Attack Detection in High-Speed Maglev Train Network Control Systems Based on Multidimensional Consistency

A train network control system is a cyber–physical system (CPS) whose safe operation depends on the integrity and timeliness of on-board sensor data. As such systems migrate towards networked and wireless architectures, the measurement link becomes exposed to replay, denial-of-service (DoS) and false data injection attacks (FDIA), against which single-dimensional detectors are no longer sufficient. This paper proposes a unified sensor-attack detection framework based on multidimensional consistency checking. A linear discrete-time state-space model and an explicit threat model are established first, and each attack is mapped component by component onto the system architecture and onto the network-level countermeasures that remain applicable. Three complementary layers then extract attack features from the algebraic, temporal-statistical and physical dimensions: analytical-redundancy parity-space detection, improved Transformer-based temporal prediction, and a physics-informed neural network (PINN) observer. For the PINN layer, a sufficient FDIA detectability condition is derived. The attack enters the physical residual through an explicitly computed attack-to-residual gain, which yields a non-centrality parameter and a closed-form bound on the attack magnitude; the gain predicted by the theory agrees with the measured gain to within 14%. Every distributional assumption behind these statistics is verified rather than taken for granted. The three statistics are then combined by a likelihood-ratio rule built from empirically estimated class-conditional densities of probability-integral-transformed statistics. This rule requires no conditional-independence assumption and additionally returns a coarse attack-type estimate whose reliability is quantified experimentally. On a single-car maglev experimental platform, the fused rule attains the highest average area under the curve (AUC, 0.943) and the highest F1 score, and it gives the shortest detection delay of all methods in every scenario (0.06–1.18 s, against 1.04–10.20 s for the individual layer). In a closed-loop Monte Carlo study it reaches an AUC of 0.996 under complete channel blocking, a regime in which the algebraic and physical layers collapse to 0.001.