A Robust Classification Method for Power Transformer Status Recognition Based on Sound Signals

Sound signals of the power transformer contain rich information of operating status. A robust classification method based on sound signals is proposed to determine whether the transformer is abnormal. Once the defect is detected, our method will automatically identify the defect category. Firstly, the raw sound signals of power transformer are preprocessed by framing and windowing before the step of Discrete Short Time Fourier Transform (DSTFT) [1]. Then, Mel filters [2] are applied to obtain the Mel spectrum of the sound data. Secondly, we establish a model to extract the features from the Mel spectrograms [3] and classify these features based on Convolutional Neural Network (CNN) [4]. The output of the fully connected layer of the CNN corresponds the probabilities of each category. The combination of adaptive moment estimation (Adam) [5] optimizer and categorical cross entropy loss [6] is applied to improve the performance of our recognition model. Finally, the test results on data collected from the transformer substation show that the method based on sound signals recognition can effectively identify power transformer anomalies.