As one of the popular adaptive algorithms in active vibration control (AVC) application, the step size of the Filtered-x least mean square (Fx-LMS) algorithm is significant to the performance. It cannot obtain the best performance both in convergence speed and steady-state error with a same step size. To improve the performance, an improved filtered x variable step size least mean square (Fx-VSSLMS) algorithm based arccot function is then proposed. In this paper, seven typical VSSLMS algorithms are induced and classified. Simulations based Adams and Simulink for the active vibration of the flexible piezoelectric cantilever beam validate the effectiveness of the proposed algorithm. Also, experiments with different noise levels are accomplished. The results show that the proposed VSSLMS algorithm has better robust performance than others.