Based on the analysis of loss distribution approach, loss events can be divided into three types: internal fraud, external fraud and illegal operation. Then, we adopt two-stage distribution to fit the loss intensity distribution of operational risk and Gibbs sampling of Bayesian theory to obtain the parameter estimates, which can reduce error caused by the insufficient low-frequency and high-loss data. In view of the correlation between different types of operational risk loss, the copula function is applied to integrating the total loss distribution. Finally, we calculate VaR and CVaR for different confidence level of the operational risk of commercial banks in China. The empirical research result shows that: Parameter estimation based on Bayesian theory takes into account priori information such as population and sample information which can reduce the estimated error. The introduced copula function and measured value of VaR and CVaR can not only consider the probability of loss events, which can also calculate potential losses of operational risk, but also get a more accurate measurement result of operational risk.