Documente Academic
Documente Profesional
Documente Cultură
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
Department of Mechanical Engineering, University of California, Berkeley, California 94720, United States
Computer Science Division and Department of Statistics, University of California, Berkeley, California 94720, United States
c
Commercial Building Systems Group, Lawrence Berkeley National Laboratory, California 94720, United States
d
Airow and Pollutant Transport Group, Lawrence Berkeley National Laboratory, California 94720, United States
b
a r t i c l e
i n f o
Article history:
Received 16 August 2011
Received in revised form 9 January 2012
Accepted 20 February 2012
Available online 27 March 2012
Keywords:
Bayesian network
HVAC systems
Air-handling unit
Energy management
Fault detection and diagnosis
Machine learning
a b s t r a c t
An air handling units energy usage can vary from the original design as components fail or fault dampers leak or fail to open/close, valves get stuck, and so on. Such problems do not necessarily result in occupant complaints and, consequently, are not even recognized to have occurred. In spite of recent progress
in the research and development of diagnostic solutions for air handling units, there is still a lack of reliable, scalable, and affordable diagnostic solutions for such systems. Modeling limitations, measurement
constraints, and the complexity of concurrent faults are the main challenges in air handling unit diagnostics. The focus of this paper is on developing diagnostic algorithms for air handling units that can address
such constraints more effectively by systematically employing machine-learning techniques. The proposed algorithms are based on analyzing the observed behavior of the system and comparing it with a
set of behavioral patterns generated based on various faulty conditions. We show how such a patternmatching problem can be formulated as an estimation of the posterior distribution of a Bayesian probabilistic model. We demonstrate the effectiveness of the approach by detecting faults in commercial building air handling units.
2012 Elsevier Ltd. All rights reserved.
1. Introduction
1.1. Overview
Air handling units account for a signicant portion of building
energy consumption and have a major impact on comfort conditions and building maintenance cost. An air handling units energy
usage can vary from the original design as components fail or fault:
dampers leak or fail to open/close, valves get stuck, and so on. Such
problems do not necessarily result in occupant complaints, as the
cascade structure of the control system would try to neutralize
the fault effect through re-adjusting other parameters and/or
changing the component loads. For instance, the effect of a damper-leakage fault may be covered by re-adjusting the position of
the hot or cold water valves. The fault may not even be recognized
to have occurred even though it may result in an increase in energy
usage. As long as the control system satises the set-points, the
building operators tend to assume that the system is working efciently in a non-faulty condition.
The topic of fault detection and diagnosis in air handling units
has been an active area of research and development for more than
Corresponding author.
E-mail addresses: massieh@ieee.org (M. Naja), dma@me.berkeley.edu (D.M.
Auslander), bartlett@cs.berkeley.edu (P.L. Bartlett), phaves@lbl.gov (P. Haves),
mdsohn@lbl.gov (M.D. Sohn).
0306-2619/$ - see front matter 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2012.02.049
348
Nomenclature
HVAC
heating, ventilation, and air conditioning
No fault no-fault condition
Reverse reverse actuator fault
OAD leak outside air damper leakage fault
RAD leak return air damper leakage fault
Stuck
stuck damper fault
Fouling fouling fault
VLV stuck valve-stuck fault
SAT
supply air temperature (F)
OAT
outside air temperature (F)
RAT
return air temperature (F)
MAT
mixed air temperature (F)
DMP
outside air damper position (F)
T_air_in temperature of entering air (F)
T_water_in temperature of entering water (F)
T_air_out temperature of outgoing air (F)
NTU
number of transfer unit (NTU) method
CFM
cubic feet per minute, measurement of air volume ow
rate
the mixing box and heating and cooling coils. As will be shown later, in such scenarios, when the sensor output is contaminated, it
could be due to the malfunction of any involved components,
and it is not necessarily straightforward to locate the malfunctioning one.
The complexity of modeling limitations and measurement constraints in air handling unit diagnostics becomes even more severe
when the possibility of concurrent faults is taken into account. A
single-fault assumption would relieve the diagnostic complexity,
but in reality, two or more faults may occur at the same time within one component or across different ones. The effect of concurrent
faults is not necessarily a linear interpolation of each individual
ones.
