a smart digital platform for airport services improving
TRANSCRIPT
A Smart Digital Platform for Airport ServicesImproving Passenger Satisfaction
Sonke Knoch, Philipp StaudtGerman Research Center for Artificial Intelligence (DFKI)
Campus D3.2, 66123 Saarbrucken, Germany
Bruno Puzzolante, Andrea MaggiCefriel - Politecnico di Milano
Viale Sarca 226, 20126 Milano, Italy
Abstract—Airport management companies are facing increas-ing passenger numbers and the pressure to provide high-qualityground services to satisfy the passengers’ needs and expectations.A new kind of information system is required, monitoringpassenger figures in real time to estimate the time and locationwhen and where a service must be provided and responsiblestaff scheduled. We suggest an event processing platform whichis built upon an industry-proven technology and aggregatesstreams of information about passenger frequency occurring inheterogeneous formats and with different frequencies to a keyperformance indicator reflecting the current state of selectedareas at the airport. Following the design research principle, weconducted a requirements analysis and developed a prototypicalsystem which efficiently fuses the data streams and generatesalerts to notify staff if passenger satisfaction threatens to drop.The system was evaluated with historical data from one of the20 largest airports in Europe. The evaluation is based on asimulation and provides evidence that the system has the potentialto improve customer satisfaction. In addition, due to the event-based architecture, a generic API allows a smooth integrationof new data sources, the parametrization of the event rulesaggregating and processing all data, and thus, an intuitive usageby the airport operator. The final tool consists of the eventprocessing module and a dashboard supporting airport managersin keeping track of the important quality attribute passengersatisfaction and scheduling staff in the right spot at the righttime.
I. INTRODUCTION
Today, the airport is a complex, multi-functional travel cen-
tre offering a wide range of services. It is a broad ecosystem
of different business units: commercial aviation, non-aviation
and operations. Each of them is aiming to reach its goals
of increasing revenues, optimizing resources, reducing costs,
and increasing the passenger satisfaction. Across the globe,
passengers are demanding higher levels of service, and the
competition among airports aims to increase the passenger
satisfaction in order to be able to attract more and more airline
companies; see for instance [1].
Airports themselves usually compare through global bench-
marking approaches, like the industry group airport councils
international’s (ACI) airport service quality (ASQ) program
which measures passengers’ satisfaction while they are travel-
ling through an airport. This type of benchmark gives each
airport a ranking with respect to competitors, highlighting
which areas are to be improved. According to a recent report
of the ACI on Airport Service Quality, airport cleanliness
(especially in restrooms) recently had the greatest effect on
travelers’ airport ratings [2], [3].
In this context, Smart DPA addresses the restroom use case
by providing an event-driven architecture integrating available
systems already available at the airport, including solutions for
queue monitoring, boarding pass scans and air quality sensors.
It is shown that a dynamic scheduling of cleaning operations
based on a KPI aggregating streaming data increases the
customers’ satisfaction compared to static schedules.
We conducted a requirements analysis at one of Europe’s
20 largest airports and developed a rule-based event process-
ing system. The system fuses heterogeneous data streams
efficiently since it is based on the industry-proven event
processing engine Apache Flink [4]. Events are mapped to
a KPI reflecting the passenger satisfaction. Based on this KPI,
alerts and forecasts are generated when restroom cleaning
becomes or probably will become necessary. A Web-based
dashboard visualizes all data and allows fast reactions to
unforeseen events, such as a sudden drop in satisfaction. The
system was evaluated based on historical data from the airport.
Due to organizational changes, a deployment at the airport was
not possible before the time of paper submission.
The paper starts with a discussion of related work including
scientific papers and commercial solutions in Section II. In
Section III, the requirements elicitation and the use case are
described. The concept for Smart DPA is presented in Section
IV and followed by the implementation in Section V. In
Section VI, the analysis of historical data and the evaluation of
the proposed system based on simulated passenger feedback
illustrates a potential improvement of passenger satisfaction.
The paper concludes in Section VII.
II. RELATED WORK
Related work is divided into scientific papers and commer-
cial solutions. Papers include similar approaches addressing
complex event processing and decision support, or analyze the
need for performance measurement and service improvement
at the airport. Commercial vendors provide solutions for
specific problems such as queue management, passenger flow
and restroom monitoring.
A. Scientific Work
Bezerra and Gomes [5] provide an overview about literature
in the field of performance measurement at the airport. 380
250
2020 IEEE 22nd Conference on Business Informatics (CBI)
2378-1971/20/$31.00 ©2020 IEEEDOI 10.1109/CBI49978.2020.00034
documents were analyzed, dating between 1970 and 2015.
Three stages were identified illustrating the evolution of the
airport business from predominantly government ownership, to
privatization and deregulation, to a complex business model
with diverse services for customers and stakeholders. The
authors conclude that benchmarking of performance improve-
ment at the airport should address the identification of or-
ganizational practices which lead to high performance and
service quality, such as the antecedents and consequences of
passengers’ satisfaction. Smart DPA was conceptualized to
satisfy exactly this requirement.
