For industrial applications, one of the most important metrics for assessing the performance of an insulating system is partial discharges (PD) in a switching reluctance motor (SRM). Because of their unique design, which permits a high degree of torque per unit volume, SRMs can achieve high torque density and high efficiency.
However, in a real-world scenario, several PD sources (single or multiple) or significant interference may hinder the signal’s acquisition, potentially leading to incorrect equipment diagnostics. Convolutional neural networks (CNNs) have outstanding classification performance and robust automatic feature extraction capabilities, making them adept at recognizing patterns. Nevertheless, the collected samples are unbalanced, which complicates the ability of the existing approaches to provide an accurate diagnosis.
This proposed work addresses two key challenges in PD pattern recognition: external interference is mitigated using an adversarial de-noise model, and a three-dimensional phase-resolved PD (PRPD) pattern is enhanced using Canny edge detection. Feature extraction is performed using a pre-trained CNN, specifically VGG19—a 19-layer deep CNN architecture known for its simplicity and strong performance in transfer learning applications. To optimize recognition performance, the hyperparameters of VGG19 are fine-tuned using a Fish Swarm Optimization (FSO) algorithm. The suggested results show that the system, which uses an optimized VGG19 architecture combined with a macro average method for industrial applications, achieves a reasonable 99% identification rate and is robust against noise and occlusion variations.
For more on this article click here.