Based on the machine learning fault diagnosis method, a random forest fault diagnosis method characterized by the time domain index of vibration signal was proposed in view of the diversified characteristics of the composite fault state feature of marine rolling bearings. The vibration time domain signals were cleaned and converted to construct the derived features of five dimensionless indicators, and the random forest algorithm with the decision tree as the basic classifier was selected to establish a training model; the ideal fault diagnosis classification model was obtained through the feature selection, evaluation testing and model optimization; the model simulation was carried out by using the rolling bearing competition data set, and 8 kinds of marine rolling bearing fault conditions in actual cases were simulated. The scientificity of the feature extraction and the validity of the fault diagnosis model were proved through the threeway vibration experiment and algorithm modeling. The results show that the data simulation diagnosis accuracy and experimental diagnosis accuracy are 98.61% and 98.85% respectively, and the fault diagnosis effect of this method is the best when the vibration acquisition is in the axial direction.