Apply and optimize machine learning algorithms for estimating battery health
With the growing demand for energy-efficient and reliable battery-powered systems, accurate estimation of battery State of Charge (SOC), State of Health (SOH), and Remaining Useful Life (RUL) must be monitored to ensure safety, performance, and longevity. Traditional estimation techniques such as Co...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Year Project / Dissertation / Thesis |
| Published: |
2025
|
| Subjects: | |
| Online Access: | http://eprints.utar.edu.my/7093/1/fyp_CS_2025_CWY.pdf http://eprints.utar.edu.my/7093/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the growing demand for energy-efficient and reliable battery-powered systems, accurate estimation of battery State of Charge (SOC), State of Health (SOH), and Remaining Useful Life (RUL) must be monitored to ensure safety, performance, and longevity. Traditional estimation techniques such as Coulomb counting and model-based approaches often suffer from error accumulation, calibration complexity, and poor adaptability to dynamic conditions.
This project investigates machine learning (ML) techniques for estimating the SOC and RUL based on Electrochemical Impedance Spectroscopy (EIS) data. A range of regression and classification models including Random Forest (RF), Support Vector Machines (SVM), Gaussian Process Regression (GPR), and Artificial Neural Networks (ANN) were evaluated on both full-frequency and single-frequency EIS inputs. Results show that full-spectrum EIS features provide superior predictive performance, with Random Forest excelling in regression tasks and ANN achieving the highest classification accuracy. For RUL estimation, ANN and CNN-SAM models demonstrated competitive accuracy compared to baseline Gaussian Process Regression, effectively capturing degradation patterns across different operating temperatures.
To enable deployment on resource-constrained embedded systems, pruning and quantization techniques were employed to compress model size while preserving predictive accuracy. Optimization reduced ANN size from 260 kB to 26 kB and CNN-SAM from 1679 kB to 158 kB, confirming that lightweight yet robust models can be achieved without significant performance loss.
The findings confirm the potential of integrating EIS data with optimized ML models for real-time battery state estimation. This work provides a pathway toward practical, efficient, and intelligent BMS capable of supporting the growing adoption of lithium-ion batteries in diverse applications. |
|---|