Pattern transition recognition based on transfer learning for exoskeleton across different terrains.

Human motion intention detection is a growing trend in wearable robots. In the study, a novel transfer learning method based on temporal convolutional network spatial attention (TCN-SA) is applied for pattern transition recognition under triple physical loads on different terrains. The proposed approach is used to recognize eight locomotion modes transition among five dynamic locomotion modes in sequence, such as level ground walking, stair ascending, stair descending, ramp ascending, and ramp descending. To address the problem of pattern transition recognition, transfer learning adapts a model from the source domain to the target domain. Temporal convolutional network (TCN) relies on local relationships in time sequence and gains steady gradient propagation. Furthermore, spatial attention (SA) provides insight into significant components in multi-dimensional feature selection. Pattern transition recognition based on a transfer learning method achieves higher accuracy and earlier prediction time (Pre-T). The accuracy of pattern transition detection reaches 97.46%, 97.62%, and 98.21% in M0, M20, and M40, respectively. In the process of pattern transition recognition, Pre-T of next locomotion mode in M0, M20, and M40 are 240-600 ms, 200-410 ms, and 120-420 ms before the step into that locomotion mode. The proportion of prediction time in a gait cycle (Pre-T/GC) in M0, M20, and M40 is 14.2-36%, 14.82-28.22%, and 7.4-29.1%, respectively. Ultimately, the results indicate that the proposed approach fulfills the expected performance in Pre-T and comparisons with TCN-attention, TCN, residual network (ResNet), and long short-term memory (LSTM) in assessment criteria. Our study early detects pattern transition, allowing the exoskeleton to traverse between adjacent terrains smoothly.