基于LSTM的Stewart平台运动学正解研究

Research on the Kinematic Forward Solution of Stewart Platform Based on LSTM

摘要: Stewart平台的运动学正解是一个强非线性、多耦合的复杂问题。传统求解方法存在计算量大、实时性差、依赖初值及多解等问题，而现有神经网络方法大多仅建立从杆长到位姿的静态瞬时映射，无法利用杆长序列蕴含的时序依赖关系，导致连续轨迹预测精度受限。为此，本文采用一种基于长短期记忆网络的运动学正解方法。将运动学正解从“瞬时静态求解”转变为“时序状态预测”，构建三层堆叠LSTM网络，通过位姿、速度和加速度预测三个维度的仿真结果，全面评估LSTM网络的求解效果，验证其在Stewart平台动态轨迹预测中的有效性和优越性。

Abstract: The forward kinematics of the Stewart platform is a highly nonlinear, strongly coupled complex problem. Traditional solution methods suffer from issues such as high computational cost, poor real-time performance, dependence on initial values, and multiple solutions. Meanwhile, most existing neural network approaches only establish a static instantaneous mapping from leg lengths to platform pose, failing to exploit the temporal dependencies inherent in leg length sequences, which limits prediction accuracy for continuous trajectories. To address this, this paper adopts a forward kinematics method based on Long Short-Term Memory networks. By transforming the forward kinematics from an “instantaneous static solution” into a “temporal state prediction”, a three-layer stacked LSTM network is constructed. The performance of the LSTM network is comprehensively evaluated through simulation results in three dimensions—pose, velocity, and acceleration prediction—validating its effectiveness and superiority in dynamic trajectory prediction of the Stewart platform.