The permanent magnet synchronous motor (PMSM) traction system adopts an axle control mode, with each power vehicle containing four traction drive units. Due to the complex electromechanical coupling characteristics, a fault in a traction drive unit can affect the entire system. Accurate detection and isolation of faulty units are crucial for train safety. This paper uses the spatial vector amplitude of the stator three-phase current as the diagnostic signal, constructs causal feature vectors using Convergent Cross Mapping (CCM), and classifies faults with the Random Forest (RF) algorithm. A CCM-RF fault diagnosis method is proposed and the simulation results of IGBT open-circuit fault and speed sensor interference fault indicate that the CCM-RF method can accurately identify the faulty traction drive unit in the traction system, verifying the effectiveness of the proposed method.