Abstract: The rise of large AI models is rapidly reshaping the smart home sector. However, these systems often struggle with scattered knowledge, scene adaptability, and personalized services. To overcome these hurdles, this paper presents a smart home scenario self-generation technique based on a knowledge tower cluster. We design a layered appliance ontology and integrate large models using a synergistic optimization approach rooted in the tower cluster's knowledge graph, bolstering model reliability and generalization in smart homes. By capturing multi-dimensional inputs from users, devices, environments, and spaces, we create precise user profiles and autonomous scene organizers with graph neural networks and transfer learning. This enhances task adaptability and self-learning, enabling proactive, customized user experiences. Testing reveals that our approach achieves a scenario generation accuracy of over 95%.

Key words: Knowledge graph, Incremental learning, Intent recognition, Scene generation