Visual Analytics for Building Management
Petr Glos and Lubomír Popelínský
Fig. 1. 3D model of the new Masaryk University Campus Bohunice
Abstract--We first shortly describe Masaryk University building management system and visual analytic tools that are used in
building management at Masaryk university. Then we introduce a novel method that combine visualization with automatic
classification for finding unexpected trends in temperature.
Index Terms-- Building management, visualization, data mining, classification, decision tree, outliers
1 INTRODUCTION
Masaryk University maintains a digital version of building passport
that currently consists of approximately 200 buildings and 17,000
rooms. For processing this digital data and its visualization, the ESRI
ArcGIS software has been successfully applied. However, for deep
understanding of running processes in a room, namely for anomaly
detection, we need an analytic tool that can predict (or at least detect)
such rare events. A rare event is a pattern that does not occur very
often but is important for precaution, e.g. a fire (or an increase of a
temperature), fast repeated switching on/off of a device, or a water
pipe disruption.
In [8] we presented an ongoing project on mining spatio-temporal
frequent patterns from building management data. We focused on
two tasks that are important for facility management - mining
frequent patterns [7] and mining rare events in spatio-temporal data
[3].
In this paper we focus on another part of the analytic tool: the
anomaly detection in temperature trends. We first describe shortly
the building management system and then we summarize
visualization tools that are currently used. In the second part of this
text we introduce an analytic method that serve for detection of
anomalies ­ unexpected behavior in temperature ­ in a room. Since
the number of temperature measurements varies (from less than 90 to
more than 80,000) depending on a way of measurement and the
actual temperature changes in the particular place, various
aggregated values are computed first. This method combines
automatic classification and outlier detection.
2 GEODATABASE OF BUILDINGS AND TECHNOLOGIES
Several distinct constructions (building primitives ­ engineering
constructions, holes in constructions and panes of wholes) were
defined for building and room representation in a geodatabase and
every building is made up of these objects. The building primitives
have polygon geometry representations with 2D coordinates and two
attributes of height (upper and bottom part) for 3D representation.
Each room, floor and building has unique identification called
position code. For example, BMA01 is the position code of Masaryk
university headquarters building, BMA01N02 is the position code of
the second floor of this building and BMA01N02040 is the position
code of one room on this floor (see Fig 2).
The resulting passport data is available to university employees,
students and even for public via the internet/intranet as well as it is
used by other university's information systems. Additionally, the
building passport is used to generate 2D maps and 3D models of the
buildings (see Fig 3).Petr Glos with Masaryk University, Institute of Computer Science, E-Mail:
glos@ics.muni.cz.
Luboš Popelínský, with Masaryk University, Faculty of Informatics, E-
Mail:popel@fi.muni.cz.
Fig. 2. Internet floor map with positions codes. High resolution image
can be found at http://gis.ics.muni.cz/bms/demo/SP_RMU.PNG.
Fig. 3. 3D model of a new campus building. High resolution image can
be found at http://gis.ics.muni.cz/bms/demo/SP_A7.PNG.
3 BUILDING MANAGEMENT SYSTEM
Masaryk University implements a Building Management System
(BMS) based on BACNet open protocol. BACnet is a data
communication protocol for building automation and control
network (http://www.bacnet.org). The BMS applications are used for
monitoring and controlling technologies of the new University
Campus buildings such as HVAC (heating, ventilation, air
conditioning and cooling), lighting, fire alarm system, intrusion
system, access system, CCTV, etc. The data from HVAC are very
interesting because room environment is very important for building
users and also operation costs depends on HVAC behaviour. BMS
provides means of storing a historical building operation data in the
relational database.
A typical record of the database contains an identification of the
measurement, time, and a value. The identification of measurement
consists of the position code (unique code of room, floor and
building) and type of value such as temperature or humidity. Tuple
[BMA01N02040, TK14, 4.9.2007 15:00:00, 20.511] represents one
measurement of room temperature (TK14) in room with position
code BMA01N02040. The room temperature was 20.511 °C on
September 4th, 2007 at 15:00.
