基于鸣声组合特征与CNN的电网危害鸟种识别

Identification of Harmful Bird Species in Power Grid Based on Combined Sound Features and CNN

为了辅助电网涉鸟故障的差异化防治,提出一种基于组合特征和卷积神经网络(Convolutional Neural Network,CNN)的电网危害鸟种鸣声识别方法.根据历史涉鸟故障的鸟种信息及输电线路周边鸟种调查结果,选择13种高危鸟类、8种微害鸟类和2种无害鸟类建立鸣声样本集;对鸟种鸣声信号进行分帧、加窗、降噪和剪裁等预处理,提取鸟鸣Mel倒谱系数(Mel-frequency Cepstrum Coefficients,MFCC)、Gammatone倒谱系数(Gammatone Frequency Cepstrum Coefficients,GFCC)和短时能量(Short-term Energy,STE)特征.针对单一特征表达能力不足的问题,将MFCC及其一阶差分、GFCC及其一阶差分和STE归一化后进行组合,生成新的鸣声特征集.搭建卷积神经网络模型对组合特征进行训练和识别,鸟种鸣声测试集的识别正确率达91.8%,较单一MFCC和GFCC特征表现更为优异.

In order to assist differentiated prevention of bird-related faults in power grid,this paper proposes a method for the identification of bird species related to power grid faults based on combined features and a Convolu⁃tional Neural Network(CNN).Firstly,based on the information from historical bird-related faults in the power grid and the investigation results of bird species around transmission lines,13 high-risk,8 low-risk,and 2 harmless bird species were selected to build a sound sample set.Then,the Mel-frequency Cepstrum Coefficients(MFCC),Gamma⁃tone Frequency Cepstrum Coefficients(GFCC),and Short-term Energy(STE)features of bird sounds were extracted after preprocessing such as framing,windowing,noise reduction,and clipping.To solve the problem of insufficient ex⁃pression ability of a single feature set,a new sound feature set was generated combining MFCC,GFCC,their first order differences,and STE features after normalization.Finally,a CNN was built to train and recognize the combined features.The identification accuracy of the test set reaches 91.8%,which is better than those with a single MFCC and GFCC feature set.