Fault prediction/diagnosis and sensor validation technique for a steam ...

18 downloads 86977 Views 406KB Size Report
ANNs are tested with unseen data sets, including combined scenarios of the partially failed system to assess ... Often they are slow to execute and hard ... failed sensor and recover the failed critical measurement if it .... drive turbine. This will ...
mesbahi.qxd

4/11/05

3:49 pm

Page 33

Fault prediction/dianosis and sensor validation technique for a steam power plant

Fault prediction/diagnosis and sensor validation technique for a steam power plant E Mesbahi, School of Marine Science and Technology, University of Newcastle upon Tyne, UK M Genrup and M Assadi, both of Lund University, Sweden An intelligent sensor validation and fault prediction/diagnosis technique for a typical steam power plant is proposed and studied. An auto-associative Artificial Neural Network (ANN) is trained to examine the consistency of the overall simulated data and allocate a confidence level to each signal.The same set is used to replace the missing or faulty data with a close approximation. For fault prediction and diagnostic system a feed-forward ANN with extra linear connections is trained to recognise faulty and healthy behaviour of the steam cycle for a wide range of operating conditions. Both ANNs are tested with unseen data sets, including combined scenarios of the partially failed system to assess fault prediction capability of the proposed ANN. It is concluded that a significantly more reliable sensor reading and a highly accurate fault prediction/diagnosis system is achieved.

INTRODUCTION

O

ver the past few years, maintenance techniques have drastically changed. New information processing technologies to reduce costs have been implemented and companies are experiencing increased competitiveness. The conventional corrective maintenance approach, which tends to fix the problem as soon as it occurs, seems to be more expensive when compared to predictive schemes where the plant down-time is minimised. Most competitive manufacturers use periodic AUTHORS’ BIOGRAPHIES Ehsan Mesbahi is a Senior Lecturer with the School of Marine Science and Technology at Newcastle University in the UK. He has a wide range of research interests mainly involved in system analysis, control and optimisation. He has also been focusing on marine system design for the environment and reducing harmful impacts of maritime technologies on the marine environment. Mohsen Assadi received his PhD from the Dept. of Heat and Power Engineering, Lund University and is currently working as associate professor. His research interest is environmental friendly fuel based energy conversion for heat and power production. Power plant monitoring and performance analysis are among research areas he has focused on. Magnus Genrup received his PhD from the Dept. of Heat and Power Engineering, Lund University. His research interest is turbomachinery design, modelling and analysis.

No. A7 2005

Journal of Marine Engineering and Technology

maintenance, which requires maintenance tasks on a timebased programme. Nevertheless, none of these regimes are optimal. Increased economic competitiveness necessitates reorganisation of power plant operations; many maintenance groups are determined towards condition-based maintenance rather than periodic or corrective maintenance. Changes in condition based calibration strategies require instruments to be physically recalibrated only when their performance is degraded. Continuous monitoring of the instrument’s calibration performance will allow power plants to reduce the efforts necessary to assure that the instrument is in calibration. Benefits of continuous sensor validation include the reduction of unnecessary maintenance and more confidence in actual sensed parameter values. Reduced maintenance will result in cost savings and reduced outage times while a better knowledge of the actual state of the process could result in increased product quality and reduced equipment damage. In order to implement preventive or condition-based maintenance techniques it is essential that early warning of developing faults is provided so that appropriate decisions may be taken and correct actions planned in advance. For this reason various condition monitoring techniques are developed which generally involve power plant health monitoring using case based reasoning, Artificial Neural Networks, performance physical modelling analysis and many more.1,9 These techniques should be able to assess engine condition, predict developing failures, diagnose the failure cause and finally, give recommendations for operation and maintenance

