Timely and accurate failure analysis of photovoltaic (PV) systems is crucial forensuring the stable operation of power grids. However, existing failure analysis and diagnosis algorithms based on deep neural networks excessively rely on high-quality failure state data collected by sensors. This is extremely difficult to achieve in real photovoltaic power plants that are commonly equipped with self-protection mechanisms. To address this issue, we propose a Digital Multi-Twin integrating Theory, Features, and Vision (TFV-DMT) for failure analysis of PV strings in PV systems. This method first constructs theoretical simulation twins, feature twins, and visual twins based on the concept of digital twins, specifically tailored for actual PV systems, and designs a multi-twin collaborative model for model updating and failure diagnosis. Secondly, to better construct the visual twin, we introduce a Two-Dimensional Gram Angle Field Transformation Algorithm (TDGAF) to achieve targeted two-dimensional mapping of PV feature data, facilitating a more direct expression of failure characteristics. Furthermore, by constructing a Swish-activated Deep Convolutional Generative Adversarial Network (SDCGAN) to achieve balanced augmentation of mapping data, the model bias of theoretical simulation twins can be reduced. Finally, we propose a Swin-LT network that incorporates a Lightweight Dual-Channel Attention Module (LDAM) to better analyze the features of the visual twin, enabling more precise fault diagnosis. The algorithm has been validated on a real 250 kW grid-connected PV system, with results indicating that the proposed digital twin model is effective, achieving a diagnosis accuracy rate of 98.8 % for string failures.