IMU- and LSTM-CNN-based multi-gait recognition for a lower limb exoskeleton

Abstract

This study presents a platform-specific and subject-independent validation of an IMU-based LSTM-CNN gait-recognition pipeline on a self-developed lower-limb exoskeleton. Ten healthy participants performed five locomotion modes, including level walking, stair ascent, stair descent, standing up, and sitting down. Each participant completed five level-walking trials and ten trials for each of the remaining four locomotion modes, resulting in 450 original recording trials. IMU signals collected from the thigh, shank, and foot were segmented using a 128-sample sliding window with a 64-sample step. To prevent information leakage from overlapping windows and to evaluate inter-subject generalization, the dataset was partitioned at the participant level before window segmentation. The LSTM layers modeled temporal dependencies in the IMU sequences, while Conv1D and pooling layers extracted local discriminative features. Under the subject-independent evaluation protocol, the model achieved an overall accuracy of 94.08%, a macro-F1 score of 93.73%, and a balanced accuracy of 93.62%. These results support the preliminary feasibility of the implemented IMU-based LSTM-CNN workflow for offline gait-mode recognition on the tested lower-limb exoskeleton platform. Further validation under larger participant cohorts and online closed-loop conditions is still required.