Abstract: In the field of the rolling bearing fault diagnosis,the remaining useful life prediction is very important.This paper proposes an approach based on genetic programming for features extraction,and multiple features are combined into a feature tree,somulti dimensional input transfers to single-dimensional input.Furthermore,using the improved fitness to estimate the quality of thefeature tree.After repeated iterations,the final output is the feature tree,whose fitness is the maximal.Besides,the curve of thisfeature tree is the closest to the linear trend in the time domain,hence,it is regarded as an independent feature named the optimization feature.This paper uses the vibration signal of the entire bearing ife to predict the remaining useful life of the bearing with theoptimization feature as the prediction model,and verifies the accuracy of prediction.