Abstract: To address the challenges posed by the complex, non-periodic, and multi-modal characteristics of climbing and working on energized poles for exoskeleton assistance control, a recognition method based on plantar pressure and deep reversible consistency learning(DRCL) is proposed. A plantar pressure data collection system is constructed, and the center of pressure(COP) and various derived features are extracted as the main inputs. Combined with the DRCL method, multi-modal feature fusion and robust recognition are achieved through selective prior learning, semantic consistency constraints, and modality-invariant representation reshaping. Experimental results show that the COP-derived features achieve F₁ values above 0.99 under both level ground walking and pre-pole-climbing preparation modes. After model training, the recognition accuracy for three subjects reaches a maximum of 98.1%, with an average accuracy of 94.4%. This method significantly improves the accuracy and robustness of mode recognition for climbing and working on energized poles, providing a reliable basis for exoskeleton robots to achieve real-time human-robot collaborative assistive control.

Key words: plantar pressure, deep reversible consistency learning(DRCL), work modes recognition, exoskeleton robot