According to the problem of traction converter fault caused by the abnormal condition of cooling filter for high-speed train, a fault diagnosis method based on multi task deep learning is proposed by comprehensively analyzing the correlation between fault classification of traction converter and filter blockage degree. Firstly, a multi-task deep neural network (MT-DNN) including the main task of traction converter fault diagnosis and the sub task of filter blockage degree is constructed. Then, in order to accurately predict the failure and degradation trend of traction converter, the prediction results of MT-DNN and self-organizing mapping (SOM) method are combined to construct the multi-task deep neural network self-organizing mapping model (MTDNN-SOM). This method defines the degradation state curve according to the evolution law of characteristic variables of historical fault data, which directly reflects the relationship between fault characteristics and degradation state, and finally realizes fault diagnosis and maintenance prediction for the cooling filter of traction converter. The experimental results show that the proposed method is superior to the single task or traditional fault diagnosis method in both accuracy and efficiency, and has achieved good results.