33

mesbahi.qxd

4/11/05

3:49 pm

Page 34

Fault prediction/dianosis and sensor validation technique for a steam power plant

ensuring the plant will operate safely and to its maximum efficiency. Existing/proposed monitoring techniques involve mathematical and/or causal models, which simulate the healthy and faulty plant/component in order to identify or verify a particular condition.10 These models have been applied with various degrees of success. Often they are slow to execute and hard to implement. Derivations from the healthy conditions are discriminated by using physical models, which are based on many simplifications, unrealistic assumptions and linearisations. However the main cause of unreliability for these techniques is uncertainty in data acquisition, either by a completely failed sensor or incorrect measurement originating within the hostile environment in which the engine and sensors are installed. Operators’ expertise, response time, tiredness and sobriety could also contribute to a wrong diagnosis. Alarms may appear in large quantities in a short time and need to be recorded and interpreted immediately. In order to improve a power plant’s operational reliability, fault diagnostic accuracy and condition monitoring precision, it is necessary to validate the acquired data, isolate any failed sensor and recover the failed critical measurement if it is required by other systems such as control or fault diagnosis mechanisms. Extended efforts have been made to apply analytical redundancy to sensor failure detection and isolation in jet engine and nuclear power plant failure diagnosis.11 In general, these approaches utilise the engine physical model and the Kalman filter (for state estimation) to detect and isolate the failed sensor(s). These techniques strongly depend upon an accurate model of the system, which may not be attainable in a time-variant dynamic system. In this paper the following topics are addressed: G Steam power plant is selected for our case study, its physical modelling and simulation is presented and used for generation of data required by ANN. G The application of ANN technology to Sensor Validation (SV) and Fault Diagnostics and Prediction (FDP) of a steam plant is studied.

ARTIFICIAL NEURAL NETWORKS (ANNs) ANNs have been used successfully in pattern classification, data management modelling and control problems. By increasing the complexity of the processing elements using input-output data samples of a continuous process, the network can take the form of an interpolation and extrapolation filter. They have the capability of storing data during learning processes and then reproducing this data during recall processes. Their pattern classification and generalisation capability, which in some cases extends to valid extrapolation, is the focus of their utilisation for mapping multi-dimensional input/output data sets as: f : Rn → Rm

(1)

where f is any continuous function, R and R are input and output data spaces of dimensions n and m as shown in Fig 1. X=(x1,x2,…,xn) and Y=(y1,y2,…,ym) are input and output vectors respectively. W1(m+1, h) and W2(h+1, n) are first and second weighting matrices where h is the number of processing elements in the intermediate layer, called the hidden layer. n

34

m

Fig 1: A fully connected feed-forward ANN with m input and n outputs The key to successful application of ANNs lies in training the network that would represent the signal association as accurately as possible. Backpropagation algorithm is an iterative gradient descent technique that minimises the global error between the output of a multi layered ANN and the desired (actual) output.

ANNs for sensor validation and fault recognition In recent years there has been an increased interest in the applications of ANNs for engineering problems. For implementations within remote sensor validation and condition monitoring the following properties of ANN are important. 1. Applicability to a majority of nonlinear systems shows that a feed-forward ANN with at least one hidden layer is capable of approximating any nonlinear function if enough training is provided.12 Due to this capability, they are also easily capable of providing reverse (effect-to-cause) model of any nonlinear system. 2. Parallel distributed processing and hardware implementations; ANNs have inherent parallel architecture, which could lead to parallel hardware implementations. These implementations also have an advantage of having in general, a high degree of fault tolerance and high processing speed due to the simplicity of their connections. 3. Learning; ANNs can be trained using past recorded data (offline learning) or current data (online learning). 4. Applicable to multivariable systems; ANNs are, by definition, multi-input multi-output entities and this naturally leads to their application to multivariable systems such as power plants. 5. Speed of response; trained ANN models are by far much faster than physical models since they do not need to perform any iterative calculations and/or search for any parameters. This could provide an online platform for live data processing and analysis.

Autoassociative neural networks Autoassociative ANNs where the same inputs and outputs signals are trained for an appropriate range of system dynamics are generally used for sensor validation.13 During training,

