IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 19, NO.

1, MARCH 2004 179

Predictive Maintenance in
Intelligent-Control-Maintenance-Management
System for Hydroelectric Generating Unit
Chuang Fu, Luqing Ye, Yongqian Liu, Ren Yu, Benoit Iung, Yuanchu Cheng, and Yuming Zeng, Member, IEEE

Abstract—The predictive maintenance within the framework of an industry, the costs of maintenance account for 15%–40%
of intelligent-control-maintenance-management system (ICMMS) of the whole production, and one-third of which is caused by
makes full use of all the information of control, maintenance, and unnecessary and inaccurate maintenance [4], [8]. Therefore, ad-
technical management aspects to make right maintenance at the
right time in the right place. In this paper, the three key elements vanced maintenance method should be implemented. In recent
of the predictive maintenance within the framework of ICMMS are years, with the rapid development of condition monitoring and
presented. The ICMMS platform for hydroelectric generating unit, forecasting, information processing, fault detecting, and artifi-
especially its maintenance function, is introduced. An artificial- cial intelligence technology, it is possible and feasible to mon-
neural-network (ANN)-based identification and diagnosis model is
set up to implement the predictive maintenance of the electrohy-
itor and forecast equipment condition and assess its health on-
draulic servomechanism in the hydroelectric generating unit. The line. In ICMMS, the three “isolated islands” of the automation
tests show that the proposed strategy can guarantee ideal perfor- system are integrated into an organic system, and maintenance
mance. will be improved by sharing information among these three.
Index Terms—Artificial neural network, failure model, hydro- In this paper, the predictive maintenance within the frame-
electric generating unit, intelligent control-maintenance-manage- work of ICMMS is presented. In Section II, the inherent de-
ment system, predictive maintenance. fects of breakdown maintenance and scheduled maintenance are
analyzed first, then the concept of predictive maintenance and
I. INTRODUCTION its three key elements, namely, monitoring and forecasting, di-
agnosis and prognosis, and maintenance decision-making are

M ODERN automation system mainly includes three tech-

nical areas, namely, control, maintenance, and technical
management. The aim of control is to achieve good quality and
presented. In Section III, the ICMMS platform for hydroelec-
tric generation unit, especially its maintenance function, is pre-
sented. In Section IV, an artificial neural network (ANN)-based
performance. As for maintenance, high reliability and avail-
identification and diagnosis model is outlined, and its applica-
ability are the goals. In the aspect of technical management,
tion to the electrohydraulic servomechanism of the water turbine
the efficiency, which includes economic and social benefits, is
governing system is introduced. In Section V, the conclusions
the everlasting objective. These three are interconnected and
are drawn and future work is pointed out.
interacted. In order to meet the comprehensive needs of per-
formance, reliability, and benefits, they should be studied as
a whole, and the intelligent-control-maintenance-management II. PREDICTIVE MAINTENANCE
system (ICMMS) has emerged with these requirements [1]–[5]. Both breakdown maintenance and scheduled maintenance
Nowadays, the development of control, maintenance, and tech- are out-of-date, and new maintenance method that accords with
nical management is out of synch. The research and application failure development procedures must be studied and applied
of control and technical management are more advanced than to the industrial areas, especially to the ICMMS of automation
that of maintenance. For example, the scheduled maintenance system. The goal is to replace “postmortem” and “blind”
is still the main method currently adopted by most hydropower maintenance with “just in time” maintenance.
plants and other industrial sector [4], [6]–[8]. Because of the
blindness of the scheduled maintenance, the breakdown mainte- A. Limitations of Breakdown Maintenance and Scheduled
nance cannot always be avoided. According to the investigation Maintenance
The breakdown maintenance is a primitive maintenance
Manuscript received July 27, 2002. This work was supported by the European
Union under Grants DC 961744 and IST-2000-28739.
method and often with great economic loss. It is only suited
C. Fu, L. Ye, and R. Yu are with the College of Hydropower and Digital to noncritical areas under the following preconditions: the
Engineering, Huazhong University of Science and Technology, Wuhan 430074, consequences of failure are slight, there is no safety risk, the
China (e-mail: [email protected]).
Y. Liu and B. Iung are with Centre de Recherche en Automatique de Nancy, failure will be identified quickly, and the repair will be quick.
University of Nancy I, Nancy B.P. 239-54506, France. Nevertheless, these preconditions nearly could not be met
Y. Cheng is with the Department of Power Engineering, Wuhan University, simultaneously in modern industrial areas.
Wuhan 430072, China.
Y. Zeng is with Nextron Corporation, Calgary, AB T2H 2H4, Canada. The scheduled maintenance is based on the following
Digital Object Identifier 10.1109/TEC.2003.816600 assumption: the failure behavior of all equipment possesses
0885-8969/04$20.00 © 2004 IEEE
180 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 19, NO. 1, MARCH 2004

Fig. 1. Effects of scheduled maintenance.