For the improvement of operational performance in an
organization, a helpful tool to handle big data and derive
conclusions in real time provides complex event processing
(CEP). One explanation is provided by Lundberg [6], who
argues that business operations typically consist of micro-
events which may be produced with high frequency from a
heterogeneous IT environment. Such events can be handled
efficiently by an event-driven architecture where a CEP en-
gine aggregates and combines such events and allows for a
mapping to key performance indicators (KPIs) reflecting the
effects on revenue, cost, and risk. The Smart DPA Restroom
Passenger Satisfaction Index (PSI) is such a KPI which is
related to the cleaning services at the airport and influenced by
several micro-events, e.g. from security control and passenger
feedback buttons.
Already in 2002, Boyer et al. [7] published a patent targeting
the event driven airport. They describe a system which notifies
entities of events and includes a source system tracking
changes in travel schedules, a publish-subscribe communi-
cation, and a receiving system notifying other applications.
Stanford’s David Luckham, known as the originator of CEP,
states in his book that the airport industry and transportation
sector was an early adopter of the event processing technology
and describes a baggage system checking transport times,
passenger connection times and flight gate assignments [8,
p. 66–67]. Both the patent and the example from Luckham’s
book also aim for service improvements and provide solutions
for specific problems such as changes in the travel schedule
and in-time baggage delivery. Similarly, Smart DPA addresses
the specific problem of scheduling cleaning activities and in
parallel aims to provide a more general and extensible solution.
In [9] Pestana et al. propose a decision support system
(DSS) with an event-driven architecture integrating multiple
sensor technologies to provide a comprehensive picture of
ground operations. The architecture is organized in different
layers and includes alerts, business rules, and map and KPI
views. The work is use case-driven and lacks an evaluation,
but a field trial was planned at the Faro airport. Smart
DPA provides a similar concept, but recent developments in
event processing, data handling and Web-based user interfaces
allowed us to provide a leaner, more flexible and scalable
architecture.
Zografos et al. [10] describe a decision-oriented modeling
framework of a DSS for the management and planning of air-
port operations. A classification into airside and landside types
of analysis and a procedure to develop a DSS are presented.
Finally, a use case focusing on capacity management and the
configurable presentation of flight data (arrivals, departures,
delays, traffic, queue length etc.) is illustrated and was built on
industry tools to foster acceptance. The approach was tested at
four European airports. The work provides interesting concepts
and focusses more on DSS. The Smart DPA system could be
combined with the suggested control mechanisms and improve
decisions at run-time.
The study by Hong et al. [1] investigates appropriate at-
tributes to measure user satisfaction at the airport. The study
was conducted using a questionnaire and showed that inter-
actional (self-service, interpersonal services, remote), physical
environment (e.g. seating arrangements, size, cleanliness), and
outcome (e.g. time delays and crowding) service quality are
supported for the passenger terminal, exemplified at the In-
cheon International Airport. The result supports the conclusion
that high service quality is a key factor for service providers to
achieve an outstanding position in the marketplace, but at the
same time is difficult to measure. Smart DPA aims to provide
a solution for this challenge.
B. Commercial Solutions
An increasing number of airports are looking for an easy so-
lution to improve passenger satisfaction. Most of the solutions
on the market are essentially vertical systems addressing the
core-business activities of the airport, such as the monitoring
of queues at security and passport control, or the estimation of
passenger flows. In this category there are solutions working
with proprietary sensors (e.g. Xovis [11], I-Sense [12]), solu-
tions able to integrate multiple sources like Wi-Fi and boarding
pass scans (e.g. SITA Queue management [13], Smart Flows
[14]) and general-purpose solutions applied to the airport and
with a specific application for passenger flow monitoring (e.g.
TIBCO [15]) leveraging the integration of multiple sources
such as ad-hoc cameras and Wi-Fi.
Recently, new solutions focusing on the monitoring of
restroom usage have also been launched on the market. These
solutions use a set of deployed ad-hoc technologies such
as counting passengers entering and exiting the restrooms
(e.g. Infax SmartRestroom [16]), smart latches and lights
(e.g. Tooshlights [17]), and a digital touchscreen monitor
(e.g. Restroomapp [18], Saniori [19]) for staff checklists and
passengers’ feedback.
All these solutions monitor the restrooms’ usage using
ad-hoc deployed devices, without providing any additional
insight. On the other hand, the Smart DPA solution integrates
with systems already deployed in the airport (through APIs)
and provides insights for the optimization of staff coordination
and for a better passenger satisfaction.
III. REQUIREMENTS ENGINEERING
The airport departments of IT, Customer Care and Opera-
tions, at one of the 20 largest airports in Europe, were con-
sulted to collect information on the passengers’ satisfaction,
inputs for the definition of the use case, and details about
251
Fig. 1: Landside check-in and airside departure areas.
solutions deployed or to be deployed in the near future that
would deliver relevant data. For this, interviews with the air-
port’s stakeholders, passenger surveys, and an analysis of the
data collected from existing airport systems were conducted. In
addition, a one-day workshop on the passenger experience was
held with more than twenty people from different departments
of the airport. The aim was to identify passengers’ “pains
and gains” and to work on the ideation of new services for
improving the passenger experience.
As a result, it was confirmed that a passenger’s time in
queue and the level of cleanliness of the restrooms are the
two main indicators that are affecting passengers’ satisfaction.
From an operational point of view, it was determined that
the management of staff in accordance with the status of the
services, i.e. the restroom cleanliness and queue at security
control, is one of the main challenges for airport managers.
As confirmed by the competition analysis in related work, the
queue management at security and passport control is widely
covered by market solutions, which lead the focus on the staff
management.
