A CNN-SVM MODEL USING IMU FOR LOCOMOTION MODE RECOGNITION IN LOWER EXTREMITY EXOSKELETON

Human activity intention is indispensable for wearable powered lower extremity exoskeleton such that ensuring the compliant control of the robot. Lots of researches have been done on gait phase detection, which served as a sub-module of locomotion mode recognition to support the follow-on task. Therefore, it is self-evident that locomotion mode recognition is of great importance. Many model-based recognition methods are usually applied in manual extraction of cumbersome features, such as the traditional neural network (NN), support vector machine (SVM), etc. In contrast, the feature mapping layer coming with the convolutional neural network (CNN) can effectively solve the above time-consuming problem. Given that the training of NN is prone to overfitting, SVM with optimal characteristics is considered. A hybrid CNN-SVM model is proposed to identify human locomotion modes by collecting multi-channel inertial measurement unit (IMU) signals and is integrated with the error correction function of the finite state machine (FSM). Therefore, the CNN-SVM model has great influence on the generalization performance and recognition accuracy. The recognition rates of five single locomotion modes and eight mixed locomotion modes reach 97.91% and 98.93%, respectively. The system meets the demand of real-time performance, and the recognition time exceeds 370 ms on heel strike.