Combining Fast Orthogonal Search with Deep Learning to Improve Low-Cost IMU Signal Accuracy.

Area of Science:

• Navigation Systems
• Sensor Fusion
• Machine Learning for Signal Processing

Background:

• Low-cost inertial measurement units (IMUs) exhibit significant drift and noise.
• Errors stem from sensor biases, scale factor instability, and nonlinear stochastic noise.
• Accurate IMU data is crucial for navigation, especially in Global Navigation Satellite System (GNSS)-denied scenarios.

Purpose of the Study:

• To develop a hybrid error compensation approach for improving IMU signal accuracy.
• To combine Fast Orthogonal Search (FOS) and Long Short-Term Memory (LSTM) neural networks.
• To reduce drift and noise errors in IMU data for enhanced navigation.

Main Methods:

• A hybrid approach combining FOS for deterministic error modeling and LSTM for time-dependent error dynamics.
• FOS predicts high-end IMU output from low-end IMU data, extending training data for LSTM.
• FOS-based pseudo-label bootstrapping for data-efficient LSTM training.

Main Results:

• The FOS-LSTM hybrid model achieved high predictive accuracy (MSE < 0.1%) in predicting high-end IMU signals from low-end signals.
• The hybrid model yielded more stable velocity estimates compared to FOS or LSTM alone.
• Significant reduction in short-term uncertainty for inertial navigation solutions.

Conclusions:

• The FOS-LSTM hybrid method offers a promising synergy between model-based identification and data-driven learning for IMU error calibration.
• This approach enhances IMU accuracy in GNSS-denied navigation systems.
• The study highlights the impact of signal correction on dead reckoning drift reduction.