Because　feature　extraction　from　electroencephalogram　(EEG)　signals　is　essential　for　cognitive　investigations,　effective　feature　extraction　approaches　are　needed　to　improve　the　practical　recognition　accuracy　of　EEG　signals.　In　this　paper,　a　strategy　is　presented　for　fusing　both　the　linear　and　nonlinear　features　from　EEG　signals　to　improve　the　accuracy　of　motor　imagery　classification.　First,　principal　component　analysis　(PCA)　is　used　to　extract　the　linear　features　from　EEG,　and　linear　discriminant　analysis　(LDA)　is　introduced　to　supplement　the　discriminant　features　by　utilizing　the　label　information　of　the　training　data.　Second,　we　use　parametric　t-distributed　stochastic　neighbor　embedding　(PTSNE)　to　extract　the　nonlinear　features　reflecting　the　original　manifold　structure　of　the　EEG　data.　Third,　these　linear　and　nonlinear　features　are　fused　to　generate　the　final　features　for　classification.　After　feature　extraction,　we　choose　the　hierarchical　extreme　learning　machine　(HELM)　algorithm,　which　has　a　high　classification　accuracy　for　EEG　signal　classification　of　motor　imagery.　To　verify　the　validity　of　the　strategy,　we　compare　the　accuracy　of　the　proposed　method　with　that　of　other　methods　on　the　motor　imagery　dataset.　We　achieve　a　high　accuracy　of　95.89%　and　an　average　accuracy　of　93.45%.　The　performance　shows　that　the　accuracy　of　the　proposed　feature　fusion　strategy　is　effective　for　classification　and　that　the　recognition　accuracy　is　improved　compared　with　other　state-of-the-art　methods.