Comparative Assessment of Baseline LSTM and GA-Tuned LSTM for RUL Estimation in Diesel Engines with Synthetic OBD-II Data

Chavez Artica R.; Mendoza Acosta A. · Proceedings of the LACCEI international Multi-conference for Engineering, Education and Technology (2026) · DOI: 10.18687/laccei2026.1.1.1927

Resumen

Unplanned diesel-engine failures generate significant operational and economic impacts in heavy-transport systems. Predictive maintenance based on remaining useful life (RUL) estimation can mitigate downtime; however, the availability of labeled degradation datasets remains limited. This work presents a simulation-based and reproducible framework for generating multivariate OBD-II-like signals and evaluating RUL prediction using a Long Short-Term Memory (LSTM) network. A Genetic Algorithm (GA) is further employed for hyperparameter optimization, using validation performance as the fitness criterion. The system is evaluated as a regression problem using R2, RMSE, MAE, and relative RMSE under an asset-based split, where a complete unseen engine is reserved for testing. Results show that both models meet the predefined acceptance criteria (R2 ≥ 0.85 and relative RMSE ≤ 15%). However, the baseline LSTM achieves lower prediction error on the test set compared to the GA-optimized model, indicating superior generalization performance. This finding highlights that improved validation fitness does not necessarily translate into better performance on unseen assets. The proposed framework provides a controlled proof of concept for reproducible evaluation of predictive models and supports future extensions using real OBD-II data and more robust validation schemes.

Instituciones: Universidad Tecnológica del Perú