In order to realize the blind separation of FH signals based on the needs of its networkstation sorting, most of existing approaches based on joint diagonalization for FH signals request the orthogonality exactly, whereas this requirement is always limited in many practical applications. In order to soften the terms, a new blind separation algorithm based on nonorthogonal joint diagonalization is proposed. The algorithm effectively obtains the autosource (timefrequency) TF points with the eigenmatrix structure through gradient norm based on noise reduction algorithm in timeslots area, computes a sequence of matrices of timefrequency distributions (TFDs), and then estimates the separate matrix through nonorthogonal joint diagonalization to realize blind source separation of mixed frequencyhopping signals. The results of the simulations illustrate that the proposed algorithm is effective in the blind separation of frequencyhopping signals, and the robustness is better than that of the other blind separation algorithms of frequencyhopping signals.