Accurate identification and prediction of a wearer's motion intention in real time are necessary to realize the compliant motion control of exoskeleton robots. Thus, we use the multi-sensor information fusion method to recognize the wearer's motion intention. The comparison of various machine learning algorithms with respect to recognition accuracy, resource consumption, and real-time processing revealed that support vector machine can recognize eight daily motion patterns (sitting, standing, walking, running, ramp ascent, ramp descent, stairs ascent and stairs descent), at an average recognition accuracy rate of 95%. The neuro-fuzzy inference theory is adopted to predict motion phase and motion switching events. On the given test set, the phase recognition accuracy rate is 99%, and the average absolute value of the deviation between the predicted and real-time state switching moments is 61.5 ms. This observation meets the requirements of exoskeleton compliance control for predicting time.