In the system of training and operation in virtual reality

the standard and normative evaluation of action has become an increasingly prominent problem. It is very important and critical to segment the standard actions of the trainer in the training process. The accuracy of segmentation directly affects the subsequent training evaluation.In this paper

a kind of method based on multi-dimensional state cloud model prediction is proposed. Firstly

convolutional neural network is used to model the multi-sensor data set of operation actions

and the multi-result data set of the next moment is formed by training and predicting the multi-sensor data set of operation actions; Then

a multi-dimensional cloud model for the predicted multi-result data set is designed; Finally

the actual actions are used in the cloud model. The corresponding probability of the distribution in the pattern is used to determine the action segmentation point. The results show that the adaptive multi-dimensional cloud model based on convolutional neural network（CNN） can select the segmentation points better

improve the accuracy of action segmentation

and can be used for continuous action segmentation in virtual reality environment.