Building and Environment 127 (2018) 268–276

Contents lists available at ScienceDirect

Building and Environment

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

Automated mobile sensing: Towards high-granularity agile indoor T

environmental quality monitoring
Ming Jina,∗, Shichao Liub, Stefano Schiavonb, Costas Spanosa
a
Electrical Engineering and Computer Sciences Department, University of California, Berkeley, CA 94720, USA
b
Center for the Built Environment, University of California, Berkekey, CA 94720, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Indoor environmental quality (IEQ) is a critical aspect of the built environment to ensure occupant health,
Indoor environmental quality comfort, well-being and productivity. Existing IEQ monitoring approaches rely on sensor networks deployed at
Mobile sensing selected locations to collect environmental measurements, and are limited in scale and adaptability due to
Spatio-temporal interpolation infrastructure cost and maintenance requirement. To enable high-granularity IEQ monitoring with agile adap-
Robotic sensing
tion to the dynamic indoor environment, we propose an “automated mobile sensing” system that dispatches a
Smart building
sensor-rich navigation-capable robot to actively survey the indoor space. Data collected in this fashion is sparse
Energy efficiency
in the joint temporal and spatial domain, and cannot be used directly for IEQ evaluation. To deal with this
special characteristics, we developed a spatio-temporal interpolation algorithm to capture the global trend and
local variation in order to use the data efficiently to reconstruct the IEQ dynamics. We compared the perfor-
mance of the automated mobile sensing with a dense sensor network in a laboratory where we measured the air-
change effectiveness (ASHRAE standard 129) for four different conditions. Results indicate that automated
mobile sensing is able to accurately estimate the parameters with a minimal sensor cost and calibration effort.
Potential applications of this system include indoor thermal comfort, lighting, indoor air quality and acoustic
monitoring, pollutant source identification, and building commissioning. We shared publicly the source codes
for robot control, sensor setup, and interpolation algorithm to encourage comparison study and further devel-
opment.

1. Introduction temperature, humidity) need to be continuously monitored when oc-

cupants take a “right-now” thermal comfort survey, according to
Smart buildings are cyber-physical energy systems (CPES) that in- AHSRAE/CIBSE/USGBC PMP [10]. LEED suggests CO2 monitoring in
tegrate sensing, data analytics, and control to provide essential services all densely occupied spaces. In addition, IEQ assessment involves con-
to the occupants. Buildings consume about 40% of primary energy in taminants sampling in all occupied spaces, such as volatile organic
the U.S. and there is a fundamental drive for buildings to be energy compound (VOC) and particulate matter (PM) [11]. Guidelines, stan-
efficient [1,2]. As people spend about 90% of their time indoors, they dards and rating systems recognize that more IEQ monitoring would be
should also be human-centric by focusing on improving human health, valuable but affordability constrain limits what is suggested to be used.
comfort, well-being and productivity, and well-being [3–6]. This could Due to complex indoor structures and dynamic environment, IEQ
be achieved effectively by monitoring and enhancing indoor environ- parameter distributions are often inhomogeneous, resulting in spatial
mental quality (IEQ), such as indoor air quality, thermal comfort, variations in thermal environment and indoor contaminant exposure
lighting and acoustics [7–9]. IEQ monitoring has been recognized as [12–14]. Furthermore, applications of personalized heating/cooling
one of the fundamental strategies to obtain credits by various guidelines devices, aiming to reduce building energy use, augment such in-
and rating systems, such as American Society of Heating, Refrigerating homogeneity [15–20]. Consequently, spatio-temporal monitoring of
and Air-Conditioning Engineers (ASHRAE)/Chartered Institution of indoor environment can provide an comprehensive IEQ assessment.
Building Services Engineers (CIBSE)/U.S. Green Building Council Key challenges in the objective IEQ assessment of commercial
(USGBC) Performance Measurement Protocols for commercial buildings buildings involve accurate, easy-to-use, and scalable sensing systems
(PMP) [10] and Leadership in Energy and Environmental Design [21]. An effective approach is to implement wireless sensor networks;
(LEED) [11]. For instance, environmental parameters (e.g., however, despite the continuous reduction in sensor cost and

∗
Corresponding author.
E-mail addresses: jinming@berkeley.edu (M. Jin), scliu@berkeley.edu (S. Liu), schiavon@berkeley.edu (S. Schiavon), spanos@berkeley.edu (C. Spanos).

https://doi.org/10.1016/j.buildenv.2017.11.003
Received 31 August 2017; Received in revised form 29 October 2017; Accepted 2 November 2017
Available online 06 November 2017
0360-1323/ © 2017 Elsevier Ltd. All rights reserved.
M. Jin et al. Building and Environment 127 (2018) 268–276

Fig. 1. Automated mobile sensing system overview. The robotic

platform can either work alone or with static sensor network to
actively estimate building context and facilitate building control.

