The　result　of　the　emotional　state　induced　by　music　may　provide　theoretical　support　and　help　for　assisted　music　therapy.　The　key　to　assessing　the　state　of　emotion　is　feature　extraction　of　the　emotional　electroencephalogram　(EEG).　In　this　paper,　we　study　the　performance　optimization　of　the　feature　extraction　algorithm.　A　public　multimodal　database　for　emotion　analysis　using　physiological　signals　(DEAP)　proposed　by　Koelstra　et　al.　was　applied.　Eight　kinds　of　positive　and　negative　emotions　were　extracted　from　the　dataset,　representing　the　data　of　fourteen　channels　from　the　different　regions　of　brain.　Based　on　wavelet　transform,　δ,　θ,　α　and　β　rhythms　were　extracted.　This　paper　analyzed　and　compared　the　performances　of　three　kinds　of　EEG　features　for　emotion　classification,　namely　wavelet　features　(wavelet　coefficients　energy　and　wavelet　entropy),　approximate　entropy　and　Hurst　exponent.　On　this　basis,　an　EEG　feature　fusion　algorithm　based　on　principal　component　analysis　(PCA)　was　proposed.　The　principal　component　with　a　cumulative　contribution　rate　more　than　85%　was　retained,　and　the　parameters　which　greatly　varied　in　characteristic　root　were　selected.　The　support　vector　machine　was　used　to　assess　the　state　of　emotion.　The　results　showed　that　the　average　accuracy　rates　of　emotional　classification　with　wavelet　features,　approximate　entropy　and　Hurst　exponent　were　respectively　73.15%,　50.00%　and　45.54%.　By　combining　these　three　methods,　the　features　fused　with　PCA　possessed　an　accuracy　of　about　85%.　The　obtained　classification　accuracy　by　using　the　proposed　fusion　algorithm　based　on　PCA　was　improved　at　least　12%　than　that　by　using　single　feature,　providing　assistance　for　emotional　EEG　feature　extraction　and　music　therapy.　Copyright　©　2017　by　Editorial　Office　of　Journal　of　Biomedical　Engineering.