Cognitive Solutions in Process Industry: H2020 CAPRI Project

Cristina Vega

, Daniel Gómez

and Aníbal Reñones

Fundación CARTIF, Parque Tecnológico de Boecillo, 205, 47151, Boecillo (Valladolid), Spain

Keywords: Process Industry, Automation, Industry 4.0, IIoT, Cognitive Platform, Innovation, Digital Transformation,

Industrial Plants, Smart Modules, Smart Industry, Open Data, Open Source, Open Science, Asphalt.

Abstract: The CAPRI project is a H2020 project that develops Cognitive Solutions (CS) to the Process Industry and a

Cognitive Automation Platform (CAP) towards the Digital Transformation of process industries. CAPRI

enables cognitive tools to provide to the existing process industries flexibility of operation, improving the

performance and quality control of its products and flows. The project is developing and testing different CS’s

at each automation level, from sensors to planning. The content of this paper is focused on the CAPRI asphalt

production applying different CS’s for the sensors and control levels. Specifically the paper discusses a

cognitive sensor for measuring filler quantity to the filter at drying process (noted as CAS2) and cognitive

control concept applied to optimize the operation of the rotary dryer (noted as CAC1). The paper explains

also how the CS’s are being integrated by means of an open source architecture based on FIWARE. The paper

provides also open access to the data and algorithms used as part of the commitment of CAPRI with open

science.

1 INTRODUCTION

Big data and artificial intelligence (AI) are giving a

huge boost to Industry 4.0. Intelligent software

solutions based on AI models can process high

volumes of data generated to identify trends and

patterns that can be used to make manufacturing

processes more efficient and reduce their energy

consumption (ElMaraghy & ElMaraghy, 2022).

An extension of this is to incorporate cognitive

features that enable sensing complex and unpredicted

behaviour and reason about dynamic strategies for

process optimization, leading to a system that

continuously evolve its own digital structure as well

as its behaviour. This way, an industry process will

have its own cognitive capabilities over time based on

the data it will collect and experience it will gain

(Abburu, et al., COGNITWIN - Hybrid and Cognitive

Digital Twins for the Process Industry, 2020).

Cognitive computing (Essa, et al., 2020) is an

interdisciplinary field, which uses a collection of

technologies to build a machine that have reasoning

capabilities like a human brain. Cognitive computing

https://orcid.org/0000-0002-7670-5088

https://orcid.org/0000-0001-6123-2401

https://orcid.org/0000-0002-4702-4590

integrates machine learning techniques to facilitate

computers to recognize the objective world and to

make decisions. Cognitive technologies have large

influence on different systems and technologies such

as cloud, mobile, wearable devices, IOT, big data, and

industrial production (Abburu, et al., Cognitive

Digital Twins for the Process Industry, 2020).

This paper is organized as follows: Section 2,

introduces the novel paradigm of what is known as

cognitive manufacturing. Section 3, shows how this

concept is present in the H2020 CAPRI project. More

specifically, the asphalt use case is shown and

presented as an industry sector where CS’s could

make a big improvement in terms of efficiency. Then,

in the following sections, two of the CS’s developed

for the asphalt use case are explained. Section 4 deals

with the Reference Architecture that is being

deployed as part of the CAP concept, based in the

open source FIWARE framework and how the

reference architecture enables an easy integration of

the CS explained for the asphalt plant. Section 5 ends

with the conclusions and next steps.

Vega, C., Gómez, D. and Reñones, A.

Cognitive Solutions in Process Industry: H2020 CAPRI Project.

DOI: 10.5220/0011562000003329

In Proceedings of the 3rd International Conference on Innovative Intelligent Industrial Production and Logistics (IN4PL 2022), pages 267-278

ISBN: 978-989-758-612-5; ISSN: 2184-9285

267

2 COGNITIVE

MANUFACTURING

One of main challenges for process industry plants is

to enable an efficient monitoring and control when

the production or environments are complex, e.g. due

to harsh conditions the system is operating in. The

basic elements of process monitoring and control

loops, including the models which can be used for

supporting this task cannot be solved easily using nor

traditional techniques from process monitoring (like

Statistical Process Control) neither solely by using

advanced AI techniques (like predictive analytics)

(Cinar, Nuhu, Zeeshan, & Korhan, 2020). This

problem requires a better understanding of the

underlying data and processes, their contexts and

their dynamics, similarly how human cognition is

building a superior situational understanding and

reasoning (Jacoby, Jovicic, Stojanovic, & Stojanović,

2021), even in very ambiguous cases. CAPRI uses the

analogy of human cognition, based on cognitive

architecture (Kaur & Sood, 2015), for addressing

above challenges. It must be emphasized that the

human cognition is extremely efficient in getting a

big picture of a situation at hand, i.e. not only what is

happening (Eirinakis, y otros, 2022), but also what is

causing the situation and what can be the

consequences before understanding what is going on

and how to react on (Sánchez Boza, Guerra, & Gajate,

2011). Complex behaviour arises from sequences of

cognitive cycles and this is exactly how CAPRI

envisions the process of monitoring/sensing and

controlling/reacting in cognitive plants.

3 H2020 CAPRI PROJECT

Digitalisation represents a new challenge for the

European process industries, which need to handle an

increasingly wide range of actions (Sharma, Kosasih,

Zhang, Brintrup, & Calinescu, 2020). Cognition

capabilities will permit the sector to improve its

flexibility and performance. The EU-funded CAPRI

project (Consortium, 2022) will establish, test and

demonstrate an advanced CAP for process industry

digital transformation. The platform will help process

industries increase its flexibility of operations and

improve performance through different indicators

and cutting-edge quality control of products and

intermediate flows. The CAP will be modular and

scalable, allowing the development and integration of

advanced applications that address manufacturing

challenges in significant process sectors such as

asphalt, steel making and pharma.