Journal of Marine Engineering and Technology

No. A7 2005

mesbahi.qxd

4/11/05

3:49 pm

Page 35

Fault prediction/dianosis and sensor validation technique for a steam power plant

interrelationships between the signals are embedded in the ANN connection weights. By using a robust training procedure, ANN is forced to rely on the information inherent in the signals correlated with a specific sensor to estimate that specific sensor value. As a result, any specific ANN output shows virtually no change when the corresponding input has been distorted by noise, faulty or missing data.3 Autoassociative ANNs generally have five layers of input, mapping, bottleneck, de-mapping and output.14 Utilisation of such networks and the number of layers are generally dependant on the application and the nature of data. In our approach, to be able to provide early warnings in condition monitoring/fault diagnosis and to implement sensor validation tasks, we have used two feedforward ANN structures, one to validate the incoming data, supposedly measured online, this model is subsequently called ‘SV unitl’. The other ANN is trained to model system faults and associate a particular pattern to each. After successful fault pattern association, the possibility of using the FDP ANN model for providing early warning (fault prediction) on the steam plant is presented.

STEAM PLANT MODELLING PROCEDURE In order to generate appropriate data for ANN training we may either use a real set of data acquired from a steam plant or we may alternatively use physical modelling tools to simulate the real scenario. Obviously, analytical models here are only used to prove the applicability of the proposed methodology.

The chosen cycle is a full arc admission (maximum power) industrial cold condensing turbine plant. This particular plant consists of five pre-heaters and the following general data: - Admission pressure: 140bar(a) - Admission temperature: 540°C - Steam flow: 50kg/s - Final feed water temperature: 287.5°C - Condenser pressure: 0.08bar(a) - Alternator output: 45 247kW.

Cycle calculation The cycle performance is calculated with an in-house code developed at Lund Technical University. The code which is a flexible heat and mass balance calculation tool (HMB program), also calculates part-load and off-design performance data. The cycle model is based on standard aerothermal methods.

Brief introduction to the turbine section model The isentropic efficiency of a turbine section is defined as: hα − hω hα − hω ,s

(2)

The part-load efficiency of a turbine section is assumed to be a function of the Parson number. The Parson number describes the average loading in the specific turbine section and is defined as:

No. A7 2005

∑u

2

Journal of Marine Engineering and Technology

(3)

∆hs

It is only the relative Parson number which is of interest when the part load efficiency is to be calculated. The efficiency is assumed to be a parabolic function of the relative Parson number and the design value. One further simplification is also possible since the shaft speed is constant for a generator drive turbine. This will give a parabolic function of the relative isentropic heat drop: ⎛ ∆h ⎞ ηs = ηs ,design ⋅ f ⎜ s ,design ⎟ ⎝ ∆hs ⎠

(4)

This approach gives a rather good approximation of the turbine section efficiency. The changes in the Parson number are small for all turbine sections except for the last stage which operates under different pressure ratios, depending on the load. It may be shown that under various load conditions the velocity triangles are more or less constant for all stages except for the last. The pressure ratio or loading in a turbine section will change if the flow path is changed eg, due to fouling. The turbine or swallowing capacity is loosely stated a gauge of the effective flow path size for a turbine section. One common definition of turbine capacity is: m = CT ,α →ω

Cycle design data

ηs =

Χ=

pα2 − pω2 pα vα

(5)

There are other ways of defining the turbine capacity but the expression above will provide sufficient accuracy for this study. The given reference provides a comparison between different capacity models.19 They are normally based on the Stodola ‘steam cone’ rule improved by eg, experience factors. It should be emphasised that various methods applied on different type of turbines generate quite different results.19 One method, which may work very well on an eg, lightly loaded reaction turbine may be quite erroneous on a choked impulse turbine. This is normally one of the weak spots in commercial heat and mass balance programs and one should use the builtin models with some caution. The uncertainty in flow when using generic correlations may under favourable conditions be in the range of 1%. The uncertainty figure mentioned previously is valid for HP turbines. LP turbines may have a much higher uncertainty figure (up to about 5%). One other way to calculate swallowing capacity is to use a more advanced calculation model, preferably an S-2 through flow program that generates turbine section characteristics. Such programs are commercially available with different loss and blade outlet angle models but they require detailed knowledge about the meridional flow path and the blades in the specific turbine. Calculation of the last stage exhaust losses in this paper are based on a method derived from DIN 1943.20 Probably the most used method in manufacturing companies is based on a calibrated mid-span calculation model

35

mesbahi.qxd

4/11/05

3:49 pm

Page 36

Fault prediction/dianosis and sensor validation technique for a steam power plant

together with an HMB program. Such in-house programs representing the company’s know-how are most likely proprietary codes and not available outside the specific company.