One approach to relieve the diagnostic complexity due to modeling limitations and measurement constraints is active diagnostics. In active-mode diagnostics, the diagnostic mechanism
actively controls or manipulates the system inputs (e.g. damper
positions, valves, etc.) to detect and isolate faults. Usually, inputs
are changed based on predened (or adaptive) test sequences to
explore various operating conditions. The tests can be structured
to explore operating points with less uncertainty or error, or in
the case of one sensor being affected by several components functionality, put neighboring components into neutral states to have
one component at a time affecting the measured variable. However, active-mode diagnostics require isolation of the system from
normal operation, an option that may not be feasible.
Conversely, in passive-mode diagnostics, there is no control on
the inputs. In this approach, the system is in a closed-loop operation manipulated by the control system based on the set-point error and so on. This is a more complicated scenario, as there is no
capability to change or manipulate the inputs to follow a test procedure or sequence. The diagnostic mechanism needs to somehow
make the best use of available data (measurements) from daily
operation.
The focus of this paper is on developing passive-mode diagnostic algorithms for air handling units that can systematically address the above constraints in a passive mode. We believe that
an ideal diagnostic solution should not only be reliable in detecting
and isolating abnormal behaviors but also have systematic solutions for constraints and challenges related to scalability and
affordability. Our proposed diagnostic algorithm is based on
VLV
IID
va
vw
Ch
Cc
Tair-in
Tw-in
a
b
l
r2
DP
Cp
d
valve position
independent and identically distributed
air velocity (ft/s)
velocity of water (ft/s)
hot uid capacity rate
cold uid capacity rate
temperature of incoming air (F)
temperature of incoming water (F)
coefcient factor
coefcient factor
mean or expected value
variance
total pressure rise across fan
specic heat of air (BTU/lbF)
density (lb/ft3)
fan combined efciency
349
Typically, an air handling unit contains three temperature sensors, the outside air temperature (OAT), return air temperature
(RAT), and supply air temperature (SAT) sensors, along with a fan
status indicator (Fig. 1). One of the main challenges in monitoring
air handling unit performance is the absence of a reliable measurement for the mixed air temperature (MAT), the temperature of the
air coming from the mixing box before going through the heating/
cooling sections. Usually, either there is no sensor in place to measure the MAT or, even if there is a temperature sensor, the sensor
readings are unreliable due to incomplete upstream mixing. This
constraint forces us to use the SAT sensor to evaluate mixing box
performance. However, as shown in Fig. 1, the SAT is also affected
by the heating/cooling coil functionality, and distinguishing the
350
each individual buildings limit the applicability of diagnostic solutions that rely on detailed models (or models that rely on conguration data that is not easily measureable or accessible) from the
scalability perspective. On the other hand, when an analysis approach employs simplied, more generic, models, the challenge
is how to differentiate between the inconsistencies due to model
misspecication errors and those due to system malfunction. In
other words, when detailed models are replaced with more simplied ones, the interpretation of model prediction differences becomes more challenging.
A strategic approach to address the complexity of employing
simplied models is to change the focus of an analysis approach instead on system behavioral patterns instead on error residuals. In
other words, instead of analyzing the difference between the system output and the model prediction at one or a few operating
points, diagnostics are made by evaluating the system behavioral
patterns over a window of operation. This lessens the dependency
of the diagnostic algorithm on model accuracy. Such an approach
has been employed by a number of diagnostic routines, particularly qualitative and semi-quantitative diagnostic approaches
[26,27]. The key here is an algorithm (inference mechanism) that
evaluates the observed behavior and compares it against a set of
predened (or even adaptive) hypotheses. Fuzzy logic has become
a popular choice for such problems due to the inherent exibility
embedded in fuzzy sets and fuzzy rules, which makes it a suitable
solution for reasoning in domains with some level of uncertainty
[44,16,17,20]. For example, Haves et al. [17] proposed a fuzzybased diagnostic routine for the fault diagnostics of VAV air handling units in which the fuzzy-based inference mechanism compares the predictions of simplied models with the air handling
unit component outputs at various operating conditions to draw
conclusions about the air handling unit health status.
However, fuzzy-based inference mechanisms have their own
limitations. As the problem complexity grows (due to the system
complexity, a large amount of disparate sensor data, the number
of potential faults, etc.), a large number of fuzzy sets and fuzzy
rules are required to analyze the system performance. Added to
this is the difculty with adjusting and tuning fuzzy sets either
manually or through other approaches.
