A hybrid Bi-LSTM model for data-driven maintenance planning

Area of Science:

• Engineering
• Computer Science
• Operations Research

Background:

• Modern industries require efficient asset maintenance under resource constraints.
• Classical maintenance methods face limitations due to estimation errors and computational complexity.
• Industry 4.0 and deep learning (DL) enable data-driven health predictions for maintenance planning.

Purpose of the Study:

• To bridge the gap between DL-based remaining useful life (RUL) predictions and maintenance plan optimization.
• To develop a scalable and accurate framework for the selective maintenance problem (SMP).
• To optimize maintenance for mission-oriented series k-out-of-n:G systems.

Main Methods:

• Development of a hybrid DL model incorporating Monte Carlo dropout for RUL predictions.
• Construction of empirical system reliability functions from RUL predictions.
• Optimization of the selective maintenance problem (SMP) using the generated reliability functions.

Main Results:

• The proposed framework effectively optimizes maintenance plans, minimizing costs while ensuring mission survival.
• Numerical experiments demonstrate superior performance compared to prior SMP methods.
• The method provides exact solutions without computationally intensive parametric reliability functions.

Conclusions:

• The developed hybrid DL framework offers a scalable and accurate solution for complex industrial maintenance scenarios.
• It enables data-driven, optimized maintenance planning, enhancing operational efficiency and reducing costs.
• The approach is applicable across various industrial contexts and system configurations.