In the context of the accelerated digital transformation of supply chains and the continuous growth of multi-source heterogeneous data, how to break through the information barriers in procurement, production, inventory, logistics, and sales, has become a key issue in intelligent supply chain management. This paper constructs a multi-source data hierarchical modeling framework, adopts single-source feature encoding, cross-source temporal alignment, dynamic weight fusion, and global state representation methods, and further designs shared decision representation, task adaptive gating, multi-task collaborative output, and confidence constraint mechanisms to achieve integrated decision-making for demand identification, inventory judgment, distribution response, and risk warning. Experiments based on 12,000 samples show that the accuracy, precision, recall rate, and F1 value of the proposed method reach 92.8%, 91.9%, 93.4%, and 92.6% respectively, which are 2.5 and 2.9 percentage points higher than the accuracy and F1 value of the Transformer model, and achieve the best comprehensive performance at a threshold of 0.80. This research provides a feasible technical path for the intelligent fusion and decision support of multi-source data in supply chains.
