The measured rolling bearing vibration signals were usually interfered by noises and had nonlinear and non-stationary characteristics, while multi-linear principle component analysis network（MPCAnet）had poor nonlinear fitting ability and poor feature clustering ability when dealing with complex non-stationary data. An improved multi-linear principal component analysis network was proposed by introducing kernel transformation, which increased the degree of difference among the training samples, further enhanced the generalization ability and classification accuracy when dealing with non-linear data. It is proved that this method has high robustness in different fault diagnosis data sets of rolling bearings and may accurately identify various faults of rolling bearings.