Efficient Joint Identification Based on Neural Networks and Its Application in the Tool–Collet–Holder System

This study aims to develop an efficient and accurate method for identifying joint parameters in assembled structures. A novel neural network-based joint identification framework is proposed. Frequency response function (FRF) datasets are generated by combining finite element simulation with frequency-domain substructure synthesis. The Uniform Manifold Approximation and Projection (UMAP) algorithm is employed for nonlinear dimensionality reduction in FRF sequences, preserving critical characteristics. A multilayer perceptron (MLP) network is then trained to regress joint parameters from the reduced-dimension FRF data. The necessity of the nonlinear dimensionality reduction within this joint identification framework is verified through comparison with the linear dimensionality reduction technique of principal component analysis (PCA). This methodology is implemented and validated using a tool–collet–holder system. Comparative studies with the global optimization method reveal that the proposed approach maintains superior identification accuracy while achieving significant improvements in computational efficiency across varying preload conditions. Furthermore, the identified joint parameters exhibit strong predictive capability when tested under tool/holder component changes, preload variations, and when coupled with a spindle, proving robustness under complex operational scenarios. This study provides a new technical pathway for the joint identification of assembly structure.