Energy and Buildings 133 (2016) 37–45

Contents lists available at ScienceDirect

Energy and Buildings

journal homepage: www.elsevier.com/locate/enbuild

A decision tree based data-driven diagnostic strategy for air

handling units
Rui Yan a , Zhenjun Ma a,∗ , Yang Zhao b , Georgios Kokogiannakis a
a
Sustainable Buildings Research Centre (SBRC), University of Wollongong, NSW, 2522, Australia
b
Institute of Refrigeration and Cryogenics, Zhejiang University, China

a r t i c l e i n f o a b s t r a c t

Article history: Data-driven methods for fault detection and diagnosis of air handling units (AHUs) have attracted wide
Received 11 May 2016 attention as they do not require high-level expert knowledge of the system of concern. This paper presents
Received in revised form 10 August 2016 a decision tree based data-driven diagnostic strategy for AHUs, in which classification and regression tree
Accepted 20 September 2016
(CART) algorithm is used for decision tree induction. A great advantage of the decision tree is that it can
Available online 21 September 2016
be understood and interpreted and therefore its reliability in fault diagnosis can be validated by both
testing data and expert knowledge. A steady-state detector and a regression model are incorporated
Keywords:
into the strategy to increase the interpretability of the diagnostic strategy developed. The proposed
Decision tree
CART
strategy is validated using the data from ASHRAE 1312-RP. It is shown that this strategy can achieve a
Feature selection good diagnostic performance with an average F-measure of 0.97. The interpretation of the diagnostic
Air handling unit decision tree using expert knowledge showed that some diagnostic rules generated in the decision tree
Fault diagnosis comply with expert knowledge. Nevertheless, the interpretation also indicated that some diagnostic rules
Interpretability generated are not reliable and some of them are only valid under certain operating conditions, which
indirectly demonstrated the importance of the interpretability of fault diagnostic models developed using
data-driven methods.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction knowledge of the system and the computational costs are generally
manageable [4].
The operation of building Heating, Ventilation, and Air- Air handling unit (AHU) is one of the important components
Conditioning (HVAC) systems is vulnerable to various faults, and in HVAC systems and a number of studies have been focused on
the occurrence of any fault could lead to increased energy con- detection and diagnosis of various faults in AHUs. For instance, a
sumption or indoor thermal discomfort [1,2]. Successful detection rule-based FDD strategy for AHU temperature sensors was devel-
and isolation of any faults in HVAC systems in a timely manner can oped by Yang et al. [7], in which a set of if-then rules were
improve building energy efficiency and reduce carbon footprint. formulated based on expert knowledge. The performance evalu-
Over the last two decades, considerable efforts have been made ation using the data from a green building and a small scale AHU
in the development of fault detection and diagnosis (FDD) strate- simulator showed that this strategy is capable of isolating AHU sen-
gies for building HVAC systems [3–5]. Fault detection is a process sor faults under different operating modes. A FDD strategy based
that determines whether or not an HVAC system is operating on Bayesian Belief Network (BBN) for AHUs was presented by Zhao
in a healthy condition while fault diagnosis aims to identify the et al. [8,9], in which a probabilistic graphical model was used to
causes of the fault. Compared to fault detection, fault diagnosis represent the relationships of probabilistic dependencies within
is more challenging and complicated. The existing FDD methods different variables. It was shown that this method can success-
can be generally categorised into model-based methods, rule-based fully isolate common faults occurred in AHUs. A drawback of this
methods and data-driven methods [6]. Data-driven methods have method is that it requires high-level expert knowledge, especially
attracted wide attention as they do not require high-level expert in determining the probability parameters. A sensor fault detection
strategy for AHUs using cluster analysis was developed by Yan et al.
[5], in which the clustering algorithm Ordering Points to Identify
the Clustering Structure (OPTICS) was used to identify the spa-
∗ Corresponding author.
tial separated data groups which possibly indicate the occurrence
E-mail addresses: [email protected], [email protected] (Z. Ma).
of sensor faults. Wang et al. [10] described a hybrid FDD method

http://dx.doi.org/10.1016/j.enbuild.2016.09.039
0378-7788/© 2016 Elsevier B.V. All rights reserved.
38 R. Yan et al. / Energy and Buildings 133 (2016) 37–45

