Systematic GAN Parameter Selection for Fault Data Generation using Particle Swarm Optimization

In real-world applications of artificial intelligence, fault data are often insufficient, making data training challenging. Generative adversarial networks (GAN) are widely recognized for addressing data generation issues. This study introduces swarm intelligence algorithms, specifically particle swarm optimization (PSO) and genetic algorithm (GA), to enhance GAN performance. The single-channel vibration signal data from a faulty motor serves as the case study. Fault vibration signals can rapidly, accurately, and comprehensively reflect the nature and extent of mechanical faults. The study compares and analyzes the results of PSO, GA, and random number (RN) optimization of GAN parameters. The findings demonstrate that PSO outperforms GA in terms of time efficiency and reducing generation errors. Swarm intelligence algorithms eliminate the need for manual experience or repetitive trials when selecting parameters. Compared to GA and RN, PSO improves performance by 80.92% and 90.44%, respectively, while also reducing optimization time by 46.55% compared to GA.