Loading...

Noise Covariance Prediction of IMU using ANN in Dead-Reckoning Navigation for Constrained Dynamics

Najjarnasab, Mohammad Hassan | 2026

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58794 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Salarieh , Hassan; Zohoor, Hassan
  7. Abstract:
  8. Dead Reckoning navigation, as an infrastructure-independent approach, plays a key role in scenarios where access to the Global Navigation Satellite System (GNSS) is unavailable; however, its performance is consistently affected by the accumulation of errors in inertial measurement unit (IMU) sensors. In this research, a novel approach is proposed to enhance the accuracy of this method. In the first stage, sensor noise modeling is performed using three innovative approaches. The first approach involves a static tuning based on an adaptive Extended Kalman Filter (AEKF). In this method, the discrete process noise covariance matrix is adaptively estimated, and the continuous-time process noise covariance is extracted using time-averaged analytical relationships. The results demonstrate that the proposed method improves position estimation accuracy by 51% for a commercial-grade inertial sensor and by 33% for a tactical-grade inertial sensor. In the hybrid optimization approach based on filter performance, a metaheuristic optimization algorithm is employed to minimize position estimation error under GNSS-denied conditions. In this framework, the initial condition is estimated based on the process noise covariance matrix. The obtained results indicate that the position and attitude estimation errors are reduced by 52% and 84%, respectively, compared to the conventional Allan variance–based tuning method. In the third approach, a convolutional neural network (CNN) is utilized to predict the Allan variance parameters of the inertial sensor and consequently to tune the process noise covariance matrix. To this end, Allan variance tests are conducted on inertial sensors with different quality grades under two conditions: laboratory environment and a stationary vehicle with the engine running. Subsequently, a CNN is trained to replace the time-consuming Allan variance test results for predicting or retuning the inertial sensor noise covariance matrix. The results demonstrate satisfactory generalization capability of the network under various conditions, particularly for commercial-grade inertial sensors. In the next stage, based on the extracted statistical characteristics of the inertial sensor noise, the vehicle kinematic velocity constraint is employed as auxiliary observations in vehicle navigation. The uncertainty and noise covariance matrix of this pseudo-measurement are also estimated using the CNN. The final contribution of this research is the successful real-time implementation of the proposed algorithm on a microcontroller platform using symbolic regression of the convolutional neural network, enabling stable and robust dead reckoning navigation. Moreover, an Extended Kalman Filter is developed to estimate the state of the nonholonomic constraint in real time, including the lever arm and the rotation angles between the body frame and the vehicle frame, in order to apply the velocity pseudo-measurement more accurately. The results of this section show that a more precise application of pseudo-measurements, considering the governing system dynamics, leads to a significant improvement in navigation accuracy.
  9. Keywords:
  10. Inertial Measurement Unite (IMU) ; Convolutional Neural Network ; Deep Learning ; Nonholonomic Constraints ; Extended Kalman Filter ; Adaptive Kalman Filter (AKF) ; Dead Reckoning Navigation ; Noise Covariance

 Digital Object List

 Bookmark

No TOC