Using artificial intelligence technique based on neural network in gas

0 downloads 0 Views 582KB Size Report
Sep 17, 2017 - Keywords: Artificial intelligence, gas turbine monitoring, neural ... the importance of gas turbines in petroleum installations and in other industrial fields, ... in contrast to the use of artificial intelligence based methods based.
The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Using artificial intelligence technique based on neural network in gas turbine monitoring Mohamed Benrahmoune (1)

(1)

, Ahmed Hafaifa

(1*)

, Mouloud Guemana

(2)

(3)

and Fatima Bekkadour

Applied Automation and Industrial Diagnostics Laboratory, Faculty of Science and Technology, University of Djelfa 17000 DZ, Algeria

Emails : [email protected], [email protected], [email protected] (2)

Faculty of Science and Technology, University of Médéa, Algeria Email: [email protected] (3)

SCAL Team, MISC Laboratory, Constantine, Algeria. Email: [email protected]

Abstract The detection of failures in an industrial plant requires a thorough knowledge of the behaviors of the system which complicate and modernize in new technology. The objective of this work aims at the design of an intelligent diagnostic tool based on the network of artificial neurons, this technique with its capacities of generalization and memorization gives a tool of diagnosis and the efficient modeling, applied to A MS 3002 gas turbine, in order to facilitate the interventions on the turbine examined as well as to avoid the prognosis of the operators and to reduce the cost of preventive maintenance. Keywords: Artificial intelligence, gas turbine monitoring, neural network, vibration. 1. Introduction Industrial production is characterized by ever-increasing complexity, which leads to complexities in diagnostic systems and can be critical throughout the industrial plant. For this purpose and in view of the importance of gas turbines in petroleum installations and in other industrial fields, these systems require a reliable diagnostic system to deal with the fugitive defects affecting these processes, which can put the turbine in critical or in a fully stopped state of the machine. In the need for speed to detect and locate these failures in these gas turbine systems, this work uses artificial intelligence techniques based on artificial neural networks for the treatment of diagnostic problems in these machines, Since neural networks have a great capacity for generating and storing information. By these techniques, this work proposes to treat the various failures, which can lead to dangerous vibrations and which degrades the operating state of these types of machines. The vibrations can then be seen as symptoms of failures which would allow to translate the mechanical state of a gas turbine through tools to aid in the detection and diagnosis of failures. Indeed, a large quantity of the breakdowns causes a decrease in the production appearing in the industrial systems, where the failure to detect breakdowns leads to a degradation of the turbine itself

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

with the redundancy of maintenance interventions. All these disadvantages are caused by the use of conventional diagnostic methods, in contrast to the use of artificial intelligence based methods based on the neural network which possesses the ability to memorize and generate knowledge by the turbine system at gas. However, the integration of the neural network in the diagnosis of industrial systems allows a reliable and immediate diagnosis of failures occurring fugitively on the system as well as ensuring a good operating condition of this turbine system and decreases the costs of preventive maintenance. 2. Gas turbine The gas turbine is a machine which converts heat energy into mechanical energy. These transformations are carried out by the basic elements of the operation of the gas turbine, whatever its design, a combustion turbine, is a machine which is currently in vogue, this machine is composed of three main elements; The axial compressor, the combustion chamber and the power turbine. The examined turbine in this work is a turbine of model MS 3002 with double shaft and single cycle, it consists of a 15-stage axial compressor with 6 combustion chamber arranged (90 °) as a function of axial direction of high pressure turbine (HP) has a single stage which is the first step and in the second stage there is the pressure-based turbine which drives the load. This turbine is installed at the gas compression station in Hassi Massoud in southern Algeria, the monitoring configuration of this machine is shown in Figure 1.

Fig 1. Gas turbine MS 3002 The input pressure to the compressor P1 is given by the pressure drop in the intake channel decreased by the pressure drop of the turbine installation area, given by:

2

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

P1 = Pa − ∆Pa

(1)

With Pa is the atmospheric pressure of the turbine installation area, ∆Pa is the pressure drop in the intake channel, determined by P2 = P1τ ,

τ

is the compression ratio.

