异常放电是电力变压器中一种潜在的危险故障，若未及时检测可能导致严重的安全事故。采用接触式拾音器收集变压器箱体内异常放电声纹信号，并提出了一种特征提取方法和一个深度神经网络结构，以实现对变压器异常放电的高效识别。首先，设计了一种结合梅尔频率提取和关键频率提取的二维声纹特征提取方法。其次，提出了一种基于卷积神经网络和Transformer网络的混合二维特征识别模型，能够在确保识别速度的同时准确辨识异常放电声纹信号。通过对110 kV三相三绕组变压器无载调压试验过程中采集的放电数据进行试验分析，所提方法相较于ResNet50识别速度增加约0.16秒/样本，同时识别效果提升了4.5%。

Abnormal discharge is a dangerous power transformer fault, which can lead to serious safety hazards if not detected in time. A method for identifying abnormal discharge in transformer is proposed by collecting the voiceprint signal in the transformer box through the contact voice pickup. And a feature extraction method and a deep neural network structure are proposed to achieve efficient identification of abnormal transformer discharge. Firstly, a two-dimensional voiceprint feature extraction method combining Mel frequency and key frequency is designed. Then, a hybrid two-dimensional feature recognition model based on convolutional neural network and Transformer network is used to accurately identify the abnormal discharge voiceprint signal while ensuring the speed. Finally, according to the experimental analysis of the discharge data collected from the 110 kV three-phase three-winding transformer in the no-load voltage regulation process, the recognition speed of the proposed method increases by 0.19 seconds per sample, and the accuracy increases by 4.5% compared with ResNet50.

新型电力系统智能运维安徽省联合共建学科重点实验室成果