Session: 08-03-01: Energy-Related Multidisciplinary I

Paper Number: 94599

94599 - Development of a Holistic Data-Driven Detection and Diagnosis Approach for Operational Faults in Public Buildings 

Buildings comprise a number of complex systems governed by various control strategies, such as thermostat control and occupancy sensors. Faults on such systems and control mainly occur as a consequence of poor maintenance, resulting in a significant increase in energy consumption, which leads to an increase in the shares of carbon emission of the buildings sector. Fault detection and diagnosis (FDD) approaches are well-established mechanisms for alleviating anomalies related to such inter-related building systems. FDD can be described as having two main pillars – first is to solely identify whether there is a fault or not through irregularities in the operation, which is regarded as fault detection and then secondly to determine the root causes of such faults, which is also known as faults diagnosis. The process of FDD can be generally achieved through model-based, knowledge-based, data-driven and hybrid forms. The model-based and knowledge-based forms primarily depend on the physical characteristics of the system as well as background experience. Unlike the aforementioned approaches, the data-driven approach prioritises operational data and does not require indepth knowledge of system background; nevertheless, it requires considerable amounts of data.

 Obtaining faulty building data is a significant challenge for researchers. As a result, employing simulated data can be beneficial in data-driven FDD analysis because it is inexpensive and can run multiple sorts of faults with varying severities and time periods, thereby alleviating previously experienced accessibility issues during investigations.The anomalies detection techniques such as forecast-based, dimensionality-reduction, as well as classification and regression trees (CART) are frequently used to detect the abnormalities in building consumption and detecting faults events. At the same time, classification approaches are mainly applied in the diagnosis phase. In order to improve the diagnosis accuracy, the features extraction methods might be implemented prior to the classification. Additionally, the predominant implementation of FDD techniques within the building sector are done at the system level. However, as useful as system level analysis is, typical buildings are comprised of multiple systems with their peculiar characteristics. Also, individualised system level based analysis makes it challenging and sometimes impossible to visualise system-system interactions. However, there is a glaring underrepresentation of literatures that explore the development of whole building models that diagnose faults over the entire building energy performance. Therefore, this paper presents a work to detect and diagnose building systems (HVAC, lighting, exhaust fan) faults in whole building energy performance within hot climate areas, using energy consumption and weather data. The detection process on the main building meter was conducted using LSTM-Autoencoders, and different multi-classes classification methods were compared for the diagnosis phase. Moreover, features extraction approaches were included in the comparison to quantify their performance in improving the diagnosis. The evaluation metrics in this work were recall score for detection and correct diagnosis rate (CDR), and misdiagnosis rate (MDR) for the faults diagnosis. The results of the study shows a comparison of using various techniques in the diagnosis process.

Presenting Author: Ashraf Alghanmi The University of Manchester

Presenting Author Biography: Mr Ashraf Alghanmi (BSc, MSc) is a postgraduate research student in the Department of Mechanical, Aerospace and Civil Engineering (MACE) at the University of Manchester. Ashraf obtained his Bachelor's in Mechanical Engineering from Yanbu Industrial College (2012) and a master's in Mechanical Engineering from the University of Dayton, Ohio (2015). Prior to joining MACE, Ashraf worked as a lecturer in the Mechanical Engineering Department of the Royal Commission Yanbu Colleges & Institutes (RCYCI). His current research interests are in building energy consumption, fault detection and diagnosis, HVAC systems and energy efficiency.

Authors:

Ashraf Alghanmi The University of Manchester
Akilu Yunusa-Kaltungo University of Manchester
Rodger Edwards The University of Manchester