Seite - 76 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3. ModelingMicrowaveHeating Neural Network Estimator Yr(k-1) ∑ - + Yr(k) e (k) U(k-1) YNN(k) (b)Online learningapproach Figure3.8. Neuralnetworkapproachesusedin thisdissertation. anofflinemode. Duringtheprocessof thecontrolling theNNestima- tor, thecontrollerbecomesmoreandmorefamiliarwith therealplant and finally it can be used to control the real plant. This method has been widely used for controller design such as in [Sch90], [WMS92], [LN95] and [GHLZ10]. In this case, an amount of historical exper- imental data can be used and the batch learning mode is preferred over incremental learning,dueto themorestable learningresultsand fasterconvergingspeed. But the limitationof thisapproachis that thecontroller trainedbythis NNestimatorisnotguaranteedtohavethesameperformanceinprac- tice as in the test. That is because in HEPHAISTOS, the real heating process is influencedbymanydifferentfactorsandit isnotpossibleto obtainexperimentaldata thatcancoveralldynamicsof theplant. De- spiteof this limitation, thisNNestimatorstillprovidesvaluable infor- mationandtheperformanceonitcanbeconsideredasthebenchmark tocompareandselectdifferentcontrolmethods. Thesecondapproach is touseaNNestimator foronlinesystemiden- tification (see figure 3.8b), which functions similarly with the online system identification algorithms introduced in the grey-box model- ing part. In this approach, the input of the NN estimatorUNN ( with the dimension (N+M)×1) contains the former temperature vector Yr(k−1)andthecontrol inputvectorU(k−1), suchas UNN(k) = [ Yr(k−1) U(k−1) ] . (3.76) 76