Entry
Name: MDX-Attfield-MC2
VAST Challenge 2016
Mini-Challenge 2
Team Members:
Simon Attfield,
University of Middlesex, S.Attfield@mdx.ac.uk
PRIMARY
Peter Passmore,
University of Middlesex, P.Passmore@mdx.ac.uk
Kai Xu, University of
Middlesex, K.Xu@mdx.ac.uk
David Windridge, University
of Middlesex, D.Windridge@mdx.ac.uk
Adrian Wagstaff,
MASS Consultants Ltd, awagstaff@mass.co.uk
Daniel Hewitt, MASS
Consultants Ltd, dhewitt@mass.co.uk
Graham Phillips, MASS
Consultants Ltd, gphillips@mass.co.uk
Greg Dash, MASS Consultants
Ltd, gdash@mass.co.uk
Richard Chapman, MASS
Consultants Ltd, rchapman@mass.co.uk
Lee Mason, MASS Consultants
Ltd, lmason@mass.co.uk
Student Team: NO
Tools Used:
Excel
POLAR Kermode A Patterns of Life analysis tool created by Graham Phillips
for MASS Consultants Ltd and modified for the VAST2016 challenge
Excel
Classifier A tool created by Peter Passmore
for the VAST2014 challenge and modified to incorporate elements of VAST2016
Data Timelines A bespoke tool created by a team of engineers at MASS
Consultants for the VAST2016 challenge
Approximately how many hours were spent working on this submission in
total?
375
May we post your submission in the Visual Analytics Benchmark
Repository after VAST Challenge 2016 is complete? YES
Video
Questions
N.B. Herein zone identifiers of the
format zone x.y refer to zone y on floor x.
MC2.1 What
are the typical patterns visible in the prox card data? What does a
typical day look like for GAStech employees?
Limit your response to no
more than 6 images and 500 words.
Daily Routine of GasTech Employees
Figure 1 shows a visualisation from POLAR Kermode where
the proximity movements of the GasTech employees are
shown for the entire period. The map view on the left of the screen shows the
locations employees were observed with colours
corresponding to those in the timeline view on the right side. Looking at the
timeline, which shows employees as rows and time along the horizontal axis,
there is a typical weekly work schedule at the company, with very few employees
entering the building during the night or over weekends.
There seem to be four
main working patterns for GasTech employees. The
majority of employees (around 100) operate in a typical day shift pattern,
arriving at work some time between 6am and 7am and
leaving around 4pm. An evening shift of around 20 employees begins at 2pm and
ends at 10pm. A pair of watchmen (Varro Awelon and
Emile Arpa) guard the building between the hours of
11pm and 6am on weekdays, spending the majority of their time in the lobby.
Finally, three employees (Dylan Scozzese, Chi Staley
and Raye Paredes) work in the loading area and spend
little time outside of it.
Figure 1: POLAR Kermode visualisation of
a the working period
In Figure 2 the Classifier tool has been used to show the estimated location of all
employees for the first day in the dataset, organised
by department. In a similar fashion to the POLAR Kermode timeline, there is a
row of data for each prox card with time on the
horizontal axis but colours show the last observed location for a given card,
rather than instantaneous observations.
For those on the day
shift there is a period of mass migration throughout the building at around 8am
where most departments head into hallway prox zones
(these areas contain break rooms). The deli is the lunch location chosen by
most employees (shown in red in Figure 2). Exceptions include Executives and HR who do not move for lunch, the
Security staff (the bottom group) who leave the building and the loading bay
employees who remain in the loading bay. In fact, once the loading bay staff
(visible as 3 blue streaks around) enter the loading bay at around 8am, they
stay inside for the remainder of their working day (approximately 5 hours).
They do not appear to deliver received goods throughout the building and the
only record in the prox dataset of another employee
entering the loading bay is Lise Carrara on the final
day.
Figure 2: Classifier visualisation of the
locations of employees during the first day in the dataset
Employee Groups
Even within departments
there is some subgrouping in the data, Figure 3 shows one such example taken from the security department. Figure 4 presents the movements of this subgroup during one of the morning
migration events observed in Figure 2. Interestingly, while their movements to the hallway area are perfectly
synchronised, their return to their original
locations is fragmented.
