Battery SoH Estimation Using Cell-Level Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing battery health estimation methods for electric vehicle battery packs rely on the modularity assumption, which fails to accurately account for cell-to-cell variations due to manufacturing variances and aging, leading to inaccurate predictions of battery health and lifetime.
Innovation Solution
A data-driven battery management system using graph neural networks that models individual battery cells and modules within the pack, capturing intra-cell, intra-module, and inter-cell interactions to generate accurate and timely state-of-health estimates without relying on empirical knowledge or modularity assumptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If domain-knowledge driven physical models with modularity assumption are used for battery health estimation, then the device complexity is reduced and ease of manufacture is improved, but the measurement precision and reliability of battery health estimation deteriorate due to cell-to-cell variations
Solution Approach 1:
The patent segments the battery pack into individual cell-level units, each with its own neural network model. Instead of treating the battery pack as a modular system where single-cell models can be extrapolated, it creates independent models for each cell that capture individual aging patterns and variations, thereby maintaining manufacturing simplicity while improving estimation accuracy.
Solution Approach 2:
The patent applies local quality by allowing each battery cell to have its own specialized neural network model with cell-specific parameters and features. This enables the system to account for local variations in aging mechanisms, manufacturing defects, and operating conditions for each individual cell, improving overall measurement precision while maintaining a standardized overall architecture.
2Measurement precision
If individual cell-level models are created for each battery cell without modularity assumption, then the measurement precision and reliability of battery health estimation are improved, but the device complexity increases
Solution Approach 1:
The patent implements universality by using a standardized neural network architecture that can be applied to every battery cell in the pack. The same model structure, input features, and training methodology are universally applied across all cells, which simplifies the overall system design and deployment while still capturing individual cell characteristics through cell-specific parameter learning.
Solution Approach 2:
The patent uses copying by replicating the same neural network model structure across all battery cells. Instead of creating fundamentally different models for each cell, it copies a proven effective architecture and trains it with cell-specific data, thereby reducing development complexity while maintaining individualized accuracy for each cell.
Data Source
AI summary
Approaches, techniques, and mechanisms are disclosed for assessing battery states of batteries. According to one embodiment, raw sensor data are collected from a battery module. The battery module includes multiple battery cells. Input battery features are extracted from the raw sensor data collected from the battery module. The input battery features are used to update node states of a GNN. The GNN include multiple GNN nodes each of which representing a respective battery cell in the multiple battery cells. Estimation of one or more battery state of health (SoH) indicators is generated based at least in part on individual output states of individual GNN nodes in the multiple GNN nodes. The individual output states of individual GNN nodes in the multiple GNN nodes are determined based at least in part on the updated node states of the GNN.