Abstract
Safety accidents involving Lithium-ion (Li-ion) batteries have increased in frequency, highlighting the critical need for accurate Remaining Useful Life (RUL) prediction and health monitoring. This paper introduces a deep learning framework, HA-FOSELM, which integrates Forgetting Online Sequential Extreme Learning Machine (FOS-ELM) with a Hybrid Grey Wolf Optimizer (HGWO) and an attention mechanism. By employing Variational Mode Decomposition (VMD) for denoising and a forgetting mechanism for real-time adaptation, the proposed method achieves superior accuracy and efficiency compared to traditional neural network approaches.
Introduction
Li-ion batteries are essential components in modern technology, ranging from electric vehicles to aerospace applications. However, monitoring their health is challenging due to complex chemical degradation processes. Predicting the Remaining Useful Life (RUL)—the number of cycles remaining before a battery's capacity falls below a safety threshold—is vital for preventing catastrophic failures. Existing prognostic methods often struggle with signal noise and the inability to update models incrementally as new data becomes available.
Methodology
The proposed HA-FOSELM framework addresses these challenges through a multi-stage pipeline:
- Denoising with VMD: Raw capacity data is often "messy" due to electrochemical instability. Variational Mode Decomposition (VMD) is used to extract the underlying degradation trend while filtering out high-frequency noise.
- The Learner (FOS-ELM): The core model is an Extreme Learning Machine (ELM), known for its high training speed. It is enhanced to be Online Sequential, allowing it to learn from data batches incrementally. A Forgetting mechanism is included to discard outdated data that no longer represents the battery's current state.
- Optimization with HGWO: To optimize the model's hyperparameters, a Hybrid Grey Wolf Optimizer is used. By combining GWO with Differential Evolution (DE), the algorithm avoids local optima and ensures stable parameter tuning.
- Attention Mechanism: A "key-value-predict" attention layer is implemented to help the model focus on critical features of the capacity curve, improving regression performance on time-series data.
Experimental Setup and Datasets
The study utilizes the NASA Ames Prognostics Center of Excellence (PCoE) public battery dataset, focusing on batteries B0005, B0006, B0007, and B0018. These batteries were cycled at room temperature until reaching a 30% capacity fade (the end-of-life threshold).
Results and Discussion
The HA-FOSELM model was benchmarked against standard Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and Relevance Vector Machines (RVM).
- Accuracy: For battery B0005, HA-FOSELM achieved a Root Mean Squared Error (RMSE) of 0.0049, significantly lower than LSTM (0.0061) and RNN (0.0165).
- RUL Prediction: The error in predicting the exact failure cycle ($E_{RUL}$) was as low as 0 to -1 cycles, demonstrating exceptional precision compared to traditional models which were off by up to 48 cycles.
- Efficiency: By avoiding iterative backpropagation, the ELM-based approach is computationally efficient, making it suitable for real-time monitoring in edge computing environments.
Conclusion
The HA-FOSELM framework provides a robust solution for Li-ion battery health management. By combining denoising, online learning, and attention-based feature weighting, it offers a high-precision, adaptive tool for RUL prediction. Future work will focus on testing the model under varying environmental temperatures and dynamic load profiles to better simulate real-world conditions.
Summary updated