On-line neuro-expert monitoring system for Borssele Nuclear Power Plant

  • Published on

  • View

  • Download


Pergamon www.elsevier.com/locate/pnucene Progress in Nuclear Energy, Vol. 43, No. 1-4, pp. 397-404, 2003 Available online at www.sciencedirect.com 2003 Published by Elsevier Science Ltd Printed in Great Britain SC~E.CI(.~DJ~mCT 0149-1970/03/$ - see front matter doi: 10.1016/S0149-1970(03)00051-9 ON-LINE NEURO-EXPERT MONITORING SYSTEM FOR BORSSELE NUCLEAR POWER PLANT K. NABESHIMA a, T. SUZUDO a, S. SEKER b, E. AYAZb,, B. BARUTCU b, E. TORKCAN b, T. OHNO c, K. KUDO c a Research Group for Advanced Reactor System, "Japan Atomic Energy Research Institute", Tokai-mura, Ibaraki-ken, 319-1195, Japan b Electrical Engineering Department, "|stanbul Technical University, ITU", 80626, Maslak, lstanbul,Turkey c Applied Quantum Physics and Nuclear Engineering, "KyushuUniversity", Hakozaki 6-10-1, Higashi-ku, Fukuoka-shi, 812-8581, Japan ABSTRACT A new method for an on-line monitoring system for the nuclear power plants has been developed utilizing the neural networks and the expert system. The integration of them is expected to enhance a substantial potential of the functionality as operators support. The recurrent neural network and the feed-forward neural network with adaptive learning are selected for the plant modeling and anomaly detection because of the high capability of modeling for dynamic behavior. The expert system is used as a decision agent, which works on the information space of both the neural networks and the human operators. The information of other sensory signals is also fed to the expert system, together with the outputs that the neural networks generate from the measured plant signals. The expert system can treat almost all known correlation between plant status patterns and operation modes as a priori set of rules. From the off-line test at Borssele Nuclear Power Plant (PWR 480 MWe) in the Netherlands, it was shown that the neuro-expert system successfully monitored the plant status. The expert system worked satisfactorily in diagnosing the system status by using the outputs of the neural networks and a priori knowledge base from the PWR simulator. The electric power coefficient is simultaneously monitored from the measured reactive and active electric power signals. 2003 Published by Elsevier Science Ltd. KEYWORDS on-line plant monitoring, neural network, expert system, pressurized water reactor (PWR), anomaly detection, plant diagnosis, electric power coefficient. 397 398 K. Nabeshim,; e; al. 1. INTRODUCTION A real-time condition monitoring at nuclear power plants (NPPs) is one of the most important tasks lbr operational safety. Conventional monitoring methods in the present NPPs can detect anomalies ~hen the monitored signals exceed their error boundary. However, it is difficult to detect the symptom of anomalies with this method because of the wide error boundary covering from zero to full power operation. Therefore, we proposed a neuro-expert methodology that is more preferable than the threshold-level-based one tbr early fault detection. The main purpose of this monitoring system is to complement the conventional alarm system and to support operators. Neural network techniques already have been applied to plant monitoring and shown good pertbrmance for early fault detection (Upadhyaya et al, 1992) (Nabeshima et al, 1998). However, the neural network itself can merely detect a deviation from the normal state, and requires an interpretation of the deviation by an expert to diagnose the cause. On the other hand, establishing independent expert systems lbr plant monitoring involves too many complicated tasks such as collecting knowledge and rules about plant design. This motivates the integration of neural networks and an expert system for plant monitoring. 