In this paper, an inverse heat conduction problem with two unknown heat sources in a square region is investigated based on the reduced order model. At first, two sets of pseudo binary random signals are used as the input heat source functions to generate snapshots. The reduced order model is then obtained by projecting the CFD model on the eigenfunctions obtained from the proper orthogonal decomposition of these snapshots. The sensitivity problem and adjoint problem needed for the minimization of the object function are also developed based on the reduced order model. The performance of the algorithm is examined by several numerical experiments, and is found to be very accurate as well as efficient.