Figure 1: H2020 CAPRI project introduction.

European process industries need to address

resources, materials and environmental constrains by

improving its flexibility and performance through

cognition capabilities, as existing in human

intelligence. Digitisation is the main enabler for such

capabilities (Auditors, 2021).

CAPRI Cognitive Automation Platform for

Process Industry enabled by cognitive tools will

provide existing process industries flexibility of

operation, improvement of performance across

different indicators (KPIs) and state of the art quality

control of its products and intermediate flows.

The CAP will encompass methods and tools for

governing six Digital Transformation pathways (6P,

Product, Process, Platform, Performance, People,

Partnership) (Salis, Marguglio, De Luca, Gusmeroli,

& Razzetti, 2022), a Reference Architecture with four

levels of cognitive human-machine interaction

(industrial IoT connections, smart events processing,

knowledge data models and AI-based decision

support), a set of reference implementations, both

commercial and open source, for batch, continuous

and hybrid process industry plants, and a toolbox of

CS’s for planning, operation, control and sensing.

The CAP will be modular and scalable, so that

advanced applications could be developed and

integrated on top of it and its validation will take place

addressing manufacturing challenges in industrial

operational environments of three outstanding

process sectors: asphalt (minerals), steelmaking, and

pharma industry (chemical). CAPRI results could be

applied to a wide range of problems and challenges in

future cognitive plants. CAP Platform and the

cognitive tools included in it can be replicable in areas

of production planning, control, automated processes

and operations of many process industry sectors.

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

268

3.1 Asphalt Use Case

The asphalt use case of the H2020 CAPRI project is

located in EIFFAGE Gerena plant (located in

southern Spain). A general overview of the asphalt

manufacturing process of the corresponding use case

is (El-Haggar, 2007) shown in Figure 2.

Figure 2: Asphalt manufacturing process diagram.

By weight, 95% of asphalt consists of gravel, sand

and filler (aggregates less than 63µm) – aggregates

that give asphalt its strength. The remaining 5%

comprises an agent that binds all of these materials

together. That agent is usually bitumen derived from

crude oil. The process begins when the stockpiled

aggregates in the cold feed are metered and conveyed

to a dryer drum where they are heated to a specific

temperature. A first collector removes large dust

particles from the gases before entering the bag

house, which removes fine particulate matter before

they are released into the atmosphere.

Hot aggregates are elevated to a vibrating screen

where they are classified by size and stored in

different bins. The aggregates, filler and other

additives are scaled and mixed with the hot bitumen

in the mixer producing the final asphalt mix. In some

asphalt mixes, RAP (Reclaimed Asphalt Pavement) is

also added. RAP is scaled, taken into account its

approximated bitumen content (measured in a

laboratory), and added to the mixer. After that, the

asphalt mix is ready to be loaded to the truck for

shipment (Sivilevičius & Šukevičius, 2009).

3.2 Objectives and Benefits

In the asphalt mix manufacturing process, most of the

measured data is not usually exploited although it

may provide very interesting information. There

could be variables that are not known how to relate

with the information obtained or whose relationship

is unknown. Even more, some variables are not

measured or measured only in the laboratory.

CAPRI project addresses the challenge of

integrating relevant information data sources as well

as knowledge of the personnel of the plant, at all the

levels: planning, operation and control of the plant

(Zhang, Huchet, & Hobbs, 2019). The results of the

project are translated in terms of costs, effectiveness,

and product quality for the asphalt mix manufacturing

process. With the development of CAPRI, for the

asphalt use case there should be five kinds of

improvements in the plant.

At production performance level, the objective is

to increase productivity by around 8% with the CS.

CAPRI will act as well in the energy efficiency.

Objective is to decrease the consumption of 15% of

electricity, 11% recycled fuel and 50% diesel. These

will be improved with the cognitive control of dryer

drum (known as CAC1). Knowing the humidity and

temperature in the input of the drum, adjustments will

be made to obtain the best conditions of output,

avoiding overheating of aggregates.

The next benefit, in the asphalt use case, is related

to the consumption of resources, like, aggregates,

bitumen and RAP. In this case, is to reduce 20% of

aggregates consumption, 20% of the consumption of

bitumen and increase in 500% RAP consumption.

Related with waste generation and the product

quality, the point to action is the control of hot

aggregates temperature. To obtain this, the cognitive

control of asphalt plant dryer drum (CAC1) is

required to optimise aggregates heating of

temperature, and the cognitive sensor of amount of

filler (CAS2) in combination with the other CS’s are

needed to control the filler present in the aggregates

and the filler needed in the final mix.

3.3 Cognitive Solution – Control of the

Asphalt Drum (CAC1)

The asphalt drying process aim is to produce a dry

solid product of desired quality at minimum cost and

maximum throughput. Good quality implies that the

product corresponds to a number of technical,

chemical and biological parameters, each within

specified limits (Yliniemi, Koskinen, & Leiviskä,

1998).

Different control techniques, at different levels,

are present in this type of equipment, ranging from

conventional industrial controls (like PID) to more

advanced control systems like model-based

feedforward-feedback until recently applied

intelligent control systems based on fuzzy logic or

Cognitive Solutions in Process Industry: H2020 CAPRI Project

269

Figure 3: CAC1 and CAS2 Basic Architecture.

neural networks applied to machine learning

techniques (Raghavan, Jumah, & Mujumdar, 2006).