simplification in deployment, infrastructure investment and main- building environment and diagnose potential faults. Static sensors are
tenance might still remain a concern in the near future, especially when deployed in a space to continuously monitor environmental parameters
considering monitoring numerous IEQ variables simultaneously. [1]; nevertheless, limited by cost, the deployment is often sparse in
Moreover, many sensors that require a significant amount of power locations or absent, especially for expensive sensors like CO2. In addi-
(e.g., hot wire anemometer) can not easily become wireless. Ad- tion, while indoor environment is often inhomogeneous and un-
ditionally, buildings might undergo several renovations in their life- predictable, the stationary sensors may not always be deployed in the
cycle, so agility is essential to adapt to the changing environment. optimal locations to reflect indoor environment. Personal monitoring
systems, such as using infrared thermography [31] and physiological
1.1. Main contributions and objectives measurements [32,33,37] can offer assessment of individual comfort
and inform building operation system of proper adjustments in real-
Differentiated from existing approaches of deploying static sensors time; however, they require users to be equipped with special instru-
for indoor monitoring, we propose a sensing paradigm of “automated ments or sensors and may involve privacy concerns. For some IEQ
mobile sensing” by leveraging a navigation-enabled sensing-capable parameters like indoor air quality, the effect on productivity and health
mobile robot (see Fig. 1 for the overall architecture). This represents a may be long-term and cannot be readily captured by physiological
paradigm of “active inference”, where the robot can plan its path to measurements.
take representative measurement samples at locations of interests, as Mobile carts, such as an instrumented chair-like cart [34] and the
compared to “passive inference” where the data collection is limited by IEQ cart [35,36] can hold multiple sensors to take measurements si-
the geolocations of static sensors. multaneously at a given location. While the results are comprehensive,
From a data analytic perspective, unlike data from static sensors, the the carts often require considerable labor cost and manual navigation.
samples taken by the robot is highly sparse in time and space, as illu- Several studies exist to deploy robots for monitoring and identifying
strated in Fig. 3. While existing interpolation mainly focuses on the pollutants both indoor and outdoor [38–41]; however, the methods do
spatial domain [22,23], we propose a data-efficient spatio-temporal not distinguish the global trend of physical parameters from their local
(ST) interpolation method that extracts local and global trends and variations, which might lower the estimation accuracy, and the results
constructs an informative visualization of IEQ. Through experimental have not been validated against a ground truth, which requires a dense
evaluations of zone air distribution effectiveness (air-change effective- sensor network for comparison.
ness, ACE), automated mobile sensing is compared with static sensing
with a dense sensor network required by the ASHRAE standard 129
[24]. Note that the air-change effectiveness experiment is only used to 2.2. Continuous interpolation from discrete measurements
demonstrate our novel platform, rather than to investigate possible
factors that influence its value, for which we refer the readers to more Data from static or mobile sensor measurements is highly sparse and
established works [20,25–27]. It is, therefore, the objective of this requires interpolation for informative visualization. Spatial interpola-
paper to describe the novel “automated mobile sensing” system for tion is a well-studied topic in geostatistical analysis and image pro-
indoor environmental quality monitoring, enabled by a sensor-rich cessing communities, where methods like Kriging and Markov random
navigation-capable robot to actively survey the indoor space. field (MRF) are among the most prominent [22,23]. Kriging has also
been combined with Gaussian MRF [42], Bayesian network [43], and
2. Brief literature review principle component analysis [44] to improve the computational effi-
ciency. In practice, this means that the algorithm can analyze a large
2.1. Indoor environmental quality assessment amount of data within limited time span, thus enabling large-scale
sensing.
IEQ assessment can be conducted using occupant surveys [5,28–30], Since Kriging is efficient with sparse data, it has been generalized to
personal monitoring [31–33], and sensor measurements [34–36]. Sur- spatio-temporal interpolation [22,45]. Shape functions have also been
veys provide subjective IEQ evaluation from occupant perspectives; introduced based on finite element mesh generation [46]. Variational
however, survey design requires systematic effort to avoid bias and Gaussian-process factor analysis is proposed to model the dynamics of
confusion, and the results can not be updated frequently due to user spatio-temporal data [47]. Prior works assume multiple time series data
fatigue. Several online or mobile tools have been developed to allow from individual sensor stations, which requires continuity in time at a
users to vote their thermal or lighting preferences in real time [17,28]; specific location; but the data from mobile sensing robot poses the
however, the responses may reveal only subjective perceptions, like challenge of high sparsity and non-continuity in time and space (Fig. 3).
“the air is stale”, but it rarely gives hints about the causes, such as Differentiated from existing interpolation methods, our method can
increased indoor pollution caused by low outdoor air flow rate or un- efficiently capture spatial and temporal dynamics by constructing
pleasant thermal environment due to malfunctioning mechanical sys- global and local trend estimators based on highly sparse data.
tems.
Objective measurements, taken by static or mobile sensors during
daily operation or performance commissioning, can accurately depict

269
M. Jin et al. Building and Environment 127 (2018) 268–276

Fig. 2. (a) Snapshot of the environmental sensing platform. (b) Schematic of sensor integration and data communication. Snapshots of (c) the robotic platform, (d) Augmented reality tag
indoor positioning, and (e) camera-enabled simultaneous localization and mapping (SLAM).

3. Methodology correction (ABC) beforehand. In particular, we calibrated the CO2

sensors (K-30) using an off-the-shelf HOBO MX1102 CO2 logger.
3.1. Integrated system architecture Programmed under the robot operating system (ROS), a collection
of software frameworks for robot software development, it runs map-
This section introduces the environmental sensing platform and the ping, positioning, and navigation algorithms autonomously. Mounted
robotic base (Fig. 2) as the essential components in the autonomous with ESP, the mobile robot (Turtlebot 2) can be controlled remotely or
sensing system. make decisions based on real-time sensing data (Fig. 2).
Based on an Arduino microcontroller board, the environmental
sensing platform (ESP) is designed on a software-level to detect and 3.1.1. Indoor positioning
report sensor faults automatically to alert users, and to work instantly In a new and dynamic environment, the robot can operate without a
once powered on, relieving laborious configuration or setup (Fig. 2). detailed floormap. Based on the depth image from Kinect camera, the
A comprehensive list of sensors are integrated in the ESP to monitor Simultaneous Localization and Mapping (SLAM) problem is solved with
indoor environmental parameters, including temperature and humidity, particle filter by tracking the robot position relative to the surroundings
light level, PM2.5, CO2, and organic volatile compound (VOC) [48] (Fig. 2e).
(Table 1). Data are sampled and uploaded to a server using a WiFi With the additional user-provided layout, a higher-precision loca-
communication link at an interval of 10 s, which are also pushed to a tion estimation is enabled by an augmented reality tag system (Fig. 2d).
front-end visualization portal (hosted at dweet.io) for real-time mon- Upon the initial positioning by SLAM, the estimate is updated by
itoring. In addition, all sensors are calibrated using automatic baseline reading the distance and direction to nearby reference points marked by

