Incremental learning of LSTM framework for sensor fusion in attitude estimation

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Fusion

Background:

• Traditional attitude estimation relies on sensor fusion methods like Extended Kalman Filters, which struggle with real-world motion dynamics and uncertainties.
• Inertial sensors (gyroscope, accelerometer, magnetometer) are crucial for attitude estimation but require sophisticated processing for accuracy.

Purpose of the Study:

• To develop a novel, robust, and efficient attitude estimation method using incremental learning of a Long-Short Term Memory (LSTM) network.
• To address the limitations of traditional methods in handling dynamic motion and sensor uncertainties.

Main Methods:

• Inertial sensor data (gyroscope, accelerometer, magnetometer) are processed using an incrementally learning Long-Short Term Memory (LSTM) network.
• The LSTM network is updated in real-time to adapt to dynamic changes in object motion.

Main Results:

• The proposed LSTM-based framework demonstrates significant improvements in attitude estimation accuracy and robustness.
• The method outperforms traditional Extended Kalman Filter and Complementary Filter approaches, especially in highly dynamic environments.
• Validation was performed using data from a commercial inertial measurement unit.

Conclusions:

• Incremental LSTM learning offers a superior approach to attitude estimation compared to conventional sensor fusion techniques.
• The developed framework is suitable for real-time applications and can be deployed on AI-supported processing modules.