Professional Documents
Culture Documents
2, 2016 97
Seshadhri Srinivasan
Kalasalingam University,
Krishnankoil, India
Email: seshucontrol@gmail.com
Guruprasath Muralidharan
FLSmidth Pvt. Ltd.,
Chennai, India
Email: guruprasath.muralidharan@flsmidth.com
Keywords: cement grinding process; system identification; state space model; prediction error
method; PEM; subspace method.
Reference to this paper should be made as follows: Sivanandam, V., Kannan, R., Srinivasan, S.
and Muralidharan, G. (2016) Comparison of subspace and prediction error methods of system
identification for cement grinding process, Int. J. Simulation and Process Modelling, Vol. 11,
No. 2, pp.97107.
Ramkumar Kannan received his PhD in Process Control from SASTRA University. He is
working in the Department of Electronics and Instrumentation Engineering, SASTRA University,
Thanjavur, India over 18 years. He has successfully completed a funded project from DRDO.
Seshadhri Srinivasan received his PhD from National Institute of Technology, Tiruchirappalli in
2010. He worked as an Associate Scientist in ABB Indian Corporate Research Center at
Banaglore, India, and Sr. Scientist in Center for Excellence in Nonlinear Systems at Institute of
Cybernetics, Tallinn University of Technology, Tallinn, Estonia. He worked as a Forscher with
Technical University of Munchen, Germany and currently leads the international research efforts
of the Kalasalingam University. He has worked in various projects of significance such as
European Union Regional Development Fund and Target Project SF 140018s0 on complex
control systems, I3RES (FP7 project) and e-Gotham (Arteimus) in Smart grids. EUCLID has
identified him as a resource person in industrial networked control systems. He was also the
recipient of MHRD fellowship for pursuing research from 20072010.
Guruprasath Muralidharan received his PhD in the field of process control from National
Institute of Technology, Tiruchirapalli, India. He has extensive experience in the automation of
cement plants over 12 years. He is currently working for FLSmidth Pvt. Ltd, Cement and
Minerals Projects India, Chennai, India.
parameters directly with the physical variables. Further, use investigation are most suited for designing predictive
a firm mathematical framework wherein the sensitivity of controllers for cement grinding circuit in the presence of
the accuracy to the designer choice is eliminated. feed-grindability variations.
Main contributions of this investigation are:
Table 1 Advantages and disadvantages of different modelling
techniques 1 two data-driven state space models for the cement
grinding process based on prediction error and
Approach Advantages Dis-advantages subspace methods that use real-time data from cement
First Based on well-defined Complex to develop. grinding circuit to map feed-grindability variations with
principle physical laws. Requires multiple product quality
models Mathematical domain expertise and is
formulation that time-consuming.
2 analyse different candidate models for both the state
directly relates physical space models and selecting the most accurate model
Relating
variables with model feed-grindability 3 model validation using real-time industrial data.
parameters. variations with product
quality is difficult.
Data Simple and easy to Direct correlation 2 Process description
driven develop. between physical
model Relate feed-grindability variable and model The schematic of the cement grinding process used in this
variations with product parameter is not study is shown in Figure 1. It consists of ball mill, bucket
quality. possible.
elevator, and separator. The feed to the process is the
Model uncertainties are Accuracy depends on clinkers along with additives such as gypsum and fly ash to
adopted. the model data and
expertise of the improve the chemical properties of the cement that are
designer. stored in hoppers. Clinkers are introduced into the ball mill
grinding process through weigh feeders. The clinkers are
To reach the objectives, this investigation proposes two ground by the impact of steel balls to obtain a powdered
modelling techniques: material that are taken by the bucket elevator to the
separator. The classifying separator separates the fine and
1 prediction error method (PEM)
coarse particles into two streams based on centrifugal force.
