It is one of the important methods of pat tern recognit ion to apply neural netw orks to target classificat ion. Forw ard propagat ion mult-i layers neural networks and it s BP a-l gorithm are used w idely . In this art icle, some measures are taken to improve BP algorithm, and to make its performance bet ter and it s convergence speed quicker. The seismic sensor is an essential sensor in bat tlefield w atching system. By test ing, a g reat number of seismic sig nals are obtained on footsteps, w heeled-vehicle and tank. These signals are processed using w avelets transform and w avelets package. The energy spectrum features of these signals are ex tracted, and tw o series connected BP neural networks ident ify them. The results of 94. 5% proper ident ification are at tained.