The Depression of Noises in Acoustic Signals of Transformer Based on Adversarial Auto-encoder

Acoustic signals produced by an operated transformer are the important parameters for the assessment of noise level of substation and condition monitoring of transformer. When acoustic signals of transformer transmit in the air, some interference components are inevitably contained in the measured acoustic signals by microphone sensors with the features of diversity and a certain degree of uncertainty. To obtain the real acoustic signals resulting from the transformer, the adversarial convolutional de-noising auto-coder (CDAE) network is proposed and trained to carefully depress the interference components of acoustic signals in an end-to-end manner. The similarity of power spectrum (SPS) is defined to evaluate the de-noising effects of the proposed method. The acoustic signals for an operated transformer in a substation are measured to build the data set. The calculated results have shown that the adversarial CDAE network is capable of eliminating the interference components contained in acoustic signals of transformer with better generalization ability and high accuracy.