t分布的异步多速率系统迁移学习滤波算法

Asynchronous multi-rate systems of a transfer learning filtering algorithm using t-distribution

中文关键词  状态估计  异步多速率传感器  t分布  迁移学习  变分贝叶斯

英文关键词  state estimation  asynchronouus multi-rate sensors  t-distribution  transfer learning  variational Bayesian

中文摘要

      针对异步多速率传感器在状态估计过程中受异常值影响问题, 本文研究了一种基于t分布的迁移学习滤波 算法. 文中结合全概率设计和多尺度系统理论, 设计了新型异步多速率系统的迁移学习方法, 源域和目标域具有不 同采样速率传感器, 通过建立多尺度模型, 将异步多速率系统转化为同步多速率系统. 使得文中所设计方法具有异 步多速率系统在最小化源域预测分布至目标域理想分布的Kullback-Leibler散度同时, 允许传感器采样速率之比为 任意正整数的优势. 考虑异常值对状态估计的影响, 源域和目标域依赖于t分布的重尾性质来对状态和观测过程建 模, 通过期望最大化和变分贝叶斯进行近似估计. 最后, 所提出方法被应用于平面位置速度系统的速度位置估计, 仿 真结果验证了其该方法的有效性.

英文摘要

      This paper investigates a transfer learning filtering algorithm based on the t-distribution to address the problem of state estimation in asynchronous multi-rate sensor systems affected by outliers. By integrating full probability design and multi-scale system theory, we propose a novel transfer learning framework for asynchronous multi-rate systems. A multi-scale model is established to convert the asynchronous multi-rate system into a synchronous one. This design minimizes the Kullback-Leibler divergence between the predicted distribution in the source domain and the ideal distribution in the target domain, while allowing the sensor sampling rate ratio to be any positive integer. To account for outlier effects on state estimation, the source and target domains leverage the heavy-tailed properties of the t-distribution to model state and observation processes, with approximate estimation achieved through expectation-maximization and variational Bayesian methods. Simulation results on a planar position-velocity system demonstrate the superior performance of the proposed method.