Aiming at the problem of possible false alarm caused by the single diagnosis targets and insufficient consideration of the strong coupling between sensor signals, a joint real-time diagnosis method of multi-sensor faults is proposed. The research begins by analyzing the diagnosability based on the existing sensor layout and constructs the minimum structured overdetermined equation set (MSOs) and fault characteristic matrix that can achieve isolation of all sensors faults. Secondly, based on the sensor signal sets corresponding to each MSO and relevant system mechanism knowledge, the input and output signals of the data-driven model, the order of model input signals, and the correlation between different input signals are determined. Subsequently, using the Extreme Learning Machine (ELM) algorithm, data-driven models for each MSO are established based on historical normal data samples, enabling effective estimation of their output values and generating residual sequences. By combining these with the fault characteristic matrices, effective detection and diagnosis of different sensor faults are achieved. Finally, a virtual and physical joint test verification platform, which employs semi-physical simulation and on-site fault scenario recording, is used to test and verify the proposed diagnostic algorithm. The verification results demonstrate that, compared to existing methods, the proposed method can achieve rapid detection and localization of multi-sensor faults in traction drive system, offering significant value for engineering applications.