Causal inference-based matching networks time-series prediction method under small sample data: battery remaining useful life prediction as a case study

Abstract

Data-driven machine learning models face challenges in accurately predicting remaining useful life in small-sample time-series datasets, primarily due to data scarcity and information leakage. Addressing these limitations, this study presents causal inference-based matching network, a novel framework synergizing causal inference with matching networks. The methodology initially employs causal convolution to strictly enforce unidirectional temporal flow, thereby preventing data leakage. Subsequently, global causal features (CFs) are extracted to model intrinsic dependencies and integrated into a stacked cross-attention module, which sharpens the model’s focus on critical drivers. To further enhance robustness, a soft dynamic time warping augmented transformer constructs an adaptive matching network for precise attention reweighting. Empirical validation in small-sample scenarios on NASA and Oxford battery datasets demonstrates superior performance: the method maintains a root mean square error (RMSE) below 0.02 and error within 6% for NASA cells, while consistently achieving a mean absolute error under 2.092 on Oxford data. Ablation experiments confirm that global CFs are pivotal, reducing RMSE by 35.2%. Finally, sensitivity analysis and controlled leakage isolation experiment show that the model requires a minimum of 30% of lifecycle data to effectively capture complex non-linear degradation features.