The temperature of the compressor output is given by:

P T2 = T1 ( 2 ) P1

(γ −1)

γ

(2)

With the isentropic exponent γ is defined by:

γ =

C p (T1 −T2 ) C p (T1 −T2 ) − r

(3)

C p (T1 −T2 ) defines the average specific heat of the air between the temperatures T1 and T2 . 3. Intelligent fault diagnosis of a turbine Neural networks are capable of generating behavioral models from the input-output data of dynamical systems. They have been widely used in the control, modeling and monitoring of industrial systems. Neural networks are nowadays a well-understood and mastered data processing technique that allows the engineer to extract, in many situations, the maximum amount of relevant information from the data he possesses : Process control, property prediction, function modeling, pattern recognition, etc. Moreover, in the diagnosis phase of industrial systems and precisely in the residue generation step between the system studied with the neural model, the network possesses the ability to obtain the behavior of the system to evaluate the residue. In the case of the gas turbine examined, detection of neural network-based defects is proposed, this approach is shown in Figure 2.

3

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Fig 2. Detection by neural network-based turbine defects In this work we select the multilayer perceptron , shown in Figure 3, which is an extension of the monolayer perceptron which has one or several hidden layers. The neurons are arranged in successive layers: the first layer that forms the vector of the input data is called the input layer, while the last layer that produces the results is called the output layer, all the other layers in the middle are called layers. The neurons in the input layer are connected only to the next layer while the neurons in the hidden layers have the particularity of being connected to all the neurons of the previous layer and the next layer. The selection of the number of hidden layers generally depends on the complexity of the problem to be solved, in theory a single hidden layer may be sufficient to solve the problems, but it is possible that having several hidden layers can solve the diagnostic problems more easily.

4

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Fig 3. Multilayer perceptron configuration The error of a neuron k in a layer j is calculated from the errors of the weight-weighted neurons in the layer j + 1 , as following :

ε k( j ) = f ( ∑ w ki x i ). i

∑ w ki( j + i ) .ε i( j +1)

i∈ j +1

(4)

4. Application results After several tests, we have developed a multi-layer neural network with a hidden layer of 18 neurons and sigmoid-type activation function and the linear type activation function output layer as well as minimizing the function of costs, we train our network 800 times, with each iteration we get an error and the network repeat the process in order to have a zero error. The Figure 4 shown the axial vibration of the turbine with the associated network model, the Figure 5 shown the mean squared error of the network modelling, the Figure 6 present the histogram error between the actual signal and the network output.

5

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Variation temporelle du vibration axial du Turbine

30 25

V ib r a t io n u m

20 15 10 Sortie réel Sortie du réseau

5 0 4

E rre u r

2 0

-2 -4 50

100

150

200

250

300

350

400

450

500

550

Time sec Fig 4. Axial vibration of the turbine with the associated network model Best Validation Performance is 0.45348 at epoch 8

3

Mean Squared Error (mse)

10

Train Validation Test Best

2

10

1

10

0

10

-1

10

0

2

4

6

8

10

12

14

14 Epochs

Fig 5. The mean squared error of the network modelling

6

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

300

Training Validation Test Zero Error

250

In s ta n c e s

200 150 100

5.574

4.973

4.371

3.769

3.167

2.565

1.963

1.361

0.7587

0.1567

-0 . 4 4 5 3

-1 . 0 4 7

-1 . 6 4 9

-2 . 2 5 1

-2 . 8 5 3

-3 . 4 5 5

-4 . 0 5 7

-4 . 6 5 9

-5 . 2 6 1

0

-5 . 8 6 3

50

Errors = Targets - Outputs

Fig 6. Histogram error between the actual signal and the network output The Figure 7 shown the radial vibration of the turbine and the associated network model, the Figure 8, present the mean squared error of the turbine radial vibration modelling and the Figure 9, shown the histogram error between the actual signal and the network output.

Variation temporelle du vibration radial du Turbine

35 30

V ib r a t io n u m

25 20 15 10

Sortie réel Sortie du réseau

5 0

E rre u r

5 0

-5 50

100

150

200

250

300

350

400

450

500

Time sec Fig 7. Radial vibration of the turbine and associated network model

7

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Best Validation Performance is 0.92448 at epoch 4

3

Mean Squared Error (mse)

10

Train Validation Test Best

2

10

1

10

0

10

-1

10

0

1

2

3

4

5

6

7

8

9

10

10 Epochs

Fig 8. The mean squared error of the turbine radial vibration modelling

180

Training Validation Test Zero Error

160 140

In s ta n c e s

120 100 80 60 40

4 .8 6 8

4 .3 4 7

3 .8 2 6

3 .3 0 6

2 .7 8 5

2 .2 6 4

1 .7 4 4

1 .2 2 3

0 .7 0 2 1

0 .1 8 1 4

-0 . 3 3 9 3

-0 . 8 6

-1 . 3 8 1

-1 . 9 0 1

-2 . 4 2 2

-2 . 9 4 3

-3 . 4 6 3

-3 . 9 8 4

-4 . 5 0 5

0

-5 . 0 2 6

20

Errors = Targets - Outputs

Fig 9. Histogram error between the actual signal and the network output The Figure 10, shown the radial vibration of the compressor and associated network model, the Figure 11 present the mean squared error of the compressor radial vibration modelling, the Figure 12 shown the histogram error between the actual signal and the network output and the Fig 13 shown the variation of the high pressure shaft rotation speed NGP model of the examined turbine.

