Robust Health Monitoring for Lithium-Ion Batteries Under Guidance of Proxy Labels: A Deep Multitask Learning Approach

State of health (SOH) estimation is foundational for effective health prognosis and management of lithium-ion batteries. However, conventional data-driven approaches struggle to ensure robust generalizability when training data lacks sufficient SOH labels. In this work, we explore the potential utility of ageing information, gathered from routine operational data beyond the standard input collection region, as proxy labels (PLs) to establish nonlinear mappings between a short charging segment and SOH. To achieve this, we develop a multitask learning (MTL) framework for battery health monitoring, which comprises two critical components: 1) a deep MTL model, and 2) a multivariate Gaussian process regression (GPR) model. The deep MTL model is designed with four distinct tasks, each with a specific goal to exploit data under different labeling conditions, facilitating the extraction of domain-invariant and target-specific features for PL estimation. Using the estimated PLs, the GPR model is employed to provide precise mappings to SOH. Validations are conducted on batteries with diverse chemistries and cycling temperatures, particularly under severe label scarcity below 5%. The results confirm the effectiveness of the proposed method, outperforming baseline methods and achieving a mean absolute error of less than 1.3% in all case studies.