Fig. 3. Scheme of predictive maintenance.

according to the actual operating conditions and experiences.

Then the appropriate diagnostic tools and methods should be
chosen to measure/compute the indicators and to predict the
trends. These values of the indicators are compared with the
thresholds, so the rate of degradation can be determined, and
the time to failure can be predicted from the trend. Finally, the
maintenance decision can be made.
Fig. 2. Failure models. The predictive maintenance has the ability to assess the cur-
rent health, predict the future health of the equipment for a pre-
the bathtub curve (Fig. 1). However, the reliability theory has determined time horizon, or predict the time to failure. Fig. 3 is
proven that scheduled maintenance is just suitable for failures the schematic diagram of the predictive maintenance. Assume
that have a clear wear-out characteristic, but unsuitable for the equipment is normal at time , three possible health
random failure [4], [8]. If the maintenance takes place in the trends are represented by curve 1, 2, and 3, respectively. Curve
wear-out phase, it will reduce the failure rate (Fig. 1). This is 1 shows that the equipment health index is above the normal
an ideal case. In fact, the maintenance period is shorter than the baseline at both time and . Curve 3 points that the equip-
ideal one for the safety reason. In other words, the maintenance ment health index is under the normal threshold at time , some
often takes place in the constant failure rate phase. After failure has occurred, the equipment must be shut down, and
maintenance activities, the bathtub curve moves horizontally to maintenance actions must be carried out immediately. Curve
right (Fig. 1), and the equipment is forced to enter into a new 2 depicts that the equipment health index is above the normal
wear-in phase and a higher incipient failure rate is brought into threshold at current time , but will be under it at time
the system. Moreover, many types of equipment do not obey and certain latent failure may occur, so maintenance should be
the typical bathtub wear-out model. To this day, seven failure predicted and arranged in advance.
models [6] have been found (Fig. 2). It is seen from Fig. 2 that
the failure rates of most of failure models (from C to F) are C. Three Key Elements of the Predictive Maintenance
invariant or grow very slowly. In other words, there is not an The predictive maintenance consists of three key elements:
explicit wear-out period, and random failures play a dominant monitoring and forecasting, diagnosis and prognosis, and main-
role. tenance decision-making.
• Monitoring and forecasting: The scope of monitoring
B. Predictive Maintenance and forecasting includes all states of the whole system.
The scheduled maintenance is time-oriented, whereas the From the discussion in the above section, the conditions
predictive maintenance is actual condition and trend-oriented. of the system can be represented by certain indices. The
Most of failures do not occur instantaneously, and usually purpose of monitoring and forecasting is to monitor
there are some kinds of degradation process or symptoms these indicators and to analyze their trends online. After
from normal states to failures. Therefore, the actual conditions comparing the indices with the references, the equipment
and their trends should be assessed and predicted during the health can be assessed. In a similar way, comparing the
degradation process, and appropriate maintenance actions forecasting value with the reference, the predictive judg-
should be taken before breakdown. This is the main target of ment about the equipment health can be made. The online
the predictive maintenance. condition monitoring and forecasting is the foundation of
To achieve above-mentioned goals, the mechanisms of the predictive maintenance.
equipment failure and deterioration should be identified, • Diagnosis and prognosis: The scope of diagnosis and
and the suitable indicators for equipment health should be prognosis includes all failures and degradations. Based on
determined. These indicators may be direct sensing data, but the monitoring and forecasting results, the states of the
in many cases, they may be indirect indices derived from equipment, namely, normal, degradation or failure, can be
several direct sensing data. Moreover, the thresholds should be identified. Based on the states of failure and degradation,
identified. These indicators and thresholds may be modified the cause and location of failure are identified. Similarly,
FU et al.: PREDICTIVE MAINTENANCE IN ICMMS FOR HYDROELECTRIC GENERATING UNIT 181

