To enhance the accuracy of partial discharge (PD) pattern recognition, this paper proposes a method integrating an Multi-Scale Convolutional Neural Network (MCNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network with a self-attention mechanism. Firstly, time-frequency analysis is conducted on PD pulse signals to produce time-frequency diagrams. Then, the MCNN-BiLSTM model with self-attention is utilized to automatically extract key features from these diagrams and perform pattern recognition. Experiments using high-frequency current signals from three typical PD faults in oil-immersed transformers show that the proposed method effectively identifies PD fault types. Unlike traditional feature extraction methods relying on expert experience, this model can comprehensively capture deep feature information in PD signals, significantly improving pattern recognition accuracy.