Beruflich Dokumente
Kultur Dokumente
A practical solution for HVAC prognostics: Failure mode and effects analysis T
in building maintenance
⁎
Chunsheng Yanga, , Weiming Shena, Qiangqiang Chenb, Burak Gunayc
a
National Research Council Canada, Ottawa, Ontario, Canada
b
Nanchang University, Nanchang, China
c
Carleton University, Ottawa, Canada
A R T I C L E I N F O A B S T R A C T
Keywords: Fault detection, diagnostics, and prognostics (FDD&P) is attracting an amount of attention from building op-
FMEA erators and researchers because it can help greatly improve the performance of building operations by reducing
Work-order energy consumption for heating, ventilation and air-conditioning (HVAC) while improving occupant comfort at
Fault Detection and diagnostics (FDD) the same time. However, FDD&P, particularly HVAC prognostics, for building operations remains with many
Prognostics
challenges due to special operation environments of HVAC systems. These challenges include “tolerance or
HVAC
ignorance” of failures in long-haul operations, lack of operation regulations, and even lack of documents for
HVAC failure mode and effects analysis (FMEA), which is a systematic method of identifying and preventing
system, product and process problems. To address some of these challenges, the authors propose an FMEA
method for common building HVAC equipment by exploring work-orders generated by building energy man-
agement systems (BEMS) using a data mining approach. With this FMEA approach, it is possible for building
operators to isolate and prognose faults practically. The FMEA approach also helps us tackle high impact failures,
for which operation data can be acquired and machine learning-based predictive models can be developed. This
paper reports some preliminary results in conducting an HVAC FMEA from a large number of work-orders
obtained from a BEMS in routine operations. The HVAC FMEA will be used as a guidance tool for data gathering
and developing data-driven models for HVAC FDD&P and as a practical solution for HVAC prognostics in case
that predictive models are difficult to develop.
1. Introduction will increase the space heating loads, and if the heating coil is not sized
to compensate for this fault, indoor temperature setpoints will not be
According to Natural Resources Canada [16] and US Department of met. Aside from physical faults, inappropriate control sequencing of the
Energy [5], about 15% of the secondary energy used in North America HVAC systems can cause suboptimalities in buildings’ energy and
is dedicated for indoor climate control. Indoor climate control is re- comfort performance [2]. It is worth noting that controls programming
sponsible for over 10% of the CO2 emissions and a driver for new en- is a custom and manual process – thus prone to human errors. For ex-
ergy infrastructure. It is estimated that 25–50% of the energy use in ample, if an AHU's supply air pressure setpoint is too low, the terminal
commercial buildings in the UK is wasted due to improper operating variable air volume (VAV) units’ dampers will remain open all the time
conditions (i.e., faults) [6]. In order to optimize occupant comfort and and zone temperature setpoints cannot be met.
energy use in commercial buildings, researchers have been in- There is an extensive research literature on HVAC fault detection
vestigating methods that can detect and isolate these faults as they and diagnostics (FDD) in buildings. Existing FDD methods can be ca-
occur and prognose them before they emerge [20]. tegorized into two groups: rule-based methods [10,11] and perfor-
In buildings, physical faults (i.e., hard-faults) can occur in sensors, mance indexing-based methods. Rule-based FDD is built on a set of
meters, actuators, and equipment. Inappropriate control decisions expected operational conditions for a building automation system
executed by (or based on) faulty building systems and components can (BAS). The rules are used to generate alarms for operators to take ne-
cause occupant discomfort and/or waste energy. For example, if an air cessary actions to fix a potential problem. For instance, a rule may
handling unit's (AHU) outdoor air intake damper is stuck open, the notify an operator to pay attention to a room's terminal HVAC equip-
fraction of outdoor air in supply air will be higher than intended. This ment if the temperature of the room increased above 25 °C. Using a
⁎
Correspondence to: National Research Council Canada, 1200 Montreal Road, Ottawa, ON, Canada K1A 0R6.
