Reliable fault diagnosis in photovoltaic (PV) arrays is essential for reducing energy losses, preventing safety hazards, and maintaining system availability; however, conventional supervised learning methods require large labeled datasets that are costly and risky to obtain from operating plants. This paper proposes a data-efficient diagnostic framework that combines convolutional neural networks, semi-supervised generative adversarial learning, and few-shot meta-learning. A 5.3-kW, 6 × 3 PV array is modeled in MATLAB/Simulink to generate healthy and faulty operating data under variations in irradiance, temperature, fault resistance, and partial shading. Four conditions are considered: normal operation, line-to-line faults, hot-spot or high-series-resistance faults, and open-circuit faults. Current, voltage, irradiance, and temperature signals are normalized and transformed using the short-time Fourier transform; their spectrograms are then concatenated into composite images for automatic feature extraction and classification. A three-layer CNN provides a high-accuracy supervised benchmark, reaching approximately 99.5% accuracy on noise-free data. To reduce dependence on labeled samples, a semi-supervised GAN exploits unlabeled and synthetically generated data and achieves 81.57% test accuracy using only 100 labeled samples on the one-dimensional noisy dataset. To further address the computational cost of GAN training and extreme label scarcity, a few-shot meta-learning model is introduced. In the 5-shot, 4-way setting, it achieves 94.03% accuracy on previously unseen samples, substantially outperforming a conventional CNN trained with limited labeled data. The main contribution is therefore a unified comparison of three complementary learning strategies and a practical few-shot solution that preserves strong diagnostic performance with minimal labeled PV fault data under realistic operating variability.