During the discussions the idea was made concrete to design
a solution able to (1) improve restroom quality service for
the passengers, (2) predict the status of the restrooms, and
(3) proactively provide information to the cleaning staff to
maintain a high level of cleanliness in all the restrooms. This
led to the following use case which forms the basis for the
evaluation. The data sources Smart DPA relies on, such as
the functional and architectural requirements, are described in
Sections III-B, III-C, and III-D.
A. Use Case
Different restroom types have been identified depending on
where the restroom is located with respect to the surrounding
area. Public airports are divided into landside and airside areas.
The landside area is open to everyone, while access to the
airside area is tightly controlled. The airside area includes
all parts of the airport around the aircraft, and the parts of
the buildings that are only accessible to passengers and staff.
Passengers and staff must be checked by security before being
permitted to enter the airside area. Conversely, passengers
arriving from an international flight (Extra-Schengen in the
case of EU airports) must pass through border control and
customs to access the landside area, where they can exit the
airport.
For the following conception, we will focus on the departure
area of an airport with segregated departure and arrival areas.
Figure 1 provides an overview of this area and divides the
restrooms into four possible types for departures, namely
landside, airside, Schengen flights and Extra-Schengen flights.
B. Data Sources
The available data sources at the airport were checked and
analyzed, which led to the following types of solutions being
identified to support the staff in the management of cleaning
services:
1) Boarding pass scanning: real-time but also statistical data
on the passengers entering the security control area.
2) Queue management: solution to estimate waiting time and
count the number of passengers passing through the metal
detectors at security control.
3) Passenger forecasting: numbers of passengers who are
departing and landing in coming hours and are to be
expected at the security gates.
4) Passenger feedback button: solution to monitor passenger
satisfaction at every touch point using a “smiles terminal”
ranging from sad to happy and providing the feedback in
the form of good, average, or bad ratings.
5) Restroom air quality monitoring: measures the air quality
in the restroom using odor sensors.
6) Restroom cleaning service monitoring: solution used to
certify the start and end of a restroom cleaning task
performed by the staff. It is usually done through a RFID
tag located in the restroom that cleaning staff scans when
they start and end a cleaning task.
C. Functional Requirements
During requirements elicitation it had become clear that
a KPI is needed to monitor service performance and to
indicate required actions if the performance has decreased or
will decrease: the location-based passenger satisfaction index(PSI). The following capabilities were identified:
• Provide an overview of the overall cleaning status of the
restrooms monitored
• Provide clear indications of the restroom blocks with the
lowest PSI
252
Fig. 2: Plot and trend of the passenger satisfaction index (PSI).
• Visualize on a map the overall cleaning status of the
restrooms monitored
This means for each location monitored:
• Visualize the PSI KPI (updated at least every 15 minutes)
• Visualize information about the last time a cleaning
activity has been performed
• Send automatic alerts to recommend a cleaning activity
• Provide an estimation of the next cleaning time
• Provide access through APIs to KPI and alerts
• Visualize the trend of PSI KPI (15-minute granularity)
• Visualize the historical data of recorded events
D. Architectural Requirements
To deploy the solution, additional static data is required,
such as a list of all restrooms at the airport with information
about location, type, and size. In addition, real-time events
about the restroom cleaning activities are necessary to reset
the PSI to 100% per restroom; see Section IV-A.
In a similar way, the system needs up-to-date information
about the passenger feedback rating per restroom, and the
number of passengers passing through metal detectors and
scanning their boarding pass at the security check as described
in Section III-B. The passenger satisfaction thereby will be
monitored, and rough information about the number of people
at the security check is known. In addition, a forecast about the
expected number of passengers departing in the coming hours
can be used. Finally, the air quality sensor delivers detailed
data with high frequency. To integrate all these various data
sources, the system needs to provide an API that allows the
ingestion of that kind of data.
IV. CONCEPT
Based on the requirements analysis which led to the func-
tional, architectural, and API requirements presented in the
previous section, a KPI was defined and is described in Section
IV-A. The required architecture is described in Section IV-B
enabling the event processing engine handling real-time and
historical events to automatically find patterns in the event
streams, as described in Section IV-C.
A. Key Performance Indicator
It is the goal of the Smart DPA platform to generate an easy-
to-understand indicator which represents the current customer
satisfaction at a certain location. Therefore, we introduce the
passenger satisfaction index (PSI), a value between 0 and
100%. It can be plotted as a time series and allows the
recommendation of actions in the future following a trend
estimation. Figure 2 shows a graph plotting this index. At
initialization, the plot always starts with 100% per restroom.
Two factors were identified during requirements engineering
as decreasing factors for this index: (1) information from data
sources, such as air quality monitoring and passenger feedback
buttons, will be aggregated to a weighted score decreasing the
current index; (2) a standard decreasing slope is introduced
reflecting the assumption that satisfaction decreases over time
if no actions are performed. This decreasing factor can be
configured and set differently for day and night. If an action
(i.e. a cleaning task) occurs, the index is reset to 100%.
The importance of predicting the next process event for
run-time management in order to identify processes at risk
has been discussed by [20], which introduced the use of
deep learning for predicting process behavior. In our case,
the recommendation of actions in the future assumes that the
threshold when the PSI reaches a critical point is known. In
Figure 2 the intersection between the index plot and a defined
cleaning threshold depicts this point. Computing the trend of
the index at an index position above the threshold allows the
forecast of the next action estimating a future intersection.