considered. The results showed that this ANN-based FDD method

Nomenclature can identify the root cause of AHU faults. An AHU FDD strategy
using ANN and wavelet analysis was developed by Fan et al. [20].
Symbols
Back-propagation neural network (BPNN) was used for AHU fault
a1 -a3 Coefficients
detection while Elman neural network coupled with wavelet anal-
␣ Complexity parameter
ysis was used for fault diagnosis. Support vector machine (SVM) is
c Total number of class
another commonly used machine learning method for AHU FDD.
F Air flow rate (m3 /min)
Liang and Du [16] proposed a diagnostic strategy for AHUs based
FN False negative
on support vector machine (SVM) method. This strategy can iso-
FP False positive
late different fault classes through finding the boundary that can
I Impurity measure
maximize the margins between the classes. Four physical mod-
N Fan speed control signal (%)
els were formulated and the residuals between the predicted and
n Number of observations
actual values were used as the input variables of the SVM classi-
P Power (W) or the proportion of the observations
fier. It was found that the diagnostic accuracy of the SVM-based
p Probability of the observations in a node
method mainly relies on the selection of the SVM kernel func-
R Misclassification rate
tion. Mulumba et al. [17] described a model-based FDD strategy
r Node misclassification rate
with SVM classification and an autoregressive model with exoge-
s Split
nous variables (ARX) for AHU FDD. The supply air humidity was
T Temperature (◦ C)
selected as a dependent variable while the supply air tempera-
t Node for splitting
ture and mixed air temperature were selected as the exogenous
TP True positive
variables for the ARX model using ReliefF-based feature selection.
U Valve control signal (%)
The performance of the SVM-based method was compared with
the classifiers such as ANN, random forest, and Naïve Bayes. It was
Subscripts
shown that the SVM-based method outperformed the other meth-
cc Cooling coil
ods in terms of the diagnostic accuracy. Wall et al. [25] described an
L Left
AHU FDD method using a dynamic Bayesian network (DBN). DBN
ma Mixed air
is similar to BBN but with an extra dimension of time. In contrast
oa Outdoor air
to the BBN-based method proposed in [8] in which the BBN param-
R Right
eters were estimated by expert knowledge, the parameters in DBN
rf Return fan
were obtained using both normal and faulty training data.
sa Supply air
Although data-driven FDD methods are promising when vali-
sf Supply fan
dated using the testing data, a key drawback of this approach is that
most data-driven methods (e.g. ANN and SVM-based) were devel-
oped based on black-box models, which means that it is almost
which combines physical models and expert rules for variable air impossible to understand how faults are isolated. As the actual fault
volume AHUs. To increase the model accuracy, genetic algorithms diagnostic accuracy of data-driven methods heavily relies on the
were used to determine the parameters of the model. quality of training data used, the factors such as lacking of unique
Data-driven based fault diagnosis methods were primarily data patterns and errors in the training data may result in invalid
developed based on pattern classification techniques [11]. The classification. In many cases, only a limited amount of training data
data-driven FDD methods for AHUs can be generally categorised is available, which usually cannot cover the full operation range of
into statistic based methods [12–14] and machine learning based the system of concern. The model developed may therefore only
methods [6,15–17]. Principal Component Analysis (PCA) and Fisher valid under certain operating conditions. Without interpretability,
Discriminant Analysis (FDA) are two common statistic based meth- it is difficult to know whether or not the model used is reliable and
ods. PCA is a multivariate analysis method which transforms the under what operating conditions the model is reliable.
data set into a new set of uncorrelated variables so that the first few This paper presents a decision tree based data-driven method
principal components retain the most of the variations [18]. Wang for fault diagnosis of AHUs. Decision tree, as a well-known classi-
and Xiao [12], for example, presented a PCA-based AHU sensor FDD fier, has been applied in prediction of building energy usage with
method, in which AHU sensor faults were detected by comparing satisfied accuracy [26,27]. Compared to other data-driven diagnos-
the PCA model prediction error with a pre-defined threshold while tic strategies, the proposed strategy is interpretable by using the
fault diagnosis was achieved by using Q-contribution plot. It was decision tree which can generate a set of if-then rules. The inter-
shown that PCA is capable of extracting major information from the pretability of the proposed strategy can be helpful in understanding
data with high dimensions and then presenting it in lower dimen- the diagnostic strategy for isolating different AHU faults. The pro-
sions. FDA is a classifier based on linear dimensionality reduction posed strategy can also automatically perform feature selection,
techniques and is optimal in terms of maximising the separations which often requires considerable efforts to analyze and define the
among different classes. Du and Jin [14] used PCA for fault detection key features in data-driven FDD strategies [28].
and FDA for fault diagnosis of AHU sensor faults. In a two-step AHU
sensor fault detection and isolation strategy developed by Padilla
and Choinière [19], PCA was used in the first step for fault detection 2. Outline of the diagnostic strategy
while active functional testing was used in the second step for fault
isolation. Among the machine learning-based methods, the use of The outline of the proposed AHU diagnostic strategy is illus-
artificial neural networks (ANNs) for AHU FDD has been extensively trated in Fig. 1. The overall strategy consists of three steps including
studied [6,20–24]. An ANN-based AHU FDD strategy was developed 1) data preparation; 2) decision tree induction and evaluation and;
by Lee et al. [22]. Seven user-defined residual variables between 3) decision tree interpretation.
the predicted and measured values were fed into a two-layer feed- The fault-free data is first used to determine the coefficients
forward ANN trained by back-propagation with both normal and of a regression model, as shown in Eq. (1) [8]. This regression
faulty data. Eight typical AHU sensor and mechanical faults were model is used to generate a residual feature in order to reduce the
 R. Yan et al. / Energy and Buildings 133 (2016) 37–45 39

where TP is the true positive, FP is the false positive and FN is the

false negative. Positive denotes the data with the targeted labels
while negative denotes the data with all other labels.
The last step is the interpretation of the post-pruned decision
tree by using expert knowledge in order to confirm whether the
diagnostic rules generated are valid and have physical meanings.