Another approach to managing modeling limitations are rulebased diagnostic routines [42,10,35,1,28,37,38]. In this approach,
a priori knowledge is formulated through a set of if-then rules coupled with an inference mechanism searching through the rules to
draw a diagnostic conclusion. Rule-based frameworks can be designed based on expert knowledge or rst principles. Their advantage is simplicity and ease of deployment; however, as discussed in
Katipamula and Brambley [26,27], as problem complexity grows or
when new/additional rules are added, the simplicity of the approach is lost quickly. Furthermore, sometimes the activation of
the rules depends on threshold(s), which may depend greatly on
model uncertainties, measurement errors, or other issues. More
discussion on this can be found in House et al. [19].
In this paper, we adopt the strategy of employing simplied
models, as we believe that dependency on complex and detailed
models is a signicant technological barrier and cause for industry
resistance to large-scale deployment. Our approach therefore relies
on more sophisticated inference mechanisms to interpret discrepancies between model predictions and the system output.
F ff1 ; f2 ; f3 ; . . . :; fn g
3:1
E fe1 ; e2 ; e3 ; . . . :; em g
3:2
Pfj je1 ; e2 ; e3 ; . . . ; em
Pf1 ; f2 ; f3 ; . . . :; fn je1 ; e2 ; e3 ; . . . ; em
f1 ...fn excludingfj
3:3
Now, using Bayes rule, we can compute P(F/E) as:
Pf1 . . . fn je1 . . . em P
3:4
m
P
i1
f1 ...fn
3:5
3. Diagnostic algorithm
We think of fault diagnostics as the process of analyzing a system behavioral pattern (observed performance) and comparing it
with a set of hypothetical patterns to nd the closest match. Each
hypothetical pattern is developed based on the assumption of the
3
The mixing box functionality, model, and diagnostic algorithm are discussed in
detail in Section 4.
4
Here, the IID assumption means that, given faults f1 . . . fn, the random variables
e1 . . . em are statistically independent and identically distributed. More on IID
sampling can be found in DasGupta [11].
351
ei Ii ; Oi
The inputs are assumed to be known and deterministic,5 and the
output is what is measured from the system behavior. For example,
in the case of the mixing box, the inputs are the outside air temperature (OAT), the return air temperature (RAT), and the outside air
damper position (DMP), and the output could be the mixed air temperature (MAT) or outside air fraction (OAF).6
Under these assumptions, P(ei|f1 . . . fn) can be written as7:
3:6
x1 h1 I; x2 h2 I; . . . ; xn hn I
3:7
5
The assumption of deterministic inputs can be dropped for more general
scenarios.
6
OAF is dened in Section 4.
7
Here we assume modeling the static behavior of the system.
8
Keep in mind that such simplication/assumption would affect only the
denominator of Equation (3.4) [or the last of part of Eq. (3.6)], which is the
normalizing factor for correct estimation of the posterior probabilities. They will not
affect the process of locating the fault combination with maximum posterior
distribution. They would change only slightly the marginal probability of faults.
l hT x where hT h1 ; h2 ; . . . ; hn T
3:8
1
1
Pyi jl; r2 p exp 2 yi li 2
2r
2pr2
9
8
<l y l2i =
y2i
1
i i
2
exp
p exp 2
: r2 ;
2r
2pr2
li yi Ali
hyi exp
2
l1 = hTx1,
where
hyi
1
p
,
2pr2
Ali
3:9
lh log
N
Q
li yi Ali
r2
hyi exp
i1
N
P
loghyi
i1
hT XN
xy
i1 i i
1 XN
r2
i1
AhT xi
3:10
N
Q
Pyi
i1
rh l
N @l @ l
P
1 XN
1
i
y li xi 2 X T y l
2
i1 i
@
l
@h
r
r
i1
i
3:11
where X = [x1, x2, . . ., xn], Y = [y1, y2, . . ., yn]T and l = [l1, l2, . . ., ln]T
And the second derivative is:
r2h l
1 Xn
r2
x xT
i1 i i
r2
XT X
3:12
352
I O
PF 1 ; F 2 jI1 ; O2 P P 2 P1
F1
F2
PF 1 PF 2 PO1 jF 1 ; I1 PO2 jF 2 ; I2
I2 O1 PF 1 PF 2 PO1 jF 1 ; I 1 PO2 jF 2 ; I2
3:14
the determination of the fault combination/status maximizing the
posterior probability.
Another issue to consider is the distinction between what we
call abrupt faults and degradation faults. Abrupt faults are faults
that arise instantaneously (e.g. stuck-damper fault, reverse-actuator fault, etc.), while degradation faults evolve over time, becoming
progressively more severe (e.g. damper leakage, valve leakage,
etc.). From a modeling and diagnostics standpoint, abrupt faults
can be thought of as binary events [the fault either exists (1) or
not (0)], while degradation faults are more like events with an
associated severity level parameter. This means that, in the case
of degradation faults, another parameter (fault stage) is required
to include the characteristics of the fault and its effects. Thus, we
enhance the diagnostic algorithm in Fig. 2 (Graph A) and Eq.