Those forecasts for different locations can be used to optimize
the action scheduling, i.e. cleaning tasks. Nevertheless, if the
index crosses the threshold line—the forecast was ignored or
false—an alert is generated warning the operator that a critical
point was reached. A second, lower cleaning threshold allows
a separation between the threshold used for trend estimation
and the threshold used for alerts.
B. Architecture
Figure 3 shows the high-level architecture for Smart DPA.
The API ingestion provides the interface to the airport systems
and the entry point to the event processing module used by
external systems. Event data ingested via the API is pre-
processed and forwarded to the event processing module. In
addition, selected data sets are written to the database for later
visualization in the web application. Two different usages of
the API are relevant in the Smart DPA use case: (1) real-time
data ingestion of live data exposed by the Airport Systems to
receive and process real-time events; and (2) historical data
from static files, such as CSV and XLSX, recorded by the
airport systems and streamed to simulate real-time events.
The latter data ingestion was used to evaluate the system as
described in Section VI.
The core of the solution is the event processing module that
contains the following three components: the event processingengine, event sources and sinks. The event processing engineis responsible for the processing of real-time and historical
events. It is used to efficiently aggregate data and to automati-
cally find patterns in the event streams. If necessary, streaming
data is enhanced with additional information, e.g. information
about the passenger’s destination. The results of the engine are
the PSI KPI, complex cleaning alerts and a forecast about the
next cleaning. Details about the event rules and algorithms are
provided in Section V-A. Event sources are used to ingest data
253
real-time stream-data ingestion via API
historical data ingestion via API (CSV
and XLSX)
airport systems Smart DPA platform
pub/subsink
DBsink
eventbroker
DB
publishsubscribe
event processing module
web application
back-end front-end
DB
API
inge
stio
n
…eventsource
data enhancement
writeread
event processing engine
airport management
Fig. 3: Design of the Smart DPA event-based architecture.
into the event processing engine. Sinks are used to stream or
persist processed data and are subdivided into pub/sub sinkswhich publish processed data to the event broker to provide it
to appropriate subscribers (e.g. the live monitoring dashboard),
and DB sinks which store data to a database for monitoring
and analytics applications.
Regarding the data storage and publishing to the web
application, the following components are used: The databaseon the right of Figure 3 stores all the data collected and
generated by the event processing engine. The database on
the left (inside the event process module) provides additional
information that is correlated to the data from sources during
event processing. The event broker is responsible for distribut-
ing messages to its clients. It is the handler of communication
between publishers and subscribers. The web application is
the end-user application showing the Smart DPA KPIs. The
web application consists of a back- and front-end. The back-
end acts as middleware between the front-end on one side
and the database and the event broker on the other. By
exposing a series of APIs, it deals with login authentication
and provides the list of restrooms with the KPIs’ actual values,
KPIs’ historical data and the history of restroom passenger
satisfaction.
C. Event Processing
The event processing engine must handle heterogeneous
data streams which occur with different frequencies. The
following will describe how these streams are fused and
aggregated to one single index, the PSI KPI. For example,
passenger feedback occurs at creation time, while the air
quality sensor delivers a measurement every ten seconds. The
content of the two messages is totally different. Figure 4 shows
how data streams from different sources (1 to N), with different
content (ty1 to tyN ), and in a different frequency (indicated by
the dotted line of the arrows) are aggregated to an individual
score, respectively. Such a score computation is common
when ranking an item, as for example suggested by [21] for
a running route recommendation system in the health and
fitness domain balancing between different route properties
such as length, slope, and location which are available offline.
Transferred to the event processing domain, where streams of
data need to be handled online, such a score is restricted to a
specific time window as follows.
A score is a numerical value between zero and one. All
scores are computed with the same frequency. The frequency
is defined by the user and defines the time in which events
from the respective source are aggregated. Each score has an
individual formula which defines the score computation. All
scores are aggregated by calculating the weighted mean. A
default decreasing slope, set by the user, specifies the decrease
per hour, since it is assumed that the cleanliness decreases over
time. Thus, the index always decreases with one exception,
which occurs when the restroom was cleaned, as indicated by
an event from data source 6 in Section III-B. In this case, the
index is immediately reset to 100%. The resulting cleanliness
index is used internally and further processed to generate alert
streams, predictions, and the final PSI output stream.
In total, four score streams are generated. The PSI streamrepresents the final PSI and is continuously streamed to the
event broker and persisted to the database. The cleaning alertstream delivers data if and as long as the cleanliness index
lies under a user defined threshold. In the same way, the lowcleaning alert stream delivers data if the cleanliness index lies
under a second user defined threshold. Finally, the predictionstream continuously streams the predicted time when the
index will probably drop under the threshold indicating the
next cleaning time based on the index trend. All streams are
repeated in a frequency defined by the user.
The internal cleanliness index used in these output streams
is based on the following scores related to the data sources
listed in Section III-B:
Boarding pass scans score (data source 1): gathers the
number of boarding passes that were scanned in front of the
security checkpoint, which affects the restrooms in front of
the security checkpoint. To estimate a score, the maximum
passenger capacity per hour must be defined by the user for
the respective area of interest.
