Optimal wavelet filter is a commonly used and effective tool for bearing fault diagnosis. To locate the informative frequency band for extracting fault-induced repetitive transients, the wavelet parameters are traditionally optimized with a single criterion such as kurtosis, smoothness index, etc. calculated from the narrow-band filtered signal or its envelope. However, in some cases it is difficult for them to fully depict the fault characters and to be robust to different background noises. In this work, a general multi-objective optimized wavelet filter is proposed to adaptively extract the bearing fault features. To take impulsiveness and cyclostationarity into consideration simultaneously, a general rule as maximum sparsity of the squared envelope and squared envelope spectrum is given to design the multi-objective fitness functions. The Pareto solutions which score better under all objectives in the sense of non-domination are utilized to estimate the informative frequency band with the help of differential evolution and a robust knee point selection strategy using kernel density estimation. A simulated and two cases of real wheelset bearing signals are applied to evaluate its performance, some comparisons with peer single-objective and multi-objective methods are also conducted to illustrate its consistency and robustness in extracting the fault-induced repetitive transients under complex interferences.