270
M. Jin et al. Building and Environment 127 (2018) 268–276

Table 1
Sensing modules of ESP.

Environmental parameter Module Performance Price

Temperature MCP9808 Accuracy: 0.25 °C typical precision over −40 °C to 125 °C range $4.68
Illuminance SI1145 Resolution: 100 mlx $9.95
CO2 K-30 Measurement range: 0–10000 ppm Accuracy: ± 3% of measurement Response time: 20 s diffusion time $85
PM2.5 SEN0177 Measuring pm range: 0–500 μg/m3 $46.90
Organic volatile compound TGS2620 Typical detection range: 50–5000 ppm Sensitivity: 0.3–0.5 in ethano $8.90

augmented reality tags, which are fiducial markers deployed in advance Sensor & High-granularity
at locations indicated on the room layout. During the positioning pro- location data interpolation
cess, the relative distances and angles are processed by the triangula-
tion algorithms to revise the location estimate [49]. This also enables
more robust estimation against dynamic changes in the room as long as

+
the augmented reality tags remain visible to the robot [49].
Step 1: Step 2: Step 3:
3.1.2. Navigation and collision avoidance
ST binning Global trend Local trend
Given a set of goal points, the robot is navigated using a global
planner to set the shortest path based on the current knowledge about
Fig. 4. Illustration of the ST interpolation algorithm, including ST binning for data
the space. However, when some moving objects, like occupants, ob- smoothing, global trend extraction, and local variation estimation. The outcome is an
struct the planned path, a local planner is employed to avoid obstacles. interpolation function fˆ that encodes a high-granularity interpolation map.
Supported by ROS, robot specs like velocity, angular speed, and goal
tolerance can be set to accommodate specific requirements.
points.
3.2. Spatio-temporal interpolation algorithm
The high-granularity map that depicts the indoor environment
evolution is embedded in the ST interpolation function, given by:
Compared to data collected by static sensor stations, measurements
from mobile sensors cover the whole space with higher granularity; fˆ (s , t ) = fˆglobal (t ) + fˆlocal (s , t ), (1)
however, for each location, the samples are conducted sequentially in
multiple locations (see Fig. 3 for an illustration). The problem is stated where fˆglobal (t ) and fˆlocal (s , t ) are the global and local trends, obtained
as below: as follows.
Given the mobile sensing data D = {(s1, t1, v1), …, (sn, tn, vn )} , where The global trend depicts the average evolution of the phenomenon
si = (x i , yi ) and ti are the spatial coordinate and timestamp, and vi is the independent of individual locations, since data pooling provides sufficient
actual value, our goal is to find a function, fˆ : 2 ×  ↦  , which es- samples for estimation. Based on the locally weighted scatterplot
timate values at unexplored locations at certain time. smoothing (LOWESS) method [50], the global trend, fˆglobal (t ), at time
A data-driven approach to spatio-temporal (ST) interpolation is t ∈ [tc − h (tc ), tc + h (tc )], is given by:
adopted based on statistical decision theory. The variation of an indoor 1
fˆglobal (t ) = β0 + β1 (t − t0) + β2 (t − t0 )2 , t ∈ [tc − h (tc ), tc + h (tc )]
environment exhibits both a global trend, as dominated by outdoor 2
weather, building envelope and Heating, Ventilation and Air (2)
Conditioning (HVAC) operation, as well as a local trend, as influenced
where tc is the center point, h (tc ) is the smoothing span, and β0, β1, β2
by occupants, inhomogeneous air turbulence, pollutant source, and
are parameters given by:
furniture.
n 2
The proposed algorithm, therefore, has three key steps (Fig. 4): 1
(β0, β1, β2) = argmin ∑ wi (tc ) ⎛vi − ⎛β0 + β1 (ti − tc ) + β2 (ti − tc )2⎞ ⎞
i=1 ⎝ ⎝ 2 ⎠⎠
1. ST binning: consider a 3D-space (xy and t axes represent space and
(3)
time, respectively) divided into 3D cubes based on spatial and
u 3 )3
temporal resolutions. Data points are binned and aggregated to re-
duce measurement errors.
where wi (tc ) = W ( ) and W (u) = ⎧⎨⎩((1 −0
ti − tc
h (tc )
u ≤1
u >1
obtain the

2. Global trend extraction: a regression trend is fitted by, e.g., locally data point weights.
weighted scatterplot smoothing (LOWESS) [50], to capture the While the global trend function is applicable to all locations, the
global variation. local function fˆlocal captures the spatial variation, which is encoded in
3. Local variation estimation: based on the residues from global trend, the residuals ri = vi − fˆglobal (ti ) . Based on the empirical risk minimization
a local variation function is approximated and applied on unknown (ERM):

a b c Fig. 3. Illustration of data characteristics in the spatial and temporal

domains for (a) static stations, (b,c) automated mobile sensing,
where (c) depicts a situation of active exploration around the spot of
critical event (star).

