In order to solve the problem of slow loading speed and high memory uage when loading the scene and model at one time in the virtual maintenance training system of machine service, we propose a scene management method  based on task relevance. The method uses the TFIDF algorithm to get the similarity of each task card in the virtual maintenance system and then classify them. The high similarity of work cards indicates that the correlation between virtual resources such as virtual maintenance scene, maintenance tools and maintenance objects is stronger. When loading scene resources and allocating memory, the virtual resources described by the task cards with a relevance greater than 68% are loaded and distributed by the buddy system. The task scenarios with relevance less than 42% apply for a memory in the buddy system and then the memory is divided into memory pools. The sub scenes with a task reference between 42% and 68% are managed by the double dynamic double linked list. The proposed method solves the problem that the traditional virtual maintenance training system has to face. That is there lacks maintenance resource related allocation management when loading resources. The allocation method has no taskspecific limitation, and avoids system allocation time for separate memory blocks. Experimental results show that the improved allocation method reduces memory usage by 17% and improves the frame rate by 17.57.