Machine learning pipeline for battery state of health estimation

Darius Roman; Saurabh Saxena; Valentin Robu; Michael Pecht; David Flynn

Source-linked AI summary

Machine learning pipeline for battery state of health estimation

Darius Roman, Saurabh Saxena, Valentin Robu, Michael Pecht, David Flynn

arXiv:2102.00837v1Feb 1, 2021cs.LG

TL;DR

Reliable real-time battery SOH estimation must handle varying degradation conditions while quantifying prediction uncertainty. This paper develops and evaluates a machine-learning pipeline using charge-curve segments and four algorithms, finding that the best algorithm depends on charging protocol and that calibrated uncertainty can trade off against error.

Problem
Real-time battery SOH estimation must account for diverse degradation paths and provide uncertainty information for safe battery management.
Method
BHUMP engineers features from charge-curve segments, selects features automatically, augments training data, and estimates SOH uncertainty with four calibrated machine-learning algorithms.
Results
The lowest error depends on charging protocol, while the deep neural-network ensemble generally provides better uncertainty calibration than random forest at higher error.
Takeaways & Limitations
Machine-learning pipelines using charge-curve segments can provide computationally efficient cell SOH estimation, but algorithms should undergo uncertainty-quantification checks before deployment.
Takeaways & Limitations
Temperature variations during charging can introduce uncertainty into charge curves and propagate it into SOH estimates.

Abstract

from arXiv · show

Lithium-ion batteries are ubiquitous in modern day applications ranging from portable electronics to electric vehicles. Irrespective of the application, reliable real-time estimation of battery state of health (SOH) by on-board computers is crucial to the safe operation of the battery, ultimately safeguarding asset integrity. In this paper, we design and evaluate a machine learning pipeline for estimation of battery capacity fade - a metric of battery health - on 179 cells cycled under various conditions. The pipeline estimates battery SOH with an associated confidence interval by using two parametric and two non-parametric algorithms. Using segments of charge voltage and current curves, the pipeline engineers 30 features, performs automatic feature selection and calibrates the algorithms. When deployed on cells operated under the fast-charging protocol, the best model achieves a root mean squared percent error of 0.45\%. This work provides insights into the design of scalable data-driven models for battery SOH estimation, emphasising the value of confidence bounds around the prediction. The pipeline methodology combines experimental data with machine learning modelling and can be generalized to other critical components that require real-time estimation of SOH.

1 Introduction

The paper addresses real-time battery SOH estimation under diverse ageing conditions by focusing on capacity fade as a health indicator. It proposes BHUMP, a scalable machine-learning pipeline that estimates SOH together with calibrated uncertainty across varied cells, chemistries, designs, and operating protocols.

Motivation: Li-ion battery degradation depends on dynamic operating conditions, while real-time degradation estimation must remain reliable across varied failure mechanisms.Ageing causes capacity loss and impedance increase under varying charge/discharge rates, voltage limits, and temperatures.
Motivation: Capacity is used as the SOH indicator because it correlates with energy-storage capability and directly affects battery runtime and life.SOH compares a battery’s capacity or impedance with its unused state.
Contribution: Existing machine-learning SOH studies provide limited or no assessment of prediction uncertainty, motivating a pipeline that estimates SOH with associated algorithm uncertainty.The proposed battery health and uncertainty management pipeline is called BHUMP.
Contribution: BHUMP processes incoming data streams hierarchically, beginning with feature engineering from segments of raw charge curves.The pipeline is designed for real-time estimation and uncertainty quantification.
Evaluation: The pipeline is evaluated on 179 cells spanning three designs, two chemistries, three charge protocols, and various discharge rates.The designs are prismatic, pouch, and cylindrical; the chemistries are LiFePO4 and LiCoO2.
Contribution: BHUMP combines improved capacity-fade estimation with metrics for accurately quantifying uncertainty and is intended to adapt across cell chemistries and operating conditions.The paper characterizes the approach as scalable and more accurate than conventional methods as batteries age.

2 Machine learning pipeline approach

BHUMP estimates battery SOH as capacity fade through a hierarchical pipeline that engineers and selects charge-curve features, augments training data, and fits four regression algorithms. It also quantifies predictive uncertainty and evaluates calibration and decision-oriented accuracy.

Feature engineering and selection: The pipeline creates 30 features and automatically selects relevant subsets using random-forest recursive feature elimination with cross-validation.Feature engineering uses domain knowledge and selected charge-curve segments; RF-RFE-CV avoids user-defined maximum-feature thresholds.
Model fitting: The augmented dataset trains four algorithms: random forest, deep neural network ensemble, Bayesian ridge regression, and Gaussian process regression.RF and dNNe are frequentist models, whereas BRR and GPR are Bayesian approaches; modified ensemble methods provide uncertainty estimates for frequentist models.
Pipeline target: Battery SOH is quantified as capacity fade relative to the first cycle’s measured capacity.Ci denotes capacity at cycle i, while C1 is capacity measured during the first complete charge-discharge cycle.
Data augmentation: Training data are augmented with FGSM adversarial examples combined with ridge-regression weight decay, using perturbations bounded at 1% of each feature’s range.The augmented examples are concatenated with the original training data, and the ridge-FGSM combination reduced confidence intervals around estimated means.
Uncertainty quantification: Probabilistic forecasts estimate a Gaussian distribution N(µ, σ^2), with µ as predicted capacity and σ^2 as associated uncertainty variance.Reliability diagnostic curves assess how predicted confidence intervals align with empirical coverage.
Uncertainty evaluation: Decision-oriented uncertainty assessment defines an accuracy zone of ±1.5% around true capacity and computes α-accuracy from predictions within that zone.The threshold α can be adjusted according to the intended application.

