2 0 1 1 Interna t io na l Sy mpo s i u m o n IN no va t io ns i n In te l l i g ent Sy sTe ms a nd Appl i ca t io ns

2 011 I n te rn a t i o na l Sy mp o s iu m o n I N no vat i o n s i n I n te l l i g en t S ys T e ms a n d A pp l i ca t i on s

O r g a n i z e d b y

I n c o o p e r a t i o n w i t h

S p o n s o r s

1 5 - 1 8 J u n e 2 0 1 1 , İ s t a n b u l - K a d ı k ö y , T U R K E Y

Copyright © 2011 by the Institute of Electrical and Electronics Engineers, Inc. All rights reserved. Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limit of US copyright law for private use of patrons those articles in this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code is paid through: Copyright Clearance Center 222 Rosewood Drive Danvers, MA 01923 For other copying, reprint or republication permission, write to: IEEE Copyrights Manager IEEE Service Center 445 Hoes Lane P.O. Box 1331 Piscataway, NJ 08855-1331 The papers in this book comprise the proceedings of the meeting mentioned on the cover and title page. They reflect the authors’ opinions and, in the interests of timely dissemination, are published as presented and without change. Their inclusion in this publication does not necessarily constitute endorsement by the editors, the IEEE Computer Society Press, or the Institute of Electrical and Electronics Engineers, Inc. ISBN: 978-1-61284-920-1 Additional copies may be ordered from the IEEE Service Center: IEEE Catalog Number: CFP1172N-PRT IEEE Service Center 445 Hoes Lane P.O. Box 1331 Piscataway, NJ 08855-1331 USA Telephone (toll-free): 1-800-678-IEEE Telephone (direct): +1-732-981-0060 Fax: +1-732-981-9667 E-mail: customer-service@ieee.org Editors: Selim Akyokuş, Adil Alpkoçak, Bülent Bolat, Fırat Doğan, Tülay Yıldırım Cover design by Fırat Doğan, Mustafa Zahid Gürbüz, Cem Kaya, Deniz Şerifoğlu, Z. Hilal Urhan

Neural Network Model of Mill-Fan System Elements Vibration for Predictive Maintenance

T. Balabanov Institute of Information and

Communication Technologies, Bulgarian Academy of Sciences

Acad. G. Bonchev str. bl.2 Sofia – 1113, Bulgaria

tdb@tbsoft.eu

P. Koprinkova-Hristova Institute of System Engineering and

Robotics, Bulgarian Academy of Sciences

Acad. G. Bonchev str. bl.2 Sofia – 1113, Bulgaria

pkoprinkova@icsr.bas.bg

L. Doukovska Institute of Information and

Communication Technologies, Bulgarian Academy of Sciences

Acad. G. Bonchev str. bl.2 Sofia – 1113, Bulgaria l.doukovska@mail.bg

M. Hadjiski University of Chemical Technology

and Metallurgy – Sofia Blvd. St. Kliment Ohridski 8,

Sofia – 1756, Bulgaria hadjiski@uctm.edu

S. Beloreshki Honeywell EOOD

Blvd. Christopher Columbus 64, Sofia – 1528, Bulgaria

beloreshki@honeywell.com

Abstract—In the present paper we focus on online monitoring system for predictive maintenance based on sensor automated inputs. Our subject was a device from Maritsa East 2 power plant - a mill fan. The main sensor information we have access to is based on the vibration of the nearest to the mill rotor bearing block. Our aim was to create a (nonlinear) model able to predict on time possible changes in vibrations tendencies that can be early signal for system work deterioration. For that purpose recently developed kind of Recurrent Neural Networks named Echo state networks were applied. The preliminary investigations showed their good approximation ability for our purpose. Direction of future work will be increasing of predications time horizon.

Keywords-vibration; Recurrent neural network; Echo state network; mill fan system

I. INTRODUCTION Fault detection is recognizing that a problem has occurred,

even if you don't yet know the root cause. Faults may be detected by a variety of quantitative or qualitative means. This includes many of the multivariable, model-based approaches. It also includes simple, traditional techniques for single variables, such as alarms based on high, low, or deviation limits for process variables or rates of change; Statistical Process Control measures; and summary alarms generated by subsystems.

A "fault" or "problem does not have to be the result of a complete failure of a piece of equipment, or even involve specific hardware. For instance, a problem might be defined as non-optimal operation or off-spec product. In a process plant, root causes of non-optimal operation might be hardware failures, but problems might also be caused by poor choice of operating targets, poor feedstock quality, poor controller tuning, and partial loss of catalyst activity, buildup of coke, low

steam system pressure, sensor calibration errors, or human error.