(3.4) by adding another node (stage node) to capture such dynamics (Fig. 2, Graph B). If a fault is a degradation fault, the stage node
species its severity level, and the system output is inuenced by
both the fault and its severity level. If the fault is an abrupt fault,
the severity will become just an identity matrix. Then, Eq. (3.4)
can be modied as:
PFjE
P
P
L PFPLjFPEjF; L
PF; LjE P P
L
F
L PFPLjFPEjF; L
3:13
where L = {l1, l2, l3, . . . , ln} is various severity levels. P(F) is the fault
prior. P(L|F) is the severity prior for each fault. Again, statistical or
intuitive approaches can be employed for the estimation of severity
priors. Here, we consider a uniform distribution for severity priors.
P(E|F,L) is, again, the likelihood function: the probability of observing E given the fault condition F and severity level L.12
The last issue to discuss is how the proposed algorithm can be
extended systematically to address measurement constraints. As
mentioned earlier, one of the challenges in air handling unit diagnostics is the absence of reliable measurements for MAT, which
forces us to rely on SAT to evaluate the functionality of both the
mixing box and heating/cooling coils.
For such scenarios, the proposed diagnostic algorithm (Bayesian
model) can be expanded to a model with a mixture of components
in which missing measurements are considered to be hidden variables (nodes) and the posterior probability is estimated by summing over all possible values of the hidden nodes.
12
To avoid complexity and present the equations in a more clear structure, we use
the Fig. 2 (Graph A) model as our reference for further derivations. However, all the
equations can be systematically expanded to the Fig. 2 (Graph B) format using
Equation (3.13).
The elements in Eq. (3.14) are either component fault priors (e.g.
P(Fi)) or a component likelihood function (e.g. P(Oi|Fi, li)). For each
measurement, the estimation of the posterior probability of each
fault combination involves considering all possible values of the
hidden (unknown) variable(s) and computing the posterior probability in each case. This has a higher computational complexity
compared to the scenarios where there is no hidden variable, especially if there are two, three, or more hidden variables. Also, note
that, in the case of a mixture of components, diagnostics are performed with less information available (some variables are not
measured and considered hidden). As we will see in the illustrative
examples in Section 4, this does not come for free. There will be
some penalty /cost for the diagnostic process that may affect diagnostic efciency.
4. Illustrative examples
In this section, we apply the proposed algorithm to air handling
unit diagnostics. We rst start with single component diagnostics
to demonstrate the capability of the proposed diagnostic algorithm
in detecting/isolating faults, handling concurrent faults, and dealing with modeling uncertainty/error. Then, we will consider mixture component scenarios to show how the algorithm can
systematically manage measurement constraint issues.
Example 1 Diagnosis of mixing box: As mentioned earlier, a mixing box mixes the air returning from the building with the outside
air based on the ratio dened by the control system (Fig. 1). The ratio is specied to minimize the energy required to heat up or cool
down the supply air and to satisfy the standard fresh air required
for occupants. A mixing box consists of three dampers: the outside
air damper, return air damper, and exhaust air damper (Fig. 1),
manipulated by one actuator in which the outside air and return
dampers operate in opposite directions. Mixing box malfunction
is a common fault in air handling unit operation. The malfunction
could be due to the leakage of the outside or return air dampers
(not fully closed in the 0% closing command position), the reverse
operation of the actuator, stuck damper(s), and so on.
Mixing box performance is usually analyzed by a dimensionless
parameter [44] outside air fraction (OAF), which is the ratio of the
difference between the mixed air temperature (MAT) and the return air temperature (RAT) over the difference between the outside
air temperature (OAT) and the return air temperature (RAT).
OAF
MAT RAT
OAT RAT
4:1
353
Fig. 4. Mixing box model (OAF variations versus OA damper) in non-faulty and various faulty conditions; Graph A is the operation under the non-faulty condition. Graph B is
related to the reverse-actuator fault condition, Graph C is an outside-air-damper-leakage fault, and Graph D is a return-air-damper-leakage fault.
l a1 f1 a2 f2 a3 f3
r2 b1 f1 b2 f2 b3 f3
f1 1;
f2
DMP
;
100
f 3 DMP=1002
where DMP is the outside air damper position. For example, a set of
possible values for a1, a2, a3 and b1, b2, b3 in non-faulty operation
(Graph 1 in Fig. 4) could be:
a1 0; a2 1; a3 0
b1 0;
b2 1;
b3 1
For each operating condition (various faulty and non-faulty conditions), a and b can be estimated through maximizing the likelihood
function of N(l, r) provided IID samples. However, as shown in Eqs.