2. MONITORING SYSTEM OVERVIEW The structure of this monitoring system is shown in Fig. 1. The Borssele NPP represented on the left of the figure is a two-loop pressurized water reactor with nominal electric power output of 477 MWe. The on-line data acquisition system sends 72 plant signals to the neuro-expert system every two seconds. Out of these, 21 most significant plant signals are selected for the inputs of neural network. The other signals are unchanging during normal power increase and decrease operation from low to full power, so that the anomaly of those signals can be easily detected by the conventional monitoring method. Recurrent and feedforward neural networks in auto-associative mode train with plant's normal operational data, then model the plant dynamics. The expert system diagnoses the plant status with the measured signals, the output of neural networks, the alarm signals from the conventional alarm system, and the operation information from the operators. The software of neural networks and the expert system are programmed in FORTRAN and executed on the PC. The advisory displays show the status of NPP diagnosed by neuro- expert system. The graphical advisory displays are programmed by Java language, so that the monitoring results can be displayed on any computers connected to the Internet (Suzudo et al., 2002). ( . . . . . . . . . . PWR S imula tor . . . . . . ,~ S ignal f . . _ t / _ _ _ ~ Z~ i ,...- P lant S ignals , I S,eam~,o* / / I I . . . . . . . . . . ~ . . . . I I , , . . . . . . I I Steam Press =leam Press I FeedV',a er Flow FeedWater Flow I ~eedW i reedWae, meJ , Ir I I I I] ' I Cotd-Le Tern Cold-Leg Tern p I , : Co~. t // Reactor C~oofan, I Pump 1 Vessel Pump-2 ,,, / Ana log ~-- - - Neuro -Exper t Sys tem - ~ -.., J Fig. l Monitoring System On-line neuro-expert monitoring system 399 3. PLANT MODELLING BY NEURAL NETWORKS The neural networks are trained by the current and past system inputs and outputs, so that it can predict the next outputs of the system. This is direct analogy with the concept of one-step-ahead prediction and can be effectively implemented by ANN, and applied to dynamic tracking. The basic principle of the anomaly detection is to monitor the deviation between process signals measured from the actual plant and the corresponding values predicted by the plant model, i.e., the neural networks. If one of the deviations exceeds the fault level, the error message will be displayed on the screen and the error log-file with the time and signal information will be written. In this application, the fault level during steady state operation is defined experimentally as 1.25.e; the maximum error ~ is the largest deviation during the initial learning. On the other hand, the fault level during transient operation is defined as 1.45.e. If the deviations between measured and estimated values are small enough, the plant is considered to be operated normally. In such cases, the neural network is adaptively trained with certain number of previous data at the adaptive learning stage. In the actual application, the latest 20 of previous points are used for the adaptive learning at each time step, and newer point data must be learned linearly more than older points. For example, newest point is trained 20 times, second newest is 19 times ...etc (Nabeshima et al, 1998). The reactor operational condition always changes because of several factors sucfi as a fuel-burn-up state, and the reactor condition. Thus the dynamics at the beginning of the fuel core cycle are completely different from those at the end of the cycle. As a result, the same model cannot be used during the whole fuel core cycle. Therefore, the adaptive learning used here can gradually change the network model to follow the actual plant status by updating the weights. The feedforward neural network has three layers: input, one hidden and output layer. The numbers of input and output nodes are 21, respectively. Here, the output signals are supposed to be the same as the input signals at the same or next time step. The number of hidden node is selected as 25 because the sum 400 K. Nabesh ima et al. Figure 3 obviously shows that the initial learning is sufficient, because the measured signals (solid line) and the predicted values (circle) overlap to each other. The dotted line shows the deviation between two. The fault severity levels at testing stage are defined the maximum error during the initial learning. These fault severity levels indicate the accuracy of the modelling by the neural network. The initial learning with the recurrent neural network also shows the similar result, although the errors of some signals are larger than the errors by the feedforward neural network. Therefore, the results by the feedforward neural network are represented in this paper. 