2 subspace identification technique. The final cement product that satisfy the desired fineness
(Blaine) get suspended in the air stream and are therefore
Both these data-driven approaches are simple and the
sucked by the separator fan. The coarse and semi-ground
models developed directly correlate the parameters and
particles that are not satisfying the desired fineness are
physical variables. The identified models were validated
collected at the bottom of the separator and re-circulated
with several techniques as discussed in Tavakolpour-Saleh
into the ball mill for further grinding. Higher the separator
et al. (1997, 2015). As a result, these approaches combine
speed better the product quality.
the advantages of the first-principle and data-driven models
and map the feed-grindability variations to the product
quality and output. Therefore, models developed in this
Figure 1 Schematic of cement grinding process (see online version for colours)
100 V. Sivanandam et al.
The feed with grindability variations entering the ball mill, with a sampling time of 1minute. The data collected has
affects the mill load (material inside the mill) and product outliers and is corrupted by noise. Data pre-processing
quality in the following two ways assuming that the techniques such sigma testing and average filtering are used
separator speed is maintained at is maximum value: to eliminate the outliers and bad-data. The processed data is
then normalised (see Figure 2) to simplify the computations.
1 low grindability feed causes increase in mill load which
In the modelling step, the normalised data is divided
may lead to the plugging phenomenon (obstruction of
into two parts: one set of data is used for fitting the model,
mill) and also affect the product quality because the
whereas the second set of data is used in validation.
semi-ground particles are collected at the outlet stream
2 high grindability feed reduces the mill load that leads to 3.1 Prediction error method
wear and tear of the balls and fine particles enter into
the reject stream. This premise of the method is to minimise the error between
the predicted and observed values of the output. The basic
To guarantee product quality in the presence of feed idea behind prediction error approach is, first describe the
grindability variations requires that the feed flow rate and model as a predictor of the next output, and then
separator fan power to be varied leading to improved parameterise the predictor. The predictor thus obtained is
product quality, productivity, and reduced plugging and used to predict the next output and the parameters are
wearing of the balls in the mill. corrected based on the predicted error using a suitable norm.
Currently, cement industries lack a measurement This method is classified under the broad category of
methodology for product quality directly. Therefore, the parameter estimation methods. It has close association with
elevator current and main drive load can be used as the maximum likelyhood approach. The predictor for this
controlled variable, whereas feed-flow rate and separator model is represented by
power are the manipulating variables. Thus, the model that
relates the manipulating and controlled variables will (
y m (k | k 1) = f Z k 1 ) (1)
capture the effect of feed grindability variations on cement
product quality. where y m (k | k 1) is the one step ahead predictor, f is the
function denoting past observed data, k is the sampling
Figure 2 Real time data acquired from cement grinding process
instant and Z is the input and output data observed from the
(see online version for colours)
cement grinding process. The predictor is represented in
20
terms of parametric vector as follows
current (A)
Elevator
( )
0
y(k | ) = f Z k 1 , (2)
-20
0 200 400 600 800 1000 1200 1400 1600 1800 2000
The parameterisation of the model is obtained from the
50
differential equation as follows
Main drive
load (kW)
0
y (k ) + a1 y (k 1) + " + an y (k n)
(3)
-50
0 200 400 600 800 1000 1200 1400 1600 1800 2000
= b1u (k 1) + " + bm u (k m)
20 where y and u are the outputs and inputs, n and m are the
rate (TPH)
0
predictor of equation (3) becomes
-20
0 200 400 600 800 1000 1200 1400 1600 1800 2000
y(k | ) = a1 y (k 1) " an y ( k n) + b1u (k 1)
(4)
20 + " + bm u (k m)
power (kW)
Separtor
0
where = [a1 ... an b1 ... bm ]T . The observed data of the
-20
0 200 400 600 800 1000 1200 1400 1600 1800 2000 cement grinding process is expressed as
Time in minutes
( )
f Z k 1 , = T (k ). (5)
y ( k ) = G (q, )u (k ) + H (q, )e(k ). (7) In the first step, the data is projected into the row space to
find the observability matrix and state estimates. The second
where e(k) is the unpredictable white noise, G is the transfer step involves finding the system matrices A, B, C and D
function of a linear time invariant system of finite order from observability matrix and state estimates. Several
(LTFID) and H is a stable, monic and minimum algorithms are there for implementing the subspace
phase transfer function corresponding to a noise model. identification techniques. All the subspace identification
Equation (7) is written as procedure start with the following input and output matrix
equation as discussed in Favoreel (2000).
