Optimization of an unscented Kalman filter for an embedded platform

Area of Science:

• Computational Biology
• Embedded Systems Engineering
• Biomedical Signal Processing

Background:

• The unscented Kalman filter (UKF) is increasingly used in biological applications.
• Implementing complex UKF systems on low-power embedded platforms presents significant design challenges.
• Optimization is crucial for achieving real-time performance and accuracy in wearable biological monitoring.

Purpose of the Study:

• To develop and present a method for optimizing UKF systems for embedded platforms.
• To minimize computation time and state reconstruction/forecasting error.
• To identify an optimized UKF variant suitable for biological modeling on resource-constrained devices.

Main Methods:

• A three-stage optimization process assessing computation time, state forecast error, and state reconstruction error.
• Evaluation of 432 UKF variants for a rat sleep-wake regulatory model.
• Application of a cost function to filter and select optimal UKF variants.

Main Results:

• An optimized UKF variant was identified, achieving 27x faster computation than the reference.
• The optimized variant maintained required state estimation and forecasting accuracy.
• Key insights into the roles of process noise and data assimilation in optimization were gained.

Conclusions:

• The proposed optimization method effectively reduces UKF computation time for embedded biological systems.
• Process noise and data assimilation offer opportunities for UKF simplification and speed-up.
• Decoupling model-dependent and structure-dependent variables accelerates the optimization process.