An online fault diagnosis method for converter transformer based on multi-modal information fusion

In response to the issues of variability and sample missingness in multimodal data for transformers, a multimodal information fusion method based on vibration signals and infrared images is proposed. This method aims to effectively and rapidly assess the fault status of power transformers using multimodal data. First, the method employs a bidirectional gated neural network to extract feature vectors from the time-domain information and frequency domain diagrams of vibration and magnetic field signals, as well as from the infrared images of the transformer, obtaining important feature vectors across different modalities. Then, a cross-attention mechanism is used to establish connections between different modalities and fuse the feature vectors. Finally, the fault status of the power transformer is output through convolutional layers and fully connected layers. Experimental data were collected from a 10kV transformer, including vibration, magnetic field signals, and infrared images of the transformer. The experimental results show that the proposed multimodal information fusion method outperforms the comparison methods in four evaluation metrics, with a fault diagnosis accuracy rate of 96%. Under different voltage and current levels, the multimodal information fusion method can achieve relatively reliable diagnostic results with high accuracy, providing a method for fault detection in multimodal data of transformers.