CSP共空间模式的介绍

CommonSpatialPattern(s)algorithm算法.TheCSPparadigmisbasedonthedesignoftheBerlinBrain-ComputerInterface(BBCI)[1],morecomprehensivelydescribedin[2],whichismainlycontrolledby(sensori-)motorimagery.ThefeaturesexploitedbythisparadigminitsoriginalformareEvent-RelatedSynchronizationandDesynchronization[3]localizedinthe(sensori-)motorcortex,buttheparadigmisnotrestrictedtotheseapplications.CSPwasoriginallyintroducedin[5]andfirstappliedtoEEGin[6].Duetoitssimplicity,speedandrelativerobustness,CSPisthebread-and-butter实用的paradigmforoscillatory振荡processes,andifnothingelse,canbeusedtogetaquickestimateofwhetherthedatacontainsinformationofinterestornot.Likepara_bandpower,CSPuseslog-variancefeaturesoverasinglenon-adaptedfrequencyrange(whichmayhavemultiplepeaks),andneithertemporalstructure时间结构(variations)inthesignaliscaptured捕捉,norareinteractions相互作用betweenfrequencybands.Themajorstrengthoftheparadigm范式isitsadaptivespatialfilter自适应空间滤波器,whichiscomputed计算usingtheCSPalgorithm.Theparadigmisimplemented实施asastandardsequenceofsignal(pre-)processing(spatial/spectral光谱filtering),featureextraction,andmachinelearning.Thefirstpreprocessing预处理stepisfrequencyfiltering,followedbyanadaptivelylearnedspatialfilter(whichisthedefiningpropery定义的性能oftheparadigm),followedbylog-variancefeatureextractionandfinallya(usuallysimple)machinelearningstepappliedtothelog-variancefeatures.Thespatialfilteringprojectsthechannelsoftheoriginalsignaldowntoasmallsetof(usually4-6)surrogate代理channels,wherethe(linear)mappingisoptimized线性映射被优化suchthatthevarianceinthesechannelsismaximallyinformativew.r.t.totheprediction预测task.TheCSPfilterscanbeobtainedfromtheper-classsignalcovariancematrices协方差矩阵bysolvingageneralizedeigenvalueproblem广义特征值问题(oftheform[V,D]=eig(Cov1,Cov1+Cov2)).CSPcanalsobeappliedtoindependentcomponentstorate评价theirimportanceorforbetterartifact工件robustness鲁棒性.Awiderangeofclassifiers分类canbeusedwithCSPfeatures,themostcommonlyusedonebeingLDA狄利克雷/一个集合概率模型.Thereexistsalargecorpus语料库ofCSPvariantsandextensions变换与拓展,mostlytogivebettercontroloverspectralfiltering,includingmultiband多波段的CSP(para_multiband_csp),SpectrallyWeightedCSP(para_speccsp)光谱加权CSP,InvariantCSP,CommonSpatio-SpectralPatterns(CSSP),CommonSparseSpectralSpatialPattern(CSSSP),RegularizedCSP,【不变的CSP，普通的时空光谱模式（CSSP），普通的稀疏频谱空间格局（CSSP），正则CSP】andseveralothers.Amoreadvanced(butalsocomputationally计算morecostly)paradigm范式thanCSPistheDual-AugmentedLagrangeParadigm双增强拉格朗日范式(para_dal/para_dal_hf).Thelengthofthedataepoch数据纪元andthechoiceofafrequencyband(defaulting默认tomotorimagerytimescales时间尺度andfrequencyranges)aretheparameters参数thataremostcommonlytunedto调谐thetask,bothofwhichcanalsobefoundviaasmallparameter参数search.Someapplicationareasincludedetectionofmajorbrainrhythmmodulations主要的大脑节奏调制(e.g.alpha,beta),forexamplerelatedtorelaxation/stress,aspectsofworkload,sensori-motorimagery,visualprocessingvs.idlingandotheridle-rhythm-relatedquestions,oremotionrecognition视觉处理与空转和其他空闲的节奏相关的问题，或情感识别。.Seealso[4].Examples:Afteranapproachhasbeendefinedasinoneofthefollowingexamples,apredictive预测的modelcanbeobtained(givenacalibration校准datasetandaspecification规格oftargetmarkers目标标记)usingbci_train:[loss,model,stats]=bci_train('Data',io_loadset('calibration_rec.eeg'),'Approach',myapproach','TargetMarkers',{'mymarker1','mymarker2'});%defineabasicCSPapproach,usingthedefaults(7-30Hzbandpass,shrinkage收缩LDAclassifier,3patternpairs)myapproach='CSP';%useanFIRfilter有限脉冲响应滤波器restrictedtothealphaband使用FIR滤波器限制在α频myapproach={'CSP''SignalProcessing',{'FIRFilter',[681415]}};%useanIIRfilterinsteadofthedefaultFIRmyapproach={'CSP''SignalProcessing',{'FIRFilter','off','IIRFilter',[571417]}};%alsorestrictthemodeltoastationarysubspace一个固定的子空间myapproach={'CSP''SignalProcessing',{'FIRFilter','off','IIRFilter',[571417],'StationarySubspace',{'StationaryDim',-0.1,'Operation','keep_stationary'}}};%useasharpFFT快速傅氏变换band-passfilter尖锐的FFT的带通滤波器myapproach={'CSP''SignalProcessing',{'FIRFilter','off','SpectralSelection',[715]}};%useasimplelogisticregressionclassifier一个简单的逻辑回归分类(variationalBayes贝叶斯变换)insteadoftheLDA狄利克雷一个集合概率模型myapproach={'CSP''Prediction',{'MachineLearning',{'Learner','logreg'}}};%useasimplelogisticregressionclassifier(sparse稀疏的variationalBayes)myapproach={'CSP''Prediction',{'MachineLearning',{'Learner',{'logreg','variant','vb-ard'}}}};%usethesparselogisticregressionclassifierbutappliedtoalargersetofpatterns范式myapproach={'CSP''Prediction',{'MachineLearning',{'Learner',{'logreg','variant','vb-ard'}},'FeatureExtraction',{'PatternPairs',6}}};%usingquadraticdiscriminantanalysis二次判别分析myapproach={'CSP''Prediction',{'MachineLearning',{'Learner','qda'}}};%usingGaussianmixturemodels高斯混合模型(variational变化的BayesianDirichletprocess贝叶斯狄利克雷过程prior之前)myapproach={'CSP''Prediction',{'MachineLearning',{'Learner','gmm'}}};%usingrelevancevectormachines相关向量机(herewithafixedkernelscale高速内核混合模式forspeed)myapproach={'CSP''Prediction',{'MachineLearning',{'Learner',{'rvm','kernel','rbf','gamma',1}}}};%usel1-regularizedlogreg(whichinvolvesaparametersearchovertheregularizationparame