To address the challenges in performance evaluation of complex equipment, such as multi-stage transition, diverse test indicators, and difficulty in integrating historical information, a multi-level dynamic evaluation (MLDE) model for fusing multi-stage test data is proposed based on the evidential reasoning (ER) rule. Firstly, some ER sub-models are constructed to tackle the challenge of indicator heterogeneity, with standardized fusion of test data at each stage achieved. Secondly, an information attenuation factor is introduced to quantify the impact of performance in historical stages on the current state, thus overcoming the limitation of insufficient utilization of in-stage information. Finally, a weight discount factor based on performance attenuation is designed, enabling adaptive fusion of multi-stage performance states. Case studies on the ZHS-2 rotor test system have demonstrated that the MLDE model not only achieves desirable evaluation accuracy but also features a transparent and traceable reasoning process, with a reliable basis provided for fault tolerance and maintenance decision-making of complex equipment.