibox detector)is used to achieve the thickness identification and monitoring of icing image.Finally,the ice image perceived in the real scene is experimented in the edge intelligent device with limited computing resources.The experiment results show that the proposed edge intelligent monitoring method can achieve the front-end recognition of ice grade and can maintain high recognition accuracy for ice images collected in the extreme weather.This method greatly avoids long-distance transmission of the ice images and achieve the edge intelligent autonomy of ice cover monitoring in the extreme weather,which has strong generalization ability and practical application value.