1 / 25

Control with Neural Networks Case studies

Control with Neural Networks Case studies. Roland Pihlakas 08. Dec 2008. Proof of concept examples. Sunspot Activity: - Classical example Hydraulic actuator for a crane: NPC > APC The issue of fast sampling for validation Pneumatic position servomechanism: - Nonlinear

Download Presentation

Control with Neural Networks Case studies

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Control with Neural NetworksCase studies Roland Pihlakas 08. Dec 2008

  2. Proof of concept examples • Sunspot Activity: - Classical example • Hydraulic actuator for a crane: • NPC > APC • The issue of fast sampling for validation • Pneumatic position servomechanism: - Nonlinear • Level in a water tank: - Direct inverse control

  3. The Sunspot Benchmark • Optimal Brain Surgeon (OBS) assists in: • Selecting the network architecture • Selecting the regressors (inputs)

  4. The Sunspot Benchmark • Fully connected network has too many adjustable parameters for the training set • OBS algorithm: • Prune input-to-hidden weights • Retrain • Remove the least salient unit from the set of units with single input

  5. The Sunspot Benchmark • The training error gets larger during pruning • Test errors will decrease due to better generalization / less overfitting ... Until some point. • Note that FPEis not too informativein current example...

  6. The Sunspot Benchmark • Matlab: “it looks as if not much is gainedby pruning. The reason for this is, however, that thenetwork has been trained using regularization.” • The result of pruning:

  7. The Sunspot Benchmark • Additional notes • The result of pruning sessions can vary a great deal. -> One must run multiple pruning sessions, each one started with a different set of network weights. • The test sets were in some sense “actively” used for pruning. A distinction is made between this type of result and so-called “genuine predictions”, where test sets are strictly used for validation. -> gives more reliable estimate of generalization error.

  8. Hydraulic Actuator • Problem with fast sampling • NPC > APC

  9. Hydraulic Actuator • Measured values: • Valve opening (input) • Oil pressure (output) • Note the oscillatory response

  10. Hydraulic Actuator • First, linear model will be estimated • This is useful as a reference against more complicated models • ARX(3, 2, 1) evaluation:

  11. Hydraulic Actuator • NNARX(3, 2, 1) • 10 network architectures, with 1-10 hidden units, 5 networks of each • Legend:x – training erroro – test error • Spread of errorsis caused by local minima

  12. Hydraulic Actuator • For comparing model structures it is absolutely vital that the training must be continued until the weights are extremely near the minimum. Else overfitting will be less pronounced. • Network with 4 hidden units was best. It is recommended to choose then a slightly larger network.

  13. Hydraulic Actuator • Next, regularization is performed. • Legend:solid – training errordashed – test errordot-dashed – simulation on test • Note that again test set was used for training...

  14. Hydraulic Actuator • NNARX simulation is better than of the linear model:

  15. Pneumatic Servomechanism • Nonlinear and has poorly damped complex pole pair in the operating point.

  16. Pneumatic Servomechanism • The system has to be operated in closed-loop when conducting the experiment. • Manually tuned PI-controller is used for stabilization of the system during the experiment.

  17. Pneumatic Servomechanism • To cover entire operating range, a high-frequency signal is applied in some periods of the experiment.

  18. Pneumatic Servomechanism • Mimimum test error was achieved with 12 hidden units, which corresponds to 121 weights. • 121 weights is small number compared to training set => no need for regularization or pruning.

  19. Pneumatic Servomechanism • NPC control. Note how controller anticipates future changes in the set-point.

  20. Pneumatic Servomechanism • APC control. The response is similar to the one of NPC. • But this time there was no more steady-state error. • APC is simpler to implement and requires much less computations than NPC.

  21. Pneumatic Servomechanism • The poles of the extracted linear models:

  22. Control of Water Level • The water input inlet is controlled. • The water outlet is uncontrolled and open. The water output flow and thus also the system is nonlinear.

  23. Control of Water Level • When linearising such nonlinear system, one gets different parameters at different operating points. • This time direct inverse control was used instead. Such controllers are very simple to implement.

  24. Control of Water Level • Conducting the experiment: • One of linear models was used for conducting the experiment in closed-loop. • There should be both small and big changes in the output. • A random signal is added to the control inputs to ensure that the model will be able to produce reliable high-frequency outputs of small magnitude.

  25. Control of Water Level • Behavior of the model: • “Bang-bang” type control - maximum or minimum control input is applied until the desired set-point is achieved. • After that the control input attains its steady-state.

More Related