Master's Thesis Simon Thul


Application of automated Fault Detection and Diagnosis at E.ON ERC's Main Building Heat Pump

Average cross-validation scores Copyright: EBC Average cross-validation scores and standard deviation for a fault dataset constructed from experimental data

In order to achieve the goals set by the german federal government in its Klimaschutzplan 2050, direct and indirect emissions of greenhouse gases caused in the building sector have to be lowered. Therefore, the use of fossil fuels as the primary energy source for building heat should also be diminished. In Germanys future energy system, heat pumps could use electricity from wind and solar plants in order to provide building heat and cooling. In order to fully use the potential of heat pumps, it is imperative that they run efficiently.

During regular use, heat pumps often have diminished system efficiency due to some faults occuring in the system. Data-driven Fault Detection and Diagnosis (FDD) can be applied to find and repair those faults early. Data-driven FDD uses machine-learning algorithms trained on collected process data in order to detect faulty behavior.

In this thesis, the application of data-driven Fault Detection and Diagnosis is being investigated for the heat pump used in the main building at RWTH Aachen University’s E.ON Energy Research Center. As the data collected there contained too little information of faults that occured in the past, it was not possible to train and validate machine learning models from that data alone. Therefore, data provided by the National Institute of Standards and Technology (NIST) was used instead. First, the data is used to show its suitability for data-driven FDD. After that, with the variables provided in both NIST’s and ERC’S heat pump data, classification models are trained by using NIST’S heat pump data.

Predictions with those algorithms on ERC’S dataset reveal that the similarity between the two systems is too small in order to achieve good overall prediction scores. Nevertheless, some of the algorithms are able to predict some of the faults that occured, at least in certain periods of time. Therefore it is concluded, that the use of foreign datasets for fault prediction could be possible, provided that both heat pumps are similar enough.

Overall, the investigation of ERC’s heat pump data showed that merely collecting large amounts of data is insufficient for the application of data-driven FDD. Additionally, the precise labeling of faulty and fault-free system states is required to successfully apply classification methods.