Seite - 75 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3.3. Black-boxModeling istakenintoaccountineachupdate. Thereforeallweightsareupdated Q times in one epoch, starting from the first date pair ( UNN(1),Y(1) ) until the lastdatapair ( UNN(Q),Y(Q) ) isaccountedfor. Online learning is a special incremental learning, with only one data pair is available at one time, therefore the cost function 3.74 can be rewritten inaonline formas JOL(k) = 1 2 ( YNN(k)−Yd(k) )T( YNN(k)−Yd(k) ) , (3.75) whereYNN(k) is theestimatedoutputfromtheNNatthecurrenttime k and Yd(k) is the desired (real measured) output of the system at time k. Online supervised learning is similar with the online system identificationintroducedinthegrey-boxmodelingpart,whichismore suitable todescribe thedynamicsof time-varyingsystems. NeuralNetworkModelingApproachesusedinHEPHAISTOS The implementation of neural networks in the temperature control system of HEPHAISTOS is straightforward compared with the grey- box modeling approaches. The aforementioned two learning strate- giesarebothusedtosolvedifferent tasks,as infigure 3.8. Neural Network Estimator Controller Trained by historical Data Set D Target temperature Yt Control input U Estimated temperature YNN (a)Batch learningapproach(offline) In the first approach (see figure 3.8a), batch learning is used to train theNNestimator. Thiswell-trainedNNisemployedasanapproxima- tionoftherealplanttotest theperformanceofdifferentsystemidenti- fication algorithms or control methods. For example, in the controller designpart,acontrollercouldbefirstlydesignedtocontrol thisNNin 75