A New In-Flight Alignment Method with an Application to the Low-Cost SINS/GPS Integrated Navigation System

Zhenglong Lu, Jie Li, Xi Zhang, Kaiqiang Feng, Xiaokai Wei, Debiao Zhang, Jing Mi, Yang LiuSensors 2020 · DOI

TL;DR

This paper introduces a novel, fast, and accurate in-flight alignment method for low-cost SINS/GPS navigation systems. It addresses the limitations of traditional methods by proposing a modified double-vector construction to mitigate sensor bias accumulation and a gradient descent-based attitude estimation for improved speed and reduced computational burden.

What Problem Does This Paper Address?

Accurate attitude determination is crucial for the performance of Strapdown Inertial Navigation Systems (SINS). While SINS/GPS integrated systems offer improved accuracy, the initial alignment process, especially during flight (in-flight alignment or IFA), remains a challenge, particularly for low-cost systems. Traditional optimization-based alignment (OBA) methods, while accurate, suffer from significant computational demands and are susceptible to accumulated errors from low-precision MEMS gyroscopes and accelerometers. This accumulation leads to inaccurate "double-vectors," which are essential for attitude estimation, ultimately degrading alignment accuracy and speed. The paper aims to bridge this gap by developing an IFA method that is both accurate and computationally efficient, making it suitable for low-cost SINS/GPS platforms.

Methods

The proposed method consists of two main components:

  1. Modified Construction Method of Double-Vectors:
    • Reduced Integration Interval: The integration time for specific force measurements is shortened to minimize the accumulation of accelerometer bias.
  2. Gyroscope Bias Identification: A linear Kalman filter is employed to identify and compensate for gyroscope bias. This is crucial because gyroscope bias, when integrated over time, significantly impacts the accuracy of the double-vectors. By identifying and feeding back this bias, the accuracy of the double-vectors is substantially improved.
  3. Fast IFA Method Based on Gradient Descent:
    • Gradient Descent for Attitude Estimation: Instead of complex matrix operations like Singular Value Decomposition (SVD) or solving quartic equations (as in QUEST), this method uses gradient descent to iteratively find the optimal attitude that minimizes an objective function. This significantly reduces computational complexity.
  4. Mini-Batch Gradient Descent (MBGD): The paper explores different batch sizes for gradient descent. While Batch Gradient Descent (BGD) offers the best accuracy and robustness, it's computationally intensive. Stochastic Gradient Descent (SGD) is fast but less stable. MBGD strikes a balance, allowing for adjustable performance based on the mini-batch size.

Datasets

The paper utilizes simulated data for initial testing and validation. For the vehicle experiment, real-world data was collected using a vehicle equipped with a high-precision navigation system (NovAtel) for ground truth.

Key Insights and Ideas

  • Bias is the Enemy: The core insight is that the accumulated bias from low-cost inertial sensors is the primary bottleneck for accurate IFA. The paper demonstrates that directly addressing this bias, both in accelerometer and gyroscope measurements, is paramount.
  • Double-Vectors are Sensitive: The accuracy of the "double-vectors" directly dictates the quality of the attitude estimation. The proposed modification to construct these vectors more robustly is a key enabler.
  • Gradient Descent for Efficiency: Shifting from complex matrix algebra to iterative gradient descent for attitude estimation is a clever way to reduce computational load, making the method practical for resource-constrained systems.
  • Mini-Batch Tuning: The performance of gradient descent is highly tunable. The choice of mini-batch size offers a trade-off between speed, accuracy, and robustness, allowing for adaptation to different operational needs.

Key Results

  • Accuracy Improvement: The proposed method significantly reduces attitude angle errors. For instance, the maximum level attitude angle error was reduced from 1.17° to 0.05°, and yaw angle error from 20.34° to 0.705° in simulations with biased sensors.
  • Speedup: Alignment time was drastically reduced, from 39.2 seconds to 7.6 seconds in simulations.
  • Vehicle Experiment Validation: Real-world experiments confirmed the superiority of the proposed method over traditional QUEST, showing improved accuracy and faster alignment times. For example, alignment time was reduced from 38.79 s to 6.08 s.
  • MBGD Performance: Simulations showed that MBGD with a mini-batch size of 60 provided a good balance between accuracy and speed, outperforming SGD and being comparable to BGD in accuracy while being faster.

Limitations and Future Work

  • GPS Measurement Noise: The paper acknowledges that GPS measurement noise is a significant factor contributing to alignment error. It suggests that using higher precision GPS could further improve accuracy.
  • Mini-Batch Selection: While the paper demonstrates the impact of mini-batch size, selecting the optimal size for specific real-world scenarios (balancing computation speed and precision) is highlighted as a practical consideration for future implementation.
  • Sensor Disturbances: The paper notes that BGD is more robust when measurements are seriously disturbed, suggesting that further investigation into robust estimation techniques for highly noisy environments could be beneficial.

Summary updated

BibTeX
@article{lu2020new,
  author = {Zhenglong Lu and Jie Li and Xi Zhang and Kaiqiang Feng and Xiaokai Wei and Debiao Zhang and Jing Mi and Yang Liu},
  title = {{A New In-Flight Alignment Method with an Application to the Low-Cost SINS/GPS Integrated Navigation System}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20020512},
}
Figure