according to the observations and nature analysis of the

degradation dynamic process, the effects of degradation
on relative equipments and the whole system are identi-
fied. Moreover, the time of failure occurrence can be pre-
dicted.
• Maintenance decision-making: The scope of main-
tenance decision-making includes the maintenance
strategies and actions, and their arrangement as well.
According to the results of monitoring and forecasting,
diagnosis and prognosis, the following judgments have
to be made: Whether the maintenance is necessary? If
it is, when it should be carried out? what material and
personnel resources have to be provided? and how the Fig. 4. Configuration of ICMMS platform.
maintenance actions would be adopted and arranged?
Obviously, actual constraints must be taken into account • Maintenance and Technical Management Station:
for the maintenance decision-making. The maintenance of the hydroelectric generating system,
These three key elements supplement each other, so that they especially which of the water-turbine governing system,
should be integrated into a global and consistent predictive has been performed. The technical management aspect
maintenance within the framework of ICMMS. involves the performance assessment, operation arrange-
ment of the unit, and so on. Combined with the results
of monitoring and forecasting, diagnosis and prognosis
III. ICMMS PLATFORM FOR HYDROELECTRIC of the intelligent electrohydraulic servomechanism, the
GENERATING UNIT proper maintenance decision is made. In addition, the
dynamic motion and failure status of the electrohydraulic
In recent years, the computer supervisory and control system, servomechanism are graphically displayed in real-time.
management information system, and computerized mainte- • Unit Simulating Platform: It is impossible to do all tests
nance management system are gradually applied in hydropower under various operating conditions in field, so the simula-
plants, respectively. However, they almost are isolated from tion studies are very important. There are different types
each other. As for the maintenance in hydropower plants, the of water-turbines in the unit simulating platform. The
scheduled maintenance is still the main method [4], [6]. dynamic responses of the processes, including startup,
In order to validate and show the advantages and feasibility synchronization, frequency (or speed) regulation, load
of ICMMS technology in a practical engineering environment, increasing/decreasing, load rejection, shutdown, etc. are
a showcase, that is, an ICMMS platform for hydroelectric gen- displayed.
erating unit, has been designed and set up in our laboratory. The • Intelligent Electrohydraulic Servomechanism: Elec-
control, maintenance, and technical management aspects in the trohydraulic servomechanism is very complicated. In
automation system for a hydroelectric generating unit are inte- order to improve its performance and reliability, a good
grated into an ICMMS in this platform. Here, special attention is way is to make it intelligent. In existing analog and digital
paid to predictive maintenance strategy for the electrohydraulic governing systems, the connection between controller and
servomechanism of the governing system. electrohydraulic servomechanism is always via analog
The ICMMS platform for electrohydraulic servomechanism signals. Whereas in the ICMMS platform, both controller
(Fig. 4) is designed according to the philosophy of function and and electrohydraulic servomechanism are treated as
intelligence distribution [9]. The higher layer includes control intelligent nodes of the Profibus network [9] (Fig. 4).
station, maintenance, and technical management station; they The conditions of the electrohydraulic servomechanism
are constructed by industrial computers with high performance; are monitored and forecasted in real-time. Some kinds
the lower layer consists of two parts: intelligent electrohydraulic of failure/degradation, such as trembling coil breakage,
servomechanism that is an intelligent actuator of the controller; zero-point creeping, and mechanical backlash, can be
and unit simulation platform that can simulate different types of diagnosed and prognosed. Failures and emergencies are
water-turbine. dealt with immediately. In addition, all of these messages
• Control Station: The control station is designed as a con- are sent to the maintenance station for further processing.
troller for water-turbine governor, and it is mainly used for
the research of the control law for the hydroelectric gen- During the design and implementation of the predictive main-
erating unit. All of the parameters can be adjusted easily, tenance in the ICMMS platform for the hydroelectric generating
such as (proportional gain), (integral gain), and unit, the following procedures are adopted. Based on the infor-
(differential gain) for the proportional-integral-differen- mation availability, define the maintenance user requirements.
tial (PID) controller. The behaviors of unit frequency (or Map the requirements to technical maintenance in order to en-
speed), network frequency, output of the controller, actual sure that the distributed system satisfies the required mainte-
stroke of the main servomotorm, etc. are shown graphi- nance [3], [4]. The maintenance functions are distributed to field
cally and in real time. devices and maintenance station.
182 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 19, NO. 1, MARCH 2004

Fig. 5. Scheme of electrohydraulic servomechanism of governing system.

Fig. 7. Responses of electrohydraulic servomechanism and ANN.

