Data-Driven Fault Detection and Diagnosis Approaches for Special Purpose Buildings

Abstract

Data-driven faults detection and diagnosis (FDD) emphasises operational data but does not require in-depth understanding of the system’s background; yet, significant volumes of data are essential, including the number of attributes and length of observations. Furthermore, the absence of building energy management systems (BEMS) in most special-purpose buildings such as mosques and churches impedes data/feature availability. Nonetheless, research on FDD with limited data/features especially at whole-building levels is still limited within current body of knowledge. Thus, a comprehensive FDD framework is built in this thesis to address the challenges brought on by insufficient data/features as well as the scope of the investigation. As a result, two mosques were employed as case studies to explore the deployment of FDD methodologies. The whole building modelling software (EnergyPlus) was used to model the targeted buildings as well as simulate different system faults. In addition, machine learning algorithms were trained and tested using aggregated hourly energy use and various meteorological data. FDD has two main pillars; detection and diagnosis, in terms of fault detection, firstly, a comparison study of three unsupervised anomaly detection approaches (one class support vector machine (OCSVM), Isolation forest, and LSTM-Autoencoders) was conducted for detecting faults of heating, ventilation, and air conditioning (HVAC) system with a limited amount of faulty data points in order to prove their ability in fault detection with limited faulty points. The LSTM-autoencoders outscored all other techniques with an average precision of about 83 % of all faults scenarios. Second, the forecast-based anomaly detection technique (seasonal autoregressive integrated moving average (SARIMAX)) was used to detect whole-building faults and then quantified the impacts of the identified faults on energy patterns. The SARIMAX provided outstanding detection accuracy for the majority of building faults with an average detection accuracy of roughly 85% for all faults, where it was also observed that the HVAC system faults had the highest impact on total building energy usage, ranging from 242 to 590 kWh/period. In terms of fault diagnosis, supervised multi-class classification techniques such as k-nearest neighbours (KNN), decision tree (DT), and random forest (RF) were employed to diagnose single-severity faults as well as fault free class on the entire building since no work has been reported employing the multi-class classifiers in this whole-building scale. Moreover, prior to the classification process, feature extraction techniques such as principal component analysis (PCA), kernel principal component analysis (KPCA), independent component analysis (ICA), and autoencoders (AE) were utilised to assess their efficacy in boosting the diagnostic process. With regards to validation and testing phases, the RF classifier as an ensemble approach had the highest classification accuracy without feature extraction with around 90% and the incorporation of feature extraction methods was not associated with the enhancement of the classification performance of most classifiers, indicating that some classifiers may perform better with high-dimensional datasets. Consequently, another case study was undertaken to detect and diagnose numerous types of building system faults at whole building scale and classify their severities using ensemble multi-class classification approaches, including RF, stacking classifier (SKC), gradient boosting classifier (GBC), and extreme gradient boosting (XGBoost). Throughout the validation and testing phases, the XGBoost classifier had the greatest classification accuracy for both fault categorisation and severity ratings with 96% on average. Additionally, the added statistical decomposition elements of energy usage greatly improved the learning and classification accuracy, especially for the severity classification, with roughly 33% on average.