<p>Multi-terminal high-voltage direct current (MTDC) systems with half-bridge modular multilevel converters (MMCs) enable high-capacity power transfer and flexible control. However, overhead DC lines in remote, harsh environments remain prone to frequent faults. Conventional detection methods are often limited by threshold sensitivity, high-resistance faults, and polarity identification. To address these shortcomings, a novel DWT-RCMDE-ET framework is proposed, in which two-level discrete wavelet transform (DWT), refined composite multiscale dispersion entropy (RCMDE), and an extremely randomized trees (ET) classifier are integrated. First, salient time–frequency features are extracted from fault voltage signals by two-level DWT. Then, subtle fault characteristics that are often overlooked by conventional techniques are captured by RCMDE. Finally, fault types and affected poles are accurately identified by the ET classifier, even under high-resistance and noisy conditions. The simulation results showed that the proposed algorithm can accurately identify faults under different fault resistances, achieving a 100% accuracy in PSCAD/EMTDC verification and a 96.3% accuracy in RT-LAB validation. Compared with convolutional neural network (CNN), long short-term memory (LSTM), and Transformer-based methods, DWT-RCMDE-ET outperforms under varying noise and sampling frequencies, with higher accuracy, F1-score, Recall, and more significant P-value. Thus, the proposed DWT-RCMDE-ET is a reliable solution for advanced MTDC fault detection, offering great potential to enhance the operational stability of future power networks.</p>