Data-driven Predictive Maintenance for Li-Ion Batteries in Autonomous Systems using Machine Learning
Pysyvä osoite
Kuvaus
Opinnäytetyö kokotekstinä PDF-muodossa.
Autonomous Mobile Robots (AMRs) in warehouse logistics, manufacturing, and industrial automation applications are powered by Li-Ion batteries. The smooth operation of this system relies on battery performance. Traditional maintenance approaches, such as run-to-failure and fixed-schedule preventive maintenance, are ineffective at addressing the progressive, non-linear capacity loss observed in Li-Ion batteries. A data-driven predictive maintenance approach that accurately predicts batteries' State of Health and Remaining Useful Life in real time provides a more efficient and environmentally friendly solution. In this thesis, such a framework is developed and tested using publicly available data from the NASA Prognostics Centre of Excellence (PCoE) battery dataset. Two machine learning models were trained. A Random Forest regression model was used as interpretable baseline and a two-layer Long Short-Term Memory (LSTM) neural network model to account for the time-dependent nature of battery degradation data. Both models predicted SoH instead of RUL directly. RUL was then calculated from the predicted SoH by determining when the predicted SoH is projected to reach 80% end-of-life threshold. The performance of the two models on the test battery B0018 shows that the LSTM model significantly outperformed the Random forest on all four metrics (RMSE, MAE, MAPE, and R²). The result translates to 78% improvements in RMSE, and demonstrates that the sequential modelling of discharge cycle history can delivery an accuracy gain over feature-based ensemble models.
The feature importance analysis highlight the discharge capacity as the most significant predictor, with a relative importance of 75.4% in line with the electromechanical knowledge that capacity fade is the main microscopic effect of lithium loss through solid electrolyte inter phase growth. This thesis result shows that a simple LSTM model architecture can be trained to achieve near perfect RUL prediction accuracy on a previously unseen battery with a very simple computer setup and Python. These findings set a practical, repeatable and academically rigorous approach to data-driven predictive battery maintenance for autonomous system, which will benefit the sustainability objectives of avoiding unplanned downtime, increasing lifespan, and reducing e-waste in industrial AMR fleets.