3. Decision tree induction

CART is a widely used decision tree induction method, in which

the decision tree is constructed by recursively partitioning the data
space and fitting a prediction model within each partition [31].
The partitions are graphically represented as a binary decision
tree. Unlike other commonly used tree induction methods such
as C4.5 and Chi-square Automatic Interaction Detector (CHAID)
which can be used for classification only, CART can be used for both
classification and regression [32]. In this study, CART is used as a
classification tree induction method for constructing the diagnostic
model.
For the classification tree induction, CART evaluates the split
point based on the goodness of the split criteria as shown in Eq. (3)
[33], where the impurity measure I used in CART is the Gini impurity
measure as defined in Eq. (4) [30]. The best split is at the point
where I (s, t) is maximised. As I(t) is a fixed value for a given point,
maximising the goodness of the fit is equivalent to minimising the
weighted average impurity measure of the child nodes [30]. If there
are no stopping criteria provided, the tree will continuously grow
until all observations in the node have the same class or the number
of the observations in the node is below a predefined threshold.
I (s, t) = I (t) − (PL (I (tL )) + PR (I (tR ))) (3)

Fig. 1. Outline of the proposed AHU fault diagnostic strategy. 

c−1
2
Gini (t) = 1 − [p (i|t)] (4)
i=0
complexity of the decision tree to be developed. The recorded fault-
free data and the data labelled with different AHU faults are then where I (s, t) stands for the impurity difference before and after
processed using a steady state detector to filter out the data in the a split, s stands for a particular split, t is the node for splitting, tL
transient period in order to improve the accuracy of fault diagnosis and tR are the left and right child nodes of t respectively, PL and PR
and reduce the complexity of the decision tree. The details of the denote the proportions of the observations at node t that go into the
steady state detector used are the same as that used in [8,29]. Using left and right child nodes respectively, p (i|t) stands for the fraction
the regression model and the filtered data, the residual between of the observations in class i at a given node t, and c is the total
the measured and predicted fan speed control signal can be gener- number of classes.
ated and a new data set with this new feature (i.e. residual) and all As mentioned before, a fully grown classification tree could
filtered data can be prepared for fault diagnosis. have two disadvantages. The first is the over-fitting problem, which
means that the model performs extremely well with the training
2
Nsf = a1 Fsa + a2 Fsa + a3 (1) data but performs poorly with the testing data [30]. The second
disadvantage is that a fully grown classification tree is difficult to
where Fsa and Nsf are the supply air flow rate and supply fan speed be understood and interpreted. CART provides a cost-complexity
control signal respectively, and a1 -a3 are the coefficients. based tree pruning strategy, which can optimize the trade-off
The second step, i.e. decision tree induction and evaluation, is between the cost of misclassification and the tree complexity. The
the core of the diagnostic strategy. Firstly, 2/3 of the data in the cost from the number of leaf nodes is measured by a non-negative
new data set is randomly selected and used as the training data for complexity parameter ␣ as defined in Eq. (5) [34].
decision tree induction based on the classification and regression
R (t) − R (T )
tree (CART) algorithm, which will be elaborated in Section 3. The ˛= (5)
decision tree is initially fully grown in order to capture all possi- |T̃ | − 1
ble diagnostic rules. However, a fully grown tree usually has too where T̃ is the leaf nodes of the subtree T, |T̃ | is the number of the
many branches which make the model interpretation difficult and leaf nodes, R(T) is the tree misclassification rate defined in Eq. (6),
may result in over-fitting. It is, therefore, necessary to prune the and R(t) is a single node misclassification rate defined in Eq. (7), in
tree by using appropriate methods. In CART, the cost-complexity which r(t) is defined in Eq. (8) [34].
measure is used for tree pruning purpose. After the decision tree is 
pruned, the rest data in the prepared new data set is then used to R (T ) = R (t) (6)
t ∈ Ť
test and validate the accuracy of the diagnostic decision tree devel-
oped based on the F-measure defined in Eq. (2). The F-measure is a R (t) = r (t) p (t) (7)
commonly used binary classification accuracy measure [30]. r (t) = 1 − maxi p(i|t) (8)
2 × TP where p(t) is the probability of the observations in the node t, and
F-measure = (2)
2 × TP + FP + FN maxi refers to the class i that results in the largest p(i|t).
40 R. Yan et al. / Energy and Buildings 133 (2016) 37–45

Fig. 2. The fully grown AHU fault diagnostic tree with the residual feature.

