See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333627447

Data Analytics Platform for the Optimization of Waste Management Procedures

Conference Paper · May 2019

DOI: 10.1109/DCOSS.2019.00074

CITATIONS READS
14 3,946

6 authors, including:

Thanasis Vafeiadis Alexandros Nizamis

The Centre for Research and Technology, Hellas Information Technologies Institute (ITI)
50 PUBLICATIONS 1,151 CITATIONS 34 PUBLICATIONS 518 CITATIONS

SEE PROFILE SEE PROFILE

Vaia Rousopoulou Dimosthenis Ioannidis

The Centre for Research and Technology, Hellas The Centre for Research and Technology, Hellas
6 PUBLICATIONS 77 CITATIONS 233 PUBLICATIONS 2,135 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Thanasis Vafeiadis on 06 March 2020.

The user has requested enhancement of the downloaded file.

 Data Analytics Platform for the Optimization of
Waste Management Procedures
Thanasis Vafeiadis∗ , Alexandros Nizamis∗ , Vissarion Pavlopoulos∗ , Luigi Giugliano† ,
Vaia Rousopoulou∗ , Dimosthenis Ioannidis∗ and Dimitrios Tzovaras∗
∗ Centre
for Research and Technology Hellas-Information Technologies Institute (CERTH/ITI),
Thessaloniki, CO 57001 Greece
† MLW, Links Foundation, Turin, Italy

Email: ∗ {thanvaf, alnizami, apavlo, vrousop, djoannid, dimitrios.tzovaras}@iti.gr, † [email protected]

