Yaw/Heading optimization by Machine learning model based on MEMS magnetometer under harsh conditions

The paper's main goal is to accomplish a high accuracy of yaw/heading by Machine Learning approach when the motion range of vehicle/device calibration is limited. The nonlinear Random Forest (RF) Regression with proper training has a high potential to deal with the magnetometer uncertainty before calibration and during iron distortion cases. The proposed solution solely requires the magnetometer without other sensor's support. A Pan Tilt Unit-C46 (PTU-C46) with high precise positioning was used as a reference heading value to label the corresponding magnetic features in the learning model. The proposed approach helps yaw estimation to be carried out under harsh conditions, which resolve many difficulties in orientation tracking since the magnetometer is susceptible to hard iron and soft iron in the environment. In addition, many mechanical devices work only within the specific range and waste their dynamic motion around two axes or more just for calibration. Thus, the research focuses on the level rotation calibration around Z-axis within the restricted range of motion for practical application. The experiment was carried out using a low-cost platform equipped with Micro-Electro-Mechanical System (MEMS) sensors as gyroscope, accelerometer, and magnetometer. The 9 Degree of Freedom (DoF) Madgwick was implemented into the Microcontroller to compare with the proposed model. The sensor fusion can track the yaw value after the level calibration despite various error conduction. The RF model accomplishes a superior result with more stability and more minor error. Under iron disturbance or calibration absence, the ML model still maintains the good tracking command with maximum Mean Square Error of about 0.3°, while the Madgwick is unsuccessful in heading measurement due to huge error in these circumstances.