Identification, Prediction and Detection of the Process Fault in a Cement Rotary Kiln by Locally Linear Neuro-fuzzy Tech...
Journal of Process Control 21 (2011) 302–308
Contents lists available at ScienceDirect at ScienceDirect
Journal of Process Control j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j p r o c o n t
Identiﬁcation, prediction and detection of the process fault in a cement rotary kiln by locally linear neuro-fuzzy technique Masoud Sadeghian a,b,∗ , Alireza Fatehi a a Advanced Process Automation & Control b
(APAC) Research Group, Faculty of Electrical and Computer Engineering, K.N. Toosi University of Technology, Tehran, Iran Mechanical Department, Sharif University of Technology, Iran
a r t i c l e
i n f o
Article history: Received 21 August 2009 Received in revised form 13 July 2010 Accepted 15 October 2010 Keywords: Cement rotary kiln Fault detection Delay estimation method Locally linear neuro fuzzy model LOLIMOT
a b s t r a c t
In this paper, we use nonlinear system identiﬁcation method to predict and detect process fault of a cement rotary kiln. After selecting proper inputs and output, an input–output model is identiﬁed for the plant.To plant.To identi identify fy thevarious thevarious ope operat rationpoint ionpointss in thekiln,locally thekiln,locally linearneuro linearneuro-fu -fuzzy(LLN zzy(LLNF) F) model model is used. used. This model is trained by LOLIMOT algorithm which is an incremental tree-structure algorithm. Then, using this method, we obtained 3 distinct models for the normal and faulty situations in the kiln. One of the models is for normal condition of the kiln with 15 min prediction horizon. The other two models are presented for the two faulty situations in the kiln with 7 min prediction horizon. At the end, we detect these these faults in validation validation data. data. The data collected from White Saveh Cement Cement Company Company is used in this study. © 2010 Elsevier Ltd. All rights reserved.
1. Introducti Introduction on
Nowadays, many researches are conducted on developing new strategies on process control, so that the automation system not only works as much as possible independent of operators but also can provide much insight and thorough information of the plant for operators. On the other hand, using advanced control theories just to control the process in its normal condition is not sufﬁcient, since during operation, there might be some situations, such as large disturbances or change of quality of input materials, which the process deviates from its normal operating point. These are known as abnormal or fault. So, it is important to not only detect them but also diagnose and properly react on this new situation during the operation, as a human operator does when encounter a faulty condition [1–3] condition [1–3].. Fault detection and diagnosis methods may be divided into signal-bas signal-base e and model model basemethods.In signalbased methods methods some feature feature of the output output signal signal like oscillation oscillation,, frequency frequency deviation deviation and so on is detected which are symptoms of a fault  fault .. For example detection of vibration in an electric motor may be a symptom of some unbalance in the rotor shaft. On the other hand, in model based methods, a model is established for the plant. Deviation of
Correspo Correspondin nding g author. author. Currentaddress:Schoolof Currentaddress:Schoolof Electricaland Electricaland ComputerEngiComputerEngineering, Oklahoma State University, Monroe, Stillwater, OK 74075, USA, Tel.: +1 405 7623979. 7623979. E-mail addresses: [email protected]
(M. [email protected]
(M. Sadeghian), Sadeghian), [email protected]
(A. [email protected]
(A. Fatehi). 0959-1524/$ 0959-1524/$ – see front matter © 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.jprocont.2010.10.009 doi:10.1016/j.jprocont.2010.10.009
estimated parameters, states or outputs from their normal values are evidences of some faults [1,3] faults [1,3].. Nonlinear models like multimodels  ,, f uzzy uzzy logic systems systems [6–8] and neu neural ral networ networks ks  have been widely widely used in faultdetection. faultdetection. Locallinear neuro-fuzz neuro-fuzzy y (LLNF) (LLNF) models models is also extensively extensively used in fault detection detection [10–13] d [10–13] due ue to the local local interp interpret retati ation on of models models andtheirpowerfullearn andtheirpowerfullearningalgoingalgorithms  rithms .. In this paper we use nonlinear system identiﬁcation method in order to predict and detect some common faulty conditions in the most important part of a cement factory, i.e. cement rotary kiln. To identify the kiln, we use LLNF model, also referred to as Takagi–Sugeno fuzzy model [12,15] model [12,15].. LOLIMOT algorithm is used  to learn both the structure and the parameters of the local models. In a previous study on the same process  process ,, a model was established for the process which could detect normal from some faults in the kiln. In this study, two other kinds of faults; ringing and coating, are not only detected but also isolated. The paper is organized as follows: In the next section, a brief description on rotary kiln is given. Also, some abnormal conditions are mentioned that may happen frequently in it. Then the input–out input–output put selection selection to detect detect faulty faulty situation situation are discussed. discussed. In Section Section 3, the input channel channel delays delays on the model model are estimated base on Lipschitz method  method .. Afterward in Section 4 Section 4,, with NNLF model and LOLIMOT learning algorithm, three models are designed for normal and faulty situation of the kiln. Section 5 is i s devoted to the discussion on detecting normal and two abnormal conditions that were observed in test and validation data. Section 6 is about fault detection tests. Conclusion comes at the end.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308 Table 1 Delays from inputs to output.
