To address the strong subjectivity of prior data in dynamic Bayesian network (DBN) for dynamic risk assessment, this study focuses on a coal⁃slurry preparation system and uses a long short⁃term memory recurrent neural network (LSTM⁃RNN) to optimize the prior data of the DBN. Through its gating mechanism and cell state, the LSTM⁃RNN dynamically screens and retains key information, effectively captures early⁃stage features in long sequences, filters redundant noise, and robustly learns long⁃term dependencies. By learning the time‑series historical operational performance data of the system, the prior data of the DBN is optimized. Bidirectional inference is then performed on the optimized DBN to achieve dynamic reliability analysis of the coal⁃slurry preparation system. The results show that forward inference yields the optimized trend of system reliability, whereas backward inference identifies key events and weak links before and after optimization.