271
M. Jin et al. Building and Environment 127 (2018) 268–276

n
1 to the breathing level within the ventilated space”. It is also known as “zone
fˆlocal = argmin ∑ l (f (si, ti), ri)
f ∈ n (4) air distribution effectiveness” [51]. ACE is calculated from the age of air.
i=1
The age of air is the “average time elapsed since molecules of air in a given
where  delineates the range of estimators, e.g., the class of linear volume of air entered the building from outside”, in other words, the age of
regressors, and l:  ×  →  is the loss function, which penalizes error air at a specific location in a building refers to the time for a bulk of
in estimation, e.g., the squared loss l (a, b) = (a − b)2 . Implicitly, we outdoor air to reach the position after entering the building. A
assume that points close in space and time are also close in values, “younger” age of air represents that the air is fresher, which is often
which is generally true for indoor environment. located around supply diffusers. An indoor space with a smaller age of
A variety of algorithms to capture local variation have been im- air on average often has a higher ventilation efficiency and air quality.
plemented in our open toolset, such as K-nearest neighbor (KNN), According to ASHRAE standard 129, the age of air can be estimated
Lasso, support vector regression (SVR), adaptive boosting (AdaBoost), using a tracer gas step-up or decay procedure (described in Section 4.3),
random forest, and extra trees [50], which are trained by ERM with which often uses non-toxic, non-flammable, and environmentally
different loss functions. The implementation code in Python is available friendly gases like CO2 as the tracer gas [24]. The ISO 16000 standard
(see the link in section 5.3). also describes the use of a single tracer gas to determine the local mean
age of air by using concentration decay or homogeneous constant
4. Experiments emission [52]. According to the decay procedure, the air age A (s ) at
location s is given by:
4.1. Environmental chamber Cavg (s )
A (s ) = (tstop − tstart ) ,
C (s , tstart ) (5)
The experiment took place in a climate chamber in the Center for
the Built Environment (CBE), which can precisely control supply air- where tstart is the time at the beginning of the decay, tstop is the time
flow rate, indoor temperature ( ± 0.5 °C) and humidity ( ± 3%). The when the procedure ends, C (s , t ) is the tracer gas concentration at lo-
ventilation system employs underfloor air distribution (UFAD), where cation s at time t, and Cavg (s ) is the time-averaged concentration be-
air is supplied from one or two linear grille diffusers on the floor (shown tween tstart and tstop . Since the samples are usually discrete in time, the
in Fig. 5). By providing cool air at a low momentum from the floor level time-averaged concentration Cavg (s ) can be estimated by
1 n
and utilizing buoyancy generated from indoor heat sources (e.g., oc- n
∑i = 1 C (s , tstart + iΔt ) , where Δt is the sampling interval, and n is the
cupants) to displace room air, UFAD often achieves a thermally strati- number of samples between tstart and tstop .
fied air distribution. During the experiment, all supply air came from Based on the air age measurement, the ACE E (s ) is given by Ref.
outside, and there was not circulated air. We assumed that outdoor air [24]:
CO2 concentration was 400 ppm. The supply air flowrate during the A (sex )
experiments was maintained at 79 ± 11 m3/hr. E (s ) = ,
A (s ) (6)
As for heat sources, two sedentary thermal manikins, each with total
sensible heat loss of 68.7 W ( ± 5%) were placed, representing female where A (sex ) indicates the age of air at the exhaust vent. By definition,
subjects for office work (Fig. 5). In addition, we placed one heater panel this metric describes an air distribution system's capacity to deliver
(size: 0.5 × 0.2 m; power: 200 W) under each table to investigate how ventilated air to an indoor space. The local ACE represents the effec-
personal heating device affected the zone air distribution effectiveness tiveness of outdoor air delivery to one specific point in a space.
and examined the robustness of the algorithms in the given experi- The ACE is typically 1.0 for a well mixing ventilation system in
mental setup. cooling and equal or higher than 1.2 for displacement ventilation [51].
Ten (10) ESPs were deployed in the occupied zone, which measured In general, a space with a higher ACE is associated with a better air
CO2 concentrations at 10 locations and provided the standard assess- distribution system.
ment of air-change effectiveness according to the ASHRAE standard
129. The sensors were placed at a height of 1.5 m, where detailed lo- 4.3. Experimental procedure
cations are shown in Fig. 5.
The experimental procedure was designed according to the standard
tracer-gas decay process by ASHRAE standard 129, as follows:
4.2. Description of air-change effectiveness
1. First, beverage-grade CO2 was injected into the chamber with a
According to the ASHRAE standard 129 [24] the air-change effec-
ceiling fan and a standing fan well mixing the injected CO2
tiveness (ACE) is “a measure of the effectiveness of outdoor air distribution

Fig. 5. Left: testbed floorplan, showing the air inlet diffusers, heaters, tables, CO2 sensor and source locations, and thermal manikins. The sensors are placed at a uniform height of 1.5 m.
Right: testbed snapshot, showing the static sensor stations (1–4), thermal manikins to model realistic heat sources (5,6), robot (7), floor heaters (8,9), and CO2 source (10).

272
M. Jin et al. Building and Environment 127 (2018) 268–276

Fig. 8. Plot of air age estimation by static stations and mobile robot with different local
variation methods in Exp B.

location and moved to the next one. Fig. 6 illustrates the moving
route of the platform that stayed at each location for 45 s for CO2
sampling. The platform moved at a speed of 2 m/s to the next lo-
cation until the completion of CO2 recording at all locations. The
moving speed was optimized so that airflow pattern was disturbed
at a minimum level and CO2 decay at 10 locations can be captured
in roughly 8 min.