Metal detectors score (data source 2): metal detectors de-
liver the number of people who were scanned. The number is
used to estimate the usage of restrooms behind the security
checkpoint using the maximum passenger capacity per hour,
254
Source 1 Aggregate 1 Score 1
raw streams individual scorestreams
Source 2 Aggregate 2 Score 2
Source N Aggregate N Score N
… … …
internal passenger satisfaction index stream
fixed window of size *
keye
d pe
r loc
atio
n
cleaning
fixed window of size *
default decreasing slope
reset index = weighted mean, * = time interval in which satisfaction percentage index is aggregated
Fig. 4: Resampling and aggregation of heterogeneous event streams to the KPI in the event processing module.
which is defined by the user for the respective area of interest.
Passengers planned score (data source 3): the estimated
number of passengers departing in the following hours is
used to estimate the number of passengers arriving at the
security checkpoint. The source is used to calculate a score
for restrooms before and after the security checkpoint using
the maximum passenger capacity per hour, which is defined
by the user for the respective area of interest.
Passenger feedback score (data source 4): uses the feedback
of users who can rate a restroom in the categories good,
average and bad. The score normalizes these ratings between
1 (100% good ratings) and 0 (100% bad ratings). A minimum
number of ratings is necessary within a time window to avoid
a strong impact of, for example, one bad rating within four
hours.
Air quality score (data source 5): delivers a value reflecting
the quality of air of the restroom in which the sensor is
installed. A formula maps the proprietary measurement to the
score.
All sources are combined using weights defining the impact
of the respective source and computing the weighted mean.
User-defined properties can be configured easily in a separate
file.
V. IMPLEMENTATION
The following describes how the event processing (V-A)
and the web application (V-B) were implemented. The actual
deployment of the ecosystem is explained in Section V-C.
A. Event Processing
The event processing engine, which was developed using
the open source stream processing framework Apache Flink
[4], is the core part of the event processing module (see Figure
3). It allows stateful computations over multiple data streams
in parallel and can evaluate complex event patterns on these
streams. Events are ingested by so-called event sources. These
sources can handle arbitrary kinds of data and serve as an
adapter interface for data ingestion into the engine. Currently,
the engine supports sources for REST (representational state
transfer) and MQTT (message queuing telemetry transport)
sources which cover state-of-the-art data transfer protocols in
the Web and in the IoT (Internet of Things) area. During
ingestion of events, which can happen in parallel and, due to
multiple sources of event emitters, even non-deterministically,
events are aggregated in equidistant time frames.The PSI is computed using multiple event sources and thus
multiple event scores per time frame of evaluation. For each
event source and time frame, a score is calculated expressing
the positiveness of the event stream according to location-
based passenger satisfaction. The following describes the score
computation of the restroom passenger feedback:For each restroom i, take all passenger feedback xi1 to xin
for each equidistant time interval tk to tk+1 and aggregate
them according to f(i, xi1, .., xin), for f(I, xi1, .., xin) =((good(xi1, .., xin) ∗ 100.0 + avg(xi1, .., xin) ∗ 25.0 +bad(xi1, .., xin) ∗ 0.0)/n)/100.0.
These scores are summed together, calculating the weighted
mean using a pre-defined weighting per score. The post-
processing function reflecting the standard decay over time
is applied afterwards. The result is the PSI KPI stream that
forms the basis for alerts. Alerts and the KPI are uniformly
emitted through a set of pre-defined event sinks. Currently
available sinks include Elasticsearch, MQTT and RabbitMQ,
but the engine is not limited to these.The database where the processed data is stored is Elas-
ticsearch, which is a NoSQL document-based database. This
database is hosted in Amazon Web Services (AWS) as a
service along with a Kibana dashboard that enables one to
browse the data in the indices where it is stored. The data
is organized in different indices, e.g., psi-index, psi-threshold,
and psi-time-to-clean-prediction.Processed events are also sent in realtime to a RabbitMQ
message broker on specific channels. The data itself is format-
ted according to the JSON format.
B. Web ApplicationThe restroom cleaning dashboard (Figures 5 and 6) has been
designed to show the restrooms’ status at a glance allowing the
255
(a) Home page.
(b) Restroom PSI.
Fig. 5: Home page and PSI visualization.
dashboard end-user (e.g. the airport manager) to organize the
cleaning teams accordingly. The home page (Figure 5a) shows,
in the top section, a summary of the average PSI KPI for all
the restrooms as well as the number of restrooms with a high,
medium and low index (respectively, “high” corresponds to a
PSI above the cleaning threshold; “low” corresponds to a PSI
below the low cleaning threshold; “medium” corresponds to
a cleanliness index between the two previous thresholds). The
remaining part of the home page provides a table summarizing,
for each restroom, the PSI KPI, the forecast for the next
cleaning time, and possible alarm raised. A filter panel on the
left allows the restrooms to be filtered according to terminal,
floor, area, restroom, gender and team. Alarms raised for a low
PSI violating the thresholds are displayed in the restrooms
table for each restroom as well as in a separate notification
area accessed by clicking on the relevant toolbar icon in the
header. By clicking on the name of the restroom in the table,
it is possible to access a pop-up page (Figure 5b) that shows
the trend of the PSI.
The map view page (Figure 6a) connects with the map
system of the customer to display, on the airport map, the
restrooms with their status (high, medium and low PSI) as
well as the other details in the pop-up window. The restroom
passenger satisfaction page (Figure 6b) shows a dashboard
with the last three weeks of restroom passenger satisfaction
indices related to the restrooms and estimated according to
the passenger feedback system (updated to the previous day).