(3.9), (3.10), (3.11), (3.12), the likelihood function of such a model is
a concave function, which means that a convex optimization
354
Fig. 5. Mixing box operation and diagnostic results. The upper graph shows the variations of OAT, RAT, MAT, and outside air damper position (DMP). SAT is used as a proxy for
MAT after correcting for the fan temperature rise. The lower graph shows the diagnostic assessment. As shown, the mixing box has a return-air-damper-leakage fault.
algorithm can be employed to locate the global maxima and estimate a and b numerically.
Applying the diagnostic model shown in Fig. 2A for mixing box
diagnostics, the input node contains the damper position and return and outside air temperatures, and the output node contains
OAF. Data from air handling unit operation at an experimental
facility, the Iowa Energy Center,13 was obtained to evaluate the performance of the diagnostic algorithm. During the test, faults were
deliberately induced to the system for analysis purposes.
An example of the mixing box operation, which contains the
variations of OAT, RAT, and MAT, as well as outside air damper position (DMP), is shown in Fig. 5 (upper graph).14 Note that, due to
poor mixing in most mixing boxes, SAT has been used as a proxy
for MAT after correcting for the fan temperature rise.15 Also, keep
in mind that the outside and return air dampers operate in opposite
directions and are controlled by one actuator.
A visual inspection of Fig. 6 (upper graph) would indicate an
abnormality in the mixing box operation. When the DMP is fully
closed, MAT is very close to RAT (as expected); however, when
the DMP is fully open, there is a gap between MAT and OAT, indi-
13
The Iowa Energy Center is an experimental facility for research, demonstration,
and education in building HVAC systems, energy efciency, and conservation (http://
www.energy.iastate.edu/).
14
The air handling unit has a design air ow of 3200 CFM with a draw-through
supply fan. Both the supply fan and the return fan motors are in stream. The design
supply fan pressure rise is 3.25 inches of water, and the return Fan pressure rise is 1.7
in. of water.
15
The fan temperature rise can be calculated by equating the increase in the
sensible heat content of the air stream to the sum of the uid work done by the fan
and the heat produced by the inefciency of the fan and other associated
components: DT dCDpPg where DP is the total pressure rise across the fan, d is the
density of air, Cp is the specic heat of air, g is the combined efciency of the fan
components in the air stream (typically the fan, belt, and motor). DP can be obtained
from the design pressure rise [usually available from test and balance (TAB)
measurements] and the efciency of the fan + motor is available from manufacturers
literature, mechanical drawings, etc. In general, the fan temperature rise is about one
degree or so. For instance, in this example, DP is 3.2 in. of water, and the efciency is
0.76, which leads to a fan temperature increase of around 1. 5 F. More can be found
in Haves [17].
cating that MAT is not solely inuenced by OAT but by RAT as well.
This suggests that the return air damper is not fully closed and has
a leakage fault.
The diagnostic result is shown in Fig. 5 (lower graph). As predicted, the system has a return-air-damper-leakage fault.16 Initially, the diagnostic mechanism makes a vague assessment
regarding most faults. Earlier assessments are inuenced signicantly by the fault prior assessments. As more data is observed,
the diagnostic belief moves away from prior belief and gets closer
to the true belief (empirical distribution). Another interesting point
is how the diagnostic mechanism nalizes its assessment. As we
know, return-air-damper-leakage faults cannot be condently conrmed unless the mixing box operates at (or near) the 100% outside
air damper position (which would be a 0% return air damper position). As shown in the gure, the mechanism does not nalize its
assessment regarding the return-air-damper-leakage fault until it
observes the system performance at around the 100% outside air
damper position.
The probability graphs shown in Fig. 5 are marginal probabilities calculated by Eq. (3.3). At each sampling time, the posterior
probability of each fault combination is rst computed based on
the current and previous observations, P(f1 . . . fn|e1 . . . ei). With the
assumption of IID sampling, the computation can be achieved
using Eq. (3.5). After the computation of posterior probabilities,
the marginal probability of each fault is computed using Eq. (3.3).
