Modeling of a flexible slider-crank mechanism with unbalance: A neural network-based inverse dynamics approach for fault identification

This study presents a neural network-based inverse dynamics method for identifying unbalance faults in flexible slider-crank mechanisms. Slider-crank mechanisms are central in reciprocating machinery, and their dynamics can be altered by unbalance and link flexibility. We develop a Lagrangian model of a flexible slider-crank with an extensible spring-damper connecting link and a parametric unbalance mass on a triangular plate, then train a feedforward neural network to recover five coupled physical parameters (unbalance mass, angular position, eccentricity, period time, and damping ratio) simultaneously from a single 51-point displacement trace. On clean synthetic data, the network achieves R >0.97 for all five parameters with sub-millisecond inference. A spatial error analysis maps prediction accuracy to unbalance position on the link, revealing that errors concentrate near the crank and slider joints where kinematic sensitivity is lowest. A noise robustness study shows that the clean-data network fails entirely under measurement noise; retraining with noise-augmented data (30–60 dB SNR) restores useful accuracy for four parameters at realistic noise levels, while unbalance mass remains vulnerable due to its weak coupling to slider displacement. These results establish quantitative bounds on the method’s applicability and identify supplementary sensing requirements for experimental implementation.