Given that the mixing matrix of underdetermined blind source separation (UBSS) changes with the recording environment, offline UBSS methods encounter difficulty in satisfying the time-varying estimation demand. Therefore, in this work, an online tensor algorithm has been proposed to estimate the time-varying mixing matrix for separating an instantaneous linear underdetermined mixture. First, we construct a canonical polyadic tensor model by assuming individually correlated sources. Second, an online tensor algorithm is applied to decompose the canonical polyadic tensor model to ensure the accuracy of the time-varying mixing matrix. Finally, two types of data, including speech and biomedical signals, have been used to substantiate the effectiveness of the proposed algorithm in estimating the time-varying mixing matrix for UBSS. The results show that the developed online tensor algorithm is significantly superior to the conventional offline UBSS methods in terms of time consumption and accuracy.
展开▼