Other elements of Operations Management Automation related to diagnosis include the associated system and user interfaces, and workflow (procedural) support to the overall process. Workflow steps that might be manual or automated include notifications, online instructions, escalation procedures if problems are ignored, fault mitigation actions (what to do while waiting for repairs), direct corrective actions, and steps to return to normal once repairs are completed.

Automated fault detection and diagnosis depends heavily on input from sensors or derived measures of performance. In many applications, such as those in the process industries, sensor failures are among the most common equipment failures. So a major focus in those industries has to be on recognizing sensor problems as well as process problems. Distinguishing between sensor problems and process problems is a major issue. Our usage of the term "sensors" includes process monitoring instrumentation for flow, level, pressure, temperature, power, vibration and so on. In other fields such as network and systems management, it can include other measures such as error rates, CPU utilization, queue lengths, dropped calls, and so on. Also, diagnosis as a decision support activity rather than a fully automated operation is common in the management of large, complex operations such as those found in the process industries and network and systems management.

Here we focus mainly on online monitoring systems, based on sensor or other automated inputs. In the presented paper we have chosen to analyze a device from Maritsa East 2 power plant - a mill fan. The main sensor information we have access to is based on the vibration of the nearest to the mill rotor bearing block. Our aim was to create a (nonlinear) model able

410

to predict on time possible changes in vibrations tendencies that can be early signal for system work deterioration.

In [1] it is proposed to use feed-forward artificial neural network (ANN) structure for mapping of vibration amplitude given time steps in the future accounting for its values for several time steps in the past. Although such approach showed good approximation abilities, it needs preliminary investigation of the needed number of backward time steps to be included in the ANN input vector. From the other hand training of known recurrent neural network (RNN) architectures needs much complex and computationally demanding algorithms such as backpropagation through time (BPTT) or Extended Kalman filter (EKF) methods [2].

The recently proposed Echo state network (ESN) structure [2, 3, 4] incorporates a randomly generated dynamic reservoir and easy trainable output neurons. The feedback connections in the reservoir allow accumulation of input history, “memorizing” resent past and “forgetting” the closest to the beginning input states. Thus ESN structure allows mapping of dynamic dependences without need to know in advance the time steps backwards needed to be accounted for.

Once generated reservoir connections weights as well as the input weights are not subject of training. The only trainable weights are that of the output connection matrix. Since the readout form the reservoir usually is linear dependence the training can be done off-line within a single step or by using Recursive Least Squares algorithm (RLS) in on-line mode [2]. In search of improving reservoir quality algorithm for initial adjustment of reservoir connections matrix was proposed in [5]. It is called intrinsic plasticity (IP) and is aimed at increasing the entropy of the reservoir neurons outputs thus stabilizing its dynamic behavior. Such combined IP-RLS training was already successfully applied in [6, 7].

Since ESN incorporates dynamic into its reservoir in natural way it is a good candidate for approximation of complex dynamics nonlinear dependences we are working with here. The preliminary investigations showed ESN good approximation ability of our purpose. Direction of future work will be increasing of predications time horizon.

II. PROBLEM STATEMENT

A. Mill Fan System In the presented paper we have chosen to analyze a device

from Maritsa East 2 power plant - a mill fan. The mill-fans are used to mill, dry and feed the coal to the burners of the furnace chamber. They are together milling and transporting devices. Mill-fans are most often used for power plants burning brown and lignite coal. In general these are large centrifugal fans which suck flue gases with temperature around 800-1000°C from the top of the furnace chamber. In the same pipe the coal is feed, thus diminishing the drying agent temperature and drying the coal prior entering the fan. The coal is being milled by the fast rotating rotor of the fan and turn into coal dust. This dust is transferred to separator which returns the bigger particles to the fan. The separator can be tuned for a desired dust granulometric size. One of the most important parameters to control is the discharge temperature of the dust-air mixture.

For the considered mill-fan it should be between 145-195°C. Lower than 145 °C may cause clogging of the mill and higher than 195°C may cause the dust to be fired in the ducts prior the burners. This temperature is also a measurement for the load of the mill. The lower the temperature the higher the load is – more coal is fed to the mill. The part which suffers the most and should be taken care of is the rotor of the mill fan. Because of the abrasive effect of the coal it wears out and should be repaired by welding to add more metal to the worn out blades.

A mill fan similar to the considered one is shown on Figure 1. This paper considers prognostic asset maintenance of the coal mill fan system of unit 3 boiler 6 at Maritsa East II complex. There are 8 power units in the complex - four units of 175MW and four units of 210MW each. The boiler which milling system is studied is a Benson type once-though sub-critical boiler. There are four mills per boiler. Each mill fan system has four radial bearings – two in the mill and two in the motor. The Distributed Control System (DCS) installed on the site is Honeywell Experion R301 Process. All the data used in the present research are obtained from the DCS historian system.

