The worktables of core components in CNC machine may experience wear and failure over extended periods of use, leading to reduced machining accuracy. To accurately predict positioning accuracy loss in workbenches under various factors as a function of usage time, this study proposes a modeling and prediction method for one-dimensional workbench positioning accuracy loss based on dynamic Bayesian network. The effectiveness of the proposed method is validated through comparative analysis of measured error data and predicted data. First, the composition of error sources is determined based on the structural analysis of the one-dimensional workbench. According to the established theoretical model of accuracy loss under complex operating conditions for the one-dimensional workbench, load, speed, temperature, and time are identified as the primary factors influencing the positioning error of the workbench. Second, experimental platform was constructed to measure positioning error data under various influencing factors. The validity of the theoretical model was verified based on cloud map results. Next, we incorporate the time dimension to construct a dynamic Bayesian network prediction model for workbench positioning errors under multi-factor influences. We sequentially determine the basic structure of the dynamic Bayesian network, its network nodes, and the ranges of variable domains. Subsequently, we employ mathematical statistics and the EM algorithm for parameter learning, obtaining the prior probability distribution for root nodes and the conditional probability for non-root nodes. Finally, using forward and backward positioning error as an example, the dynamic Bayesian network clustering inference algorithm was employed to predict workbench positioning error. Simultaneously, the predicted error was compared with the measured error under identical conditions. Results indicate that both the predicted forward and backward positioning error curve and the measured error curve generally increase over time, exhibiting similar trends. The maximum absolute error reached 1.63 μm, while the maximum relative error was 13.471%. This validates the effectiveness of the prediction model.