As part of the RFID flip chip package development, the high speed manipulator has obvious nonlinear and time-variable characters， so a nonlinear adaptive inverse control is needed. The key to this method is to identify the high speed manipulator by using a third-order Volterra nonlinear model in limited time and with sufficient accuracy. However,it is hard to satisfy real-time requirement with a conventional method. This paper proposes a fast identification algorithm to resolve the problem. Firstly, a high-order input vector is constructed from a low-order input vector according to the structural character. Next, it speeds up the estimates of high-order kernels based on low-order kernels according to their correlation. Finally, it uses a linear variable step-size LMS strategy in a nonlinear algorithm and proves convergence with the Lyapunov global stability theorem.In experiments with a manipulator based on conventional and proposed methods, respectively, the results show that this algorithm reduces the identification time from 100 ms to 30 ms, improves convergent speed 3.3 times and reduces misadjustment by 93.3%, as well as having great precision. It can satisfy both requirements of real-time and identification precision .