State-of-Health Estimation and Anomaly Detection in Li-Ion Batteries Based on a Novel Architecture with Machine Learning

Junghwan Lee, Huanli Sun, Yuxia Liu, Xue Li, Yixin Liu, Myung-Jun KimBatteries 2023 · DOI

TL;DR

This paper introduces a novel three-tiered architecture (ECU, edge, cloud) for Li-ion battery State-of-Health (SOH) estimation and anomaly detection. By combining a generalized DNN for broad applicability and a personalized DNN for vehicle-specific accuracy, the system achieves a remarkable 0.33% RMSE for personalized SOH estimation and uses Mahalanobis distance to identify abnormal degradation patterns.

What Problem Does This Paper Address?

Accurately estimating the State-of-Health (SOH) of Li-ion batteries (LIBs) is crucial for their safe and efficient operation, especially in demanding applications like electric vehicles. However, machine learning models, particularly Deep Neural Networks (DNNs), often struggle with significant errors due to variations across individual battery cells, modules, packs, and even different vehicles. Training models with massive datasets that capture all these variations is frequently impractical due to resource and time constraints, especially during the early stages of product development. This limitation can lead to suboptimal models that don't generalize well or fail to detect subtle, abnormal degradation patterns.

Methods

The core of this work is a novel three-layer architecture designed to tackle the challenges of LIB SOH estimation and anomaly detection:

  • Electronic Control Units (ECUs): Handle real-time data processing and safety-critical functions.
  • Edge Computers (DCU/CCU): Provide temporary data storage and run personalized machine learning algorithms. This is where the "personalized DNN" resides, trained on data from a single vehicle.
  • Cloud: Offers large-scale data storage, analysis, and hosts the "generalized DNN." This DNN is trained on data from multiple vehicles, aiming for broader generalizability.

The proposed system leverages two types of DNNs:

  1. Generalized DNN: Trained on a smaller dataset to ensure reasonable estimation accuracy during cross-validation, making it suitable for initial deployment before extensive online training. This DNN aims for generalizability across different battery conditions.
  2. Personalized DNN: Trained using data from a specific vehicle, allowing it to capture unique driving and storage conditions for higher accuracy.

An outlier detector, specifically using the Mahalanobis distance, is employed to compare the outputs of the generalized and personalized DNNs. This comparison helps identify deviations that might indicate abnormal degradation, going beyond simple error metrics. The Mahalanobis distance is chosen for its ability to account for covariance between variables, providing a more robust measure of dissimilarity than Euclidean distance.

Datasets

The primary dataset used in this study is the NASA Prognostics Center of Excellence battery dataset. This dataset contains charge and discharge cycle data for LIBs. The authors performed significant data cleansing, including removing voltage jumps after discharge termination and addressing abnormal voltages during charging. They also noted that the observed SOH regeneration in this dataset does not align with the physical behavior of LIBs, a phenomenon they could not fully investigate due to data limitations.

Key Insights and Ideas

  • The Power of Hybridization: Combining a generalized DNN (for broad applicability) with a personalized DNN (for specific accuracy) is crucial for robust SOH estimation in real-world scenarios. Neither approach alone is sufficient.
  • Generalizability is Key for Early Deployment: A generalized DNN, even when trained on limited data, must demonstrate good cross-validation performance to be useful before online training can refine it.
  • Mahalanobis Distance as a Subtle Anomaly Detector: This metric proves effective in detecting abnormal degradation by analyzing the differences between generalized and personalized SOH predictions, capturing subtle deviations that might be missed by simpler error metrics.
  • Data Cleansing is Non-Negotiable: The study highlights how crucial data cleaning is for accurate SOH estimation, particularly when dealing with datasets that may contain artifacts or unusual behaviors like SOH regeneration.

Key Results

  • The personalized DNN achieved a remarkably low Root Mean Square Error (RMSE) of 0.33%.
  • The generalized DNN achieved an RMSE of 4.6%.
  • The study evaluated 65 DNN models, exploring various hyperparameter configurations.
  • The Mahalanobis distance was successfully used to detect abnormal degradation patterns, with specific examples noted for LIB B5 and B7 at the 38th cycle.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Dataset Scope: The generalized DNNs were not trained or evaluated on various cell types or operation conditions, and the analysis did not account for variations at the module, pack, or vehicle level, as the datasets were from cell tests.
  • Distinguishing Causes of Differences: It was challenging to definitively attribute observed differences to cell variations or training issues of the generalized DNN, as no experiments were conducted specifically on cell variations.

Future work is planned to address these limitations:

  • Utilizing real-world, large-scale datasets from vehicles to train and evaluate generalized DNNs across diverse cell types and degradation conditions, while also considering module, pack, and vehicle variations.
  • Developing and prototyping an embedded controller for personalized DNNs and outlier detectors for use in electric vehicles.
  • Implementing cloud-based SOH labeling, as real-time SOH is difficult to measure directly in vehicles.
  • Training generalized and personalized DNNs using labeled data from multiple and individual vehicles, respectively, with the expectation of improved generalized DNN performance and a more robust outlier detector.

Summary updated

Abstract

Variations across cells, modules, packs, and vehicles can cause significant errors in the state estimation of LIBs using machine learning algorithms, especially when trained with small datasets. Training with large datasets that account for all variations is often impractical due to resource and time constraints at initial product release. To address this issue, we proposed a novel architecture that leverages electronic control units, edge computers, and the cloud to detect unrevealed variations and abnormal degradations in LIBs. The architecture comprised a generalized deep neural network (DNN) for generalizability, a personalized DNN for accuracy within a vehicle, and a detector. We emphasized that a generalized DNN trained with small datasets must show reasonable estimation accuracy during cross validation, which is critical for real applications before online training. We demonstrated the feasibility of the architecture by conducting experiments on 65 DNN models, where we found distinct hyperparameter configurations. The results showed that the personalized DNN achieves a root mean square error (RMSE) of 0.33%, while the generalized DNN achieves an RMSE of 4.6%. Finally, the Mahalanobis distance was used to consider the SOH differences between the generalized DNN and personalized DNN to detect abnormal degradations.
BibTeX
@inproceedings{lee2023state,
  author = {Junghwan Lee and Huanli Sun and Yuxia Liu and Xue Li and Yixin Liu and Myung-Jun Kim},
  title = {{State-of-Health Estimation and Anomaly Detection in Li-Ion Batteries Based on a Novel Architecture with Machine Learning}},
  booktitle = {Batteries},
  year = {2023},
  doi = {10.3390/batteries9050264},
}
Figure