Power transformers are crucial for power grid reliability, and Dissolved Gas Analysis (DGA) is widely used for fault diagnosis. However, traditional methods like the IEC three-ratio method suffer from insensitivity to early faults and reliance on manual experience. With the popularization of intelligent sensors, massive chromatographic time-series data urgently need efficient analysis methods. Nevertheless, the proportion of unlabeled data exceeds 90%, restricting the application of supervised learning methods. This paper proposes a full-process health assessment framework integrating self-supervised contrastive learning and attention mechanism. It employs Temporal-aware Adversarial Augmentation and Generative Adversarial Networks (GANs) with physical constraints to address data scarcity, and a Dual-stream Contrastive Attention Network (D-CAN) combining ResNet1D and Transformer for robust feature extraction. The Dynamic Health Index (DHI) quantifies health status via unsupervised clustering and Mahalanobis distance. Experiments show that DHI achieves “zero delay” fault detection, outperforming IEC and Isolation Forest, with 3D visualization verifying health status distribution. This framework systematically enhances data quality, feature representation, and indicator interpretability, offering a novel path for transformer health monitoring.