500 400 V l,.,.. 300 0 13_ 200 O .~ 100 0 o 0 3O 2O 10 i i I i i I o t~ - - Measured _ -10 .~ 0 Predicted - . . . . . . Deviation ~ -20 ' ' ' ' " ' ' ' ' ' '8 o' ' i o 'o6 ' ' ' ' ' ' ' . . . . 30 200 600 1200 1400 1600 Time ( x l0 4 s) V c- O Fig. 3 Initial Learning Result 4. DIAGNOSIS BY EXPERT SYSTEM The expert system is used as a decision agent that works on the information space of both neural networks and human operators. The information of other sensory signals is also fed to the expert system, together with the outputs that the neural networks generate from the plant signals. A major advantage of an expert system is to process a lot of operator's knowledge and to derive useful information for complex decision environments. The expert systems can treat almost all known correlation between plant status patterns and operation modes as a priori set of rules. The database of anomaly detection patterns by neural networks are created using PWR simulator because it is impossible to get a lot of anomaly data from the real NPP. If we assume that the response of the neural network during anomaly is similar in the same type of PWR, we can utilize the simulation data from the other PWR plant. Using the compact simulator of Surry-l, U.S.A, many kinds of malfunctions caused by equipment failure during steady state and transient operation were simulated for the purpose of the testing. The time interval of the simulation is 2 seconds. The neural network initially learns the plant modelling with the normal operation data, same as in the previous section. After the initial learning, the neural network tests the anomaly data, and the responses of the network are stored in the database. The anomaly detection time and channels during many anomaly cases are shown in Table 1. Those malfunctions are much smaller than those for the significant accident cases. Most of the testing results indicated that the neural networks could detect the symptom of small anomaly much faster than the conventional alarm system embedded in the plant simulator. The malfunction case of "Small Reactor Coolant System Leak" is small leak caused by the coolant boundary failure in Reactor Temperature Detector Well. If the leakage was larger than 18.9 l/min, the neural network detected this anomaly with the deviation of pressurizer level (Ch.6) exceeding the fault 48.0 On-line neuro-expert monitoring system ~ o o y T T [ 1 3.0 alfunction Started (1:00) ooooooooooooooooooooooooooooo~ 2.0 ! . . - , - . . . . . . . . . - . . . . . . . . . . . . . . . . . , . . . . . . . . 1.0 .op " - -~ . . I 46.0 ~" w E n" . 7 ' ,,, . ;_.{..= o.o o N - - ~: 42.0 t ~ Fault Level I-- ~) Anomaly Detected (1:18) -1 .0 ~-~ 03 [, ] > (/) ~ Measured : W UJ 40.0 l ~ Predicted ~ -2.0 3 r'v' l . . . . . . . . . Deviation n I, i 38.0 I . . . . . 1 L . . . . . . . -- . . . . . . . . . . . . . . . . J _3. 0 0 2 4 6 8 10 T IME (Min.) Fig. 4 Pressurizer Level Signal (Ch.6) during Small Reactor Coolant System Leak 401 severity level owing to rapid decrease of pressurizer level. Figure 4 shows the monitoring result of pressurizer level signal in the case of 56.7 l/min leakage. The deviation (dotted line) between measured signal (solid line) and predicted value (circle) exceeded the fault level (broken-dot line) at 18 seconds after the leakage started. The fault of feedwater pressure signal (Ch. 16) was detected secondarily. "Leakage of atmospheric steam dump valve" causes rapid decrease of steam flow and pressure which is followed by the average steam temperature decrease and the power decrease. To supplement the loss of power, the control rod was driven out by control system. However, no alarm was given by the conventional alarm system even after 10 minutes if the leakage of valve capacity is less than 50%, because over reactor power by rod withdrawing caused only small shift of power balance. The neural network immediately detected the anomaly of steam flow signals (Ch.I l&12), as shown in Fig. 