Fig. 6. Serial-parallel/parallel identification and diagnosis model based on

ANN.

IV. IMPLEMENTATION OF PREDICTIVE MAINTENANCE IN THE

ICMMS PLATFORM FOR HYDROELECTRIC GENERATING UNIT Fig. 8. Error between responses of electrohydraulic servomechanism and
A simplified structure of electrohydraulic servomechanism ANN.
is shown in Fig. 5, and it can be simplified as a second-order
system with several nonlinear components [4]. After a
lengthy period of operation, a series of knowledge concerning
failures/degradations has been gathered. Together with the
existing knowledge about frequent failures, failures can be
more precisely identified, and their treatment will be better.
In addition, the study in ANN indicates that the forward
neural networks with three layers can be applied to the system
dynamic identification [4], [10]–[12]. Therefore, during the
design and implementation of the predictive maintenance of
the electrohydraulic servomechanism, an ANN-based approach Fig. 9. Scheme of monitoring and forecasting, diagnosis, and prognosis based
is adopted and combined with historical failure knowledge on ANN.
obtained from field experience.
dorandom signal M-serials or Sinusoidal signals are used as
A. ANN-Based Identification and Diagnosis Model exciting signals. The responses of electrohydraulic servomech-
Fig. 6 can be used for either serial-parallel identification anism and which of the ANN are shown in Fig. 7, and their dif-
model (when the switch S is set on 1) or parallel identification ference is shown in Fig. 8. The inputs are as follows: positive
model (when S is set on 2). According to the requirements of step signal with amplitude of 0.4 (at 5 s); ramp signal with a
the ANN identification, the serial-parallel identification model slope of 0.02 (at 15 s); negative step signal with amplitude of
is used in the ICMMS platform to realize dynamic identifi- 0.3 (at 25 s); band-limited white noise (at 35 s).
cation of the electrohydraulic servomechanism; the parallel As discussed above, the condition change of the electrohy-
identification model is used in emergency case to treat some draulic servomechanism is reflected through corresponding in-
kinds of failures. A 4 7 1 forward ANN has been adopted, dicators, which may be direct sensing values or indirect values
and its main characteristics are as follows: simple architecture, derived from several direct sensing data. Here, the direct de-
speedy approximation capability, and easy implementation. tected values include the feedback coil current, working coil
The simulation studies have proven this ANN can meet the current, trembling coil current, oil pressure, oil level, oil temper-
needs of the dynamic identification of the electrohydraulic ature, and so on (Fig. 9); the indirect indicators are mainly ob-
servomechanism [4]. This ANN has four inputs: U , tained through ANN-based serial-parallel identification model.
U(k), Y and Y(k), where U is the control, Y is the
stroke of the main servomotor, k and signify present B. Condition Monitoring and Forecasting Based on
and previous time. The output of ANN [i.e., ], is the ANN Model
predictive value of the main servomotor stroke. The electrohydraulic servomechanism is dynamically iden-
In order to obtain the initial weighting values of the ANN, tified through ANN in order to monitor its condition online
the physical model is utilized to train this ANN offline [4]. The (Fig. 9). In this case, serial-parallel identification model is ap-
experiments indicate that the results are satisfactory when pseu- plied. The historical values required by ANN are the actual his-
FU et al.: PREDICTIVE MAINTENANCE IN ICMMS FOR HYDROELECTRIC GENERATING UNIT 183

torical feedback values of the electrohydraulic servomechanism.

In every sampling time, the output of ANN is compared with
that of the electrohydraulic servomechanism. There exist three
possible cases which are