Condenser performance Calculation of condenser heat transfer coefficient is based on a method derived by the Heat Exchanger Institute (HEI).21 This method gives a simple tool for calculation of the heat transfer coefficient according to: (6) U =C C C C V 1

2

3

4

Where: C1 = factor depending on the outer tube diameter C2 = factor depending on the cooling water inlet temperature C3 = factor depending on the tube material and thickness C4 = fouling margin V = cooling water velocity in the tubes at the inlet (m/s). The off-design heat transfer coefficient is calculated according to the equation: U act = U des

C2,act C2,des

m cooling ,act m cooling,des

Simulation of faults Three different fault scenarios were simulated in the HMB program (these results have not been validated on the real system): 1. Deposits in the turbine section situated between the top HP heater and the 1st HP heater 2. Solid particle erosion (SPE) in the 1st stage nozzle box 3. Condenser fouling. The deposits may be a result of massive Na or SiO2 ‘carry over’ due to a problem with the make-up water treatment plant. This fault is simulated by reducing both the swallowing capacity by 5% and the calculated turbine efficiency by 2% units in the referred turbine section. The SPE is modelled by increasing the capacity in the 1st turbine section and reducing the efficiency by 2% units. The modelled condenser heat transfer capacity reductions are: 10, 20, 30 and 40%. All changes to the cycle are carried out at nominal admission data and three different cooling water temperatures (10, 15 and 20°C) for the first two cases.

Operational instruments (7)

Off design calculation For off-design calculation the set of necessary equations is solved iteratively using a Newton-Raphson matrix solver. The preheater terminal temperature difference (TTD or grädigkeit) is assumed to be constant at part-load in this study.

The plant is assumed to be furnished with sufficient instruments according to the industrial standard of today. Most modern turbine plants have a Distributed Control System (DCS) system for control and operation. Some plants even have a system for long-term saving of parameter values. 1. Alternator output 2. Net station output (alternator minus parasitic consumption)

Fig 2: Schematic diagram of the steam power plant

36

Journal of Marine Engineering and Technology

No. A7 2005

mesbahi.qxd

4/11/05

3:49 pm

Page 37

Fault prediction/dianosis and sensor validation technique for a steam power plant

3. Admission pressure 4. Admission temperature 5. Steam flow (assumed accurate in this study) 6. Control valve position 7. Cross-over pipe pressure 8. Cross-over temperature 9. Condenser pressure 10. Inlet cooling water temperature 11. Outlet cooling water temperature 12. Condensate temperature 13. Condensate temperature after the 1st low pressure preheater 14. Condensate temperature after the 2nd low pressure preheater 15. Dearator pressure 16. Temperature after the 1st high pressure preheater 17. Temperature after the 2nd high pressure preheater 18. Feed water pressure 19. Shell pressures in all surface heaters.

operational range and simplicity of implementation for realtime use has been addressed.15 20 data sets are chosen from 50 sets and used for training purposes with 1000 training iterations only. Input data was normalised between 0.3 and 0.7 (instead of 0 and 1), which proved to be immensely successful in speeding up the learning procedure as well as its generalisation capability outside the training limits. Using such simple auto-associative ANN also provided us with a single functional relationship for further sensitivity analysis and simple hard/software practical implementations.1 (x1, x2, …, xm) = f (x1, x2, …, xm)

(8)

where xm is measured data and f is the non-linear relationship between sensor readings and themselves. In this case study m=25. Function f in equation (8) can be represented as: ⎡ h ⎤ ⎛ n ⎞ xm = sig ⎢ ∑ sig ⎜ ∑ x j W1i , j + ci ⎟ W 2 m ,i + bm ⎥ ⎠ ⎢⎣ i =1 ⎝ j =1 ⎥⎦

(9)

