Analysis of Thermographic Patterns using Open CV

Case Study: A Clinker Kiln

Villie Morocho

, Eliezer Colina-Morles

, Sebastian Bautista

, Alfredo Mora

and Mara Falconí

Departamento de Ciencias de la Computación, Universidad de Cuenca, Cuenca, Ecuador

Departamento de Ingeniería Eléctrica, Electrónica y Telecomunicaciones, Universidad de Cuenca, Cuenca, Ecuador

Departamento de Investigación y Desarrollo, Industria Cementera Nacional, Azogues, Ecuador

Keywords: Cement Kiln, Clinker, Pattern Recognition, Two Bag-of-Words Classifiers.

Abstract: The core of the cement production process is the clinker kiln. Proper operation of the kiln depends on

factors such as the timely monitoring of its thermal behavior under different operation conditions. This work

includes a systematization of empirical knowledge of skilled kiln operators, linking it with the analysis of

thermography images of the kiln using Open CV. The paper includes an integration of interventions

implemented by the operators, in terms of a log described in natural language. The work highlights potential

uses of the knowledge of experienced operators, when this is combined with techniques based on image

analysis and artificial intelligence.

1 INTRODUCTION

A cement factory consists of several areas of

different processes that influence the final cost of

production. The area between the preheaters section

to the cooler, which includes the clinker kiln, is

undoubtedly the most important among the cement

production processes (Deolalkar, 2009). Figure 1

illustrates the elements of that area. The control

actions performed on the kiln have much influence

on productivity levels and the quality characteristics

of cement. Therefore, it is vital to maintain

appropriate operating conditions, including real-time

monitoring of thermal behavior of the walls of the

kiln. In many cement installations, the operators that

control the kiln typically decide their actions on

empirical grounds, that have improved based on

experience or verbal knowledge transfer.

This paper contains a proposal that allows

systematizing empirical actions taken by operators,

based upon graphical information of the thermal

reading of the kiln walls obtained from a

thermograph. The data capture is done via Open CV

(Open CV, n.d.), which also facilitates the analysis

of its graphic content. Also, the relationship between

thermal images and description of the actions taken

by the operators is done in natural language, in a log.

In addition, a validation process through a decoding

algorithm and extraction of information stored in the

thermograph is performed using Open CV. This

allows generalizing the methodology to existing

industrial thermographs.

Figure 1: Illustration of the area between the preheaters to

the cooler.

The analyzed information facilitates the search

for patterns of operation of the kiln that permits

establishing rules of inference, relating process

variables, thermographic patterns and the log. This

result represents a step forward to build a rule-based

479

Morocho V., Colina E., Bautista S., Mora A. and Falconi M..

Analysis of Thermographic Patterns using Open CV - Case Study: A Clinker Kiln.

DOI: 10.5220/0005535304790484

In Proceedings of the 12th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2015), pages 479-484

ISBN: 978-989-758-123-6

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

decision support system, with online reading of the

thermography images and the values of the process

variables, to enhance the operation of the kiln.

2 EXISTENCE OF

UNSTRUCTURED EMPIRICAL

KNOWLEDGE

The gap between experience and phenomenological

understanding of the processes make the control

actions taken in the kiln "reactive", based on a mind

map of past experiences. On this map there may be

“holes” (ie. Tendency to not quite remember the

entire context where the new experience was

developed) and a systematization of knowledge is

therefore paramount to support decisions about the

operation of the kiln. For example, in a study

conducted on the considered plant, a Piping and

Instrumentation Diagram (P&ID) of the considered

area was built. This diagram included seventy-seven

variables in the preheater and oven and eighty

variables in the cooler. Figure 2 illustrates the part of

the P&ID that corresponds to the clinker kiln. In a

given operational condition, the operator only

considers a small map of variables; empirically

determining those affecting in higher degree the

action to be taken. In this way, a system for

establishing relationships among variables of the

process with control actions would greatly help in

the kiln operation.

Figure 2: Clinker kiln P&ID.

3 SYSTEMATIZED

THERMOGRAPHIC MAPS AS A

SOURCE OF KNOWLEDGE

Figure 3 corresponds to a thermal image of the kiln

over its 58 metres in length and constitutes an

important source of information. The thermography

image reads temperatures in a range that varies

between 80° Celsius to 500° Celsius, around the

360° of rotation of the kiln.

Empirically, it has been determined that an area

with temperatures of 400° Celsius or higher for a

long period of time would have higher risk of

damage to the refractory and might cause a forced

stop of the kiln.

Capturing thermographic information with

OpenCV

Open CV (Open CV, n.d.) is software under the

BSD license, which allows free use for commercial

and academic purposes.

Figure 3: Thermographic image of the clinker kiln.

Usually, a thermography is presented from a

propietary system as an image (see Figure 3).

