臺灣博碩士論文加值系統

　　本研究探討結構與轉移函數對類神經網路(ANNs)運用於輻射測溫之影響。分別針對六種不同鋼材(AISI 420、AISI 630、AISI A2、AISI A6、AISI H10、AISI H13)推測出在表面溫度為700 K、800 K與900 K時的放射率，並同時預測試件之溫度，以了解不同結構與轉移函數對溫度預測的影響，進而找出較佳的神經網路架構。
　　藉由類神經演算法，以紅外線光譜儀對實驗溫度下所量測得不同波長之試件輻射強度為輸入之修正資料，利用數值方法推導出每一層權重與偏壓值之修正式，經過不斷學習修正後以推測出表面之放射率與溫度。此外，透過與實驗真實放射率比較以檢測此測溫法之準確性，並利用本研究找出之較佳的神經網路架構與多光譜輻射測溫法兩者所預測之溫度比較，探討其於溫度預測上之優劣性。最後比較在真空環境與大氣環境下六種鋼材應用此輻射測溫法的預測溫度，以了解氧化效應對此測溫法之影響。
　　由研究結果可知：(1)當學習速率為0.0005時，在溫度預測及時間花費上有較好的表現；(2)對於不同鋼材，選取最大波長數可以得到較好的溫度預測結果；(3)就不同的轉移函數而言，非線性轉移函數(本研究取用對數S型轉移函數)比線性轉移函數有較好的預測結果；(4)熱作工具鋼在部分溫度下會因放射率變化過大導致誤差大於5K；(5)由神經網路推得之放射率越接近真實放射率時，越能準確地預測出試件表面溫度；(6)增加一層隱含層可提升神經網路的準確度，但因修正過程所需的時間過長，因此在可容許的誤差範圍下可選擇單一隱含層的神經網路；(7)無論在真空環境或大氣環境下，類神經網路運用於輻射測溫法都比多光譜輻射測溫法(MRT)還要準確，且誤差較不受不同加熱溫度影響；(8)氧化效應對於類神經演算法幾乎沒有影響，真空環境與大氣環境下之溫度預測誤差幾乎都小於1%。

The effect of structure and transfer function on Artificial Neural Networks (ANNs) used in radiation thermometry is discussed in this study. The spectral emissivity was inferred for six types of steels (AISI 420, AISI 630, AISI A2, AISI A6, AISI H10, AISI H13) at three temperatures (700 K, 800 K and 900 K). The sample temperatures were predicted simultaneously to comprehend the effect of structure and transfer function on temperature prediction and then the better structure of neural networks was found.
In ANNs algorithm, the sample’s spectral intensity measured by spectrometer was used as revised data to derive the modified relationships for the weights and biases by numerical method. Then, the surface emissivity and temperature could be inferred after incessant learning and modification. Besides, the accuracy of this thermometry was examined by the comparison between real and inferred emissivity. Temperature prediction by using the better structure of neural networks found in this study was compared to the multispectral radiation thermometry and the accuracy was discussed. Finally, the inferred temperatures for open-air and high-vacuum experiments were compared to find out the oxidation effect on this thermometry.
From results in the research, (1) there are better performances on temperature prediction and time cost when the learning rate is 0.0005; (2) for different steels, increasing number of wavelength will get better results; (3) for different transfer functions, the nonlinear transfer function (Log-Sigmoid transfer function used in this research) has better results than linear transfer function; (4) emissivity of some hot work tool steels changes dramatically and causes the temperature errors larger than 5K; (5) the closer the inferred emissivity to the real emissivity, the more accurate temperature prediction; (6) the network can be improved by adding one more hidden-layer. However, the time cost of two hidden-layers is much higher. Since the errors of one hidden-layer are rationally acceptable, thereby the neural networks with one hidden-layer is a better choice; (7) for both open-air and high-vacuum experiments, ANNs is more accurate than MRT and is less affected by different heating temperatures; (8) the oxidation effect on ANNs is not obvious and almost all the temperature errors for both open-air and high-vacuum experiments are smaller than 1%.

