Abstract
Accurate prediction of failure in industrial machinery and engines is critical for minimizing unexpected downtimes and enabling cost-effective maintenance. Existing predictive models often struggle to generalize across diverse datasets and require extensive hyperparameter tuning, while conventional optimization methods are prone to local optima, limiting predictive performance. To address these limitations, this study proposes a hybrid optimization framework combining Harris Hawks Optimization (HHO) and Wild Horse Optimization (WHO) to fine-tune the hyperparameters of ResNet, Bi-LSTM, Bi-GRU, CNN, DNN, VAE, and Transformer-GRU models. The framework leverages HHO's global exploration and WHO's local exploitation to overcome local optima and optimize predictive performance. Following hybrid optimization, the Transformer-GRU model consistently outperformed all other models across four benchmark datasets, including time-to-failure (TTF), intelligent maintenance system (IMS), C-MAPSS FD001, and FD003. On the TTF dataset, mean absolute error (MAE) decreased from 0.72 to 0.15, and root mean square error (RMSE) from 1.31 to 0.23. On the IMS dataset, MAE decreased from 0.04 to 0.01, and RMSE from 0.06 to 0.02. On C-MAPSS FD001, MAE decreased from 11.45 to 9.97, RMSE from 16.02 to 13.56, and score from 410.1 to 254.3. On C-MAPSS FD003, MAE decreased from 11.28 to 9.98, RMSE from 15.33 to 14.57, and score from 352.3 to 320.8. These results confirm that the hybrid HHO-WHO optimized Transformer-GRU framework significantly improves prediction performance, robustness, stability, and generalization, providing a reliable solution for predictive maintenance.