A voiceprint pattern recognition method of smoothing reactor based on CNN

Abstract

In order to accurately identify the operating condition of the smoothing reactor， a deep learning method based on CNN （convolutional neural network） is introduced. A voiceprint pattern recognition model for reactor windings using Mel spectrogram is developed. The sound signals are collected using dry smoothing reactors as the experimental object. The Mel filter method is used to convert the collected sound signals into a spectrogram with different working conditions used as the labels of the data set. The CNN algorithm is used to identify the working conditions corresponding to the different signals. The results show that CNN can be used to accurately identify dry voiceprint patterns of smoothing reactors. The optimized neural network can achieve an accuracy of 98.4% in recognition of voiceprint signals under sinusoidal excitation， harmonic excitation and DC bias excitation. The research results provide a potential technical solution for realizing intelligent detection of power grid signals.

Keywords

• smoothing reactor
• operating condition
• winding
• mel spectrogram
• convolutional neural network