Many insulation issues in high voltage direct current (HVDC) gas insulated switchgear (GIS) originate from partial discharge (PD) activity induced by defects within the equipment. Accurate identification of defect types responsible for PD events is therefore critical for proactive maintenance and reliability assurance. In partial discharge defect classification, two primary methodologies have historically dominated: traditional manual feature extraction and data-driven deep learning. Early works typically employed these approaches in isolation, with manual features relying on domain knowledge and deep learning emphasizing automated feature discovery. This paper presents a framework for collecting PD signals from four typical defects and performing systematic feature extraction across the time domain, frequency domain, and time-frequency domain to leverage the full information potential of these signals. These features are integrated with deep learning through a novel dual-input neural network (DINN) that fuses traditional feature engineering with data-driven deep learning methods. By combining these approaches, the model achieves robust pattern recognition of PD signals in DC GIS. Experimental results demonstrate that single-domain representation yields insufficient classification performance, whereas the proposed dual-input architecture enhances recognition accuracy, enabling reliable classification and identification of PD signals associated with different DC GIS defects.