The Job Shop Scheduling Problem is a classic combinatorial optimization problem with significant theoretical and practical implications for real-world production activities. Existing methods typically utilize disjunctive graphs combined with Graph Neural Networks for feature extraction. However, disjunctive graphs offer limited representational capacity, and edges for different relationships can interfere with each other, posing challenges in depicting complex operational dependencies. To address this, this paper proposes an algorithm based on multi-view feature fusion. This approach decouples different relationships among operations, expands the topological structure by incorporating prior knowledge during the scheduling process, and employs a multi-head self-attention mechanism for feature fusion to provide richer information, thereby enhancing algorithm performance. Extensive experiments conducted on the TA benchmark datasets demonstrate that the proposed algorithm achieves superior average solution performance compared to other baseline methods.