Partial Discharge (PD) analysis is a critical metric for evaluating the insulation performance of Switching Reluctance Motors (SRMs) in industrial applications. However, PD signal acquisition is often hindered by noise and interference, leading to inaccurate diagnostics. This work proposes a robust PD pattern recognition framework that combines an adversarial de-noising model with enhanced feature extraction from phase-resolved partial discharge patterns using a Canny edge detection method. To address class imbalance, both weighted and macro average techniques are applied, with feature extraction performed via a pre-trained VGG19 Convolutional Neural Network (CNN). Fish Swarm Optimization (FSO) is employed to fine-tune the model's hyperparameters. The proposed method achieves a high identification accuracy of 99%, demonstrating strong resilience against noise and signal occlusion, making it highly suitable for reliable PD detection in industrial SRM systems.