Experion is a cost-effective open control and safety system that expands the role of distributed control. It addresses critical manufacturing objectives to facilitate sharing knowledge and managing workflow. Experion provides a safe, robust, scalable, plant-wide system with unprecedented connectivity through all levels of the plant as illustrated. The Experion unified architecture combines DCS functionality and a plant-wide infrastructure that unifies business, process, and asset management to:

•Facilitate knowledge capture

•Promote knowledge sharing

•Optimize work processes

•Accelerate improvement and innovation.

Experion system is capable to provide data to external applications. It supports OPC, Microsoft Excel Data Exchange (HWL AddIn to Excel), ODBC. It is also possible to develop custom tasks running on the EPKS server using C++ and Visual basic.

On the next Figure 2 the structure scheme of the boiler as well as its main parameters is shown. These include: Steam output 690t/h; Steam pressure - at boiler outlet 15,4 (157) MPa (kgf/cm2) and at re-heater outlet 3,5 (35,7) MPa (kgf/cm2); Steam temperature - primary steam temperature 540°C and reheat steam temperature 540°C; Secondary steam flow 600t/h; Design fuel: Lignite; Design gross efficiency of the boiler 89,9%; Weight content of NOx in flue gas (at α=1.4) 3.315mg/Nm.

Prognostic maintenance of a mill fan is considered in the presented paper. It is based on the vibration of the nearest to the mill rotor bearing block. The observation period is 16.12.2010 – 16.01.2011. On 31.12.2010 the rotor of the mill-fan is changed. After the replacement it has been working for 378 hours. The period chosen allows for vibrations analyzes before and after the replacement. It is observed that after the

411

replacement the vibrations with new rotor have higher amplitudes than with the worn out one. This is because of the abrasive wear out of the rotor – the blades become thinner and the rotor becomes lighter. The new rotor is heavier so the vibrations are more intensive even though the rotor has been carefully balanced.

Figure 1. Mill fan system.

Figure 2. Boiler Ep-690-15,4-540 LT.

B. Echo State Network Basics ESNs are a kind of recurrent neural networks that arise

from so called “reservoir computing approaches” [4]. The basic ESN structure is shown in Figure 3 below. It includes a vector of inputs in, feed to all the hidden neurons as well as to the output; a hidden layer called “reservoir” R that contains a set of neurons with randomly generated feedback connections among them and an output vector out that receives inputs form all the reservoir neurons as well as form the input layer.

The ESN output vector out(k) for the current time instance k is usually a linear (usually identity) function fout of its weighted sum of inputs as follows:

( ) ( ) ( )[ ]⎟⎠⎞⎜

⎝⎛= kRkinoutWoutfkout , (1)

Here, Wout is a trainable )( Rninnoutn +× matrix (where nout, nin and nR are the sizes of the corresponding vectors out, in and R).

The neurons in the reservoir have a simple sigmoid output function fres (usually tanh) that depends on both the ESN input in(k) and the previous reservoir state R(k-1):

( ) ( ) ( )⎟⎠⎞⎜

⎝⎛ −+= 1kRresWkininWresfkR (2)

Win and Wres are innRn × and RnRn × matrices that are randomly generated and are not trainable.

Figure 3. Echo state network structure.

There are different approaches for reservoir parameter production [4]. A recent approach used in the present investigation is proposed in [5]. It is called intrinsic plasticity and suggests initial adjustment of these matrices, aiming at increasing the entropy of the reservoir neurons outputs.

ESN training can be done in an off-line or an on-line mode. For on-line training, the RLS algorithm [3] was proposed. It is claimed that it converges fast and it is less computationally expensive in comparison to BPTT-EKF methods [8].

out(k) Σ

reservoir

Wres

R(k) Win Wout

in(k)

412

III. RESULTS AND DISCUSSION

A. Data Pre-processing Since in real industrial conditions there are a lot of

vibration sources, the on-line measurement data contain a lot of noise.

Our first experiment was to try to map the data without noise filtration. As can be seen form the Figure 4, the ESN tries to fit to upper picks of the data and the fitting error is quite big although the tendency of data changes is kept.

Next wavelet de-noising method was applied [9]. In our case it is done by soft heuristic thresholding the wavelet coefficients using symlets for wavelet decomposition at level 5.

Figure 5 shows parts from the data before and after signal de-noising.

200 400 600 800 1000 1200 1400 1600 1800 20001

2

3

4

5training data matching

t, minutes

vibr

atio

n am

plitu

de, m

m

200 400 600 800 1000 1200 1400 1600 1800 20001.5

2

2.5

3

3.5testing data matching

t, minutes

vibr

atio

n am

plitu

de, m

m

Figure 4. Non-filtered data mapping for one minute ahead prediction. The NRMSE errors are 0.78694 and 0.53355 for the training and testing data sets

respectively.