There are a finite number of ␣ values corresponding to a AHU faults considered in this study. These faults are among the typ-
sequence of pruned trees T0 , T1 ,. . .,Tn . Each Ti in this sequence is ical AHU mechanical faults and are commonly considered in AHU
characterised by a different value ˛i . To generate the sequence, FDD studies.
CART adopted an iterative algorithm starting with the fully grown The diagnostic strategy developed was validated based on the
tree (T0 ) and the first pruned tree (T1 ) is then developed by finding data collected under the AHU mechanical cooling mode tested dur-
the node t ∈ T0 that minimises Eq. (5) [35]. ing the summer period. In the mechanical cooling mode, the supply
The minimum value of this function is the value of ␣0 corre- air temperature was controlled at 12.8 ◦ C and the indoor temper-
sponding to T0 and a new pruned tree T1 is generated by pruning ature was maintained at 22.2 ◦ C. The supply fan was controlled to
tree T0 at the node t. The same procedure will be repeated based maintain the supply air static pressure at 348 Pa while the return
on tree T1 and generates ˛1 and T2 . The iteration terminates until fan operating speed was maintained at 80% of the supply fan oper-
the root node is reached. ating speed [40].
Once the sequence of the minimal cost-complexity trees is gen-
erated, the final pruned tree can be determined by plotting the 4.2. Classification tree induction
relative error (RE) versus the complexity parameters. The relative
error of a decision tree is defined as in Eq. (9). The optimal prune Based on the randomly selected 2/3 training data from the new
point is the point where RE does not have any significant decrease data set, a fully grown decision tree with a minimum number of 20
if further decreasing the cost of complexity. observations in a leaf node was generated, as illustrated in Fig. 2.
ncorrect The total number of observations corresponding to each fault and
RE = 1 − (9) fault free conditions used as the training data is presented in Table 1.
ntotal
The number of the observations in a leaf node is constrained in order
where ncorrect is the number of correctly classified observations and to avoid generating a too large decision tree and also for making the
ntotal is the total number of observations. visualization easier. The fully grown decision tree has a total of 26
splits. The root and internal nodes were labelled with the test condi-
4. Validation and interpretation of the AHU diagnostic tions in order to split the observations with different characteristics
strategy while the leaf nodes were labelled with the classification results.
The left branches of the root and internal nodes are the branches
In this study, the proposed strategy was implemented in R [36] that met the node splitting criteria while the right branches did not.
with CART implemented using the rpart [37] and rpart.plot [38] The number in the leaf node represents the number of the correctly
packages. The majority of figures presented were generated using classified observations out of the total number of the observations
R package ggplot2 [39]. classified to this leaf node. For instance, 333/334 in the cooling
coil valve stuck at 65% opening (i.e. CCVS65%) indicated that 333
4.1. Experimental data observations were correctly classified to CCVS65% fault while one
observation was mis-classified.
The experimental data from ASHRAE project 1312-RP [40–42] From Fig. 2, it can be seen that some faults such as cooling coil
was used to test and evaluate the proposed AHU fault diagnos- valve stuck at fully closed (i.e. CCVSFC) and AHU duct leakage before
tic strategy. ASHRAE 1312-RP tested various AHU faults based on supply fan (i.e. DLBSF) can be easily isolated after only a few node
an experimental AHU facility. Each fault was tested for 12 h from splitting steps. However, isolating some faults such as the heating
6.00a.m. to 6.00p.m. while the fault-free test was conducted for coil valve leakage (i.e HCVL) and exhaust air damper stuck at fully
three days. The sampling rate was one minute. The details of the closed (i.e. EADSFC) required considerable partitions, which sig-
experimental AHU facility and how the faults were introduced and nificantly decreased the interpretability of the diagnostic decision
tested can be found in [41–43]. Table 1 summarizes the different tree. Some leaf nodes only included a very small proportion of the
 R. Yan et al. / Energy and Buildings 133 (2016) 37–45 41

Table 1
Summary of AHU faults considered and their abbreviations.

Case No. Fault description Abbreviation Number of data points used as training data

Fault 1 Heating coil valve leakage HCVL 330

Fault 2 Cooling coil valve stuck at fully closed CCVSFC 474
Fault 3 Cooling coil valve stuck at 65% opening CCVS65% 364
Fault 4 Return fan fixed at 30% speed RFF30% 261
Fault 5 Return fan completely failed RFCF 285
Fault 6 Outdoor air damper stuck at fully closed OADSFC 266
Fault 7 Exhaust air damper stuck at fully closed EADSFC 313
Fault 8 AHU duct leakage before supply fan DLBSF 343
Fault-free Normal operation / 699

However, this relative error measure cannot reflect the impact of

the tree pruning on the diagnostic accuracy of each individual fault.
To confirm whether the diagnostic accuracy is acceptable for
individual faults, the performance of the pruned tree is evaluated
in terms of F-measure using the testing data and the results are
compared with that of the pre-pruned tree. The F-measures of the
pre-pruned tree and post-pruned tree are summarized in Table 3. It
can be seen that both pre-pruned and post-pruned diagnostic trees
can completely isolate five faults including the cooling coil valve
stuck at fully closed (Case 2), return fan fixed at 30% speed (Case 4),
return fan completely failed (Case 5), outdoor air damper stuck at
fully closed (Case 6) and AHU duct leakage before supply fan (Case
8), from the other faulty and fault-free cases. The F-measures of the
other four cases considered were ranged from 0.94 to 0.99 for the
pre-pruned decision tree and from 0.90 to 0.98 for the post-pruned
decision tree. The largest decrease in the diagnostic accuracy was
Fig. 3. Number of splits in the sequence of the pruned subtrees with different costs for the exhaust air damper stuck at fully closed (Case 7) and the cor-
of complexity and the corresponding relative errors with the residual feature. responding F-measure was decreased from 0.96 for the pre-pruned
tree to 0.90 for the post-pruned tree. The diagnostic accuracy of
observations which indicated that the contributions from these this fault is considered as acceptable in terms of the F-measure, as
branches to the diagnostic accuracy of the decision tree are not sig- the symptom of this fault is very insignificant to be detected and
nificant and can be generally neglected. Pruning these branches can isolated, as mentioned by Li and Wen [40].
result in a more interpretable classification tree while still remain- The decision tree was also developed without including the
ing an acceptable diagnostic accuracy. residual feature in the new data set in order to demonstrate the
Fig. 3 illustrates the number of splits in the sequence of the necessity and importance of using this residual feature in the pro-
pruned subtrees with different costs of complexity as well as the posed strategy. Fig. 5 shows the number of splits in the sequence
corresponding cross validated (X-val) relative errors. The X-val of the pruned subtrees with different costs of complexity as well
relative errors were calculated based on 10-folds cross valida- as the corresponding relative errors. Compared to Fig. 3, the size
tion. It should be noted that both horizontal axes are not in scale. of the decision tree is increased to 34 splits. Fig. 6 shows the post-
The dashed line indicates the minimum relative error that can be pruned tree which was pruned at the point where the relative error
achieved. More accurate values of the complexity parameters and is close to the relative error of the post-pruned decision tree with
their corresponding relative errors are summarized in Table 2. From the residual feature. It is obvious that without considering the resid-
Fig. 3 and Table 2, it can be seen that the cost of the complex- ual feature, the decision tree is hard to be interpreted due to the
ity parameter between 0.00759 (i.e. corresponding the number of increased number of splits. Some faults such as AHU duct leakage
splits of 13) and 0.00531 (i.e corresponding number of splits of 14) is before supply fan (i.e. DLBSF), and outdoor air damper stuck at fully
a good pruning point as the improvement of classification accuracy closed (i.e. OADSFC), can be easily isolated by the decision tree gen-
in terms of the relative error is insignificant if further decreasing the erated with the residual feature. However, they required several
cost of the complexity parameter. The pruned decision tree with 14 splits to be successfully isolated without considering the residual
splits is shown in Fig. 4. Compared to the fully grown tree, the over- feature.
all fault diagnostic error was increased by 3.0% after the tree prune.
Table 3
F-measures of different faults for pre-pruned and post-pruned diagnostic trees.
Table 2
Number of splits, corresponding complexity parameter and relative error. Case No. Fault description F-measure

