Identification of friction in joints for multibody dynamics: a Neural-DAE based approach

Accurate prediction of multibody system dynamics hinges on precise modeling of joint friction moments. Conventional approaches rely on predefined analytical friction models with calibrated parameters, limiting flexibility and introducing selection bias. This work proposes a Neural Differential-Algebraic Equation (Neural-DAE) framework to directly identify friction moments in revolute joints from experimental data. By embedding physics-informed artificial neural networks (ANNs) within the multibody dynamics equations, the method eliminates reliance on ad hoc friction models. The ANN architecture incorporates three key innovations: input transformations using Lagrange multipliers and relative joint velocities, a decoupled structure separating normal forces and friction coefficients, and odd symmetry constraints to enforce antisymmetric friction-velocity relationships. A discrete adjoint method compatible with the generalized-α integrator enables efficient gradient computation, significantly reducing training time and memory costs compared to auto-differentiation. Experimental validation on planar and spatial double pendulums demonstrates that the Neural-DAE robustly identifies friction moments from noisy measurements by preserving underlying physics, generalizes to diverse initial configurations with consistent prediction accuracy, and outperforms baseline ANN architectures lacking physics-informed constraints in convergence speed. The framework’s integration of data-driven learning with mechanistic principles offers a paradigm shift for joint parameter identification in complex mechanical systems.