Abstract
Railway Turnout Machines (RTMs) are critical components of railway infrastructure, where failures can lead to severe safety risks. This paper introduces CEC-PA, a cloud-edge collaborative framework designed for real-time, 24/7 fault diagnosis of RTMs. The system employs a hierarchical ensemble of deep learning models—including MLP, DAE, and TCN—fused via fuzzy logic. To address the computational demands and reliability requirements of edge deployment, CEC-PA utilizes a reinforcement learning-based pipeline parallelism strategy and an election consensus mechanism. Experimental results on a large-scale real-world dataset demonstrate a 97.4% diagnostic accuracy and a significant inference speedup of up to 7.93x, ensuring robust performance even under node failure conditions.
Introduction
Railway Turnout Machines (RTMs) are the mechanical switches responsible for directing trains by moving tracks. Operating in harsh outdoor environments, these machines are prone to wear and failure, necessitating continuous monitoring to prevent derailments. However, existing AI-driven solutions face a dual challenge: high-accuracy deep learning models are often too heavy for local edge devices, while cloud-only processing introduces latency and creates single points of failure. This research proposes a framework that balances low-latency real-time alerts with high reliability through distributed execution and fault-tolerant coordination.
Methodology
1. Modular Fault Diagnosis Model
The proposed diagnostic system moves away from monolithic architectures in favor of a hierarchical ensemble approach. The process begins with Phase Segmentation, where a Gated Recurrent Unit (GRU) identifies the three operational stages of a turnout: Starting, Transition, and Indication.
Following segmentation, three specialized sub-classifiers analyze the data:
- MLP: Extracts traditional time-domain features such as peak current.
- Denoising Auto-Encoder (DAE): Captures morphological similarities in the current waveform's shape.
- Temporal Convolutional Network (TCN): Learns deep sequential dependencies.
A late-fusion layer using Fuzzy Logic then combines these outputs to manage sensor data ambiguity and produce a final classification.
2. The CEC-PA Framework
To execute these models efficiently across distributed resources, the Cloud-Edge Collaborative Pipeline Parallelism (CEC-PA) framework was developed. It features:
- Strategic Partitioning: The model is broken into sub-tasks that can be executed in parallel across cloud and edge nodes.
- PPO-based Resource Allocation: A Proximal Policy Optimization (PPO) agent dynamically decides task offloading based on real-time network and hardware conditions.
- Election Consensus Mechanism: Inspired by the Raft protocol, this mechanism ensures that if a coordinator node fails, a new one is elected automatically, maintaining system availability.
Experimental Setup and Dataset
The study was validated using a real-world dataset from the Nanjing Metro Bureau, comprising 10,000 samples from 227 turnouts across 12 metro lines (2021–2023). The data includes three-phase AC current sequences (25Hz), GPS, and rotation data. Labels cover a "Normal" state and 11 fault types, ranging from "Hidden Dangers" (maintenance indicators) to "Critical Faults" (immediate risks).
Results and Discussion
The CEC-PA framework achieved a state-of-the-art 97.4% accuracy, outperforming traditional baselines like SVM (69.0%) and GBDT (84.2%). Key performance insights include:
- Inference Speedup: The pipeline parallelism achieved a speedup of 1.98x to 7.93x compared to non-parallelized deployments.
- Scalability: Under high loads (200 requests/s), the system maintained a response time of ~1709ms, whereas competing methods exceeded 20,000ms.
- Robustness: The election mechanism successfully handled node failures, ensuring zero-downtime diagnostic services.
The authors found that partitioning at the component level (pipeline parallelism) is significantly more efficient for edge computing than neuron-level parallelism, as it minimizes the communication overhead between nodes.
Conclusion and Future Work
CEC-PA provides a scalable and reliable solution for railway safety. Future research will focus on optimizing the Markov Decision Process (MDP) for offloading decisions, further reducing the latency of the re-election phase during total cloud outages, and incorporating unsupervised learning techniques to detect novel, previously unseen fault types.
Summary updated