SENSOR VALIDATION The concept of validating a measurement suggests that a comparison must be made with some known and trustworthy information or that some logical conclusions can be drawn from the data themselves. It is also understood that any single value read by a sensor must be consistent with the other sensor readings. If a valid correlation between all measured data at different operational conditions including faulty conditions, is established this could lead to a valid interpretation of the available readings. In reality this correlation is represented by the system itself. A single auto-associative feed-forward ANN with 16 nodes in input and output layers and 10 nodes in a single hidden layer is selected for sensor validation purposes. By this selection many technical and industrial objectives such as online, timely sensor assessment, applicability on a wide

where h, m and n are hidden, output and input nodes respectively (m=n in this case), W1, W2, c and b are network weights, which are frozen after training, sig(.) is a sigmoidal function presented in equation (3): sig(x) =

1 1 + e-x

(10)

To investigate the performance of the trained ANN for sensor validation, 8 measurements (out of 16) are chosen and their actual reading are manually changed by 0%, 20%, 50%, 70% and 130% to account for the fact that a failed sensor does not necessarily transmits a value of zero. (100% is considered for a healthy sensor operating at normal condition.)

Definition of confidence level (CL) Equation (11) is used for calculation of a confidence level for each reading:

Failure of Valve Sensor 1.2

Confidence Level

1 0.8 0.6 0.4 0.2 0 pcond pco valve flow php2 php1 plp2 plp1 tcwout ALT 0%

Fig 3: Auto-associative ANN for sensor validation

No. A7 2005

Journal of Marine Engineering and Technology

20%

50%

70%

130%

Fig 4: Sample of sensor validation results

37

mesbahi.qxd

4/11/05

3:49 pm

Page 38

Fault prediction/dianosis and sensor validation technique for a steam power plant

Healthy Fault 1 Fault 2 Fault 3

Output 1 Output 2 Output 3 Output 4 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1

Table 2: Binary output values associated with faulty and healthy conditions Sˆ i -Sri CL i = 1Sˆ

Table 3: Fault diagnosis results for four patterns of unseen data

(11)

i

where Sˆ is the predicted value by ANN, Sr is the sensor reading and i is the sensor number. A confidence level close to one indicates a highly reliable sensor and a value close to zero describes a failed sensor. During the test, CL for all 16 sensors is calculated, (Fig 3). Some results are presented in Fig 4.

Data recovery For reliable condition monitoring, control and engine fault diagnosis, it is imperative to have access to highly reliable measurements. If a sensor in a closed loop of a control system fails the control mechanism has no option other than to respond to the new condition and to act accordingly which in some cases may be disastrous. In this section the potentiality of data recovery is addressed. A comparison between the predicted value by the combination of forward and reverse ANN models with the healthy sensor reading is presented in Table 1. Although not entirely accurate all predicted (recovered) values are very close to the correct value. This may temporarily replace the readings to avoid unexpected control and monitoring system responses whilst raising the alarm for mainteTable 1: Comparison between recovered (predicted) and correct reading with a faulty sensor for10 sample measurements

Sensor failed Net electric output Final feed watertemp FW temp after HP heater No1 Feed water temp after DEAE/feed water tank Condensate temp after LP heater No 2 Condensate temp after LP heater No 1 Condenser pressure Cross-over pressure (between HP and LP turbine module) Control valve position Admission flow

38

Steam Plant Pattern 1 Pattern 2 Pattern 3 Pattern 4 Condition Healthy 99% 0% 1% 0% Fault 1 1% 98% 0% 0% Fault 2 0% 0% 99% 0% Fault 3 0% 2% 0% 100%

nance action. If the same set of signals is used for fault diagnostic purposes recovered data set shall prevent arrival at incorrect conclusions by such systems.

STEAM PLANT FAULT DIAGNOSIS In order to investigate the fault diagnostic capability of an ANN in our case study, each input data set is associated with a particular binary output pattern, representing faults 1, 2, 3 and healthy pattern, represented in Table 2. From 50 data sets available, 20 sets are randomly selected for network training while the rest is kept aside for validation purposes. It is noticed that only 500 iterations are sufficient to achieve a reasonably low training-error criterion. By using a feedforward ANN with additional linear connections between inputs and outputs and applying a conventional backpropagation training algorithm reasonably acceptable solutions are achieved. All unseen data patterns (those not used during ANN training) could be easily recognised by numerical comparison, some of which are presented in Table 3.

