As is known that nonunitary joint diagonalization (JD) has some advantages over the unitary one in terms of system identification accuracy. However, the existing nonunitary JD algorithms are prone to converge to degenerate (even singular) solutions, which result in deteriorated identification performance. Moreover, the existing algorithms usually seek a square diagonalizing matrix, which greatly limits their application in overdetermined system identification scenario. In order to overcome these drawbacks, we reformulate the nonunitary JD as a multicriteria optimization model. The resulting algorithm can converges to a nonsquare well-conditioned diagonalizing matrix.