Source-linked AI summary
Feature Analyses and Modelling of Lithium-ion Batteries Manufacturing based on Random Forest Classification
Kailong Liu, Xiaosong Hu, Huiyu Zhou, Lei Tong, W. Dhammika Widanage, James Marco
TL;DR
Battery manufacturing contains strongly coupled variables, but their importance and interactions are difficult to analyze reliably. The paper applies a random-forest framework using OOB predictions, Gini changes, and PMOA to classify electrode mass load and porosity while quantifying feature importance and correlations. The framework provides reliable classification and identifies interpretable manufacturing-feature effects, although the reported porosity classification remains weaker than mass-load classification.
Problem
Battery manufacturing variables are numerous and interdependent, while prior work often emphasizes prediction accuracy without systematically analyzing feature importance and interpretability.
Method
The study applies an RF framework to three mixing features and one coating parameter, using OOB predictions, Gini changes, and PMOA to analyze electrode-property classification, feature importance, and correlations.
Results
The framework classifies electrode mass load and porosity and quantifies the effects, importance, and correlations of the four manufacturing features.
Takeaways & Limitations
The RF framework offers a data-driven approach for feature analysis, sensitivity analysis, and possible model-dimension reduction in battery manufacturing.
Abstract
from arXiv · showhide
Lithium-ion battery manufacturing is a highly complicated process with strongly coupled feature interdependencies, a feasible solution that can analyse feature variables within manufacturing chain and achieve reliable classification is thus urgently needed. This article proposes a random forest (RF)-based classification framework, through using the out of bag (OOB) predictions, Gini changes as well as predictive measure of association (PMOA), for effectively quantifying the importance and correlations of battery manufacturing features and their effects on the classification of electrode properties. Battery manufacturing data containing three intermediate product features from the mixing stage and one product parameter from the coating stage are analysed by the designed RF framework to investigate their effects on both the battery electrode active material mass load and porosity. Illustrative results demonstrate that the proposed RF framework not only achieves the reliable classification of electrode properties but also leads to the effective quantification of both manufacturing feature importance and correlations. This is the first time to design a systematic RF framework for simultaneously quantifying battery production feature importance and correlations by three various quantitative indicators including the unbiased feature importance (FI), gain improvement FI and PMOA, paving a promising solution to reduce model dimension and conduct efficient sensitivity analysis of battery manufacturing.
I. INTRODUCTION
Battery manufacturing generates many coupled variables whose effects remain difficult to understand, while existing studies emphasize prediction more than interpretability. The paper therefore develops an interpretable RF framework to classify electrode properties and quantify feature importance and correlations.
- Battery manufacturing produces numerous coupled process variables that affect intermediate products and ultimately final battery performance.
- Existing battery-manufacturing studies largely focus on applying common prediction methods and model accuracy, with limited systematic analysis of interpretability.
- The proposed RF framework uses bagging and OOB predictions to classify five electrode-property classes while reducing overfitting.
- The framework quantifies feature importance and correlations for mixing and coating variables using three evaluation criteria.
- The manufacturing chain includes mixing, coating, calendering, cutting, cell assembly, formation, and ageing, with mass load and porosity treated as electrode properties.
- The study analyzes four features from mixing and coating stages and classifies mass load and porosity using five labels.
III. METHODOLOGY
The methodology builds a random forest from bootstrap-sampled, decorrelated decision trees and uses their out-of-bag predictions for evaluation. Majority voting produces classifications, while bagging, random feature selection, and OOB samples support generalization assessment.
- Random forests ensemble decision trees trained on different bootstrap samples to reduce classification variance through tree diversity.
- The RF predicts a new observation by selecting the class receiving the maximum number of decision-tree votes.
- Each tree selects a random subset of features for node splitting, producing decorrelated trees that can be grown without pruning.
- Out-of-bag samples, which are excluded from a tree’s bootstrap sample, provide an internal basis for evaluating classification performance without external test data.
- The methodology includes RF training, OOB prediction and generalization-error calculation, followed by classification-model construction and performance evaluation.
B. Feature Importance and correlation
The framework quantifies feature importance using OOB permutation effects and Gini impurity gains, while PMOA measures correlations through surrogate decision splits.
- Feature importance: Unbiased FI uses OOB predictions to estimate how classification error changes when a feature’s values are randomly permuted.Local importance is averaged within classes, then aggregated across observations.
- Feature importance: Gain improvement FI sums Gini impurity decreases from tree splits to measure each feature’s contribution to classification.A larger IG(xk) indicates greater impurity improvement for the target classification.
- Feature correlation: PMOA compares an optimal split on one feature with surrogate splits on another to quantify feature-pair correlations.The best surrogate split has the maximum PMOA value.
- Feature correlation: PMOA values greater than zero identify worthwhile surrogate splits, and larger PMOA values indicate more highly correlated feature pairs.PMOA values range within (−∞, 1].
C. Classification model structure and framework
The RF framework classifies electrode mass load or porosity from mixing and coating variables, then uses importance and correlation analyses to reduce the feature set and reconstruct the model.
- Model structure: The RF model uses AMMC, StoLR, viscosity, and coating CG as inputs, with labelled electrode mass load or porosity as the output.The framework targets intermediate-feature analysis alongside electrode-property classification.
- Model construction: RF hyperparameters are tuned by randomized search, balancing classification accuracy, generalization, computational effort, tree strength, and tree correlation.The tuned parameters are the number of decision trees and the number of features considered at each split.
- Feature analysis: The framework applies unbiased FI through OOB permutation and gain improvement FI through Gini changes caused by feature splits.These indicators quantify feature effects on mass-load and porosity classification.