Contents
摘要....................................................................................................................... i
Abstract...................................................................................................................... iii
Acknowledgments...................................................................................................... v
Contents..................................................................................................................... vi
List of Tables.............................................................................................................. ix
List of Figures............................................................................................................ xi
Nomenclature............................................................................................................. xiv
Ch1. Introduction....................................................................................................... -1-
1-1 Research Motivation and Background............................................................. -1-
1-2 Introduction of Artificial Neural Networks..................................................... -2-
1-3 Literature Review............................................................................................. -7-
1-4 Research Goal.................................................................................................. -16-
1-5 Research Construction..................................................................................... -16-
Ch2. Experimental Foundation.................................................................................. -19-
2-1 Structure of the Open-air Experiment.............................................................. -19-
2-1.1 Spectrometer............................................................................................. -19-
2-1.2 Heating Assembly..................................................................................... -21-
2-1.3 Data Acquisition System........................................................................... -24-
2-2 Structure of the High-vacuum Experiment...................................................... -24-
2-2.1 High-vacuum System................................................................................ -24-
2-2.2 Cooling System......................................................................................... -28-
2-3 Samples Employed.......................................................................................... -28-
2-4 Procedure of the Experiment........................................................................... -30-
2-4.1 Procedure of Open-air Experiment........................................................... -30-
2-4.2 Procedure of High-vacuum Experiment................................................... -32-
2-5 Uncertainty of the Experiment......................................................................... -33-
Ch3. Fundamental Principles..................................................................................... -36-
3-1 Radiation Fundamental Principles................................................................... -36-
3-2 Radiation Intensity Measurement Theory........................................................ -40-
3-3 Principles of Artificial Neural Networks......................................................... -43-
3-4 Combination of Artificial Neural Networks and Radiation Thermometry...... -52-
3-5 Introduction of Multispectral Radiation Thermometry (MRT)....................... -53-
Ch4. Results and Discussion...................................................................................... -58-
4-1 Comparison of Results with Different Learning Rates……………………… -58-
4-2 Comparison of Results with Different Number of Wavelengths……………. -60-
4-3 Comparison of Results with Linear and Nonlinear Transfer Function……… -62-
4-4 Comparison of Results with Different Types of Steels……………………… -67-
4-5 Relationship between Inferred Emissivity and Temperature Error………….. -67-
4-6 Structure Effect on Artificial Neural Networks……………………………... -70-
4-7 Comparison of Results of ANNs and MRT for Open-air Experiment………. -74-
4-8 Comparison of Results of ANNs and MRT for High-vacuum Experiment…. -77-
4-9 Oxidation Effect on ANNs…………………………………………………... -80-
Ch5. Conclusions and Future Work........................................................................... -83-
5-1 Conclusions...................................................................................................... -83-
5-2 Future Work..................................................................................................... -85-
References.................................................................................................................. -86-
Appendices................................................................................................................. -89-
Vita............................................................................................................................. -98-
List of Tables
Table 1.1 The tested emissivity models (Chun-ling Yang et al. [8]). ………… -11-
Table 2.1 Material sizes of heating assembly. ………………………………….. -25-
Table 2.2 The constituents of steels used in this experiment (provided by Gloria Material Technology Corporation ). ………………………….. -29-
Table 2.3 The applications of samples used in this experiment (provided by Gloria Material Technology Corporation ). ………………………….. -31-
Table 2.4 The uncertainty analysis of the experiment. …………………………. -35-
Table 4.1 Absolute inferred temperature error and converged step number with different learning rates. ………………………..……...……………… -59-
Table 4.2 Absolute inferred temperature error with different number of wavelengths. …………………………………………………………. -61-
Table 4.3 Relative inferred temperature error with different transfer functions. . -64-
Table 4.4 Absolute inferred temperature error and converged step number with different number of hidden-layers. ……………………..……………. -71-
Table 4.5 Relative inferred temperature error of MRT (from previous research [17]) and ANNs for six oxidized steels at 700K, 800K and 900K. ….. -75-
Table 4.6 Relative inferred temperature error of MRT (from previous research [18]) and ANNs for six unoxidized steels at 700K, 800K and 900K. ………………………………………………………………… -78-
Table 4.7 Relative inferred temperature error of ANNs for oxidized and unoxidized steels at 700K, 800K and 900K. ………………………… -81-
List of Figures
Fig. 1.1 Structure of ANNs. ……………………………………………………. -3-
Fig. 1.2 Different types of transfer function. …………………………………. -5-
Fig. 1.3 Learning process of the back-propagation algorithm. ………………… -6-
Fig. 1.4 Flow chart of HI learning algorithm (Xiang-mei Li et al. [7]). ……….. -10-
Fig. 1.5 Emissivity data and simulation curve (Chun-ling Yang et al. [8]). ……. -12-
Fig. 1.6 Multi-reflection of a 60 degrees V-groove (Haugh [14]). ……………... -15-
Fig. 2.1 The diagrammatic sketch of open-air experiment. …………….……… -20-
Fig. 2.2 Field of view of the slit and the maximum effective measuring area. ... -22-
Fig. 2.3 The diagrammatic sketch of heating assembly. ……………………...... -23-
Fig. 2.4 The diagrammatic sketch of high-vacuum experiment. ……………..... -26-
Fig. 3.1 The spectral radiance distribution of real surface and blackbody at the same temperature in different wavelengths. ……………………..……. -37-
Fig. 3.2 The blackbody radiation intensity at different temperatures and wavelengths of Plank’s law, Wien’s law and Wien’s displacement law. -39-
Fig. 3.3 The composition of radiation intensities measured by spectrometer. …. -41-
Fig. 3.4 Structure of Artificial Neural Networks constructed in this research. … -44-
Fig. 3.5 The modified weights and biases between output layer and the second hidden layer. …………………………………………………………… -46-
Fig. 3.6 The modified weights and biases between two hidden layers. …………………………………………………………….……. -48-
Fig. 3.7 The modified weights and biases between the first hidden layer and input layer. ………………………………………………..…………… -50-
Fig. 3.8 Methodology of Artificial Neural Networks used in radiation thermometry. ……………………………….………………………….. -54-
Fig. 3.9 The schematic diagram of exact technique and least-squares technique. ….…………………………………………………………... -56-
Fig. 3.10 Methodology of Multispectral Radiation Thermometry. ……………… -57-
Fig. 4.1 Average relative inferred temperature error with different number of wavelengths for each steel. …..………………………………………... -63-
Fig. 4.2 Relative inferred temperature error for different steels with two different transfer functions. ……………………...…………………….. -66-
Fig. 4.3 Absolute inferred temperature error for different steels with log-sigmoid transfer function. ………..……………………………….. -68-
Fig. 4.4 The real emissivity and the inferred emissivity of AISI A6 at 700K, 800K and 900K. ……………………………………………………….. -69-
Fig. 4.5 Average of absolute inferred temperature error for different steels with different number of hidden-layer. …………………………….……….. -73-
Fig. 4.6 Relative inferred temperature error of ANNs and three better models of MRT (from previous research [17]) for six oxidized steels. ………………………………………………………………….. -76-
Fig. 4.7 Relative inferred temperature error of ANNs and three better models of MRT (from previous research [18]) for six unoxidized steels. ………………………………………………………………….. -79-
Fig. 4.8 The inferred emissivity and real emissivity of AISI A6 at 900K for oxidized and unoxidized steels. ……………………………………….. -82-