Example 2 Concurrent faults: In this example, we consider a
concurrent fault scenario in the mixing box. The mixing box characteristics are the same as in the previous example, and its operation is shown in Fig. 6 (upper graph). It may not be as easy as in the
previous example to detect faults by visual inspection. However,
the graph indicates that, as the DMP closes, MAT gets closer to
OAT instead of RAT, which suggests a reverse-actuator fault. In
addition, when the outside air damper is completely closed, there
is still a gap between MAT and OAT, which is a red ag for a leakage
16
In fact, a 10% return-air damper-leakage fault was intentionally induced to the
system.
355
Fig. 6. Mixing box operation and diagnostic results. The upper graph shows the mixing box operation. The lower graph shows the diagnostic analysis. In this example, the
mixing box has two concurrent faults: reverse-actuator and return-air-damper-leakage faults.
4:2
1 expNTU c
1 c expNTU 1 c
4:3
C min
; C min minC h ; C c ; C max maxC h ; C c
C max
4:4
Ch is the hot uid capacity rate; Cc is the cold uid capacity rate.
NTU
UA
;
Cm
1
;
rt
r t ra v 0:8
r m r w v 0:8
a
w
4:5
where va is the velocity of the air and vw is the velocity of the water.
In his paper, Holmes provided the suggested values for ra, rm, and rw.
Similar to the mixing box case, here we have used a simplied model of the heat exchanger, requiring parameters that can be obtained
356
Fig. 8. Heating coil operation and diagnostic results. The upper graph shows the variations of entering air temperature (T_air_in), entering water temperature (T_water_in),
outgoing air temperature (T_air_out), and valve position (Valve). The lower graph shows the diagnostic results. As shown, the coil has a valve-leakage fault.
18
A leakage fault of ve to ten percent was deliberately induced in the system. The
coil has 12 circuits, the tube inside diameter is 0.5 in., the face height is 1.93 ft, and
the face width is 3.33 ft. Also, the valve authority is 50% and the expected thermal
transfer at low ow for the valve-coil system at normal operation is shown in Table 1.
Fig. 9. Mixture of components diagnostic model for the air handling unit. The rst
component is related to the mixing box, and the second component is related to the
heating coil. MAT, which is the output of the rst component and part of the inputs
to the second component, is not measureable and is considered hidden (unknown).
357
3
7.1
6
13.2
9
18.8
12
24
15
28.9
18
33.4
21
37.7
24
41.7
27
45.5
30
49.1
Fig. 10. Air handling unit operation and diagnostic results. The upper graphs show the operation of the mixing box and the heating coil and the lower graphs show the
diagnostic results.
penalty paid for the absence of MAT. As the algorithm relies on SAT
to interpret the functionality of both the mixing box and the heating coil, the uncertainty associated with each component model is
aggregated into SAT readings. In this case, the algorithm requires
more data, which basically means more time, to make a solid
assessment. In other scenarios, the constraint may affect the condence level of the diagnostic assessment.
The last issue to discuss is sensor error. Normally, in air-handling units, sensor error could be up to 2 F. The question is
how this would affect the diagnostic performance. In general, as
the diagnostic algorithm is less dependent on measurement at
any individual point and takes into account the system behavioral
pattern over a window of operation, it has more exibility in dealing with sensor errors. We know that the algorithm already takes
into account some level of uncertainty/error due to the employment of simplied models. To account for sensor error in a systematic way, the uncertainty can be expanded to include both
modeling errors and sensor errors, such as in the likelihood function. If the employed mode is an extension of an analytical model
with added uncertainty, the sensor error can be included as part of
the analytical process. If other statistical procedures are used to develop the model, the sensor error can be assumed to be statistically
independent and aggregated accordingly. One easy way could be to
think of the sensor error as a random variable with normal distribution (or a distribution recommended by the sensor manufacturer), add that to the modeling error random variable, and
compute the joint variance.
Incorporation of sensor error results in a higher level of uncertainty associated with each measurement. As a result, the diagnostic mechanism would require more data to reach a solid
5. Conclusion
Relying on accurate or detailed models and requiring the measurement of parameters/variables that are not easily measureable/
accessible in practical applications have been the main drawbacks
of diagnostic solutions for air handling units from the scalability
and affordability perspectives. The aim of this paper was to develop diagnostic algorithms that are less dependent on model accuracy and more exible with respect to measurement constraints.