Within CAPRI project, CAC1 Cognitive Control

Solution objectives are to obtain a dry product at an

optimum temperature and fumes (combustion gases)

at the possible lowest temperature, on one hand not to

damage the baghouse filter and on the other to

minimize energy consumption, thus increasing the

efficiency of the drying process. The main objective

is to decrease the consumption of electricity, recycled

fuel and diesel. This way, knowing the humidity and

temperature in the input of the drum, adjustments will

be made to obtain the best conditions of output,

avoiding overheating of aggregates.

This solution has been developed based on a

control algorithm where sensors and actuators are

used to calculate the optimum values for the different

variables that run the drum. Currently, a dynamic

modelling of the rotary drum is being created through

model-based identification methods (Ljung, 1998)

running several experimental tests performed at the

asphalt plant taking into account some of the main

variables: temperatures, humidity, load to dry, burner,

drum speed, combustion gas flow. This identified

model will be required like an input for the Model

Predictive Control (MPC) (Schwenzer, Ay, Bergs, &

Abel, 2021), advanced method of process control that

is used to control a process while satisfying a set of

constraints. It is in this control solution where the

rotary drum optimized control calculations are

performed.

The Cognitive Algorithm will be executed in real

time by providing the setpoints: drum burner power

(%), drum rotation speed (%) and exhaust damper

opening (%), to obtain the optimal temperature of the

hot aggregates coming out of the drum and to

guarantee in this way the desired temperature of the

final asphalt mix and also the gas combustion

temperature. In addition, this is intended to minimize

the combustion gases temperature and to improve

energy efficiency and reduce pollution.

CAC1 Data Model and Algorithm

Different attempts have been made to model rotary

dryer drums, ranging from physical equations as in

(Rubio, Bordons, Holgado, & Rivas, 2001), (Le

Guen, Huchet, & Tamagny, 2011), numerical

analysis as in (Li, Yao, & Zhao, 2017) and energy and

exergy equations (Zhang, Huchet, & Hobbs, 2019).

Regarding control algorithms for rotary dryer

drums, several approaches can be found through the

literature, from basic algorithms as in (Rubio,

Bordons, Holgado, & Rivas, 2001) to an advanced

control based on fuzzy logic, (Yliniemi, Koskinen, &

Leiviskä, 1998), (Koskinen, 1998), variable structure

controller as in (Mahmoud, El-Kasassy, & Areed,

2020). More recently, intelligent control applied to

rotary dryer drums has also been approached from

different intelligent perspectives: a decisive control

module (Pang, Jia, Ding, Yu, & Liu, 2021), rule-

based expert and neural networks (Raghavan, Jumah,

& Mujumdar, 2006) and a more sophisticated control

based on cognition with self-X capabilities as in

(Haber, Juanes, Del Toro, & Beruvides, 2015) in a

more general way.

CAC1 algorithm applied in CAPRI consists of an

identified data-based model and an MPC

programmed in MATLAB environment using both

MATLAB scripts and SIMULINK function blocks.

The identified model has been done using the data

CRUDE GAS

CLEAN GAS

Temperature

Humidity

Flow

Weight/level

Sensors

Control

Set-points

Depression

Speed

DRUM

FAN

Electric power

Cold

aggregates

Extracted

filler

Dust

emission

Sensors

CAC1 -

Algorithm

Burner control

MES (CAP1)

Dried

aggregates

Burner

SPx

Drum speed control

Gas composition

CAS2

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

270

with certain predefined conditions and with different

tests performed at the asphalt plant. The manipulated

variables are the ones used to the tests for

identification purposes: First set of experiments were

based varying the rotary dryer burner power (and

leaving the rest of the variables as constant as

possible) and a second set of tests where the varying

variable was the dryer drum rotary speed. The CAC1

experimentally data-based identified model, from all

sensor measurements and the dynamics (data-based

model) in the production chain and related process

variables calculates the setpoint SP1 of the drum

temperature controller. The setpoint of combustion

gases temperature controller leaving the rotary drum

dryer SP2, modulating the speed of rotation of the

drum will also come from this AC algorithm.

The MPC calculates and changes in real time the

setpoints of the plant PLC slave controllers from the

setpoints generated by the CAC1 algorithm.

All related CAC1 files can be accessed at Zenodo

open CAPRI link (Gómez & Diego, 2022).

3.4 Cognitive Solution – Cognitive

Sensor for Amount of Filler (CAS2)

The asphalt plants contain different types of sensors

in order to be able to monitor and control the different

stages of the production process like temperature

sensors, humidity sensors, pressure sensors, load cells

and more.

However, not all these sensors per se give a smart

understanding and approach to the process. During

process assessment in the CAPRI project, some CS

have been identified to find an optimal behaviour and

reaction to the manufacturing of asphalt mixes to give

a high-level cognition reaction to optimize and detect

variations and have a cognitive sensing and support

of the process that commercial sensors cannot give.

One of these identified CS is: Cognitive sensor of

amount of filler (known as CAS2).

This cognitive sensor is developed to estimate and

measure the fine filler quantity that goes out of the

aggregates drying to the baghouse filter (Figure 3

with position of CAS2 solution). The high-level

outcome of this cognitive sensor is to obtain the real

amount of filler present in the cold aggregates, which

allows then wasting less energy in the rotary drying

drum and in the filtering (baghouse) process.