Number of splits Relative error Complexity parameter Pre-pruned Post-pruned

9 0.09219 0.01081 1 Heating coil valve leakage 0.94 0.90

13 0.04895 0.00759 2 Cooling coil valve stuck at fully closed 1.00 1.00
14 0.04135 0.00531 3 Cooling coil valve stuck at 65% opening 0.99 0.98
16 0.03073 0.00512 4 Return fan fixed at 30% speed 1.00 1.00
18 0.02049 0.00266 5 Return fan completely failed 1.00 1.00
19 0.01783 0.00190 6 Outdoor air damper stuck at fully closed 1.00 1.00
21 0.01404 0.00076 7 Exhaust air damper stuck at fully closed 0.96 0.90
23 0.01252 0.00057 8 AHU duct leakage before supply fan 1.00 1.00
25 0.01138 0.00038 9 Normal operation 0.97 0.93
26 0.01100 0.00000 Average 0.98 0.97
42 R. Yan et al. / Energy and Buildings 133 (2016) 37–45

Fig. 4. Post-pruned AHU fault diagnostic tree with the residual feature.

4.3. Decision tree interpretation

In this section, the rules generated by the decision tree are inter-
preted in order to confirm the reliability of the proposed strategy.
It is worthwhile to note that the following analysis is based on the
faults investigated in this study (Table 1) without considering other
possible AHU faults.
The five faults with F-measure of 1.0, which are also the top five
branches of the diagnostic tree were first analyzed. The first split
in Fig. 4 indicated that the AHU suffered from the cooling coil valve
stuck at fully closed (i.e. CCVSFC) when the supply air temperature
(Tsa ) is not less than 19.2 ◦ C. This rule is easy to understand as there
is no chilled water supply when the cooling coil valve is fully stuck,
and therefore the supply air cannot be cooled down to the required
set-point.
The next split in Fig. 4 isolated the AHU duct leakage before
Fig. 5. Number of splits in the sequence of the pruned subtrees with different costs of supply fan (i.e. DLBSF) fault based on the residual between the pre-
complexity and the corresponding relative errors without using the residual feature. dicted and actual supply fan speed control signal. If the difference
between the predicted and the actual supply fan speed control sig-
nal is not less than 5.23%, the AHU will then be deemed as suffering

Fig. 6. Post-pruned AHU fault diagnostic tree without the residual feature.
 R. Yan et al. / Energy and Buildings 133 (2016) 37–45 43