Figure 3: POLAR Kermode visualisation of
an example group of employees
Figure 4: POLAR Kermode movement of the
example employee group during one morning migration event
MC2.2
Describe up to ten of the most interesting patterns you observe in the building
data. Describe what is notable about the pattern and explain what you can about
the significance of the pattern.
Limit your response to no
more than 10 images and 1000 words.
Daily HVAC Variations
Zones have similar
thermostat temperature profiles in different areas of the building and even
across floors. This suggests that the thermostat set point is controlled
automatically. There is regular variation in heating and cooling setpoint which is synchronised with the Night Manager
Control Status building sensor. Thermostat target temperatures are relaxed
(heating point lowered and cooling point raised) during weekends and at night
and the building temperature control systems are correspondingly less active.
In Figure 5 the cooling setpoint of
both zones are raised between the hours of 10pm and 6am and the HVAC mass flow
rates drop noticeably during the period.
Figure 5: Overview of daily HVAC variation
seen in the summary spreadsheet
Figure 6: Regular variation in thermostat
set point shown in the Data Timeline visualisation tool
Hourly HVAC Variations
In Figure 6 the VAV REHEAT Damper Position
and SUPPLY INLET Mass Flow Rate respond to demand but during the night the system
enters a low power mode with a spike in operation every hour.
Figure 7: Fine grained variation in a
number of HVAC elements shown in the Data Timelines tool
Anomalous Thermostat Behaviour
Details of the anomalies
are discussed during question 3 but their existence in different sensors (and
over different floors) suggests systematic problems with the HVAC system and
that the thermostat controls may be drawn from a number of decentralised
systems. Note that artefacts occurring during the 1st, 2nd,
7th, 8th and 9th days are seen inconsistently
across different thermostat setting channels.
Figure 8: Variation in anomalous thermostat events across
building zones suggesting complex coordinated control structure
Dry Bulb Temperature Variation
The dry bulb temperature varies
between the low 20s at night to the high 20s/30 deg. during the day. The transitions
are smooth so we assume that this is a measure of natural temperature variation
outside the building. This measure does not seem to be responsible for or
affected by HVAC temperature fluctuation anomalies discussed in question 3.
Figure 9: Dry bulb measurements of
external temperature shown using the overview spreadsheet
HVAC Sensor Clustering
Figure 9 shows the results of sensor clustering: there are
8 clusters (each has a different colour) and each dot is a sensor. H1-H4 are
the four Hazium sensor readings.
The clustering is
derived by applying t-distributed stochastic neighbourhood embedding [1] to the
time series data of each sensor and projecting into 2D. Cluster labelling is
derived from K-means clustering of Hazium values with
k=8. As a result, the x and y axis are not directly interpretable.
The regular
spacing among the dots is not an artefact of the clustering algorithm. This
indicates an interesting pattern of the sensor readings. Also, the boundary of
different clusters is very clear (i.e., no overlapping). This shows which sensors
behave similarly and their similarity to the Hazium
readings.
[1] van der Maaten, L.J.P.; Hinton, G.E.
(Nov 2008). "Visualizing High-Dimensional Data Using t-SNE" . Journal of Machine Learning Research 9: 25792605
Figure 10: HVAC sensor similarity based on
t-SNE clustering
MC2.3 Describe up to ten
notable anomalies or unusual events you see in the data. Describe when and
where the event or anomaly occurs and describe why it is notable. If you have
more than ten anomalies to report, prioritize those anomalies that are most
likely to represent a danger or serious issue for building operation.
Limit your response to no
more than 10 images and 1000 words.
Cards left overnight
There is a
synchronised leaving of cards at work overnight, as detailed in Table 1 and displayed in Figure 9. The main offender seems to be Clemencia
Whaley although a number of other evening shift workers seem to leave their
cards overnight at the same time as Clemencia. From a
security perspective the unattended cards represent a significant potential for
abuse and the fact that the timing of these events is shared by many employees
is also worrying.
