Abstract: Aiming at the limitations of current converter transformer condition monitoring-such as heterogeneous data misalignment, over-fitting under sample scarcity, and high false alarm rates caused by pure data-driven models lacking physical constraints—this paper proposes an intelligent diagnostic method integrating physics-informed constraints and bidirectional cross-modal attention. First, a time-masking and interval-encoding strategy is designed for spatiotemporal alignment, and a semi-empirical dual-modal dataset driven by real-world DGA is constructed with rigorous temporal-leakage prevention. Second, a bidirectional cross-modal network is established to extract the deep coupled representations between long-term chemical insulation degradation and transient mechanical anomalies. To overcome the physical invalidity of black- box models, this study innovatively transforms discrete classification outputs into continuous differentiable risk expectations and derives a joint physics-constrained loss incorporating gas production penalties and short-term state smoothing. This explicitly guides the network to learn prior-mechanism-compliant decision boundaries. Experiments demonstrate that the proposed method balances high diagnostic accuracy with an extremely low false alarm rate under harsh environments. Integrated with a constructed heterogeneous interpretable mechanism and a hierarchical alarm matrix, it provides rule- consistent, quantitative decision support for the predictive maintenance of power equipment.