Disclosed is a method for behavior pattern analysis, in which the behavior signals of a subject are continuously detected using tri-axis accelerator, and the behavior signals are divided as a plurality of time windows based on a predetermined time length. The plurality of time windows are configured as an X-Y matrix, and cross analysis is performed on the time windows as X-axis and Y-axis to obtain a plurality of correlation coefficients as being the preferred correlation values. A pattern with bright spots and dark spots is plotted according to the preferred correlation values, and behavior pattern is evaluated from the number and distribution of bright spots and the pattern. The behavior consistency of a subject can be efficiently evaluated, and his/her behavior abnormality, disease development or other situations can be determined accordingly.