Fig. 1. A cement rotary kiln plant.
2. Cement rotary kiln
Cement is a substance which is made of grinded gypsum and cement clinker which itself is produced from a burned mixture of limestone and clay in certain percentages. Cement is used to bind other materials together. Since cement factory is much expanded and it is consisted of different instruments and various processes in each part, modern condition monitoring methods are seemed suitable to be used in order to prevent abnormal conditions which end in a loss. Cement rotary kiln is the most vital part of a cement factory whose outcome is cement clinker. A rotary kiln is a cylinder with a length of around 70m and a diameter of around 5 m in a factory with a capacity of producing about 2000tons of clinker in a day. The kiln is rotated by a powerful electrical motor. The temperature in the hottest point in the kiln is up to 1400 ◦ C (Fig. 1). The kiln works nonstop and an impending fault may cause inferior product at the end of the line or a halt in a large part of the factory with irreparable damages to equipments. Sudden kiln stop could damagevarious pointsof thekiln basedon a high heat degree ﬂuctuation. Hence, it is essential to use some methods in order to prevent such faults. Many of the abnormal conditions in the plant are detected andreportedby the plant automation andsafety system such as high temperature of cyclones, lack of pressure in hydraulic systems and so on. There are, however, other abnormal conditionswhich are not detected by conventional automation systems. In these cases, none of the measured variables are beyond their limitations, but the overall behavior of the plant is abnormal. An expert operator can recognize these conditions by comparing the current behavior of the plant by what was expected from the normal condition behavior. What we are concerned about in this paper is these types offaults orabnormalitywhich cause poor product or are the origin of a halt in the process. For instance, some of the common abnormalities in the kiln are • Coating
• Ringing. • Super
heated or super chilled.
In this paper, in order to continue the previous attempts for fault detection in kiln , f o r the ﬁrst time, the procedure of identiﬁcation, prediction and detection of two common, meanwhile damaging fault in the process, these are, ringing and coating, are introduced. In this procedure, during the identiﬁcation process and analyzing the kiln data, we found that these two faults show a special behavior in cement rotary kiln output, upon occurrence. We use systemidentiﬁcation approaches for the sake of abnormal condition detection. The output that is going to be identiﬁed is the
Martial feed rate Fuel feed rate Kiln speed I.D fan speed Secondary air pressure
18 4 36 0 0
temperature of the ﬁrst point at the beginning of the kiln, which is called back-end temperature. It is in the calcining zone of the kiln and has a signiﬁcant role on the quality of the clinker. The inputs are material feed rate, fuel feed rate, kiln speed, I.D. fan speed and secondaryair pressure.The reason of using these inputsand output is themore affectionconﬁrmedby experts andprocess engineersof the factory. In their point of view, the back-end temperature illustratesthe internalcondition of thekiln andby means of theselected inputs it is possible to recognize whether the process is going well or something undesirable is taking place. We have been collecting and analyzing the data of a fourteenweek periodof operationof thekiln.Accordingto thecollecteddata, in two cases the kiln operation changes. One case is when there are some mechanical and electrical defects in the system; the other is when the operators change some of the inputs in accordance with operational policies. It means as an abnormal condition happens; operators detect it and make proper reaction to overcome the condition. Therefore, the period which an abnormal condition stays is short and we have to detect it in this short period. By studying the operators note and discussion with process engineers, we separatefaulty data from normal data ofthe kiln data. Here, with each setof data we make an effortto eliminateconstant, repetitive and faulty operational points. The reason why we eliminated the constant and repetitive points is that in these situations the variables are affected more by the small noise and disturbance. Therefore, in practice, theidentiﬁed dynamics by these data arenot the main dynamics of the process. On the other hand, if the volume of the used data for making a model is high in one operating point, this point gets more weight; as a consequence increases error in other operating points for making a model. After removing invalid data and pre processing on them . They are divided to three parts: 50% of that is used as the training set, 20% as the test set and the rest of it as the validation data set. 3. Inputs channels delay estimation
