Work

Beyond cycle-aware segmentation: Spectral–temporal relational learning for irregular battery RUL prediction

Year: 2026
Type: article
Abstract: Remaining useful life (RUL) prediction for lithium-ion batteries is difficult when the available measurements are irregularly sampled, organized by operating cycles, and coupled across sensor variables in ways that change during degradation. Many irregular time-series methods rely on fixed-window patching and mainly model temporal correlations, which can split charge–discharge cycles and miss frequency-domain information. This paper proposes the Cycle-Aware Spectral–Temporal Graph Neural Network (CAST-GNN), which uses each operational cycle as a dual-domain modeling unit for battery prognostics. Raw multivariate records are first partitioned into cycle-aware patches using consecutive discharge-onset events. For each sensor variable within a cycle, CAST-GNN builds a raw irregular temporal view and a resampled spectral view. The two views are encoded by a Transformable Time-aware Convolution Network and by a fast Fourier transform followed by a multilayer perceptron (FFT–MLP), respectively. The resulting variable-wise embeddings are contextualized across cycles by a Transformer. Inter-variable dependencies are then modeled through a Dynamic Time-Aware Graph and a smoothed Spectral Coherence Graph, fused by a cycle-conditioned gate, and aggregated by hierarchical attention pooling for battery-level RUL regression. Experiments on the NASA, CALCE, and UNIBO benchmarks show lower prediction errors than representative baselines, and an additional experiment on commercial cells illustrates the model behavior under heterogeneous operating conditions.
Language: English
Open Access status: hybrid
Estimated APC paid: $4,730
OpenAlex ID: W7237356002
FWCI:
Cites: 56
Cited by:
Related to: