손실함수를 재정의한 GRU 기반 학습 알고리즘을 통한 보행위상 예측 성능 향상

Abstract

In this study, we propose a user-defined loss function to enhance a learning-based algorithm for gait phase prediction. The algorithm utilizes gait phase labels and time-series data from a 6-axis IMU as inputs, employing a Gated Recurrent Unit (GRU) model for the prediction. A novel loss function is introduced, incorporating both gait phase error and error regulation to improve performance. The trained model autonomously generates two distinct outputs to enhance gait phase prediction, which are combined with the atan2 function to estimate the gait phase. Phase differences between the labeled and predicted gait phases are calculated and regulated within the loss function. To validate the proposed method, a pilot experimental study was conducted involving overground walking tests with five subjects and a treadmill walking test with one subject at various walking speeds. The results demonstrate precise gait phase predictions throughout the entire gait cycle, outperforming conventional approaches. This GRU-based model shows promise for implementation in embedded systems.