Spoken Language Understanding (SLU) stands as a pivotal element within task-oriented dialogue systems, where it leverages the dialogue context to steer intent detection at the utterance level and slot filling at the token level. Nonetheless, the challenge of judiciously assimilating dialogue context into multi-turn SLU persists as a formidable hurdle. This difficulty is compounded by the inherently dynamic distribution of conversational information in real-world settings, where key details, such as shifts in intent and behavior, can be overshadowed by a deluge of less critical data. To address this issue, we introduce a novel SLU model tailored for intent detection and slot filling in multi-turn interactions. Central to our approach is the incorporation of a dynamic graph convolutional network that selectively amalgamates essential historical information into the dialogue context, thereby enhancing the model's sensitivity to contextually relevant cues. We subject our model to rigorous evaluation on three widely recognized benchmark datasets: SIM, CMCC and FewJoint. The experimental outcomes underscore the model's superior performance, and further results from real-scene datasets strengthen the effectiveness of the method.