E-mail address: Chunsheng.Yang@nrc.gc.ca (C. Yang).
https://doi.org/10.1016/j.jobe.2017.10.013
Received 25 July 2017; Received in revised form 29 October 2017; Accepted 31 October 2017
Available online 10 November 2017
2352-7102/ Crown Copyright © 2017 Published by Elsevier Ltd. All rights reserved.
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
BAcnet International [1] control network, one can program a system to of these challenges is that there is no failure mode and effects analysis
monitor a subsystem that have direct impacts on that room's tempera- (FMEA) tool for one to select the failure types for prognostic modelling.
ture, such that the cause of the unexpected temperature rise can be FMEA has been used for fault identification, isolation, and prevention in
monitored. However, such rules may generate too many false positives the maintenance industry as a systematic method, since it was origin-
(rule too sensitive) or too many false negatives (rule too insensitive) – ally developed by NASA to enhance the reliability of space program
because they have no connection with the physics of the problem that it hardware [4]. Theoretically, FMEA provides a foundation for qualita-
is attacking. Another limitation of rule-based FDD is that rules cannot tive reliability, maintainability, safety and logistic analysis; it docu-
be adapted automatically while building usage changes; operators have ments the relationships between failure causes and failure effects [24].
to update the rules when any changes are made in buildings. FMEA has been widely used in safety-related industries such as aero-
Performance indexing-based FDD presents a complex building per- space [26] and transportation. In building engineering and operation,
formance metric as a simple threshold. Most of these FDD methods rely there has been some research to apply FMEA to building maintenance.
on the statistical characteristics of the prediction residuals to identify For example, Lair et al. [13] and Talon et al. [19] used FMEA to define
the performance index. There are two kinds of modelling techniques to failure modes for common building products for estimating reliability
compute residuals: greybox modelling [22,28] and blackbox modelling and remaining useful life. Pride [18] provided a solution about how to
[7]. In general, the greybox models are developed as thermal network use the identified building failure mode to derive reliability centered
models with equivalent thermal resistances and capacitances. These maintenance (RCM) program. Unfortunately, there were only a few
models are generic and simplified representations of a building's ther- publications on developing FMEA for HVAC systems due to the diffi-
mophysical characteristics. With sensor and meter data, these models culties and unavailable resources since 2010. Some work attempted to
predict the thermal response of a building – whereby the differences apply FMEA method to model predictive control [30,33] and risk as-
between a model's timestep ahead predictions and the measurements sessment [32] for HVAC systems. The only work which tried to apply
are the residuals. The blackbox models [15] are data-driven models FMEA for HVAC FDD and assurance of reliability was Lv and Liang
[25] to predict or estimate the heating or cooling load by using machine [31]. A big technical gap exists between the state of the art of FMEA
learning algorithms [9,17,21,23] such as linear regression [3], artificial and HVAC maintenance. One of the motivations of this work is to
neural network (ANN), and support vector machine (SVM). Most of the bridge this gap by developing an FMEA tool for HVAC systems from
blackbox models are used to predict the energy consumption of build- work-order logs available from building operation. Therefore, the first
ings. Few are able to detect / predict failures from the system behaviour task in this work is to develop an FMEA for HVAC systems, which is
point of view. able to guide us in focusing on the high impact faults and acquire the
Recently, inverse modelling has been proposed as an approach to relevant operational data for prognostic modelling.
build blackbox and greybox models whose parameters are intended to This paper extends upon previous research [27] by providing new
characterize the performance of a building instead of estimating space results obtained from a case study in developing an FMEA for HVAC
conditioning loads. For example, the parameters of a blackbox model systems. The case study demonstrated the feasibility of an FMEA-based
which links the effect of outdoor temperature to heating load intensity HVAC prognostics methods. This FMEA for HVAC systems was devel-
can be an indicator of performance deviations. One can compare these oped upon a large work-order dataset from 44 buildings in Ottawa,
parameters across a building cluster to establish a performance Canada. The main contributions are summarized as follows. (1) This is a
benchmark. Both greybox and blackbox models are mainly used to first attempt to develop FMEA for HVAC systems from historic work-
provide the performance indices for FDD at the building level. It is order dataset. (2) A symptomatic data mining-based approach is pro-
desirable that FDD&P is able to provide a solution for building opera- posed for creating an FMEA from work-order dataset; (3) An FMEA-
tors to detect, identify, or predict failures at the system level or com- based method for prognostics of HVAC systems is proposed.
