LOCALIZATION SYSTEM ON AUTONOMOUS TRAM USING VIRTUAL SENSOR BASED ON LONG SHORT-TERM MEMORY AND UNSCENTED KALMAN FILTER AND DEVELOPMENT OF THE DIGITAL TWIN MODEL

• Main Author: Tania Mulyadi, Yolanda
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/65143
• Institution: Institut Teknologi Bandung

The recent development of autonomous vehicles has the potential to bring revolution in transportation sector. One of the vehicles that is currently being developed is the autonomous tram. These vehicles have been developed for reasons of being energy-friendly, low emission, and eliminating human factor that accounts for most accidents. One of the important keys in the development of the tram is the localization system. In this research, a further development of the localization system is proposed using a virtual sensor that is integrated with a controller system and a digital twin model. The proposed virtual sensor consists of a Long Short-Term Memory (LSTM) to replace the GNSS sensor in the case of mssing position information and an Unscented Kalman Filter (UKF) to estimate the tram position. The output of this virtual sensor is then used as one of the inputs to the controller which is built using the Stanley controller algorithm to keep the tram on the specified track. The entire autonomous tram system can be monitored and controlled via digital twin model that is connected in real time. The localization system using the proposed virtual sensor was simulated first through the CARLA Simulator. The results show that in the case of missing position information, the localization system using a virtual sensor with LSTM can improve the correction of position prediction by more than 97% compared to a virtual sensor system without LSTM. The simulated algorithm is then implemented on a tram prototype along with the digital twin model and controller. The digital twin model on the implementation of the tram prototype shows that the localization system using the proposed virtual sensor has succeeded in reducing the position estimation error in the localization system without LSTM by more than 54%. Meanwhile, in terms of control, the results show that the tram with the proposed system can still move on its path independently of the availability of position information. The experiment showed that the position error of the tram towards the target path was 10.83 cm and 13.49 cm using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics.