Methodologies And Applications Of Weak Signal Detection Based On Blind Source Separation

The weak information capture and measurement in the strong noise backgroundhas got the extensive need in the disciplines of science and engineering applications.For example, the inter-satellite ranging and communication signal detection in thedeep-space exploration, weak characteristic signal extraction in fault diagnosis, smalltarget detection in emergency rescue, and Fetal Electrocardiograph (FECG) andelectroencephalogram (EEG) signal detection in the biomedical etc. However, theweaker the information we need to measure is, the easier it is corrupted by other signalsand complex background noise. Due to the complexity and randomness of noise, thetraditional method which based on time domain waveform and spectrum analysis, it isdifficult to achieve the extraction of weak signal. Therefore, we need to in-depthanalysis the characteristics and mixing mechanism of weak signals and noise, andresearch new detection theory and method for weak signal extraction in the complexnoise background.Varity of weak signal detection methods and applications are analyzed in this thesis,and the deficiencies and limitations of existing methods are pointed out. Aiming at theseparation of weak signal and noise, the characteristics and mixing mechanism of themare discussed. For the four kinds of typical mixing classes: over-determined linear,over-determined nonlinear, single channel linear, and under-determined linear, the blindsource separation based detecting models and algorithms are proposed, respectively.The main research results are:For the detection of weak signal from the over-determined linear mixtures, noisesources are divided into two classes according to their types and characters. The maindifference of the two classes is that whether their frequencies are within the usefulsignal band. In order to remove the noise outside the signal band, the VariableParameters Amplification and Filtering Circuit (VPAFC) is employed. The circuit canbe adjusted automatically according to the spectrum of the signal, thus the noise can bemaximally reduced. While, for the noise inside the signal band, the signal extractionmethod based on Denoising Source Separation (DSS) is proposed. It separates the noisesources sequentially from observed signals by using the statistical independencebetween the weak signal and noise. Numerical simulation and FECG signal extractionexperiments illustrate that the proposed approach can effectively extract the weak signal components from the over-determined linear mixing of signal and noise sources.In order to extract the weak signal from the over-determined nonlinear mixtures,the source separation method based on interval optimization is proposed. Throughanalyzing the mixing property of weak signal and noise sources, the nonlinear inversemapping from observations to estimated source signals is modeled by a multilayerperceptron neural network, while their weights are replaced by interval numbers. Then aBranch-and-Bound global optimization algorithm is constructed by using intervalanalysis. It can overcome the problems of easily get trapped in the local minimum andunsatisfactory convergence speed, which would otherwise be severed in unsupervisedlearning with nonlinear models. The experimental results show that the proposedalgorithm can efficiently separate the weak signal and noise components.An Iterative Reweighted Continuous Basis Pursuit (IRCBP) based signal and noiseseparation approach is proposed. It can be used to extract the weak signal from thesingle-channel measurements which corrupted by linear interferences. Under theassumption of signal transformation invariance, the mechanism of single-channel linearmixture is systematically analyzed, and the continuous basis pursuit model for signalrecovery is obtained. Thus the single-channel blind source separation problem istransformed into a sparse optimization problem. The weak signal component andstructure noise components are extracted from the observations by make use ofalternating iterative optimization. The simulation of weak aperiodic pulse signal andextraction of weak spikes in EEG show that the proposed approach can successfullyrecover the underlying useful weak signal from the heavily corrupted observations.A signal and noise separation approach based on parallel tensor decomposition isproposed to solve the weak signal detection problem of general underdetermined linearinterference. The method analyses the character of time-delay correlation matrices ofobservations. The principle of tensor decomposition and nonlinear programming ofsource separation is analyzed systematically. Meanwhile, the traditionalunderdetermined blind sources separation methods based on tensor decompositionmeets the problem of computational inefficiency when dealing with high dimensionalsignal. In order to overcome this problem, the proposed method introduces OpenCLbased heterogeneous parallel computing. The simulation of gearbox weak fault signaldetection and weak audio signal extraction show that the proposed approach can fastseparate weak signal component from noise environment.For the experiments of backscattering and target space stimulated scattering light energy detection, we design and construct a weak light energy detection system appliedfor Inertial Confinement Fusion (ICF) target physical experiment. The experimentalresults show that the designed equipment and proposed method can extract weak energysignal from complex environmental noise even in small energy experiment, and canprovide accurate data source for accuracy measurement of laser coupling efficiency.