In this way, we can obtain a spatiotemporal data of the room
environment and HVAC operation and we are able to use the data for
analyses and visualizations.
4 GIS VISUALIZATIONS
The objects of BMS database and the geodatabase of buildings are
linked through the position code and we can apply GIS tools for
analysing the BMS spatiotemporal data and visualization. We use
ESRI ArcGIS for creating 2D and 3D thematic maps to show either
the actual operation values or the aggregating values over a given
time interval. We utilize 2D and 3D temporal animations for
visualization of operation values for given time period. Especially
temporal animations are very interesting because they make it
possible to see the history of long-term values quickly and detect the
trends and anomalies of a building operation.
4.1 2D Thematic Maps
They are two typical scenarios for preparation of 2D thematic maps.
In the first scenario we create floor maps or maps of buildings for a
given value (e.g. temperature, humidity, differential pressure,
electricity consumption) in a given instant of time. For this purpose
we prepare data view of a required feature class (e.g. geometries for
a ground plan of rooms) and required values from the BMS database
(e.g. temperatures of rooms for given instant of time). The records
from these views are joined through the position code.
In the second scenario we create floor maps or maps of buildings for
aggregated values (e.g. minimum or maximum of temperatures,
electricity consumption, etc.) for a given time interval. For this
purpose we prepare a query that gathers the required data and then
we join the query results (e.g. electricity consumption of a given day
for all campus buildings) and the required feature class from the
geodatabase again using the position code.
After that, we prepare the desired symbology and legend for the
requested map and do the GIS analyses on the BMS data. The
resulting map can be exported in the desired data format (e.g. PDF,
JPG or PNG). The map (Fig. 4) represents temperatures measured by
ceiling sensors and sensors positioned on top of computer racks in a
computer room. Blue colour represents the lowest temperature and
red colour is used for the highest temperature.
Fig.4. Computer room temperatures. High res. image can be found at
http://gis.ics.muni.cz/bms/demo/VIZ_2D_UVT_sal_teploty.PNG.
4.2 3D Thematic Maps
Similarly to the 2D, there are also two typical scenarios for
preparation of 3D thematic maps. The difference is in the method of
visualization. We need to prepare 3D objects that represents the
buildings and rooms. For this purpose we can use the height
attributes (upper and bottom heights of the building primitives) for
the extrusion method of ArcScene software. The extrusion method
allows to create a 3D model of rooms and buildings from building
primitives.
Using the retrieved 3D representation we can create the required 3D
map for given values as in the case of 2D maps and we can also
export the result ­ there are the data formats available: VRML, PDF,
JPG or PNG). The map (Fig. 5) illustrates the values of selected
room temperatures in a given building. Dark blue colour is used for
the lowest temperature while light ping colour represents the highest
temperature.
Fig. 5. 3D model of a new campus building. High resolution image can
be found at http://gis.ics.muni.cz/bms/demo/VIZ_3D_A5_teploty.jpg.
4.3 2D Temporal Animations
Again, there are two typical scenarios. We can create animations of
2D maps using the BMS data for either values measured in a given
time frame (e.g. history of room temperatures of one floor of given
building) or aggregated values in a given time interval (e.g. history
of daily electricity consumption for all building of campus). For this
purposes we prepare a data view of the requested feature class and
the required BMS data for the given time interval (e.g. room
temperatures of required floor of building for one week or month)
and again join them using the position code.
We can use the same symbology as for the thematic maps and export
the resulting animation to a video format AVI. The animation
(http://gis.ics.muni.cz/bms/demo/fi_sal_video.avi - 28 MB)
represents a two-day history of temperatures in a computer room.
You can see that the temperatures are relatively steady, but the
temperatures in north-western racks were increased between 20:00
and 24:00 on August 23rd.
