The Adaptive Learning Rate Blind Source Separation Method

The purpose of blind source separation(BSS)is to recover the unknown source signals from their linear mixtures without the knowledge of mixing coefficients.It has broad application prospect in various fields such as speech recognition,image processing and various biomedical signal processing.In real-time BSS,the step-size has great influence on the separation performance.The choice of step-size for real-time BSS problem was researched in this thesis.The main work can be expressed as follows:(1)For real-time linear mixtures BSS problem,an adaptive step-size algorithm based on improved particle swarm optimization(PSO)and grading learning was proposed.The whole signal separation process is divided into two stages: the rapid stage and the precise stage.In the rapid stage,the improved PSO is used to choose the step-size.It ensures that the outputs get an acceptable degree of separation with less data sets or iterations.In the precise stage,a smooth function is used to determine the step-size.It ensures that the outputs have high stable performance.The proposed learning rate determined algorithm takes full advantage of the rapid convergence of PSO in the early stage of separation and overcomes the instability of PSO in the precise stage.Simulation experiments demonstrate that significant improvements of the convergence speed and accuracy are achieved by the proposed algorithm compared with fixed learning rate,learning rate determined by PSO,grading learning and adaptive combination algorithm.(2)For real-time convolutive mixtures BSS problem,an adaptive step-size algorithm based on fuzzy system was proposed.First,by executing short-time Fourier transformation,the time domain signal is converted to a group of frequency signals.Owing to the different separating state of frequency bins,the required learning rates for different frequency bins are different.So,a vector learning rate has been used to replace a scalar one for independent vector analysis(IVA).Second,the correlation indicators of current frequency bin and the whole output signals are introduced to measure the separating state of the output signals.Finally,we develop a fuzzy inference system to determine the learning rate of real-time IVA.The fuzzy inference system consists of two inputs(correlation indicators of output signals in current frequency bin and the whole output signals)and one output(the learning rate of current frequency bin).Through applying the fuzzy inference system in every frequency bin,a proper learning rate can be obtained.Simulation experiments demonstrate that significant improvements of the convergence speed and accuracy are achieved by the proposed algorithm in speech blind source separation problem.