Real-world application of machine-learning-based fault detection trained with experimental data
Introduction
Buildings today account for up to 40% of the final energy demand in the EU and the US [1,2]. In Germany, their consumption is responsible for up to 30% of greenhouse gas emissions [3]. A big share of these emissions is directly linked to space heating and cooling. Heat pumps are able to provide heating and cooling power using electric energy from renewable energy sources. Hence, an increase in the number of installed of heat pumps is expected to reach the reduction goals set by the German Government [4].
To fully exploit the saving potential, the operation of heat pumps needs to be as efficient as possible. However, studies have shown that up to 40% of the energy consumption of a system is incurred due to faults occurring during the operation of energy systems [[5], [6], [7]].
Major faults completely stop the operation of the energy systems and are noticed very easily, either by simple alarm systems or if the desired energy services (heating or cooling) are not delivered to the user. In this case, since the system is not working, it does not consume energy in this state, and the downtime will not be considered when calculating key performance indicators.
Minor faults in the system, like fouling in heat exchangers or small leakages, will, however, not stop the system operation. As long as the system continues to deliver the required energy service, the user may be less impacted, and therefore not intervene. The impact of the faults on the system may, however, push the operating point out of the desired conditions into an inefficient system state, thereby increasing the energy consumption of the system. These kinds of faults unfortunately tend to go unnoticed until the next maintenance or energy bill [6,8].
In order to detect these faults before they affect the long-time system performance, many methods of automated fault detection and diagnosis (FDD) have been developed. However, recent literature reviews and market studies have found that very few of these techniques and algorithms have been tested on real-world data and ever see any application outside the science domain [[9], [10], [11], [12], [13]].
In this paper, we investigate whether we can transfer FDD systems, that were trained using data from laboratory experiments, to a real-world application in an office building. In the next section, we describe both the laboratory training data set as well as the data set collected from the real-world application. In section 3, we explain how we trained the FDD system and what steps are necessary to apply it to a different data set. The results of the FDD training and testing are presented in section 5.
For the implementation of the algorithms we used the implementation provided in the scikit learn package [14]. All visualisations were done with the matplotlib library [15].
In recent years, promising results have been achieved with methods from the field of machine learning (ML) [7,[16], [17], [18]]. In these methods, data is used for the training of algorithms to detect and identify faults. Their main drawback is the need for extensive training databases including the faults to be considered in the final system. Hence, many of the developed methods work on a specific development and test case, but have not been transferred into real-world application, due to the vast amount of different energy systems, configurations and boundary conditions found in the building sector [19]. Additionally, the sensor networks used to monitor the operation, and therefore generating the input of the algorithms, are usually designed based on control requirements and financial considerations and not may not fit the requirements of FDD algorithms [20].
Instead, many commercially available FDD systems still rely on simple rules and thresholds to generate alarm messages if certain values exceed a predefined value [21]. [22] highlights the insufficiency of available methods to detect faults at early stages. However, depending on the system engineer defining these values, the resulting FDD system may perform poorly, because the thresholds set are unsuitable for inefficient system states to be recognized, or too close to the normal operation range, so that the operator is overwhelmed with alarm messages while the system is still in an acceptable state [23]. For larger systems, the identification and localization of faults becomes even more difficult and expensive [24].
[25] investigate quantitative methods for refrigerant charge faults in heat pumps using convolutional neural networks. The aim of the study was to provide more insights into the occurred fault by predicting the amount of refrigerant. Refrigerant charge faults on experimental data gathered in a R410A heat pump were also investigated regarding their effects on heat pump performance [26]. The authors of this study use correlations in temperature conditions and a model of the refrigerant charge ratio detect faults. It is not investigated how the approach would perform on data from a different heat pump or under new operation conditions.
A method using incomplete data using back-propagation and maximum likelihood estimation to learn parameters of Bayesian networks. They applied the resulting network to a solar-assisted heat pump was proposed by Refs. [27]. Bayesian networks are also employed by Ref. [28] for chiller fault analysis. They evaluate their method using the ASHRAE RP-1043 experimental data set. Using combined Baysiean networks [29], were able to achieve accurate detection results for single faults, but not for multiple faults occurring at the same time. This lead the authors to the conclusion that multiple sensor informations are required to correctly asses the state of a system.
[30] employ clustering-based principal component analysis (PCA) models in sensor fault detection for water source heat pump systems. They tested their algorithm on experimental data. They are able to enhance the sensitivity of the detection compared to traditional PCA models.
Fault detection of a sewage heat pump system were detected using multi-mode PCA and Gaussian mixture models by Ref. [31]. They find their algorithm to be more accurate than conventional clustering algorithms using real measurement data for their evaluation.
