Research On DOA Estimate Method Based On PF

In the third generation mobile communication, in order to develop the using offrequency spectrum,enlarge the capability of the system,expend the radiation of thebase station,improve the quality of the communication and reduce theelectromagnetic pollution, an important technique is proposed which is namedsmart antenna based on array signal processing.The DOA(Direction ofArrival)method is key technology of smart antenna.In this paper, we make anintroduction on the appearance background of the Smart Antenna technology andits relative theory and concept. After that, we do some research work on theestimation arithmetic of the signal DOA(Direction Of Arrival) using SmartAntenna Array. The research work is based on the ULA(uniform linear array)antenna.1. Beamforming algorithm(1) delay-and-sum algorithmClassical DOA eatimation technique is base on the delay-and-sumbeamformer.This technique is simple to implement but requires a number ofsensors to achieve a high resolution.(2) Capon beamforming Minimum Variance algorithmCapon Minimum Variance algorithm is also named Minimum VarianceDistortion-less Response (MVDR) algorithm. Capon algorithm can provide betterresolving power than the delay-and-sum algorithm,but it also has manydisadvantage:If there are other signals correlated with the interested signal, Caponalgorithm will not react and the computational complexity is increase obviously.2.Subspace-based techniques(1)MUSIC algorithmMultiple Signal Classification (MUSIC) algorithm was first published bySchmidt in 1979,and was republished in IEEE in1986.It is a principal candidate inthe field of spectral estimation and rather widely referenced in the literature.It isassumes that the incident signals and the noise are uncorrelated.Thus we mayjustifiably consider the subspace spanned by the noise is orthogonal to the subspacespanned by the incident signals.By calculating the eigenvalue of the covariancematrix of the waveforms received at the antenna array elements,we can find theminimum eigenvalues and the eigenvectors that belongs to the noise subspace.Weconsider that the subspace which is orthogonal to the eigenvectors associated withthe minimum eigenvalues is just the subspace apanned by the incident signals.Bycalculating the spatial domain spectrum function,we can estimate the directions ofarrival of multiple plane waves.However,it is not efficient for coherent multipathsignals.In this paper,two modified MUSIC algorithms are presented bydedution,which can estimate the DOA of coherent signals and adjacent signals withsmall SNR.Finally,Some simulations have proved the effectiveness of these newmodified MUSIC algorithms.(2)ESPRIT algorithmCompare with MUSIC method,ESPRIT method needn't search for spectralpeak,so it has advantage of less calculation,higher precision and suitable forrealtime application,thus attracted extensive attention.However,in case of lowSNR,the estimation error can be high,and when the signals are coherert or theDOAs of signals are close,the reliability of DOA estimation will fall.3. Maximum Likelihood(ML)The Maximum Likelihood(ML)direction-of-arrivial(DOA)estimation methodwas one of the first to be investigated.For a long time,the complexity andcomputation load of maximizing the multivariable,highly nonlinear likelihoodfunction prevented it from popular.In theory,the ML method gives a superiorperformance compared to other methods,especially in low signal-to-noise ratioconditions,providing asymptotically unbiased and efficient estimates,especially inthe threshold region.It is a nearly optimal technique.4.PFParticle filters are modern Bayesian methods based on numericallyapproximating the posterior distribution.This method is very useful to estimateparameters when the system model and measure model are not linear and measurenoises are non-Gaussian.Such a filter consisits of essentially two stages:predictionand update.The prediction stage uses the system model to predict the state pdfforward from one measurement time to the next.Since the state is usually subject tounknown disturbances(modeld as random noise),prediction generallytranslates,deforms,and spreads the state pdf.The update operation uses the latestmeasurement to modify the prediction pdf.This is achieved using Bayestheorem,which is the mechanism for updating knowledge about the target state inthe light of extra information from new data.(1) SISThe sequential importance sampling(SIS) algorithm is a MonteCarlo(MC)method that forms the basis for most sequential MC filters developedover the past decades.The key idea is to represent the required posterior densityfunction by a set of random samples with associated weights and to computeestimates based on these samples and weights.As the number of samples becomesvery large,this MC characterization becomes an equivalent representation to theusual functional description of the posterior pdf,and the SIS filter approaches theoptimal Bayesian estimate.But a common problem with SIS method is that after afew iterations, most particles have negligible weight. It means the weight isconcentrated on a few particles only. we call this problem as the degeneracyphenomenon. At this time variance of weight is very large. Amount degeneracy canbe estimated based on variance estimation of weights. The simplest method toreduce degeneracy effect is to use a very large N. But it will increase thecomputational load. It is often impractical.So we usually chooce good importantdensity or resample step.(2) SIRIn SIR algorithm there are two characteristics: importance density is the priordistribution and the resampling step is used every time and don't need to judgedegeneracy parameter. The advantage of SIR is that the importance density can beeasily sampled and the process for updating weight is simple. But we know that itis not a optimal choice. The resampling procedure can solve the degeneracyproblem. But new problem arises. Particles with high weight are selected more andmore often. Others will disappear slowly. It will cause loss of diversity or sampleimpoverishment.5. Other nonlinear fliter algorithm(1) EPFThe Extended Kalman Filter(EKF)has unquestionably been the most widelyused estimafion algorithm for nonlinear systems.However,the EKF is simply basedon linearizes all nonlinear functions to the first-order by using the Taylor seriesexpansions approximations of state transition and observation equations about theestimated state trajectory.The EKF provides an insuficiently accurate representationin many cases,and it is difficult to implement and tune.Many of these dificultiesarise from its use of Taylor linearization.Based on PF method,process update useEKF at every sampling particle can achieve the EPF algorithm.(2) UPFUnscented Kalman Filtering(UKF) is a typical nonlinear filter method,thesubstance is Uncented Transform(UT).Unlike the EKF method,UKF is notapproximate the nonlinear model,but to approximate the state probability densityfunction(PDF).Differ with PF method,it is not the random sample,the UKF is thesampling with sigma point.The basic of UPF algorithm is use UKF based on PFmethod to achieve better important probability density function.Everytimesampling partical is update by UKF method,the mean and variance is used for nexttime sampling.