Scholarlink

<p>To achieve the seamless robust and accurate drone localization, an adaptive filter is developed, in which the expectation-maximization (EM) method is integrated with a residual unbiased finite impulse response (rUFIR) filter to enhance performance using long short-term memory (LSTM). The EM method is used to provide adaptive estimation of noise covariances via the Mahalanobis distance and the drone’s position over the ultra wide band (UWB) data. The residual information is extracted from the rUFIR filter. When UWB fails, the system switches to a dual operation mode. When LSTM is oriented towards, errors in the inertial navigation system (INS) position are corrected along one direction, and the rUFIR filter output is adjusted accordingly. It is shown experimentally that the proposed approach outperforms the adaptive Kalman filter (AKF)+long short term memory (LSTM), FIR+LSTM, dual maximum correntropy KF (mcKF)+least squares-support vector machine (LS-SVM), and dual FIR+LSTM in terms of accuracy in position estimation. Experimental testing shows that the average position error reduction can be achieved at the level of 90% compared to the FIR+LSTM method.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive expectation maximization-based FIR filter with dual LSTM assisted correction for seamless INS-based indoor drone localization

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Yuan Xu,
  • Yuriy S. Shmaliy,
  • Xiyuan Chen,
  • Yanli Gao,
  • Shunyi Zhao

摘要

To achieve the seamless robust and accurate drone localization, an adaptive filter is developed, in which the expectation-maximization (EM) method is integrated with a residual unbiased finite impulse response (rUFIR) filter to enhance performance using long short-term memory (LSTM). The EM method is used to provide adaptive estimation of noise covariances via the Mahalanobis distance and the drone’s position over the ultra wide band (UWB) data. The residual information is extracted from the rUFIR filter. When UWB fails, the system switches to a dual operation mode. When LSTM is oriented towards, errors in the inertial navigation system (INS) position are corrected along one direction, and the rUFIR filter output is adjusted accordingly. It is shown experimentally that the proposed approach outperforms the adaptive Kalman filter (AKF)+long short term memory (LSTM), FIR+LSTM, dual maximum correntropy KF (mcKF)+least squares-support vector machine (LS-SVM), and dual FIR+LSTM in terms of accuracy in position estimation. Experimental testing shows that the average position error reduction can be achieved at the level of 90% compared to the FIR+LSTM method.