부산대학교 도서관

This paper proposes a new approach for power signal disturbances (PSDs) classification using a two-dimension (2D) deep convolutional neural network (CNN). The data preprocessing stage introduces a conversion method from signal to the 2D grayscale image. Firstly, the signal is divided into multiple cycles. The zero-crossing rate is adopted to specify a cycle's start and endpoints. Then, the cycles are transformed into matrices. Next, the matrices are merged into a new form matrix. Lastly, the matrix is converted into the 2D image grayscale. The obtained 2D image preserves information and waveform the sinusoidal of the signal. The experiment was carried out on datasets containing 14 different disturbance categories with the same model learning structure. The results show that the 2D deep CNN performs better than the one-dimension (1D) deep CNN. According to this result, the 2D deep CNN can improve the PSDs classification effectiveness. Furthermore, the proposed method outperforms the conversion method used in previous studies.