| 引用本文: | 贺才郡,李开成,董宇飞,等.基于知识蒸馏与RP-MobileNetV3的电能质量复合扰动识别[J].电力系统保护与控制,2023,51(14):75-84.[点击复制] |
| HE Caijun,LI Kaicheng,DONG Yufei,et al.Power quality compound disturbance identification based on knowledge distillation and RP-MobilenetV3[J].Power System Protection and Control,2023,51(14):75-84[点击复制] |
|
| 摘要: |
| 针对复合电能质量扰动(power quality disturbance, PQD)识别中特征提取复杂、识别正确率低和模型难以轻量化等问题，提出一种利用递归图(recurrence plot, RP)对PQD信号可视化方法和基于知识蒸馏的模型训练方法。首先，基于RP挖掘PQD信号隐含特征并构建图像数据集，并利用深度残差收缩网络(deep residual shrinkage network, DRSN)对图像数据集进行更深层次特征提取并完成自主分类。然后，基于知识蒸馏(knowledge distillation, KD)让已训练的DRSN指导轻量化网络MobileNetV3进行训练，通过蒸馏实现知识的跨网络传输。最后，仿真实验和硬件实验表明，利用知识蒸馏训练的MobileNetV3能实现高精度且轻量化的复合扰动识别，同时在30 dB噪声环境下正确率能提升1.06%，对实际扰动信号识别效果良好，具有良好的噪声鲁棒性。 |
| 关键词: 电能质量扰动 递归图 图像 深度残差收缩网络 知识蒸馏 MobileNetV3 |
| DOI：10.19783/j.cnki.pspc.221856 |
| 投稿时间：2022-11-25修订日期：2023-02-09 |
| 基金项目:国家自然科学基金项目资助(522077089) |
|
| Power quality compound disturbance identification based on knowledge distillation and RP-MobilenetV3 |
| HE Caijun,LI Kaicheng,DONG Yufei,SONG Zhaoxia,XIAO Xiangui,LI Beiao,LI Xuan |
| (School of Electrical and Electronic Engineering, Huazhong University of Science and Technology, Wuhan 430074, China) |
| Abstract: |
| To solve the problems of complex feature extraction, low recognition accuracy and model weight reduction in the recognition of complex power quality disturbance (PQD), a method of PQD signal visualization using recurrence plot (RP) and a model training method based on knowledge distillation are proposed. This method first mines the hidden features of PQD signals based on RP and builds image data sets. It then uses a deep residual shrinkage network (DRSN) to extract deeper features of image data sets and complete classification. Finally, based on knowledge distillation (KD), it lets the trained DRSN guide the training of the lightweight network MobilenetV3, and realizes cross network transmission of knowledge through distillation. The simulation experiment and hardware experiment show that MobileNetV3 trained by knowledge distillation can achieve high-precision and lightweight composite disturbance recognition, and the accuracy can be improved by 1.06% in a 30 dB noise environment. It has good recognition effect on actual disturbance signals and good noise robustness. |
| Key words: power quality disturbances recurrence plot image deep residual shrinkage network knowledge distillation MobileNetV3 |
