Study on predicting double cracks coalescence behavior based on hyperparameter optimization of the kernel extreme learning machine classification model

Purpose The study aims to develop an intelligent prediction framework for dual-crack coalescence behavior and enhance the prediction accuracy and generalization performance of crack state judgment under multiple crack morphologies. Design/methodology/approach Four typical machine learning classifiers including LR, KNN, SVM and KELM are comprehensively compared on datasets of linear, elliptical and circular dual cracks. The Kernel Extreme Learning Machine (KELM) with the best overall performance is selected as the base model. A Reinforced Multi-strategy Parrot Optimization (RMPO) algorithm is employed to globally optimize KELM's hyperparameters, taking regularization coefficient C and RBF kernel parameter as optimization variables. The negative average validation-set accuracy from inner-fold cross-validation is adopted as the fitness function for adaptive searching within a nested stratified cross-validation workflow. Findings KELM outperforms other machine learning models in predicting three types of dual-crack coalescence. After hyperparameter tuning by RMPO, the hybrid RMPO-KELM model achieves prediction accuracies of 0.973, 0.911 and 0.824 for linear, elliptical and circular dual cracks respectively. Compared to grid search (GS) and the Parrot Optimization Algorithm (PO), the optimized model achieves a more accurate nonlinear fit between crack geometric inputs and coalescence labels, and maintains reliable prediction accuracy on independent test subsets. Research limitations/implications This study only considers dual cracks in homogeneous plates under predefined loading conditions. The dataset is derived from numerical simulations, and further physical tests are required to validate the model. Future work can extend the model to heterogeneous materials, multiple cracks and complex load cases. The feature selection strategy may be further refined for broader engineering scenarios. Practical implications This proposed RMPO–KELM model enables fast identification of dual-crack coalescence states in homogeneous plates, avoiding heavy finite-element iterative computation. It offers an efficient auxiliary tool for structural safety assessment. The feature screening scheme also provides a reference for selecting dimensionless crack parameters. Engineers can adopt this framework for preliminary crack risk evaluation in similar plate structures. Social implications Accurate and rapid prediction of crack coalescence helps prevent sudden structural failures in infrastructure, reducing safety hazards and casualties. Reliable structural risk assessment extends service life of engineering components, cuts maintenance costs and lowers resource consumption. This predictive framework supports safer civil and mechanical infrastructure, benefiting public safety and sustainable asset management. Originality/value An RMPO–KELM hybrid intelligent prediction model is proposed for multi-morphology dual-crack coalescence prediction. The RMPO algorithm realizes efficient global searching of KELM hyperparameters, which significantly improves prediction precision and generalization capacity. This work supplies a data-driven solution for structural safety assessment involving crack propagation and coalescence.