|
Card Owner |
Zone |
Date |
|
Clemencia Whaley |
2.1 |
31st |
|
Dante
Cognian |
2.1 |
31st |
|
Effie
Davies |
2.1 |
31st |
|
Isak Baza |
2.1 |
31st |
|
Valeria
Morlun |
2.1 |
7th |
|
Clemencia Whaley |
2.4 |
7th |
|
Adan
Morlun |
2.4 |
7th |
|
Dante
Cognian |
2.1 |
7th |
|
Ermina Bello |
2.1 |
8th |
|
Clemencia Whaley |
2.1 |
8th |
Table 1: Prox cards left overnight
Figure 11: Classifier overview of prox card data for the whole measurement period showing
cards left overnight
Late Night Visits
On two nights a prox card registers in the ground floor lobby at exactly
midnight and is then not registered again for many hours. This registration
could indicate a malfunction in the prox card system.
|
06/06/2016 17:02 |
gflorez005 |
1 |
9 |
Administration |
|
07/06/2016 00:00 |
gflorez005 |
1 |
1 |
Administration |
|
07/06/2016 07:56 |
gflorez005 |
1 |
4 |
Administration |
Table 2: Anomalous prox
observation for Geneviere Florez
|
01/06/2016 12:03 |
jsanjorge001 |
1 |
9 |
Executive |
|
01/06/2016 14:00 |
jsanjorge001 |
1 |
1 |
Executive |
|
02/06/2016 00:00 |
jsanjorge001 |
1 |
1 |
Executive |
|
02/06/2016 14:02 |
jsanjorge001 |
1 |
4 |
Executive |
Table 3: Anomalous prox
observation for Sten Sanjorge
jr.
Wandering Facilities Card
Patrick Young (83
and 84 in Figure 11) loses his card after the second day and then
appears to use the second card the rest on the time. His first card appears to
move around spending most time in level 2.2 (magenta) and 2.7(dark green)
(though going on short trips elsewhere in between) before its final move on the
second friday.
Figure 12: Loss of prox
card by Patrick Young and its continuing movement throughout the building
Building Shortcuts
Figure 12 shows an estimated adjacency graph for all proximity zones in the GasTech building. Adjacency values were produced by counting
the total number transitions between every pair of nodes. In this graph the
line thickness is the fifth root of adjacency, heavily skewing thickness
towards lower values. For reference, the thickest line in the graph (prox zone 1.4 -> prox zone 2.4)
represents the movement of many thousands of people and the thinnest lines
represent around 5 observed transitions. The extreme variation in the highest
and lowest values might suggest sensor malfunctions, employees tailgating
through doors in the building or the existence of other methods of transit
through the building.
Figure 13: Estimated adjacency graph of
building prox zones.
Systematic HVAC System Malfunctions
On day 2 at around
1pm there is very high spike in air inlet temperature in zones 1.2, 2.6, 2.14,
3.5 and 3.6. The supply inlet temperature 1.2 shoots up from 12 degrees to 40
(2.6) in a 5 minute period then drops down again. The
temperature soon falls away but it coincides with the beginning of a three-hour
high in thermostat cooling setpoints in zone 1.2,
2.6, 2.14, 3.5 and 3.6 and an increase in temperature in those areas. This is
not a natural change. Possibly this is a sensor malfunction. This is followed
by a pattern of activity in Inlet Mass Flow Rate and Reheat Damper position.
Figure 14: Systematic HVAC anomaly seen
in zone overview information
On day 3 at around
1pm there is a very high spike in supply inlet temperature on floor 3 in zone
1. The supply inlet temperature shoots up from 10 degrees to 36 in a 5 minute period and remains at high levels until early the
following morning. This pattern repeats every day for the remainder of the
dataset.
At around 9am on
day 3 the thermostat cooling setpoint in a number of
zones increases along with an increase in temperature. The zones are 2.2, 2.4,
2.6, 2.8. These zones are located around the outside of the building (image
shows zones 2.2 to 2.6 on day 3).
Figure 15: Day 7 thermostat set point
anomalies
On day 8 every
HVAC zone in the building (except zone 3.1 which is already in an irregular
cycle) begins to experience temperature fluctuations similar to those seen on
day 7 (above). The cycle seen on day 7 repeats itself until midnight on day 9
for most zones, at which point the building returns to the normal operation
observed at the beginning of the dataset. Zone 3.1 is unaffected and zone 3.8
experiences a recurrence on day 9 and then returns to normal operation.
