Model learning of timed automata (TA) aims to build a formal model of software and hardware systems by external inputs and outputs. Learning of deterministic one-clock TA is one of the important research directions, but current algorithm has some limitations and is difficult to be applied to complex systems. Therefore, an improved learning algorithm is proposed, which uses logical-time classification tree instead of logical-time observation table as the internal data structure of the learning algorithm, effectively reducing the number of membership queries and the space complexity of the algorithm. In addition, it can efficiently construct hypothetical automata. Finally, relevant experiments have been carried out, and the experimental results show that the improved algorithm proposed in this study reduces the number of member queries by 60% and the number of equivalent queries by 5%. At the same time, in this experiment, the learning speed of the improved algorithm can be increased by more than 50 times at most.