As known, it is essential to carefully tune the parameters of non-local means (NLM) in order to take it into full play. The adaptively determination of NLM’s parameters for a signal of interest has not been reported so far, which will significantly weaken the NLM in bearing fault diagnosis. Aiming at such a dilemma, a novel fault diagnosis method for rolling element bearings is proposed based on non-local means with particle swarm optimization (PSO). PSO algorithm is used to obtain optimal values of parameter λ, M and P with a superior performance with respect to global optimization and convergence speed. Then an optimalfilter is acquired with the resultant optimal parameters, which can suppress noises and enhance cyclic impact feature hidden in vibrations of faulty bearings after filtering. Finally, fault diagnosis can be achieved by means of the envelop spectrum of the filtered signal. The viability of the proposed method is demonstrated through a series of simulation data and experimental data.