Hazium Buildup
At many times during the
recorded period Hazium concentration sensors measure
non-zero levels as shown in Figure 13. These events may represent benign sensor malfunctions or attempted
chemical attacks on the building. Figure 14, though, seems to show a relationship between Hazium
concentration readings and building airflow which might suggest a gas is
responsible for the increased readings. Further, if Hazium
is truly responsible for the elevated readings the existing systematic HVAC
system anomalies may artificially reduce building airflow in the future,
compounding the effect of a Hazium gas event.
Figure 16: Hazium
concentration readings through the building during the recorded period
Figure 17: Relation between Hazium sensor reading and HVAC airflow for HVAC zone 2.2
Weekend Power Usage
Around 6pm on day
11 (Friday) the thermostat set points for heating and cooling are both set to
24 degrees. This continues over days 12 (Saturday) and 13 (Sunday) for until
around 4.55am on day 14 (Monday). The effect of this is to create an overly
narrow hysteresis sending the system into oscillations as shown (in Figure 13) by the behaviour of the REHEAT COIL Power, SUPPLY
INLET Mass Flow Rate, REHEAT Damper Position. The overall effect is an increase
in overall building power usage.
Figure
18: HVAC system power usage and relevant factors
MC2.4 Describe up to five observed relationships between the proximity card
data and building data elements. If you find a causal relationship (for
example, a building event or condition leading to personnel behavior changes or
personnel activity leading to building operations changes), describe your discovered cause and
effect, the evidence you found to support it, and your level of confidence in
your assessment of the relationship.
Limit your response to no
more than 10 images and 1000 words.
Automated Building Lighting
In the GasTech building the lighting systems appear to be
automated and are likely motion activated. The tagging system of the Data
Timelines tool allows a user easy access to proximity card information which
may be related to HVAC building zones and vice versa. Figure 15 shows the relationship between the occupancy of areas of the building
and lighting power. For major hallways in the building (like HVAC zone 2.12A)
the lights are permanently switched on for the duration of the dataset while in
others (like zone 2.6) the lighting is switched on if and only if there is at
least one person present in the zone. In the example shown in Figure 15, a prox card is left in the zone overnight
yet the lights automatically turn off, implying that they are reacting to
physical presence and that the card has been abandoned.
Figure 19: Building lighting responding to the
presence of GasTech employees
Building Equipment
Zone equipment power (Figure 16) tells a slightly different story to the lighting power. In particular,
the equipment is not on by default when the zone is occupied implying that the
systems for powering zone equipment are less automated than those for lighting.
Secondly, we see that the evening shift of GasTech
employees seem to interact with the equipment differently, having a much lower utilisation time despite having similar total numbers of
employees present in the zone to some times during the day shift.
Figure 20: Interaction between building
equipment and GasTech employees
Peak Hazium Building Visit
On the evening of the
second Saturday in the dataset the Hazium sensors in
all parts of the building show the highest reported values of the whole dataset
(see Figure 17).
Figure 21: Recorded Hazium
concentrations thoughout the recording period
The Data Timelines tool
provided a platform to examine this event and we noticed that a pair of
employees enter the building at the same time immediately before the Hazium level begins to rise. These employees (Orhan Strum and Mat Bramer) visit
prox zones adjacent to the office of S. Sanjorge, the location of the floor 4 Hazium
sensor.
Figure 22: Access of the building by GasTech employees coincident with the largest Hazium recordings
Investigating the issue
further it was found that there is no equivalent visit to Hazium
sensor related zones on floor 2 during the same period. At this point it seems
unlikely that the suspect prox carded employees are
responsible for the Hazium outbreak.
Figure 23: Observed excessive Hazium concentration with lack of employee cause
Wandering Facilities Card
As noted in question 3
the prox card of Patrick Young moves through the
building on multiple occasions after Patrick himself is issued a replacement
card. In Figure 20 this prox card can be seen entering prox zone 2.2 (where it remains for some time) at a similar
time to maintenance work carried out to repair the VAV REHEAT Damper in the
corresponding HVAC zone 2.8. This might imply that the card is still being used
by members of the Facilities team or, more benignly, is at the bottom of a tool
bag. This inference, being based on a single datapoint,
is far from certain and it would be wise to investigate the movement of this
card at the time of other maintenance operations.
Figure 24: Patrick Young prox
card movement at time of building system repair