ponent level [10–12,29]. In particular, a comprehensive FDD&P solu- Following this section, the paper discusses the major challenges
tion for HVAC systems is still needed. Unfortunately, in the reviewed facing us in developing prognostic techniques for HVAC systems. Then,
literature, there were not any publications regarding prognostics for Section 3 presents an overview of the work-order data from routine
building operation. building operations or BEMS. Section 4 introduces a data mining-based
As one of the objectives of our on-going project, the authors are method for generating an FMEA for HVAC systems; Section 5 presents
contemplating to develop a novel FDD&P technology for HVAC systems. the results obtained from a case study; Section 6 discusses the results
The first objective is to develop models which can characterize the and develops recommendations for future work.
normal behaviour of a component and help us identify failures. Another
objective is to develop machine learning-based prognostic models
2. Challenges
which are able to predict a failure before it occurs and to estimate the
remaining useful lifetime, such that a proactive action can be taken to
Building operation is in many ways different from the operation of
fix the failure before it detrimentally affects energy and comfort per-
other complex systems such as mechanic-electronic systems and aircraft
formance of a building. To this end, many challenges are facing us. One
engines. It is challenging to develop FDD&P for building operation
27
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
systems. These challenges can be summarized as follows: Today, more and more BEMSs operating on top of BASs are being
developed and installed in commercial buildings to improve the
(1) Ignorance of failure features or symptoms performance of building operations by monitoring BASs more clo-
The performance of a building is usually measured or benchmarked sely and scientifically. These BEMSs collect operation data and store
by energy use intensity and occupant comfort. When HVAC systems them in a well-designed database. However, such database contains
suffer any failures, the symptoms of such failures can be easily ig- many dropping data, missing values, and incomplete data due to
nored by operators and occupants. For example, when the room different reasons such as power outage, network communication
temperature of an office is too high, occupants may open a window failures, sensor failures, etc. Sometimes, data tagging cannot be
to cool down the temperature; on the contrary, occupants may in- correctly interpreted as well. BEMS do not necessarily know the
crease their clothing insulation. Consequently, the failure of HVAC sensor label tags used by BAS vendors. So, a BEMS does not know
systems can be ignored and not reported, even though these ab- what that data represent in the BAS level and it merely records such
normal room temperatures could be an indication of a high energy data without proper interpretation. The quality of data is not
impact failure. Similarly, when an AHU's outdoor air intake damper guaranteed. For example, as mentioned above, sensors may be in-
is stuck open, the space conditioning loads are expected to increase. stalled at improper locations and not correctly calibrated, the data
However, with the redundancy in HVAC subsystems (e.g., heating/ from such sensors are not trustable, which introduces further
cooling coils, zone reheat coils and air dampers, perimeter heaters) challenges to the data analysis for failure effects and modelling for
and the practice of oversizing these subsystems, the indoor tem- FDD&P.
peratures can still be maintained within operating setpoints. And,
abnormal operation with damper stuck fault may last for a long These challenges slow down and even block the development of
period without being noticed until an energy audit is performed. FDD&P technologies for building operations, particularly the develop-
Based on authors' experience in working with several large building ment of prognostics technology. For example, due to the lack of an
clusters, collecting normal operation data from building systems is FMEA tool, it is difficult to identify the high impact faults that one
very difficult because the ignorance of the faults directly affects the should pay more attention when developing FDD&P algorithms or
determination of its onset point (a starting point for failure data). models. This work aims at addressing this issue by developing an FMEA
(2) Improper calibration of sensing systems for HVAC systems from work-orders and operation data collected
In existing buildings, the commissioning work in automation and during routine building operations. The developed FMEA will provide a
control systems is typically conducted before and/or just after the practical solution for HVAC prognostics in building operation.
