Multi-time scale feature extraction for early prediction of battery RUL and knee point using a hybrid deep learning approach

Accurate remaining useful life (RUL) prediction is crucial for the reliable and safe operation of lithium-ion batteries. However, the nonlinear degradation of lithium-ion batteries and the variability in characteristics across different batches make it highly challenging to predict the RUL using limited early cycle data. To address this issue, we propose a multi-time scale feature extraction method and a hybrid deep learning method. Specifically, we extract health indicators (HIs) both across cycles and within each cycle from the first 100 cycles, employing a sliding window strategy to maximize the utilization of aging information. A library of 445 features is generated using this method as input for feature selection which eventually produced a subset of the 320 most impactful features for model input. Subsequently, the processed features are fed into a hybrid model based on multi-head attention mechanisms and a multi-layer perceptron (MLP), which can capture aging information across different time scales. This provides a more comprehensive insights into short-term and long-term trends, allowing for accurate RUL prediction. Additionally, a snapshot ensemble learning strategy is introduced to further enhance the model's generalization ability without increasing any additional training cost. We use a total of 123 batteries to validate our method. The mean absolute percentage errors (MAPE) on the primary test set and the secondary test set are 7.77% and 9.82%, representing improvements of 9.0% and 13.2% compared to the benchmark. This study highlights the promise of feature engineering and deep learning networks for early prediction of battery RUL.