Abstract—Today, the use of IoT devices, the interconnectivity support and optimization of planning to the waste management
of machines and systems, and the development of novel analytic company. The scalable design of the platform enables its
algorithms which are main characteristics of Industry 4.0 are incorporation by further waste managements companies. This
also transferred to other domains connected to the industrial
one. The advances on industrial domain have impact on waste paper describes this platform's functionalities by emphasizing
management one as these domains are strictly connected. The on data analytic tools and algorithms.
connection with sensors on industrial partners premises, the big The remainder of this paper is structured as follows. In
data availability and the significant advances on data analytics, Section II, related work regarding waste management is given
enable the waste management companies to smarten their domain while in Section III a brief description of the waste man-
and automate many of their solutions and processes. In this work,
we present a data analytics platform for the optimization of agement platform and the embedded tools (fill level analysis,
waste management procedures. By using this platform, a waste tonnage forecasting, price forecasting, state prediction engine,
management company is able to monitor and analyze sensors optimal routes and statistical analysis tools) is provided. Fi-
data from industrial partners bins and optimize its planning nally, in Section IV the conclusions are drawn.
based on different analytic tools and historical data. Besides
the platform description, special focus is given on data analytics II. L ITERATURE R EVIEW
algorithms and methods.
Index Terms—data analytics, Industry 4.0 time series, smart A growing body of literature has been examined towards
waste management, deep learning, IoT platform smart waste management and leading-edge solutions. As a first
consideration, waste management monitoring is the starting
I. I NTRODUCTION point of waste management initiation and optimization. A
Smart waste management is a very important procedure in monitoring system provides, among others, transparency of
Industry 4.0 and it can be a profitable asset of both waste waste spend, access to historical data and the capability of
producers and waste management providers. Based on reports evaluating and improving the waste management through an-
of World Bank1 the total volume of generated wastes is going alytics. In the work of [1] a smart solid waste monitoring and
to get doubled by 2025. Moreover, Smart Waste Management collection system is presented, able to monitor the solid waste
Market was estimated over 1 billion USD in 2016 and it is collection and control the transportation system of disposal.
projected to reach 4 billion USD by 2025 based on the same Moreover, the advent of IoT technology is definitely im-
resources. These numbers indicate the importance of the smart proving waste management operations. Sensor-based bins and
waste management procedures as they can reduce the overall dedicated waste management applications are forcefully re-
transport and collection cost by 50%. placing obsolete methods. A monitoring system deploying
Today the smart waste management procedures are based IoT technology is presented in [2]. It is a smart garbage
on IoT sensor devices and data analytics tools. Variant factors management system which also makes the garbage collection
such as the continuously decreasing cost of IoT sensors, more efficient by monitoring the dust bin level. Likewise,
the growth of fast machine-to-machine connections and the [3] is a study for the development of a recycle management
constantly increasing performance of data analytic tools have information platform based on Internet of Things technology.
enable the adoption of smart waste management solutions. A more thorough solution is the one proposed in [4]. With
This work presents a data analytics platform regarding op- the use Smart-M3 platform, real-time monitoring of the level
timization problems of a waste management company, which of bins fullness through sensors is achieved and the incoming
participates as a pilot partner of the COMPOSITION2 EU data are provided to decision support algorithms. Last but
project. The platform aims to provide smart solutions that foremost, if we direct the attention to waste management, the
will enable monitoring processes based on sensors, decision IBM Intelligent Waste Management Platform [5] cannot go
unmentioned. This platform is addressed to both municipality
1 http://www.worldbank.org and waste management companies and provides them with
2 https://www.composition-project.eu analytics to improve operational processes.
 III. DATA A NALYTICS P LATFORM • MQTT protocol for data transmission.
A. Platform Overview The use of the aforementioned devices provides services for
different type of bins located on industrial plants through
The proposed data analytics platform offers a variety of
the platforms interfaces monitoring. The solution is based on
analytic services to end-users related to waste management
activities in order to enhance the decision-making and optimize
planning.
The platform was designed as a completely web-based
tool. It is developed in AngularJS3 framework, which takes
care of all the basic and content management features of
the application. Moreover, the MongoDB4 is used for the
tools storing requirements. The tool offers a wide variety
of advanced visualizations and it can be considered as a
Fig. 1. Bins Fill Level Monitoring.
visual analytics tool as well. The graphical representations are
enabled by the use of the Chart.js5 and D3.js6 visualization low power hardware offers over the air firmware update and
libraries. These libraries are ideal for producing dynamic, enable remote and real time monitoring. The monitoring is
interactive data visualizations in web browsers. In contrast a valuable asset for waste management companies as they
to many other, they allow great control over the final visual are continuously informed of their customers bins' fill level
results by using the widely implemented SVG7 , HTML58 and and they are able to act quickly, re-allocate resources(such as
CSS9 standard. From communication perspective both MQTT drivers and trucks) and re-organize their planning.
and HTTP communication protocols are supported. By using
these protocols the platform is able to connect to IoT devices C. Fill Level Analysis
and different analytic services. Both authentication and the Fill Level Analysis is a tool that utilizes a methodology
authorization services are supported by the platform for the called Slope Statistic Profile, denoted hereafter as SSP [6].
secure data exchange. Users log-in services are supported as SSP is a methodology that aims to estimate the change point
well. T from the profile of a linear trend test statistic, computed
The introduced platform supports different technologies as on consecutive overlapping time windows along the time
well as different types of features to the end-users. The main series. The test statistic for linear trend estimation is selected
platform features are: based on its high power compared to other test statistics for
• Monitoring of bins fill level based on IoT sensors. both correlated and white noise residuals. Modifications and
• Analysis of the bins fill level. extensions of SSP methodology has been widely used to deal
• Forecasting about the tonnage of wastes that is going to with anomaly detection problems in industrial sector. The real
be transported by a waste management company. time approach of SSP (Real Time - Slope Statistic Profile)
• Calculator for optimal pair of routes and tonnage. utilizes the information of two interdependent time series, real
• Price forecasting for various waste types/materials. time classification of their linear trend profiles signals and a
• Statistical analysis and visualization for better data ex- self-adaptive sliding window [7], [8], [9].
ploration and understanding of data.
All these main features are analyzed in the following sub-
section especially from data analytics and algorithms perspec-
tive.

B. Bins Fill Level Monitoring

The introduced IoT platform enables the bins fill level
monitoring for industrial customers of a waste management
company. The solution is based on smart IoT fill level sensors.
The core technologies of these IoT devices are:
• Ultrasonic sensors for fill level measurement.
10
• Long range communication provided by LoRa .
Fig. 2. Trend Analysis tool that utilizes the SSP methodology.
3 https://angularjs.org/
4 https://www.mongodb.com/ SSP method is applied on the time series of recordings
5 https://www.chartjs.org
(percentages) of a fill level sensor so as to receive notifications
6 https://d3js.org/
7 https://www.w3.org/Graphics/SVG/
of the fill level, deployed on waste containers of a waste
8 https://www.w3.org/TR/html5/ management factory. The interval between two recordings is
9 https://www.w3.org/Style/CSS/Overview.en.html 5 seconds (on average). The selected size of sliding window
10 https://lora-alliance.org for SSP is 20 elements, based on intensive simulations and
experiments. Fig. 2 shows the linear trend profile (right axis)
and the fill level time series (left axis). When the linear trend
of fill level time series is steady, the linear trend profile points
out that there are not significant changes indeed. In this test
case, the time series has many ups and downs in small time
intervals. This fact has as a result, the SSP methodology to
detect many changes on the linear trend (see Fig. 2 , right
axis). In the case of fill level sensor, the SSP methodology, is
set to detect only the time points when the linear trend has
upward changes, a fact that indicates a state of bin fullness.
In general, the real time approach of SSP method has the
ability to detect multiple change points in linear trend of a time
series and it is most effective in cases when the change(s) in
a time series is(are) not abrupt, as happens in most natural
variations.

