Personalized Gait Generation Using Convolutional Neural Network for Lower Limb Rehabilitation Robots

In walking training, tailoring the gait trajectories to accommodate individual patient conditions is pivotal for improving the effectiveness and comfort of rehabilitation. This paper proposes a novel gait generation model based on 1D convolution techniques, capitalizing on the correlation between gait patterns and physiological parameters to learn personalized gait patterns. Firstly, the method applies Fourier fitting to the gait trajectories, incorporating data from hip and knee joints, thereby transforming fitted parameters into a comprehensive representation of the gait trajectory. To improve the network’s capability of feature extraction, a multi-branch architecture is utilized, incorporating the squeeze and excitation (SE) blocks to fine-tune the channel weights. Furthermore, the model undergoes structural re-parameterization, a process that converts the multi-branch structure into a single-path structure. This procedure significantly reduces memory usage and computational demand without sacrificing accuracy. Comparative analysis reveals that our model outperforms existing benchmarks, achieving the lowest average loss of 0.1124, thereby demonstrating exceptional generalization capabilities. The optimized model further achieves a 35.63% reduction in inference time, with a negligible average error discrepancy. The generated gait trajectories are intended to serve as reference paths for lower limb rehabilitation robots, thereby offering the groundwork for delivering personalized training to patients.