For each week, in addition to a daily chart, the worst and
best restrooms are shown in relation to the satisfaction index.
(a) Airport map.
(b) Passenger feedback.
Fig. 6: Map overview and historical data about passenger
feedback.
From this page, the airport manager can see at a glance the
ongoing trend of the index and the restrooms performing best
and worst.
C. Deployment
The various software components are integrated with each
other in a Kubernetes cluster. The cluster is hosted in a
virtual machine on AWS and comprises a master node that
also enables the execution of docker containers. There is
a deployed pod for every software component, namely: air
quality API data extraction, API module for data ingestion,
event processing engine, RabbitMQ and Smart DPA web
application.
Elasticsearch and Kibana are not hosted inside the Kuber-
netes cluster, but rather on AWS as a managed service. The
pods can interact with each other in a private network inside
the cluster. Each pod exposes an endpoint for communication
with the other pods. The web application pod is also exposed
publicly because it needs to be accessed from outside the
cluster in order to display the Smart DPA dashboard to the
end user.
VI. EVALUATION
In the following, Section VI-A describes the evaluation
approach. Section VI-B provides some reflections on the data
which was identified as crucial for the implementation and
evaluation of the system. The underlying models for the
simulation are developed in Section VI-C, followed by the
results in Section VI-D.
256
A. Approach
To test the implemented event processing module, the
logged data that was received from the airport was replayed.
To do this, the static files were read and streamed to the
Smart DPA solution through the API ingestion. The stream-
ing process was accelerated, but the relative time distance
between the events was kept. In the evaluation, we target a
comparison between the PSI based on recorded data (cleaning
operations as they occurred) and the PSI that is achieved
using the Smart DPA solution (simulated cleaning operations).
Simulated cleaning activities are scheduled when alerts have
been triggered based on the PSI. Artificial passenger feedback
is generated based on trends discovered in historical feedback
logs. We thereby focus on the impact of cleaning activities
on PSI represented by passenger feedback (good, average and
bad), computed as described in Section V-A. The following
approach for the validation has been followed:
1) Analysis of logged restroom passenger feedback data
before and after cleaning, aiming to identify a function
able to estimate the trend of restroom passenger feedback.
2) Simulation of new cleaning activities with the Smart DPA
solution and estimation of the new Restroom Passenger
Feedback using the trend function defined in 1.
3) Comparison of the new restroom passenger satisfaction
against the logged restroom passenger satisfaction.
B. Data Assessment and Preparation
Before simulation, the data was checked for correlations
which potentially affect the event processing. The following
possible relations were considered and investigated in the
available data sets as potential candidates for the evaluation:
1) Cleaning duration and restroom size: For the analysis
of a correlation between the duration of a cleaning activity
and the size of the restroom, we had a data set ranging
from January 2018 to February 2019 from 147 restrooms. In
this data, the correlation between rising cleaning duration and
restroom size was analyzed to improve the alert timing taking
into account the probable duration. Only a weak positive
correlation (corr=0.25) between these data points was found.
Restroom size is one factor which affects cleaning time. It
would be interesting to include information about the physical
room layout, and the number of people moving in and out, in
this calculation.
2) Air quality comfort and effective passengers: The air
quality sensor delivers a comfort value according to the
current air conditions every ten seconds. A high comfort value
means good air quality. This comfort value was compared to
the effective number of passengers who passed through the
security gates. The data set ranges from January to August
2019. The results can be seen in Figures 7a and 7b. The scatter
chart shows high variance. In the heat map, a correlation
between the two properties is shown. A high comfort means
few passengers, and vice versa, supporting the assumption that
air quality decreases when the number of people increases.
(a) Scatter. (b) Heat map.
(c) Scatter. (d) Heat map.
Fig. 7: a and b plot normalized air quality (x-axis) against
the number of passengers (y-axis) at the security checkpoint
close to the measured restroom; c and d plot the normalized
time after cleaning (x-axis) against the accumulated passenger
feedback (y-axis).
3) Time after cleaning and accumulated passenger feed-back: The development of the passenger feedback over time
after a cleaning activity is of interest since it is closely related
to whether passengers are satisfied with the cleaning service
at the airport. The assumption was that there is a negative
correlation between a rising time after one cleaning with no
cleaning in between and a declining accumulated passenger
feedback.
A data set covering the time frame from the beginning of
January 2019 to the end of February 2019 containing data
about 15 restrooms of the selected airport area was available.
Complete data showing a good coverage of the selected data
properties was only available for this subset. The correlation
between time after cleaning, i.e. the accumulated seconds after
a recent cleaning without additional cleaning, was compared
to the accumulated passenger feedback in that time period.
Positive feedback was counted as +1, negative feedback -1
and neutral feedback as 0. The result is shown in Figures
7c and 7d. In the figures it can be seen that feedback is
tendentially negative (around a normalized 0.4, where 0.5
indicates neutral feedback) and occurs with higher frequency
briefly after cleaning due to the fact that cleanings were in
most cases separated by a normalized time smaller than 0.2.
In the charts it can be seen that a negative trend was not found
in the whole data set. Only selected restrooms showed such a
trend. The reason lies in missing data points, because many
cleaning events were not properly recorded, or were recorded
with wrong data, e.g. a cleaning duration of zero seconds.
