Second-level heterogeneous retired battery type identification using pulse-test-enabled federated learning with output-level privacy preservation

The imminent retirement of large numbers of lithium-ion batteries creates an urgent need for efficient reuse and recycling to realize lifecycle economic and environmental benefits. In this context, reliable battery type information is essential for downstream sorting, reuse evaluation, and recycling process selection, because different cathode chemistries are associated with different material values, processing requirements, and safety considerations. However, such information is often unavailable in practice due to long-term label degradation and restricted access to sensitive battery data, especially under non-independent and identically distributed (non-IID) conditions across stakeholders. To address this challenge, we propose a privacy-preserving federated learning framework for retired battery type identification using only second-level pulse-test data collected locally by clients. The framework incorporates an expert-weighted aggregation mechanism, in which client-specific autoencoders quantify local expertise through reconstruction error, while only classification probabilities and expert indices are transmitted for collaboration. This output-level design reduces privacy exposure by avoiding the exchange of raw data, gradients, and model parameters. Experiments on a heterogeneous retired-battery dataset containing 8 battery types, 640 retired batteries, and 10,184 pulse-test records across LFP, LMO, NMC811, and NMC622 cells with nominal capacities ranging from 10 Ah to 68 Ah show that the proposed framework achieves an average classification accuracy of 96.3% across 100 stochastic non-IID partitioning scenarios. It consistently outperforms distributed aggregation baselines, including Average Aggregation, Class-Count Weighting, and Mahalanobis Distance Weighting, while remaining close to the centralized reference performance. By addressing label-distribution and data-quantity skew under a controlled non-IID setting, the proposed framework provides a methodological proof-of-concept for recovering chemistry-relevant battery type information from historically untraceable retired batteries, while external validation under independently collected datasets remains necessary for deployment-level assessment. • Second-level pulse data enable retired battery type identification. • An expert-weighted federated framework is proposed for non-IID settings. • Only output probabilities and expert indices are shared across clients. • The method preserves privacy without exchanging raw data or gradients. • It achieves 96.3% average accuracy across 100 stochastic scenarios.