Abstract: Accurate and reliable transient stability assessment is of great significance for the secure and stable operation of power systems. Traditional theoretical judging methods and criteria still face some difficulties when applied to complex power systems, while problems such as poor interpretability still hold in the artificial intelligence related methods. In this paper, based on response-driven maximum Lyapunov exponent induced transient stability judgment mechanism, GAT-GRU coupled network and diffusion kernel density estimation methods are used to predict the probability distribution of maximum Lyapunov exponent to replace the calculation process. The gradient computation algorithm is deduced for KL divergence-based distance measurement. Then, neighborhood approximation estimation with adaptive weights is proposed as stability metrics of probability distribution in substitution of fixed threshold criterion. Results from case studies show that the framework of TSA proposed in this paper can improve the applicability of theoretical judging methods as well as enhance the interpretability of intelligent methods, which is able to realize accurate transient stability judgments with a very small amount of initial transient response information.