Abstract: Theory-data fusion is an emerging paradigm for intelligent decision-making and control in complex industrial systems. By synergistically integrating theory-driven models—rooted in physical mechanisms, experimental analysis, and domain knowledge (Theory)—with data-driven approaches for modeling uncertainties from real-world industrial data (Data), this paradigm overcomes the limitations of traditional single-approach methods. At its core, Theory-Data Fusion establishes a “dual-core, triple-mechanism” framework: a dynamic industrial knowledge graph serves as the knowledge center (Theory), enabling structured cognition and logical reasoning, while a cluster of physics-guided hybrid intelligence models function as the computational center (Data), facilitating pattern recognition and compensatory computation. The three key mechanisms include: intelligent scheduling to enhance the precision of AI interventions, safety verification to ensure physical consistency and reliability, and co-evolution to promote autonomous learning and continuous optimization between the knowledge and computational centers. This approach combines the high reliability and interpretability of physical models with the nonlinear fitting and adaptive capabilities of data-driven models, effectively addressing the limitations inherent to each. Theory-Data Fusion provides critical support for advancing industrial systems from automation to autonomy, and is of great significance for enhancing the adaptability, reliability, and self-evolution capabilities of such systems.