EMI and IEMI Resistant Signalling and Communication Systems for High-speed Rail

Safety-critical railway operations increasingly rely on wireless communications to convey signals relating to train control, supervision, and passenger services under stringent reliability and ultra-low latency constraints. As railway electrification expands and services accelerate, the signal transmission is exposed to electromagnetic interference (EMI) from natural and onboard sources and to intentional EMI (IEMI) from hostile jammers. These disturbances threaten the security of communicating control messages and can trigger emergency braking if they are not promptly detected and mitigated. This thesis investigates the end-to-end impact of EMI/IEMI on modern railway wireless communications and proposes detection, classification, and localization methods designed for next-generation 5G-R/FRMCS systems operating at high speed.

First, a system-level network-based modeling framework is developed to represent EMI/IEMI sources, coupling paths, and vulnerable subsystems as nodes and edges, enabling unified reasoning from physical coupling to link-level performance. Four representative EMI source categories, pantograph catenary arcing, onboard power electronics, public cellular co-existence, and diverse jammers, are analyzed to expose coupling paths and performance degradation mechanisms.

Building on these insights, a real-time classification pipeline is introduced for high-mobility 5G-R links. Time-series signal features are extracted with fine time-frequency resolution using both real and imaginary components. An Attention-BiLSTM architecture performs adaptive multi-class detection of EMI/IEMI types under rapid channel variations. Assessed on four representative railway scenarios with the train speed up to 500 km/h, the method achieves 94.98% accuracy with a 7.43 ms decision latency; validation on measurement data from a 5G-R test facility attains 92.5% accuracy. These results confirm the sub-10 ms safety detection requirements while maintaining robustness across different scenarios.

To address evolving or previously unseen disturbances, an unsupervised anomaly detector based on an AE–BiLSTM learns nominal operations and flags deviations based on the reconstruction loss. The approach reduces dependence on labelled anomalies, improves accuracy by approximately 5% over strong baselines (93.24% overall), and achieves 4.51ms online decision speed. The framework demonstrates stable performance across different speeds and environments, supporting early warning before service degradation propagates to safety-critical functions.

Finally, a localization is proposed which combines a CNN–BiLSTM angle-of-arrival estimator using a circular 16-element array with multi-AOA fusion and a Kalman filtering for 3D triangulation. The system delivers 0.71° AOA RMSE at 2.69ms inference latency and achieves sub-3m position error across scenarios (sub-2m in 92% of cases), enabling timely site isolation and targeted mitigation.

To conclude, the contributions demonstrate that deep-learning-based, time-series-aware methods can meet the latency, accuracy, and robustness requirements in EMI and IEMI-resistant wireless communications for high-speed rail. The proposed models form a practical foundation for interference-aware detection, classification, and localization within operational time budgets, and provide actionable pathways toward resilient 5G-R/FRMCS deployments in intelligent transportation systems.