We compared ACE at various air distribution conditions measured

by the mobile sensing with that obtained according to ASHRAE stan-
dard 129 to validate the performance of the automobile platform.
Table 2 describes the investigated scenarios by altering the intensities
of supply air momentum and heat sources, which influences zone air
distribution effectiveness as represented by ACE. It is worth noting that
the main goal for varying indoor configurations was to validate the
reliability of the platform rather than thoroughly examining how ACE
could be influenced accordingly. In addition, we repeated one scenario
twice (Exp B &C) to assess the repeatability of the indoor air evaluation
and the robustness of mobile sensing. Supply airflow rate remained
Fig. 6. (a) Spatio-temporal trace of mobile measurements. (b) Plane view of the visiting constant for all the scenarios.
trace of the robot.

5. Results and discussion

Table 2
Experimental conditions for vents and heater status. We conducted the experiments listed in Table 2 with the robot,
where the stationary sensors were used as the standard results for
Experiment i.d. Exp A Exp B, C Exp D Exp E
comparison. For each experiment, the robot collected a dataset, D ,
Supply diffusers Diffuser 1 Diffuser 1 & 2 Diffuser 1 Diffuser 1 & 2 which was used to train both the global and local trend estimators in
Floor heater status Off Off On On (1).

throughout the chamber. It was assumed that CO2 was well mixed in 5.1. Age of air
the chamber short after the injection period [53].
2. When indoor CO2 concentration was elevated to approximately Firstly, we demonstrate that the mobile sensing captures the actual
3000 ppm, we turned off the CO2 injection and ceiling fans. The dynamics of CO2 concentration in the space. Since the measurements
injection period took roughly 5 min. from the mobile robot change in both space and time, we visualize CO2
3. After the ceiling fans were off, the indoor airflow pattern was re- concentrations of mobile sensors and compare them with the stationary
established after 5 min [54]. data, shown in Fig. 7. While the mobile data reveals the overall decay
4. The automobile platform measured CO2 concentration from the start trend, it also effectively differentiate the spatial distribution of CO2 at
different locations, which is helpful to build the local trend estimator

Fig. 7. Left: collocation of the sensor measurements from the static

station S5 (lines) and the mobile robot (crosses). Right: comparison
of station and robot measurements at the same location and time.
This indicates that the robot measurements are representative of the
spatial temporal concentrations.

273
M. Jin et al. Building and Environment 127 (2018) 268–276

Fig. 9. Comparison of air age estimation given by the stations (blue)

and robot (red) at 7 distinct locations in the repeated experiments
(Exp B and C). The box indicates the min/max range of estimation
across two experiments. (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version
of this article.)

Table 3
Comparison results of RMSE for ACE estimation and the corresponding Pearson correlation coefficients (reported in brackets) between the static stations and robot estimation. Mean:
baseline estimation. Other algorithms represent different local variation estimators.

Mean OLS Ridge AdaBoost SVR KNN RandForest ExtraTrees

Exp A 0.077 [0.00] 0.074 [0.73] 0.074 [0.73] 0.058 [0.81] 0.071 [0.82] 0.076 [0.63] 0.072 [0.74] 0.070 [0.76]
Exp B 0.061 [0.00] 0.027 [0.96] 0.027 [0.96] 0.038 [0.77] 0.039 [0.97] 0.031 [0.99] 0.033 [0.91] 0.030 [0.98]
Exp C 0.089 [0.00] 0.078 [0.79] 0.078 [0.79] 0.072 [0.82] 0.066 [0.88] 0.074 [0.82] 0.068 [0.85] 0.071 [0.88]
Exp D 0.043 [0.00] 0.036 [0.62] 0.036 [0.62] 0.031 [0.76] 0.041 [0.84] 0.029 [0.81] 0.031 [0.84] 0.029 [0.81]
Exp E 0.044 [0.00] 0.044 [0.80] 0.044 [0.80] 0.050 [0.79] 0.039 [0.77] 0.045 [0.82] 0.044 [0.73] 0.047 [0.73]

0.065 (Section 3.2). Indeed, by comparing the robot measurements with the
ACE estimation error

station data collected at the same time and location, we can see that
they are almost identical (right plot of Fig. 7).
0.06
To evaluate the accuracy of different interpolation algorithms out-
lined in Section 3.2, we compare the age of air (see Equ. 5) estimated by
0.055 the robot to the estimation by static sensors, as is shown in Fig. 8. The
baseline method, denoted as “Mean”, conducts the global trend estima-
tion (Equ. (2)) but disregards the local variation (Equ. (4)), thus creating a
0.05
uniform estimation across the space. On the contrary, methods based on
KNN, random forests, and extra trees distinguish the spatial distribu-
0.045 tions and produce estimation very close to the static measurements
ExtraTrees
Mean
OLS

KNN
RandForest
AdaBoost

(Fig. 7). Furthermore, for the same ventilation setups, the robot esti-
Ridge

SVR

mation exhibits little variation in the repeated experiments (Exp B and

C), comparable to the station performance (Fig. 9), implying that the
influence of robot movement on CO2 distribution is negligible.

Fig. 10. Average estimation error of ACE for different methods, where Mean is the 5.2. Air-change effectiveness
baseline method. The RMSE of Mean reflects the standard deviation of ACE across the
space. The reduced RMSE enables accurate differentiation of ACE at different locations.
As for the ACE evaluation, results indicate that the proposed

Fig. 11. Visualization of spatial distribution of ACE parameters estimated by the static sensors (top plots) and the robotic platform (bottom plots). The five experimental conditions are
listed in Table 2.