Fig. 8. Density distribution of the return fan power under the faulty and fault-free
conditions.
Fig. 7. Relationship between supply fan speed control signal and supply air flow
rate under the faulty and fault-free conditions.
From Fig. 7, it can be seen that, for the remaining non-isolated faults,
only the data of CCVS65% fault was not overlapped with the fault-
from the DLBSF fault. This also complies with the expert knowledge free data, which indicated that CCVS65% fault may change the air
that the duct leakage before the supply fan leads to a decreased duct duct resistance. However, this symptom did not match with the
resistance before the supply fan as the static pressure before the common knowledge that abnormal operation of the cooling coil
supply fan is negative. Fig. 7 illustrates the relationship between the valve should not result in a variation in the air duct resistance.
AHU supply air flow rate and supply fan speed control signal under ASHRAE 1312-RP report [40] mentioned that the control strategy
different faulty and fault-free conditions, where the black curve of the test AHU facility opened the heating coil valve to maintain
represents the predicted supply fan speed control signal under the the supply air temperature if it is under the over-cooled condition.
fault-free conditions. A large deviation can be observed between Nevertheless, the control strategy had a malfunction which led to
the data points under the AHU normal operating condition and the fully close the outdoor air damper and exhaust air damper and fully
operating condition suffering from the DLBSF fault, which clearly open the return air damper. Due to the malfunction of the test AHU
illustrates why the residual feature was selected by CART for iso- under CCVS65% condition, the training data of this fault was con-
lating this fault. The patterns of other faults will be discussed later. sidered as invalid and was not representative to the AHU CCVS65%
Once CCVSFC and DLBSF faults have been isolated, the return fan fault.
power was then selected as the next splitting feature. Fig. 8 shows To isolate the exhaust air damper stuck at fully closed (i.e.
the density distribution of the return fan power under the fault-free EADSFC) and heating coil valve leakage (i.e. HCVL) faults, CART algo-
condition and all faulty conditions considered except the above two rithm further selected the outdoor air temperature (Toa ) as a new
faults that have been isolated. It should be noticed that the density splitting feature. If the outdoor air temperature is less than 26.8 ◦ C,
distribution of the return fan completely failed (i.e. RFCF) fault is the AHU was more likely suffering from the HCVL fault. The isola-
not visible in Fig. 8 as the power of the return fan of all observations tion of faults based on the outdoor air temperature alone does not
under the RFCF fault was 0. It is clearly shown that the boundary have any physical meaning, but could be meaningful when con-
of the return fan power of 182 W can isolate the return fan fixed at sidered with the previous rule of the cooling coil valve opening
30% speed (RFF30%) and RFCF faults from the other faulty and fault- (Ucc ) that is not less than 64.3%. In another word, if the outdoor
free conditions while the splitting point 77.9W was further picked temperature is not very high but the cooling coil valve opening
up to separate the RFCF fault from the RFF30% fault. Actually, it is is relatively large, the AHU might experience the HCVL fault. This
obvious that 77.9W is not the only point that can separate both combined rule complies with the common knowledge that an HCVL
faults as there is a large gap in the return fan power distribution fault will result in a larger cooling coil valve opening, which will be
under the RFCF fault and RFF30% fault conditions. more significant under the low cooling load conditions. However,
The residual feature was then further used to isolate the outdoor the cooling coil valve opening is not only influenced by the out-
air damper stuck at fully closed (i.e. OADSFC) fault from the other door air temperature but also is influenced by the factors such as
non-isolated faulty and fault-free conditions. When the difference supply air flow rate, supply water temperature, etc. Therefore, a
between the predicted and actual supply fan speed control signal is conclusive fault diagnostic result cannot be drawn based on these
less than −4.09%, the CART determined that the outdoor air damper rules. Fig. 9 shows the isolated fault boundaries in terms of the cool-
was stuck at fully closed. This is because that the fully closed out- ing coil valve control signal and outdoor air temperature. The right
door air damper increased the duct resistance, resulting in a higher bottom corner area with the light red color labelled as Condition
supply fan speed at a given air flow rate. This relationship can also 1 is the data space that the cooling coil valve opening (Ucc ) is not
be observed in Fig. 7. less than 64.3% and the outdoor air temperature (Toa ) is less than
After isolating all faults with F-measures of 1.0, further tree split- 26.8 ◦ C. As the majority data points in this region were with the
ting isolated the following remaining faults, including the cooling heating coil valve leakage (i.e. HCVL) fault and some of them with
coil valve stuck at 65% opening (i.e. CCVS65%) fault, exhaust air the exhaust air damper stuck at fully closed (i.e. EADSFC) fault, it is
damper stuck at fully closed (i.e. EADSFC) fault and heating coil therefore highly possible that the AHU is under the HCVL fault but
valve leakage (i.e. HCVL) fault, from the fault-free condition. The this conclusion is not definite. Another leaf node labelled with the
next split picked up the cooling coil valve control signal (Ucc) as HCVL fault (see Fig. 4) was isolated in the same way, which labelled
the splitting feature, but without resulting in a leaf node. The left as Condition 2 in the area with the light brown color (Fig. 9). It
branch first isolated CCVS65% fault based on the residual feature. deemed that the AHU is also under the HCVL fault if the cooling
44 R. Yan et al. / Energy and Buildings 133 (2016) 37–45

