Aiming at the problem of slow convergence and control failure due to complex frequency components of input signals in the an active vibration isolation system, a variable-step wavelet packet least mean square algorithm (WPx-LMS) is proposed. At first, the input signal is decomposed into a series of non-overlapping band frequency through wavelet packet transform. And then, an independent control filter and iterative variable-step are designed for each of the decomposition signals. The decomposition results of different wavelet basis are compared to determine the optimal wavelet basis in control system. By using error signals and decomposition signals to design the variable step strategy, the sensors and the controller are used to build the two-layer vibration isolation control test bench. The simulations experiments are carried out to verifily that the WPx-LMS algorithm, for the two-layer vibration isolation platform with complex input signals, can achieve better robustness and control effect than the conventional LMS algorithm.