| 摘要: | 本研究開發一款14頻道燈箱的光譜生成器，向光譜生成器輸入光參數數值，可獲取符合輸入設定之光譜。透過基因演算法生成數據集，使用數據集訓練神經網路，評估預測表現良好的模型將作為光譜生成器。期望本研究可以幫助使用者在短時間內獲取特定目標的情境光源。
使用基因演算法生成相關色溫(Correlated Color Temperature, CCT)為3000 K、4000 K與5000 K且照度(Illuminance, Ev)為500 lux的數據集，其中包含CCT、色偏差值(Delta u-v, Duv)、平均演色性指數(General Color Rendering Index, Ra)與晝夜節律刺激值(Circadian Stimulus, CS)等光參數與頻道權重。研究過程中對基因演算法的設計進行多次修正，實驗一與二皆使用CCT作為適應函數，但實驗二只使用10頻道。實驗三將適應函數改為Duv。因實驗三的CCT偏差大，所以實驗四設計適應函數為uv座標。
計算染色體的光參數以及移除不可行解後，將數據集分成訓練集、驗證集與測試集，並對訓練集與驗證集進行過採樣。使用訓練集與驗證集訓練神經網路，透過測試集比較輸入光參數Xparam與預測結果的光參數Y’param的平均誤差，評估神經網路的效能。
研究結果顯示CS會受到藍色與黃色顏色機制(b-y)的影響，縮小Duv範圍後，4000 K的CS範圍會低於3000 K與5000 K的CS。數據集的光譜普遍具有高Ra。使用uv座標作為適應函數時，可以有效減少CCT預測誤差。當輸入數值落在數據集的光參數範圍內時，光譜生成器均可以產生相對應的光譜。但是光譜生成器的穩定性不足，未來仍需改善。;In this study, we develop a spectrum generator for a fourteen-channel illuminator. The aim is to derive a spectrum that corresponds to input optical parameters from the spectrum generator. After generating datasets via genetic algorithm, we use the datasets to train and evaluate a neural network. The model with the best performance will be chosen as the spectrum generator. We hope to assist users with obtaining specific lighting scenarios within a short time.
We apply genetic algorithm for collecting datasets with a fixed illuminance (Ev) at 500 lux and different correlated color temperature (CCT) at 3000, 4000, and 5000 K. The datasets include channel weights and optical parameters, such as CCT, delta u-v (Duv), general color rendering index (Ra) and circadian stimulus (CS). The design of the genetic algorithm has been revised several times during the research. In the first and second experiments, the fitness function is CCT. In contrast with the first experiment, we only use ten channel weights in the second experiment. The fitness function for the third experiment is changed to Duv. Due to the large deviation of the predicted CCT in the third experiment, we change the fitness function to uv coordinates in the fourth experiment.
After calculating optical parameters of chromosomes and remove infeasible data, the datasets are divided into training dataset, validation dataset and test dataset. Subsequently, the training dataset and validation dataset undergo oversampling. We use the training dataset and validation dataset to train neural networks and evaluate the trained neural networks by comparing the average error between the input optical parameters Xparam and the predicted optical parameters Y’param from the test dataset.
The results show that CS distributions are affected by blue versus yellow color mechanism (b-y). After reducing Duv range, the dataset with the CCT of 4000 K displays a CS distribution with lower values than that of 3000 and 5000 K. The spectra of the datasets generally have high Ra. When using uv coordinates as the fitness function, the deviation of the predicted CCT can be effectively decreased. The spectrum generator can successfully predict the channel weights of desired lighting scenario within the range of the datasets. As long as the input optical parameters are in the the range of optical parameters from the datasets, the spectrum generator is capable of producing corresponding spectra. However, the stability of the spectrum generator is not enough and requires improvement in the future. |
