本論文在光源光分佈的二維函數擬合曲線下藉由對稱旋轉所產生的理想三維光分佈函數曲面為基礎，透過疊合這些理想的三維光分佈函數曲面所形成的函數曲面，做為本研究用來模擬真實物理世界的三維光場的基底函數，透過調整基底函數的尺度母數與位置母數，與光學模擬軟體相比較的平方絕對誤差可以控制在 5%以下。 本論文以 UV-LED 曝光機為基礎研究對象，以三維光場分佈函數藉由準直 LED 模組單元的角度參數偏轉量的演算法，在曝光機最佳光源間距下，針對不同光源到目標平面距離 D 和光源直徑 S 的比率(距離直徑比，DSR)，DSR=10、12、14、16、18 的平面之照度均勻度皆大於 90%的情況下，透過光源排列的對稱性與隨機角度偏轉下進行照度均勻度的分辨分析，以局部光源的偏轉加上均勻度變化幅度，結合神經網路學習演算程式，從而分辨每顆光源的偏轉情形，此演算法可以得出每個準直 LED 模組單元角度變動對於目標平面照度的變動並且與實際的光源在偏移角度於一定的範圍下，兩者之間的誤差可以達到 5%以下。 再來以目標平面的照度變動做為神經網路的學習對像，並且提取訓練資料的特徵使得準確率能夠提升，最終成效可以達到 75%以上，以此訓練所分類的角度變動，可以做為調整組裝誤差的依據，藉以提昇每個準直 LED 模組單元的組裝定位精度，進而提昇整個 UV-LED 平行曝光機之光源模組的光學品質。 ;This thesis is based on the ideal three-dimensional light distribution function surface generated by symmetrical rotation under the two-dimensional function fitting curve of the light source light distribution, and the function surface formed by superimposing these ideal three dimensional light distribution function surfaces is used as the basis Study the basis function used to simulate the three-dimensional light field of the real physical world. By adjusting the scale mother number and position mother number of the basis function, the square absolute error compared with the optical simulation software can be controlled below 5%. This thesis takes the UV-LED exposure machine as the basic research object, and uses the three-dimensional light field distribution function to collimate the angle parameter deflection algorithm of the LED module unit. Under the optimal light source spacing of the exposure machine, for different DSR=10 When the uniformity of illuminance on the planes of, 12, 14, 16, and 18 are all greater than 90%, through the symmetry of the light source arrangement and random angle deflection, the illuminance uniformity is distinguished and analyzed, and the deflection of the local light source is added to the uniformity. The range of change is combined with the neural network learning algorithm to distinguish the deflection of each light source. This algorithm can obtain the angle change of each collimated LED module unit to the target plane illuminance change and offset from the actual light source. When the angle is ithin a certain range, the error between the two can reach 4% or less. Then use the illumination change of the target plane as the learning object of the neural network, and extract the characteristics of the training data to improve the accuracy, and the final effect can reach more than 75%. The angle change classified by the training can be used as The basis for adjusting the assembly error is to improve the assembly and positioning accuracy of each collimated LED module unit, thereby improving the optical quality of the light source module of the entire UV-LED parallel exposure machine.
