Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling

Area of Science:

• Battery Engineering
• Machine Learning for Energy Systems
• Electrochemical Systems Analysis

Background:

• Accurate battery state-of-health (SOH) estimation is crucial for reliable operation, especially with variable charging protocols and unknown polarization history.
• Existing data-driven methods often lack robustness due to reliance on handcrafted indicators or fixed sampling conditions.
• Incomplete charging data presents a significant challenge for precise SOH assessment.

Purpose of the Study:

• To develop a robust transfer-learning framework for accurate SOH estimation using incomplete battery charging data.
• To leverage voltage prediction pre-training to capture latent polarization dynamics for improved SOH accuracy.
• To evaluate the framework's performance on a diverse dataset under varied charging conditions.

Main Methods:

• Developed a transfer-learning framework utilizing a long short-term memory (LSTM) network.
• Pre-trained the LSTM network on a voltage-prediction task to encode battery polarization dynamics.
• Fine-tuned the pre-trained LSTM for SOH estimation using the Severson fast-charging dataset.
• Evaluated performance on partial charging segments (20% capacity) at varying start positions.

Main Results:

• Transfer learning reduced SOH estimation Mean Absolute Error (MAE) from 1.75% to 0.91%.
• Root Mean Square Error (RMSE) for SOH estimation decreased from 2.35% to 1.30% compared to direct LSTM training.
• The framework demonstrated improved data efficiency and robustness under varied charging conditions.

Conclusions:

• Mechanism-informed pre-training via transfer learning significantly enhances SOH estimation accuracy and robustness.
• The proposed LSTM-based transfer learning approach effectively captures latent polarization dynamics from charging data.
• Further validation across diverse datasets and unified protocols is recommended for broader applicability.