Before identifying the 5 inputs and 1 output model of the kiln, we should estimate its input channels delays. The reason that we estimatethe inputs delays witha free-modelapproachis thatdetermining themduring the identiﬁcation,makes this taskburdensome and increases some computational volume. Therefore determining the input channel delays shrinks the search space to a high extent andmakes therestof theidentiﬁcationphaseeasier andmoreaccurate. The approach that we use is based on Lipschitz method which was presented by Makaremi et al. [17,19]. The results are shown in Table 1. 4. Identiﬁcation and prediction of rotary kiln
In the preceding section, we estimated the input channel delays of the kiln. Knowing these parameters, the search space for the identiﬁcation shrinks and it is easier to do the rest of the job, i.e. determining the suitable number of dynamics on each input and output, and approximating the best function which represents the behavior of the kiln as well.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308 Table 2 Actual sampling rate for different variables.
Sampling time (s)
Material feed rate Fuel feed rate Kiln speed I.D fan speed Secondary air pressure Back end temperature
150 90 300 60 60 60
Table 3 The best number of dynamics for the inputs and output.
Fig. 2. Network structure of a local linear neuro-fuzzy model .
We use locally linear neuro-fuzzy (LLNF) network to identify the normal and faulty condition of kiln and the LOLIMOT algorithm to ﬁnd the best structure and parameters of the network. In the following, LLNF networks and the LOLIMOT algorithm is reviewed brieﬂy. Then the result of applying them on kiln data is presented. 4.1. Locally linear neuro-fuzzy network
The most important reasons why LLNF network is selected are as follow: (1) High accuracy. (2) Robustness. (3) Computational efﬁciency and (4) Smooth switch for multiple models. LLNF model is among the best and most popular nonlinear model structure that is used to identify a process. Fig. 2 shows the structureof a LLNF model.The strategy of LLNF is that thenonlinear modelof theprocessis divided into a numberof smaller andsimpler linear models . Therefore, each neuron represents a Local Linear Model (LLM) which its validity domain is veriﬁed by a validity function. The validity function that is selected in this study is Gaussian function. Based on these validity functions a validity weight is assigned to each model; theoutput of themodelis computed based on the weighted sum of the LLM outputs. The strategy to divide the behavioral space of the process plays a key role in order to success of LLNF model. The algorithm that is selected to solve the problem is called LOcally LInear MOdel Tree (LOLIMOT) [20,21]. LOLIMOT is an additional tree structure algorithm that splits up the input space by orthogonal axis. To identify a process allpossible axis orthogonal divisions that split each input space of a local model into two equal spaces are evaluated by estimation of the parameters of the local models for each division. A division that hasminimum overall error is selected as a new branch on the model tree. For each division a Gaussian validity function is assigned on the middle of the represented sub-space. The parameters of LLMs areoptimizedbasedon a locally weighted least squares error. The LOLIMOT algorithm starts by one neuron, i.e. a single linear model, and continues to reach the best number of neurons which can satisfy the overall minimum error. We used LLNF network with LOLIMOT learning algorithm to identify the kiln. The ﬁrst problem we faced is to determine a commonsample time. Each input has itsparticular duration of effect on the output. For instance, materials feed to the kiln; have a delay of about 18 min. Besides the variation of temperature inside the kiln, that is a consequenceof changing fuel ﬂowstart toeffect after4 min.