In the proposed diagnostic algorithm, we think of fault diagnostics
as the process of analyzing a system behavioral pattern (observed
performance) and comparing it with a set of hypothetical patterns
to nd the closest match. Each hypothetical pattern is developed
based on the assumption of the existence of none, one or more
faults in the system. We demonstrated how such a problem can
be formulated as a posterior estimation problem of a Bayesian
358
model. It was shown how effective the problem diagnostic algorithm could be in detecting/isolating faults, dealing with measurement constraint challenges, and managing the complexity of
concurrent faults.
Although the focus of this paper was on air handling unit diagnostics, we believe that the proposed algorithm has the potential
to be applied in other applications with similar restrictions. Further
research and development may be required for more complex scenarios. A commercial building, for example, contains several air
handling units and other components, which increases the system
complexity as a network of components. Research needs to be
done to determine what challenges exist because of the added
complexity (for example, dimensionality) and what further development might be needed to make the proposed approach applicable to such systems.
References
[1] Andersen KK, Reddy TA. The error in variables (EIV) regression approach as a
means of identifying unbiased physical parameter estimates: application to
chiller performance data. Int J Heat, Ventil, Air Condit Refrig Res 2002;8(3):
295309.
[2] Ardehali M, Smith TF, House JM, Klaassen CJ. Building energy use and control
problems: an assessment of case studies. ASHRAE Trans 2003;109:2.
[3] Baker N, Steemers K. Energy and environment in architecture. New York: E&FN
Spon; 2000.
[4] Brambley MR, Pratt RG, Chassin DP, Katipamula S. Automated diagnostics for
outdoor air ventilation and economizers. ASHRAE J 1998;40(10)::4955.
[5] Buildings energy data book. US Department of Energy (DOE); 2003.
[6] DOE plan projects savings for new building energy costs. ASHRAE J; 2000. pp6.
[7] Carling P. Comparison of three fault detection methods based on eld data of
an air-handling unit. ASHRAE Trans 2002;108(1).
[8] Cibse GH. Building control systems. Oxford: Butterworth-Heinemann; 2000.
[9] Claridge DE, Culp CH, Liu M, Deng S, Turner WD, Haberl JS. Campus wide
continuous commissioning of university buildings. In: Proceedings of the
ACEEE summer study. Washington, DC: ACEEE; 2000.
[10] Castro N. Performance evaluation of a reciprocating chiller using experimental
data and model predictions for fault detection and diagnosis. ASHRAE Trans
2002;108(1).
[11] DasGupta A. Asymptotic of statistics and probability. Springer Tests in
Statistics; 2008.
[12] Dexter AL, Ngo D. Fault diagnosis in air-conditioning systems: a multi-step
fuzzy model-based approach. Int J Heat Ventil Air Condit Refrig Res
2001;7(1):83102.
[13] Dobson AJ, Barnett A. An introduction to generalized linear models. Texts in
statistical science. 3rd ed. Chapman & Hall/CRC; 2008.
[14] Du Z, Jin X, Yang Y. Fault diagnosis for temperature, ow rate and pressure
sensors in VAV systems using wavelet neural network. J Appl Energy
2009;86(9):162431.
[15] Glass AS, Gruber P, Roos M, Todtli J. Qualitative model-based fault detection in
air-handling units. IEEE Control Syst Mag 1995;15(4):1122.
[16] Haves P, Salsbury T, Wright JA. Condition monitoring in HVAC subsystems
using rst principles. ASHRAE Trans 1996;102(1):51927.
[17] Haves P, Kim M, Naja M, Xu P. A semi-automated commissioning tool for VAV
air handling units: functional test analyzer. ASHRAE Trans 2007;113:1.
[18] Holmes MJ. The simulation of heating and cooling coils for performance
analysis. In: Proceeding of the conference on system simulation in building,
Liege; 1982.
[19] House JM, Vaezi-Nejad H, Whitcomb JM. An expert rule set for fault detection
in air handling units. ASHRAE Trans 2001;107(1).
[20] Hyvrinen J, Krki S. International energy agency building optimisation and
fault diagnosis source book. Espoo, Finland: Technical Research Centre of
Finland, Laboratory of Heating and Ventilation; 1996.
[21] Incropera FP, DeWitt DP, Bergman TL, Lavine AS. Fundamentals of heat and
mass transfer. Chapman and Hall/CRC; 2011.
[22] Jordan MI. Graphical models, statistical science, vol. 19; 2004. p. 14055
[Special Issue on Bayesian Statistics].
