The present disclosure relates to estimation methods of actuator faults based on Bayesian learning. An actuator fault is modeled based on a random walking model, and a joint posterior probability distribution of a system state variable and the actuator fault is represented using two mutually independent hypothesis distributions based on a variational Bayesian theory; a system state variable and an actuator fault of a system at a moment are predicted at a moment; and a predicted system state variable and a predicted actuator fault are iteratively updated at the moment according to the Bayesian theory to output an estimated value of the system state variable at the moment, a variance of the estimated value and an estimated value of the actuator fault at the moment and a variance of the estimated value.