HybridCube: Integrating Multi-Sensor Cross-Modal Data for Early Fault Detection in Power Transformers

In recent years, the reliability of power transformers has become increasingly critical due to their central role in the electrical grid. Unexpected transformer failures can lead to significant economic losses and service interruptions. Traditional monitoring methods often rely on limited data sources and modalities, which may not provide a complete picture of the transformer's health. Due to the existence of convergence patterns in the detection data before transformer faults occur, different modal data are used to complement the information and corroborate the conclusions. Therefore, we propose HybridCube, a novel early fault detection framework for power transformers, consisting of three stages: data acquisition, cross-modality fusion, and trusted early fault detection. We design multi-view infrared image and vibro-acoustic acquisition matrices to create a unimodal monitoring cube for simulating the monitoring environment. Cross-modal data fusion is achieved using a hybrid cube based on information entropy theory. The time-series feature vector representation obtained through an encoder is fed into an attention mechanism-based decoder. In terms of reasoning, anomaly detection identifies multi-view anomalous regions enabling predictive spatial reconstruction and providing information about fault type, location, and predicted time. Solid case studies show that HybridCube provides a holistic view enabling early identification of potential issues while improving the reliability and efficiency of transformer operation. This framework demonstrates significant potential in predictive maintenance and reducing unexpected outages.