In order to address the model performance degradation caused by sample sparsity and class imbalance in multimode fault diagnosis of automatic guided vehicle (AGV) power batteries, a generative adversarial network incorporating gradient penalty and an independent classifier (GPSC-GAN) was proposed. First, the discriminator loss function was optimized through the introduction of Wasserstein distance with gradient penalty, whereby the distribution divergence between generated samples and real battery fault features was constrained to enhance training stability. Second, an independent classifier module was designed to decouple classification and discrimination functions, enabling high-quality multimode fault samples to be generated for constructing a balanced dataset. Experimental results on AGV battery fault datasets demonstrated that the generated samples achieved structural similarity index measure (SSIM) and maximum mean difference (MMD) values of 0.5 and 0.2 respectively, showing 25% to 66% improvements compared with WGAN-GP, ACGAN, SMOTE, ADASYN, Diffusion GAN and StyleGAN-T models. After integrating synthetic samples, the F1-score of the fault diagnosis model was enhanced to 0.938, with precision and recall rates reaching 94.6% and 93.0% respectively. These metrics outperformed six other comparative methods, verifying the proposed method’s efficiency and robustness in complex industrial scenarios.