<p>Considering the inevitable manufacturing and assembly errors, as well as environmental disturbances, this study investigates the uncertainty quantification and dynamic model updating of the parallel helical gear transmission system (PHGTS) based on the Bayesian approach. The primary computational bottleneck in updating the PHGTS model stems from the repeated numerical integration of its complex gear dynamics, requiring multiple calls to Runge–Kutta solvers during each sampling iteration. This leads to extremely high computational costs and slow convergence in model updating. To address this issue, this study proposes an adaptive reinforcement learning-enabled Bayesian approach (ARLEB). First, a bending-torsional-axial coupling dynamic model of the PHGTS is established. By introducing a probability model with hyper parameters, the uncertainty is embedded in the model parameters. Furthermore, an adaptive reinforcement learning algorithm is embedded within the Bayesian framework, combined with the transition Markov Chain Monte Carlo method (TMCMC) to obtain the posterior distribution of hyper parameters. A three-degree of freedom mass-spring example is first employed to verify the performance of the ARLEB approach and generate the Q-table for dynamic model updating of the PHGTS. The results demonstrate that the ARLEB approach not only promotes the convergence of TMCMC but also improves the accuracy. Subsequently, the dynamic model updating is performed according to the vibration response obtained from the gear test bench, and the impact of uncertain parameters on the dynamic response is analyzed. The research can guide the dynamic performance analysis of the PHGTS.</p>