LSTM-Assisted SINS/2D-LDV Tightly Coupled Integration Approach Using Local Outlier Factor and Adaptive Filter

Abstract

The tightly coupled integration of strapdown inertial navigation system (SINS) and two-dimensional laser Doppler velocimeter (2D-LDV) enhances system robustness by directly using raw 2D-LDV measurements, making it well-suited for land autonomous navigation. However, this approach struggles with long-term failures of individual or both 2D-LDV beams and is sensitive to violations of the vehicle's lateral zero-velocity constraint, which can degrade performance. To address these limitations, this paper proposes a novel approach involving a long short-term memory (LSTM)-assisted SINS/2D-LDV tightly coupled integration, incorporating a local outlier factor (LOF) and an adaptive filter. The LOF is introduced to evaluate the anomaly degree of the system’s measurements, while offline datasets constructed and classified from historical normal data improve detection accuracy and reduce computational load. The LSTM is used to predict the 2D-LDV measurements and the vehicle’s lateral velocity, substituting these predictions for anomalous data to mitigate their impact. Furthermore, an adaptive filter is used to adjust the measurement noise covariance matrix of the navigation filter to avoid the adverse effects of potential errors in LSTM predictions. The effectiveness of the proposed method is validated through two groups of experiments, demonstrating satisfactory performance under both normal conditions and prolonged single or dual beam failures in the 2D-LDV.

Keywords

• Land autonomous navigation
• Strapdown inertial navigation system/ two-dimensional laser Doppler velocimeter (SINS/2D-LDV) tightly coupled system
• Local outlier factor
• Long short-term memory network
• Adaptive filter

Recommended citation: Z. Xiang, Q. Wang, X. Nie, S. Jin, and J. Zhou, "LSTM-Assisted SINS/2D-LDV Tightly Coupled Integration Approach Using Local Outlier Factor and Adaptive Filter," IEEE. Trans. Instrum. Mea, vol. 74, p. 1-15, 2025.