Abstract: This paper proposes a novel approach to quantitative evaluation of actual fault diagnosability for dynamic systems. This approach, which can quantify the difficult level to detect and isolate a fault without designing a diagnosis algorithm, provides a guidance and reference for the technical purpose of improving fault diagnosis ability in the system design phase. First, fault diagnosability evaluation for the dynamic system described by a state space model is converted to a statistical probability problem for distinguishing the similarity of different multivariate distributions through the model standardization and the parity space method. KLD (Kullback-Leibler divergence) is introduced to present the principle of evaluating fault diagnosability based on the criterion of distance similarity, and the shortages of this method are pointed out through rigorous proofs. To make up these shortages, a new method for quantitative diagnosability evaluation is designed, which is derived from the probability distribution of fault vectors and the cosine similarity between two different fault vectors in the view of directional similarity. Finally, the validity and superiority of the proposed approach are verified by numerical simulations.