Physics-informed inverse identification and dynamic reconstruction of high-speed gear systems with asymmetric mesh stiffness

Gear transmission systems are critical components in engineering applications, yet their dynamic modeling under high-speed conditions remains challenging. This study develops a physics-informed learning framework for inverse identification and model diagnosis in nonlinear gear dynamics. Rather than using PINNs only as a parameter-estimation tool, the proposed framework is employed to examine the structural validity of the classical symmetric mesh-stiffness assumption. Benchmark tests first confirm that the PINN can accurately recover effective stiffness and damping parameters. When applied to ADAMS-generated transmission-error data, the identified symmetric stiffness successfully reconstructs the low-speed heavy-load response but exhibits a clear mismatch under the high-speed light-load condition, showing that parameter adjustment within the classical symmetric formulation is insufficient. Additional contact-state information extracted from ADAMS reveals that the low-speed response remains predominantly on a single contact branch, whereas the high-speed light-load response repeatedly switches between the drive-side and coast-side contact branches. This independent contact-state evidence, together with the mechanical nonequivalence of the two tooth flanks, motivates the manual formulation of a bilaterally asymmetric mesh-stiffness model. The proposed model provides a more consistent reconstruction of the high-speed light-load response and demonstrates that direction-dependent stiffness switching can provide an independent source of non-smoothness, including grazing-type transitions in the zero-backlash limit.