building is occupied. Continuous commissioning and post-occu-
pancy commissioning are seldom employed [14]. It is worth noting 3. Work-order overview
that there are no regulations to enforce to monitor building systems’
operation by installing proper sensors. Therefore, some sensors Modern building operations generate a large number of work-orders
were sometimes installed improperly. And, the sensor data were in routine operations and maintenance. These work-orders contain
never validated or checked routinely. Some sensors may defunct, some valuable information, including dates, fault symptoms, fault
and they may remain unnoticed for a long period. Therefore, the types, problem descriptions, actions taken to fix the problems, fault
data collected from such sensors may not be trustable and ac- locations (sub-systems or components), causes, etc. Currently there are
countable. two kinds of work-orders: operators’ logbooks and event reports from
(3) Lack of failure datasets BEMSs. This study explored the work-orders collected from a university
Buildings remain operational despite some failures causing loca- campus. It provides much richer and valuable information for devel-
lized discomfort and energy inefficiencies. Such failures may cause oping FMEA. An illustrative example for such a work-order is shown as
performance deteriorations, but buildings almost never stop being Table 1. The main fields are “description”, “priority”, “date/time’, and
operational during their service life. As most failures were not re- “status”. The meaning of each field is listed in Table 1. In ongoing
ported for years, it is quite challenging to establish these datasets. commissioning of buildings, large amounts of such building main-
Therefore, there are not many failures reported in the routine tenance data are generated every day. Many of them are in regard to
building operation. This creates a challenge for us to obtain com- routine operations such as changing lights, fixing door, and changing
prehensive fault symptom datasets from existing BAS. locks, and so on; only a tiny fraction of the work-orders relates to HVAC
(4) Lack of knowledge of failure mode and its effect on HVAC systems operation and maintenance.
In maintenance of complex systems there usually is a well-defined
maintenance vehicle called FMEA which is a useful tool for op- 4. FMEA for HVAC
erators to perform systems maintenance – i.e., identifying failures
and isolating the faults at the system or component level. However, By definition, an FMEA contains useful information presented as
in the buildings sector, an FMEA tool does not exist for building parameters, including Critical Severity (CS) of a failure, Failure Rate
operators to perform failure analysis and identify the failures when
a building system suffers any failures. Without FMEA it is also a Table 1
challenge for us to concentrate on the vital failures and to gather A work-order sample.
relevant historic operation data to develop models for prognostics
Item Meaning
of HVAC systems.
(5) Lack of operation data Work order ID Record the ID of work order. For example, 3789
Existing BASs are installed from original builders or facility op- Type The type of work order
erators for regular operations. They provide temporary storage in Location Operation zone code
Description AHU5 appears to be experiencing a return
terms of trend logs. However, the buffer to store such data is limited AirLoeb AHU2, have Brian verify that the cold deck
in a small size (e.g., 2k bytes). Therefore, large amounts of data Temperature on this unit is calibrated and in use.
have been lost and not collected routinely. Permanent data ar- Job Assignment Who is assigned for the work
chiving is a new concept in building operations, and in many Status Case statue, for example, close
Priority The work importance, (1, 2,..0.5)
buildings, meter data instead of HVAC operational data have been
Created Time 2015-02-09T10:52:48-05:00
collected. Updated Time 2015-12-29T16:17:19-05:00
(6) Incomplete data and low-quality data
28
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
(FR), Failure Mode Probability (FMP), and Mean Time between Failure • Failure category
(MTBF), for determining the effects of each failure mode on system • Component (location)
performance. These parameters are useful in determining the policy of • Failure mode
maintenance either for fault isolation or fault prevention. Based on the • Failure effects
standard FMEA definition from NASA [4], the authors propose to create • CS
an FMEA for HVAC systems in building operation. The following is the • FR
definition of the FMEA parameters and methods for computing them • FMP
from the building work-orders. • MTBF
• RPN
4.1. Definition of main FMEA parameters
The main task is to find out the FMEA parameters defined above
Focusing on the work-orders as described above, the authors de- from the work-order dataset based on the proposed procedure as shown
veloped an FMEA for HVAC systems by generating FMEA parameters: as Fig. 1. This requires a large amount of data to statistically estimate
CS, FR, FMP, MTBF, and Risk Priority Number (RPN). These parameters these parameters. Some parameters such as MTBF require more historic
are defined in Table 2 and expressed using Eqs. (1)–(5). data. Operational data are mainly used to compute and evaluate the CS
by estimating the impact on energy consumption or thermal comfort. In
1 n
CS =
Nl
∑i =1 Ci Nli (1)
this work, only FMEA parameters defined in Table 2 are computed.
