UAV MONITORING SYSTEM WITH MACHINE LEARNING BASED PREDICTIVE MAINTENANCE AND DIGITAL TWIN METHODS

• Main Author: Syauqi Buldan, Rasis
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/50392
• Institution: Institut Teknologi Bandung

Market value of UAV products and researches has been rapidly increasing in the past 5 years and is expected to grow by 300% in 2026. But it is yet to found that for every accident involving a piloted plane, 100 more accidents involving a UAV have happened, where 50% of the accidents are caused by machine failures. Its propulsion system is the biggest contributing factor of the machine failures, with motor as the system’s main component. UAV motor failures that are only known after the failures have occurred will increase the asset’s downtime and are very disadvantageous in terms of time and money. Thus, it is very important for the UAV user to be informed of the condition of the UAV motors and its predicted time of failure. This final project aims to develop a UAV monitoring system that is able to estimate the condition of UAV motors and predict the time of failure of the motors, with the implementation of predictive maintenance and digital twin methods that are powered by machine learning. Modification of the UAV is done with the installation of vibration sensors on the motors as the parameter that will be monitored, and a microcontroller to stream the vibration data, Vibration data is then processed with vibration signal analysis to produce statistical features that represent the motor’s health degradation. The statistical features will then be transformed with PCA and EWM to a one-dimensional health indicator (HI) of the motor. UAV motor condition (normal/abnormal) estimation model and HI prediction model is developed using LSTM-NN because of its ability to learn time series data. The model will also generate Remaining Useful Llife (RUL) of the UAV motors as the implementation of predictive maintenance method. A UAV motor vibration estimation model is also developed as the implementation of digital twin, and its potential prospect to develop the monitoring system that will not require installing an additional sensor and microcontroller on the UAV. A UAV monitoring system with three main computation models and a web-based interface was built based on acquired vibration and PWM data of Parrot AR Drone 2.0 on hover mode. A total of 350.000 data points were then processed into time domain statistical features, where 20% of the data was separated as a validation data to test the performance of the models. UAV motor condition estimation model has an accuracy rate of 95.5% and an F1 score of 95.7%. UAV motor health indicator predictor model has an RMSE rate of 17% and an average error of 2.7 minute for the RUL prediction. A UAV motor vibration estimation model was also developed with LSTM, with PWM and vibration as input and output, respectively. The model has an average RMSE of 1,3192 for 5 time domain statistical features of UAV motor vibration.