STEAM PLANT FAULT PREDICTIONS In order to evaluate the capability of the trained ANN for fault prediction, fault 3 was selected for further investigations. Simulation of the steam plant physical model as described above was run with the faulty condition but with consideration of 20%, 50% and 70% of the fault 3. This could provide us with the data belonging to the system ‘approaching’ towards fault 3. Numerical interpretation of response of FD

Recovered value when actual measurement 20% 50% 70% reading reading reading 42793 43222 43594 268 276 283

is

Correct Reading 43922 287

0% reading 41227 260

226

198

213

215

221

229

172

146

153

164

168

172

123

99

113

116

120

127

80 0.079939

67 0.059467

71 0.064983

74 0.068528

78 0.071552

83 0.081589

8.919924 0.030028 49.99849

6.357698 0.021193 37.21498

7.113528 0.023784 39.12651

7.835673 0.026003 42.69921

8.435802 0.028204 47.86432

9.236001 0.032811 53.97187

Journal of Marine Engineering and Technology

130% reading 44323 291

No. A7 2005

mesbahi.qxd

4/11/05

3:49 pm

Page 39

Fault prediction/dianosis and sensor validation technique for a steam power plant

Steam plant condition Healthy Fault 1 Fault 2 Fault 3

Pattern 1 20% of Fault 3 90% 5% 2% 3%

Pattern 2 50% of Fault 3 33% 8% 27% 32%

Pattern 3 70% of Fault 3 4% 1% 7% 88%

Table 5: Results of FD ANN when data belonging to 20%, 50% and 70% of fault 3 is presented to the network ANN to these new patterns are summarised in Table 5. It is clearly observed that in the first case, ie, 20% of fault 3, ANN is not able to distinguish recognise the faulty condition although the possibility of the system being healthy is reduced to 90%. In the second case of 33% the system is recognised not to be healthy but fault diagnosis is not distinctive. In the last case Pattern 3, ANN is able to clearly point to the highest possibility, ie, fault 3.

FAULT PREDICTIONS WITH ANN GRAPHICAL INTERFACE Considering pattern recognition capabilities of ANNs and to provide a user-friendly interface for operational purposes another feedforward ANN is to be trained in such a way as to graphically present the steam plant condition where recognition between healthy and faulty system conditions as well as the tendency of the system moving towards a fault can be graphically observed. This will be an added value for operators to recognise and diagnose the faults at early stages of development. This additional ANN will receive binary outputs from previous network, described above and translate those output into 50 digit (or pixel) which are arranged into a 10 rows and 5 columns structure and graphically present numbers such as 1, 2, etc, as well as words such as H standing for healthy. The outcome of this additional interface is presented in Figs 5 and 6 each captioned with the respective system condition.

Fig 5: Interface patterns recognised by ANN as faults 1, 2, 3 and healthy condition for unseen data presented in Table 4.

Fig 6: Interface patterns corresponding to 20%, 50% and 70% fault 3 conditions from left to right respectively and as presented in Table 5.

CONCLUSIONS Application of ANNs for sensor validation, data recovery and engine fault diagnosis/prediction by using the simulated steam plant data is studied. It was observed that an auto-associative

No. A7 2005

Journal of Marine Engineering and Technology

3-layer ANN is successful in examination of data consistency; each reading can be tagged with a ‘confidence level’ describing the reliability of data measured. It was also experienced that if the confidence level is very low, ANN can firstly alarm the operator and consequently replaces the faulty data with a close approximation for the other users such as control or fault diagnostic systems. A single feed-forward ANN with an additional linear layer is trained for system fault diagnosis purposes and tested against untrained data where results of more than 98% correct fault prediction is achieved. The proposed fault diagnostic programme is then tested, numerically and graphically to recognise the trend of an approaching fault; a fouled compressor inlet in this case. It was concluded that when in more than 50% closeness to the actual fault it will announce irregularity in the plant, whilst when around 70% approach to fault 3 it can clearly alarm operators that the system is moving towards a recognisable fault.

ACKNOWLEDEGMENTS This work was sponsored by the Swedish National Energy Administration, within the frame of the research program ‘Methods for Analysis and Optimisation of Thermal Power Plants’.