D. Tonnage Forecasting
This algorithm helps us to forecast new observations based Fig. 3. Time Series Forecasting tool that utilizes moving average models.
on a time series. This algorithm uses smoothing methods. The
moving average, denoted hereafter as MA, algorithm is used
only on time series that does not have a trend. This method is actually very simple to find out. Given a transition matrix
is by far the easiest and it consists of making the arithmetic P , this can be determined by calculating the value of entry
mean of the last n observations contained by the time series to (i, j) of the matrix obtained by raising P to the power of
forecast the next observation. We use the following formula: M . For the time series of median tonnage per month for all
Pt customers of the waste management factory, we define two
xi
M At + 1 = i=t−n possible states: “Up” and “Down”, which means that a value
n of the time series at its current time can be followed by a
where t is the period of time (or the size of the time series) larger (“Up”) or a smaller (“Down”) value. We could also
and xi the actual time series. In some cases, we need to find the define the a third state when the current and the future value
optimal number n of observations to be used in the forecast. are the same (Steady), but the form of the time series does not
We can find it by checking the square error mean of multiple provide any clues for the existence of this state. Thus, based
n observations. Basically, the range for forecasting elements on the assumptions described above, the transition matrix P
is to start at two observations up to half of the data set size for the median tonnage per month time series is calculated
plus one. The moving average model has been applied on the and given below:
time series of tonnage weights per month for n = 2, 3, 4, 5,  
0.222 0.778
which means that the predictions based on moving average P =
0.857 0.143
model will take place from 2 to 5 months ahead (see Fig. 3).
From the transition matrix P one can see that the probability
E. State Prediction Engine of being at state “Up” and the next state to be the same is 0.222
Markov chains are a fairly and simple common way to and the probability of being at state “Up” and the next state
statistically model random processes. Markov Chains are con- to be “Down” is 0.778. Moreover, the probability of being
ceptually quite intuitive, and are very accessible in that they at state “Down” and the next state to be “Up” is 0.857 and
can be implemented without the use of any advanced statistical the probability of being at state “Down” and the next state to
or mathematical concepts. Formally, a Markov chain model be “Down” is 0.143. Based on transition matrix P it is easy
is a probabilistic automaton. The probability distribution of to calculate the probabilities of futures states over M steps.
state transitions is typically represented as the Markov chains For example, the probability of being at state “Up” and after
transition matrix P . If the Markov chain has N possible one month to be at state “Down” the probability is 0.777 (see
states, the matrix will be an N xN matrix, such that entry Fig. 4).
(i, j) is the probability of transitioning from state i to state j.
Additionally, the transition matrix must be a stochastic matrix, F. Price Forecasting
a matrix whose entries in each row must add up to exactly one. A Deep Learning algorithm is implemented at this point,
This makes complete sense, since each row represents its own able to derive the latest prediction on the price per ton
probability distribution. at which users are likely to accept to buy/sell scrap metal
Markov chain models can be used to determine the proba- within a fixed timeframe in the future. The Price Forecasting
bility of moving either from state i to state j for a single step is a prediction model based on historical price data. The
or from state i to state j over M steps. As it turns out, this initial dataset consists of aggregated sales statistics spanning
 prices pattern with the help of continuous learning. Hence,
we come to the conclusion that if the circumstances in which
the network has been trained are the same with the ones used
to predict or to apply continuous learning, the network will
converge. The Price Forecasting of the presented platform is
shown in Fig. 5. A price forecasting analysis is presented for
the future including the prediction values and the coefficient
of determination [10] expressed as accuracy rate in the same
diagram.

