改进MobileViT轻量级网络的电网危害鸟种鸣声识别
Sound Recognition of Bird Species Related to Power Grid Faults Based on Improved MobileViT Lightweight Network
宋超 1邵明玉1
作者信息
- 1. 安庆师范大学 数理学院,安徽 安庆 246133
- 折叠
摘要
准确识别输电线路上的危害鸟种对电网涉鸟故障的差异化防治具有重要意义.目前,基于深度学习的鸟鸣识别方法得到了广泛应用,但现有模型存在参数大、效率低等问题,为此,本文通过引入轻量高效的MobileViT网络在MV2模块中融入压缩和激励(Squeeze Excitation,SE)机制,提出了具有更强识别能力的改进模型——MobileViT-SE模型.该研究以典型电网危害鸟种的音频作为实验样本,对MobileViT和MobileViT-SE模型的识别能力进行了评估.结果表明,相比其他典型的深度学习模型,MobileViT和MobileViT-SE模型在参数量更少的条件下识别平均准确率更高,分别达到95.5％和97.2％,且后者提高了1.7％.
Abstract
Accurate recognition of bird species that are harmful to transmission lines is of great significance to the differ-entiated prevention and control of bird-involved faults in power grids. Currently, bird song recognition methods based on deep learning have been widely applied. However, existing recognition deep models have problems of a large number of parameters and thus low training efficiency. To address this issue, we introduce the MobileViT network, a lightweight and efficient model to identify bird species by their sounds. By incorporating the Squeeze Excitation (SE) mechanism into the MV2 module of the standard MobileViT model, we propose the MobileViT-SE model, an improved version of the MobileViT model with stronger recognition capability. We evaluate the recognition ability of both the MobileViT model and the MobileViT-SE model with au-dio samples of typical bird species that are harmful to transmission lines. Experimental results show that compared with other typical deep learning models, the MobileViT model and the MobileViT-SE model have higher average recognition accuracy while using fewer number of parameters. Specifically, the average test accuracy of the MobileViT model is 95.5％, and the av-erage test accuracy of the MobileViT-SE model is 97.2％, which is 1.7 percentage higher than the former.
关键词
电网涉鸟故障/鸟鸣识别/MobileViT-SE模型/压缩和激励机制Key words
bird-involved faults in power grids/bird song recognition/MobileViT-SE model/Squeeze Excitation mechanism引用本文复制引用
基金项目
安徽省科学研究重点项目(2022AH051046)
出版年
2024
国家科技期刊平台
NSTL
万方数据