FAULT DETECTION IN COAL MILLS USED IN POWER PLANTS Peter Fogh Odgaard Babak Mataji

Department of Control Engineering, Aalborg University, Aalborg, Denmark, [email protected] Elsam Engineering A/S, Kraftvrksvej 53, DK-7000 Fredericia, Denmark, [email protected]

Abstract In order to achieve high performance and eciency of coal-red power plants, it is highly important to control the coal ow into the boiler in the power plant. This means suppression of disturbances and force the coal mill to deliver the required coal ow, as well as monitor the coal mill in order to detect faults in the coal mill when they emerge. This paper deals with the second objective. Based on a simple dynamic model of the energy balance a residual is formed for the coal mill. An optimal unknown input observer is designed to estimate this residual. The estimated residual is following tested by measured data of a fault in a coal mill, it can hereby be concluded that this residual is very useful for detecting faults in the coal mill. Keywords: Fault Detection, Coal-red Power Plants, Optimal Unknown Input Observer

1. INTRODUCTION The performance of a coal-red power plant is highly dependent on the supplied coal ow into the boiler. In order to achieve good performance of the power plant, this coal ow should be known and controllable. In other words the performance requirements of the power plant can be transfered to the coal mill which delivers the pulverized coal to the burners inside the power plant boiler. One thing is to guarantee the required coal ow, in terms of suppressing plant disturbances. It is, however, another thing to monitor the coal mill for detecting faults. In order to avoid additional failures and decreased performance in the power plant unit, it is important to detect any fault as soon as possible, and hopefully as the fault emerges. In most coal mills it is not possible to measure the pulverized coal ow out of the mill. The coal ow into the mill is only indirectly mea-

surable. Implying that the early fault detection based on detection of drops in the coal ow of the coal mill, is not so simple as it seems. A critical example on a fault in the coal mill is caused by a blocking in the raw coal inlet pipe, a coal mill is illustrated in Fig. 1. This blocking of the raw coal inlet will with time result in a stopped coal ow from the coal mill to the boiler, and since the actual coal ow into the boiler cannot be measured, the fault in the coal mill cannot be directly detected. Dynamic modeling and nominal control of these coal mills have been the topic of numerous of publications. Some examples dealing with modeling of coal mills are (Rees and Fan 2003), (Zhang et al. 2002) and (Tigges et al. 1998). Controllers for the coal mill are designed in (Rees and Fan 2003) and (Hasselbacher et al. 1992). High order dynamic models and observer design for coal mills

are the topics in (Fukayama et al. 2004). However, the interest has not previously been addressed on the monitoring of the coal mills with the purpose of detecting any emerging faults. The temperature of coal dust ow after the classier is a good indicator of the health of the coal mill, since a non-intendant change in the in and out ows of the coal mill will result in a variation in the temperature. However, the measured temperature cannot be used to detect faults directly, since a control loop is formed to keep the measured temperature at 100 C, in order to evaporate the moisture in the coal. Meaning that the fault can be seen by combining the temperature and the energy in the primary air. I.e. it might be useful to monitor an energy balance state space model of the coal mill, which has the energy ows in and out of the coal mill as inputs, and the temperature as output. A fault in the coal mill is in this setting an extra energy input to this model. If this extra energy input could be estimated, it would be a useful residual for monitoring the coal mill, with the intension of detecting eventual faults in the coal mill. This residual can be estimated by introducing a state representing the fault in the energy model. This state is driven by an unknown input which is ltered through the fault model. This described estimation problem can be solved by the use of an optimal unknown input observer, see (Chen and Patton 1999). Such an optimal unknown input observer is as consequence designed for estimating the residual/fault signal and the temperature. A threshold method is following applied to the residual with the purpose of detecting the fault in coal mill, it is as well interesting to investigate the sign of this energy balance, since it can be used to separate the possible faults into two classes which help in the fault isolation. This fault detection method is subsequently applied to data sampled during a fault in a coal mill. This experiment shows the potential of the proposed residual estimated based on the energy balance model of the coal mill.