Fig. 4. State Prediction Engine tool that utilizes Markov Chain approach.

from 2016 to 2018, specifically, price values over time and

records of respective fluctuations. All records refer to the same
customer, but differentiate in terms of material type (paper,
PET, HDPE, scrap metal) and timespan. The samples have
the following features: material type, date start, date end, tons,
price per ton. In order to produce a reasonable amount of data
for training the neural network, the initial dataset has been Fig. 5. Price Forecasting tool that utilizes deep learning algorithm.
resampled on daily basis. This intended to improve the quality
of the dataset, since the provided data included only one entry G. Optimal Routes Calculator
per month. The resulting datasets for each material consist of
People use search engines to find information they desire
three parts: (1) the price trend of the material, (2) the trend
with the aim that their information needs will be met. Infor-
of daily amount of good exchange in tons and (3) the price
mation retrieval (IR) is a field that is concerned primarily with
variations over time, which is a synthetic feature.
the searching and retrieving of information in the documents
Predicting prices of raw materials is a very tough task for
and also searching the search engine, online databases, and
Machine Learning algorithms considering the lack of pattern
Internet. Genetic algorithms (GAs) are robust, efficient, and
and the many parameters affecting it, which cannot easily be
optimized methods in a wide area of search problems moti-
listed. Normally, the network seems effectively able to learn
vated by Darwins principles of natural selection and survival
the price trend, but it often predicts a value equal to the value
of the fittest. An IR system searches for the matches in the
at the previous time point, though it still minimizes the loss
document databases and, thus, retrieves search results of the
function. This behavior, named “mimicking”, is very harmful
matching process. However, based on the relevance, the user
especially with dataset composed by sequential constant size
will then evaluate and display the search results. The relevance
steps like the ones described above. In order to prevent the
of the document is very important to the user. If the user feels
“mimicking” behavior a new hand-crafted feature has been
that it is a relevant document, he finishes the search while
added, which consists of a boolean feature taking the value one
else user continues to search in the document database by
every time that a new set of values start to learn, i.e. the third
reformulating the query until the relevant documents that will
part of the dataset described above. Furthermore, the network
satisfy users information needs are retrieved.
architecture and hyper-parameters has been tuned to cope with GA is a probabilistic algorithm simulating the process of
this particular issue. However, the best way to improve the natural selection of living organisms and finally coming up
results would be to add features that go beyond historic prices with an approximate solution to a problem [11], [12], [13]. In
alone. The initial network for regression is composed of four GA implementation, the search space is composed of candidate
hidden Long Short Term Memory11 (LSTM) layers with 64, solutions (called individuals or creatures) to an optimization
32, 24 and 8 neurons respectively and hyperbolic tangent as problem to evolve better solutions; each represented by a string
activation function. is termed chromosome. Each chromosome has an objective
It is almost impossible to understand the long-term im-
function value, called fitness. A set of chromosomes together
plications on the network behavior because they are strictly
with their associated fitness is called a population. This
dependent on the input data that will be provided to the
population, at a given iteration of the genetic algorithm, is
network. Nevertheless, it is safe to assume that, if the boundary
called a generation. In each generation, the fitness of every
conditions do not change from the time of obtaining the data
individual in the population is evaluated from the current
to the time of using new data, the network will learn the
population based on their fitness value and modified to form a
11 https://www.researchgate.net/publication/13853244 Long Short- new population. The new population is then used in the next
term Memory iteration of the algorithm.
 GA terminates when either a maximum number of genera-
tions has been produced or a satisfactory fitness level has been
reached for the population. If the algorithm has terminated due
to a maximum number of generations, a satisfactory solution
may or may not have been reached. The working of the genetic
algorithm depends upon the constraint of how well we choose
our initial random keywords.
The Optimal Routes Calculator is applied on monthly
recordings of routes and tonnage weights of brown paper of
the waste management factory for 48 months, between 2015
and 2018. As a result, it provides 10 simulated options for the
routes and tonnage weights that the waste management factory
should follow in order to have several solutions (see Fig. 6).
An initial version of the approach has been described in [14],
as a tool of a decision support system.