(1)
(2)
(3)
Fig. 10. Relationship between steady-state error and feedback factor.
where is the difference between the outputs of ANN and
the electrohydraulic servomechanism; is a degradation
threshold and is a failure threshold for the electrohydraulic
servomechanism, and they are determined according to the
historical knowledge and simulation studies.
Equation (1) indicates that the dynamic characteristics of
ANN are consistent with that of the electrohydraulic servo-
mechanism, and ANN training is not necessary. Equation
(2) shows that there are certain changes in the characteristics
of the electrohydraulic servomechanism, so ANN training is
necessary to follow these changes. As the bounds to be adjusted
are very small each time, the ANN training time is very short.
The simulation results have shown it can meet the requirement Fig. 11. Results of feedback factor creeping test.
of real-time application. After training, the weights of ANN
represent the characteristic information of the electrohydraulic servomechanism. For example, if the feedback coil breaks, the
servomechanism. Equation (3) shows that there are drastic steady-state output attains 100%; if feedback factor creeps, the
changes that occurred in the parameters and/or structure of the error of the steady-state output has continuous and monotone
electrohydraulic servomechanism, namely, abrupt failures. relationship with the creeping value. The relationship between
In the ICMMS platform, other monitored signals include oil the steady-state error and the feedback factor creeping is
pressure, oil level, oil temperature, trembling coil state, unit shown in Fig. 10, as the input is 0.5. Therefore, the error of
frequency, grid frequency, etc. The following processing ap- the steady-state output can be considered as an indicator of the
proaches are adopted: feedback failures.
• discrimination based on parameter limits; In the ICMMS platform, it is assumed that the feedback factor
• discrimination based on parameter change rate; of the electrohydraulic servomechanism decreases at a rate of
• discrimination based on parameter function relationship; 0.003/s, with input of 0.5. ANN has been trained just 54 times
• adaptive filtering weighted moving average forecasting al- during a period of 180 s to follow the characteristic changes of
gorithm. the electrohydraulic servomechanism. The results are shown in
Fig. 11, where Curve 1 is the output of the ANN, Curve 2 is the
C. Failure/Degradation Diagnosis and Prognosis Based on output of the electrohydraulic servomechanism, and Curve 3 is
ANN Model the steady-state error. Based on Figs. 10 and 11, the creeping
The abnormal phenomena can be divided into abrupt failures value of the feedback factor can be identified. The results are
(e.g., actuator blockage, feedback coil breakage) and degrada- shown in Fig. 12, where Curve 1 represents the steady-state
tions (e.g., feedback factor creeping). Generally, degradations error identified by ANN, and Curve 2 denotes the actual value.
have certain processes and symptoms. These failures and degra- The polynomial exponential smoothing forecasting algorithm
dations reflect internal characteristic changes in the electrohy- has been adopted to predict the future creeping value. The re-
draulic servomechanism. According to the above discussion, the sults are shown in Fig. 13, where Curve 1 is the forecasting value
internal characteristics of the electrohydraulic servomechanism of steady-state error, Curve 2 is the actual steady-state error, and
can be identified by ANN. Curve 3 is the error between forecasting value and the actual
1) Degradation Diagnosis and Prognosis: During the value. According to the actual values and their trends, the feed-
normal operation, ANN is not connected with the input of back failures can be diagnosed and prognosed.
the electrohydraulic servomechanism, and a special exciting 2) Abrupt Failure Diagnosis: If drastic changes happen in
signal (pseudorandom signal M-serials) is put into ANN the parameters and/or structure of the electrohydraulic servo-
(Fig. 9). ANN uses historical outputs of the ANN instead of mechanism, the difference between the output of ANN and that
the electrohydraulic servomechanism outputs (i.e. the parallel of the electrohydraulic servomechanism increases suddenly
identification model) [Fig. 6] is applied. In this case, charac- and exceeds (0.05%). Under this condition, emergency mea-
teristic indices can be obtained from the inputs and outputs of surements should be taken. ANN training is stopped at once, and
the ANN. Based on these indices and their trends, degradation ANN takes the form of parallel identification model instead of
diagnosis and prognosis can be performed. This method is serial-parallel identification model (Fig. 6). In this case, the fail-
very useful to identify degradations of the electrohydraulic ures are isolated and the states before failures are conserved in
184 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 19, NO. 1, MARCH 2004

Fig. 12. Identified results of steady-state error.

Fig. 14. Treatment of feedback coil breakage.

Fig. 15. Responses in the case of the blockage failure (close direction).

