A machine learning-based battery management system for state-of-charge prediction and state-of-health estimation for unmanned aerial vehicles

Abstract

Unmanned aerial vehicles (UAVs) are becoming more popular as they start to be utilized in many applications, such as surveillance, agriculture, and military. However, due to their limited battery capacity, a proper battery management system (BMS) is required to avoid flight delays and crashes, which can be highly expensive in terms of cost and time. SoC prediction and SoH estimation can help ensure the arrival of UAVs to their destination and increase the lifetime and efficiency of the battery. A UAV BMS is based on machine learning (ML), where the ML models predict the SoC and estimate the SoH based on the voltage and current of the battery, and ambient temperature. Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) are utilized for SoC prediction as a regression problem, and Random Forest (RF) was utilized for SoH estimation through a classification problem with four classes. The results verified the reliability of the ML models due to their high accuracy. The estimation of SoC using the DNN model had a low mean squared error of 7.6E−4 and a high explained variance score of 0.98. In addition, the prediction of SoC using the LSTM model had a low mean squared error of 0.023 and a high explained variance score of 0.97. Moreover, the RF model achieved a high accuracy of 0.92 at classifying SoH. Regarding the practical implementation, the system was deployed through the utilization of a drone, ESP32 microcontrollers, a Raspberry Pi gateway, and a cloud server, which proved the reliability and effectiveness of the ML-based BMS.