257
Since the trend after cleaning was not approved, we ex-
amined the change of feedback, having a look at the trend
before and after cleaning. Figure 8 shows passenger feedback
aggregated from the same 15 restrooms in 10-minute buckets
before and after a cleaning activity. Data in the interval
between 10 and 15 minutes after cleaning was resampled due
to a lack of data in this time window. In the chart, the good
feedback slightly increases and the average and bad feedback
clearly drops, which indicates a positive impact of the cleaning
activities on passenger feedback.
Fig. 8: Mean number and trend of passenger ratings in 10-
minute buckets 30 minutes before and after cleaning
C. Simulation Model
Since it was not possible to deploy the Smart DPA so-
lution due to ongoing organizational changes at the airport,
we had to simulate the customer behavior in order to test
the hypothesis that dynamically scheduled cleaning based on
the PSI improves the passenger satisfaction. Thus, the event
processing engine was tested in isolation by using the historic
data of restroom passenger satisfaction events as above. Two
simulation approaches were checked, machine learning and
regression analysis.
1) Machine Learning: In the learning approach, we wanted
to simulate passenger feedback through a machine learning
approach. This included the construction of a neural network
machine learning model using logistic regression, of a k-
nearest neighbors algorithm (k-NN classifier), and of a random
forest classifier based on decision trees. The available data set
of passenger feedback was divided into a training set (70%
of the available data) and a test set (30% of the available
data). For training purposes, AWS (Amazon Web Services)
and scikit-learn was used.
During testing, the machine learning models turned out to be
not very precise. The neural network using logistic regression
provided a precision of 54%, the k-NN classifier 52% (k = 5),
and the random forest classifier 57% (n estimators = 500) for
this three-class problem. Due to our experiment, we assume
that the amount of training data is too low. The models tend
to overfit due to a lack of variation in the data set (it was
dominated by good ratings). More data needs to be collected
that has an impact on the passenger feedback to train the
model.
2) Regression Analysis: Since the machine learning ap-
proach based on the available data set did not provide good
results, we decided to use a trend and forecast estimation based
on the same set of restrooms accepting the lack of variance.
For this, we used the least squares method to compute the trend
of existent data and to forecast the trend for non-existing data,
a standard approach in regression analysis.
The average number of passenger ratings was computed in
buckets of 30 minutes after cleaning. The resulting trend of
passenger feedback was estimated for buckets between 0 and
270 minutes on existing numbers and continued into a forecast
for buckets between 270 and 1050 minutes. Figure 9 shows
the results.
Fig. 9: Trend and forecast for average restroom passenger
satisfaction numbers after a restroom was cleaned.
3) Simulation Parameters: In simulation mode, one hour
corresponds to 0.5 seconds, which allows us to simulate two
months in twelve minutes. Passenger feedback was linearly
distributed from the three categories good, average and bad
within a 30-minute bucket. The cleaning threshold on the
restroom cleanliness index (used to estimate the next cleaning
activity) was set to 85% in the first and to 90% in the second
simulation run. Once a necessary cleaning was detected,
restroom passenger satisfaction was reset to the beginning of
the respective trend line.
We selected the restroom with the air quality sensor for
the simulation, a restroom in the landside check-in area, since
it provides a complete log and an interesting test case with
variation between the three passenger feedback categories with
a correlation between feedback and the time passed after a
cleaning activity of corr=-0.55.
D. Result
In Table I the results from both simulation runs are sum-
marized and compared against the logged data of the selected
restroom.
Simulation 1: In the first simulation, 153 cleaning activities
were generated, cleaning on average 3 times a day. Compared
to the recorded data of the selected restroom, that is 3 times
less. Nevertheless, the restroom passenger satisfaction, based
on the formula described in V-A, was 48% compared to 42%,
i.e. six percentage points better.
258
Data set Cleaning Decreasing Cleaning Avg. Pax.thresh. slope activities index satisf.
Rec. data 0.85 (w=1) 0.05 (w=0.1) 312 0.57 42%Simulation1 0.85 (w=1) 0.05 (w=0.1) 153 0.82 48%Simulation2 0.90 (w=1) 0.05 (w=0.5) 296 0.66 51%
TABLE I: Comparison between the simulation runs and
recorded data from the restroom used as reference.
Simulation 2: Increasing the threshold to 90% and the
weight of the standard decreasing slope from 0.1 to 0.5 in
the second simulation run, we achieved a higher frequency
of cleaning activities. In this run, a satisfaction of 51%
was achieved with 296 generated cleaning activities. With
an almost equal number of cleanings, the simulated restroom
passenger satisfaction was nine percentage points higher with
an almost equal number of cleaning activities.
It is important for the scope of this analysis to focus on
the percentage increase we measured rather than the absolute
percentage value of this data that depends on the airport
and on the selected restrooms, i.e. apart from the cleanliness
of the restroom, there could be other reasons behind the
feedback given by customers, for example the size, fixtures
and crowdedness of the restroom.
The outcome of the simulation supports the hypothesis that
cleaning at the appropriate time has the potential, even if
conducted less frequently, to improve the restroom passenger
satisfaction. The Smart DPA solution enables such an improve-
ment and supports the improvement of satisfaction and the
effective scheduling of cleaning activities.
VII. CONCLUSION
Smart DPA provides a configurable event-driven architec-
ture efficiently processing and mapping heterogeneous data
streams occurring with different frequencies at the airport.
