Denoising-Based Improved Deformation Estimation Method for Self-Balancing Lower Limb Exoskeleton

This paper proposes a denoising-based improved deformation estimation method for a self-balancing lower limb exoskeleton (SBLLE). A denoiser constructed by a residual network is utilized to improve the accuracy of deformation estimation for SBLLE. Specifically, we introduce a residual network-bidirectional long short-term memory network (ResNet-BiLSTM) denoising deformation estimator (RBDDE) to estimate and compensate for the deformation of SBLLE, based on force and torque (F/T) data. F/T data and deformation data are collected separately using F/T sensors embedded in the sole of SBLLE and a motion capture system. Initially, the empirical mode decomposition method is employed on the training set to construct a dataset for training the ResNet denoising network, enabling the denoising network with denoising capabilities. Subsequently, the ResNet-BiLSTM network is trained using both F/T and deformation data, enabling the estimation and compensation of SBLLE deformation at its center of mass. Comparative experiments with a deformation estimation network without denoising show that RBDDE improves the accuracy of deformation estimation.