Fault Detection in Solar Thermal Systems using Probabilistic Reconstructions
Florian Ebmeier, Nicole Ludwig, Jannik Thuemmel, Georg Martius, Volker H. Franz · 2025
Read Fault Detection in Solar Thermal Systems using Probabilistic Reconstructions by Florian Ebmeier, Nicole Ludwig, Jannik Thuemmel, Georg Martius, Volker H. Franz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-10296.txt.
Introduction
Florian Ebmeier, Nicole Ludwig, Jannik Thuemmel, Georg Martius, and Volker H. Franz detect faults in domestic solar thermal systems from the sensor data the systems already collect. Poor installation or maintenance can quietly slash efficiency or damage a system, and per-home human monitoring is too expensive. Their framework flags abnormal states by probabilistic reconstruction: a model rebuilds normal behavior, and poor reconstructions signal trouble.
Evaluated on the public PaSTS dataset of real operating systems with diverse fault types, reconstruction methods detected faults and generalized to systems never seen in training. The approach beat simple and more complex deep learning baselines, and estimating uncertainty that varies with conditions proved essential to performance. The authors also discuss the engineering overhead honestly and argue for simple models.
Catalog record
| CREATOR | Florian Ebmeier, Nicole Ludwig, Jannik Thuemmel, Georg Martius, Volker H. Franz |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-10296
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fault Detection in Solar Thermal Systems using Probabilistic Reconstructions Florian Ebmeier, Nicole Ludwig, Jannik Thuemmel, Georg Martius, Volker H. Franz Abstract Solar thermal systems (STS) present a promising avenue for low-carbon heat generation, with a well-running system providing heat at minimal cost and carbon emissions. However, STS can exhibit faults due to improper installation, maintenance, or operation, often resulting in a substantial reduction in efficiency or even damage to the system. As monitoring at the individual level is economically prohibitive for small-scale systems, automated monitoring and fault detection should be used to address such issues. Recent advances in data-driven anomaly detection, particularly in time series analysis, offer a cost-effective solution by leveraging existing sensors to identify abnormal system states. Here, we propose a probabilistic reconstruction-based framework for anomaly detection. We evaluate our method on the publicly available PaSTS dataset of operational domestic STS, which features real-world complexities and diverse fault types. Our experiments show that reconstruction-based methods can detect faults in domestic STS both qualitatively and quantitatively, while generalizing to previously unseen systems. We also demonstrate that our model outperforms both simple and more complex deep learning baselines. Additionally, we show that heteroscedastic uncertainty estimation is essential to fault detection performance. Finally, we discuss the engineering overhead required to unlock these improvements and make a case for simple deep learning models. [orcid=0000-0002-1518-9593, ] [ orcid=0000-0003-3230-8918 ] [ orcid=0009-0004-4209-6239 ] [ orcid=0000-0002-8963-7627 ] [ orcid=0000-0002-8457-0611 ] keywords Solar thermal systems ,Fault detection ,Anomaly detection ,Time series ,Deep learning ,Uncertainty estimation † † credit: Writing – original draft, Software, Data curation, Methodology, Investigation † † credit: Supervision, Funding acquisition, Writing – review & editing, Methodology Conceptualization † † credit: Software, Writing – review & editing, Methodology, Investigation † † credit: Funding acquisition, Writing – review & editing, Methodology, Conceptualization † † credit: Supervision, Funding acquisition, Writing – review & editing, Methodology, Conceptualization † † corresponding: Corresponding author 1 Motivation Domestic solar thermal systems (STS) present a promising approach to decarbonising the energy system. They require minimal natural resources, do not add strain to the electrical grid, and can be more space-efficient than photovoltaic systems. However, adoption has been hampered by low yields due to poor monitoring and undetected faults [ 32 , 33 ] . Monitoring these systems can be prohibitively expensive if it involves additional sensors, complex system modelling, and continuous oversight by human experts. With increased digitalisation, substantial sensor data are now available, creating an opportunity to develop data-driven fault and anomaly detection methods that reduce costs and human supervision. Data-driven time series anomaly detection (TSAD) has attracted considerable attention in machine learning. Among the prevailing paradigms, reconstruction- and forecasting-based methods are particularly popular, and recent years have seen a surge in the application of