Firstly,an App design method is designed,which is based on smart phone sensor to obtain acceleration and other data,sets 15 action and mobile phone position combination categories to be recognized,and collects 750 000 motion data records.Secondly,the sliding window technology is used to segment the time-series data,and the targeted time-domain and frequency-domain characteristic indexes are constructed to form a new sample series with different window sizes and steps.Finally,four kinds of algorithm with excellent performance are selected through the classification results,which demonstrates the good performance of deep learning in feature construction,and discusses the influence of sliding window size and step size on the recognition results.These results have reference value for the construction of human behavior recognition system.