Pulverized coal Rotating Classifier

11 00 11 00 1 0 11 00 1 0 111 000 1 0 111 000 1 0 111 000 1 0 111 000 1 0 1 0 1 0 111 000 11 00 111 000 11 00 11 00 111 000 1 0 111 000 1 0 11 00 1 0 11 00 1 0 1 0 11 00 1 0 11 00 11 00 11 00 11 00 11 00 1 0 1 0 1 0 11 00 11 00 Roller 11 00 1 1 0 0 1 1 0 0 1 1 0 0
Grinding table

Raw coal Inlet

Roller

Primary Air
Figure 1. An illustration of the coal mill. by the furnace. The ratio of these air ows are used to control the temperature of the primary air ow. Coal particles which in the pulverizing have been small enough will pass through the classier and out through the outlet pipes into the boiler.

2.1 Control and measurements References to coal ow and primary air ow are given by the general power plant controller, as well as rotational speed of the classier. The temperature of the primary air is used to control the temperature in the coal mill at the classier. The temperature controller is often required to keep temperature constant at 100 C in order to evaporate the moisture content in the coal. A coal mill is a harsh environment in which it is dicult to perform measurements, this means that all the variables are not measurable. E.g. the actual coal ows in and out of the coal mill are not measurable. However, the primary air ow and temperature are, as well as the temperature at the classier.

2. THE COAL MILL The work presented in this paper, is based on a Babcock MPS 212 coal mill used at Elsams Nordjyllandsvrktet Unit 3. However, the method proposed in the paper is so generic that it can be applied to other types of coal mills. The coal mill is illustrated in principles in Fig. 1. The coal is fed to the coal mill through the central inlet pipe. The coal is pulverized on the rotating grinding table by the rollers. The pulverized coal is then blown up and the moisture content is evaporated by the hot primary air. The primary air is mixed by cold outside air and heated outside air, which is heated

104 nonlinear model measurements linear model 102

Qair Pmotor

T mm

Qcoal Qmoisture
degrees C

100

98

Figure 2. An illustration of energy balance in the coal mill, where T is the temperature in the mill, Qair is the energy in the primary air ow, Pmotor denotes the power delivered by the roller motors, Qcoal is the energy in the coal ow, and Qmoisture is the energy in the coal moisture. 2.2 Faults A number of dierent faults can occur in the coal mill, and if the fault leads to a decrease in the output coal ow from the coal mill, this can lead to drop out of the entire power plant unit. Some examples on critical faults are: choking of the raw coal inlet, faults in primary air supply both the fan and the temperature controller, and nally all the sensors, etc. All these are faults which it are urgent to detect.

96

94

92

90

0.5

1 samples [n]

1.5

2.5 x 10
4

Figure 3. A plot of the non-linear and linear model response compared with measurements of a step response on the coal mill. are dierent, however the model error due to heat of steam to a couple of degrees above 100 C is neglectable in this context. The dynamic non-linear model is subsequently given by mm Cm T (t) =mpa (t)Cair (TPA (t) T (t)) + mc (t)Cc (Ts T (t)) + (t)mc (t)Cw Ts (t)mc (t)Hst T (t) + Pmotor (t), (2)

3. ENERGY BALANCE MODEL OF THE COAL MILL A simple energy balance model of the coal mill is derived based on (Rees and Fan 2003). In this model the coal mill is seen as one body with the mass mm , as illustrated in Fig. 2, in which T is the temperature in the mill, Qair is the energy in the primary air ow , Pmotor denotes the power delivered by the grinding table, Qcoal is the energy in the coal ow, and Qmoisture is the energy in the coal moisture. It is also assumed that the input coal ow is equal the output coal ow. Even though this assumption is only entirely true for steady state, it is assumed in this paper for simplifying the model. A more detailed model which takes dierent coal ows into account might result in better results with respect to even earlier detection of the faults. The energy balance is given by (1). mm Cm T (t) = Qair (t) Qcoal (t) Qmoisture (t) + Pmotor (t). (1)

