Xiaoyan Liu, Y. He
Abstract Accurate health assessment of power transformers is challenged by the scarcity of state labels, a difficulty further compounded by persistent stochastic disturbances arising from high renewable energy integration in China’s power grid. To address this issue, this paper presents a physics-informed time-frequency graph contrastive learning (PTF-GCL) method that leverages online monitoring data. The approach begins by converting multi-sensor signals into visibility graphs that capture their respective temporal characteristics. A graph fusion strategy is then applied to integrate the adjacency matrices of these individual graphs, yielding a unified interactive topology with enhanced and more balanced connectivity. From this fused structure, the PTF-GCL framework learns a joint representation by enforcing alignment between time-domain and frequency-domain views through a time-frequency relational contrastive loss, thereby enabling frequency-domain structural patterns to consistently regularize time-domain learning. This alignment maximizes feature discriminability by increasing the representational distance between graphs corresponding to healthy and degraded conditions. A quantitative integrated health index (IHI) is subsequently derived from the Euclidean distance between graph-level readout vectors. The proposed method incorporates physics-informed experimental data from power transformers, and validation on vibration data collected under identical operating conditions demonstrates superior performance, achieving 96.07% accuracy, 96.13% precision, 96.07% recall, and 96.08% F 1-score, with benchmark comparisons confirming its consistent advantage over existing baselines.