minimize the impurity in the leaf nodes. However, these rules do

not have a physical meaning. Furthermore, CART selected the resid-
ual feature to isolate the EADSFC fault from the normal operation
as shown in Fig. 4. However, as illustrated in Fig. 7, the data pat-
tern of the EADSFC fault is virtually completely overlapped by the
data points in the normal operating conditions, and these two rules
were therefore considered as unreliable.
In conclusion, the proposed CART-based AHU diagnostic strat-
egy is able to easily isolate the AHU CCVSFC, DLBSF, RFCF, RFF30%
and OADSFC faults. For the AHU HCVL fault, the proposed method
cannot provide a conclusive diagnostic result, but the diagnostic
outcome is highly valuable and can be used as a reference for fur-
ther investigation. The proposed method is not able to successfully
isolate the EADSFC fault as the symptom of this fault is insignificant.
The interpretation of the developed decision tree-based AHU
diagnostic strategy demonstrated that, similar to many data-driven
fault diagnostic strategies, the diagnostic accuracy and reliability
of the proposed strategy rely on the information embedded in the
training data. If the training data does not embody the unique data
Fig. 9. Isolated fault boundaries in terms of the cooling coil valve control signal and pattern of a particular fault, this fault is unlikely to be isolated,
outdoor air temperature. although the diagnostic accuracy of the strategy validated using the
testing data may still be acceptable. The isolation of the CCVS65%
fault showed that if the training data cannot reflect the actual oper-
coil valve opening is in the range of 45.6-64.3% and the outdoor air ation of a specific fault, the developed diagnostic strategy can still
temperature (Toa ) is less than 21.9 ◦ C. isolate this fault if it is only validated by the randomly selected test-
If the outside air temperature is not less than 26.8 ◦ C while the ing data. However, it might not be able to identify the target fault
cooling coil valve opening (Ucc ) is not less than 64.3%, it will be during the actual operation.
deemed that the AHU is suffering from the exhaust air damper stuck Therefore, the interpretability of a fault diagnostic strategy is
at fully closed (i.e. EADSFC) fault (see Fig. 4). These rules were also essential to help understand how different faults are isolated. The
illustrated in Fig. 9 as Condition 3 labelled with the light blue. This proposed decision tree-based data-driven fault diagnostic strategy
diagnostic result might not be reliable as the maximum outdoor provides a meaningful way to develop an interpretable diagnostic
air temperature under the normal operating condition used as the strategy through generating a set of if-then rules in a binary tree
training data is only 30.2 ◦ C. Under the EADSFC fault condition, all form, which can help to validate the reliability of the developed
return air will be re-supplied to the AHU and therefore a smaller diagnostic strategy.
cooling coil valve opening than that in the fault-free operation will
be resulted in, under the mechanical cooling mode. The EADSFC
fault was also isolated if the outdoor air temperature is not less 5. Conclusion
than 30.2 ◦ C and the cooling coil opening signal is less than 64.3%
(Condition 4 in Fig. 9). This result may also not reliable due to the This paper presented a data-driven AHU fault diagnostic strat-
same reason as mentioned above. egy based on the Classification and Regression Tree (CART) method,
The next split used the mixed air temperature (Tma ) as the split- which can automatically perform feature selection and therefore
ting feature and isolated the normal condition from other faulty doesn’t require considerable efforts to analyze and define the key
conditions. This is because that the fault-free case has more data features. Compared to black-box model based data-driven fault
points in this range than the other two faulty cases as shown in diagnostic strategies which do not have explicit physical mean-
the histogram plot in Fig. 10. Splitting in this way can therefore ings, the fault diagnostic results from the proposed strategy are
interpretable. The interpretation of the developed diagnostic tree
demonstrated that for most faults considered in this study, the fault
diagnostic rules generated by the CART algorithm matched well
with expert knowledge. It is also shown that the proposed CART-
based fault diagnostic strategy can successfully identify relevant
features to isolate different AHU faults. However, for some faults
such as heating coil valve leakage and exhaust air damper stuck at
fully closed, the rules generated were not valid or only valid under
certain operating conditions, although a high diagnostic accuracy
of the proposed strategy validated by using the testing data is still
achieved.
The overall results demonstrated the importance of the inter-
pretability in data-driven AHU diagnostic strategies. Due to the
inherent limitation of data-driven diagnostic strategies, AHU faults
that do not have a unique data pattern or are outside the oper-
ating conditions covered in the training data are difficult to be
isolated. In general, data-driven methods are superior in extract-
ing the useful information from large data sets and modelling the
behavior of HVAC systems. Expert knowledge can play a role in
Fig. 10. Histogram plot of the number of the observations against different mixed assisting in interpreting and validating the information and knowl-
air temperature. edge discovered by the data-driven methods. Therefore, combining
 R. Yan et al. / Energy and Buildings 133 (2016) 37–45 45

