A Bayesian posterior distribution-based abnormal data processing GM （0， N） model is proposed for complex equipment cost estimation to enhance the accuracy and reliability of cost prediction for sophisticated systems. Taking rocket systems as a case study， potential abnormal data samples were first eliminated using the three-sigma criterion to improve dataset quality and model precision. Subsequently， Bayesian estimation methods were employed to calculate the mean and variance of critical development cost-per-unit parameters （e.g.， takeoff mass）， providing a scientific foundation for subsequent outlier detection. A fractional-order accumulation GM （0， N） model was subsequently constructed， where the unknown parameters were determined through the least squares method， enabling precise development cost estimation for target rocket configurations. Experimental results demonstrate that compared with multivariate linear regression models and conventional GM （0， N） models， the proposed method achieves superior estimation accuracy and enhanced robustness. Specifically， the prediction error of this methodology was reduced to 2.637 5%， whereas the errors of multivariate linear regression and conventional GM （0， N） models reached 20.716 9% and 14.212 8% respectively. Furthermore， this methodology not only applies to liquid rocket development cost estimation but can also be extended to cost prediction problems in other complex engineering systems， providing valuable references for scientific research and technological development in related fields.