Bias Normalization for Sensors in Smart Devices

Modern electronic devices, such as smartphones and drones, integrate various sensors to enable diverse sensor-based applications. Yet, sensor measurements exhibit significant variations across different device models, even under the same environment. These variations arise from sensor biases, which occur in three different types: offset bias (additive constant errors), scale bias (multiplicative proportional errors), and drift bias (time-dependent or temperature-dependent errors). Among the biases, in this paper we specifically target offset bias, which has the greatest impact in typical smartphone usage scenarios. This generally leads to performance degradation in sensor-based applications across various device models and instances. To understand the characteristics of the offset bias, we categorize sensors into sensors with and without absolute reference values. Sensors with absolute references enable direct calibration using theoretical true values, while sensors with relative references require different approaches depending on how sensor applications process the data. For scalar-based applications that determine the current state by comparing a sensor measurement against a pre-defined reference, the offset biases can be removed by the existing procedures using reference devices. However, for sequence-based applications that determine the current state by analyzing relative changes in a sequence, the offset bias issue has not been addressed yet. We propose initial value removal and mean removal algorithms that statically and dynamically remove the offset biases from the sensor data sequences for these sequence-based applications. We evaluate our bias normalization algorithms for two different use cases in a geomagnetic-based indoor positioning system (IPS). First, we evaluate the impact of our bias normalization algorithms on the positioning performance of our LSTM-based IPS. Without bias normalization, although the reference device (Galaxy S23 Plus) showed an average positioning error of 0.6 m, the other three smartphone models (Galaxy S22 Plus, iPhone 15, and iPhone 16 Pro) exhibited much worse positioning performance, with errors of 2.48 m, 18.21 m, and 13.13 m. However, after applying our bias normalization, the average positioning errors of all models dropped below 0.68 m. Second, we also evaluate the impact of the bias normalization on detecting whether the position of a smartphone is in a pocket or in a hand-held state. For this, we analyze the sequence of light sensor measurements. We improved the detection accuracy from 42.3% to 97.6% with bias normalization across all device models without requiring individual threshold settings.