Data-driven reverse identification of spindle structures using machine learning approaches: a proof-of-concept study

This study presents a data-driven framework for reverse identification of spindle structures using frequency response functions (FRFs) and machine learning (ML). A comprehensive database was generated through Receptance Coupling Substructure Analysis (RCSA) simulations to map spindle geometrical and contact parameters to their dynamic responses. Random Forest (RF) and Neural Network (NN) models were trained to infer these parameters inversely from FRF data. The methodology was validated through two case studies: one purely simulation-based and another experimentally implemented on a simplified spindle prototype. The results demonstrated that ML can recover spindle parameters with reasonable accuracy—achieving coefficients of determination (R²) up to 0.99 for simulated data and up to 0.87 for experimental data—while reproducing the measured FRFs with strong agreement. The findings reveal that contact and bearing stiffness parameters exert the most significant influence on the FRF, whereas geometric parameters are less sensitive and more difficult to infer. Despite inherent non-uniqueness in inverse modeling, the developed approach proved robust to experimental noise and parameter uncertainty. The study establishes the foundation for extending the proposed framework toward real machine tool spindles, focusing on fault detection and predictive maintenance.