Lightweight Attention Temporal Convolutional Network for Locomotion Scene Recognition of Lower-Limb Exoskeletons

Abstract

A lightweight attention-based temporal convolutional network (LTA-TCN) is proposed for lower-limb exoskeleton locomotion-scene recognition under resource constraints. A database was collected from ten healthy participants using bilateral motor-encoder and inertial measurement unit angle signals. After synchronization, filtering, differentiation, and standardization, 12-dimensional kinematic features were segmented into 0.5-s windows. LTA-TCN combines depthwise separable dilated residual convolutions with attention-weighted temporal pooling. Under leave-one-subject-out cross-validation with five random seeds per fold, the model achieved an accuracy of 96.18±0.84% and a Macro-F1 score of 95.91±0.93% with 0.024 M parameters. Its mean CPU forward-inference time was 2.96±0.19 ms and the complete processing cycle required 5.2 ms. Compared with standard TCN, LTA-TCN reduced the parameter count by 66.20% while improving accuracy by 0.55 percentage points, indicating a favorable accuracy-efficiency trade-off on the evaluated CPU platform while leaving direct embedded-hardware validation for future work.