It is necessary to improve the visual robust fault identification ability of roller bearing in non-stationary operating condition. To achieve it, LMD-CM-PCA approach was proposed. First, based on roller bearing vibration acceleration signals, local mean decomposition(LMD) was applied to extract product function (PF) sample matrix. Second, the discrete correntropy and Pearson product moment correlation coefficient (PPCC) of PF and primary signal were calculated. Correntropy was modified by PPCC as the amplitude modulation (AM) of correntropy.Then, the correntropy matrix (CM) of the samples was constructed with AM-correntropy being itselements. Finally, principal component analysis (PCA) was employed to implement the integration of CM with the largest variance accumulated contribution rate as the evaluation index. Integrated CM (ICM) of vibration datum under mixed operating conditions was calculated in slight fault and serious fault situations both. The visual results indicated that ICM could isolate operating condition better and separate faults under different fault severity levels more robustly than traditional characteristics, such as energy moment and spectral kurtosis, do. Above all, application of ICM ,like roller bearing fault features provides more effective technical support for roller bearing fault intuitively diagnosis so that it can support applications in fault diagnosis and safety early warning fields.