5. Secondarily, the signals of neutron flux and feedwater flow exceeded the fault severity levels. O ...J I1 402 K. Nabeshima et al. However, the deviation of the feedwater pressure signal (Ch.16) never exceeded the fault severity level in the cases of anomalies at pressurizer. The anomalies of Turbine Governor Valve showed the early fault detection at turbine impulse signal, feedwater flow, steam flow and electric power, because this valve is dominating factor for power controller. The last five cases in Table 1 are some kind of controller failures, so that the conventional alarm system detected those failures immediately. However, it is difficult to identify the cause of those anomalies by the alarm information. On the other hands, the neuro-expert system can detect them at the next time step and identify them as the controller failures because only the deviation of failed signal was much lager than the others. Table 1 - Anomaly Detection Time and Channels No. Malfunction Small Reactor Coolant System Leak (56.7 I/min) Leakage of Atmospheric Steam Dump Valve (5%) Partial Loss of Feedwater 3 (90.7 ton/hr) 4 Pressurizer Spray Control Valve Fails Open Both Pressurizer Spray Control Valve Fails Close 6 Backup Pressurizer Heater Fails On One of the Turbine Governor Valves Fails Open Turbine Governor Valves Fails 8 Closed 10 Volume Control Tank Level Control Fails Low Volume Control Tank Level Control Fails High Steam Generator Level Control 11 Fails High Steam Generator Level Control 12 Fails Low 13 Detection Channel No. First Ch.6 (0:18) Ch.11,12 (0:02) 3h.16 (0:04) Ch.6 (0:10) Ch.6 (0:24) Ch.6 (0:22) Ch.8,22 (0:02) Ch.8,11,12, 22 (0:02) Ch.7 (0:02) Dev. 2.31 Ch.7 (0:02) Dev. 2.33 Ch.14 (0:02 ~, Dev. 0.85 Ch.14 (0:02 Dev. 0.35 Second Ch.16 (2:26) Ch.13,14 (o:o4) Ch.13,14 (0:10) Ch.22 (1:04) Ch.8 (3:32) ! Ch.22 (2:56) Ch.10, 11,12 (o:o4) The Others (0:04) Ch.12 (0:14) Dev. 0.01 Ch.1,2,3, 12 (0:14) Dev. 0.22 Ch.1-4,10,11,13, 16,21,22 (0:04) Dev. 0.28 Ch.16,10,13, 16,21,22 (0:04 Dev. 0.02 Third Ch.8,22 (3:32) Ch.2,4 (0:06) Ch.2,3 (0:32) Ch.8 (1:24) Ch.22 (3:36) Ch.8 (3:14) Ch.13,14 (0:06) Ch.1,2,3 (0:20) Others Others Ch.1,3,4... Others (0:04) Ch.5 (0:02) Ch.8,11,12, Temperature Failure High in 16,21,22 (0:14) Others Cold Leg Dev. 0.81 Dev. 0.01 - Ch. No. Without Alarm Ch.9 Ch.5,9,10, 17,18 Ch.5,6,9,10, 13,17,18 Ch.9,10, 13 -17,19,20 Ch.5,6,9,11 -14,17,18 Ch.9 Conventional Alarm No Alarm No Alarm No Alarm 0:34 Pressurizer Press. Low 3:59 Pressurizer Press. High No Alarm No Alarm 0:00 VCT Level High 0:00 SG Level Error 1:40 Turbine Trip 0:00 SG Level Errol 1:02 SG-C Alarm Rod Stop 0:02 Tav/Tref Deviation 0:00 VCT Level Low 0:02 High Steam Line Flow On-line neuro-expert monitoring system 403 It is clear from Table 1 that the patterns of anomaly detection channels from the discrepancy between the predicted and actual outputs depend on the kind of anomaly. Furthermore, the monitoring results during steady state and transient operation showed almost same anomaly detection property. Therefore, the intbrmation of detection signal, time and deviation is very useful to identify anomalies by rule-based expert system. From the above results, the expert system can easily diagnose the plant status and identif} the type of anomalies by using the output of neural networks and the simple "if-then" rules in FORTRAN. The diagnosis process by the neuro-expen system and the simple expert system called DISKET is compared in Table 2 (Yokobayashi et al., 1991). DISKET diagnoses the plant status with the intbrmation from the conventional alarm system. DISKET identified the leakage from the primary loop three minutes after the malfunction, and SG tube rupture four minutes after. On the other hands, the neuro-expen system could identified SG tube rupture six second after, because the neural network detected this anomaly vdth the pressurizer level and SG flow and the detected pattern of these two signals shows "SG tube rupture" only. The neuro-expert system can easily identify the anomaly because the influence of anomaly had not propagated to the neighbours at the beginning. Table 2 - Plant Diagnosis at SG Tube Rupture DISKET (expert system) Neuro-expert System 0' 00" Radiation Leak 1' 