y(k ) = I H 1 (q, ) y (k )
(8) y f = Ti X i + H d u f + H s M f + N f (12)
+ H 1 (q, )G (q, )u (k ) + e(k )
where the observability matrix Ti is given by
The prediction error is the difference between the predicted
output by the model and the actual measurement of the C
process. Ti = ,
CA
The prediction error can be written as
the lower triangular matrix Hd is defined as
e(k ) = y (k ) y ( k ) = H 1 [ y ( k ) Gu (k ) ] (9)
D 0
where y(k) and y(k ) denote the actual measurement and the CB D ,
output predicted by the model respectively and u(k) is the
input. The identification criterion has to be chosen so that a and Hs is the stochastic triangular matrix defined as
best model fit can be obtained. Generally least square
0 0
criterion will be used. The cost function of least square C 0 .
criterion can be obtained directly
1 The input Hankel matrix is
N
min VN ( ) = e( k , ) 2 (10)
N k 1
u0 u1
u u2
where is parametric vector and it is found as discussed in 1
Ljung (1999) and N is the length of data.
The state space model using this parameters is y y1
and the output matrix is 0
written as y1 y2
x(k + 1) = A( ) x(k ) + B( )u (k ) yf yfuture is the predicted output
x(0) = x0 ( ) (11)
Xi is the estimate of state vector
y (k ) = C ( ) x(k ) + D( )u (k )
u is the input.
The system matrices of equation (11) A, B, C and D are
Mf and Nf are process and measurement noise, respectively.
parameterised and found using canonical form
The matrix with input up and output yp is known as Wp.
representation.
That is
State estimate
102 V. Sivanandam et al.
order to check the dependency, correlation analysis is being representing the cement grinding process with guaranteed
performed. accuracy of 95%.
First the auto correlation among the residuals is tested. Among the winner models SSP8 from the PEM and the
All the correlation functions except at lag 0 are lying within SSS4 from the subspace method, the best model which suits
the confidence region of 95% as shown in Figure 4. It is the cement grinding process that relates the feed grindability
obvious that the correlation function will hit 1 at lag 0 variations with product quality has to be chosen. For this
because the dependency of the signal at the same time analysis, the testing dataset is introduced as input for the
samples is maximum. The cross correlation function winner models SSP8 and SSS4. Figure 10 shows the
between the residuals of the output (difference between the comparison of measured outputs with that of prediction
observed output and model output) and the inputs are error-based state space model SSP8 where as Figure 11
analysed in order to check the dependency among them. shows the comparison between the measured and model
Figure 5 shows the cross correlation function of output of subspace state space model. A close look of these
residuals of the output elevator current with the inputs feed Figures 10 and 11 reveal that the model output
flow rate, separator power. corresponding to SSS4 (state space model found using
Figure 6 shows the cross correlation function of subspace method) is tracking the observed (measured)
residuals of the main drive load with the inputs. A close output. Hence the model SSS4 found with the technique of
look at Figures 5 and 6 show that the correlation functions subspace method is a good representative for the cement
lie within the 95% confidence region, which means that the grinding process under feed grindability variations.
winner state space model SSP8, identified using PEM is
Figure 4 Auto correlation function among the residuals for both the outputs for the selected candidate model
(see online version for colours)
For elevator current
0.2
0.1
correlationfunction
-0.1
-0.2
0 100 200 300 400 500 600 700 800 900 1000
lag
0.1
correlationfunction
-0.1
-0.2
0 100 200 300 400 500 600 700 800 900 1000
lag
Figure 5 Cross correlation function between residuals of the first output elevator current with two inputs viz. feed flow rate and separator
power (see online version for colours)
