From bearings to substations: Transfer Learning for fault detection in district heating

Fault Detection and Diagnosis (FDD) in District Heating (DH) systems is vital for improving operational efficiency. As DH networks evolve towards Fourth-Generation District Heating (4GDH), their reliance on lower operational temperatures intensifies the need for robust FDD. However, implementing effective FDD faces challenges due to the lack of labelled fault data and the complexity of DH substations. This paper introduces a novel FDD methodology using Transfer Learning (TL) to bridge the gap between faulty bearings, controlled DH substation experiments and real-world operational data. We propose a fault signature method that aligns the data to reduce the domain gap, revealing similarities between faulty bearings’ vibration patterns and Δ T readings in DH substations. The TL-enhanced models demonstrated robust performance, achieving F1 scores up to 98% on lab data and 91% on real-world operational data, respectively. These results mark a notable advancement in FDD of DH substations, as our method offers accurate fault detection and valuable insights across diverse operational contexts, ranging from valve issues, faulty sensors, wrong control strategies and normal behaviour. Notably, temperature dynamics resemble behaviour akin to faulty bearing vibrations, highlighting their potential as a critical indicator of a faulty substation, enabling more effective FDD in DH systems. • Transfer learning framework using faulty-bearing, lab-emulated district heating, and real-world district heating data for substation fault detection. • A novel daily-signature generation pipeline bridging mechanical vibration and thermal temperature patterns. • The temperature signatures exhibit high variance and vibration-like patterns when a substation is faulty, as opposed to a stable, normal operational substation. • Models achieve high fault detection performance with an F1 score of 98% on lab emulation and 91% on real-world DH substation data.