274
M. Jin et al. Building and Environment 127 (2018) 268–276

interpolation methods significantly outperform the baseline model effectiveness in terms of supply air short-circuit is not uncommon for
(Table 3 lists the mean squared error for all the experiments, and HVAC operated at heating mode [55]. The proposed system could as-
Fig. 10 illustrates the average performance). Since the typical values of sess indoor ACE without deploying dozens of trace gas sensors [24].
ACE are around 1, the reduction of RMSE from 0.063 (Mean) to 0.049 Nevertheless, the platform may not be suitable in residential houses
(ExtraTrees) in Fig. 10 is significant from a practical point of view, where compact furniture and stairs may limit the moving range of the
which also enables more accurate depiction of the spatial differences of platform. The source codes for robot control, sensor setup, and ST al-
ACE. The Pearson correlation coefficients are close to one for the list of gorithm are made available at http://www.jinming.tech/software/ to
algorithms, indicating that the ACE estimation from interpolated ro- encourage comparison study and further development in this area.
botic measurements follow the same trend as the station estimation.
However, some algorithms like Ridge had relatively low correlation 6. Conclusion
coefficient of 0.62 for Exp D, which was outperformed by Random
Forest with coefficient of 0.84. This indicates that Random Forest had a This study proposed the “automated mobile sensing” paradigm for
better generalization capability than Ridge for interpolation at un- high-granularity, agile and scalable indoor environmental quality
known points. In addition, it can be observed that the correlation monitoring. The integrated sensing system consists of a mobile base and
coefficients of Random Forest are relatively low for Exp A and E as an environmental sensing platform which is capable of measuring a
compared to the others, since the spatial variation of ACE in these two range of environmental parameters. To derive actionable insights from
experiments was not substantial as indicated by the static sensor mea- the collected data that are sparse in both spatial and temporal domains,
surements in Fig. 11. Note that the correlation for Mean method is 0 we developed a spatio-temporal interpolation algorithm that leverages
because the prediction from Mean is a constant, which fails to capture a hierarchical approach to reconstruct continuous mapping of the in-
the spatial variation of ACE across the chamber. door environment. We demonstrated the mobile platform in a labora-
The ACE measurements shows the ventilation effectiveness of the tory experiment of measuring air-change effectiveness. By comparing
different system configurations. In Fig. 11, ACE results for the static and the measurements from the mobile platform and those from a standard
the robot case are shown. For the robot data the interpolation is done dense sensor network, we showed that the automated mobile sensing
using KNN to produce a high-granularity ACE mapping. The visuali- approach was able to determine the air-change effectiveness with high
zation indicates that ACE improves with reducing mixing effects with spatial granularity and accuracy.
more supply vents (Exp B, C, E) and adding heating sources (Exp D and
E)for the given indoor configuration in this study. The results are ex- Acknowledgements
pected. The air distribution system with two supply diffusers reduces
supply air momentum and mixing effects, resulting in a higher thermal This research is funded by the Republic of Singapore's National
stratification in the room. In addition, the heaters under the tables also Research Foundation through a grant to the Berkeley Education
increase thermal stratification, while minimizing the potential disrup- Alliance for Research in Singapore (BEARS) for the Singapore-Berkeley
tion in terms of thermal plumes to room airflow. The increased thermal Building Efficiency and Sustainability in the Tropics (SinBerBEST)
stratification helps generate displacement ventilation in the room and Program. BEARS has been established by the University of California,
improve ACE. Berkeley as a center for intellectual excellence in research and educa-
tion in Singapore. We would like to express our gratitude to Fred
Bauman for his suggestions during the lab experiments.
5.3. Potential applications
References
The proposed method is able to capture indoor environmental
parameters both spatially and temporally. The approach can sig- [1] K. Weekly, M. Jin, H. Zou, C. Hsu, A. Bayen, C. Spanos, Building-in-briefcase (bib),
nificantly reduce the sensing infrastructure cost.1 It could be applied to arXiv preprint arXiv:1409.1660.
assess indoor environmental quality and for continuous commissioning. [2] M. Jin, N. Bekiaris-Liberis, K. Weekly, C.J. Spanos, A.M. Bayen, Occupancy detec-
tion via environmental sensing, IEEE Trans. Automation Sci. Eng. 99 (2017) 1–13.
In the future deployment, the robot can rover around a space to con-
[3] M. Jin, H. Zou, K. Weekly, R. Jia, A.M. Bayen, C.J. Spanos, Environmental sensing
tinuously monitor indoor IEQ, and can meanwhile interact with users by wearable device for indoor activity and location estimation, 40th Annual
and facilitate automatic building control, which allows for active in- Conference of the IEEE Industrial Electronics Society, 2014, pp. 5369–5375.
[4] Indoor climate and productivity in offices, REHVA guidebooks 6, in: P. Wargocki,
ference and decision making in real-time building operation. Since the
O.A. Seppnen (Eds.), REHVA, Federation of European Heating and Air-conditioning
robot can move at a speed of about 2 m/s, one single robot is needed to Associations, Brussels, Belgium, vol. 6, 2006.
cover a typical office building floor, though it cannot yet climb the [5] M. Frontczak, S. Schiavon, J. Goins, E. Arens, H. Zhang, P. Wargocki, Quantitative
stairs or navigate between floors. relationships between occupant satisfaction and satisfaction aspects of indoor en-
vironmental quality and building design, Indoor air 22 (2) (2012) 119–131.
It has capacities to navigate across large indoor spaces to search [6] M. Jin, R. Jia, C. Spanos, Virtual occupancy sensing: using smart meters to indicate
locations where local environment is unacceptable. Pollutant sources your presence, IEEE Trans. Mob. Comput. 16 (11) (2017) 3264–3277.
(e.g., particles from occupants or penetrated through open windows) [7] J. Sundell, On the history of indoor air quality and health, Indoor air 14 (s7) (2004)
51–58.
may be hard to capture using a limited number of stationary sensors. [8] K. Parsons, Human Thermal Environments: the Effects of Hot, Moderate, and Cold
This platform can access an indoor space extensively and it could be Environments on Human Health, Comfort, and Performance, CRC press, 2014.
modified to identify pollutant sources through active sampling (see [9] S. Prasow, Acoustics in green buildings: refining the concept of environmentally
quality while improving occupant health and productivity synergistically, J. Acoust.
Fig. 3c). Soc. Am. 123 (5) (2008) 3095–3095.
Building system commissioning could also employ the platform to [10] ASHRAE/CIBSE/USGBC, Performance Measurement Protocols for Commercial
diagnose the efficacy of HVAC and lighting systems. For instance, non- Buildings, ASHRAE Transactions, American Society of Heating, Refrigerating and
Air-Conditioning Engineers, January 1, 2010 ISBN-13: 978-1933742793, ISBN-10:
uniformity of air temperature in the underfloor air distribution (UFAD)
1933742798, 298 pages.
plenum indicates an ineffective system design or implementation. The [11] US Green Building Council, Leadership in Energy and Environmental Design LEED,
platform can map the temperature distribution inside the plenum which Building Design and Construction, V4.
[12] W. Zhang, K. Hiyama, S. Kato, Y. Ishida, Building energy simulation considering
is often not easily accessible for traditional temperature monitoring
spatial temperature distribution for nonuniform indoor environment, Build.
(e.g., thermistor probe deployment). Moreover, low ventilation Environ. 63 (2013) 89–96.
[13] A. Bulińska, Z. Popiołek, Z. Buliński, Experimentally validated cfd analysis on
sampling region determination of average indoor carbon dioxide concentration in
1
The robot employed in the study is TurtleBot 2, which costs about 2000 USD: http:// occupied space, Build. Environ. 72 (2014) 319–331.
www.turtlebot.com/turtlebot2/ [access date: 10/2017]. [14] M. Jin, N. Bekiaris-Liberis, K. Weekly, C. Spanos, A.M. Bayen, Sensing by proxy:

