PSRONet: A Deep Reinforcement Learning-Based Sensor Configuration Framework in Railway Point Machines Fault Diagnosis

Abstract

Despite rapid progress in data-driven fault diagnosis for Railway Point Machines (RPMs), most studies tacitly assume a fixed sensing setup: predefined sensor types, placements, channels, and sampling rates. This leaves a critical gap: the sensing layer (i.e., the sensor configuration of the monitoring system) itself is rarely optimized for the diagnostic task, cost constraints, or robustness to real-world disturbances. We address this gap by formulating sensor-configuration optimization as a Markov Decision Process (MDP), framing it as a task-aligned decision problem. To solve this, we introduce the Policy-guided Sensor Reduction and Optimization Network (PSRONet), a Deep Reinforcement Learning (DRL) framework that learns a deployable sensing policy (selected channels and sampling interval) to maximize diagnostic accuracy under acquisition budgets. The PSRONet optimizes a multi-objective reward that balances accuracy and sensor cost (quantified by data overhead), and explicitly incorporates robustness terms reflecting sensor anomalies and environmental interference encountered in field deployments. Trained offline with a fast “Fidelity Oracle” surrogate of the downstream classifier, the PSRONet outputs a deterministic configuration that is simple to implement on existing hardware. Experiments on real-world RPM data across diverse operating conditions show that the PSRONet establishes a stronger accuracy–cost Pareto frontier than many existing methods. It retains near 91% of full-data diagnostic performance with over 92% lower acquisition overhead and remains robust to sensor anomalies and environmental interference. To our knowledge, this is the first RPM-oriented, DRL-based framework that directly optimizes the sensor layer for fault diagnosis, providing a practical path toward cost-aware and robust PHM in railway operation and maintenance.

Topics & Concepts