<p>Early fault diagnosis in gearbox systems remains a critical and challenging task in industrial condition monitoring, particularly within the key health state interval H ∈ [0.7, 1.0] (where H ∈ [0, 1] with 1 as perfect health and 0 as failure), where conventional single-modal approaches often fail to provide reliable detection. This study proposes a novel multi-modal graph neural network (GNN) framework that effectively fuses oil analysis and vibration monitoring data through a physics-informed graph construction. This graph explicitly models intrinsic relationships between tribological phenomena and mechanical dynamics, enabling enhanced feature integration via an advanced cross-modal attention mechanism. Experimental validation on a comprehensive dataset comprising 5000 synchronized multi-modal samples demonstrates significant improvements over current state-of-the-art methods, including recent Transformer-based baselines. The mean absolute error (MAE) is reduced by 16.9 %, detection accuracy reaches 93.7 % with a specificity of 91.2 %, and early fault detection sensitivity exceeds 90 % within the critical health state range H ∈ [0.7, 1.0]. While validated primarily in a laboratory, the framework exhibits robustness to simulated industrial noise (&gt; 88 % accuracy at SNR = 10 dB) and was successfully applied to a real-world ironmaking plant reducer for field validation. The model achieves an inference latency of 28.5 ms (on NVIDIA RTX 3080), which, coupled with inline oil sensors, facilitates near real-time industrial applications. Statistical significance testing (Bonferroni-corrected) confirms the robustness of these gains. These results offer a promising paradigm for multi-modal condition monitoring in rotating machinery and highlight the value of physics-informed machine learning in industrial fault diagnostics, further validated on real industrial cases demonstrating consistent early detection.</p>