基于LSTM与Integral-RLS的遥操作机器人环境刚度辨识与力反馈研究

Research on Environment Stiffness Identification and Force Feedback of Teleoperation Robots Based on LSTM and Integral-RLS

摘要: 针对遥操作机器人环境刚度辨识中，传统递归最小二乘法(RLS)在Hunt-Crossley模型下因对数线性化处理产生的零点奇点爆炸及参数耦合问题，提出一种融合LSTM时序增强与滑动窗口能量积分RLS (Integral-RLS)的辨识算法。首先，利用LSTM网络对含噪力觉信号进行预处理，在20 dB高斯白噪声环境下将信噪比提升了约15 dB，有效保留了接触瞬态特征。其次，创新性地构建了能量积分辨识算子，将瞬态动力学方程转化为宏观能量平衡方程，新算法较传统RLS的辨识震荡幅度降低了约85%，收敛时间缩短至0.3 s以内。最后，在AUBO-E5机械臂平台上进行验证，实验显示主从位置追踪误差在辨识反馈开启后降低了21%，显著提升了遥操作系统的力觉透明度与交互稳定性。

Abstract: In the identification of environment stiffness for teleoperated robots, traditional Recursive Least Squares (RLS) under the Hunt-Crossley model encounters issues of singularity explosion at zero and parameter coupling due to logarithmic linearization. This paper proposes an identification algorithm that integrates LSTM temporal enhancement with sliding window energy-integral RLS (Integral-RLS). First, the LSTM network is used to preprocess force feedback signals containing noise, improving the signal-to-noise ratio by approximately 15 dB under a 20 dB Gaussian white noise environment, effectively preserving transient contact features. Second, an innovative energy-integral identification operator is constructed, transforming transient dynamic equations into macroscopic energy balance equations. The new algorithm reduces the identification oscillation amplitude by about 85% compared to traditional RLS and shortens the convergence time to within 0.3 seconds. Finally, verification is carried out on the AUBO-E5 robotic arm platform. Experiments show that the master-slave position tracking error decreases by 21% after enabling identification feedback, significantly enhancing the force transparency and interaction stability of the teleoperation system.