1. Fault Prediction in Electrical Valves Using Temporal Kohonen Maps Luiz F. Gonçalves, Eduardo L. Schneider, Jefferson L. Bosa, Renato Ventura B. Henriques, Paulo M. Engel, Marcelo S. Lubaszewski 11 th LATW Punta del Leste - March 28-31 2010 UFRGS

2. 2 OUTLINE  Introduction  Maintenance scheme  Mathematical model  Signal processing  Temporal Kohonen maps  Experimental results  Conclusions

3. 3 INTRODUCTION

4. 4  The prediction of certain phenomena, processes or failures (or time series prediction) is particularly interesting and useful in many cases  It has been the subject of research in several areas:  Medicine (saving lives)  Meteorology (predicting the rain precipitation)  Engineering (increasing equipment reliability)  Economics (predicting changes in the stock market)  Main motivation: is the need to predict the future conditions and to understand the underlying phenomena and processes of the system under study Building models of the system using the knowledge and information that is available

5. 5 INTRODUCTION  Many methods for system prediction have been developed with very different approaches  Statistics:  Autoregressive  Autoregressive Moving Average  Neural networks:  Multi-Layer Perceptrons  Radial Basis Networks  Self-Organizing Maps (SOM) In the last years, models based on self-organizing maps have been raising much interest

6. 6 INTRODUCTION  Self-organizing map algorithms perform a vector quantization of data, leading to representatives in each portion of the space  The temporal models, built from SOM such as:  Temporal Kohonen maps (TKM)  Merge self-organizing maps (MSOM)  Recurrent self-organizing maps (RSOM) Use a leaky integrator memory to preserve the temporal context of the input signals

7. 7 INTRODUCTION  In this work, a proactive maintenance scheme is proposed for fault prediction in electrical valves Electrical valves Model Signals of torque and position Predicting the faults Oil distribution network Proactive maintenance scheme Wavelet packet transform & Temporal Kohonen maps

8. 8 MAINTENANCE SCHEME

9. 9 PROACTIVE MAINTENANCE  Recent advances in:  Electronics  Computing  Proactive ≠ corrective, preventive or predictive To automate and integrate proactive (also know as intelligent) maintenance tasks into embedded system That are based either on post-failure correction or on off-line periodic system checking Focuses on fault prediction and diagnosis based on component lifetimes and on system on-line monitoring

10. 10 MAINTENANCE SCHEME Mathematical model Wavelet packet transform Temporal Kohonen maps

11. 11 MATHEMATICAL MODEL  Electrical actuator Main components Forces

12. 12 MATHEMATICAL MODEL  Electrical actuator model Differential and algebraic equations Fault injection Position Torque

13. 13 SIGNAL PROCESSING  Wavelet packet transform  Preserves timing and spectral information  Suitable for the analysis of non-stationary signals  Capable of decomposing the signal in frequency bands  Energy (spectral density)  Torque  Position  The energy is used by the self-organizing maps Divided into N frequency bands The WPT runs in a PC station during the training phase During on-line testing, the WPT shall be part of the embedded system

14. 14 SELF-ORGANIZING MAPS  SOM or Kohonen maps (class of neural networks)  Unsupervised learning paradigm based on:  Competition (search the winner neuron)  Cooperation (identify direct neighbors)  Adaptation (update synaptic weights) The goal of a SOM is, after trained, mapping any input data from a R n space representation into R 2 lattice-like matrix Energy vector Synaptic weight vector

15. 15 TEMPORAL KOHONEN MAPS  The temporal Kohonen map (TKM)  Unsupervised approach for prediction derived from the SOM algorithm  Uses leaky integrators to maintain the activation history of each neuron  These neurons gradually loose their activity and are added to the outputs of the other normal competitive units  These integrators, and consequently the decay of activation, are modeled through the difference equation: Where: Euclidean Distance Temporal Activation