Thermographic systems store information in a

confidential manner, i.e. they use algorithms of

condensation, encryption, and others, which do not

allow the extraction of information for direct

analysis of their images. In addition, being

proprietary systems, any possibility of integration

with their databases can lead to malfunctioning of

the system or could even be subjects to security

risks.

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

480

In this proposal, regardless of the themographic

system implanted in plant, the use of an interface

under Robot (Oracle, n.d.) allows to capture the

thermography image and to store it in an array of

pixels which is subsequently treated under Open

CV.

Through Open CV (.Mat, .Highgui) the

information in the array of pixels is extracted and

presented using SWT (ECLIPSE, n.d.) (Pulli, et al.,

2012) (see Figure 4). Next, the necessary

information is stored, so that it can be used for

image analysis.

The post-processing allows locating longitude,

latitude and temperature on any point of the

extracted image; which determines values that will

be used to search for patterns that describe thermal

behaviors. The position of the matched pattern is

stored in a database whose design allows for fast

data recovery. It uses a standard indexation search

method and facilitates the identification of refractory

zones which may have been damaged for having

been exposed to high temperature for large periods

of time. The positions patterns storage allows for a

semi-automated post-processing capability.

Figure 4: Image extracted from the Thermographic system

and processed with Robot, Open CV and SWT.

3.1 Patterns Matching using Open CV

with Bag of Words Approach

(BoW)

The Probabilistic Latent Semantic Analysis (PLSA)

(Fergus, n.d.) and the Dirichlet Allocation Latent

Technique (Hofmann, 1999) used for text analysis

were introduced in the visual domain methods (Blei

& Jordan, 2003) (Fei-Fei & Perona, 2005) (Sivic, et

al., 2005). In the code developed under Open CV

there are implementations of PLSA, including all

stages of pre-processing.

There are two possibilities to search for patterns

that identify areas of high temperatures: use a loop

that examines all the pixels in the images, or use a

BoW approach.

By using a BoW approach it is possible to choose

a pattern on one of the images of interest and search

its repetition in the bank of images, obtained using

an automated image capture.

The studied images, which can be currently

obtained from the thermograph, are 350x330 pixels.

However, the dimensions of the refractory areas

which may be affected have a minimum length of 10

cm; as a result, image analysis should be based in

affected regions of that size.

A disadvantage of using the BoW approach in

Open CV is that coding is performed under

MatLab®, whereas in the proposal free software is

preferred.

3.2 Open CV Template Matching

The study of options for pattern analysis led to try

other search alternatives. Under Open CV there is a

Match Template function, with various methods of

searching for similarities. The aim of these methods

is to find a piece of specific image (Template Image)

within an image source (Source Image). In the case

of thermography images, this allows to determine

the pattern of high temperatures, within the full

image on the thermography of the oven.

The methods integrated into Open CV for

searching patterns based on the Match Template

function are:

method= CV_TM_SQDIFF

method=CV_TM_SQDIFF_NORMED

method=CV_TM_CCORR

method=CV_TM_CCORR_NORMED

method=CV_TM_CCOEFF

method=CV_TM_CCOEFF_NORMED

Each method is based on a mathematical

formulation, which can be found in the

documentation of Open CV (Open CV, n.d.), which

allows formulating an algorithm for determining the

most appropriate results, according to the

characteristics of the desired images. In the case of

thermography images, specific features such as its

imprecise contours should be considered.

The tests were conducted on thermographic

images obtained a day before a forced stop of the

kiln, in a period of two years. Different methods of

patterns search were used.

The methods implemented of the “matching”

function, considered the template on the image

search and tried to find the most similar. A

AnalysisofThermographicPatternsusingOpenCV-CaseStudy:AClinkerKiln

481

disadvantage of using these methods in

thermographic images is that hot spots showing the

refractory wear may have different shapes. In spite

of this, the necessary tests were carried out and

obtained the following results: Figure 5 represents

the source image, taken one day before a forced stop

of the kiln. On the source image may be appreciated

a small yellow area, which has been extended in

Figure 6, and corresponds to the template image.

This image symbolizes an area of high temperature

on the wall of the oven.

Figure 5: Source image.

Figure 6: Template image.

After using the matching methods, the following

results were obtained: Figure 7 corresponds to the

result of applying the method CV_TM-SQDIFF,

which provided a wrong pattern (upper left corner).

None result were obtained by applying the method

CV_TM_SQDIFF_NORMED, as it is illustrated in

Figure 8. Figure 9 illustrates the wrong result from

applying matching method CV_TM_CCORR, as

seen in the upper left corner.

Figure 7: CV_TM-SQDIFF matching method result (upper

left corner).

Figure 8: CV_TM_SQDIFF_NORMED matching method

result (none pattern found).

Figure 9: CV_TM_CCORR matching method result

(upper left corner).

Figure 10 depicts a set of wrong results after

using the matching method CV_TM_CCOEFF.