cutting-edge deep learning (DL) architectures to this task [ 1 , 2 , 9 , 42 , 35 , 45 ] . To our knowledge, no system applies state-of-the-art deep learning anomaly detection frameworks to real-world domestic STS. This is particularly severe when considering the variety of system realisations in operational domestic STS. Generalization remains the key challenge: to build robust models that can adapt to the variety of real-world system designs, operational states, and environmental factors common in domestic solar thermal setups. Furthermore, concerns have been raised regarding the soundness of experimental protocols and benchmarking practices in many DL studies [ 47 , 24 , 38 ] . For example, 38 demonstrate that simple linear methods such as principal component analysis (PCA) can outperform recent, more complex DL models; thus challenging the presumed superiority of advanced architectures. The finding that PCA substantially outperforms DL models in TSAD is surprising, as there is no theoretical basis for this shortcoming of DL models. Previous findings for Variational Autoencoders (VAEs) show that they learn local PCA representations [ 36 ] and can be interpreted as non-linear PCA. Thus, we hypothesize that this performance gap is not due to the inherent shortcomings of DL models, but rather to the added complexity of optimizing overly complicated DL models. Hence, we focus on simpler methods that require little optimisation while keeping the strengths of DL models. More specifically, we show that probabilistic reconstruction-based methods make fault detection in STS feasible, and explicit heteroscedastic uncertainty estimation is a simple yet effective method to enhance fault detection accuracy. Our experiments on the PaSTS dataset [ 12 ] will explore these hypotheses, offering insights into how data-driven techniques can be effectively integrated into domestic STS monitoring workflows. The structure of the paper is as follows: First, we briefly introduce and discuss the relevant literature on fault detection in STS and TSAD, identifying the research gap and our research contribution (Section 2). We then introduce the methods we use to close this gap (Section 3). Following this, we introduce the Experiments, in particular the dataset and experimental setup, and then show our results (Section 4). In section 5, we interpret these results and discuss the key takeaways of the study as well as possible improvements for the future. In a nutshell, this paper demonstrates that reconstruction-based anomaly detection methods are both qualitatively and quantitatively successful in fault detection in STS. Our method outperforms simple statistical baselines, such as PCA, as well as both simple and complex DL baselines. Additionally, uncertainty estimation further improves detection performance, with heteroscedastic uncertainty estimation leading to the best detection performance without requiring additional hyperparameters. 2 Background 2.1 Fault detection in solar thermal systems Fault detection in STS has been approached in various ways, as discussed in several works [ 8 , 13 ] . These approaches can be broadly categorized into rule-based and expert systems, modelling or parity-space approaches, and data-driven approaches, with some overlap between these categories. In practice, operational fault detection often combines elements of these approaches, leveraging domain expertise and automated detection tools to identify and diagnose failures. Rule-based and expert systems . Rule-based and expert systems detect faults by checking if a system’s behaviour deviates from predefined rules. These rules are crafted using detailed system knowledge to model specific failure cases individually tailored to the system type. As a key advantage, they often directly classify the specific fault whenever a rule is violated. A simple example is sensor fault detection: if a sensor produces highly implausible outputs, such as jumping several orders of magnitude outside its dynamic range, it is considered a faulty sensor. However, this approach only detects anticipated failures encoded in the rules. Any anomalies or failures not considered during rule creation will go undetected. Expert and rule-based systems have been employed, for example, in the IP-Solar project [ 11 , 31 , 21 ] , the FeDet research project [ 26 , 40 , 16 ] , as well as the Solar-Check project [ 39 , 17 ] . Modelling approaches . Modelling approaches, including parity-space and physics-based methods, focus on replicating system dynamics through simulations or analytical equations. This is done either through simulation tools, simulating the expected behaviour of the system [ 7 ] , or by applying physical equations to estimate the yield of the collector loop directly [ 32 , 33 ] . Large deviations from the estimations are then flagged as faults. These approaches thus detect any