Material feed rate Fuel feed rate Kiln speed I.D Fan speed Secondary air pressure Back end temperature
Number of dynamics Normal condition
5 3 2 11 3 10
5 3 2 4 2 11
2 3 3 2 4 4
Kiln speed variation may effect after 30 min on back-end temperature. Also, the time constant from the inputs to the output are not thesame.For thefast affectinginputs smaller sampling rate is more appropriation. For the slow affecting inputs bigger sampling rate is more appropriation. Using different sampling times, we avoid increasing the space of the inputs of the model without addition of any information to it; while, using various sampling times makes problem in analyzing and modeling process. To solve this problem, we use a basic constantsampling timeof 30s, where the used input samples with slower affection is a coefﬁcient rate of the basic constant sampling time . Therefore, we resample each input with a different rates. Table 2 shows sampling time for each of them. For instant, the 2 samples of the kiln speed equals 600 s that equals 20 samples of the back-end temperature. The lastproblem was toﬁnd the numberof dynamics ofthe outputand theinputs.Withregard tothe pre-knowledge aboutthe kiln properties, the range of inputs dynamism is obtained then through trial and error during identiﬁcation, the best number of inputs and output dynamics are obtained. The best numbers of dynamics used for identiﬁcation are presented in Table 3. Whereas our goal is abnormal condition detection, the prediction horizon in the identiﬁcation is 7 min for abnormal conditions and 15min for normal condition to increase the prediction horizon in order to predict kiln conditions some minutes in advance. Fig. 3 shows the error on train and test sets with respect to the number of the neurons in the normal model. It shows that a LLNF with two neurons can model the plant adequately in the normal situation. The one neuron linear neuro-fuzzy model, in fact, is a globally linear model. Fig.3 shows that this globallylinear model hashigherror
Fig. 3. Error on train and test data respect to different number of neurons.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308 Table 4 Root mean square error (RMSE).
Fig. 4. Normal model with 10 min prediction horizon.
Prediction horizon (min)
1 3 5 7 10 15
0.00004 0.00027 0.00092 0.00256 0.00456 0.00974
0.00018 0.00049 0.00062 0.00078 – –
0.00018 0.01032 0.03803 0.07997 – –
rotates, sometimes the ring is collapsed from top of the kiln. This occurrence causes a temperature increase, in the back-end zone. This event will be taking place from time to time; as a result, the back-end temperature showsa ﬂuctuation behavior.The otherfault is calledcoating fault which is more harmful forthe kiln andclinker producing. The most important fact about coating fault is that it may cause the back-end temperature decreases with a linear negative slope. The reason why it shows this behavior is that the kiln wall surface is completely covered bycoating layersin theback-end zone. The coating preventsthe ﬂow ofhot air tothe back-end zone; after passing some periods of time the coating becomes thicker and thicker andif it does notdetectedby theoperatoron time, they will be forcedto stop thekiln. Naturally, theoperation will resumeafter the kiln is cleaned. 4.2. Fault detection in the cement rotary kiln
Fig. 5. Coating model with 5 min prediction horizon.
in test data. Therefore, it is impossible to make an accurate model by one globally linear model. Fig. 4 shows the response of the normal model output and the real output for 10 min prediction horizon and Fig. 5 shows the response of the coating fault model output and the real output for 5 min prediction horizon for test data. Also, Fig. 6 shows the response of the ringing fault model output and the real output for 5 min prediction horizon for test data. Fordifferent operations among these models we calculated root mean square error (RMSE) for test data which are shown inTable 4. Oneof themostimportantthingsthatwe realizedafterthe identiﬁcation and analysis of the faulty models is the behavior of faulty condition when they occur.We understood that when ringing fault is going to occur the back-end temperature begins to ﬂuctuate rapidly. The reason why it shows this behavior is that, the feed materialwhich go through thekiln are rich with alkalescency characteristic in comparison with normal materials. They begin to stick to the kiln wall that ends in formation the ring. In this condition, the ring prevents the ﬂow of hot air to the back-end zone; consequently, the back-end temperature is decreased. When the kiln
Fig. 6. Ringing model with 5min prediction horizon.