16. 16 TEMPORAL KOHONEN MAPS  The internal processing of SOM and TKM algorithms can be simplified and divided in three different steps: 1. Start up 2. Training 3. Recovery  Winner neuron:  SOM: the neuron with the shortest distance  TKM: the neuron with the highest activation Except for the determination of the winner neurons (recovery step), all other steps of the TKM are the same as in the SOM

17. 17  For fault prediction, in recovery step, the map is colored such that the distance between neighboring neurons can be seen  The distance is given by the difference between the synaptic weights of neighboring neurons  Closer neurons will appear clustered in the map and will be assigned the same color  Different colors will denote neurons under different operation conditions: normal, degraded or faulty  Once the winner neuron is computed for a particular input vector, E, the current system status can be identified in the colored map and, in deviated behavior, the degradation trajectory can be visualized in the map TEMPORAL KOHONEN MAPS

18. 18 TEMPORAL KOHONEN MAPS  In the TKM the system state can be visualized as a trajectory on the map and it is possible to follow the dynamics of the process This trajectory is described based on the winning neurons In a normal operation mode, the winners ought to follow a path inside the normal behavior region When a failure occurs, the winner will deviate from the normal region

19. 19 EXPERIMENTAL RESULTS

20. 20 EXPERIMENTAL RESULTS  Steps to generate the results: 1. Generate data (W) for normal (N), degraded (D) and faulty (F) behavior (obtained from the model) 2. Obtain the classification map (N, D and F data) using temporal Kohonen maps 3. Generate new N, D and F data (E) for three faults 4. Obtain the prediction map for each kind of faulty

21. 21 EXPERIMENTAL RESULTS  A lot of simulations is performed to obtain typical values of torque and opening position under N, D and F valve operation to train the fault prediction map  The fault simulation is needed to generate the F and D data (some parameters are gradually incremented) K M deviations simulate the elasticity loss of the valve spring along time C a deviations simulate an increase of friction between the valve stem and seal 100 operation cycles K R simulates the degradation of the internal valve worm gear, till it breaks

22. 22 MODEL RESULTS  Fault simulation Torque Position

23. 23 CLASSIFICATION RESULTS  Fault classification map of faults in K R, K M and C a

24. 24 CLASSIFICATION RESULTS  Fault classification map of faults in K R, K M and C a

25. 25 CLASSIFICATION RESULTS  Fault classification map of faults in K R, K M and C a

26. 26 CLASSIFICATION RESULTS  Fault classification map of faults in K R, K M and C a

27. 27 CLASSIFICATION RESULTS  Fault classification map of faults in K R, K M and C a During the on-line testing phase, a winner neuron computed for a measured input vector can be easily located in this map Each cluster is assigned a different color

28. 28 PREDICTION RESULTS  Fault prediction map of faults in K R

29. 29 PREDICTION RESULTS  Fault prediction map of faults in K M

30. 30 PREDICTION RESULTS  Fault prediction map of faults in C a

31. 31 PREDICTION RESULTS  It can be seen in these figures, three different paths (one for each simulated fault)  The trajectories started from neurons classified as normal, passed through neurons classified as degradation, and arrived to a neuron that represents the failure  It is noteworthy that in this work, the temporal Kohonen map is just used as a visualization tool

32. 32 CONCLUSION

33. 33 CONCLUSIONS  A proactive maintenance scheme is proposed for the prediction of faults in electrical valves, used for flow control in an oil distribution network  This is the first attempt to apply a proactive maintenance methodology to this sort of valves  A implementation of temporal Kohonen maps is proposed to solve the valve maintenance problem

34. 34 CONCLUSIONS  An system implements these maps for the prediction of faults in this valves  This technique can clearly be extended to any type of maintenance scheme including the on-line testing of heterogeneous chip with some kind of electro- mechanical systems (sensors or actuators) or other, for example

35. 35 CONCLUSIONS  The results obtained point out to a promising solution for the maintenance in electrical valves  Acknowledgements  CNPq  CAPES  Petrobrás

36. Thank you! luizfg@ece.ufrgs.br