1.Herve ́ Abdi, “Neural Networks, The University of Texas at Dallas , 2003.
2.R.A. Jacobs, “Increased Rates of Convergence Through Learning Rate Adaption, Neural Networks, Vol. 1, pp. 295-307, 1988.
3.R.B. Boozarjomehry, W.Y. Svrcek, “Automatic design of neural network structures, Computers and Chemical Engineering 25, pp. 1075–1088, 2001.
4.X.G. SUN, J.M. DAI, Z.X. CHU, “Study of Multispectral Thermometry by Artificial Neural Network, Acta Metrologica Sinica, Vol. 22, No. 1, pp. 29-34, 2001.
5.X.G. SUN, J.M. DAI, “Multispectral thermometry based on neural network, Journal of Harbin Institute of Technology (New Series), Vol. 10, No.1, 2003.

6.C.L. Yang, J.M. Dai, and Y. Hu, “Optimum Identifications of Spectral Emissivity and Temperature for Multi-Wavelength Pyrometry, CHIN. PHYS. LETT. Vol. 20, No. 10, pp. 1685-1688., 2003.
7.X.M. LI, T.Y. ZHAO, “Study on Improving the Convergence of Genetic Neural Networks, Computer Applications, Vol. 25, No. 12, pp. 2789-2791, 2005.
8.C.L. Yang, Y. Yu, D.Y. Zhao, and G.L. Zhao, “Study on Modeling of Multispectral Emissivity and Optimization Algorithm, IEEE Transactions on Neural Networks, Vol. 17, No. 1, 2006.
9.C.Y. Wang, “A Study of Artificial Neural Networks Applied on Radiation Thermometry for Steel, 2011.
10.Touloukian, Y. S. and DeWitt, D.P., “Thermal Radiative Properties, Metallic Elements and Alloys, Thermophysical Properties of Matter, Vol. 7, Plenum Corp., New York, 1970.
11.Peacock, G. R., “A Review of Non-contact Process Temperature Measurement in Steel Manufacturing, Part of the SPIE Conference on Thermosense XXI, pp. 171-189, April 1999.
12.Pellerin, M. A., “Multispectral Radiation Thermometry for Industrial Application, PhD Thesis, Purdue University, school of Mechanical Engineering, Dec 1999.
13.Pellerin, M. A., “Multispectral Radiation Thermometry: Emissivity Compensation Algorithm, M.S. Thesis, Purdue University, school of Mechanical Engineering, 1990.
14.Haugh M. J., “Radiation Thermometry in the Aluminum Industry, Theory and Practice of Radiation Thermometry, D. P. DeWitt and G. D. Nutter, ed., John Wiley and Sons, New York, pp. 905-971, 1988.
15.Kanayama, K., Apparent Directional Emittance of V-Groove and Circular-Groove Rough Surface, Heat Transfer Japn. Res. 1, 11, 1972.
16.Operation Manual for ES200 Imaging Spectrometer, SPECTRALINE Document No. ES2001, Ver. 2.01, Spectraline Inc., 2003.
17.C.T. Lu, “Study of Emissivity Characteristics and Assessment of Multispectral Radiation Thermometry (MRT) Emissivity Models for Steel, 2007.
18.F.M. Chang, “Oxidation Effect on Steel Emissivity Behaviors and Examination of Multispectral Radiation Thermometry Emissivity Models, 2010.