To address the insufficient discriminability of fault-section features caused by disturbances such as train starting, load fluctuations, and rectifier ripple in metro DC traction networks, an active harmonic injection-enhanced CNN-LSTM method is proposed for short-circuit fault location. Harmonic signals are actively injected into the traction network to enhance the fault-discriminative information contained in the input features. The injected harmonic response is then extracted using the empirical wavelet transform and employed as the fault-location feature. A convolutional neural network is used to extract local waveform features, while a long short-term memory network is introduced to learn temporal variation characteristics, thereby enabling the classification of train-starting conditions and different fault sections. The results show that, compared with the original traction current input, using the harmonic response induced by active injection as the input significantly improves the fault-location performance of all evaluated models. Among them, the CNN-LSTM model achieves both an accuracy and an F1-score of 98.03%, effectively distinguishing train-starting conditions from short-circuit faults and accurately identifying different fault sections. The proposed method provides an effective approach for short-circuit fault-section location in metro DC traction networks.