where: Cm is the specic heat of the mill, T (t) is the mill temperature at the classier, mpa (t) is the primary air mass ow in and out of the mill, Cair is the specic heat of air, TPA (t) is the temperature of the inlet primary air, mc (t) is the coal mass ow, Cc is the specic heat of the coal, Ts is the surrounding temperature, (t) is the ratio of moisture in the coal, Cw is the specic heat of the moisture, Hst parameter combining the latent heat of the steam and specic heat of the water, and Pmotor (t) is the power delivered by the mill motor. All parameters in this model are found in data books except mm Cm which is identied based on measurements of a step response on the coal mill. The model response is compared with measurements as well as the response of a linearized model in Fig. 3. From this gure it can be seen that the responses of both models are quit similar to the large dynamical changes as the measurements show. However, it is dicult to validate the details in the response due to the way the signals are sampled. A dead band on one per cent is applied to these measurements meaning that the signals shall have changes of a given size before this change is sampled. The non-linear model (2) is subsequently linearized and transformed into a state space representation, see (3), the motor power is also neglected from this state space model since it is

The heating and evaporation of the moisture in the coal is modeled by a combined heating coecient. The temperature is due to the control loop is kept at 100 C. The latent energy of the evaporation dominates the energy required for a few degrees heating of the moisture. The combined heat coecient, Hst , is following dened as Hst = Cw + Lsteam /100, where Cw is the specic heat of the water, and Lsteam is the latent heat. This combined heat coecient does not deal with the fact that the specic heat of water and steam

much smaller than the other powers in the equation. mPA (t) TPA (t) (3) T (t) = AT (t) + B mc (t) + q(t), (t) Tm (t) = CT (t) + r(t), (4) where a given signal is linearized by = o , o is the operation point of , q(t) is the normal distributed process disturbances, r(t) is the normal distributed measurement noises, Tm (t) is the measured temperature and (mPA,o Cair mc,o (Cc + o Hst )) A= , mm Cm (5) T Cair (TPA,o To ) mm Cm Cair mPA,o mm Cm B = C (T T ) + (C T H T ) s o o w s st o c mm Cm mc,o (Cw Ts Hst To ) mm Cm (6) C = I. 4. RESIDUAL GENERATION The energy balance given by (2) and in a linear state space version in (3-7) would in case of a fault deviate from the coal mill which indicates an unbalance in the model. The unbalance in the model can be represented by an additional fault input. It is in addition a possible measures of the moisture in the coal which is not valid during dynamic behavior of the coal mill. By combining these two model inputs to one unknown input, which represents both the fault energy and coal moisture, an estimate of this unknown signal can be used as a residual for detection faults in the coal mill. It can be assumed that the moisture content is changing much slower than a fault can occur or emerge. However, in order to estimate the need energy/fault signal it is more convenient to represent it by an internal state in the model. The simplest representation is a rst order lowpass lter which is included in the model. This results in the linear model given by (8-13), mPA (t) (t) (t) T T m = Aq Q(t) + Bq TPA (t) Q(t) (8) mc (t) + Eq Qn (t) + q(t), T (t) + r(t), Tm (t) = Cq Q(t) (9) (7)

to balance the model), Qn (t) is the generic unknown input which is low-pass ltered in order to represent the residual, and 1 A (10) Aq = p , 0 p Bq = B1(13) , 0 (11) Cq = C 0 , 1 , Eq = 1 (12) (13)

where B1(13) denotes the three rst elements of B. The model represented by (8-13) is discritizied before an observer is designed to estimate the states in the model, and hereby the residual, see (14-15). mPA [n] T [n] T [n + 1] + Bd TPA [n] m = Ad Q[n] Q[n + 1] (14) m [n] ,
c

+ Ed Qn [n] + q[n],

T [n] + r[n], Tm [n] = Cd Q[n]

(15)

where Ad , Bd , Cd , Ed are the discritizied representations of Aq , Bq , Cq , Eq . This model is a system with an unknown input, disturbances and measurement noises, i.e. an optimal unknown input observer would be an obvious observer to use for estimating the residual.