Different technologies and approaches can be

found to tackle this measurement: Laser technology

that uses a time-of-transition technique to measure

particle size distribution (Measuring Coal Particles in

the Pipe, 2022); machine vision to analyse particulate

material on conveyor belts as in (Andersson, 2010);

also, techniques applied using intelligent vision with

camera images applied to different structures of new

neural networks to image processing and estimate the

granulometric distribution of small and medium size

aggregates, (Fernández, Viennet, Goles, Barrientos,

& Telias, 1998); more recent techniques based on 3D

particle tracking velocimetry in up and down flows in

pipes, (Oliveira, Van Der Geld, & Kuerten, 2017);

miniaturized sensors based on nanofibers to

determine vibrations and then analyse the possible

flow in different structures, (Singh, Lye, & Miao,

2019); also, new capacitive sensors and electrodes

using calibration-based and tomographic approaches

have also been recently presented to measure

particulate flow in pipes, (Suppan, Neumayer,

Bretterklieber, Puttinger, & Wegleiter, 2022).

All these techniques are not appropriate to be

deployed in asphalt production due to the required

harsh conditions of this process: high temperature,

pressure, and concentration of abrasive particles.

Eventually CAS2 solution has two kinds of

physical sensors, one is a commercial solution, that

has never been used under these conditions. The

second sensor is a custom sensor based on another

commercial sensor, not intended to measure

concentrations of particles, but to measure

disturbances of the flow, which can then be used to

estimate the amount of filler flow through the pipe of

baghouse. This second sensor is a research and

innovation action of this project. It is based on a

vibration measurement that has been validated under

laboratory conditions providing an actual

measurement of filler flow at a smaller scale process.

Thanks to the knowledge that this sensor will

provide (actual mass flow of filler trough baghouse

aspiration pipe), the needed filler addition and

extracted will be minimized and added only if it is

detected that there is less amount of filler than the

final hot mix needs.

Therefore, the outcome of this cognitive sensor,

the estimation of filler present in the cold aggregates,

will help to avoid excessive recirculation and/or

unnecessary addition of filler, making the process

more energy efficient. In the development of this

cognitive sensor, different steps were taken. At early

stages, laboratory measures and test were performed

with the following results, where all the referred files

are openly available at CAPRI Zenodo repository

(Vega & Reñones, Cognitive sensor for amount of

filler, 2022):

At lab scale, different tests were performed, where

different parameters were taken into account. Also, a

set of vibration sensors were installed at the actual

plant baghouse inlet pipe and compared against a

Cognitive Solutions in Process Industry: H2020 CAPRI Project

271

Figure 4: Comparison of sensors measures in Gerena use case plant (Seville, Spain). The process variables of aggregates flow,

pressure at the baghouse and temperature at the baghouse are compared with CAS2 cognitive sensor.

commercial solution not usually used at this location

due to the harsh conditions present at these points.

CAS2 vibration parameter is compared with the

results offered by the commercial sensor tested in

parallel (measured in ppm) (Figure 4). The data also

contains relevant process variables from the control

of the plant like the SP of aggregates flow into the

drying drum (in T/s) the aspiration pressure at

baghouse input pipe (in mm H20) and the temperature

at such input. The sequence of operation of the

baghouse and drying drum is the following (Table 1

and Figure 4). File named CAS2_Data_4.csv is a

dataset file that represents the amount of vibration

measured in the EIFFAGE asphalt plant during the

drying process of aggregates in that sequence of

operation.

CAS2 sensor aims to provide an estimation of

flow of filler during the drying process of the

aggregates. As such, the raw measurements need to

be adjusted to compensate the undesired noise when

the aspiration takes place but there are no aggregates

into the drum to be dried. To compensate this noise a

model based on the actual aspiration pressure has

been created.

Table 1: Important events of comparison the process

variables from Gerena Plant and the CAS2.

TIME

EXPLANATION

Aspiration of the baghouse starts

1,2

Baghouse depressure maintains still, sensors

starts to measure the flow without filler

Flow of aggregates starts

2,3

Production, measure of filler

Flow of aggregates stops

3,4

Measure of filler, production off, baghouse on

Baghouse depressure decreases, flow of filler

stops

4,5

No aggregates, baghouse with constant

depressure, low vibration of sensors

Asphalt plant stops

CAS2 Data Model and Algorithm

All the performed calculations are explained

following the files that can be openly found at (Vega,

Reñones, & Sanz, Cognitive sensor for amount of

filler [CAS2] - INTEGRATED, 2022). File named

CAS2_dataset_5.RData is a dataset of raw data used

for the creation of the model of CAS2. This model

Instant

0,2

0,4

0,6

0,8

1,2

1,4

1,6

1,8

100

150

200

250

300

6:50:24 7:19:12 7:48:00 8:16:48 8:45:36 9:14:24

Production set-

point (Tn/h)

Baghouse pressure

(mm/H20)

Vibration (gRMS)

2000

4000

6000

8000

10000

12000

14000

100

150

200

250

300

6:50:24 7:19:12 7:48:00 8:16:48 8:45:36 9:14:24

Production set-

point (Tn/h)

Baghouse pressure

(mm/H20)

COMMERCIAL

SENSOR (PPM)

0,5

1,5

2000

4000

6000

8000

10000

12000

14000

6:50:24 7:19:12 7:48:00 8:16:48 8:45:36 9:14:24

COMMERCIAL

SENSOR (PPM)

Vibration (gRMS)

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

272

Figure 5: CAP Reference Architecture.

tries to estimate the vibration measured (ACEL1_20-

25 [gRMS]) based on the aspiration pressure (variable

named as RPA2100 [mmH2O]).

