This companion ICICIS 2023 paper explores how Differentiable Neural Computer (DNC) memory modules extend CNN and CNN-LSTM predictors for lithium-ion battery remaining useful life estimation.
Approach
- Augments convolutional feature extractors with DNC memory to retain long-horizon degradation patterns.
- Investigates stand-alone CNN-DNC and hybrid CNN-LSTM-DNC configurations.
- Training pipeline emphasises stability to prevent catastrophic forgetting when new cycling regimes are introduced.
Findings
- Memory-augmented models improved long-term prediction accuracy compared with baselines lacking DNC components.
- CNN-LSTM-DNC achieved the lowest mean absolute percentage error, especially in late-life trajectories.
- Results suggest structured external memory can complement sequence models for prognostics.