Between feed flow rate and elevator current
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
104 V. Sivanandam et al.
Figure 6 Cross correlation function between residuals of the first output main drive load with two inputs viz. feed flow rate and separator
power (see online version for colours)
correlation function between the feed flow rate and residuals from the elevator current
0.1
confidence region
correlation function
correlationfunction 0.05
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
crorrelation function between the separaton power and resdiuals from main drive load
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
Figure 7 Correlation function among the residuals of the winner model for both outputs elevator current and main drive load
(see online version for colours)
autocorrelation of residuals for elevator current
0.2
confidence region
correlation function
0.1
correlationfunction
-0.1
-0.2
0 100 200 300 400 500 600 700 800 900 1000
lag
-0.1
-0.2
0 100 200 300 400 500 600 700 800 900 1000
lag
Figure 8 Correlation function between the output elevator current and two inputs viz. feed flow rate and separator power for the winner
model (see online version for colours)
correlation between feed flow rate and residuals from elevator current
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
correlation function between separator power and residuals from elevator current
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
Comparison of subspace and prediction error methods of system identification for cement grinding process 105
Figure 9 Correlation function between the output main drive load and two inputs viz. feed flow rate and separator power for the winner
model (see online version for colours)
correlation function between the residuals of output main drive load and input feed flow rate
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
correlation function between the residuals of output main drive load and input separator power
0.1
confidence region
correlation function
0.05
correlationfunction
-0.05
-0.1
0 100 200 300 400 500 600 700 800 900 1000
lag
Figure 10 Comparison of model output and observed output for the winner SSP8 model (see online version for colours)
4
2
Elevator current (A)
-2
measured output
-4 model output
100 200 300 400 500 600 700 800 900 1000
Time (minutes)
20
15
aindriveload(kW)
10
5
0
-5
M
-10
measured output
-15 model output
100 200 300 400 500 600 700 800 900 1000
Time (minutes)
Figure 11 Comparison of model output and observed output for the winner SSS4 model (see online version for colours)
4
2
Elevator current (A)
-2
measured output
-4 model output
100 200 300 400 500 600 700 800 900 1000
Time (minutes)
20
15
aindriveload(kW)
10
5
0
-5
M
-10
measured output
-15 model output
100 200 300 400 500 600 700 800 900 1000
Time (minutes)
106 V. Sivanandam et al.
Table 3 Model validation test for different candidate models Austin, L.G., Klimpel, R.R. and Luckie, P.T. (1984) Process
for subspace method-based state space models engineering of size reduction: ball milling, Society of Mining
Engineers of the AIME, pp.260291, 458470.
Sl. Candidate Model Austin, L.G., Luckie, P.T. and Seebach, H.M.V. (1976)
FPE AIC MSE
no. model order Optimization of a cement milling circuit with respect to
1 SSS1 1 64.355 4.165 12.6055 particle size distribution and strength development, by
simulation models, Fourth European Symposium
2 SSS2 2 4.606 1.527 9.8168
Zerkleinern, Dechema Monographein, Verlag Chemie,
3 SSS3 3 4.549 1.516 9.8129 Weinheim, pp.519537.
4 SSS4 4 4.190 1.442 8.7435 Austin, L.G., Luckie, P.T. and Wightman, D. (1975) Steady state
5 SSS5 5 5.439 1.696 10.2336 simulation of a cement milling circuit, International Journal
of Mineral Processing, Vol. 2, No. 2, pp.127150.
6 SSS6 6 4.629 1.536 9.6033
Austin, L.G., Weymont, N. P. and Knobloch, O. (1980)
7 SSS7 7 4.707 1.554 9.2741 The simulation of air swept cement mills, Part 1. The
8 SSS8 8 4.405 1.489 9.2741 simulation model, European Symposium on Particle
Technology, pp.640655.
9 SSS9 9 4.332 1.552 9.8456
Benzer, H., Ergun, L. Oner, M. and Lynch, A.J. (2001)
10 SSS10 10 4.992 1.609 9.1239 Simulation of open circuit clinker grinding, Minerals
Engineering, Vol. 14, No. 7, pp.701710.