275
M. Jin et al. Building and Environment 127 (2018) 268–276

occupancy detection based on indoor co2 concentration, in: proceedings of the [36] H. Kim, et al., Field-test of the new ashrae/cibse/usgbc performance measurement
international conference on mobile ubiquitous computing, Syst. Serv. Technol. protocols for commercial buildings: basic level, ASHRAE Trans. 118 (2012) 135.
(2015) 1–10. [37] S. Takada, S. Matsumoto, T. Matsushita, Prediction of whole-body thermal sensa-
[15] A.K. Melikov, Personalized ventilation, Indoor Air 14 (s7) (2004) 157–167. tion in the non-steady state based on skin temperature, Build. Environ. 68 (2013)
[16] S. Schiavon, B. Yang, Y. Donner, V.-C. Chang, W.W. Nazaroff, Thermal comfort, 123–133.
perceived air quality, and cognitive performance when personally controlled air [38] Y. Chen, H. Cai, Z. Chen, Q. Feng, Using multi-robot active olfaction method to
movement is used by tropically acclimatized persons, Indoor air 27 (3) (2017) locate time-varying contaminant source in indoor environment, Build. Environ. 118
690–702. (2017) 101–112.
[17] L.J. Ratliff, M. Jin, I.C. Konstantakopoulos, C. Spanos, S.S. Sastry, Social game for [39] A. Lilienthal, T. Duckett, Building gas concentration gridmaps with a mobile robot,
building energy efficiency: incentive design, IEEE Annual Allerton Conference on Robotics Aut. Syst. 48 (1) (2004) 3–16.
Communication, Control, and Computing, 2014, pp. 1011–1018. [40] M. Reggente, A. Mondini, G. Ferri, B. Mazzolai, A. Manzi, M. Gabelletti, P. Dario,
[18] S. Liu, L. Yin, W.K. Ho, K.V. Ling, S. Schiavon, A tracking cooling fan using geofence A.J. Lilienthal, The dustbot system: using mobile robots to monitor pollution in
and camera-based indoor localization, Build. Environ. 114 (2017) 36–44. pedestrian area, 3rd Biannual International Conference on Environmental Odour
[19] H. Zhang, E. Arens, Y. Zhai, A review of the corrective power of personal comfort Monitoring and Control, vol. 23, 2010, pp. 273–278.
systems in non-neutral ambient environments, Build. Environ. 91 (2015) 15–41. [41] S. Thrun, C. Martin, Y. Liu, D. Hahnel, R. Emery-Montemerlo, D. Chakrabarti,
[20] D. Faulkner, W.J. Fisk, D.P. Sullivan, D.P. Wyon, Ventilation efficiencies of desk- W. Burgard, A real-time expectation-maximization algorithm for acquiring multi-
mounted task/ambient conditioning systems, Indoor Air 9 (4) (1999) 273–281. planar maps of indoor environments with mobile robots, IEEE Trans. Robotics
[21] D. Heinzerling, S. Schiavon, T. Webster, E. Arens, Indoor environmental quality Automation 20 (3) (2004) 433–443.
assessment models: a literature review and a proposed weighting and classification [42] L. Hartman, O. Hössjer, Fast kriging of large data sets with gaussian markov random
scheme, Build. Environ. 70 (2013) 210–222. fields, Comput. Statistics Data Analysis 52 (5) (2008) 2331–2349.
[22] N. Cressie, C.K. Wikle, Statistics for Spatio-temporal Data, John Wiley & Sons, 2015. [43] J. Dearmon, T.E. Smith, Gaussian process regression and bayesian model averaging:
[23] S. Geman, D. Geman, Stochastic relaxation, gibbs distributions, and the bayesian an alternative approach to modeling spatial phenomena, Geogr. Anal. 48 (1) (2016)
restoration of images, IEEE Trans. pattern analysis Mach. Intell. 6 (1984) 721–741. 82–111.
[24] ASHRAE Standard 129, Measuring Air-change Effectiveness, ASHRAE, Atlanta, GA, [44] D. Mendez, M. Labrador, K. Ramachandran, Data interpolation for participatory
US, 1997, pp. 1–20. sensing systems, Pervasive Mob. Comput. 9 (1) (2013) 132–148.
