Generation of training data for fault detection and diagnosis algorithms using fault simulation and parameter uncertainty

• Erzeugung von Trainingsdaten für Fehlererkennungs- und Fehlerdiagnosealgorithmen durch Fehlersimulation und Parameterunsicherheit

Bode, Gerrit Thorben; Müller, Dirk (Thesis advisor); Monti, Antonello (Thesis advisor)

Aachen : E.ON Energy Research Center, RWTH Aachen University (2023)
Book, Dissertation / PhD Thesis

In: E.On Energy Research Center : EBC, Energy efficient buildings and indoor climate 113
Page(s)/Article-Nr.: 1 Online-Ressource : Illustrationen

Dissertation, RWTH Aachen University, 2023

Abstract

Over the past decades, many efforts have been made to reduce the consumption of building energy systems. However, upon closer inspection of the building stock, many newly constructed or retro-fitted buildings, do not perform as expected due to faults in construction and equipment. To overcome this problem, methods have been developed to automatically detect faults occurring in the system. In the past years, the focus of research has shifted to applications of machine learning, as the data collected in buildings has risen exponentially, while advances in computing capabilities as well as machine learning research have made advanced data science widely available to non-domain experts. However, the application of these methods requires highly specific training data, which is difficult to obtain for individual and complex building energy systems. The use of simulation models as a data source has been proposed, but did not find application due to the complex and time-intensive modelling process. In this thesis, I investigate how the use of automated model generation and uncertainty in the parameter sets can be applied to generate training data. First, I introduce a fault model library that can be automatically combined with pre-existing building energy system models to enable the simulation of fault behaviour. I then use Uncertainty Analysis to investigate if the application of uncertainty in model parameters during the training process can be used to increase independence of the fault detection and diagnosis (FDD) model from these parameters. Subsequently, I train three common machine learning approaches for FDD with data created with an increasing number of uncertain parameter sets. Finally, I present a field test where I evaluate the proposed method on data collected from a test bench. The developed fault models are capable of qualitatively representing the fault behaviour, and the simulation model of the test bench is subsequently able to correctly represent the unit's reaction to the fault. During the simulation trial, the increasing the number of uncertain parameter sets during the creation of the training data significantly increased the performance of the FDD models. However, during the field test, no such effect could be observed. Generally, the performance on the field test data is comparable to the benchmark performance. In conclusion, generating training data with simulation model is a viable approach to mitigate the problem of missing training data. Automatically adding faults from a library to existing models reduces the manual effort greatly. The use of more parameter sets does not lead to an improvement in performance at the moment, but may become viable once the general FDD performance increases. The investigated FDD model architectures are suitable for simulation data, but need to be further improved for the use on real measurement data.