Figure 6 shows a collection of temporary

moments in time when the baghouse is running, but

there is no material flow, so it is just vacuuming air.

Figure 7 shows an operation of the baghouse during

one day of production of the asphalt plant and how

the different segments used for the creation of the

model are selected (marked with red rectangles).

The file CAS2_sourcecode_1.R file is an

algorithm programmed in an open source, R

programming, environment and language. This

algorithm creates a model that relates the aspiration

(x variable) and vibration variables (f function) in the

suction process with the dataset described above. The

developed model creates a piecewise linear

relationship between the two variables for aspiration

values as can be seen in the figure below. It must be

also noted that for a certain pressure below a

threshold (27 mmH

O) the output of the model is

forced to 0 as the baghouse does not operate and the

flow should be 0.

𝒇(𝒙)=

{

0 𝑥<27

0.00355 ∗ 𝐱 − 0.07125 27 <𝑥≤36

0.0119 ∗ 𝐱 − 0.2872 𝑥>36

(1)

The model is divided in three intervals as the

variation of the variable to model is not continuous

but with abrupt changes as shown in the Figure 7.

With this model ready, the next measurements of

vibration will be compared with the ones provided by

the model and the difference among them

corresponds to the vibration due to the aspiration of

actual filler through the baghouse pipe.

Figure 6: Dataset used for the creation of the model.

Figure 7: Model aspiration values between 0 to 40 mm H

4 CAP REFERENCE

ARCHITECTURE

In process industries, due to harsh conditions the

system is operating in, some sensors might be

operating improperly (de-calibrated), or some

EDGE CLOUD

Cognitive Solutions in Process Industry: H2020 CAPRI Project

273

Figure 8: Cognition-driven process monitoring and control loop (cognitive plant) (ToBe).

parameters might be very deviating (instable) in a

period of time. On the other hand, the production

processes have to be under strict control ensuring

stability - otherwise some small issues might be

escalating very quickly.

Figure 8 (cf. grey boxes) shows basic elements of

a process monitoring and control loop, including the

models that can be used for supporting this task.

As explained in section 2, this problem requires a

better understanding of the underlying data and

processes, similarly how human cognition builds a

superior situational understanding and reasoning,

even in very ambiguous cases.

Therefore, the analogy of human cognition for

resolving above-mentioned challenges is used for an

efficient process control in process industry plants.

Since one of the most critical issues in

understanding/analysing process stability is to

observe variations, this artificial (or machine)

cognition should be based on a complex,

comprehensive but yet very efficient sensing,

analysing and understanding variations, including

their root causes, as well as their impacts (Wagner,

Milde, Barhebwa-Mushamuka, & Reinhart, 2022).

This is exactly how CAPRI envisions the

monitoring/sensing and controlling/reacting in

cognitive process plants (Zaeh, y otros, 2008).

In a cognitive plant, there is a need for monitoring

a broader context of the data that is collected and

processed in, as well as for a deep multivariate

analysis of the variation in data, to be able to detect

and react properly to unexpected events. The

realization of the cognitive plant is supported by

Cognitive components as depicted in Figure 8 (cf.

light-blue and blue coloured boxes).

The Cognitive capabilities and corresponding

Cognitive components are briefly illustrated:

Cognitive sensing enables getting accurate data

from sensors (IoT) or software sensors, even in the

cases when the sensing system is malfunctioning (e.g.

uncertainty, inconsistency, missing data). It will be

realized through Smart IoT Connection component,

which is responsible for establishing and maintaining

the connection to the production system.

Cognitive control enables reacting on various

situations of interest, even if the data is huge,

multivariate or changing (i.e. the process is instable).

It will be realized through Smart Event Processing

component, which is responsible for detecting

complex situations of interest in real-time data and

reacting correspondingly, e.g. in the context of

product/process quality control.

Cognitive operation supports the realization of

complex operations in a plant, even in the case of

unplanned delays or other types of deviations. It will

be supported by Smart Event Processing component,

applied on the intra-factory data streams, to guarantee

better performance and quality at organization level

(and nod only at multi-step or multi-stage levels).

Cognitive planning enables logistics, planning

and rescheduling capabilities in the inter-

organizational context. It will be supported by Smart

Decision Support component, which realizes

complex decision-making plans in the form of

heterogeneous processing pipelines (Salis,

Marguglio, De Luca, Gusmeroli, & Razzetti, 2022).

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

274

All Cognitive capabilities will be boosted by

Smart Knowledge Modelling component, which is

responsible for the overall technical modelling of the

plant and the aggregate interdependencies. It will be

basis for the Digital Twin (Rožanec, y otros, 2021),

as the collection of the digital assets (data, model,

services) belonging to the plant.

4.1 CAP Reference Architecture

The CAP Reference Architecture is structured for the

development of an advanced cognitive software

solution. As a digital enabler, it is an Open-Source

solution which is applicable to wide range of use

cases, supporting at the same time, a large variety of

applications. The design becomes ever harder in the

real industrial environment, for this reason, it was

done thanks to an iterative process started in the

report called D2.1, (Project Deliverables — Capri,

2022), where as a first step, there were a phase of

functional and non-functional requirements

collection followed by a continuous validations from

the pilots. The selected Reference Architecture

underlines the concept of edge and cloud cognitive

computing with the aim of solving business

challenges, creating new value from data and

improving the product quality.

