Pantograph-catenary pantograph-catenary arc faults pose a serious threat to the safe operation of trains, and the accurate identification of such faults is crucial for ensuring the reliability of railway power supply systems. Aiming at the problems that traditional methods have low recognition accuracy under complex working conditions and are difficult to balance the extraction of temporal and spatial features, this paper designs a pantograph-catenary arc fault detection method based on the Markov transition field （MTF） and an improved dream optimization algorithm. First, the original current signal is denoised by robust empirical mode decomposition. The signal-to-noise ratio of the denoised signals under four working conditions is all above 36 dB, and the correlation coefficient is greater than 0.985, which preserves the fault transient characteristics to the greatest extent. Then, the one-dimensional time-series signal is mapped into a two-dimensional feature matrix using MTF, and a dual-branch architecture combining DarkNet19 and Gated Recurrent Unit is constructed to extract deep visual features and temporal dynamic features respectively, realizing multi-modal information complementarity. Furthermore, the improved dream optimization algorithm is introduced to adaptively optimize key parameters such as learning rate and the number of GRU neurons. Meanwhile, the multi-head self-attention mechanism is optimized, and redundant information is suppressed by L1 sparse regularization and an adaptive dynamic scaling factor for redundancy. In addition, transfer learning is integrated to break through the limitation of scarce fault samples. Experimental results show that the proposed model achieves a fault detection accuracy of 99.58% under four different working conditions, with precision and recall both exceeding 99.5%.Compared with six comparative models including SVM, 1D-CNN and ResNet, the accuracy is improved by 3.75%~6.67%.In the anti-noise interference experiment, the detection accuracy of the denoised signal is 6.71% higher than that of the noisy signal. It exhibits stronger robustness and provides an effective technical solution for the accurate diagnosis of pantograph-catenary arc faults.