Diagnostic-free onboard battery health assessment

TL;DR

This work tackles the challenge of onboard battery health assessment without disruptive diagnostic cycles by formulating an interpretable physics-constrained encoder–decoder. By mapping partial charging data to latent electrode-state variables $(C_p, C_n, x_0, y_0)$ and enforcing a differential voltage analysis–based mechanistic constraint, the model reconstructs full $OCV$ trajectories, enabling both current health diagnosis and future prognosis without historical degradation data. Validation across three datasets (422 cells total) shows robust performance with $MAE$ in key quantities around 45.1 mAh for capacity and 10.1 mV for voltage, and SOH errors typically below 7%, while providing physically meaningful aging modes such as loss of lithium inventory ($LLI$) and loss of active material on cathode/anode ($LAM_p$, $LAM_n$). The framework transfers across varying C-rates and dynamic operating conditions via fine-tuning, enabling rapid onboard health management and predictive maintenance without offline testing or extensive history.

Abstract

Diverse usage patterns induce complex and variable aging behaviors in lithium-ion batteries, complicating accurate health diagnosis and prognosis. Separate diagnostic cycles are often used to untangle the battery's current state of health from prior complex aging patterns. However, these same diagnostic cycles alter the battery's degradation trajectory, are time-intensive, and cannot be practically performed in onboard applications. In this work, we leverage portions of operational measurements in combination with an interpretable machine learning model to enable rapid, onboard battery health diagnostics and prognostics without offline diagnostic testing and the requirement of historical data. We integrate mechanistic constraints within an encoder-decoder architecture to extract electrode states in a physically interpretable latent space and enable improved reconstruction of the degradation path. The health diagnosis model framework can be flexibly applied across diverse application interests with slight fine-tuning. We demonstrate the versatility of this model framework by applying it to three battery-cycling datasets consisting of 422 cells under different operating conditions, highlighting the utility of an interpretable diagnostic-free, onboard battery diagnosis and prognosis model.

Figures (6)

• Figure 1: Data illustration.a) Protocol for battery aging in the van Vlijmen et al. dataset, which includes the cycling and diagnostic test. The cycling test uses different C-rates and SOC windows for the charging and discharging, while the diagnostic cycle includes a reset test, HPPC test, and RPT tests with three different C-rates. b) Summary of the testing conditions and the degradation curves of the test batteries. c) Capacity distributions obtained by different discharge C-rates and the variations during aging van2023interpretable. d) Protocol for the different C-rates test generated in this work, where 8 different rate tests are included. e) Capacities distributions of the cells at different discharge C-rates. f) Protocol design for battery aging in the Geslin dataset geslin2024dynamic, which includes the cycling and diagnostic test. g) Cycling design and the degradation curves for the 92 cells in the Geslin dataset geslin2024dynamic.
• Figure 2: Data sampling and diagnostic model.a) Randomly sampled portions of charging curves. b) Distributions of the voltage windows for the input data sampling of the cycling dataset. c) The SOC ranges of the extracted partial charging data. d) The mapping between the measured capacity from the partial charging and the reference C/5 RPT capacity. e) Mechanistic constraint model for battery health diagnosis using randomly sampled portions of charging curves. f--g) The errors of the $\hat{Q}$ sequence and $\hat{V}$ sequence of the predicted OCV curves for all the test samples (composed of all the RPT cycles from the test cells), respectively. h) The errors of the derived voltage of the OCV curves, which are calculated based on the physics-constrained mechanistic states in the machine learning model. The predicted discharge capacity ($\hat{Q}$) and voltage ($\hat{V}$) MAE errors were 45.1 mAh and 10.1 mV, respectively, with derived voltages ($V^{\text{d}}$) showing MAE error below 18.4 mV.
• Figure 3: Health diagnosis.a) Illustration of the predicted OCV curves, derived OCV curves based on the constrained mechanistic states, and the fitted OCV based on the states obtained by the particle swamp optimization for the predicted OCV curves. b) The differential voltage curves for the corresponding curves in a. c) The predicted curves of the mechanistic states obtained from cycling cycles and the comparisons from those obtained from diagnostic cycles. d--f) The predicted aging modes, which are directly obtained based on the constrained mechanistic states, for onboard diagnosis of LLI, LAM at the cathode (LAM$_\text{p}$), and LAM at the anode (LAM$_\text{n}$). g) The SHAP analysis results for the encoder, which represents the impacts (absolute values) of the input information from the partial charging curve on the prediction of mechanistic states. h) The SHAP analysis of the decoder for the OCV curve predictions, where the impacts of the constrained mechanistic states on the predicted OCV (combination impacts of the capacity and voltage) are illustrated.
• Figure 4: Future health prognosis.a) Prediction of future capacity degradation curves. b) Sequence prediction of the cycle life from the last value of the predicted future degradation curve. c) Point prediction of the cycle life based on the decoder. d--e) The feature impacts analysis of the future degradation predictions based on SHAP for the predicted EFC and capacity values, respectively. f) The mean feature impacts on the predicted future EFC variations. g) The feature impacts on the last prediction step, i.e., the cycle life, of the predicted EFCs. h) The mean feature impacts on the predicted capacities.
• Figure 5: Rate test validation study.a) SOH distribution (C/5 discharge) of all the cells. b) Discharge curves with different C-rates. c) Predicted capacity for different C-rate validations using measurement from 1C discharge within 20 min. d) Measured OCV and predicted OCV curves using partial 1 C discharge curves under C/80 at four different SOH levels. e) dV/dQ curves derived from the OCV curves in d. f) Prediction distributions for the discharge capacities using different C-rates and the corresponding constrained mechanistic states.