The CAP Reference Architecture in CAPRI

project is designed with many horizontal layers able

to guarantee the interoperability, privacy, protection

and data sovereignty. In Figure 5 the core of the

architecture is depicted, since it contains the

brokering, the storage and the data processing

capabilities, including also cognitive process

analytics and simulation systems. Data in Motion,

Data at Rest and Situational Data are represented

using standard information models and made

available using standard APIs, (Salis, Marguglio, De

Luca, Gusmeroli, & Razzetti, 2022). The sensor layer

and the control layer use open-source technologies

from Apache (Livy, Spark, StreamPipes, Kafka) and

FIWARE (Draco, Cosmos, Orion Context Broker,

OPC UA Agent) foundations, (FIWARE - Open APIs

for Open Minds, 2022).

4.2 CAP Asphalt Use Case

The platform developed for the Asphalt domain is

comprised of the following modules:

Based on previous FIWARE Reference

Architecture, following the previous considerations,

the asphalt domain platform has been implemented in

a Linux Ubuntu distribution based server where the

different modules communicate and interact among

each other but deployed using the Docker platform

(Docker, 2022). The basic structure of this

architecture can be seen on Figure 9.

From the Asphalt Plant, real time data (with

sample times from 1 or 5 seconds, depending on the

data source) is received from a WAGO PLC

datalogger using MQTT protocol. This real time data

consists of production, event per batch of asphalt mix

and IoT data coming from other process sensors not

used for production control (e.g. weather station

data). In the asphalt domain CAP platform, a

mosquitto-based broker (Eclipse Mosquitto, 2022)

receives those data and it is redirected through an IoT

Agent for JSON (a bridge between HTTP/MQTT

messaging (with a JSON payload) and NGSIv2). This

IoT Agent has been customized to meet the asphalt

domain requirements of data flow. This IoT Agent

communicates and send the corresponding data to the

Orion Context Broker module (Generic Enabler that

provides the FIWARE NGSI v2 API, a Restful API

enabling to perform updates, queries, or subscribe to

changes on context information). This Broker is the

core of the whole FIWARE-based Reference

Architecture implemented in the Asphalt domain,

(FIWARE - Open APIs for Open Minds, 2022).

From this broker, a Draco (Fiware-Draco, 2022)

module has been set up, which is a Generic Enabler

that is a data persistence mechanism for managing the

history of context. It is based on Apache NiFi and is

a dataflow system based on the concepts of flow-

based programming. In this case, it manages the data

coming through the Orion Context Broker and sends

them to a MySQL database which is used for data

persistence within the CAP platform.

4.2.1 CAC1 Integration in CAP

CAC1 algorithm reads data coming from the plant

directly from the MySQL database, reading the last

data set received from the asphalt plant. The

calculated outputs are then sent to the Orion Context

Broker through a MATLAB S-function used at the

Simulink environment.

It is used a library function that, through https,

make a POST request to update the corresponding

entity via curl.

From this point and through the Broker, using the

Draco module, calculated outputs are written to the

CAC_Outputs table of the database.

From here, the corresponding visualization

module is used for the outputs to be shown at the

actual asphalt plant which is accessed through a web

interface (Figure 9).

Cognitive Solutions in Process Industry: H2020 CAPRI Project

275

Figure 9: CAC1 & CAS2 Integration in CAP Asphalt Use Case Platform.

The Visualization Module, based on Apache

Superset, is connected to the CAC1_Outputs table

stored in the used MySQL database, and the

corresponding burner power of the dryer drum and

the rotary speed are plotted on a time-series chart

alongside two numeric fields showing the last current

value.

It can be then accessed through a web-based

interface for the plant operators to see CAC1 output

data.

At this step, it is the plant managers/operators

responsibility to apply the displayed value or ignore

it based on their experience

4.2.2 CAS2 Integration

In the Asphalt domain, all output dataset coming from

CAS2 solution is sent using the MQTT protocol and

received at the previous CAP platform. This data is

stored in MySQL database with the structure shown

on Figure 9. The data is then stored in a table called

CAS2 with fields with self-explanatory names.

To make the measurements of CAS2 sensor

available for the process it is needed to store and

integrate the outputs of the sensor appropriately. The

file available at (Vega, Reñones, & Sanz, Cognitive

sensor for amount of filler [CAS2] - INTEGRATED,

2022), CAS2_sourcecode_2.sql contains the source

code which shows how to integrate the sensor

measurement into project’s CAP (cognitive

automation platform). The calculations directly

populate the CAS2 solution data persistence storage

at the MySQL database. The code uses a trigger

database, which is a procedural code that it is

automatically executed in response to certain events

on a particular table or view in a database.

In the case of CAS2, each time a new MQTT

output is sent to the CAP together with the pressure

measured, the trigger function ‘processStreamCAS2’

is fired and it applies the model estimated (section

3.4) and populates the CAS2 table.

The objective is that once all the asphalt CAPRI

solutions are running the results will be available for

different purposes like showing them on an interface

available to the plant operator or making them

available to other CS’s for further processing.

Providing continuous decision support, the plant

operator will not need to actively engage with the CS.

The warnings or alerts that have to be displayed on

the screen will be to increase or decrease the

depressure of the baghouse, to work with the best

magnitude of depressure, whose final objective is to

extract only the filler extracted not necessary and to

lose as less energy as possible. Before the deployment

of this CS, plant operator extracts nearly all the filler

after the process of drying in the drum, and the

necessary filler for the recipe of asphalt in the mixer

is added afterwards. This added filler is cold and leads

to an unavoidable loss of energy and raw materials.

CAS2 Visualization Module, based on Apache

Superset, is connected to the mentioned CAS2 table

stored in the MySQL database and then, through a

web-based interface, shown for the plant operators to

see CAS2 information.