Other information such as location, failure mode and failure effect are
NFc textual data derived from work-orders directly.
FR =
UT (2)
• NFc: The number of failures of a given component in HVAC systems; This work uses a work-order dataset collected from 44 buildings in a
• UT: The total operating hours of a HVAC in a building; university campus, Ottawa, Canada over seven years. Among 44
• RN: The total number of failures observed in the work-orders; buildings, 16 were used for academic activities, two for administrative
• T : The running time for failure l at time i; and
li
activities, six for athletics, four for laboratories, and nine for student
• N : The occurrence times of failure l.
l
housing. The rest were ancillary buildings such as cafeterias, service
restaurants and a clinic. The buildings were built between the 1960s
In developing an FMEA for complex systems such as an aircraft, and 2010s with wide range of the sizes. The detailed characteristics of
usually, FMEA documents are generated during the system design. the building are referred in Gunay et al. [8].
Therefore, the parameters above are estimated based on the safety re- The work-order records collected in routine building operation
quirements and accelerated test of systems and physical components. contained brief descriptions created by different operators regarding
For HVAC systems, there is no such regulation to require industry to the nature of the maintenance work, the building name, and a time-
generate FMEA documents during the design phase. Therefore, it is stamp as shown in Table 1. In addition, a work-order type label (e.g.,
necessary to compute these parameters from a large number of work- carpentry, administrative services) was attached to each log. Over
orders. It assumes that it is possible to obtain sufficient information seven years, a total of 87,865 work-orders were logged with custom
from the work-orders. For example, in the case study, the work-orders descriptions. Of them, 20% were related to HVAC, and room lighting
over seven years have been collected. Technically to analyse the FMEA and power plugs. The rest were other facilities management activities
parameters, it is desirable that the operation data corresponding to such as administrative services, carpentry, locksmith, furniture moving,
work-orders data are available. For instance, to find out CS it is required plumbing, water treatment, vehicle maintenance, painting, and custo-
to gather building operation information such as energy consumption dial services. This paper is concentrated on the work-orders that were
and maintenance costs. However, some work-orders provide the in- issued to address problems in HVAC, and room lighting and power
formation directly by giving the critical level of failure events. In this
work, such information is used to compute the CS instead of operation Table 2
The definition of FMEA parameters.
data.
Failure parameter Definition
4.2. The proposed FMEA for HVAC
CS Defining the impact and severity of a failure, which can be
computed based on the failure priority and importance
As mentioned above, the proposed FMEA focuses on HVAC at the FR Defining the rate that a failure occurs in 1000 h
system level or component level. From the view point of failure mode, FMP Defining the probability of a failure occurrence given a
the proposed FMEA only deals with failure modes of hard faults in component
building operation. The soft faults and building envelope failures are MTBF The average time between two failures occurring on a
component or subsystem
beyond the scope of this paper. In terms of the definition of FMEA
RPN Defining the priority of a failure or component by computing
parameters, our proposed FMEA for HVAC systems contains the fol- the products of CS, FR, and FMP
lowing main elements for a given failure mode or component:
29
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
systems
Work-order dataset
computation Building
For HVAC
Operation Data
MTBF)
plugs. Hereon, the term work-order refers to this group only. How to of the ranking numbers (last column). This list provides useful in-
extract such work-orders for creating FMEA will be described. formation for facility managers to look into these failures. At least it
indicates an order of components to investigate when they receive a
failure report.
