A Robust Multi-Scale Depthwise Separable With Dual-Reservoir Bi-LSTM Model for Gait Phase Recognition Across Complex Terrains

Gait phase recognition plays a pivotal role in improving wearable exoskeletons by providing phase-specific assistance to enhance human movement. This study proposes a novel model, the Multi-Scale Depthwise Separable Convolution with Dual-Reservoir Bi-directional Long Short-Term Memory (MSDSC-DR-BiLSTM), which recognizes four key gait sub-phases: heel strike, foot-flat, heel-off, and swing. The model employs a multi-scale depthwise separable convolution layer to capture both local and global gait features, while a dual-reservoir BiLSTM network effectively models temporal dependencies for precise phase transitions. To tackle class imbalance, a Focal Loss function is employed. We validated the model on a dataset of gait data from 25 healthy subjects across five terrains: level ground walking, stair ascent, stair descent, ramp ascent, and ramp descent. Experimental results show that the MSDSC-DR-BiLSTM model achieves over 95% accuracy in subject-independent 5-fold cross-validation. In leave-one-out cross-validation, the accuracy for each terrain is as follows: level ground walking (95.45%), stair ascent (98.69%), stair descent (96.94%), ramp ascent (99.27%), and ramp descent (97.3%). These results demonstrate the model’s robustness and adaptability, establishing a reliable foundation for phase-specific assistance in wearable lower-limb exoskeletons.