Assured time series forecasting using inertial measurement unit, neural networks, and state estimators

Abstract

Pedestrian dead reckoning (PDR) technology has be come an important method for predicting the position of an object or person. Sensor-based positioning is widely used because of its readily available hardware and acceptable accuracy, especially with PDR algorithms integrated with machine learning and deep learning. There are two challenges in this context. Conventional state-estimator methods suffers from dynamics, making the deployment and management of nonlinear dynamics become difficult. Training an effective neural network model with a few inertial measurement unit (IMU) samples is also challenging. This study investigates the integration and comparison of advanced state estimation algorithms such as the Kalman filter (KF), extended Kalman filter (EKF), and sigma point Kalman filter (SPKF) with deep neural networks, including multi-layer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM). The aim is to improve the reliability of forecasting and prediction tasks, particularly when processing IMU data. This study conducts a comprehensive performance comparison between state estimators integration with deep learning models, evaluating their effectiveness in addressing the challenges of estimation and prediction. The preliminary results show that the feature forecasting rate of the proposed method can reach a root mean square error (RMSE) value of 0.31 (EKF-LSTM) and 1.50 (SPKF-LSTM).