Motion State Detection Model for Rollover-Assistive Robots Based on Pressure Features

To address resource waste and insufficient real-time performance resulting from the continuous operation of a dual-model system, this paper proposes a robot motion state detection model based on pressure features and Support Vector Machine (SVM) to detect the robot's motion state and dynamically select models. First, the SVM algorithm takes advantage of the kernel function's ability to perform spatial dimensionality expansion, which is particularly useful for handling nonlinear classification problems. Next, the pressure feature acquisition sites were determined by collecting data from eight pressure sensors located at the shoulder and hip branches, forming an eight-dimensional pressure feature set. A hybrid strategy combining “grid search + particle swarm optimization” (PSO) was then used to optimize the SVM kernel function parameters (penalty coefficient C and γ values). Experimental results showed that the RBF kernel SVM model achieved the highest accuracy of 98.96% on the test set, outperforming both the polynomial and Sigmoid kernels. Ablation experiments further validated the effectiveness of feature point selection and PSO optimization. Finally, a comparative experiment with three other typical machine learning algorithms confirmed the applicability of the overall model. This model can effectively distinguish the robot's motion state and guide the selection of dual estimation models, providing both theoretical support and practical reference for the intelligent control of nursing robots.