Finally, the right pattern was found applying the

matching method CV_TM_CCOEFF_NORMED, as

it is presented in figure 11.

Figure 10: CV_TM_CCOEFF matching method result (a

set of wrong patterns are found).

The previous results were checked with other

sample data and allow to suggest that, given the

particular characteristics of the thermographic

images, the matching method

CV_TM_CCOEFF_NORMED is the most

appropriate to distinguish thermal patterns

associated with high temperatures on the walls of the

kiln.

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

482

Figure 11: CV_TM_CCOEFF_NORMED matching

method result (the right pattern is found).

4 ANALISING NATURAL

LANGUAGE TO RELATE TO

PATTERNS

While kiln operators have a log to describe

operational conditions, manually every hour, it is

required an analysis of this description to determine

its suitability. Beyond that, log analysis is important

to search for command patterns, taken by operators

according to their criteria and personal expertise, to

meet operational deviations. With this analysis it is

possible to formulate decision rules, which were

validated with elicitation sessions with operators.

This knowledge base and support from an

artificial intelligence tool, allow determining which

actions must be implemented based on conditions of

the refractory, displayed on the thermography; for

instance switch on a particular external air quencher

fan or reduce the kiln angular velocity. Additional

information to be used is the semi-automated

historical analysis of the thermography. In order to

infer from the log written in natural language, this

work is testing MYCIN (Buchanan & Shortliffe,

1984).

5 VALIDATION OF RESULTS

To validate the information obtained under the Open

CV application described in this article, a code was

developed to extract information from the

thermograph in an automated manner. The thermal

information storage uses some "compression and

encoding" techniques, especially to prevent the

growth of the database in a disproportionate way. In

the case studied here, the thermograph stores

information in a hexadecimal format. Each

hexadecimal data represents the temperature at a

point on the kiln, longitude and latitude (x, y). To

find the correlation between the stored value and the

represented decimal value, a hexadecimal to decimal

transformation algorithm, considering four cases of

data compression, was designed. These cases allow

saving space in the string that stores the

thermography. Temperatures are expressed in

degrees Celsius and to save space, normally the data

is stored divided by 10. The algorithm varies with the

used thermograph, but once decoded it was possible

to connect various applications such as the one

developed in this project. When the result of the

transformation was obtained, it was possible to

search for information directly and compare it with

the images read by Open CV.

6 CONCLUSIONS

In this paper, empirical knowledge extraction from

operators of a clinker kiln has been addressed, from

computational perspectives. Firstly, a link to a set of

thermal image readings was built, using a non-

invasive approach in an industrial case of study. It

has been possible to capture graphic information that

has been processed and analyzed by a coupled

system, without computational interference.

It has been built an application to establish

relationship between decisions made by operators

and thermographic images, corresponding to the

operating conditions under which these decisions are

implemented.

The work has used images classification

techniques based on a classification of texts

approach. In particular, the BoW approach has been

proved with satisfactory results.

ACKNOWLEDGMENTS

The authors wish to express their thanks for the

support obtained from the National Cement

Company Union, plant Guapan, Azogues, the

Research Office of the University of Cuenca and the

Prometeo Project of the Secretariat of Higher

Education, Science, Technology and Innovation of

Ecuador.

REFERENCES

Blei, D. & Jordan, M., 2003. Latent Dirichlet Allocation.

Journal of Machine Learnign Research, January, Issue

3, pp. 993-1022.

AnalysisofThermographicPatternsusingOpenCV-CaseStudy:AClinkerKiln

483

Buchanan, B. G. & Shortliffe, E. H., 1984. Rule-Based

Expert Systems: The Mycin Experiments of the

Stanford Heuristic Programming Project. s.l.:MA:

Addison-Wesley Publishing Co., Inc..

Deolalkar, S. P., 2009. Handbook for desinging cement

plants. BS Publications, pp. 73-85.

ECLIPSE, n.d. SWT library. [Online] Available at:

http://www.eclipse.org/swt/ [Accessed November

2014].

Fei-Fei, L. & Perona, P., 2005. A Bayesian Heirarchical

Model for Learning Natural Scene Categories. s.l., s.n.

Fergus, R., n.d. Two bag-of-words classifiers. [Online]

[Accessed September 2014].

Hofmann, T., 1999. Probabilistic Latent Semantic

Analysis, s.l.: UAI.

Open CV, n.d. [Online] Available at: http://opencv.org

[Accessed February 2015].

Oracle, n.d. Robot, JAVA Library. [Online] Available at:

http://docs.oracle.com/javase/7/docs/api/java/awt/Rob

ot.html [Accessed November 2014].

Pulli, K., Baksheev, A., Kornowakov, K. & Eruhimov, V.,

2012. Realtime Computer Vision with OpenCV.

ACMQueue, April, 10(4), p. 40.

Sivic, J. et al., 2005. Discovering objerct categories in

image collection. Beijing, s.n.

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

484