Railway infrastructure maintenance efficiency improvement using deep reinforcement learning integrated with digital...

Area of Science:

• Railway Engineering
• Artificial Intelligence
• Predictive Maintenance

Background:

• Railway maintenance is complex, with traditional methods (corrective, preventive) often inefficient.
• Ineffective maintenance leads to increased costs, safety risks, and reduced passenger comfort.
• Predictive maintenance offers higher efficiency but requires advanced tools for implementation.

Purpose of the Study:

• To propose a novel approach for improving railway maintenance efficiency.
• To integrate deep reinforcement learning (DRL) with digital twin technology for predictive maintenance.
• To address limitations of traditional and other machine learning techniques in railway upkeep.

Main Methods:

• Developed a DRL model using the Advantage Actor Critic (A2C) algorithm.
• Utilized four years of real-world field data covering 30 km of railway track.
• Incorporated track geometry, component defects, and maintenance activities as model parameters.
• Defined rewards/penalties based on maintenance costs and defect occurrences.

Main Results:

• Reduced maintenance activities by 21%.
• Decreased occurring defects by 68%.
• Demonstrated superior performance of A2C over traditional methods like Deep Q-learning (DQN).

Conclusions:

• The integration of DRL and digital twins offers a breakthrough for efficient railway maintenance.
• This approach significantly cuts costs, reduces downtime, and improves safety and passenger comfort.
• Presents a novel guideline for applying AI in railway infrastructure management.