Accurate recognition of gait phases and reliable prediction of gait cycles are critical for adaptive exoskeleton control. Although these two tasks are inherently coupled in both temporal and spatial domains, they have often been investigated in isolation. Such separation constrains the robustness and responsiveness of assistive devices in real-world applications. To bridge this gap, we propose a multi-task learning framework that integrates a Stacked Denoising Autoencoder with a Long Short-Term Memory network (SDA-LSTM). Inertial measurement unit (IMU) signals are employed for gait analysis, enabling both discrete gait phase classification and continuous estimation of gait cycle percentage during forward walking. The framework was evaluated against several established models, including Support Vector Machines (SVM), XGBoost, and standalone LSTM networks, under cross-subject validation. For phase classification, SDA-LSTM achieved 97.3% accuracy, with only a small reduction of 1.31% compared with its training performance. Robustness was further demonstrated under severe noise conditions (signal-to-noise ratio, SNR = 5 dB), where the model maintained 95.68% accuracy, surpassing all baseline methods. For cycle prediction, SDA-LSTM also showed strong stability. At SNR = 5 dB, error metrics increased only slightly, with RMSE and MAE rising by 0.083 and 0.075, while R 2 and PCC decreased marginally by 0.0173 and 0.0089. These results highlight the effectiveness of SDA-LSTM in capturing the spatiotemporal synergy of gait. The framework demonstrates high accuracy, robustness, and generalization. Its performance underscores strong potential for deployment in exoskeleton systems, paving the way toward reliable and adaptive human–robot interaction in daily locomotion.
Pang et al. (Sun,) studied this question.