4.4 3D Temporal Animations
We use 3D representations of rooms and buildings and prepare
temporal animations with same scenarios as in the case of 2D
temporal animations. The 3D animation give us better imagination of
the chosen characteristics of building operation because we can see
and analyze all floors of the building together. It is also simple to
rotate and pan the model and switch the visibility of floors on or off
to get a complex view of the monitored situation.
The animation (http://gis.ics.muni.cz/bms/demo/UKB_spotreby.avi -
4MB) illustrates a three-month history of daily electricity
consumptions of selected buildings of the campus. The dark brown
colour represents the highest consumption and the light brown colour
is used for the lowest values of consumption. We can produce the
video from these animations in the AVI format too.While the
geographic visualizations are applicable to obtain overview
information about buildings, for deep analyses we use data mining
techniques.
5 MINING AND VISUALIZING EXCEPTIONS
5.1 Exceptions
For efficient building management it is very important to find such
parts of a building that strongly differs ­ e.g. in terms of
consumption of resources, like heating ­ from other parts that are
supposed to be similar. Such events are actually very rare. The main
idea is following. To obtain information about rare events that may
happen in building management we first have to define similarity on
parts of building (rooms). This similarity measure generates classes
of similar objects. Then, using learning techniques, we learn a
classifier that with high accuracy classifies all rooms to those classes.
We use the C4.5 algorithm (J48 implementation). Rooms that are not
classified correctly are candidate places where a rare event has
occurred. Those rooms are visualized together with an evidence for a
rare event ­ a logical formula (e.g. a branch of a decision tree) for
which that room is as an exception. It helps not only to find the place
but also to explain the reason for a possible anomaly.
5.2 Data and Data Aggregation
Attributes of the data about rooms extracted from the building
management system can be divided into four groups. The first group
concerns room identification like the room position code and room
description (text). The second group describes a position of a room
in a building ­ its orientation (east, west), the area, etc. The third
group provides information about the way how a temperature is
measured. Since the number of temperature measurements can vary
from tens to tens of thousands, the fourth group contains computed
attributes that are actually aggregates ­ a minimum and maximum
temperature, a difference between the max. and min. temperature
and also a number of temperature measurements.
5.3 Algorithm
Below is an algorithm for finding exceptions in building
management data. The algorithms are influenced by the following
parameters, MinAcc - minimum accuracy of the decision tree on the
learning data, MaxNoice - ratio of incorrectly classified examples =
<incorrectly classified>/<majority class in the leaf>
0. Define similarity classes
1. Learn classifiers
For each class,
find a model (decision tree) that reaches
accuracy MinAcc : 100%> accuracy >= MinAcc]
3. Find exceptions:
3a) find all branches in all trees such that
MaxNoice >= noice > 0
3b) sort them by a number of appearance (and
than by noice) in ascending order
3c) take first MaxExceptions examples and
display them together with the
corresponding branch and layers
5.4 User Settings
It is clear that the process of finding anomalies can be hardly
automatic. A user has to define at least a similarity or neighborhood
measure [1] that serves for dividing data into similarity classes. It is
necessary that the class is nominal. In the case that the similarity
measure is not discrete, the class is automatically discretized into 2
to 5 classes, either equidistant or with equal frequency of examples,
and the discretization that maximize the accuracy is chosen. A user
can also set the parameters mentioned above, especially a level for
outlier detection, which is highly data-dependent. Additionally,
setting MaxExceptions, the maximal number of exceptions to
display, prevents the algorithm from finding and displaying too
many candidates for outliers.The generation of classifiers is fully
automatic. The result is then displayed with all layers where an
outlier has appeared. Places that have been detected most frequently
are displayed first.
