Back to search

University of North Dakota

Deep Learning For Lithium-Ion Battery Status Monitoring: Robust State-Of-Charge And Trustworthy State-Of-Health Estimation

Abstract

dc:description.abstract

In electric vehicle and energy storage applications, the reliable operation of safety-critical battery management systems (BMS) relies on accurate, real-time estimation of the state of charge (SOC) and state of health (SOH). Although data-driven methods show great potential, their practical application is primarily constrained by two major challenges. First, battery operation under real-world conditions involves continuously varying factors—such as temperature, aging stages, and dynamic cell characteristics—making it impossible to exhaustively cover all scenarios during laboratory data collection. This discrepancy often leads to a significant decline in estimation performance due to the domain shift problem. Second, engineering practice requires a balance between high estimation accuracy and computational efficiency, as highly complex models are difficult to deploy in a lightweight manner on resource-constrained microcontrollers (MCUs). To address these challenges, this dissertation presents five independent studies proposing a series of lightweight, physics-informed deep learning frameworks specifically designed for SOC or SOH estimation, particularly overcoming the notoriously difficult flat voltage plateaus of Lithium Iron Phosphate (LFP) batteries. In the research on SOC estimation, we first investigated domain shift issues, demonstrating the feasibility of domain adaptation algorithms in this context. Building on this, we proposed a novel dual-stream architecture integrating convolutional neural networks (CNNs) and recurrent neural networks (RNNs), enabling accurate SOC estimation with lightweight models across multiple temperature ranges. Regarding SOH estimation, we also targeted key pain points. Initially, we employed an architecture integrating the LIME framework with an LSTM model, acting as an Explainable AI (XAI) solution. This approach not only accurately predicts SOH but also generates heatmaps to visualize the influence of various factors, enhancing the transparency of the black-box model. Subsequently, we introduced a comprehensive model for charging-based estimation tasks. This model incorporates a thorough modal analysis of the constant-current (CC) and constant-voltage (CV) phases and employs an attention-based soft-gating mechanism to adaptively fuse multimodal features while suppressing noise interference in cross-battery applications. Ultimately, this research establishes a practical bridge between advanced artificial intelligence algorithms and real-world electrical engineering applications. Extensively validated on public cross-cell datasets and dynamic driving cycles, the proposed frameworks strictly bound the estimation errors within safety margins. By optimizing for edge computing and maintaining an extremely low parameter count, this dissertation offers a reliable, efficient, and lightweight technical solution for next-generation battery state estimation systems.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy (PhD)
Level thesis:degree_level
Dissertation
Discipline thesis:degree_discipline
Electrical Engineering

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Lei, Mingwei
Contributors dc:contributor
  • Jielun Zhang

Subjects

dc:subject × 1

Identifiers

dc:identifier.*
Repository record dc:identifier
https://commons.und.edu/theses/13891

Chain of custody

source
Harvested from
University of North Dakota
Base URL
commons.und.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Lei, Mingwei. Deep Learning For Lithium-Ion Battery Status Monitoring: Robust State-Of-Charge And Trustworthy State-Of-Health Estimation. Dissertation thesis, https://commons.und.edu/theses/13891