Accurate positioning algorithms are crucial for autonomous vehicle navigation and robotics. The fusion of data from GNSS, INS, and odometers can provide comprehensive positioning results across various environments. However, effectively integrating data from sources with varying reliability levels remains a significant challenge. To address this challenge, we propose a fusion positioning framework that dynamically optimizes the weights of navigation sources. This framework leverages a plug-and-play factor graph algorithm and utilizes a padding mask to flexibly extract features from opportunistically acquired sensor data. It learns the relative fusion weights of different navigation systems based on these data features, thus offering more robust and accurate positioning results in complex and dynamic urban environments. Comprehensive experiments and evaluations demonstrate the effectiveness and superiority of our algorithm.