The Application And Research On Blind Source Separation Technology Based On Independent Component Analysis

Blind source separation refers to a process that the various components of the source signals is estimated only by observed signals and source signals'some prior knowledge (such as probability density function) under the condition that unknown source signal and unknown transmission channel characteristics. Blind source separation,as an efficient signal processing method rises in recent years,has a broad prospects for development in the antenna array signal processing,the medical signal processing, image processing and ship vibration test and so on, in addition, it has gradually become a hot spot in the world signal processing and neural network community. Based on the instantaneous linear mixed model of Blind signal processing the major work done by the writer is summarized as follows:1. In the beginning, the writer expounded the major solutions to the problems occurred in the blind source separation-Independent Component Analysis (ICA). Based on the mathematical model of a signal the writer introduce the details of the basic ICA assumptions and uncertainties, speaking briefly the knowledge of ICA preparation information theory and the measurement criterion of independent degree, summarizing the commonly used algorithm and performance index.2. Secondly, the writer discusses the offline batch processing and online self adaptation processing algorithm of the blind source separation algorithm. Based on negentropy FastICA algorithm has dealt with a batch of data samples acquired with fast convergence, but the tracking performance is poor; the online self adaptation processing algorithm is on real-time when dealing with a single sampled data. In a stable channel, the separation performance of the former was obviously stronger than the latter. in an unstable channel, only the online self adaptation processing algorithm can be adopted, while the online learning algorithm uses a fixed learning step, so attention could not be given to both the convergence speed and precision in steady state at the same time. In order to achieve best convergence and stability, the optimization of the learning step is the core of this paper.3. Directed at the main factors affect the online algorithms-----activation function and learning steps. In the beginning, online of kurtosis is introduced in order to make activation function with parameters meet the combination of different types of signal source,and the paper analyzes the characteristics of the gradient variable step of the natural gradient algorithm-- a better real-time tracking performance, however, the choice of learning step is supported by the auxiliary variables, and has nothing to do with the degree of separation between the output signals, for this reason, the separation accuracy of this algorithm is unsatisfactory because of channel instability and larger initial step length. An improved variable step size self adaptation algorithm is proposed based on the Gradient variable step. Defined a similarity measure which expresses the separation degree between the neural network output. To make self adaptation step adjustment according to the signal separation state reflected by the degree of similarity, and establish a nonlinear relationship between the learning step and the similarity measure variation. Compared with the previous algorithms, the algorithm has not only fast convergence but small steady-state error characteristics and suitable to the time-varying environment in which the channel is not constant.4. Preliminary discussion has been made on the application of blind source separation algorithm (based on Independent component analysis) in vibration signals of ships. Deep analysis has been made in the line spectrum modeling of the ship noise, continuous spectrum modeling, and modulation envelope modeling. Finally, the paper summarizes the mathematical model of ship radiated noise and the feasibility of the simulation algorithm separation in the source of ship noise.