Purpose: With the development of artificial intelligence technology, the intelligent interaction of human health monitoring is gradually being realized. Smartphones have become widely available and can be used to collect information, but extracting valid information components from large amounts of information requires an effective and versatile analysis method. Methods: The common motion signals of the human body were collected by the built-in sensors in the mobile phone, and the feature extraction method combined with the convolutional neural network and the autoregressive model was used to identify the human motion. Results: Using this method, the experimental data and two types of public data sets UCI HAR and WISDM were tested respectively, and the recognition accuracy rate was more than 90%. Conclusion: The proposed model eliminates the defect that the commonly used pattern recognition requires manual design of eigenvalues. The introduction of autoregressive models effectively reduces the computational complexity of large-scale stacked convolutional layers. In addition, the efficiency and feasibility of the method are also demonstrated.