Session: 02-01 Intelligent Modeling and Predictive Analytics for Energy Systems

Paper Number: 173932

173932 - Lithium-Ion Battery State of Charge Monitoring Using Low Frequency Stress Waves and Machine Learning Methods 

Lithium-ion batteries (LIB) play increasingly critical roles in the oil and gas industry, particularly as electrification replaces hydraulic, pneumatic, and engine power systems, and they require appropriate management of operating conditions to maintain safety, ensure reliability, and achieve long lifespan. This requires monitoring battery state of charge (SoC) to prevent overcharge or over-discharge, which drastically reduce immediate and long-term reliability and safety. Accurate SoC measurements (within correct charge limits) are also critical to ensuring the reliability of LIB-powered systems.

    Although SoC cannot be measured directly, estimation methods based on voltage and current measurements exist, however their accuracy suffers when batteries are in use and under dynamic loads. Deficiencies of existing estimation systems have caused injuries, deaths, property damage, and electronic waste. Recently new estimation methods based on battery stress wave response have emerged, and while high frequency ultrasound can produce predictions with high accuracy, equipment costs and complexity favor low-frequency stress wave (LFSW) approaches. To compensate for the significant complexity of material property shifts, their relationship with electrochemical battery dynamics, and the interpretation of mixed response signals, recent research applies machine learning (ML) to enhance estimation accuracy.

    LFSW with ML approaches are evaluated, by subjecting an LIB to  Gaussian-modulated sine pulse excitations and collecting stress wave response signals while charging/discharging the LIB  with a battery tester to control SoC. The Mel-Frequency Cepstral Coefficients (MFCC) were extracted from the response signals and used to train Support Vector Machine (SVM) and Back-Propagation Neural Network (BPNN) models, and compared with Multi-Rocket model performance for SoC prediction. 

    While SVM and BPNN models predicted SoC well, with RMS errors of 11%, 8% respectively (10% error is common in automotive BMS), the Multi-Rocket model excelled with a prediction error of only 0.72%. Although MFCC-derived feature datasets are usable for SoC classification, the MFCC feature extraction method requires additional optimization and/or augmentation with other features for high-precision SoC regression. Comparing these shallow- and deep-learning model performances with Multi-Rocket suggests that enhancing the accuracy of LFSW with ML SoC predictions will require further enhancing feature extraction and selection processes, as well as a larger training dataset.

Presenting Author: Thomas Hannan University of Houston

Presenting Author Biography: Thomas "Tico" Hannan is a PhD mechanical engineering student researcher at the University of Houston developing low frequency stress wave based real-time battery monitoring methods, sensing and controls for shape memory alloy based actuators, and developing decommissioned wind turbine blade structural reuse applications, all enabled and enhanced with advances in efficient machine learning. The stress wave based battery monitoring breakthroughs have won grand prizes in the National Academy of Inventors - Genspiration Foundation as well as UH Energy - Chevron Innovation Challenge competitions in 2024.

Prior to beginning PhD studies, he worked in wired/wireless/fiberoptic telecommunication network engineering and Unix/Linux server infrastructure security engineering in the ISP/datacenter industry, and contributes more than two decades of signals, controls, and data management experience to Dr Gangbing Song's Smart Materials and Structures Laboratory research group at the University of Houston.