3 Dataset

The dataset comprises 179 lithium-ion cells divided into three groups by charging protocol. This separation supports distinct model training, feature selection, and performance assessment.

Dataset composition: 179 Li-ion cells are grouped by charging protocol into Group I (47 cells, CC-CV), Group II (124 cells, CC), and Group III (8 cells, 2-step fast charge).The groups are defined by constant current–constant voltage, constant current, and 2-step fast-charge protocols, respectively.
Dataset composition: Charging-protocol grouping enables separate model training, feature selection, and model performance assessment.

4 Algorithm performance

Algorithm performance varies by dataset: RF or BRR often delivers the strongest accuracy, while dNNe more consistently provides reliable calibration, including under fast-charging and dynamic-cycling conditions.

Group I: 18 of 30 engineered features were selected for Group I, with voltage thresholds Vh = 4.2V and Vl = 3.9V.RF-RFE-CV selected the optimum feature count for Group I.
Group I: 54.70% average calibration error was achieved by RF in Group I, but its uncertainty estimates were considered insufficiently conservative.The calibration error was attributed possibly to the Infinitesimal Jackknife variance estimate.
Group I: 1.13% in MAPE and 1.56% in RMSPE were sacrificed by dNNe relative to RF to obtain better-calibrated predictions in Group I.dNNe was, on average, less than 4% under the ideal calibration score.
Group II: Less than 2% RMPSE was achieved by all four models on Group II, while RF had the highest accuracy but over-confidence in calibration.Group II varies in charge profiles but holds discharge conditions at 4 C-rate, highlighting charging-protocol effects.
Group II: 79.72% Cscore was recorded for RF in Group II, whereas dNNe provided more reliable calibration at an accuracy cost.RF was best from an accuracy and sharpness perspective, while GPR performed worst overall.

5 Discussion on practical applicability of BHUMP

BHUMP is presented as a practical complement or alternative to conventional battery-management estimation methods, using charge-profile sections to estimate capacity fade under partial-discharge and variable operating conditions. Temperature variation remains a source of uncertainty, motivating temperature-aware training or additional sensory inputs.

Practical applicability: BHUMP can complement battery management systems for SOC and SOH estimation and potentially replace traditional equivalent-circuit models.The passage contrasts BHUMP with conventional approaches that measure capacity under static full charge-discharge conditions.
Practical applicability: BHUMP estimates capacity fade from sections of the charge profile, accommodating partial discharge and random or high discharge-rate conditions.This enables estimation beyond static full charge-discharge measurements.
Practical applicability: Temperature variations during charging can introduce uncertainty into charge-curve measurements and propagate into SOH estimation.Suggested mitigation includes using temperature during training or adding in-situ or operando sensory information.

6 Conclusion

The conclusion presents BHUMP as a machine-learning alternative to conventional online SOH models, combining computational efficiency with accurate estimation and uncertainty quantification. Results show that algorithm performance depends on charging protocol, while charge-curve segments can capture degradation in under 15 minutes.

6 Conclusion: BHUMP is introduced as a machine-learning alternative to equivalent-circuit and electrochemical models for online battery SOH estimation.The pipeline is motivated by the difficulty of balancing accuracy and computational efficiency in conventional online methods.
6 Conclusion: The pipeline evaluates Bayesian ridge regression, Gaussian process regression, random forest, and deep ensembles for error and uncertainty quantification.All four algorithms serve as base algorithms for BHUMP and are assessed on both prediction error and uncertainty estimation.
6 Conclusion: Random forest achieves the lowest error for constant current–constant voltage and fast-charge protocols, while the best algorithm depends on charging protocol.The passage also reports Bayesian ridge regression as lowest-error for another protocol, but its protocol label is truncated in the supplied text.
6 Conclusion: Under 15 minutes, charge-curve segments can capture battery degradation, supporting computationally efficient and accurate cell SOH estimation.The conclusion emphasizes that uncertainty quantification checks remain necessary before field deployment, even when prediction errors are low.

Supplementary material

The supplementary material details how charge-curve segments and limited discharge information are transformed into normalized, protocol-compatible features for battery SOH estimation. It also documents the mathematical feature-engineering methods used to represent curve statistics, distances, and entropy.

Feature extraction: Charge-curve segments reduce diagnostic time to as little as 15 min and enable SOH estimation under partial discharge conditions.The approach addresses limited availability of complete charge curves in real-world applications.
Feature engineering: Discharge data contribute one feature, pseudo linear resistance, calculated as the observed voltage drop divided by the applied load current.This feature captures the instantaneous voltage drop associated with internal battery impedance when load current is applied.
Feature extraction: Recorded charge curves are normalized over a defined voltage interval, enabling batteries with different types and designs to share a training dataset under the same charging protocol.The interval is bounded by charge cut-off voltage and a lower voltage defined using a predefined voltage range.
Feature engineering: Engineered features include pattern statistics, distances from reference lines using Directed Hausdorff and Frechet measures, and entropy of charge-curve representations.The supplementary material describes signal mean, kurtosis, skewness, curve distances, and curve entropy as feature families.