电力系统及其自动化学报2025，Vol.37Issue(10)：43-52,10.DOI:10.19635/j.cnki.csu-epsa.001659
变压器局部放电识别的轻量化深度学习模型
Lightweight Deep Learning Model for Transformer Partial Discharge Recognition
摘要
Abstract
To address the problems in the existing deep learning-based models for transformer partial discharge recogni-tion such as large parameter sizes and poor deployability in resource-constrained environments,an enhanced light-weight model AMNetV3 derived from an MobileNetV3-Small backbone is proposed in this paper.First,the channel pa-rameters of the expansion layer in the bottleneck modules are reconfigured to reduce the intermediate feature channels while preserving consistency in the output channels.Second,a parameter-free channel attention mechanism is intro-duced,which derives the response weights based on the mean and variance of feature channels,thereby emphasizing the informative channels and mitigating redundancy.Third,channel pruning guided by the L2 norm is applied to convo-lutional layers with high parameter densities,with 10％and 5％channel pruning implemented to enhance the model sparsity.Finally,just-in-time compilation is employed to accelerate inference,and the model is transformed into Torch-Script format to enable cross-platform deployment.Meanwhile,ablation experiments are conducted to verify the individ-ual contributions of each module.Experimental results demonstrate that the AMNetV 3 model achieves a recognition ac-curacy of 97.92％,with a 22.11％reduction in memory consumption and a 53.08％decrease in recognition time,con-firming its application potential for deployment in resource-limited platforms.关键词
变压器/局部放电/无参数通道注意力机制/神经网络/通道剪枝Key words
transformer/partial discharge/parameter-free channel attention mechanism/neural network/channel pruning分类
动力与电气工程引用本文复制引用
欧蓉姗,朱永利..变压器局部放电识别的轻量化深度学习模型[J].电力系统及其自动化学报,2025,37(10):43-52,10.基金项目
河北省自然科学基金资助项目(F2022502002). （F2022502002）