5.2. Generation of FMEA
Following the proposed methods described above, an FMEA docu- 5.3. Use case of FMEA in building maintenance
ment for HVAC systems and components is developed using the work-
order dataset. The first step was to extract the HVAC-related work-or- The original purpose of developing FMEA for complex systems in
ders from a large dataset. This work is focused on the subsystems and reliability research community is to provide a vehicle for maintainers to
components of common HVAC systems in buildings. Therefore, the perform fault isolation and fault prevention given a failure model.
keywords that represent the component names were used to retrieve Therefore, the FMEA parameters can be used to help isolate faults by
relevant work-orders from the dataset. The keywords include %AHU%, ranking fault detection models [26] and failure prediction by tracking
%vent%, %Vent%, % Humidifier %, %a/c%, %Cold Air%, %AHU%, % the remaining useful life for a component based on MTBF. Simply put,
temperature%, %Fan%, %fan%, %Damper%, %damper%, %VAV boxes FMEA has two main applications in maintenance industry: fault iden-
%, %Vav boxes%, %Valve%, %valve%, %leaking%, %Leaking%, tification/isolation and fault prognostics. Therefore, in building op-
%Controller%, %controller%, %Coil%, %coil%, %chiller%, %Chiller%, eration environment, FMEA can also be applied to fault identification/
%Boiler%, and %boiler%. From 87,865 work-order records, 6298 re- isolation and HVAC prognostics.
cords related to HVAC are extracted. The distribution of failure events
Table 3
for each component is shown as in Table 3. The priority numbers re-
the distribution of 6298 HVAC work-orders.
present the relative importance of each event. For example, given
equipment “FAN” in the HVAC systems, there were 1695 failure events Components Count Priority set for maintenance
over seven years operation. Among these failure events, priority num-
bers are set up at six levels. In other words, there were 59 events for 1 2 3 4 5 6
priority level one, and five events for level 2, 645 events for level 3, and FAN(VSD) 1695 59 5 645 28 0 958
so on. As expressed in Eq. (1), the priority level information is used for VALVE 1015 22 5 461 28 4 495
the 6298 work-orders related to HVAC. The FMEA parameters are PUMPS 773 36 3 392 3 3 336
computed using the Eqs. (1)–(5). The Table 4 shows the FMEA para- HUMIDIFIER 580 39 5 281 10 0 245
DAMPER 541 5 1 69 6 2 458
meters obtained from the work-order dataset. With this dataset, it was CHILLER 477 11 7 151 7 0 301
impossible to compute MTBF at this stage because the dataset contains COIL 448 2 0 86 1 0 359
insufficient data to statistically analyse MTBF yet. For instance, the AHU OTHER 426 14 1 229 12 0 170
existing work-order dataset is component or subsystem-based main- BOILER 264 3 1 111 3 0 146
VENT 58 4 1 33 1 0 19
tenance records. To compute MTBF, failure modes oriented work-or-
CONTROLLER 15 0 0 6 0 0 9
ders are required. VAV 6 0 0 4 0 0 6
In Table 4, the bold rows represent the top five components in terms
30
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
Table 4
the FMEA created for HVAC systems and components.
CS FR FMP RPN
31
C. Yang et al. Journal of Building Engineering 15 (2018) 26–32
possible to compute this powerful parameter for FMEA from the current [3] T. Catalina, J. Virgone, E. Blanco, Development and validation of regression models
work-orders dataset. As mentioned above, this parameter is extremely to predict monthly heating demand for residential buildings, Energy Build. 40
(2008) 1825–1832.
important for performing prognostics. To compute this parameter, [4] H.C. Chen, Failure Modes and Effects Analysis Training Manual, Personal commu-
failure mode-based FMEA is absolutely requested. This is planned as nication, Hen Technology Inc, USA, 1996.
future work. The work-orders will be categorized into groups corre- [5] DOE, Building Energy Data Book, Department of Energy, United States, 2011.