600 800 1000 1200 1400 1600 1800 2000 2200 24001.6

1.8

2

2.2

2.4

2.6

2.8

3

3.2

3.4

3.6data filtration

t, minutes

vibr

atio

n am

plitu

de

Figure 5. Signal denoising. The red line presents original signal while the blue one – the filtered signal.

After filtration the data are divided into two sets - for ESN training and testing.

B. Training of ESN for vibration prediction The aim was to train ESN able to predict vibrations

amplitude several time steps ahead using current measurements as input. For that purpose our ESN has one input and one output – for the current and future time vibration amplitude respectively. We’ve trained several ESNs starting with prediction for 30 minutes ahead through 5 hours ahead. In all cases the reservoir contains only 10 neurons and initial spectral radius was 0.9. The training procedure was on-line RLS and reservoir is initially improved by IP pre-training.

Table 1 below summarizes the Normalized Root-Mean-Square Error (NRMSE) for all considered cases. With increasing of prediction horizon the NRMSE increases but the approximation abilities of ESN remain relatively good even in the last case. Figures 6 through 9 below present the data fitting results for the training and testing data sets. The red line presents real measurements data for training and testing data sets while the blue one – ESN output predictions in both cases.

TABLE I. TRAINING AND TESTING ERRORS.

Predicted time ahead

NRMSE - training data

NRMSE - testing data

30 minutes 0.43484 0.6177

1 hour 0.48196 0.77934

2 hours 0.46545 0.50958

5 hours 0.57177 0.5878

10 20 30 40 50 60 70 80 90 1001

2

3

4

5training data fitting

t, hours

vibr

atio

n am

plitu

de, m

m

10 20 30 40 50 60 70 80 90 1001.5

2

2.5

3testing data fitting

t, hours

vibr

atio

n am

plitu

de, m

m

Figure 6. Training and testing data fitting for 30 minutes ahead prediction.

413

4 1 4

Figure 7. Training and testing data fitting for one hour ahead prediction.

Figure 8. Training and testing data fitting for two hours ahead prediction.

IV. CONCLUSIONS In present investigations we've tested a novel RNN

architecture - Echo state network - for vibrations prediction of a mill fan system. In contrast to previous works, where the NN input incorporates data from several past time steps, here due to the specific recurrent structure of EsN we needed data from only one time step backwards at the NN input.

The obtained results are encouraging. Our next aim will be to improve the ESN predictions quality for bigger time horizon by increasing the reservoir size or using of global feedback connection within EsN architecture.

The long term aim of the presented work is to connect the EsN with the Experion system in order to create a combined predictive maintenance and control system of the mill fan.

ACKNOWLEDGMENT

This work has been supported by the Ministry of Education, Youth and Science under the Project BG051PO001-3.3.04/40/28.08.2009.

REFERENCES

[1] F.E. Ciarapica and G. Giacchetta, "Managing the comdition-based maintenance of a combined-cycle power plant: An approach using soft computing techniques", Journal of Loss Prevention in the Process Industries, vo.19, pp.316-325, 2006.

[2] H. Jaeger, Tutorial on Ttraining Recurrent Neural Networks, Covering BPPT, RTRL, EKF and the "Echo State Network" Approach, GMD Report 159, German National Research Center for Information Technology, 2002.

[3] H. Jaeger, "Adaptive nonlinear system identification with echo state networks", In: Advances in Neural Information Processing Systems 15 (NIPS 2002), MIT Press, Cambridge, MA, 2003, pp. 593-600.

[4] M. Lukosevicius and H. Jaeger, "Reservoir computing approaches to recurrent neural network training", Computer Science Review, vol.3, pp.127-149, 2009.

[5] B. Schrauwen, M. Wandermann, D. Verstraeten, J.J. Steil, "Improving reservoirs using intrinsic plasticity", Neurocomputing, vol.71, pp.11591171, 2008.

[6] P. Koprinkova-Hristova and G. Palm, "Adaptive critic design with ESN critic for bioprocess optimization", Book Series Computer Science, Artificial Neural Networks - ICANN 2010, Lecture Notes in Computer Science, 2010, vol. 6353/2010, pp.438-447.

[7] P. Koprinkova-Hristova, M. Oubbati, G. Palm, "Adaptive critic design with echo state network", 2010 IEEE Int. Conf. on Systems, Man and Cybernetics SMC'2010, Oct. 10-13, 2010, Istanbul, Turkey, pp.10101015.

[8] D. Prokhorov, "Echo state networks: appeal and challenges", In: Proc. of Int. Joint Conf. on Neural Networks (IJCNN), Montreal, Canada, 2005, pp.1463-1466.

[9] D.L. Donoho, "De-noising by soft-thresholding", IEEE Trans. on Inf. Theory, vol.41, No3, pp.613-627, 1995.

Figure 9. Training and testing data fitting for five hours ahead prediction.