Deep reinforcement learning approaches for flexible job shop scheduling often overlook critical edge features between operations and processing times on operation-machine edges when leveraging heterogeneous disjunctive graphs for feature extraction and processing. To overcome this challenge, a novel heterogeneous progressive attention network is introduced, incorporating a progressive attention mechanism to process different types of node features independently while integrating weighted edge information. This design allows the model to simultaneously capture both node characteristics and edge relationships, enhancing the accuracy and adaptability of scheduling decisions. The proposed approach is evaluated through comparative experiments using CPLEX and OR-Tools. Experimental results demonstrate superior performance over traditional methods in multi-objective FJSP optimization, achieving a more balanced trade-off among multiple objectives while significantly enhancing scheduling quality and computational efficiency.