Page - 178 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

5. ExperimentalResults As mentioned in chapter 4, theQ(λ) based RLC is only tested in the flat-temperature period. The setup 5.21 is used in these experiments and the target temperature curve is the same as in figure 4.11. Be- fore the controller is tested in real experiments, it was also tested in a numberofsimulations. Thecontrol task inthesimulations isdifferent to the task 4.30 mentioned in chapter 4 because there is no way to simulate the whole temperature distribution. The same well-trained NN used in the NNC simulations 5.31 is also used here. Instead of the maximum and the minimum temperatures of the whole thermal picture, the maximum and the minimum temperatures from the five simulated temperatures are used as the controlled temperatures. Cor- respondingly, the indexes of these two temperatures are considered as the location states. For instance, T1 is the highest temperature is equivalent to the idea that the hot spot is the region R1 (illustrated in figure 4.13). Thissimulationcanbeconsideredas thesimplifiedversionof thereal control task. The purpose of this simulation is to test if all five tem- peraturescanbelimitedwithinacertainrangebyonlycontrollingthe maximum and the minimum temperatures. The corresponding simu- lationresultsareshowninfigure 5.36. Theentiresimulationlasts9000 s, but the controller takes around 3000 s (3000 control periods) to re- ducethetemperaturewindowfrommorethan10◦Ctoabout1◦C. This result reflects that theQ(λ) RLC method is effective to affect all other uncontrolled temperatures by only controlling the maximum and the minimumtemperatures. Based on this result, the hybrid control structure (figure 4.14) was tested in realexperiments. The total controlledarea is slightly smaller than the one shown in figure 4.13 to prevent any uncontrollable lo- cations in the corners or edges. The corresponding control results are showninfigure 5.37. In the first trial 5.37a, the raising-temperature period is controlled by using the linear MPC controller. In the flat-temperature period, the temperature window between the maximum and minimum temper- atures is not effectively reduced. Because theQ(λ) controller has to learn the correct state-action values during the controlling and it also has to take many exploration control actions. The exploration proba- bility usedintheQ(λ)controllerdecreases from0.9 inthebeginning 178