Journal of South China University of Technology(Natural Science Edition) ›› 2022, Vol. 50 ›› Issue (2): 129-136.doi: 10.12141/j.issn.1000-565X.210531

Special Issue: 2022年能源、动力与电气工程

• Energy, Power & Electrical Engineering • Previous Articles     Next Articles

Recognition of Bird Sounds Related to Power Grid Faults Based on Mel Spectrogram and Convolutional Neural Network

QIU Zhibin1 LU Zuwen1 WANG Haixiang1 KUANG Yanjun2   

  1. 1.Department of Energy and Electrical Engineering,Nanchang University,Nanchang 330031,Jiangxi,China;
    2.Power Research Institute of State Grid Jiangxi Electric,Nanchang 330096,Jiangxi,China
  • Received:2021-08-22 Revised:2021-11-16 Online:2022-02-25 Published:2022-02-01
  • Contact: 邱志斌（1991-），男，博士，校聘教授，主要从事输变电设备外绝缘、电网图像识别与人工智能等研究。 E-mail:qiuzb@ncu.edu.cn
  • About author:邱志斌（1991-），男，博士，校聘教授，主要从事输变电设备外绝缘、电网图像识别与人工智能等研究。
  • Supported by:
    Supported by the Jiangxi “Double Thousand Plan” Innovative Leading Talents Long-Term Project(jxsq2019101071) and the Jiangxi Natural Science Foundation for Young Scholars(20192BAB216028)

PDF

58

Abstract: In order to achieve targeted prevention of bird-related outages in power grid, this paper proposed a bird sound recognition method based on Mel spectrogram and convolutional neural network (CNN). The study established a sound signal sample set of 40 kinds of representative bird species related to the common faults and preprocessed the bird sound signals by framing, windowing, and denoising, thus to calculate the signal energy of each frame in each Mel filter. The Mel spectrograms of bird sound signals were extracted according to the mapping relationship between the energy and the color depth. Moreover, the Mel spectrograms of the bird species related to po-wer grid faults were used as inputs, and the CNN was used for feature extraction of the Mel spectrograms by perfor-ming the convolution-pooling process repeatedly. Several iterations of training were carried out to adjust the internal parameters of the network, thus to obtain the optimal model for bird recognition. A case study results show that the recognition accuracy of 40 bird species reaches 96.1%, and that the proposed method is better than other transfer learning models. This study can provide references for transmission line operators to identify the related birds correctly and carry out differential prevention of bird-related outages.

Key words: transmission line, bird-related outages, bird recognition, Mel spectrogram, denoising, convolutional neural network

CLC Number: