Abstract:The efficient and accurate motion pattern recognition is the basis and prerequisite of flexible prosthesis control. The key is to recognize locomotion-mode (level-ground walking stairs ascent. stairs descent. upslope. downgrade). In order to overcome the problem that a single information source cannot recognize locomotion-mode, human locomotion-mode multi-source information system was set up. hip joint angle、and acceleration signals and plantar pressure were used as the major source. Human locomotion-mode were decomposed into four segments using the plantar pressure information. hip joint angle features and acceleration signal features were fused into one feature vector according to the user’s locomotion modes characteristics. the features were reduced (at the cost of information content) using principal component analysis (KPCA). These data are initially used to train ELM(Extreme Learning Machine) models, which classify the patterns as upslope, downgrade, stairs ascent, stairs descent or level-ground walking. The average classification accuracy of level-ground walking .stairs ascent .stairs descent. upslope .downgrade is 96.8%. The average recognition time is 0.87 s, The results showed that the strategy was superior to SVM 、BP method