REFERENCES 1. Volponi AJ, Depold H, Ganguli R and Daguang C. The use of kalman filter and neural network methodologies in gas turbine performance diagnostics: a comparative study, Proceedings of ASME TURBOEXPO 2000, Munich, Germany. 2. Roemer MJ and Atkinson B. Real-time engine health monitoring and diagnostics for gas turbine engines, 53rd Machinery Prevention Technologies (MFPT) Conference, Virginia, USA, 1999. 3. Mesbahi E. An intelligent sensor validation and fault diagnostic technique for diesel engines, Journal of Dynamic systems, Measurement and Control, Trans. ASME, March 2001. 4. Hargarve SM and Flemming PJ. The use of case-based reasoning technology to aid fault isolation in a modern gas turbine engine design, Proceedings of ASME International Gas Turbine & Aeroengine Congress and Exhibition, Stockholm, Sweden, 1998. 5. Thompson BD and Raczkowski R. Development of a diagnostic tool to troubleshoot LM2500 performance & controls problems, Proceedings of ASME International Gas Turbine & Aeroengine Congress and Exhibition, Birmingham, UK, 1996. 6. Peng PY and Yang MT. Neural networks based diagnosis for mistuned bladed disk, Proceedings of ASME TURBOEXPO 2000, Munich, Germany. 7. Ghiocel DM. Refined stochastic field models for jet engine vibration and fault diagnostics, Proceedings of ASME TURBOEXPO 2000, Munich, Germany. 8. Tsalavoutas A, Aretakis N, Mathioudakis K and Stamatis. Combining advanced data analysis methods for the constitution of an integrated gas turbine condition monitoring and diagnostic system, Proceedings of ASME TURBOEXPO 2000, Munich, Germany. 9. Bergman JM, Boot P and Woud JK. Diagnosis of

39

mesbahi.qxd

4/11/05

3:49 pm

Page 40

Fault prediction/dianosis and sensor validation technique for a steam power plant

marine diesel engine faults by pattern recognition of acoustic sound, ICMES 93, Conference in Marine System Design and Operation, Hamburg-Harburg, IMarE Trans., vol. 105, No. 3, 171-179. 10. Kao M and Moskwa JJ. Turbocharged diesel engine modelling for nonlinear engine control and state estimation, Journal of Dynamic Systems, Measurement and Control, Trans. ASME, Vol 117, 20-30, 1995. 11. Merril WC, Delaat JC and Burton WM. Advanced detection, isolation and accommodation of sensor failures, real-time evaluation, AIAA J. Guidance, Control Dynamics, 11, 6, 1988. 12. Cybenko G. Approximation by superposition of sigmoidal functions, Mathematics of Control Signals and Systems, 2, 4,1989 , pp 303-314 13. Kramer MA. Nonlinear principal component analysis using autoassociative neural networks, AIChE Journal, 37(2), pp 233-243, 1991. 14. Hines JW and Uhrig R E. Use of autoassociative neural networks for signal validation, Journal of Intelligent and

40

Robotic Systems, 21, pp 143-154, 1998. 15. Mesbahi E. An intelligent sensor validation and fault diagnostic system for diesel engines, ASME Journal of dynamic systems, measurement and control, 2001. 16. Mesbahi E, Assadi M and Torisson T. An online and remote sensor validation and condition monitoring system for power plants, 23rd CIMAC Congress, Hamburg, Germany, 2000. 17. Malloy DJ, Chappel MA and Biegel C. Real-time fault identification for developmental turbine engine testing, Proceedings of ASME International Gas Turbine and Aeroengine Congress and Exposition, Sweden, 1997. 18. Traupel W. Thermische Turbomaschinen pp.51-66, ISBN 3-540-07939-4 Springer-Verlag, Berlin, 1977. 19. Cordes. Strömungstechnik der gasbeaufschlagten Axialturbine pp.234-274 20. DIN 1943. Thermal acceptance tests of steam turbines, 1975. 21. El-Wakil MM. Powerplant Technology pp.233-234, ISBN 0-07-019288- McGraw-Hill, New York, 1984.

Journal of Marine Engineering and Technology

No. A7 2005