the output of the electrohydraulic servomechanism will keep

on increasing. Under this situation, the difference between the
output of ANN and electrohydraulic servomechanism increased
Fig. 13. Forecasting results of steady-state error. continuously. According to this result, it can be concluded that
an actuator blockage has occurred and the further emergency
ANN. The controller uses the outputs of ANN instead of the ac- treatments should be taken. What actions should be taken
tual feedback signals of the electrohydraulic servomechanism. depends on the unit operation mode, and further discussion will
After the failure is eliminated, ANN returns to follow the char- be presented in the next section. In this scenario, no action was
acteristics of the electrohydraulic servomechanism. taken, and the output of the electrohydraulic servomechanism
Example 1 (feedback coil breakage): In the platform, the reached the minimum stroke. The responses of ANN and the
input of the electrohydraulic servomechanism is 0.2 and the electrohydraulic servomechanism in this scenario are shown
feedback signal disappears suddenly at 50 s, the difference in Fig. 15, Line 1 denotes the outputs of the electrohydraulic
between the output of ANN and that of the electrohydraulic servomechanism, and Line 2 denotes that of ANN.
servomechanism increases rapidly and exceeds the given
threshold (0.05%). Under this condition, ANN switches D. Maintenance Decision-Making Based on ANN Model
to parallel identification model (Fig. 6), and the parameters The information stored in the ANN weights helps to diagnose
keep the same values as before failure. ANN stabilizes its and prognose the type and cause of the failures as soon as pos-
output to 0.205 at 52 s around. In Fig. 14, the dashed line 1 sible. In the deregulated power market scenario, the units should
indicates the trend without emergency treatment. The real line be kept running when the market clearing price of electric en-
2 in Fig. 14 gives the actual response of the electrohydraulic ergy is high. Maintenance will be carried out only when the
servomechanism after failure treatment. Therefore, a feedback market price is low. Therefore, the process of maintenance de-
coil breakage failure is detected, and a corresponding treatment cision depends on the failure types and the unit operation mode
is taken. at the same time. This aspect has been taken into account in the
Example 2 (actuator blockage): As for the actuator blockage maintenance decision-making process in the ICMMS platform.
failure, the results are different from that of the feedback coil Example 1: The feedback factor creeping can be diagnosed
breakage. In this case, the initial output of the electrohydraulic and prognosed through ANN at the operating period. If a degra-
servomechanism is 0.2 and the pilot valve is blocked at the dation of feedback factor creeping was detected, but the predic-
closed position at 50 s, the difference exceeds the given tive value does not exceed the predetermined threshold, the unit
threshold (0.05%). The ANN switches to parallel identifica- continues running and the maintenance station sends a message
tion model (Fig. 6), and keeps the same parameters as before the to inform the degradation. Combined with additional informa-
failure, so ANN stabilizes its output to 0.205 at 52 s around, but tion about the governing system, such as the oil temperature-
FU et al.: PREDICTIVE MAINTENANCE IN ICMMS FOR HYDROELECTRIC GENERATING UNIT 185