Boulvin, M. and Vande Wouwer, A. (2003) Modeling and control
6 Conclusions of cement grinding processes, IEEE Transactions on Control
Systems Technology, Vol. 11, No. 5, pp.715725.
This investigation presented two data driven modelling Boulvin, M., Renotte, C., Vande Wouwer, A., Remy, M.,
approaches for cement grinding circuit that maps the Tarasiewicz, S. and Cesar, P.(1999) Modeling, simulation
product quality with feed grindability variations. The first and evaluation of control loops for a cement grinding
modelling approach was based on prediction error, whereas process, European Journal of Control, Vol. 5, No. 1,
subspace identification was the second method. The models pp.1018.
were identified from real-time data collected from a cement Favoreel, W., Moor, B.D. and Van Overschee, P. (2000)
grinding circuit using SCADA systems of FLSmidth. Subspace state space system identification for industrial
processes, Journal of Process Control, Vol. 10, Nos. 2/3,
Different candidate models using both subspace and pp.149155.
prediction error were compared. The model giving a good
Ljung, L. (1999) System Identification, Theory for the User,
trade-off between accuracy and computation complexity 2nd ed., Prentice-Hall, New Jersey.
was selected as the candidate model. The identified
Lynch, A.J., ner, M. and Benzer, H. (2000) Simulation of closed
candidate models were validated using several tests such as cement grinding circuit, Zement-Kalk-Gips, Vol. 53, No. 10,
FPE, AIC, and MSE, and two models were selected from pp.560564.
each category of prediction error and subspace method. Magni, L., Bastin, G. and Wertz, V. (1999) Multivariable non
Two models SSP8 and SSS4 were selected as good linear predictive control of cement mills, IEEE Transaction
candidate models from the available choices that use both on Control Systems Technology, Vol. 7, No. 4, pp.502508.
these methods for modelling. Correlation analysis was used Massei, M., Poggi, S., Tremori, A. and Ferrando, A. (2014)
to analyse the two models and to draw conclusions on the Innovative models for supporting operational planning,
most suitable modelling approach. Our results showed that International Journal of Simulation and Process Modelling,
the subspace identification provides more accurate models Vol. 9, No. 3, pp.181194.
for cement grinding circuit than the PEM for the cement Muthukumar, N., Seshadhri, S., Ramkumar, K., Kavitha, P. and
industry studied in this investigation. Designing model Valentina, E.B. (2015) Supervisory GPC and evolutionary PI
controller for web transport systems, Acta Polytechnica
predictive controllers (MPC) using the proposed approaches
Hungarica, Vol. 12, No. 5, pp.135153.
is the future prospect of this investigation.
Paradis, P.L., Ramdenee, D., Ilinca, A. and Ibrahim, H. (2014)
CFD modelling of thermal distribution in industrial server
centres for configuration optimisation and energy efficiency,
References International Journal of Simulation and Process Modelling,
Vol. 9, Nos. 1/2, pp.6380.
Al-Zuheri, A., Luong, L. and Xing, K. (2014) Using simulation in
verification of a mathematical model for predicting the Pillonetto, G., Chiuso, A and De Nicolao, G. (2011) Prediction
performance of manual assembly line occupied with flexible error identification of linear systems: A nonparametric
workforce, International Journal of Simulation and Process Gaussian regression approach, Automatica, Vol. 47 No. 2,
Modelling, Vol. 9, No. 4, pp.270284. pp.291305.
Apling, A.C. and Ergin, H. (1994) Validation of a grinding model Prado, R.R., Pereira, D.C., Del Rio Vilas, D., Monteil, N.R. and
for a full scale air swept cement mill, 5th International Del Valle, A.G. (2014) A parameterised model of
Mineral Processing Symposium, A.A. Balkema Publishers, multimodal freight transportation for maritime services
Turkey, pp.527532. optimisation, International Journal of Simulation and
Process Modelling, Vol. 9, Nos. 12, pp.3345.