Based on historical data from the airport, it was shown that
the platform has the potential to improve passenger satisfaction
when services are scheduled according to the alerts generated
by the platform.
For monitoring restrooms, it would have been helpful to
integrate a system counting passengers entering and exiting
restrooms, which was not available. A system integration with
the airport infrastructure was started and could be used in the
future to show the effect of the approach in real life. This
would allow a study gathering qualitative feedback, e.g. neg-
ative feedback due to a restroom being closed at the moment.
A new situation is the detection of people infected with a
virus, indicated e.g. by fever based on a temperature-sensitive
camera. The detection of one such person would make clean-
ing immediately necessary again even it was cleaned minutes
before. In our system a respective score computation and
weighting may lend emphasis to the importance of such a
data source, but a new challenge arises for staff scheduling
when dealing with such exceptions. A transfer of the solution
to other domains, for example manufacturing, would be quite
interesting and is supported due to the reusable and extensible
design of the software.
ACKNOWLEDGMENT
This research was funded in part by the EIT Digital Tech
Activity 19194 (project SmartDPA) and by the German Fed-
eral Ministry of Education and Research under grant number
01IS19022E (project BaSys4.2).
REFERENCES
[1] S.-J. Hong, D. Choi, and J. Chae, “Exploring different airport users’service quality satisfaction between service providers and air travelers,”Journal of Retailing and Consumer Services, vol. 52, p. 101917, 2020.
[2] R. Curley. (2018) Airports turn to technology to keep airport toiletscleaner. [Online]. Available: https://www.businesstraveller.com/business-travel/2018/10/09/airports-turn-to-technology-to-keep-airport-toilets-cleaner/
[3] S. Vora. (2018) Believe it or not, airport bath-rooms are getting better (and cleaner). [Online]. Avail-able: https://www.nytimes.com/2018/10/05/travel/airport-bathrooms-cleanliness-apps.html
[4] P. Carbone, A. Katsifodimos, S. Ewen, V. Markl, S. Haridi, andK. Tzoumas, “Apache flink: Stream and batch processing in a singleengine,” Bulletin of the IEEE Computer Society Technical Committeeon Data Engineering, vol. 36, no. 4, 2015.
[5] G. C. Bezerra and C. F. Gomes, “Performance measurement inairport settings: A systematic literature review,” Benchmarking: AnInternational Journal, vol. 23, no. 4, pp. 1027–1050, Jan 2016.[Online]. Available: https://doi.org/10.1108/BIJ-10-2015-0099
[6] A. Lundberg, “Leverage complex event processing to improve opera-tional performance,” Business Intelligence Journal, vol. 11, no. 1, p. 55,2006.
[7] S. G. Boyer, J. R. Offutt, and K. P. Farrow, “Event driven airport,”U.S. Patent 20 020 198 747, December, 2002. [Online]. Available:http://www.freepatentsonline.com/y2002/0198747.html
[8] D. Luckham, Event Processing for Business: Organizing the Real-TimeEnterprise. Wiley, 01 2012.
[9] G. Pestana, S. Heuchler, A. Casaca, P. Reis, and J. Metter, “Complexevent processing for decision support in an airport environment,” Inter-national Journal on Advances in Software Volume 6, Number 3 & 4,2013, 2013.
[10] K. G. Zografos, M. A. Madas, and Y. Salouras, “A decision supportsystem for total airport operations management and planning,” Journalof Advanced Transportation, vol. 47, no. 2, pp. 170–189, 2013. [Online].Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/atr.154
[11] Xovis AG. (2020) Xovis PTS. [Online]. Avail-able: https://www.xovis.com/fileadmin/dam/documents/Xovis-brochure-Airports.pdf
[12] ANA Aeroportos de Portugal. (2017) I-sense. [Online].Available: https://www.routesonline.com/airports/8214/ana-aeroportos-de-portugal/routes-tv/5597756299001/
[13] Sita. (2020) Sita QueueAnalyzer. [Online]. Available:https://www.sita.aero/pressroom/news-releases/orlando-international-airport-eases-stress-at-security-with-sita-tech
[14] Smart Flows. (2020) Make customer-driven decisions to improve yourbusiness based on the flow within your airport. [Online]. Available:http://www.smart-flows.com/p/solutions-details/airport/
[15] Tibco. (2018) Airports: How to get insights from passengers. [On-line]. Available: https://www.tibco.com/sites/tibco/files/resources/WP-airports-how-to-gain-insights-from-passengers-final.pdf
[16] Infax. (2020) TRAX SmartRestroom. [Online]. Available:https://www.infax.com/solutions-airports.html#air-trax
[17] Tooshlights. (2020) Smart restrooms. [Online]. Available:https://tooshlights.com/modus-systems-software/
[18] RestroomApp. (2020) Real-time monitoring of restrooms. [Online].Available: http://restroomapp.com/
[19] Saniori. (2020) Digital toilet monitor. [Online]. Available:https://www.saniori.com/
[20] J. Evermann, J.-R. Rehse, and P. Fettke, “A deep learning approach forpredicting process behaviour at runtime,” in International Conferenceon Business Process Management. Springer, 2016, pp. 327–338.
[21] A. Emrich, A. Theobalt, F. Leonhardt, S. Knoch, D. Werth, and P. Loos,“A pervasive mobile assistance system for health and fitness scenarios,”in 2014 47th Hawaii International Conference on System Sciences.IEEE, 2014, pp. 2898–2907.
259