Aiming at the problems that the traditional state evaluation methods focus on the key production units and do not consider the influence of the causal relationship between the monitoring variables on the evaluation results， which leads to the inaccurate results， a performance safety evaluation method for system driven by multivariate causality is proposed. The generalized partial directed coherence method is used to analyze the causal relationship of monitoring variables in the frequency domain， and a causal network model reflecting the running state of the system is established. Based on this model， the key characteristics of the system are extracted from the perspective of multi-dimensional statistics using the average path length， clustering coefficient and network structure entropy. Besides， a multi-dimensional feature fusion index reflecting the performance state of the system is established and the validity of the proposed method is verified by utilizing the fault data of a chemical enterprise. The result shows that compared with single index， the fusion feature can reflect the performance state of the system more comprehensively and accurately.