Under the background of high-speed train production performance degradation, aiming at the problem of insufficient sensitivity of high-speed train traction system for early fault detection, we propose an early fault detection method based on multi-task learning. By using the multi-task network structure, this method obtains the probability density function of the data with the help of kernel density estimation and detects the fault from the data and the probability density function value. By comparing and analyzing the detection results of two fault detection indexes, Hotelling’s T2 statistic and Kullback-Leibler divergence statistic, the identifiability of fault indexes is enhanced. In addition, the adopted multi-task learning network structure enables the networks between two different tasks to share information effectively so that the performance of the overall network for fault detection is improved. Finally, the traction drive control system platform developed by Central South University is used to verify the effectiveness of the proposed method.