4.1 Optimal unknown input observer The optimal unknown input observer is described in (Chen and Patton 1999). For discrete time systems with unknown inputs and disturbances which can be represented by x[n + 1] = An x[n] + Bn u[n] + En d[n] + q[n], y[n] = Cn x[n] + r[n], (16) (17)

an optimal unknown input observer of the following form can be derived z[n + 1] = Fn+1 z[n] + Tn+1 Bn u[n] + Kn+1 y[n], x[n + 1] = z[n + 1] + Hn+1 y[n + 1]. (18) (19)

where p is the pole of internal residual model, Q(t) is the estimated residual (need energy ow

The basic idea in this observer is to eliminate the dependency of the unknown input from the estimation error by matrix transforms, and subsequently design a Kalman estimator for the transformed system. A positive side eect of this, is that the estimator gain is recomputed at each sample meaning that the model can be changed such that the point of operations can be updated. The schemes for computing the matrices in the optimal unknown input observer can be seen in Appendix A.

20000 observed calculated

x 10

observed

15000

2.5

2
10000 energy

1.5 energy
5000

0.5
0

0
5000

200

400

600 Samples [n]

800

1000

1200

0.5 1100

1150

1200

1250 1300 Samples [n]

1350

1400

1450

Figure 4. Comparison of the estimated energy due to the coal moisture and the static calculated of this energy. The variance of the disturbance and measurement noises Q[n] and R[n], as well as the internal fault model parameter p are all found by trial and error, based on experimental data, in the way that the observer estimates the energy signal in the fault free situation well, see Fig. 4. From this gure it can be seen that the observer estimates the fault signal due to the coal moisture content well, and it is hereby concluded that the observer and model is well tuned. 4.2 Detection rules In this section two methods for detecting the faults based on the residual are described. The rst method is to use a simple threshold , and compare the absolute value of Q[n] with a threshold as in (20). This means that a fault is de tected based on this rule if Q[n] is larger than the threshold . In the following fd [n] denotes a signal which is equal 1 in case of a fault and 0 elsewhere. 1 if Q[n] > , fd [n] = (20) 0 else were. Another method partly removes the inuence by coal moisture from Q[n] by dividing it with the energy ow contribution from the coal, before the threshold is applied, see (21). Q[n] 1 if > , fd [n] = mc [n] Cc (Ts T [n]) 0 else were. (21)

Figure 5. A plot of the signal used in fault detec tion method 1, Q[n]. 4.3 Fault isolation Earlier in this paper it was mentioned that the sign of the energy balance can be used to separate the possible faults in two groups. A stopped coal ow will result in a positive energy balance. This means that the sign of the energy unbalance can be used for fault isolation if it is combined with additional information. 5. EXPERIMENTS The introduced fault detection methods based on the estimated residual can detect a number of dierent faults in the coal mill. In practice only data of one given fault type is available. It is a fault where the coal inlet pipe has been stopped by the raw coal. The sample frequency in this experiment is 60s. By visual inspection of the data from the coal mill the beginning of this fault can be detected at sample 1115. The estimated residual Q[n], can for this given fault be seen in Fig. 5. The other suggested method is to normalize the residual by the energy ow due to the coal ow, mc [n]CQ[n]s T [n]) . This c (T normalization has the eect that variations due to process variations is out compensated, and it is hereby much more certain that a given detection is due to a fault in the coal mill and not a disturbance e.g. in terms of variations in the coal ow. The detection signals of both detection methods are illustrated in Fig. 7, from where it can be seen that method 1 detects the fault earlier than the visual inspection did, at sample 1110, and that method 2 did detect the fault 1 sample later than the visual inspection, i.e. 1116. 6. CONCLUSION A simple energy balance model is formed of a coal mill with the purpose of monitoring the coal mill, for detecting faults in the mill. A state is introduced representing the fault. The monitoring problem can be viewed as a system with an

The dierences between these two methods are that the rst method reacts faster on a variation in Q[n], however, a detection based on method 2 is more certain to actually be a detection of a fault, and not due to a variation in the coal ow or in the moisture content in the coal. For both methods the threshold is found, such that it gives a detection of the beginning of the fault as early as possible, with the constraint that false detection of faults are avoided.

