Improving displacement estimation by mems accelerometer using deep q-learning algorithm

Abstract

Precise estimation of movement trajectories using cheap MEMS accelerometers could have a wide range of applications, such as simplified industrial robot control, drone navigation, etc. Cheap accelerometers, however, have significant errors in acceleration measurements. These compound to extremely large deviations when calculating displacements which result in estimated trajectories becoming completely different from the real ones. Traditional filtering techniques ranging from simple threshold filter to more complex Kalman filter can be used to improve the measurement accuracy with varying levels of success. For this application, however, these filters either make the whole system insensitive or require dedicated tuning of filter parameters for each trajectory measurement case. Applying machine learning (ML) algorithms for improving both accuracy and consistency of trajectories estimated based on accelerometer data is a promising area of research. There already are studies confirming that both LSTM-based ML algorithm and ML-Kalman filter combinations offer improved performance compared to traditional filtering methods. To supplement that research, this work explores the application of deep q-learning algorithm to improve displacement accuracy evaluation based on accelerometer data. The method involves training the algorithm on ideal displacement data and the displacement datasets calculated from raw acceleration data. The algorithm outputs corrected displacement value for each datapoint. To smooth out the displacement curve, approximating filter is applied. One of preliminary results of this research is analyzed in this work. It's visualized as displacement curves of moving a MPU6050 accelerometer in a straight line along one axis at constant velocity for 0.4 m. distance. There is a complete mismatch between the uncorrected displacement (calculated from raw acceleration data) and the one that was performed in reality (ideal case). Meanwhile, the displacement corrected by a deep q-learning algorithm and then passed through an averaging filter matches the ideal case closely. The average deviation from the ideal case is around 5%, while maximum deviation is around 18%, which is a significant accuracy improvement for such type of accelerometer. However, while the results are promising, further research is required to make the technique applicable in practise: the algorithm needs to be made less reliant on ideal displacement data, while its performance for motion in complex trajectories needs to be investigated.