This paper presents a neural network-based method to detect the impact location in composite plates. Piezoceramic sensors were mounted in the corner of a quasi-isotropic graphite/epoxy composite plate. A square plate with fixed boundary condition was impacted by impact hammer. To detect impact locations, two methods were adopted. First, the arrival- time differences of acoustic wave generated by impact were used as inputs of neural networks for detecting impact location. This neural network was trained by a set of 36 simulated impact locations, and tested on the 10 random impacts. Average RMS error of detected impact locations was 1.52 radial centimeters. Second, amplitudes of natural frequencies obtained from FFT of piezosensor data were added to inputs of the first neural networks. This added information can enhance precision of detection networks and average RMS error was reduced to 1.08 radial centimeters.