1 observed

REFERENCES Chen, Jie and R. J. Patton (1999). Robust modelbased fault diagnosis for dynamic systems. rst ed.. Kluwer academic publishers. Fukayama, Y., K. Hirasawa, K. Shimohira and H. Kanemoto (2004). An adaptive state esitmator for pulverizer control using moments of particle size distribution. IEEE Transaction on Control System Technology 12, 797811. Hasselbacher, R., M. Lang and G. Lausterer (1992). Reglerentwurf fr ein kohlenmhlenu u modell unter bercksichtingung der veru faharenstechnischen randbedingungen. Automatisierungstechnik 40, 148157. Rees, N.W. and F.Q. Fan (2003). Modelling and control of pulverised fuel coal mills. In: Thermal power plant simulation and control (D. Flynn, Ed.). rst ed. Institution of Electrical Engineers. Tigges, K.D., W. Bischo and T. Steinhage (1998). Walzenschsselmhlen als kompou u nenten moderner feuerungstechnik. VGB KraftwerksTechnik 78, 7788. Zhang, Y.G., Q.H. Wu, J. Wang, G. Oluwande, D. Matts and X. X. Zhou (2002). Coal mill modeling by machine learning based on onsite measurements. IEEE Transactions on Energy Conversion 17, 549555.

2 energy 3 4 5 6 1100

1150

1200

1250 1300 Samples [n]

1350

1400

1450

Figure 6. A plot of the signal in fault detection method 2. mc [n]CQ[n]s T [n]) . c (T

Method 1 Method 2 1

0.8

0.6 fd[n] 0.4 0.2 0

0.2 1100

1105

1110

1115 Samples [n]

1120

1125

1130

Appendix A. OPTIMAL UNKNOWN INPUT OBSERVER A necessary and sucient condition for the existence of a solution to the given observer problem is in (Chen and Patton 1999) given as: An optimal unknown input observer solution exists if and only if: rank (Cn+1 En ) = rank (En ). The computation of the matrices in the observer is also given in (Chen and Patton 1999) as: (1) Set initial values: P0 = P (0), z0 = x0 + C0 E0 (C0 E0 ) y0 , H0 = 0. + (2) Compute Hn+1 = En (Cn+1 En ) . (3) Compute 1 T T 1 , K1 n+1 = An+1 Pn Cn Cn Pn Cn + Rn (4) and Pn+1 = Pn K1 Cn Pn A1 n+1 n+1 Compute Tn+1 = I Hn+1 Cn+1 , Fn+1 = An Hn+1 Cn+1 An K1 Cn , n+1 1 K2 n+1 = Fn+1 Hn , and Kn+1 = Kn+1 + 2 Kn+1 . Now compute z[n + 1] = Fn+1 z[n] + Tn+1 Bn u[n] + Kn+1 y[n] and x[n + 1] = z[n + 1] + Hn+1 y[n + 1]. T , Compute Pn+1 = Pn K1 Cn Pn A1 n+1 n+1 and following T +Tn+1 Qn TT + Pn+1 = A1 Pn+1 A1 n+1 n+1 n+1 T Hn+1 Rn+1 Hn+1 . Set n = n + 1 and jump to step 2.

Figure 7. A plot of the fault detection based on the two proposed methods. unknown input and disturbances and measurement noises. An optimal unknown input observer is subsequently designed. Thresholds are chosen such that they do not result in false detection and at the same time detect the fault as early as possible. The method is following applied to a fault example where the coal inlet pipe is stopped by the coal. Using the estimated residual results in a detection of fault 5 samples (5 minutes) earlier than a detection based on visual inspection. In addition a normalized residual is used for the detection as well. This method, however, results in detection of the beginning of the fault 1 samples (1 minute) later than the visual inspection did. The proposed observer based method has showed to be a promising method for detecting faults in coal mills.

(5)

(6) 7. ACKNOWLEDGMENT The authors acknowledge the Danish Ministry of Science Technology and Innovation, for support to the research program CMBC (Center for Model Based Control), grant no2002-603/4001-93.

(7)