Page - 78 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3. ModelingMicrowaveHeating GiventhedatasetD, thecompletelearningprocessofthestandardon- lineSGBPalgorithmisshownbelow. Thenotationsusedintheupdate ofweightsareshowninfigure 3.9anddefinedin table 3.3. zp = ∑ wp,j xj Np Output Layer YNN,p = fo (zp) Nj Hidden Layer xj = g (zj) i zj = ∑ wj,i uNN,i or zj = ∑ wj,i xi i j Figure3.9. Notations used to denote the input and the output nodes. 1. Initialization: initializeallweightsanddetermineactivationfunctionsfoandg. 2. Ateachtimekwithk≥2: arrange the inputandoutputof theNNestimatoras UNN(k) = [ Yr(k−1) U(k−1) ] , Yd(k) = Yr(k). (3.77) Forward calculation: calculate the input x and output z of all nodes in this neural network using the current weights and input vectorUNN(k), accordingtodefinitions infigure 3.9andtable 3.3. Backward error propagation: starting from the output layer, cal- culate the localgradientvector δo(k) = [ δo1(k), δ o 2(k) . .. δ o N(k) ] , (3.78) where theelement isgivenas [Hay98] δoi(k) = ( YNN,i(k)−Yd,i(k) ) f′o(zi(k)), 1≤ i≤N. (3.79) 78 ofdifferent