Search

In order to solve the permutation problem of frequency-domain blind source separation, this paper studied the influence of discontinuity of demixing matrix and distance between frequency bins on permutation problem, and proposed an improved method. It used weight coefficients to represent the influence of each frequency bin to the undetermined frequency bin, and used the demixing matrix of the previous frequency bin to initial the demixing matrix of the latter one. This paper presented the processes of the improved method. The improved method was applied in occasions of instantaneous mixtures, convolutive mixtures and real-room conditions. The results show the proposed method has better signal to noise ratio, speed and stronger robustness, comparing with Murata method. [ABSTRACT FROM AUTHOR]

Two-step method is a commonly used solution to the underdetermined blind source separation of sparse source signals, K-means clustering algorithm is usually utilized to estimate mixing matrix firstly, and then, the shortest path algorithm is used to recover source signals. The source signal's number needed in the K-means clustering algorithm is usually unknown in fact, so have to estimate it. This paper studied clustering validity index :BWP index, combined with particle swarm optimization, proposed a new algorithm to determine the source signal's number. Then it introduced the proposed algorithm into the underdetermined blind source separation. The results of experiment show that the proposed algorithm has good separation precision meanwhile it can shorten the separation time and save some memory this advantage is more obvious when the observation data is huge. [ABSTRACT FROM AUTHOR]

To solve the problems of complex fault location process and heavy workload before test using traditional large-scale circuit fault diagnosis methods in the multi-fault conditions, this paper presented a new approach for fault diagnosis using group decomposition. Low dimensional fault eigenvectors could be acquired by decomposing the large-scale analogue circuit in accordance with its topological property and group decomposition criteria. Based on ideas of pattern recognition, neural network which had highly parallel classify capability was selected as classifier and wavelet function which had fast convergence property was selected as hidden layer's transfer excitation function. The simulation results show that this method can achieve fast fault feature vector classification and fault diagnosis results. Comparing with the existing multiple-test-condition (MTC) method and intersection tearing method, this approach has smaller amount of work before test, fewer diagnostic times, less computation, better capability of diagnosing multiple faults and stronger engineering practicality. [ABSTRACT FROM AUTHOR]