Abstract
To mitigate mode collapse in Deep Convolutional Generative Adversarial Networks (DCGANs), this study proposes an innovative deep learning model integrating Principal Component Analysis (PCA)—the PCA-enhanced Deep Convolutional Generative Adversarial Network (PCA-DCGAN)—to mitigate mode collapse during sample generation. By introducing a PCA module prior to the generator, principal components of input samples are extracted and fed into the generator as structured noise input. The resulting components are then fed back into the generator as part of the noise input. This approach breaks from the traditional random selection of generator input noise and provides a more principled direction for optimizing the generator’s parameters. The approach effectively alleviates mode collapse and reduces computational complexity. Experimental results show that the PCA-DCGAN model achieves Fréchet Inception Distance (FID) scores 35.47 and 12.26 lower than those of DCGAN and Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), respectively. It also significantly improves classification accuracy and reduces loss when training with augmented datasets, thereby validating the effectiveness of the proposed method. The model is applied to generate pantograph-catenary arc electromagnetic interference (EMI) signals, addressing challenges of transient signals and sample scarcity in high-speed railways. This provides an innovative solution for data-scarce domains.