Accurate estimation of lithium-ion battery (LIB) state of health (SOH) is crucial for safe and efficient battery management systems (BMS). LIBs are experiencing phenomenal growth in applications like electric vehicles, but their degradation presents challenges. SOH, typically defined as the ratio of present to initial capacity, is difficult to measure directly, requiring full charge/discharge cycles. Existing methods rely on lifelong battery degradation data with measured SOH labels, making data collection time-consuming and resource-intensive. These methods often involve feature engineering based on our understanding of battery degradation mechanisms. Deep neural networks (DNNs) offer an advantage by automatically extracting features from raw operating data, but they still require the same laborious data collection process. Transfer learning is a promising approach to reduce this burden, but it typically still requires a significant portion of labeled target data. This research aims to develop a deep learning framework capable of estimating battery SOH without any target-labeled data, significantly accelerating the development of BMS for new battery technologies.

Numerous studies have focused on estimating battery SOH by extracting features correlated with SOH degradation and mapping them to the SOH. These traditional methods heavily rely on the availability of target-labeled data which is costly and time-consuming to obtain. Various features have been explored including electrical, electrochemical, acoustic, mechanical, and thermal characteristics. The introduction of deep neural networks (DNNs) has improved the performance of SOH estimation by automating feature extraction from raw data. However, these DNN-based methods still necessitate the collection of target-labeled data, limiting their widespread applicability. Transfer learning techniques have been explored to alleviate the data collection burden, yet still require a substantial amount of target-labeled data. This research tackles the challenge of estimating battery SOH without the need for additional target-labeled data, addressing a major limitation of existing methods.

This study proposes a novel deep learning framework for SOH estimation that does not require any target labels. The framework utilizes a swarm of DNNs, each initialized differently but with identical hyperparameters. The training procedure involves simultaneously minimizing the SOH estimation loss of source domain samples and the gap between source and target domain feature vectors. This simultaneous optimization enables domain adaptation, transferring knowledge from the source domain to the target domain without relying on target labels. Partial charging curves are used as input, gridded with a voltage interval of 10mV to reduce data burden. One-dimensional (1D) convolutional neural networks (CNNs) extract features, followed by fully connected (FC) layers for regression. A key component is the inclusion of a middle fully connected (MFC) layer specifically for target domain feature vector reconstruction, helping quantify the domain gap and provide target domain pre-estimates. The estimation procedure involves selecting a subset of the trained DNNs based on their performance metrics (mean and standard deviation of estimations) using quartile thresholds. Finally, a trimming strategy is employed to balance the SOH distribution in the source domain, enhancing estimation performance. The methodology involves training numerous DNNs, which is detailed in the provided equation (9) describing the composite loss function. The selection of effective DNNs after the training is based on the criteria presented in equation (10). The evaluation metrics used are given in equation (11). The data preprocessing includes normalizing the partial charging curves using nominal capacity. This framework leverages the strengths of both domain adaptation and the swarm approach, enabling reliable estimation even without target labels. The trimming process is further defined by the equations (2), (3), and (4), which focus on minimizing the skewness of the source domain data distribution to ensure better generalization.

The proposed framework demonstrated high accuracy in cross-dataset SOH estimation without target labels. Across 80 validation cases encompassing 71,588 samples from 65 batteries with varying chemistries and manufacturers, the framework achieved an absolute error of less than 3% for 89.4% of samples and less than 5% for 98.9% of samples, with a maximum absolute error of less than 8.87%. The impact of the source domain SOH distribution was revealed, showcasing that trimming the source domain to achieve a more balanced distribution significantly improved estimation accuracy. Comparison with existing methods (GPR, RF, SVR, CNN) showed that the proposed framework outperformed them significantly in target label-agnostic scenarios, reducing the MAE and maximum absolute error by more than half. Ablation studies confirmed the crucial roles of both the swarm-driven and domain adaptation strategies. Analysis of the DNN swarm showed that the pre-estimation RMSEs of DNNs before selection were widely distributed, but after selection, the RMSE bandwidth was significantly reduced. The study also analyzed the relationship between the mean, standard deviation, and RMSE of the pre-estimates for each DNN, revealing the effectiveness of the proposed selection criteria. The visualization of explanation maps confirmed the effectiveness of the domain adaptation strategy in minimizing the domain gap between source and target domains. Further analysis showed that the swarm size of 50 is sufficient to provide an accurate estimation, and ReLU is the optimal activation function, while the number of channels in the CNN less than 128 and more than one layer is sufficient to provide an accurate estimation. The hyperparameter tuning of the proposed model is presented in Figure 7.

The results demonstrate the successful application of the proposed deep learning framework for SOH estimation in various battery types and chemistries without requiring additional degradation experiments. The framework's ability to achieve high accuracy in the absence of target labels significantly reduces the time and resource requirements of traditional SOH estimation methods. The superior performance compared to existing methods highlights the synergistic effects of the swarm-driven strategy and domain adaptation techniques. The visualization of explanation maps provides insights into the inner workings of the DNNs and confirms the effectiveness of the domain adaptation strategy in bridging the domain gap. This research offers a significant advancement in battery management system development, paving the way for rapid and efficient development of algorithms for new generation batteries by leveraging existing experimental data. This research overcomes a key limitation of existing methods by eliminating the need for extensive degradation testing for new battery types.

This study presents a novel deep learning framework for accurate lithium-ion battery SOH estimation without relying on target-labeled data from degradation experiments. The framework leverages a swarm of DNNs with domain adaptation and a trimming strategy to achieve high accuracy and robustness. The results demonstrate significant improvements over existing methods, paving the way for faster and more cost-effective BMS development for next-generation batteries. Future work could explore more sophisticated trimming strategies, extend the framework to other SOH metrics and input signals, and investigate applications to large-scale, unlabeled field data.

While the proposed framework demonstrates impressive results, there are limitations to consider. The performance is dependent on the quality and diversity of the source domain data. The trimming strategy, while effective, could be further refined to optimize performance. The generalization ability to completely unseen battery types needs further investigation. Although the framework shows remarkable accuracy, its applicability on large-scale field data which contains noise needs further investigation. Lastly, the interpretability of the model could be improved.