[23] Jordan MJ, Wainwright MJ. Graphical models, exponential families, and
variational inference; 2008.
[24] Kumar S, Sinha S, Kojima T, Yoshida H. Development of parameter based fault
detection and diagnosis technique for energy efcient building management
system. Energy Convers Manage 2001;42:83354.
[25] Katipamula S, Pratt RG, Chassin DP, Taylor ZT, Gowri K, Brambley MR.
Automated fault detection and diagnostics for outdoor-air ventilation systems
and economizers: methodology and results from eld testing. ASHRAE Trans
1999;105(1).
[26] Katipamula S, Brambley MR. Methods for fault detection, diagnostics, and
prognostics for building systems a review part I. HVAC&R Res 2005;11:n1.
[27] Katipamula S, Brambley MR. Methods for fault detection, diagnostics, and
prognostics for building systems a review part 2. HVAC&R Res 2005;11:n2.
[28] Kumar SS, Kojima T, Yoshida H. Development of parameter based fault
detection and diagnosis technique for energy efcient building management
system. Energy Convers Manage 2001;42:83354.
[29] Lee WY, Park C, Kelly GE. Fault detection of an air-handling unit using residual
and recursive parameter identication methods. ASHRAE Trans 1996;102(1):
52839.
[30] Lee WW, House JM, Kyong NH. Subsystem level fault diagnosis of a buildings
air-handling unit using general regression neural networks. J Appl Energy
2004;77(2):15370.
[31] McCullagh P, Nelder JA. Generalized linear models. Monographs on statistics &
applied probability. 2 nd ed. Chapman & Hall/CRC; 1989.
[32] Norford LK, Wright JA, Buswell RA, Luo D, Klaassen CJ, Suby A. Demonstration
of fault detection and diagnosis methods for air-handling units (ASHRAE 1020RP). HVAC&R Res 2002;8(1):4171.
[33] Pakanen J, Sundquist T. Automation-assisted fault detection of air-handling
unit; implementing the method in a real building. Energy Build 2003;
35:193202.
[34] Peitsman HC, Bakker V. Application of black-box models to HVAC systems for
fault detection. ASHRAE Trans 1996;102(1):62840.
[35] Reddy TA, Niebur D, Andersen KK, Pericolo PP, Cabrera G. Evaluation of the
suitability of different chiller performance models for on-line training applied
to automated fault detection and diagnosis. Int J Heat Ventil Air Condit Refrig
Res 2003;9(4).
[36] Riemer PL, Mitchell JW, Beckman WA. The use of time series analysis in fault
detection and diagnosis methodologies. ASHRAE Trans 2002;108(2).
[37] Rossi TM. Detection, diagnosis, and evaluation of faults in vapor compression
cycle equipment. Ph.D. thesis, School of Mechanical Engineering, Purdue
University, West Lafayette, Indiana; 1995.
[38] Rossi TM, Braun JE. A statistical, rule-based fault detection and diagnostic
method for vapor compression air conditioners. Int J Heat Ventil Air Condit
Refrig Res 1997;3(1):1937.
[39] Sohn MD, Small MJ, Pantazidou M. Reducing uncertainty in site
characterization using Bayes Monte Carlo methods. J Environ Eng 2000;
126(10):893902.
[40] Sohn MD, Reynolds P, Singh N, Gadgil AJ. Rapidly locating and characterizing
pollutant releases in buildings. J Air Waste Manage Assoc 2002;52:142232.
[41] Salsbury TI, Diamond RC. Fault detection in HVAC systems using model-based
feed-forward control. Energy Build 2001;33:40315.
[42] Wagner J, Shoureshi R. Failure detection diagnostics for thermo-uid systems.
J Dyn Syst Meas Contr 1992;114(4):699706.
[43] Wang S, Chen Y. Fault-tolerant control for outdoor ventilation air ow rate in
buildings based on neural networks. Build Environ 2002;37:691704.
[44] Xu P, Haves P, Kim M. Model-based automated functional testingmethodology and application to air handling units. ASHRAE Trans 2005;
111(Pt. 1):979989 [LBNL-55802].
[45] Yoshida H, Iwami T, Yuzawa H, Suzuki M. Typical faults of air-conditioning
systems, and Fault detection by ARX model and extended Kalman lter.
ASHRAE Trans 1996;102(1):55764.
[46] Yoshida HS, Kumar S. ARX and AFMM model-based on-line real-time data base
diagnosis of sudden fault in AHU of VAV system. Energy Convers Manage
1999;40:1191206.