Fig. 7. . Statistical Analysis tool.

specific materials (brown paper, scrap metal, plastic etc.), price

forecasting and the optimal values for routes and transported
tonnage. The techniques and methodologies used to provide
these solutions come from the field of statistics, probability,
deep learning, time series modeling and linear trend analysis
and can be easily re-used for similar problems and cir-
cumstances. By means of advanced visualization techniques,
the platform presented here is an easy-to-use tool appealing
mainly to waste management companies. The monitoring
functionality and deployment of IoT technology provides
supervised control of waste level and access to historical
data. The plentiful analytics methods provide the potential
of planning and optimizing waste management practices and
ultimately, gaining financial and environmental benefits.
Fig. 6. Optimal Routes Calculator tool that utilizes genetic algorithm. ACKNOWLEDGMENT
This project has received funding from the European Unions
H. Statistical Analysis Horizon 2020 research and innovation programme under grant
Statistical analysis is a component of data analytics and in agreement No 723145 - COMPOSITION. This paper reflects
the context of business intelligence it involves collecting and only the authors views and the Commission is not responsible
scrutinizing every data sample in a set of items from which for any use that may be made of the information it contains.
samples can be drawn. The Statistical Analysis tool presented
here, gives only a brief description of the nature of the data and R EFERENCES
carries out some basic statistical measures and visualizations [1] S. A. Hassan, N. G. M. Jameel and B. ekerolu, “Smart Solid Waste
so as the end-user has the ability to draw some initial inference Monitoring and Collection System.” International Journal, 6(10), 2016.
[2] S. Bhutada, S. Deshmukh, M. Keny, W. Memon, G. Hegde, “ Smart
from the data to be analyzed. The Statistical Analysis tool is Waste Management using IOT, International Journal of Research in
applied on overall recordings of price and tonnage weights of Engineering, Science and Management (IJRESM), 1(4), 2018.
the waste management factory, between 2016 and 2018 (see [3] C.Tao and L.Xiang, “Municipal solid waste recycle management infor-
mation platform based on internet of things technology.” 2010 Interna-
Fig. 7). tional Conference on Multimedia Information Networking and Security.
IEEE, 2010.
IV. C ONCLUSIONS [4] V. Catania and D. Ventura, “An approch for monitoring and smart
This work presents the current version of an IoT data planning of urban solid waste management using smart-M3 platform.”
Proceedings of 15th conference of open innovations association FRUCT.
analytics platform for the optimization of several procedures IEEE, 2014.
in a waste management factory. On-board there are data [5] “IBM Intelligent Waste Management Platform”, IBM Sales and Distri-
analytics solutions (algorithms and methodologies) regarding bution White Paper, 2015.
[6] T. Vafeiadis, E. Bora-Senta, and D. Kugiumtzis, “Estimation of linear
key aspects on waste management, such as waste bins fill level trend onset in time series, Simulation modelling Practice and Theory,
monitoring and analysis, forecasting of transported tonnage for 19 (5), 1384 1398, 2011.
 [7] T. Vafeiadis, S. Krinidis, C. Ziogou, D. Ioannidis, S. Voutetakis and
D.Tzovaras, “Robust malfunction diagnosis in process industry time
series, IEEE 14th International Conference on Industrial Informatics,
111-116, 2016.
[8] T. Vafeiadis, S. Krinidis, C. Ziogou, D. Ioannidis, S. Voutetakis, D. Tzo-
varas and S. Likothanassis, “Real-time incident detection: An approach
for two interdependent time series, 24th European Signal Processing
Conference, 1418-1422, 2016.
[9] T. Vafeiadis, C. Ziogou, G. Stauropoulos, S. Krinidis, D. Ioannidis, S.
Voutetakis, D. Tzovaras and K. Moustakas, “Early malfunction diagnosis
of industrial process units utilizing online linear trend profiles and real
time classification, International Journal of Adaptive Control and Signal
Processing, 32(9), 1313-1325, 2018.
[10] R. Taylor, “Interpretation of the Correlation Coefficient: A Basic Re-
view”, Journal of Diagnostic Medical Sonography, 6(1), pp. 3539, 1990.
[11] K. A. DeJong, “An Analysis of the Behaviour of a Class of Genetic
Adaptive Systems”, University of Michigan, 1975.
[12] D. E. Goldberg, “Genetic Algorithms in Search, Optimization, Machine
Learning”, Addison Wesley, 1989.
[13] J. H. Holland, “Adaptation in Natural and Artificial Systems”, The
University of Michigan Press, Ann Arbor, Mich, USA, 1975.
[14] T. Vafeiadis, D. Kalatzis, A. Nizamis, D. Ioannidis, K. Apostolou, I.
Metaxa, V. Charisi, C. Beecks, G. Insolvibile, M. Pardi, P. Vergori,
and D. Tzovaras, “Data analysis and visualization framework in the
manufacturing decision support system of COMPOSITION project”,
Procedia Manufacturing, 28, 57-62, 2018.

View publication stats