A Mixed Optimization Method Based on Adaptive Kalman Filter and Wavelet Neural Network for INS/GPS During GPS Outages

Xiaokai Wei, Jie Li, Kaiqiang Feng, Debiao Zhang, Pengyun Li, Lening Zhao, Yubing JiaoIEEE Access 2021 · DOI

Abstract

Maintaining navigation accuracy during GPS outages is a critical challenge for integrated Inertial Navigation System (INS) and Global Positioning System (GPS) platforms. This paper presents a hybrid optimization method, RFR-WNN+AKF, which combines an Adaptive Kalman Filter (AKF) with a Wavelet Neural Network (WNN) optimized by Random Forest Regression (RFR). The AKF utilizes an adaptive factor to suppress environmental noise, while the RFR-WNN ensemble provides high-precision pseudo-observations during GPS signal loss. Experimental results demonstrate that this approach significantly reduces positional drift and maintains stability even with limited training data, outperforming traditional single neural network and standard Kalman filtering methods.

Introduction

Integrated navigation systems typically pair an INS with GPS to provide continuous positioning. However, in environments like tunnels or dense urban areas (known as "urban canyons"), GPS signals are often lost. During these outages, the INS accumulates errors rapidly due to sensor drift. While Kalman Filters and Neural Networks are often used to bridge these gaps, they frequently suffer from non-linear noise sensitivity and overfitting. This study introduces a robust ensemble learning approach to mitigate these issues by combining adaptive filtering with optimized neural networks.

Methodology

The proposed system switches between two operational modes based on GPS availability to ensure continuous correction of the INS.

Training Phase (GPS Available)

When GPS signals are strong, the system calculates the difference between INS and GPS outputs (position and velocity). This data is used to train the RFR-WNN model, while the AKF simultaneously corrects INS errors in real-time using the GPS data as a reference.

System architecture when GPS signals are available for training.

Prediction Phase (GPS Outage)

During a GPS outage, the trained RFR-WNN ensemble takes raw INS data—specifically angular velocity and specific force—to predict the INS errors. These predictions serve as "pseudo-observations" for the AKF, allowing the filter to continue updating and correcting the INS output despite the lack of external GPS data.

System architecture during GPS outages, utilizing RFR-WNN predictions.

Key Innovations

  • Adaptive Kalman Filter (AKF): Unlike standard filters, the AKF uses an adaptive factor based on the innovation sequence to dynamically adjust to sensor outliers and complex environmental noise.
  • Wavelet Neural Network (WNN): Wavelet functions are used for activation, offering faster convergence and better modeling of non-linear uncertainties than standard sigmoid functions.
  • Random Forest Regression (RFR) Integration: By using an ensemble of WNNs trained on different data subsets (Bootstrap sampling), the system avoids the overfitting problems inherent in single-network architectures. This "Forest" approach is particularly effective when training samples are scarce.

Experimental Results

Vehicle experiments were conducted using MEMS-based sensors and a tactical-grade reference system to validate the method across different road conditions.

  • Accuracy Gains: In scenarios with sufficient training data (Scheme 1), eastern velocity error was reduced by 90% compared to pure INS (from 1.192 m/s to 0.101 m/s).
  • Robustness to Data Scarcity: In Scheme 2, where training data was reduced by 80%, the RFR-WNN+AKF maintained a northern position error of just 1.223m. Standard Back-Propagation Neural Network (BPNN) models showed significant performance degradation under the same conditions.
  • Computational Performance: The method is computationally feasible for modern navigation hardware, with processing times only slightly higher than standard filtering techniques, which is a worthwhile trade-off for the massive gain in accuracy.

Conclusion

The RFR-WNN+AKF method offers a significant advancement in robust navigation for GPS-denied environments. Its ability to generalize from limited data and adapt to complex noise makes it ideal for autonomous vehicles operating in challenging urban landscapes. Future research will focus on real-time embedded implementation and handling prolonged intermittent outages.

Summary updated

BibTeX
@article{wei2021mixed,
  author = {Xiaokai Wei and Jie Li and Kaiqiang Feng and Debiao Zhang and Pengyun Li and Lening Zhao and Yubing Jiao},
  title = {{A Mixed Optimization Method Based on Adaptive Kalman Filter and Wavelet Neural Network for INS/GPS During GPS Outages}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3068744},
}
Figure