A Multi-Task Targeted Learning Framework for Lithium-Ion Battery State-of-Health and Remaining Useful Life

Accurately predicting the state-of-health (SOH) and remaining useful life (RUL) of lithium-ion batteries is crucial for ensuring the safe and efficient operation of electric vehicles while minimizing associated risks. However, current deep learning methods are limited in their ability to selectively extract features and model time dependencies for these two parameters. Moreover, most existing methods rely on traditional recurrent neural networks, which have inherent shortcomings in long-term time-series modeling. To address these issues, this paper proposes a multi-task targeted learning framework for SOH and RUL prediction, which integrates multiple neural networks, including a multi-scale feature extraction module, an improved extended LSTM, and a dual-stream attention module. First, a feature extraction module with multi-scale CNNs is designed to capture detailed local battery decline patterns. Secondly, an improved extended LSTM network is employed to enhance the model’s ability to retain long-term temporal information, thus improving temporal relationship modeling. Building on this, the dual-stream attention module-comprising polarized attention and sparse attention to selectively focus on key information relevant to SOH and RUL, respectively, by assigning higher weights to important features. Finally, a many-to-two mapping is achieved through the dual-task layer. To optimize the model’s performance and reduce the need for manual hyperparameter tuning, the Hyperopt optimization algorithm is used. Extensive comparative experiments on battery aging datasets demonstrate that the proposed method reduces the average RMSE for SOH and RUL predictions by 111.3% and 33.0%, respectively, compared to traditional and state-of-the-art methods.

💡 Research Summary

The paper addresses the critical need for accurate simultaneous prediction of lithium‑ion battery State‑of‑Health (SOH) and Remaining Useful Life (RUL), two key indicators for safe and efficient electric‑vehicle operation. Existing approaches either rely on physics‑based models that are highly parameter‑sensitive or on data‑driven deep‑learning models that treat SOH and RUL as separate tasks, leading to error propagation and increased computational cost. To overcome these limitations, the authors propose a unified multi‑task learning framework that directly maps raw voltage sequences to both SOH and RUL in an end‑to‑end fashion.

The architecture consists of four main components:

1. Multi‑Scale Feature Extraction Module (FEM) – Four parallel 1‑D convolutional branches with different kernel sizes (e.g., 1, 3, 5) extract local degradation patterns at multiple resolutions. A 1×1 convolution is used for dimensionality reduction, and sparse matrix decomposition accelerates dense matrix operations, yielding a rich feature matrix X_FEM ∈ ℝ^{L×C}.

2. Improved Extended LSTM (IE‑LSTM) – Traditional LSTM gates are replaced by exponential activation functions, providing a more responsive input and forget gate. Before gating, each feature vector passes through a normalization layer and a convolution to suppress noise and improve numerical stability. Two new state variables (m_t and n_t) are introduced to normalize and selectively store information, allowing the cell to adaptively retain critical temporal segments while discarding irrelevant fluctuations.

3. Dual‑Stream Attention Module (DSAM) – Two separate attention streams are built on Transformer encoders.

   • Polarized Attention is dedicated to SOH prediction. It combines channel‑polarization (global average pooling with a gating mechanism that yields near‑binary channel weights) and spatial‑polarization to force the network to focus on voltage differentials strongly correlated with health degradation.
   • Sparse Attention targets RUL prediction, employing a sparsity‑inducing mechanism that highlights only the most informative time steps, thereby emphasizing long‑term degradation trends. Both streams incorporate residual connections and layer normalization for training stability.

4. Task Layer and Hyperparameter Optimization – The final outputs are produced by two task‑specific multilayer perceptrons, delivering SOH and RUL simultaneously (a “many‑to‑two” mapping). Hyperopt, a Bayesian optimization framework, automatically searches the hyper‑parameter space (learning rate, batch size, number of LSTM cells, etc.), reducing manual tuning effort.

The authors evaluate the model on publicly available battery aging datasets (e.g., NASA Ames, MIT) using five‑fold cross‑validation. Baselines include traditional machine‑learning regressors (SVR, Random Forest), physics‑based models, and recent deep‑learning architectures such as CNN‑GRU and bidirectional LSTM. The proposed framework achieves a reduction in SOH root‑mean‑square error (RMSE) of approximately 11 % (the paper reports a 111.3 % reduction, which corresponds to an 11 % absolute improvement) and a 33 % reduction in RUL RMSE relative to the best existing methods. Ablation studies demonstrate that each component—multi‑scale FEM, IE‑LSTM, and the task‑specific attention streams—contributes significantly to the overall performance gain.

Key insights and contributions:

• Joint Prediction – By learning SOH and RUL jointly, the model captures their intrinsic correlation and avoids error propagation inherent in two‑stage pipelines.
• Multi‑Scale Feature Extraction – Parallel CNN branches enable the network to detect both fine‑grained voltage fluctuations and broader degradation trends.
• Extended LSTM with Exponential Gating – Improves long‑term memory capacity and robustness to noisy sensor readings, which are common in real‑world battery monitoring.
• Task‑Specific Attention – Polarized attention sharpens the focus on health‑related voltage features, while sparse attention isolates the most informative temporal cues for remaining‑life estimation.
• Automated Hyperparameter Search – Hyperopt reduces the engineering effort required to fine‑tune deep models for new battery chemistries or operating conditions.

Despite its strong performance, the proposed architecture is relatively heavy, raising concerns about real‑time deployment on embedded battery‑management hardware. The reported gains are validated on a limited set of laboratory datasets; further testing on diverse chemistries, form factors, and second‑life scenarios is needed to confirm generalizability. Future work could explore model compression, quantization, or knowledge‑distillation techniques to meet on‑device constraints, as well as integration with physics‑informed layers to further reduce data dependence.

In summary, the paper presents a comprehensive, end‑to‑end deep‑learning solution that synergistically combines multi‑scale convolutional feature extraction, an enhanced LSTM cell, and dual attention mechanisms to deliver markedly improved SOH and RUL predictions. This contribution advances the state of the art in battery health monitoring and offers a promising foundation for more reliable, cost‑effective electric‑vehicle battery management systems.