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

276

5 CONCLUSIONS

H2020 CAPRI project develops and promotes digital

transformation through a CAP involving a Reference

Architecture (mainly based on the FIWARE

framework) with four levels of cognitive human-

machine interactions and a set of reference

implementations both commercial and open source.

This CAP coordinates a set of specific CS’s at the

various levels of functional organization of the

automation (from planning to sensors).

The asphalt domain shows as one of the main

process industry sectors where the CAP provides

flexibility of operation, improvement of performance

across different indicators (KPIs) and state of the art

quality control of its products and intermediate flows.

The CAP architecture and their different modules

have been presented in this domain and two of the

CS’s, CAS2, Sensors Layer Implementation and

CAC1, Control Layer Implementation, which are

under refinement, have been explained.

The open source architecture proposed based on

FIWARE represents a comprehensive and useful

platform to facilitate the integration of different

components that needs to interact with data coming in

real time from MQTT streams and needs to show their

results through an easy to sue webpage.

From here, next steps involve the final integration

of the rest of the “layers” of the reference architecture

and the final validation to be developed at last project

stages, addressing manufacturing challenges in

industrial operational environments of the three

chosen process sectors, and providing useful

feedbacks and lessons learnt.

Different KPI’s will be calculated and deployed to

see if initial target objectives are met with an

evaluation period (6-month minimum) of the

performance improvements thanks to the different

implemented CS’s. This will provide effective stories

for replication purposes and dissemination. It is

expected that results like the reference architecture

will be replicated in other sectors with similar

challenges from the point of view of CS’s applied to

similar unitary processes.

ACKNOWLEDGEMENTS

CAPRI project receives funding in the European

Commission’s Horizon 2020 Research Programme

under Grant Agreement Number 870062.

The authors would like to thank their colleagues

from EIFFAGE, ENGINEERING, NISSATECH and

CARTIF partners of the project for contributing with

some of the examples shown in the paper.

REFERENCES

Abburu, S., Berre, A., Jacoby, M., Roman, D., Stojanovic,

L., & Stojanovic, N. (2020, 4). Cognitive Digital Twins

for the Process Industry. 68-73.

Abburu, S., Berre, A., Jacoby, M., Roman, D., Stojanovic,

L., & Stojanovic, N. (2020, 6). COGNITWIN - Hybrid

and Cognitive Digital Twins for the Process Industry.

Proceedings - 2020 IEEE International Conference on

Engineering, Technology and Innovation, ICE/ITMC

2020.

Andersson, T. (2010). Estimating particle size distributions

based on machine vision. Lulea, Sweden:

Universitetstryckeriet.

Auditors, E. (2021). Digitising European Industry : an

ambitious initiative whose success depends on the

continued commitment of the EU, governments and

businesses. Special report No 19, 2020. Publications

Office.

Cinar, Z., Nuhu, A., Zeeshan, Q., & Korhan, O. (2020).

Digital Twins for Industry 4.0: A Review. 193-203.

Consortium, C. (2022, 7 29). Cognitive Automation

Platform for European PRocess Industry digital

transformation. Retrieved from Cognitive Automation

Platform for European PRocess Industry digital

transformation: https://www.capri-project.com/

Docker. (2022, 07 28). Retrieved from

https://www.docker.com/

Eclipse Mosquitto. (2022, 07 28). Retrieved from

https://mosquitto.org/

Eirinakis, P., Lounis, S., Plitsos, S., Arampatzis, G.,

Kalaboukas, K., Kenda, K., Stojanovic, N. (2022, 1).

Cognitive Digital Twins for Resilience in Production:

A Conceptual Framework. Information, 13(1), 33.

Retrieved from https://www.mdpi.com/2078-

2489/13/1/33

El-Haggar, S. (2007). Sustainability of Industrial Waste

Management. Elsevier.

ElMaraghy, H., & ElMaraghy, W. (2022, 5). Adaptive

Cognitive Manufacturing System (ACMS) – a new

paradigm. International Journal of Production

Research, 1-14.

Essa, E., Hossain, M., Tolba, A., Raafat, H., Elmogy, S., &

Muahmmad, G. (2020, 5). Toward cognitive support for

automated defect detection. Neural Computing and

Applications, 32(9), 4325-4333.

Fernández, R., Viennet, E., Goles, E., Barrientos, R., &

Telias, M. (1998, 1). On-Line Coarse Ore

Granulometric Analyzer Using Neural Networks.

Industrial Applications of Neural Networks, 59-67.

FIWARE - Open APIs for Open Minds. (2022, 07 22).

Retrieved from https://www.fiware.org/

Fiware-Draco. (2022, 07 28). Retrieved from

https://fiware-draco.readthedocs.io/en/latest/

Cognitive Solutions in Process Industry: H2020 CAPRI Project

277

Gómez, D., & Diego, M. (2022, 7 20). Cognitive control of

Asphalt Drum [CAC1] - integrated.

doi:https://zenodo.org/record/6867477

Haber, R., Juanes, C., Del Toro, R., & Beruvides, G. (2015,

5). Artificial cognitive control with self-x capabilities:

A case study of a micro-manufacturing process.

Computers in Industry, 74, 135-150.

Jacoby, M., Jovicic, B., Stojanovic, L., & Stojanović, N.

(2021, 5). An Approach for Realizing Hybrid Digital

Twins Using Asset Administration Shells and Apache

StreamPipes. Information, 12(6), 217. Retrieved from

https://www.mdpi.com/2078-2489/12/6/217

Kaur, N., & Sood, S. (2015, 12). Cognitive decision making

in smart industry. Computers in Industry, 74, 151-161.

