基于定点化自适应选择卷积神经网络的电力缺陷识别方法

Power Defect Recognition Method Based on Fixed-point Adaptive Selection Convolution Neural Network

无人机(unmanned aerial vehicle,UAV)电力巡检时的图像数据量急剧增加,为使深度卷积神经网络(deep convolutional neural network,DCNN)在低功耗的前提下仍能准确识别出电力缺陷,将重点改进传统DCNN算法,以减少机载前端平台中DCNN模型的计算成本,有效提高电力缺陷识别的运行速度,进而延长巡检无人机的续航里程。首先将卷积网络中的浮点运算进行定点化近似,然后通过快速机器学习算法对DCNN模型的输入图像进行自适应选择,最后通过实验对所提方法进行了验证。实验结果证明,DCNN模型经8比特定点优化和自适应选择选择策略后的准确率达88.2%,推理时间缩短了65.9%,能耗减少了71.9%,查准率提高了9.8%。所设计的定点化DCNN模型自适应选择策略不仅能节约电力巡检系统的功耗,而且能提高电力缺陷识别的精度。

The image data of unmanned aerial vehicle(UAV)during power inspection increase sharply.In this paper,we focus on improving the traditional deep convolutional neural network(DCNN)algorithm to reduce the computational cost of DCNN model in airborne front-end platform,to effectively improve the speed of power defect identification,and to extend the range of inspection UAV.Firstly,the floating-point operation in convolution network is approximated by fixed-point,and then the input image of DCNN model is adaptively selected by fast machine learning algorithm.Finally,the proposed method is verified by experiments.The experimental results show that the accuracy of the DCNN model which is optimized by 8 bit fixed-point and adaptive selection strategy is 88.2%,the reasoning time is reduced by 65.9%,the energy consumption is reduced by 71.9%,and the precision is improved by 9.8%.The adaptive selection strategy of the fixed-point DCNN model designed in this paper can not only save the power consumption of power inspection system,but also improve the accuracy of power defect identification.