[6] G.H. Gibas, Building Control Systems, Butterworth-Heierman, Oxford, 2000.
sponding to distinct failure modes. Then, an FMEA for each failure [7] H.B. Gunay, W. O’Brien, I. Beausoleil-Morrison, Control-oriented inverse modeling
mode can be created. With such a failure mode-based FMEA, it will of the themal charateristics in an office, Sci. Technol. Built Environ. (2016) (online
likely be easier for operators to conduct prognostics for HVAC systems. first).
[8] B. Gunay, W. Shen, C. Yang, Characterization of a building's operation using au-
tomation data: a review and case study, J. Build. Environ. 118 (2017) 196–210.
6.4. FMEA evaluation [9] M. He, W.J. Cai, S.Y. Li, Multiple fuzzy model-based temperature predictive control
for HVAC systems, Inf. Sci. 169 (2005) 155–174.
[10] S. Katipamula, M.R. Brambley, Methods for fault detection, diagnostics, and prog-
In reliability community, an FMEA is created as part of the system
nostics for bui8llding systems – a review, Part I, HAVC R Res. 11 (1) (2005) 3–25.
design to meet certain safety requirements. It is not validated until the [11] S. Katipamula, M.R. Brambley, Methods for fault detection, diagnostics, and prog-
system is put into real use. The FMEA validation, in which the main task nostics for bui8llding systems – a review, Part II, HAVC R Res. 11 (2) (2005)
is to validate the correction of each FMEA parameter, is based on the 169–197.
[12] J. Liang, R. Du, Model-based fault detection and diagosis of HVAC systems using
data collected in practice. In the practice of building maintenance, not support vector machine method, Int. J. Refrig. 30 (2007) 1104–1114.
all work-orders were recorded scientifically. Some of them may be is- [13] J. Lair, J. Le Teno, D. Boissie, Durability assessment of building products, in: A.
sued verbally on the site. Such work-orders may not be recorded in a Sarja (Ed.), Proccedings of the International RILEM/CIB/ISO Symposium on
Integrated Life Cycle Design of Materials and Structure, 2000, pp. 382–387.
formal work-order management system. On the other hand, most of the [14] E. Mills, Building Commissioning: A Golden Opportunity for Reducing Energy Costs
work-orders at the zone level were issued upon occupants’ complaints, and Greenhouse-gas Emissions, Lawrence Berkeley National Laboratory, 2010.
such as too hot, too cold, too stuffy, and too noisy. However, as dis- [15] M. Najafi, Fault Detection and Diagnosis in Building HVAC Systems (Ph.D. thesis),
University of California, Berkeley, 2010.
cussed in Section 2, some faults may not be reported by occupants since [16] NRCan, Energy Use Data Handbook 1990 to 2010, National Resurce Canada, 2013.
the discomfort level is not high enough to trigger complaint calls. These [17] T. Olofsson, S. Andersson, R.A. Ostin, Energy load prediction for buildings based on
situations represent further challenges for researchers to create a new a total demand perspective, Energy Build. 28 (1998) 109–116.
[18] A. Pride, Reliability Centered Maintenace (RCM)”, Whole Building Design Guild,
FMEA or to verify an existing FMEA. In this work, it is impossible to National Institute of Building Science, 2000, www.wbdg.org.
have an extra dataset for FMEA validation. Thus, the FMEA validation is [19] A. Talon, D. Boissier, J. Chevalier, J. Hans, Temporal quantification method of
left for future work. Considering these limitations, the readers should be degradation scenarios based on FMEA, in: Proceedings of the intenational con-
ferecnece on durability of building materials and components, Lyon, France, 2005.
cautious in extending the FMEA results of this study to perform routine
[20] TIAX LLC, Energy Impact of Commercial Building Controls and Performance
FDD&P tasks. Diagnostics: Market Characterization, Energy Impact of Building Faults and Energy
Savings Potential, 2005.
7. Conclusion [21] G.K.F. Tso, K.K.W. Yau, Predicting electricity energy consumption: a comparison of
regression analysis, decision tree and neural networks, Energy 32 (2007)
1761–1768.
In this paper, the authors put forward a failure mode and effects [22] S. Wang, Q. Zhoy, F. Xiao, A system-level fault detection and diagnosis strategy for
analysis (FMEA) method for building HVAC systems in order to provide HVAC systems involving sensor faults, Energy Build. 42 (2010) 477–480.
