The intelligent decision-making system has significant application value in fields such as autonomous driving and industrial control. However, the performance of existing decision systems largely depends on the quality of information acquisition and analysis, which leads to the following limitations. Firstly, high-precision information processing requires the configuration of high-performance sensors and computing units, which increases the system deployment cost and makes it unsuitable for resource-constrained scenarios. Secondly, in dynamic scenarios where the granularity of information gradually becomes clearer over time, the early coarse-grained information is not effectively utilized, causing the overall decision-making benefits to fall short of the theoretical upper bound. Therefore, this paper first designs a temporal scenario with progressively clarifying objectives, which simulates the dynamic characteristic of information granularity gradually improving as the agent approaches the target. Additionally, a risk assessment mechanism is incorporated into the decision-making process. Based on this scenario, a Two-Stage Multi-Objective Dynamic Policy Network (TSMODPN) is proposed. In this model, the first stage involves offline training of a strategy pool with diverse preferences, while the second stage employs an online dynamic policy network. This network first employs a dynamic distillation module to efficiently extract features from initial fuzzy information. Thereafter, the temporal preference selection module dynamically selects optimal strategies from the offline strategy pool based on these extracted features. The experimental results explicitly show that TSMODPN surpasses other baseline models in its ability to approach the theoretical upper bound for both safety benefits and task duration across varying risk levels.