Multi-modal information analysis for fault diagnosis with time-series data from power transformer.

• A multi-model information analysis is proposed for transformer fault diagnosis. • The MIA consists of a SKnet, a BGRU, and a cross attention mechanism. • The cross-attention mechanism learns mutual effects among diverse modalities. • The algorithm has good performance in fault diagnosis of the power transformer. Fault diagnosis is important to the timely repair of the power transformer. However, machine learning has not been exploited effectively for fault diagnosis due to the limitation of multi-modal heterogeneity of data and the ratio of missing samples. To solve this problem, a novel multi-modal information analysis method is presented to effective and speedy evaluate power transformer fault with time sequences and multi-modal data. The proposed method consists of a Selective Kernel Network, a bidirectional gated recurrent unit, and a cross attention mechanism. The proposed approach is verified by datasets of dissolved gas and infrared image modes which come from real power transformers and the historical data. The results show the advantage and efficiency of the proposed method for its higher diagnostic accuracy and shorter diagnostic time than those of the comparison approaches. [ABSTRACT FROM AUTHOR]