8

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Variation temporelle du vibration axial du Compresseur

40 35

V ib r a t io n u m

30 25 20 15 Sortie réelle Training Targets

10 5

E rre u r

0 10 0

-10

50

100

150

200

250

300

350

400

450

500

550

Time sec Fig 10. Radial vibration of the compressor and associated network model Best Validation Performance is 0.70946 at epoch 5

3

Mean Squared Error (mse)

10

Train Validation Test Best

2

10

1

10

0

10

-1

10

0

1

2

3

4

5

6

7

8

9

10

11

11 Epochs

Fig 11. The mean squared error of the compressor radial vibration modelling

9

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

300

Training Validation Test Zero Error

250

In s ta n c e s

200 150 100

7.869

7 .2 2

6.571

5.922

5.273

4.624

3.976

3.327

2.678

1 .3 8

2.029

0.7313

-0 . 5 6 6 4

0 .0 8 2 4 3

-1 . 2 1 5

-1 . 8 6 4

-2 . 5 1 3

-3 . 1 6 2

-4 . 4 6

0

-3 . 8 1 1

50

Errors = Targets - Outputs

Fig 12. Histogram error between the actual signal and the network output

R o t a t io n a l S p e e d %

120

100

80

Radial Vibration HP rotation Axial Vibration

A m p lit u d e u m

60

40

threshold

Alarme

20

0

0

100

200

300 Time (sec)

400

500

600

Fig 13. High pressure shaft rotation speed NGP model

10

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

In this system, in view of the fact that the turbine is subjected to a radial vibration which exceeds the tolerance threshold, this detection is made by the neural network which possesses the control capacity of any signal. In view of their frequent use in industry and their important role in the national economy, we develop a neural model with the ability to monitor the turbine safely using data from transmitters and sensors. This study can realize to make another integrated control system in the turbine. 5. Conclusion According to the results obtained, note that our network is playing an important role in the monitoring of large rotational range machines, as well as the neural network capable of generating information from the processed signals. The second interest is to integrate the artificial intelligence methods based on the neural network to the super- vision of industrial processes to facilitate maintenance interventions and to ensure the reliability of the turbine. The results obtained encouraging to integrate intelligent techniques in the monitoring and diagnosis of the gas turbine. References [1].

Ahmed Hafaifa, Ahmed Zohair Djeddi and Attia Daoudi, Fault detection and isolation in industrial control valve based on artificial neural networks diagnosis. Journal of Control Engineering and Applied Informatics CEAI, Vol.15, No.3 pp. 61-69, 2013.

[2].

Ahmed Hafaifa, Mouloud Guemana and Attia Daoudi, Fault detection and isolation in industrial systems based on spectral analysis diagnosis. Intelligent Control and Automation Journal, 2013, February 2013, Vol. 4, pp. 36-41.

[3].

Ahmed Hafaifa, Mouloud Guemana, and Attia Daoudi, Vibration supervision in gas turbine based on parity space approach to increasing efficiency. Journal of Vibration and Control, doi: 10.1177/1077546313499927.

[4].

Asgari, H., Chen, X.Q., and Sainudiin, R. (2012). Analysis of ANN-Based Modelling Approach for Industrial Applications. International Conference on Industrial Applications and Innovations (ICIAI 2012). Hong Kong. 5 pp. (Conference Contribution - Full conference paper).

[5].

Asgari, H., Chen, X.Q., and Sainudiin, R. (2013). Analysis of ANN-Based Modelling Approach for Industrial Applications. International Journal of Innovation, Management, and Technology (IJIMT). 4 (1), 165-169. DOI: 10.7763/IJIMT.2013.V4.383

[6].

Asgari, H., Venturini, M., Chen, X.Q., and Sainudiin, R. (2014). Modelling and Simulation of the Transient Behaviour of an Industrial Power Plant Gas Turbine. ASME Journal of Engineering for Gas Turbines and Power. 136(6), 061601. 10 pages. DOI: 10.1115/1.4026215

[7].