In the previous section, three distinct models for the kiln have been developed and introduced. One of the models is a normal model and the others are related to ringing and coating fault. In this part, we want to detect and extract abnormal conditions that exist in validation data. The procedure that we use in detecting an abnormality is based on the fact that it has more lasting effect on the output rather than those of noise or disturbance. Therefore, we give three validation data where each set relates to one of the kiln conditions, then we give each validity data to the three models that have been developedto check whether themodels could accurately track thereal output; forexample,when thenormal validation data is givento the three models, the normal model must have the minimum error in comparison with the real output of the other fault models. If the ringing validation data is given to the three models the ringing model must have the minimum error compared to the normaland coating models. These conditions are distinguished with their major error in a long lasting period. Figs. 7–9 show the evaluation of the three models with normal validity data. As you see the faulty models could not identify and recognize the validity data. Figs. 10–12 show the evaluation of the three models with ringing validity data. Figs. 13–15 show the evaluation of the three models with coating validity data.
Fig. 7. Test of the normal model with the normal validity data.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308
Fig. 8. Test of the ringing model with the normal validity data.
Fig. 12. Test of the coating model with the ringing validity data.
Fig. 13. Test of the normal model with the coating validity data. Fig. 9. Test of the coating model with the normal validity data.
Fig. 10. Test of the normal model with the ringing validity data.
For comparing the three models and their response to the various validity data, we calculated root mean square error (RMSE) for each of them. They are shown in Table 5. As it is shown in this table for each situation of the kiln behavior; the respected model has lower RMSE. Therefore, comparing RMSE of the models errors determines the condition of the kiln.
Fig. 11. Test of the ringing model with the ringing validity data.
Fig. 14. Test of the ringing model with the coating validity data.
4.3. Fault detection test
In the previous section, we show that three distinct models can distinguish different kiln operation points. Now, we want to detect faulty situation from normal condition. To do this, we select
Fig. 15. Test of the coating model with the coating validity data.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308
Table 5 Root mean square error for different models in various modes.
Normal Ring Coating
0.000459 0.003278 0.042313
0.430821 0.006542 0.012692
0.030940 0.338630 0.000116
Fig. 18. J R (t ) with =0.73.
Fig. 16. Block diagram for fault detection test.
Fig. 19. J C (t ) with = 0.73.
3 days of data of the back-end temperature. This data is simultaneously given to the three models andthe errorsbetween each model and real output data are obtained by comparing the following cost function among these models.
J (t ) =
(t − ) ∗ e2 ;
0 < ≤ 1
The model that has minimum errors shows the condition of the kiln operation. The block diagram for fault detection test is shown in Fig. 16. Figs. 17–19 show the cost of each model in the fault detection test. After comparing the cost of each model and selecting the model with the least cost, we obtained Fig. 20 that shows the result of the fault detection test. According to this ﬁgure twofaults occurred during these 3 days. The ﬁrst one which starts from time 2400 min is ring error and the second which is coating error is affected the kiln from around 3500 min.
Fig. 17. J N (t ) with = 0.73.
Fig. 20. The result of fault detection test.
In this paper, nonlinear system identiﬁcation method was used for identiﬁcation, prediction and detection of the fault process in the cement rotary kiln in White Saveh Cement factory. Back-end temperature was used as the process monitor of the various conditions. The special character of this variable is that it can show the normal and abnormal conditions inside the kiln. At ﬁrst, for this purpose, the effective inputs were selected. Then, to ease the identiﬁcation of the kiln, we estimated input channel delays based on Lipschitz numbers. Next, the effective dynamic is obtained for each input corresponding to the output. After that, with LLNF models and LOLIMOT learning algorithm, three nonlinear models were developed for the healthy and faulty condition of the kiln. All models could predict their respected situation with proper prediction horizon. Finally, by means of these models, we could distinguish faultand normalconditionsin validation data.The resultof thefault detection algorithm performance indicates that we can predict the fault occurrence 7min in advance.
