We consider the dynamics of a harmonic crystal in d dimensions with n components, d, n \ 1. The initial date is a random function with finite mean density of the energy which also satisfies a Rosenblatt-or Ibragimov-Linnik-type mixing condition. The random function is translation-invariant in x 1 ,..., x d − 1 and converges to different translation-invariant processes as x d Q ± ., with the distributions m ± . We study the distribution m t of the solution at time t ¥ R. The main result is the convergence of m t to a Gaussian translation-invariant measure as t Q .. The proof is based on the long time asymptotics of the Green function and on Bernstein's ''room-corridor'' argument. The application to the case of the Gibbs measures m ± =g ± with two different temperatures T ± is given. Limiting mean energy current density is − (0,..., 0, C(T + − T − )) with some positive constant C > 0 what corresponds to Second Law.