Physics-informed acoustic tribo-informatics operator learning for condition-indexed abrasive belt wear monitoring under varying grinding conditions

Continuous monitoring of abrasive belt wear is essential for maintaining surface integrity, material removal stability, and process reliability in grinding. Existing data-driven monitoring methods usually estimate wear through pointwise mappings from sensor features to discrete wear labels, which limits their ability to describe the cumulative and condition-dependent wear trajectory. Compounding this challenge, changes in abrasive-belt specifications and grinding parameters can alter the belt–workpiece contact state, acoustic emission response, and wear accumulation pathway. To address this issue, this study proposes a physics-informed acoustic tribo-informatics operator learning framework for abrasive belt wear monitoring with non-negative wear-rate regularization. Acoustic emission is regarded as an observable representation of the evolving acoustic tribo-state, while abrasive belt wear is formulated as a condition-indexed function-valued response over time. A conditional Deep Operator Network is developed, in which the branch network encodes acoustic emission history and a factorized trunk network represents the trajectory coordinate and grinding conditions. To improve physical consistency, a non-negative wear-rate network is coupled with the trajectory operator through a residual constraint inspired by Preston-type wear evolution. This design regularizes the predicted cumulative wear trajectory through a learned non-negative wear-rate field, providing an interpretable wear-rate representation without imposing a fixed empirical formulation. Experiments on a robotic grinding platform validate the proposed framework under varying grinding conditions. The results show that burst morphology and spectral band-energy features provide effective acoustic tribo-state representations. Within-belt tests achieve NRMSE/range values below 4% on both A100 belts, while leave-one-belt-out tests obtain mean NRMSE/range values of 8.95 ± 0.48%, 6.87 ± 0.62%, and 7.79 ± 0.67% on A45, A100–1, and A100–2, respectively. Few-shot adaptation reduces belt-specific magnitude offsets, with A100–1 reduced to 1.31 ± 0.31% using six leading calibration passes. The learned wear-rate field captures the transition from initial high-rate wear to quasi-steady wear, providing an interpretable rate representation for knowledge-intensive tool-maintenance decisions. To support reproducibility and further research, the data is publicly available at GitHub.