M. Sadeghian, A. Fatehi / Journal of Process Control 21 (2011) 302–308
Authors are grateful to Saveh Cement Company for their kind help and technical assistance during this project. References  R. Isermann, Fault Diagnosis Systems. An Introduction from Fault Detection to Fault Tolerance, Springer, Berlin, 2005.  J. Korbicz, J.M. Kos’cielny, Z. Kowalczuk, W. Cholewa (Eds.), Fault Diagnosis. Models, Artiﬁcial Intelligence, Applications, Springer-Verlag, Berlin, 2004.  R.J. Patton, J. Chen, Robust Model-Based Fault Diagnosis for Dynamic Systems, Kluwer Academic Publishers, London, 1999.  D. Neumann, Fault diagnosis of machine-tools by estimation of signal spectra, Proc. IFAC SAFEPROCESS Symposium, (1991) 73-78.  M. Rodrigues, D. Theilliol, M. Adam-Medina, D. Sauter, A fault detection and isolation scheme for industrial systems based on multiple operating models, Control Engineering Practice 16 (2008) 225–239.  M. Kowal, J. Korbicz, Fault detection under fuzzy model uncertainty, International Journal of Automation and Computing 4 (2) (2007) 117–124.  R. Isermann, On fuzzy logicapplications forautomatic control, supervision,and fault diagnosis, IEEE Transactions on Systems, Man and Cybernetics Part A 28 (2) (1998) 221–235.  P. Balle, D. Fuessel, Closed-loop fault diagnosis based on a nonlinear process model and automatic fuzzy rule generation, Engineering Applications of Artiﬁcial Intelligence 13 (2000) 695–704.  S. Rajakarunakaran, P. Venkumar, D. Devaraj, K. Surya Prakasa Rao, Artiﬁcial neural network approach for fault detection in rotary system, Applied Soft Computing 8 (2008) 740–748.
 R.J.Patton, J.Korbicz,Advancesin computational intelligencefor fault diagnosis systems, International Journal of Applied Mathematics and Computer Science (Special Issue) 9 (3) (1999).  M. Ayoubi, R. Isermann, Neuro-fuzzy systems for diagnosis, Fuzzy Sets and System 89 (3) (1997) 289–307.  D. Rutkowska, Neuro-Fuzzy Architectures and Hybrid Learning, Springer, New York, Heidelberg, 2002.  E. Zio,G. Gola, A neuro-fuzzytechnique forfault diagnosisand its applicationto rotatingmachinery,ReliabilityEngineeringandSystemSafety94 (2009) 78–88.  O. Nelles, Nonlinear System Identiﬁcation, Springer-Verlag, Berlin, 2001.  D. Rutkowska, L. Zadeh, Neuro-fuzzy and soft computing, (Special Issue) International Journal of Applied Mathematics and Computer Science 10 (4) (2000).  I. Makaremi, A. Fatehi, B. Nadjar-Araabi, Abnormal condition detection in a cement rotary kiln with system identiﬁcation methods, International Journal of Process Control 19 (9) (2009) 1538–1545.  I. Makaremi, A. Fatehi, B. Nadjar-Araabi, Lipschitz in numbers: a medium for delay estimation, in: 17th IFAC World Congress, Seoul, Korea, July 6-11 2008, 2008.  Y. Zhu,MultivariableSystemIdentiﬁcationfor Process Control,Elsevier Science Ltd., 2001.  I. Makaremi,IntelligentCondition Monitoring of a Cement Rotary on Kiln,M.Sc. thesis, KN Toosi Univ. of Tech, February 2007.  O. Nelles, Local linear model tree for on-line identiﬁcation of time variant nonlinear dynamic systems, in: Proc. of International Conference on Artiﬁcial Neural Network, Bochum, Germany, 1996, pp. 115–120.  O. Nelles, R. Isermann, Basis function networks for interpolation of local linear models,in: Proc. of IEEEConferenceon Decisionand Control,Kobe,Japan,1996, pp. 470–475.  R.M. Milasi, M.R. Jamali, C. Lucas, A Bioinspired and Multi-objective Approach, International Journal of Control, Automation, and Systems 5 (4) (2007) 436–443.