Koskinen, J. (1998). Fuzzy modelling of a pilot plant rotary

dryer. UKACC International Conference on Control

(CONTROL '98), 515-518.

Le Guen, L., Huchet, F., & Tamagny, P. (2011, 6). Drying

and Heating Modelling of Granular Flow : Application

to the Mix-Asphalt Processes. Journal of Applied Fluid

Mechanics, 4(2), pp 71-80. Retrieved from

https://hal.archives-ouvertes.fr/hal-00612503

Li, M.-L., Yao, Y.-Y., & Zhao, H.-X. (2017). Numerical

Analysis of Solid Materials Transport in a Rotary

Dryer. Proceedings of the 3rd Annual International

Conference on Mechanics and Mechanical Engineering

(MME 2016).

Ljung, L. (1998). System Identification. Retrieved from

http://link.springer.com/10.1007/978-1-4612-1768-

8_11

Mahmoud, K., El-Kasassy, M., & Areed, F. (2020, 11).

Variable Structure Controller for a Rotary Dryer

(Dept.E). MEJ. Mansoura Engineering Journal, 36(2),

13-21.

Measuring Coal Particles in the Pipe. (2022, 07 21).

Retrieved from https://www.power-eng.com/emissi

ons/measuring-coal-particles-in-the-pipe/

Oliveira, J., Van Der Geld, C., & Kuerten, J. (2017, 7).

Concentration and velocity statistics of inertial particles

in upward and downward pipe flow. Journal of Fluid

Mechanics, 822, 640-663.

Pang, S., Jia, J., Ding, X., Yu, S., & Liu, Y. (2021, 9).

Intelligent Control in the Application of a Rotary Dryer

for Reduction in the Over-Drying of Cut Tobacco.

Applied Sciences, 11(17), 8205.

Project Deliverables — Capri. (2022, 07 19). Retrieved from

https://www.capri-project.com/project-deliverables

Raghavan, V., Jumah, R., & Mujumdar, A. (2006, 11).

Control of Industrial Dryers. Handbook of Industrial

Drying, Third Edition.

Rožanec, J., Lu, J., Rupnik, J., Škrjanc, M., Mladenić, D.,

Fortuna, B., Kiritsis, D. (2021, 3). Actionable Cognitive

Twins for Decision Making in Manufacturing.

Retrieved from http://arxiv.org/abs/2103.12854

Rubio, F., Bordons, C., Holgado, J., & Rivas, I. (2001, 4).

Modelling and H∞ control of a rotary dryer. 2001

European Control Conference, ECC 2001, 33(4), 733-

738.

Salis, A., Marguglio, A., De Luca, G., Gusmeroli, S., &

Razzetti, S. (2022, 2). An Edge-Cloud based Reference

Architecture to support cognitive solutions in the

Process Industry.

Sánchez Boza, A., Guerra, R., & Gajate, A. (2011, 3).

Artificial cognitive control system based on the shared

circuits model of sociocognitive capacities. A first

approach. Engineering Applications of Artificial

Intelligence, 24(2), 209-219.

Schwenzer, M., Ay, M., Bergs, T., & Abel, D. (2021, 11).

Review on model predictive control: an engineering

perspective.

Sharma, A., Kosasih, E., Zhang, J., Brintrup, A., &

Calinescu, A. (2020, 11). Digital Twins: State of the Art

Theory and Practice, Challenges, and Open Research

Questions. Retrieved from http://arxiv.org/abs/

2011.02833

Singh, R., Lye, S., & Miao, J. (2019, 8). PVDF Nanofiber

Sensor for Vibration Measurement in a String. Sensors,

19(17), 3739. Retrieved from https://www.mdpi.com/

1424-8220/19/17/3739

Sivilevičius, H., & Šukevičius, Š. (2009, 6). Manufacturing

technologies and dynamics of hot-mix asphalt mixture

production.

Suppan, T., Neumayer, M., Bretterklieber, T., Puttinger, S.,

& Wegleiter, H. (2022, 1). A Model-Based Analysis of

Capacitive Flow Metering for Pneumatic Conveying

Systems: A Comparison between Calibration-Based

and Tomographic Approaches. Sensors, 22(3), 856.

Retrieved from https://www.mdpi.com/1424-

8220/22/3/856

Vega, C., & Reñones, A. (2022, 3 18). Cognitive sensor for

amount of filler. doi:https://doi.org/10.5281/zenodo.

6367575

Vega, C., Reñones, A., & Sanz, L. (2022, 7 28). Cognitive

sensor for amount of filler [CAS2] - INTEGRATED.

Retrieved from https://zenodo.org/record/6922993

Wagner, S., Milde, M., Barhebwa-Mushamuka, F., &

Reinhart, G. (2022). Digital Twin Design in Production.

Lecture Notes in Mechanical Engineering, 339-346.

Yliniemi, L., Koskinen, J., & Leiviskä, K. (1998, 9).

Advanced Control of a Rotary Dryer. IFAC

Proceedings Volumes, 31(23), 119-124.

Zaeh, M., Beetz, M., Shea, K., Reinhart, G., Bender, K.,

Lau, C., . . . Herle, S. (2008, 11). The Cognitive Factory.

Zhang, K., Huchet, F., & Hobbs, A. (2019, 5). A review of

thermal processes in the production and their influences

on performance of asphalt mixtures with reclaimed

asphalt pavement (RAP). Construction and Building

Materials, 206, 609-619.

ETCIIM 2022 - International Workshop on Emerging Trends and Case-Studies in Industry 4.0 and Intelligent Manufacturing

278