Comparison of subspace and prediction error methods of system identification for cement grinding process 107
Prasath, M.G., Bodil, R., Chidambaram, M. and Jrgensen, J.B. Van Overschee, P. and Moor, B.D. (1994) N4SID:
(2013) Soft constrained based mpc for robust control of a subspace algorithms for the identification of combined
cement grinding circuit, Preprints of the 10th IFAC deterministic-stochastic systems, Automatica, Vol. 30, No. 1,
International Symposium on Dynamics and Control of pp.7593.
Process Systems, The International Federation of Automatic Van Overschee, P. and Moor, B.D. (1995) Choice of state-space
Control, December 1820, Mumbai, India. basis in combined deterministic-stochastic subspace
Ramdenee, D., Ilinca, A., Ion Minea, S. and Hussein, I. (2013) identification, Automatica, Vol. 31, No. 12, pp.18771883.
Modelling of aerodynamic flutter on a NACA 4412 airfoil Venkatesh, S., Ramkumar, K., Seshadhri. S. and Guruprasath, M.
with application to wind turbine blades, International (2015) Data driven models for cement grinding circuit,
Journal of Simulation and Process Modelling, Vol. 8, No. 1, International Journal of Advanced Intelligence Paradigms, in
pp.7987. press.
Schoech, R., Schmid, S., Hillbrand, C. and Fleisch, R. (2013) Viswanathan, K. (1988) Computer based models for grinding and
Optimising plant layout decisions based on emulation industrial case studies, Aufbereitungs-Technik/Mineral
models technical framework and practical insights, Processing, Vol. 10, pp.560572.
International Journal of Simulation and Process Modelling,
Vol. 8, Nos. 2/3, pp.92103. Viswanathan, K. and Narang, K.C. (1988) Computer simulation
and optimisation of ball mills/circuits, World Cement,
Solis, A.O., Longo, F., Nicoletti, L., Caruso, P. and Fazzari, E. Vol. 19, No. 4, pp.143148.
(2014) A modelling and simulation approach to assessment
of a negative binomial approximation in a multi-echelon Wigren, T. (2006) Recursive prediction error identification and
inventory system, International Journal of Simulation and scaling of non-linear state space models using a restricted
Process Modelling, Vol. 9, No. 3, pp.146156. black box parameterization, Automatica, Vol. 42, No. 1,
pp.159168.
Tajini, R., Elhaq, S.L. and Rachid, A. (2014) Modelling
methodology for the simulation of the manufacturing Yang, B. and Ren, B. (2015) Parallel spatio-temporal model for
systems, International Journal of Simulation and Process emergency evacuation simulation International Journal of
Modelling, Vol. 9, No. 4, pp.285305. Simulation and Process Modelling, Vol. 10, No. 1, pp.1018.
Tavakolpour-Saleh, A.R., Nasib, S.A.R., Sepasyan, A. and Zhang, P. and Krieger, A.M. (1993) Appropriate penalties in the
Hashemi, S.M. (2015) Parametric and nonparametric system final prediction error criterion: a decision theoretic approach,
identification of an experimental turbojet engine, Aerospace Statistics & Probability Letters, Vol. 18, No. 3, pp.169177.
Science and Technology, June, Vol. 43, pp.2129. Zhang, Y., Napier Munn, T.J. and Kavetsky, A. (1988)
Tavakolpour-Saleh, A.R., Shumway, R and Tsai, C.L. (1997) Application of comminution and classification modeling to
The model selection criterion AICu, Statistics & Probability grinding of cement clinker, Transaction of Institution of
Letters, Vol. 34, No. 3, pp.285292. Mining and Metallurgy, December, Vol. 97, pp.C207C214.
Van Breusegem, V., Chen, L., Bastin, G., Wertz, V., Zhao, Y., Huang, B., Su, H. and Chu, J. (2012) Prediction error
Werbrouck, V and de Pierpont, C. (1996) An industrial method for identification of LPV models, Journal of Process
application of multivariable linear quadratic control to a Control, Vol. 22, No. 1, pp.180193.
cement mill circuit, IEEE Transaction on Industry
Applications, Vol. 32, No. 3, pp.670677.