Proceedings Article | 15 January 2024
Peng Zhang, Guoliang Zhang, Xiaoyu Fan, Jinrui Gan, Peng Wu, Fei Zhou, Shunli Lv, Long Chen, Xin Liu
KEYWORDS: RGB color model, Data modeling, Transformers, Histograms, Data fusion, Power grids, Feature extraction, Reflection, Pulse signals, Image processing
Transformers are key hub equipment in the power system, and their operation reliability is directly related to the stability of the power grid. In order to ensure the safe and stable operation of the transformer, monitoring devices such as dissolved gas analysis (DGA), partial discharge (PD), and voiceprint vibration are installed to obtain multimodal data such as structured data, visible light pictures, atlas pictures, and voiceprint data. However, existing fault diagnosis methods cannot integrate these multi-modal features to evaluate the health state of the transformer. In order to solve these problems, a transformer multi-modal data knowledge fusion coding method based on attention model is proposed. For structured state quantities such as online monitoring state quantities and test state quantities, Bi-LSTM is used for encoding to obtain structured state quantity codes. For the partial discharge state quantity, multi-dimensional structured data is extracted from the n-q-φ spectrum and discharge waveform spectrum, and Bi-LSTM is used to obtain the partial discharge structured state quantity codes. For RGB images, an encoding structure including a convolution layer, a pooling layer and a coding layer is constructed to obtain partial discharge image codes. Then, structured state quantity codes, partial discharge structured state quantity codes and partial discharge image codes were fused based on the Attention model, and the fused multi-modal data knowledge codes were used in the fault diagnosis process of the transformer. Multimodal data knowledge encoding technology is applied to transformer fault diagnosis. The application results show that the fault diagnosis accuracy can be significantly improved after fusion and encoding of multimodal data. Compared with the traditional structured state quantity diagnosis model, the accuracy of fault diagnosis is improved. 8.75%, up to 92.50%.