- Feature analysis: PMOA values are plotted as an M × M heat map to analyse pairwise feature correlations, after which important features are selected for RF reconstruction.The reconstructed RF uses a reduced feature set for new classifications.
- Framework scope: The framework can classify electrode mass load and porosity while analysing feature importance and correlations, and can be extended when more manufacturing data become available.The stated extension supports correlation analysis, important-feature discovery, and model simplification.
D. Performance metrics
Classification performance is assessed with confusion matrices, per-class precision, recall, and F-measure, together with macro-averaged metrics and overall correct classification rate.
- Per-class metrics: Precision measures correct classifications for an interested class, while recall measures the rate of all corresponding conditions identified for that class.The measures are defined using true and false positive or negative outcomes.
- Per-class metrics: F-measure is the harmonic mean of precision and recall for each class.It combines the two per-class measures into one score.
- Overall metric: OCCrate is the proportion of correctly classified observations among all observations.It uses the total correctly classified outputs divided by the total number of observations.
- Confusion matrix: The confusion matrix places predicted classes in rows and actual classes in columns, with diagonal entries representing correct classifications.Its final row and column report class-level recall and precision, while the bottom-right entry reports OCCrate.
- Aggregate metrics: Macro-precision, macro-recall, and macro F1-score average the corresponding per-class measures across the five classes.The metrics evaluate overall performance in the battery-manufacturing multiclass problem.
IV. RESULTS AND DISCUSSIONS
The results section uses the RF framework to classify electrode mass load and porosity while quantifying feature importance, correlations, and their effects on classification.
- Results scope: The designed RF-based framework is applied to classify battery electrode mass load and porosity while quantifying feature importance and correlations.The analysis examines how manufacturing features affect classification of electrode properties.
A. RF classification model for battery mass load
The mass-load classification model uses four manufacturing features as RF inputs and quantifies their importance through two FI measures.
- Four features—AMMC, StoLR, viscosity, and CG—are used as RF inputs to classify labelled electrode mass load.
- Feature importance for mass load is evaluated with unbiased FI based on OOB and FI based on gain improvement.
1) Feature analyses:
CG is the dominant feature for mass-load classification, while AMMC and StoLR show the strongest feature correlation.
- CG is most important, with unbiased FI 4.78 and gain improvement FI 0.037, whereas viscosity contributes least with 0.67 and 0.022.
- StoLR and AMMC rank second and third in both FI measures, with unbiased FI values of 1.18 and 0.91.
- AMMC and StoLR have the largest mass-load feature correlation, with PMOA 0.72.
- The RF framework provides engineers with a way to interpret correlations among battery-manufacturing feature variables.
2) RF-based model:
Using all features, the RF model classifies mass load with a 90.2% OCCrate; a reduced three-feature set retains nearly comparable performance.
- 90.2% OCCrate is achieved for mass-load classification using all features, with 100% Prate for the ‘very high’ and ‘very low’ classes.
- Four cases test mass-load classification with different combinations of three features against the all-feature model.
- Case 1 using CG, AMMC, and StoLR achieves 86.6% macroP, 89.8% macroR, and 90.0% macroF1.
- Case 1 performs only 3.3%, 1.9%, and 0.1% below the all-feature case for macroP, macroR, and macroF1, respectively.
- Removing CG in Case 4 substantially reduces performance, indicating its importance for mass-load classification.
B. RF classification model for battery porosity
For porosity classification, StoLR and viscosity are the most contributing features, but the four-feature RF model achieves only 70.7% OCCrate and reduced-feature performance remains lower.
- Feature analyses: StoLR and viscosity are the most contributing features for porosity classification, while AMMC is the least contributing.
- Feature analyses: AMMC and StoLR have the highest porosity feature correlation, with PMOA 0.84.
- RF-based model: Using all features yields a 70.7% OCCrate for porosity classification, with several misclassified ‘high’ samples.
- Performance comparison: Among reduced-feature tests, Case 3 using StoLR, CG, and viscosity achieves the best results: 59.4% macroP, 60.8% macroR, 59.7% macroF1, and 68.3% OCCrate.
- Performance comparison: Replacing other features with AMMC reduces classification performance, and additional IPFs and PPs are recommended for improving porosity classification.
1) Hyper-parameters tuning:
The study tunes the RF classifier’s key hyper-parameters using randomized search and compares the resulting model with three established classification approaches. The framework classifies electrode mass load more effectively than porosity and identifies feature importance and correlations relevant to both tasks.
- Hyper-parameters tuning: Randomized search tunes RF hyper-parameters J and m using mean cross-validated accuracy for electrode mass-load and porosity classification.The search considers J values from 40 to 100 in steps of 20 and m candidates of 2 and 3.
- Benchmark comparison: Decision tree, k-nearest neighbors, and support vector machine are used as benchmark classifiers for evaluating the RF model.The benchmarks represent tree-based, instance-based, and kernel-based classification approaches.
- Implications and scope: The RF framework quantifies feature importance and associations for battery electrode-property classification and can be extended to additional manufacturing stages.The proposed extension includes mixing, drying, and calendering, although collecting the required data is time-consuming and equipment-dependent.
- Classification results: The four investigated features classify electrode mass load well, achieving a macroF1 of 90.1%.CG is the most important feature for mass-load classification, associated with a nearly 60.7% decrease when its contribution is removed.
- Classification results: Porosity classification achieves a macroF1 of 66.4%, indicating that additional feature variables should be considered.The discussion relates porosity to drying parameters such as rate, temperature, and pressure.