00" Pressurizer Level Low 2' 00" Pressurizer Press. Low 3' 00" Leak from Primary Loop 4' 00" SG Tube Rupture 0' 00" Radiation Leak 0' 04" Pressurizer Level Error 0' 04" SG Flow (C) Error 0' 06" SG Tube Rupture Referenced Rule : 175 Referenced Rule : Number of Anomaly Cases Fitted Rule : 19 Fitted Rule : 1 5. FOLLOWING REACTOR POWER VIBRATION BY NEURAL NETWORKS In electric power system technique, there are three kinds of powers: active power (P), reactive power (Q) and complex power (S) respectively. They are given in the following relations; S = sqrt (P2 + Q2), cos 0 = P /S . Using the relationship between the active and reactive power, factor cos 0 can be defined as above and it is named as "power factor'. Also, same factor is used to present reactive power compensation in the system using an external reactive power source because the increasing in the reactive power usage causes to voltage drops in an electric power system. Or, decreasing the angle between current and voltage (0--)0~'), it increases the effectiveness of active power. For this reason, monitoring the power factor variations in electric energy systems plays a very important role in terms of electric energy quality. According to Figure 6, factor cos 0 follows the reactive power variations. But rinsing operation is observed in active power variations between the patterns of 4000 and 5000. A small increasing in the reactive power level is determined while the active power variation has a small decreasing at the beginning of the rinsing operation and the power factor cos 0 also compensates the reactive power variations in the same data region. Hence the increasing of reactive power in the system can be prevented with this way. In the figure, cos 0 followed the modelling by feedforward neural network as well as two recurrent neural networks (Elman and Jordan). In this example power operation of Borssele NPP is followed from 350 MWe to full power operation of 480 MWe. While the cos 0 at continuous power operation is around 0.98, the error of the three ANN results are indicating negligible deviations. 404 K. Nabeshima et al. 500 450 ; 41 ~,. t 400 .~" " Rinsing Training 350 J '# Operation 200 >. 100 ~ ' - - "~ 0 o E lman nnet results C j , , 0 .9 ~ " ~" - " . . . . . ~ Jordan nnet results ..~ 0.04 ~ 0~; ~L . . . . . . . . ~ . . . . . . . . . . . . . , , ~ . . , ~1-0.04 Feedforward nnet results ~ 0.04 i, "~ 0 . . . . ~ ~' -004 0 1000 2000 3000 4000 5000 6000 7000 Fig. 6 Cos ) Calculations by Neural Networks 8000 i 0.06 10.04 002 '~ -0.02 -O.O4 6. CONCLUSIONS The neuro-expert monitoring system has been applied to Borssele NPP and the on-line PWR simulator. From the off-line and on-line test results, it was shown that the neural network successfully detected the symptoms of anomalies in the early stage. The expert system correctly recognized the operation modes, and diagnosed the plant status, because the detection patterns by the neural networks are dependent on the kind of anomalies, and independent of the amount of anomaly or operation modes. REFERENCES Jordan M. I. (1986), Attractor Dynamics and Parallelism in a Connectionist Sequential Machine, Proceedings of the 1986 Cognitive Science Conference, 531-546. Nabeshima K., Suzudo T., Suzuki K. and Tfirkcan E. (1998), Real-time Nuclear Power Plant Monitoring with Neural Network, Journal of Nuclear Science and Technology, Vol. 35, 93-100. Rumelhart D. E., McClellan& J. L. and the PDP Research Group. (1986), Learning Internal Representation by Error Propagation, Parallel Distributed Processing, Volume 1, MIT Press, 318-362. Suzudo T., Nabeshima K. and Takizawa H. (2002), Software Integration for Monitoring System with High Flexibility - Application to ANNOMA System -, submitted to 8 ~h Symposium on Nuclear Reactor Surveillance and Diagnostics, G6teborg, Sweden. Upadhyaya B. R. and Eryurek E. (1992), Application of Neural Networks for Sensor Validation and Plant Monitoring, Nucl. Tech.. Vol.97, 170-176. Yokobayashi M. and Yoshida K. (1991), Comparison of capability between two versions of reactor transient diagnosis expert system 'DISKET' programmed in different languages (in Japanese), Nippon-Genshiryoku-Gakkai- Shi., Vol.23, 695-702.


View more >