Catálogo de publicaciones - libros

A neural-based method for source separation in nonlinear mixture is proposed in this paper. A cost function, which consists of the mutual information and partial moments of the outputs of the separation system, is defined to extract the independent signals from their nonlinear mixtures. A learning algorithm for the parametric RBF network is established by using the stochastic gradient descent method. This approach is characterized by high learning convergence rate of weights, modular structure, as well as feasible hardware implementation. Successful experimental results are given at the end of this paper.

An adaptive natural gradient algorithm for blind source separation based on convolutional mixture model is proposed. The proposed method makes use of cost function as optimum criterion in separation process. The update formula of separation matrix is deduced. The learning steps for blind source separation algorithm are given, and high capability of the proposed algorithm has been demonstrated. The simulations results have shown the validity, practicability and the better performance of the proposed method. This technique is suitable for many applications in real life systems.

Extracting independent source signals from their nonlinear mixtures is a very important issue in many realistic models. This paper proposes a new method for solving nonlinear blind source separation (NBSS) problems by exploiting particle swarm optimization (PSO) algorithm and natural gradient descent. First, we address the problem of separation of mutually independent sources in post-nonlinear mixtures. The natural gradient descent is used to estimate the separation matrix. Then we define the mutual information between output signals as the fitness function of PSO. The mutual information is used to measure the statistical dependence of the outputs of the demixing system. PSO can rapidly obtain the globally optimal coefficients of the higher order polynomial functions. Compared to conventional NBSS approaches, the main characteristics of this method are its simplicity, the rapid convergence and high accuracy. In particular, it is robust against local minima in search for inverse functions. Experiments are discussed to demonstrate these results.

This paper deals with the employment of Echo State Networks for identification of nonlinear dynamical systems in the digital audio field. The real contribution of the work is that such networks have been implemented and run in real-time on a specific PC based software platform for the first time, up to the authors’ knowledge. The nonlinear dynamical systems to be identified in the audio applications here addressed are the mathematical model of a commercial Valve Amplifier and the low-frequency response of a loud-speaker. Experimental results have shown that, at a certain frequency sampling rate, the ESNs considered (after the training procedure performed off-line) are able to tackle the real-time tasks successfully.

When the source signals are known to be independent, positive and well-grounded which means that they have a non-zero pdf in the region of zero, a few algorithms have been proposed to separate these positive sources. However, in many practical cases, the independent assumption is not always satisfied. In this paper, a new approach is proposed to separate a class of positive sources which are not required to be independent. These source signals can be separated very quickly by using genetic algorithm. The objective function of genetic algorithm is derived from uncorrelated and some special assumptions on such positive source signals. Simulations are employed to illustrate the good performance of our algorithm.

The speech enhancement method based on blind source separation and post-processing in subband [1] is an effective method in noise and reverberation environments. Performance analysis and computer simulations indicate that its performance is degraded under uncorrelated or mild correlated noise cases, and sometimes it might cause distortion of the enhanced signals. To apply the method in real environment, some improvements have been made on it. These are that adaptive noise cancellers are only used in the subbands with poor separation results and the independent component analysis (ICA) operations in low frequency bands are replaced by the efficient time-frequency masking method. Experimental results show the effectiveness of the proposed method.

The sequential blind signal extraction method with a linear predictor is proposed. The Kalman filter is introduced to overcome the problem of the choice of the linear predictor coefficients. Using a deflation technique, the proposed algorithm is able to sequentially recover the source signals one-by-one. Simulation results verify the validity and performance of the proposed algorithm.

This paper presents a new approach to speed up the operation of time delay neural networks for fast code detection. The entire data are collected together in a long vector and then tested as a one input pattern. The proposed fast time delay neural networks (FTDNNs) use cross correlation in the frequency domain between the tested data and the input weights of neural networks. It is proved mathematically and practically that the number of computation steps required for the presented time delay neural networks is less than that needed by conventional time delay neural networks (CTDNNs). Simulation results using MATLAB confirm the theoretical computations.

Nonnegative Matrix and Tensor Factorization (NMF/NTF) and Sparse Component Analysis (SCA) have already found many potential applications, especially in multi-way Blind Source Separation (BSS), multi-dimensional data analysis, model reduction and sparse signal/image representations. In this paper we propose a family of the modified Regularized Alternating Least Squares (RALS) algorithms for NMF/NTF. By incorporating regularization and penalty terms into the weighted Frobenius norm we are able to achieve sparse and/or smooth representations of the desired solution, and to alleviate the problem of getting stuck in local minima. We implemented the RALS algorithms in our NMFLAB/NTFLAB Matlab Toolboxes, and compared them with standard NMF algorithms. The proposed algorithms are characterized by improved efficiency and convergence properties, especially for large-scale problems.

A novel sparse measure of signal is proposed and the efficient number of sources is estimated by the best confidence limit in this work. The observations are classified by SVM trained through samples which are constructed by direction angle of sources. And columns of the mixing matrix corresponding to clustering centers of each class are obtained based on the sum of samples belong to the same class with different weights which are adjusted adaptively. It gets out of the trap of the initial values which interfere k-mean clustering quite a lot. Furthermore, the online algorithm for estimating basis matrix is proposed for large scale samples. The shortest path method is used to recover the source signals after estimating the mixing matrix. The favorable simulations show the stability and robustness of the algorithms.