Towards Automatic Thermography-Based Leak Detection in District Heating Systems

As one of the greatest contributors to anthropogenic climate change, the building sector and its associated operational emissions need to be limited. District heating systems (DHSs) can be part of the solution by supplying climate neutral heat to urban environments via networks of mostly subterranean pipelines. However, continuous operation under extreme conditions makes these systems prone to leaks, which cause persistent losses and potential safety hazards due to the lack and imprecision of existing monitoring methods. Thermography-based leak detection (TLD) has emerged as a non-invasive, pipeline-agnostic alternative, pinpointing underground leaks through their associated elevated surface temperatures. Combined with airborne platforms such as unmanned aircraft systems (UASs), TLD can enable large-scale monitoring of DHSs. However, its wider adoption is limited, in particular, by the effort required to manually evaluate the large amount of resulting thermal infrared images. An automatic analysis is required that is robust, usable, generally applicable, and economically viable to bridge the gap between scientific research and real-world implementation.
... mehr
The present dissertation addresses these four objectives and central research question through five studies. Reliability is advanced by developing and comparing various algorithms for thermal image processing, thermal anomaly detection, and false alarm removal (Studies A-D). Two novel deep learning (DL) models prove particularly effective for these core tasks (Studies C and D), while data quality is revealed to be a key determinant of performance (Studies A, B, and D). Usability is achieved by fully automating the analysis, including photogrammetric preprocessing, and designing easily interpretable outputs (Study A). Representativity is addressed through the creation and publication of novel UAS-based datasets, thereby establishing new benchmarks for TLD (Studies A and B) as well as DL-based semantic segmentation of spectral imagery (Studies C and D). Finally, a first economic break-even analysis for TLD demonstrates its viability as a cost-effective leak detection strategy, particularly when implemented with the described automation (Study E).

Together, these contributions establish an effective analysis and methodological foundation for automatic TLD, paving the way for its adoption for large-scale DHS monitoring and integration into future smart city applications.