6 RESULTS
The algorithm has been tested on the data collected at Masaryk
University Campus Bohunice in 2009 for all 267 rooms. Information
about temperature has been first aggregated as described above. The
data contained the following attributes: RoomID - room
identifier
Corner - yes/no
Orientation - east/west/other
Floor position ­ inner/outer
Description - text
Floor - -1,1,2,3
Area ­ in square meters
Vzorkovani_typ - COV for sampling data by change
of value (temperature)
­ POL for sampling data by change
of time
Vzorkovani_hodnota ­ in °C for COV, in minutes for POL
NumTempMeas - number of temperature measurement
Temperature_min - minimal temperature
Temperature_max - maximal temperature
Difference_min_max ­ difference between min. and max.
temperature
We defined four similarity measures ­ floor, floor position
(inner/outer), number of temperature measurements for POL
sampling (discretized into 3 bins equidistant, NumTemp-POL),
number of temperature measurements for COV sampling (discretized
into 5 bins with equal frequency in a bin, NumTemp-COV). No limit
has been set to MaxExceptions.
Class MinAcc MaxNoice num. of outliers
floor 0.95 0.20 6
floor position 0.90 0.40 9
NumTemp-POL 0.75 0.20 5
NumTemp-COV 0.75 0.20 9
The most interesting patterns are below.
floor position
Difference_min_max <= 9.0082 AND
Difference_min_max > 7.0286 AND
not to west =>
outer floor (2 outliers)
Difference_min_max > 9.0082: =>
outer floor (7 outliers)
POL data, class: number of measurement
Difference_min_max <= 17.606 =>
num of measurements < 45681.333333 (3 outliers)
COV data, class: number of measurement
outer AND
area > 10.43 AND
Difference_min_max <= 9.1335
=> -1st floor (1 outlier)
outer AND
area > 10.43 AND
Difference_min_max > 9.1335
=> 3rd
floor (1 outlier)
floor = -1 AND not to east
=> num of measurements < 1614 (1 outlier)
7 CONCLUSION
We described visual analytics tools that are used in the building
management system of Masaryk University Brno. We introduced a
novel algorithm for finding anomalies in temperature measurements.
The detected anomaly may be caused by a defect of a measurement
tool or the network that transmits data. Actually, most of the
anomalies discovered in our experiments were of that kind.
Nevertheless, even in that case the information about anomaly is
useful. Moreover, the discovered rules have been evaluated as
reasonable by an expert.
ACKNOWLEDGMENTS
We thank Michal Batko for his collaboration. This work has been
partially supported by Faculty of Informatics, Masaryk University
and also by the Grant Agency of the Czech Republic under the
Grant No. MSM 0021622418 Dynamic Geovisualization in Crisis
Management and by Faculty of Informatics, Masaryk University
Brno.
REFERENCES
[1] Ester M. et al. Spatial Data Mining: A Database Approach. Advances in
spatial databases: 5th international symposium, SSD '97, Berlin, 1997.
[2] Hodge V., Austin J.: A survey of outlier detection methodologies.
Artificial Intelligence Review Volume 22, Number 2, October 2004,
Springer 2004
[3] Huang Y., Pei J., Xiong H. Mining C-Location Patterns with Rare
Events from Spatial Data Sets. Geoinformatika 10, 2006, pp. 239-260.
[4] Glos, P. Building and Technology Passport of Masaryk University. In
Proceedings of 2008 ESRI International User Conference. San Diego,
California : ESRI Press, 2008
[5] Glos, P. Using ArcGIS for Visualizing Historical Data from BMS. In
Proceedings of 2009 ESRI International User Conference. San Diego
California : ESRI Press, 2009
[6] John G.H.: Robust Decision Trees: Removing Outliers from Databases.
In Proceedings of KDD-95 Conference, AAAI 1995.
[7] Popelínský L., J. Blat'ák: Toward mining of spatiotemporal maximal
frequent patterns. In Proceedings of ECML/PKDD Workshop on Mining
Spatio-Temporal Data (MSTD), Porto 2005.
[8] Popelínský L., P. Glos: Facility management and mining spatiotemporal
data. Proceedings of Znalosti 2010 Czech-Slovak AI
conference, Jindřichův Hradec, 2010.
[9] Simoff S. J., Böhlen, M. H., Mazeika A. (eds.) (2008) Visual Data
Mining. LNCS 4404 Springer Verlag 2008.