District heating (DH) networks are vital in the transition to sustainable energy systems. Maintaining their performance requires continuous monitoring to reduce heat losses, avoid user discomfort, and support efficient operation. However, automatic fault detection and diagnosis (FDD) in DH substations remains challenging due to limited labelled data, class imbalance, heterogeneous operating conditions, privacy constraints, and the lack of standardisation. This thesis aims to develop learning-based approaches for automatic FDD in DH substations that can operate under these real-world industry constraints. Therefore, this work focuses on scalable representations, transferability across domains, and robustness against scarce and imbalanced fault data, and includes eight studies that jointly develop and validate learning-based FDD methods for DH substations. Across these studies, the work advances unsupervised locality-based anomaly detection, hybrid and augmentation-enhanced fault diagnosis, transfer and cross-modal learning for label-scarce settings, privacy-preserving semi-supervised federated learning, and streaming representations for continuous fleet-scale monitoring. The results demonstrate that reliable, scalable FDD in DH systems can be achieved despite severe field constraints. By combining topology inference for local analysis, data representation, knowledge transfer and data augmentation, this thesis advances practical, deployable FDD intelligence to support more resilient, efficient, and data-smart DH networks aligned with the climate and digitalisation goals.