Cross-Domain Transfer Learning for Reliable Condition Monitoring of Primary Batteries Under Discharge-Only Operation

Abstract

Primary lithium-based coin-cell batteries are widely used in embedded systems, low-power sensors, and Internet of Things devices due to their long operational life and maintenance-free characteristics. In these applications, accurate estimation of the remaining battery life is essential to ensure system reliability. However, conventional methods for estimating battery health rely on repeated charge and discharge cycles, which are not applicable to primary batteries. This study presents a deep learning-based transfer learning framework to infer the condition of primary coincell batteries using discharge-only data. A new dataset is introduced, consisting of voltage versus time profiles obtained under constant current discharge from batteries manufactured by four different brands. The prediction target, referred to as the discharge progression indicator (DPI), is formally defined as the normalised elapsed discharge time and shown to be mathematically equivalent to Depth of Discharge under constant-current operation, making it fully observable from a single discharge event without requiring cycling data or explicit capacity measurements. Neural network models are initially pretrained on large-scale lithium-ion datasets and then adapted to the new dataset through a partial transfer strategy with input layer re-initialisation, enabling crosschemistry knowledge transfer across fundamentally different battery chemistries. A range of model architectures is evaluated, including fully connected networks, convolutional networks, recurrent memory-based models, and attention mechanisms. The results demonstrate that temporal models, particularly those with memory structures, achieve superior predictive performance and robustness against domain-induced variability. A sensitivity analysis further confirms that standard 8- bit or 10-bit analogue-to-digital converters are sufficient for reliable DPI prediction, supporting deployment in resource-constrained embedded systems. The proposed framework enables early and accurate condition estimation in the absence of charging data or domain-specific calibration.