With the development of industrial automation and intelligence, the use of machine learning techniques for blast furnace (BF) fault diagnosis has become increasingly important. The decision tree model, due to its intuitive and easy-to-interpret characteristics, has been widely applied in the field of BF fault diagnosis. However, for the ironmaking process there are high dimension, nonlinear and Strong Coupling, the construction of the traditional decision tree model is easy to fall into the local optimal solution, with low efficiency and high complexity. To tackle the above issues, this paper firstly introduces the trace distance function and proves that any local optimal solution is also a global optimal solution in the trace distance function. It then proposes a decision tree model based on trace distance partitioning for the node splitting process in decision trees, referred to as TraceTree. On the one hand, this model evaluates the division effect of a node more quickly and reduces the complexity of the decision tree model effectively. On the other hand, it can identify the features that contribute the most to fault diagnosis and obtain higher diagnosis accuracy. Finally, the comparison with other improved models shows that the model can achieve optimal diagnosis of BF faults with less training time, and monitor and diagnose the BF conditions in a timely manner.