[25] W.J. Fisk, D. Faulkner, D. Sullivan, F. Bauman, Air change effectiveness and pol- [45] K.V. Mardia, C. Goodall, E.J. Redfern, F.J. Alonso, The kriged kalman filter, Test 7
lutant removal efficiency during adverse mixing conditions, Indoor Air 7 (1) (1997) (2) (1998) 217–282.
55–63. [46] L. Li, P. Revesz, Interpolation methods for spatio-temporal geographic data,
[26] M. Krajčík, A. Simone, B.W. Olesen, Air distribution and ventilation effectiveness in Comput. Environ. Urban Syst. 28 (3) (2004) 201–227.
an occupied room heated by warm air, Energy Build. 55 (2012) 94–101. [47] J. Luttinen, A. Ilin, Variational gaussian-process factor analysis for modeling spatio-
[27] S. Schiavon, F.S. Bauman, B. Tully, J. Rimmer, Chilled ceiling and displacement temporal data, Advances in Neural Information Processing Systems, 2009, pp.
ventilation system: laboratory study with high cooling load, Sci. Tech. Build 1177–1185.
Environ. 21 (7) (2015) 944–956. [48] D. Fox, W. Burgard, F. Dellaert, S. Thrun, Monte carlo localization: efficient position
[28] L. Zagreus, C. Huizenga, E. Arens, D. Lehrer, Listening to the occupants: a web- estimation for mobile robots, AAAI 2 (1999) 343–349.
based indoor environmental quality survey, Indoor Air 14 (s8) (2004) 65–74. [49] H. Liu, H. Darabi, P. Banerjee, J. Liu, Survey of wireless indoor positioning tech-
[29] E. Gossauer, A. Wagner, Post-occupancy evaluation and thermal comfort: state of niques and systems, IEEE Trans. Syst. Man, Cybern. Part C Appl. Rev. 37 (6) (2007)
the art and new approaches, Adv. Build. energy Res. 1 (1) (2007) 151–175. 1067–1080.
[30] A. Leaman, F. Stevenson, B. Bordass, Building evaluation: practice and principles, [50] J. Friedman, T. Hastie, R. Tibshirani, The Elements of Statistical Learning vol. 1,
Build. Res. Inf. 38 (5) (2010) 564–577. Springer series in statistics Springer, Berlin, 2001.
[31] A. Ghahramani, G. Castro, B. Becerik-Gerber, X. Yu, Infrared thermography of [51] ANSI/ASHRAE, ANSI/ASHRAE 62.1, Ventilation for Acceptable Indoor Air Quality,
human face for monitoring thermoregulation performance and estimating personal ASHRAE, Atlanta, GA, US, 2013.
thermal comfort, Build. Environ. 109 (2016) 1–11. [52] Indoor air technical committee, Part 8: Determination of Local Mean Ages of Air in
[32] C. Dai, H. Zhang, E. Arens, Z. Lian, Machine learning approaches to predict thermal Buildings for Characterizing Ventilation Conditions, Geneva: International
demands using skin temperatures: steady-state conditions, Build. Environ. 114 Organization for Standardization (ISO) 16000.
(2017) 1–10. [53] A. Gadgil, C. Lobscheid, M. Abadie, E. Finlayson, Indoor pollutant mixing time in an
[33] S. Lee, I. Bilionis, P. Karava, A. Tzempelikos, A bayesian approach for probabilistic isothermal closed room: an investigation using cfd, Atmos. Environ. 37 (39) (2003)
classification and inference of occupant thermal preferences in office buildings, 5577–5586.
Build. Environ. 118 (2017) 323–343. [54] L. Peeters, A. Novoselac, Impact of Human Activity on Unsteadiness of Airflow and
[34] C. Benton, F. Bauman, M. Fountain, A field measurement system for the study of Convective Heat Transfer in Indoor Environmental Studies vol. 1, (2011), pp.
thermal comfort, ASHRAE Trans. 96 (1990) 623–633. 495–500.
[35] C.-M. Chiang, P.-C. Chou, C.-M. Lai, Y.-Y. Li, A methodology to assess the indoor [55] S. Liu, A. Novoselac, Air diffusion performance index (adpi) of diffusers for heating
environment in care centers for senior citizens, Build. Environ. 36 (4) (2001) mode, Build. Environ. 87 (2015) 215–223.
561–568.

276