Bias Normalization for Sensors in Smart Devices.

Area of Science:

• Sensor fusion and calibration
• Mobile device sensing technologies
• Machine learning for sensor data analysis

Background:

• Modern electronic devices rely on diverse sensors for various applications.
• Sensor measurements vary significantly across different device models due to biases (offset, scale, drift).
• Offset bias, a common issue, degrades performance in cross-device sensor applications.

Purpose of the Study:

• To address the unaddressed issue of offset bias in sequence-based sensor applications.
• To propose and evaluate novel algorithms for removing offset biases from sensor data sequences.
• To enhance the performance and accuracy of sensor-based applications across diverse smartphone models.

Main Methods:

• Categorization of sensors based on absolute or relative reference values.
• Development of initial value removal and mean removal algorithms for offset bias normalization.
• Evaluation in a geomagnetic-based indoor positioning system (IPS) using LSTM models and light sensor data.

Main Results:

• Bias normalization drastically reduced average positioning errors in an IPS from up to 18.21 m to below 0.68 m across different smartphone models.
• Detection accuracy for smartphone states (pocket vs. hand-held) using light sensor data improved from 42.3% to 97.6% with bias normalization.
• The proposed algorithms demonstrated effectiveness across various device models without requiring model-specific threshold tuning.

Conclusions:

• Offset bias significantly impacts sequence-based sensor applications, particularly in indoor positioning and state detection.
• The proposed initial value removal and mean removal algorithms effectively mitigate offset bias across diverse devices.
• This work enables more robust and accurate sensor-based applications on smartphones, regardless of device model variations.