摘要
边缘计算技术使在智能电力设备端开展基于神经网络的就地智能诊断成为可能,但存在电力智能终端资源受限与局部放电诊断模型高资源占用之间的矛盾。为解决此问题,该文提出了基于深度-广度联合剪枝的电力设备局部放电轻量化诊断方法。该方法将MobileNetV2作为基础模型,在训练中引入可迭代重要度因子,“端到端”地感知并裁剪模型中的冗余模块,实现深度方向的结构压缩;采用几何中值滤波器剪枝(FPGM)进一步去除各卷积层的冗余滤波器,并提出增强型模拟退火搜索算法(ESA)自主求解各层的剪枝比例,循环搜索直至获得最大限度的无损压缩模型。结果表明,该方法可以在数据驱动下自主设计高精度、轻量化、低延迟的局部放电深度诊断模型,相较于现有的深度模型,资源占用大幅降低、推理速度显著提升,为资源受限的电力设备边缘侧部署提供了技术支持。
Partial discharge(PD)is an early indicator on insulation deterioration that will cause catastrophic failure on the power system,so PD diagnosis is a significant approach to monitor the operating status of the electrical equipment.Recently,deep learning(DL)has gradually reached the mainstream in the field of PD diagnosis and the increasing intelligent terminals near power equipment maybe serve as the carrier for such DL models.However,the existing DL-based PD models tend to occupy higher computing resources,while the current power intelligent terminals usually have small memory space and limited calculation capacity.To address it,a lightweight PD diagnosis method based on depth-width joint pruning is proposed in this paper,which can effectively compress the computational resource consumption of DL model while ensuring the accuracy of PD diagnosis.Firstly,a set of 1D PRPD matrix is constructed based on the discharge amplitude,phase and pulse number,including four types of PD defect such as point discharge,surface discharge,air-gap discharge and suspended discharge.Then,this method selects MobileNetV2 as the basic model,and an iterable importance factorαis inserted in the training process to assess the importance of each convolution module.According toα,several modules with low importance factors(close to 0)are pruned to simplify the basic model in depth direction.Finally,to further compress this model,it adopts a filter-level pruning approach called filter pruning via geometric median(FPGM)to remove redundant convolution filters,in which the pruning ratio of filters in each layer is adaptively calculated by an enhanced simulated annealing search(ESA).Through cyclical search,a highly compressed model can be generated with almost no loss of accuracy,while greatly reduces the computational cost and time.The experimental results show that,with the premise of remaining the diagnosis accuracy,the proposed method can automatically design an efficient PD diagnosis model with lightweight architecture and less diagnosis ti
作者
张翼
朱永利
Zhang Yi;Zhu Yongli(School of Electrical and Electronic Engineering North China Electric Power University,Baoding 071003 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2023年第7期1935-1945,1955,共12页
Transactions of China Electrotechnical Society
基金
河北省自然科学基金(F2022502002)
特高压工程技术(昆明、广州)国家工程实验室开放基金资助项目。
关键词
局部放电
深度学习
自动化剪枝
轻量化诊断
结构设计
Partial discharge
deep learning
automatic pruning
lightweight diagnosis
architecture design