rise, the aging of the component, etc. the causes of the degra- VI. CONCLUSIONS
dation can be identified. Therefore, the proper time to maintain
the corresponding equipments can be planned in advance. Currently, the control, maintenance, and technical man-
agement within the production process are three “isolated
Example 2: When the feedback coil break failure happens,
islands.” The ICMMS is an integrated and effective solution
there are various treatments under different conditions. If the
to this problem. Three key elements for the predictive main-
unit runs in the network, with the help of ANN, the servomotor
tenance within the framework of ICMMS (i.e., monitoring
keeps its former position, and the maintenance station gives cor-
and forecasting, diagnosis and prognosis, and maintenance
responding failure message to repair, but the unit cannot change
decision-making), are proposed and presented in this paper.
its load. In other words, the unit loses its regulating ability auto-
An ANN-based identification and diagnosis model is put
matically. In this case, the maintenance station informs the oper-
ators of this failure information and claims that this unit should forward and implemented in the predictive maintenance for
be shutdown manually and maintained at appropriate time. If the the electrohydraulic servomechanism of the water-turbine
unit operates in no-load mode, the unit should be shutdown man- governing system. The tests show that the performance is
satisfactory. Moreover, similar models can be established and
ually, and the maintenance station gives corresponding failure
message to repair. applied to other equipments/components in the hydropower
Example 3: ANN can identify the actuator blockages, but it plants.
has no ability to prevent the main servomotor from moving to The newly added predictive maintenance system can utilize
the extreme position. Under these circumstances, the mainte- the information existed in the computer supervisory and con-
nance station will call the operators to interface with the unit trol system, management information system. The investment
operation. If connected to the grid, the unit keeps running with for a predictive maintenance is moderate and the ICMMS for
a given load and the maintenance activities will be taken at ap- hydropower plant can be set up by the integration of these three
propriate time. If in no-load condition, emergent shutdown is systems.
taken and the servomechanism is repaired at once. Within the framework of ICMMS, the conditions of equip-
ment involve not only the technical conditions, but also econom-
ical conditions including costs of investment, operation, condi-
tion assessment, condition forecasting, the direct and indirect
V. BENEFIT ANALYSIS failure costs. The economical condition is one of the main re-
search problems of technical management within the frame-
The economic benefit brought by the predictive maintenance work of ICMMS, so in the predictive maintenance, this factor
is obvious: eliminating unnecessary repairs as in scheduled must be taken into account. More work about this has been per-
maintenance, prolonging the service life of equipments, re- formed and a new performance evaluation method for hydro-
ducing accident shutdown, and prolonging the production time. electric generating unit is presented in [14].
According to the statistics in China, the cost of a major overhaul
for a hydroelectric generating unit of 200 MW is about 15
ACKNOWLEDGMENT
million RMB (U.S.$ 1.87 million), this cost almost accounts
for 30% of the total benefit of a hydroelectric generating unit The authors gratefully acknowledge the SIEMENS Company
[13]. If the predictive maintenance can be applied, this cost can for providing some Profibus products during the setup of the
be reduced by 25–50% [14]. ICMMS platform for the hydroelectric generating unit.
The investments of the application of predictive maintenance
in hydropower plants are quite limited. The computer supervi-
sory and control system and management information system REFERENCES
have been widely set up in most hydropower plants; a lot of [1] L. Ye, Z. Li, Y. Liu, and Y. Zhang, “Intelligent Control-Maintenance-
information required by the predictive maintenance can be ob- Management System and its application on hydropower,” in Proc. Int.
tained from the computer supervisory and control system and Federation Automat. Contr. 2nd Manage. Contr. Prod. Logistic Conf.,
2000, pp. 609–614.
management information system. Therefore, the existing com- [2] R. Yu, L. Ye, and Y. Zhang, “Intelligent control-maintenance-manage-
puter supervisory and control system, management information ment system (ICMMS) and its application to power system—Part one:
system, and newly added predictive maintenance can be inte- ideology, constitution and characteristics” (in Chinese), Autom. Electr.
Power Syst., vol. 23, pp. 50–54, Dec. 1999.
grated into an ICMMS. This can save money. [3] J. B. Leger, B. Iung, A. Ferro de Beca, and J. Pinoteau, “An innova-
To date, the predictive maintenance within the framework tive approach for new distributed maintenance system: application to
of ICMMS has been applied to design or retrofit automation hydropower plants of the REMAFEX project,” Comput. Ind., vol. 38,
pp. 131–148, Mar. 1999.
systems. In Europe, a European project, REmote MAintenance [4] R. Yu, “Theory of predictive maintenance and its application within
for Facility EXploitation (REAFEX) has been applied to framework of intelligent control-maintenance-management system,”
BERDROLA Hydropower plant in Spain and EDP hydropower Ph.D. dissertation, Dept. Hydropower. Eng., Huazhong Univ. Sci.
Technol., China, 2000.
plant in Portugal according to ICMMS principles [3]. In China, [5] L. Ye, S. Wang, F. Bing, O. P. Malik, and Y. Zeng, “Control/maintenance
a maintenance system has been combined with the existent strategy failure tolerant mode and reliability analysis for hydro power
computer supervisory and control system, management infor- stations,” IEEE Trans. Power Syst., vol. 3, pp. 340–345, Aug. 2001.
[6] X. Teng and G. Tang, “Condition monitoring to equipment of hy-
mation system to form an ICMMS in Geheyan Hydropower dropower plant” (in Chinese), Autom. Electr. Power Syst., vol. 24, pp.
Plant [1]. 45–48, Dec. 2000.
186 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 19, NO. 1, MARCH 2004