Asgari. H., Chen. X.Q., and Sainudiin, R. (2013). Modelling and Simulation Approaches for Gas Turbine System Optimization. In M.K. Habib and J.P. Davim (Eds.), Engineering Creative Design in Robotics and Mechatronics. Ch. 14, pp. 240-264, IGI Global. DOI: 10.4018/978-14666-4225-6.ch014.

11

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

[8].

Ben Rahmoune Mohamed, Ahmed Hafaifa and Guemana Mouloud, Étude la fiabilité d’une turbine à gaz pour l’amélioration de leur disponibilité. The 1stInternational Conference on Applied Automation and Industrial Diagnostics (ICAAID 2015), Djelfa on 29-30 March 2015, Algeria.

[9].

Ben Rahmoune Mohamed, Ahmed Hafaifa and Guemana Mouloud, Gas turbine supervision based on neural networks in degraded mode: Presence of vibrations. The 1stInternational Conference on Applied Automation and Industrial Diagnostics (ICAAID 2015), Djelfa on 29-30 March 2015, Algeria.

[10]. Ben Rahmoune Mohamed, Ahmed Hafaifa and Guemana Mouloud, Vibration modeling improves pipeline performance, costs. Oil & Gas Journal, Mars 2015, pp. 98-100. [11]. Ben Rahmoune Mohamed, Ahmed Hafaifa and Guemana Mouloud, Vibration monitoring in gas turbine speed using artificial neural networks. The 3rd International Conference on Information Processing and Electrical Engineering (ICIPEE’14), TEBESSA on 24-25 November 2014, Algeria. [12]. Chen Y.M., Lee M.L. (2002). Neural networks-based scheme for system failure detection and diagnosis. Mathematics and Computer sin Simulation, vol. 58, no. 2, pp.101-109. [13]. Chii-Shang Tsai, Chuei-Tin Chang. (1995). Dynamic process diagnosis via integrated neural networks. Computers & Chemical Engineering, vol.19, no. 1, pp. 747-752. [14]. Feyzullah Temurtas. (2009). A comparative study on thyroid disease diagnosis using neural networks. Expert Systems with Applications,vol.36, no. 1, pp. 944-949. [15]. Hamid Asgari (2014) Modelling, Simulation and Control of Gas Turbines Using Artificial Neural Networks. thesis Doctor of Philosophy. University of Canterbury Christchurch ,New Zealand [16]. Hamid Asgari, XiaoQi Chen, Mohammad B Menhaj, Raazesh Sainudiin, Artificial Neural Network Based System Identification for a Single-Shaft Gas Turbine, Journal of Engineering for Gas Turbines and Power, 2013, Vol 135, No 9, pp 092601. [17]. Hamid Asgari, XiaoQi Chen, Raazesh Sainudiin, Modelling and simulation of gas, International Journal of Modelling, Identification and Control, Vol,20, No 3, pp 253-270. [18]. Jaroslaw BEDNARZ, Tomas BARSZCZ, (2011) Tadeusz UHL. NARX MODEL IN ROTATING MACHINERY DIAGNOSTICS. MECHANICSANDCONTROL; Vol 30 No.2 2011. [19]. José Maria P. Junior and Guilherme A. Barreto (2007) Long-Term Time Series Prediction with the NARX Network: An Empirical Evaluation [20]. Joseph McGhee, Ian A. Henderson, Alistair Baird. (1997).Neural networks applied for the identification and fault diagnosis of process valves and actuators. Measurement, vol. 20, no. 4, pp. 267-275. [21]. Mohamed Ben Rahmoune, Ahmed Hafaifa and Mouloud Guemana, Elaboration of a faults tree in gas turbine using expert system based on neural network,

International Conference on

energy systems, Dec 23-25 2015 Istanbul Turkey [22]. Mohamed Ben Rahmoune, Ahmed Hafaifa and Mouloud Guemana, Neural network monitoring system used for the frequency vibration prediction in gas turbine. The 3rd International

12

The 2sd International Conference on Applied Automation and Industrial Diagnostics, Djelfa on 16-17 September 2017, Algeria

Conference on Control, Engineering & Information Technology CEIT’2015, on 25-27 May 2015 Tlemcen, Algeria. [23]. Neural Network Toolbox™ R2013a Mark Hudson Beale Martin

T. Hagan Howard B. Demuth.

www.mathworks.com/contact_TS.html [24]. Rahmoune MB, Hafaifa A, Guemana M. Fault diagnosis in gas turbine based on neural networks: Vibrations Speed Application. Book Chapter in Advances in Acoustics and Vibration, Volume 5 of the series Applied Condition Monitoring 2016: 1-11.

13

Suggest Documents