Though not sure if this or a similar approach is leveraged, but I always wondered how Apple provided an estimate of battery life in its devices.

Accidentally stumbled upon this paper (https://lnkd.in/gDJeWgNv) while looking for an LSTM-CNN combination for a totally different type of use case. But if you analyze the approach, this could be extrapolated for RUL prediction in digital twin setups as well.

Traditional physical (mechanistic) models and many data-driven models each have limitations: the physical ones struggle with complex real-world conditions; many data-driven ones don’t always capture multi-scale features or generalise well.

This research proposes a new deep-learning architecture to improve the accuracy and generalization of battery RUL prediction.

The researchers design a multi-scale feature fusion convolutional neural network with jump (skip)-connections (termed MSFFCNN in the paper). The key aspects of the proposed architecture are:

Multi-scale extraction: They use convolution kernels of different sizes (e.g., 1×3, 1×2, 1×4) to capture local and global degradation features of the battery health indicators at different scales.

Jump/skip connections: Inspired by residual networks, they fuse features from multiple convolutional layers (so that not only deep global features but also lower-level local features contribute).

Improved loss function: They propose a modified mean square error (MSE) loss that weights the error more heavily when the predicted remaining life is longer than the true remaining life (i.e., predicting too optimistically is penalized more) because overestimating RUL is riskier than underestimating (which leads to early replacement).

Health-factor selection: They identify indirect health indicators for the battery (for example, times to certain voltage/current transitions, number of cycles) that strongly correlate (Spearman’s >0.9) with capacity degradation.

Two benchmark datasets were used for the training: the NASA battery dataset and the CALCE/University of Maryland (CACLC) battery dataset.