Aiming at the problems such as low positioning accuracy, map drift and poor reliability of single-sensor simultaneous localization and mapping (SLAM) technology in designed environments, this paper proposes a refined multi-sensor fusion SLAM (RMF-SLAM) algorithm based on the R3LIVE framework. Firstly, a kinematic model enhanced by stochastic processes is designed, and the IMU measurement is modeled as the output. This method can accurately locate and reliably map the intense motion even when the IMU measurement motion is saturated. Secondly, a LiDAR and visual degradation perception module based on the discrimination of eigenvalue degradation of the Hessian matrix is constructed. This module dynamically adjusts the weight ratio of different sensor information and screens high-value visual observation frames by evaluating the system status and sensor reliability in real time. When both vision and LiDAR are extremely degraded, the system slumps the current map preventing positioning failure, and reactivates the dormant map when the sensor works normally again. Finally, a method for similar detection of the map using global descriptors is proposed, integrating the corresponding sleeping map into the current active map, thereby forming a highly accurate global map after the system operation is completed. By comparing with the classic SLAM algorithm on the public dataset, and verifying in the private dataset and real scenarios, the proposed algorithm can effectively suppress the negative impact of the degraded environment on trajectory estimation and map construction, and its accuracy and reliability are superior.