Machine Learning-Based framework for battery Remaining Useful Life Prediction

Abstract

Lithium-ion batteries find great use in critical power requirement in electric vehicles, renewable energy storage systems, aerospace and aviation, medical devices and substation DC systems. These systems require reliable operation and higher safety considerations thereby preventing catastrophic failures. Real time monitoring of critical battery parameters such as capacity, voltage, current and temperature is paramount for predictive maintenance. Several Machine Learning based techniques such as Decision Tree Regression,Random Forest, Support Vector Regression, Gaussian Process Regression and Long-Term Short Memory can be used to predict the Remaining useful life (RUL) of batteries. In this study, three machine-learning models are considered. These are the Random Forest (RF), which represents the ensemble methods class of machine learning. Support Vector Regression (SVR), representing the classical regression models class and the Long-Term Short Memory (LSTM) representing the deep learning/sequence models class. The selected models are on the basis of being good representative of each class of Machine – Learning models. The methodology used include downloading and loading in MATLAB the online NASA data set. Exploratory data analysis in MATLAB, preparing the Data for Machine Learning, Implementing the three Machine Learning Models, Comparing the Models and making Remaining Useful Life (RUL) predictions. The performance parameters such as the Root Mean Square Error (RMSE) and the Statistical Correlation Coefficient ? ? are analysed to find the Model performance in predicting RUL. The LSTM performed better than Random Forest in accuracy and long term prediction. The technique is complex and slower to train. However, the SVR model performs better with hyper- parameter tuning. The study contributes to the increasing body of Machine Learning techniques in predictive maintenance. Routine checking done practically usually requires work force. Predictive maintenance, allow for real time monitoring according to the model developed and corrective action taken. The study provides techniques for predictive maintenance, for batteries and other system with measurable raw operational data. Further research may be required on the integration of different Machine Learning based techniques in predicting the RUL. This improves the prediction accuracy, robustness and adaptability.