data-driven methods with expert knowledge might be a possible [18] I.T. Jolliffe, Principal Component Analysis, Springer-Verlag, New York, 1986.
solution for developing effective data-driven based fault diagnostic [19] M. Padilla, D. Choinière, A combined passive-active sensor fault detection and
isolation approach for air handling units, Energy Build. 99 (2015) 214–219.
strategies. [20] B. Fan, Z. Du, X. Jin, X. Yang, Y. Guo, A hybrid FDD strategy for local system of
AHU based on artificial neural network and wavelet analysis, Build. Environ.
Acknowledgment 45 (12) (2010) 2698–2708.
[21] Z. Du, X. Jin, Y. Yang, Fault diagnosis for temperature, flow rate and pressure
sensors in VAV systems using wavelet neural network, Appl. Energ. 86 (9)
The authors would like to sincerely thank ASHRAE for granting (2009) 1624–1631.
the permission to use the data from ASHRAE 1312-RP to validate [22] W.Y. Lee, J.M. House, C. Park, G.E. Kelly, Fault diagnosis of an air-handling unit
using artificial neural networks, Trans.-Am. Soc. Heating Refrigerating Air
the strategy developed. Conditioning Eng. 102 (1996) 540–549.
[23] Z. Du, B. Fan, X. Jin, J. Chi, Fault detection and diagnosis for buildings and
References HVAC systems using combined neural networks and subtractive clustering
analysis, Build. Environ. 73 (2014) 1–11.
[24] W.-Y. Lee, J.M. House, N.-H. Kyong, Subsystem level fault diagnosis of a
[1] V. Gunes, S. Peter, T. Givargis, Improving energy efficiency and thermal
building’s air-handling unit using general regression neural networks, Appl.
comfort of smart buildings with HVAC systems in the presence of sensor
Energ. 77 (2) (2004) 153–170.
faults, in: IEEE 17th International Conference on High Performance
[25] J. Wall, Y. Guo, J. Li, S. West, A dynamic machine learning-based technique for
Computing and Communications (HPCC), 2015, 2015, pp. 945–950.
automated fault detection in HVAC systems, ASHRAE Trans. 117 (part 2)
[2] Z. Ma, S. Wang, Fault-tolerant supervisory control of building condenser
(2011) 449–456.
cooling water systems for energy efficiency, HVAC&R Res. 18 (1–2) (2011)
[26] Z. Yu, F. Haghighat, B.C.M. Fung, H. Yoshino, A decision tree method for
126–146.
building energy demand modeling, Energy Build. 42 (10) (2010) 1637–1646.
[3] Z. Ma, S. Wang, Online fault detection and robust control of condenser cooling
[27] K. Basu, L. Hawarah, N. Arghira, H. Joumaa, S. Ploix, A prediction system for
water systems in building central chiller plants, Energy Build. 43 (1) (2011)
home appliance usage, Energy Build. 67 (2013) 668–679.
153–165.
[28] K. Bruton, P. Raftery, B. Kennedy, M. Keane, D.T.J. O’sullivan, Review of
[4] S. Katipamula, M.R. Brambley, Methods for fault detection, diagnostics, and
automated fault detection and diagnostic tools in air handling units, Energy
prognostics for building systems - a review (Part I), HVAC&R Res. 11 (1)
Effic. 7 (2) (2014) 335–351.
(2005) 3–25.
[29] S. Li, A Model-Based Fault Detection and Diagnostic Methodology for
[5] R. Yan, Z. Ma, G. Kokogiannakis, Y. Zhao, A sensor fault detection strategy for
Secondary HVAC Systems, Ph.D. Thesis, College of Engineering, Drexel
air handling units using cluster analysis, Autom. Constr. 70 (2016) 77–88.
University, 2009.
[6] Z. Du, B. Fan, J. Chi, X. Jin, Sensor fault detection and its efficiency analysis in
[30] P.N. Tan, M. Steinbach, V. Kumar, Introduction to Data Mining, Pearson
air handling unit using the combined neural networks, Energy Build. 72
Addison Wesley, USA, 2006.
(2014) 157–166.
[31] W.Y. Loh, Classification and regression trees, Wiley Interdiscip. Rev.: Data
[7] H. Yang, S. Cho, C.S. Tae, M. Zaheeruddin, Sequential rule based algorithms for
Min. Knowl. Discov. 1 (1) (2011) 14–23.
temperature sensor fault detection in air handling units, Energy Convers.
[32] L. Rokach, Data Mining with Decision Trees: Theory and Applications, World
Manage. 49 (8) (2008) 2291–2306.
Scientific, Singapore, 2008.
[8] Y. Zhao, J. Wen, F. Xiao, X. Yang, S. Wang, Diagnostic bayesian networks for
[33] Y. Yohannes, P. Webb, Classification and regression trees, in: CART: A User
diagnosing air handling units faults − Part I: faults in dampers, fans, filters
Manual for Identifying Indicators of Vulnerability to Famine and Chronic Food
and sensors, Appl. Therm. Eng. (2016), http://dx.doi.org/10.1016/j.
Insecurity, International Food Policy Research Institute, 1999.
applthermaleng.2015.09.121, available online.
[34] L. Breiman, J. Friedman, R.A. Olshen, C.J. Stone, Classification and Regression
[9] Y. Zhao, J. Wen, S. Wang, Diagnostic Bayesian networks for diagnosing air
Trees, Wadsworth, Inc, USA, 1984.
handling units faults −Part II: Faults in coils and sensors, Appl. Therm. Eng. 90
[35] L.F.R.A. Torgo, Inductive Learning of Tree-based Regression Models, Ph.D
(2015) 145–157.
Thesis, Department of Computer Science University of Porto, 1999.
[10] H. Wang, Y. Chen, C.W.H. Chan, J. Qin, J. Wang, Online model-based fault
[36] R. Development Core Team, R: A Language and Environment for Statistical
detection and diagnosis strategy for VAV air handling units, Energy Build. 55
Computing [Software], R Foundation for Statistical Computing, Vienna,
(2012) 252–263.
Austria, 2008.
[11] L.H. Chiang, R.D. Braatz, E.L. Russell, Fault Detection and Diagnosis in
[37] T. Therneau, B. Atkinson, B. Ripley, rpart: Recursive Partitioning and
Industrial Systems, Springer London, Great Britain, 2001.
Regression Trees, 2015.
[12] S. Wang, F. Xiao, AHU sensor fault diagnosis using principal component
[38] S. Milborrow, rpart.plot: Plot ’rpart’ Models: An Enhanced Version of
analysis method, Energy Build. 36 (2) (2004) 147–160.
’plot.rpart’, (2015).
[13] Z. Du, X. Jin, L. Wu, Fault detection and diagnosis based on improved PCA with
[39] H. Wickham, Ggplot2: Elegant Graphics for Data Analysis, Springer, New York,
JAA method in VAV systems, Build. Environ. 42 (9) (2007) 3221–3232.
2009.
[14] Z. Du, X. Jin, Multiple faults diagnosis for sensors in air handling unit using
[40] S. Li, J. Wen, Description of fault test in Summer of 2007, In: ASHRAE 1312
Fisher discriminant analysis, Energy Convers. Manage. 49 (12) (2008)
Report, 2007.
3654–3665.
[41] S. Li, J. Wen, X. Zhou, C.J. Klaassen, Development and validation of a dynamic
[15] W.Y. Lee, J.M. House, D.R. Shin, Fault diagnosis and temperature sensor
air handling unit model, Part 1, ASHRAE Trans. 116 (1) (2010) 45–56.
recovery for an air-handling unit, Trans.-Am. Soc. Heating Refrigerating Air
[42] S. Li, J. Wen, X. Zhou, C.J. Klaassen, Development and validation of a dynamic
Conditioning Eng. 103 (1997) 621–633.
air handling unit model, part 2, ASHRAE Trans. 116 (1) (2010) 57–73.
[16] J. Liang, R. Du, Model-based fault detection and diagnosis of HVAC systems
[43] L.K. Norford, J.A. Wright, R.A. Buswell, D. Luo, C.J. Klaassen, A. Suby,
using support vector machine method, Int. J. Refrig. 30 (6) (2007) 1104–1114.
Demonstration of fault detection and diagnosis methods for air-handling
[17] T. Mulumba, A. Afshari, K. Yan, W. Shen, L.K. Norford, Robust model-based
units, HVAC&R Res. 8 (1) (2002) 41–71.
fault diagnosis for air handling units, Energy Build. 86 (2015) 698–707.