[23] M. Yalcintas, S. Akkurt, Artificial neural networks applications in building energy
a new and robust solution for HVAC prognostics. The proposed method predictions and a case study for tropical climates, Int. J. Energy Res. 29 (2005)
for generating an FMEA from a building cluster's work-orders was 891–901.
presented. It addressed some of the current challenges while employing [24] C. Yang, S. Létourneau, J. Liu, Q. Cheng, Y. Yang, Machine learning-based methods
for TTF estimation with application to APU prognostics, Int. J. Appl. Intell. (2016),
FDD&P strategies in building HVAC systems and components. Using a http://dx.doi.org/10.1007/s10489-016-0829-4 (Online first).
work-order dataset obtained from 44 buildings in Ottawa, Canada, the [25] C. Yang, S. Létourneau, H. Guo, Developing data-driven models to predict BEMS
work demonstrated how a component-based FMEA was generated fol- energy consumption for demand response systems, in: Proceedings of the 27th
International Conference on Industrial & Engineering Applications of Artificial
lowing the proposed procedure. The results demonstrated that the de-
Intelligence & Expert Systems (IEA/AIE-2014), Taiwan, 2014.
veloped FMEA represents a useful tool for building operators to perform [26] C. Yang, S. Letourneau, Y. Yang, J. Liu, Data Mining-based fault isolation with
HVAC fault isolation and prognostics. Because of the limitations of the FMEA rank: a case study of APU fault identification, in: Proceedings of 2013 IEEE
International Conference on Prognostics and Health Management, Washington DC,
dataset, a powerful parameter, mean time between failures, was not
USA, 2013.
estimated, and this is left for future work. With the developed FMEA, [27] C. Yang, W. Shen, Q. Chen, B. Gunay, Toward Failure mode and effect analysis for
researchers can focus on high-impact failures on HVAC systems and heating, ventilation and air-conditioning, in: Proceedings of the 20th IEEE
develop advanced algorithms and models for FDD&P of HVAC systems International Conference on Computer Supported Cooperative Work in Design
(CSCWD 2017), Wellington, NZ, 2017, pp. 409–413.
in building operation. [28] B. Yu, D.H.C. Van Paassen, General modeling for model-based FDD on building
HVAC system, Simul. Pract. Theory 9 (2002) 376–397.
Acknowledgement [29] Y. Yu, D. Woradechiumroen, D. Yu, A review of fault detection and diagnosis
methodologies on air-hnadling units, Energy Build. 82 (2014) 550–562.
[30] A. Afram, F. Janabi-Sharifi, Theroy and application of HVAC control systems – a
This work is supported by Natural Resource Canada under Program review of model predictive control(MPC), Build. Environ. 72 (2014) 343–355.
of Energy Research and Development (PERD 006605.03). [31] Y. Lv, Y. Liang, Application of FMEA based on fuzzy multi-criteria decision-making
for HVAC in a pharmaceutical plant, J. Chem. Phaemaceeutical Res. 6 (6) (2014)
1116–1123.
References [32] A. Shukla, A. Katole, N. Jain, C. Karthikeyan, F. Methta, P. Trivedi, A risk assess-
ment approach: qualification of a HVAC system in aseptic processing area using
building management system, J. Qual. Assur. 44 (2011) (2011) 40–49.
[1] BAcnet International, Intruduction to BACnet for Building Owners and Engineers,
[33] R. Kwadzogzh, M. Zhou, S. Li, Model predictive control for HVAC systems – a re-
Available at: 〈http;//www.bacnetinternational.org〉.
view, in: Proceedings of 2013 IEEE Intenational Conferecne on Automation Science
[2] Q. Bi, W.J. Cai, Q.G. Wang, et al., Advanced controller auto-tuning and its appli-
and Engineering 9CASE, Madison, WI, USA, 2013, pp. 442–447.
cation in HAVC systems, Control Eng. Pract. 8 (2000) 633–644.
32