Residual Analysis of Predictive Modelling Data for Automated Fault Detection in Building’s Heating, Ventilation and Air Conditioning Systems

Abstract

Faults in Heating, Ventilation and Air Conditioning (HVAC) systems affect the energy efficiency of buildings. To date, there rarely exist methods to detect and diagnose faults during the operation of buildings that are both cost-effective and sufficient accurate. This study presents a method that uses artificial intelligence to automate the detection of faults in HVAC systems. The automated fault detection is based on a residual analysis of the predicted total heating power and the actual total heating power using an algorithm that aims to find an optimal decision rule for the determination of faults. The data for this study was provided by a detailed simulation of a residential case study house. A machine learning model and an ARX model predict the building operation. The model for fault detection is trained on a fault-free data set and then tested with a faulty operation. The algorithm for an optimal decision rule uses various statistical tests of residual properties such as the Sign Test, the Turning Point Test, the Box-Pierce Test and the Bartels-Rank Test. The results show that it is possible to predict faults for both known faults and unknown faults. The challenge is to find the optimal algorithm to determine the best decision rules. In the outlook of this study, further methods are presented that aim to solve this challenge.

A Review of Data-Driven Approaches and Techniques for Fault Detection and Diagnosis in HVAC Systems

“…In addition, various Bayes-based algorithms have been mentioned as successfully applied in the selected literature: Bayesian classifier [ 90 , 92 ], diagnostic Bayesian network [ 86 , 94 ], Bayesian inference with Markov Chain Monte Carlo [ 85 ], and Naive Bayes [ 90 ] with combination of decision trees (DTs) and RF [ 55 ]. Other popular models were: DTs elaborated in [ 88 , 95 ], RF [ 68 , 87 ], a hybrid RF with SVM [ 59 ], classifier chains integrated with RF [ 58 ], and multi-class SVM [ 77 ].…”

“…According to the work reviewed, it is common to use physics-based models to generate synthetic data (both faulty and healthy data), followed by data-driven modeling, i.e., supervised learning of fault prediction models [ 80 , 81 , 87 , 88 , 91 , 92 , 93 , 102 ]. Simple physical models were used by Chintala et al [ 104 ] where the Kalman filter in EnergyPlus was tested using only thermostat and outdoor temperature to perform FDD of equipment deterioration.…”

“…The use of virtual/soft sensors [ 80 , 81 , 92 ] and residual analysis [ 87 , 88 , 93 , 101 , 103 , 105 ] are also techniques worth mentioning in the context of the hybrid approach. All examples from the selected literature dealing with the hybrid approach are listed in Table 7 .…”

“…RF is more difficult to interpret, but has minimal problems with overfitting and provides an efficient nonlinear modeling strategy. Parzinger et al [ 87 ] extended their earlier research by developing an algorithm to determine the best decision rule for determining errors. Wu et al [ 58 ] employed a novel hybrid method of classifier chains with integrated RF method (CC-RF) to treat concurrent faults in RLT units as a multi-label problem.…”

Anomaly Detection in Photovoltaic Production Factories via Monte Carlo Pre-Processed Principal Component Analysis

“…In particular, two PCA models were trained, respectively, on the IQR and MC pre-processed datasets. Both models highlighted an anomalous condition almost two weeks before the equipment failure by producing KPIs (residuals) above a reference threshold which was used to discriminate between healthy and anomalous states of the equipment as done in [36].…”