Section snippets
Method
The aim of our investigations is to find a working fault detection and diagnosis algorithm for the heat pump of the E. On ERC main building. To apply an algorithm from the literature, a training data set is needed, which contains labels for fault data. While we do have access to several years of operation data from the building, we do not have the required set of labels. On the other hand, we cannot assume fault-free operation, since we know that the heat pump has shut down on several
NIST data set
The experimental data set used in this paper was kindly provided by the National Institute of Standards and Technology (NIST) [41,42]. They investigated the operation of an unitary split heat pump for residential applications. The unit used a thermostatic expansion valve to expand the refrigerant R410A. The system power rating was 8.8 kW for nominal cooling capacity. For more details on the device and the experimental set-up please refer to Ref. [41]. The following faults were artificially
Training on experimental data
Fig. 5 shows the result of the models prediction on the test data set for models using different sets of features. In almost all cases, feature selection improves the result of the prediction compared to using all available input features. The best results are obtained using LR-RFECV, hence this method is investigated further.
In Fig. 6, we show the Accuracy score and the MCC score of the model prediction. To reduce the random influence of the split in training and test data, we use 5-fold
Conclusion
In this paper, we investigated the transfer of machine learning models for fault detection and diagnosis trained on an experimental data set to a real-world building data set, to investigate whether models were transferable and if experimental data sets could mitigate the problem of missing or incomplete training data from buildings. To this end, we used experimental data of a heat pump with artificially induced faults to train and optimize several algorithms for fault detection. The algorithms
CRediT authorship contribution statement
Gerrit Bode: Conceptualization, Methodology, Investigation, Writing - original draft, Visualization. Simon Thul: Methodology, Software, Formal analysis, Investigation. Marc Baranski: Writing - review & editing, Project administration, Funding acquisition. Dirk Müller: Supervision, Project administration, Funding acquisition.
Declaration of competing interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
We gratefully acknowledge the financial support provided by the BMWi (Federal Ministry for Economic Affairs and Energy), promotional reference 03SBE006A. We would like to further thanks the National Institute of Standards and Technology in Maryland, USA for kindly providing the experimental data set.
References (43)
- et al.
Building energy-consumption status worldwide and the state-of-the-art technologies for zero-energy buildings during the past decade
Energy Build
(2016) - et al.
Analysis of an information monitoring and diagnostic system to improve building operations
Energy Build
(2001) - et al.
A review of fault detection and diagnosis methodologies on air-handling units
Energy Build
(2014) - et al.
Artificial intelligence-based fault detection and diagnosis methods for building energy systems: advantages, challenges and the future
Renew Sustain Energy Rev
(2019) - et al.
Data-driven fault detection and diagnosis for HVAC water chillers
Contr Eng Pract
(2016) - et al.
Knowledge discovery of data-driven-based fault diagnostics for building energy systems: a case study of the building variable refrigerant flow system
Energy
(2019) - et al.
Application of machine learning in the fault diagnostics of air handling units
Appl Energy
(2012) - et al.
Residential HVAC fault detection using a system identification approach
Energy Build
(2017) - et al.
Life-cycle maintenance cost analysis framework considering time-dependent false and missed alarms for fault diagnosis
Reliab Eng Syst Saf
(2019) - et al.
Faults in district heating customer installations and ways to approach them: experiences from Swedish utilities
Energy
(2019)
Refrigerant charge fault detection method of air source heat pump system using convolutional neural network for energy saving
Energy
Fault diagnosis for a solar assisted heat pump system under incomplete data and expert knowledge
Energy
A practical chiller fault diagnosis method based on discrete Bayesian network
Int J Refrig
Multi-source information fusion based fault diagnosis of ground-source heat pump using Bayesian network
Appl Energy
Sensor fault detection and diagnosis for a water source heat pump air-conditioning system based on PCA and preprocessed by combined clustering
Appl Therm Eng
Alarm safety and alarm fatigue
Clin Perinatol
A call to alarms: current state and future directions in the battle against alarm fatigue
J Electrocardiol
Comparison of the predicted and observed secondary structure of T4 phage lysozyme
Biochim Biophys Acta Protein Struct
Performance of a residential heat pump operating in the cooling mode with single faults imposed
Appl Therm Eng
Mode and storage load based control of a complex building system with a geothermal field
Energy Build
Energy Technology perspectives 2017: catalysing energy Technology transformations
Cited by (63)
Data-driven approach for the detection of faults in district heating networks
2024, Sustainable Energy, Grids and NetworksMITDCNN: A multi-modal input Transformer-based deep convolutional neural network for misfire signal detection in high-noise diesel engines
2024, Expert Systems with ApplicationsA combined genetic algorithm and active learning approach to build and test surrogate models in Process Systems Engineering
2024, Computers and Chemical EngineeringEnabling fire source localization in building fire emergencies with a machine learning-based inverse modeling approach
2023, Journal of Building Engineering