Automated Fault Detection and Diagnosis (FDD) in District Heating (DH) is essential for reducing return temperatures, improving operational efficiency, and supporting the transition to low-temperature, low-carbon integrated energy systems. Practical deployment remains difficult because labelled fault data are scarce and heterogeneous across domains, while unlabelled operational data are abundant, creating a practical mismatch. To address this, we propose a Self-supervised Time Series Transformer (STST) for low-label FDD using routinely available primary-side measurements. We evaluate the method on six datasets spanning laboratory fault emulations, simulation, and real-world DH networks. The results show that primary-side temperatures contain strong fault-discriminative information, while flow measurements add sensitivity to hydraulically driven faults. The proposed method performs competitively in scarce-label and heterogeneous-field settings, achieving F1 values between 0.77 and 0.98. Cross-domain experiments indicate that fault-relevant structure is transferable across substations within the same network, although transfer remains asymmetric across networks. Overall, the results suggest that self-supervised pre-training and regularised fine-tuning can improve transformer-based DH fault detection under stricter field conditions, while CNN-based and hybrid baselines remain highly competitive in cleaner or more separable regimes. Practical deployment across new networks, therefore, still requires attention to source–target compatibility, local validation, and remaining domain shift.