Improving Inertial Sensor by Reducing Errors using Deep Learning Methodology

Abstract

Autonomous cars are the future of automotive industry. Autonomous or driverless cars must be able to rapidly and accurately sense and analyze their immediate environment to take appropriate maneuvering/navigation actions under various driving conditions. To undertake such challenging navigation task, autonomous vehicles make use of a multitude of sensor systems and networks including (i) Global Positioning System (GPS), (ii) Inertial Navigation System (INS), (iii) Inter-vehicle and cellular wireless networks, (iv) Video cameras, and (v) Light Detection And Ranging (LiDAR) system [1–3]. Sensor fusion algorithms are applied to fuse the data from the various devices, which ultimately enables the vehicle to take accurate navigation decisions. The INS system takes a dominant role in the vehicular sensor fusion algorithm in GPS-denied environments such as dense urban places, multi-level parking structures, and areas with thick tree-coverage. The INS unit incorporates dead-reckoning principle via mechanization equations to process the linear acceleration and angular velocity data. Here, Micro Electro Mechanical Systems (MEMS) based Inertial Measurement Unit (IMU) sensors are preferred due to their low cost and resistance to shock and vibration. However, MEMS inertial sensors are prone to various errors [4–6]. Thus, development of calibration and compensation techniques for errors reduction from sensors, both systematic and stochastic/random, are essential. In this paper, we describe for the first time a novel deep learning-based methodology to simultaneously remove many errors sources in the sensor signals, under laboratory environment. By correctly identifying the classification signal, we developed a Convolutional Neural Network (CNN) algorithm that inherently removes the sensor error sources. To test the efficiency of our algorithm, the results are compared with traditional approaches like Six-Position Static Test and Rate Test. Here, we achieved an accuracy of 80 % in correctly identifying the accelerometer and gyroscope signals.