[7] IEEE Task Force, Risk, and Probability Applications Subcommittee, Yongqian Liu was born in Inner Mongolia, China, on October 3, 1965. He
“The present status of maintenance strategies and the impact of main- received the Master’s degree in 1992 in hydropower engineering from North
tenance on reliability,” IEEE Trans. Power Syst., vol. 16, pp. 638–646, China Institute of Water Conservancy and Hydroelectric Power, Zhengzhou,
Nov. 2001. China. He is currently pursuing the Ph.D. degree at Huazhong University of Sci-
[8] H. Lofsten, “Measuring maintenance performance—in search for a ence and Technology, Wuhan, China, and the University Henri Poincare, Nancy,
maintenance productivity index,” Int. J. Prod. Econ., vol. 63, pp. 47–58, France.
2000. His research interests include the automatic control and system integra-
[9] C. Fu, Y. Cheng, R. Yu, Z. Li, and L. Ye, “The application of fieldbus tion, especially for the intelligent-control-maintenance-management system
technology in Intelligent Control-Maintenance-Management System (ICMMS) in hydropower engineering and technical management.
(ICMMS) platform” (in Chinese), Process Automat. Instrum., vol. 22,
pp. 35–38, July 2001.
[10] A. U. Levin and K. S. Narenda, “Control of nonlinear dynamical systems
using networks—Part II: observation, identification, and control,” IEEE
Trans. Neural Networks, vol. 7, pp. 33–42, Jan. 1996. Ren Yu received the Ph.D. degree in hydropower engineering from the College
[11] M. O. Efe and O. Kaynak, “A comparative study of neural network struc- of Hydropower and Digital Engineering at Huazhong University of Science and
tures in identification of nonlinear systems,” Mechatronics, vol. 9, pp. Technology, Wuhan, China, in 2000.
287–300, Apr. 1999. Currently, he is pursuing the postdoctoral period in the Centre de Recherche
[12] O. S. Turkay and K. K. Safak, “Experimental identification of universal en Automatique de Nancy (CRAN), University of Nancy, Nancy, France, where
motor dynamics using neural networks,” Mechatronics, vol. 10, pp. his research focuses mainly on smart organization of e-maintenance. His re-
881–896, Dec. 2000. search interests include automatic control, fault diagnosis, and prognosis.
[13] L. Ye and X. Yu, “Modernized technology of equipment mainte-
nance—condition maintenance” (in Chinese), Mech. Sci. Technol., vol.
18, pp. 457–462, May 1999.
[14] Y. Liu, “Contributions to the methodologies and technologies for the in-
telligent control-maintenance-technical management systems (ICMMS) Benoit Iung received the Master’s degree in manufacturing and process con-
in hydropower plants,” Ph.D. dissertation, Univ. Nancy I, Nancy, France, trol, an engineering degree in automation engineering, and the Ph.D. degree in
2002. manufacturing engineering from the University of Nancy I, Nancy, France, in
1986, 1987, and 1992, respectively.
Currently, he is an Associate Professor at Computer-Integrated Manufac-
turing (C.I.M.) engineer school École Suprieure Informatique et di Applications
de Lorraine (E.S.I.A.L.), Nancy, France. His research interests include the
dependability of the production system by promoting the methodological and
technological compromise between integration and distribution of “intelligent”
Chuang Fu was born in Sichuan Province, China, on September 27, 1973. He
components.
received the Bachelor’s degree from Sichuan Institute of Technology, Chengdu,
China, in 1996, and the Master’s degree in electrical power system and automa-
tion from Huazhong University of Science and Technology, Wuhan, China, in
1999. He is currently pursuing the Ph.D. degree in the College of Hydropower
and Digital Engineering at Huazhong University of Science and Technology.
Yuanchu Cheng received the Master’s degree from China Institute of Water
His research interests include the intelligent-control-maintenance-manage-
Resources and Hydropower Research, Beijing, China, in 1988. He is currently
ment system (ICMMS) and its application in hydropower plants and application
working toward the Ph.D. degree in the College of Hydropower and Digital
of predictive maintenance in hydropower plants.
Engineering, Huazhong University of Science and Technology, Wuhan, China.
Currently, he is an Associate Professor with the Department of Power En-
gineering at Wuhan University, Wuhan, China. His research interests include
intelligent system applications in hydroelectric generating process and intelli-
gent-control-maintenance-management system (ICMMS).

Luqing Ye was born in Zhejiang Province, China, on July 1, 1935. He graduated

from Huazhong University of Science and Technology, Wuhan, China, in 1958.
Currently, he is a Professor at Huazhong University of Science and Tech-
nology and is Chairman of China Net of CMMS-IAMS network committee. Yuming Zeng (M’93) was born in Wuhan, China, on March 13, 1960. He re-
He was a Visiting Scientist from 1979 to 1981 at the Institute National Poly- ceived the B.Sc. and M.Sc. degrees in electrical engineering from Huazhong
technique de Grenoble and Derection des Etudes et Recherches, Electricite de University of Science and Technology, Wuhan, in 1982 and 1985, respectively,
France, Paris, France. His research interests include intelligent system applica- and the Ph.D. degree from the University of Calgary, Calgary, AB, Canada, in
tions in hydroelectric generating process and intelligent-control-maintenance- 1991.
management system (ICMMS). Currently, he is the Engineering Manager at Nextron Corporation, Calgary.