慢性傷口照護通常需要長時間且持續性的治療，而光生物調節療法（Photobiomodulation Therapy， PBMT）可藉由特定波長之光照調節細胞生理反應，具有應用於傷口照護之潛力。然而，傳統光療系統多採用固定波長或預先設定之光照條件，當治療需求改變時，仍需透過人工方式調整光源參數，增加光譜設計與控制之複雜度。 本研究旨在建立一套可依據目標光譜搜尋LED控制參數之智慧化光療系統。研究以3×3 RGB LED 陣列作為光源，透過拉丁超立方抽樣（Latin Hypercube Sampling， LHS）建立LED控制參數與光譜輸出之實測資料，並利用高斯過程回歸（Gaussian Process Regression， GPR）建立光學響應之近似模型，作為深度強化學習之模擬環境。進一步結合近端策略最佳化（Proximal Policy Optimization， PPO），使代理人能依據目標光譜搜尋適當之LED控制參數，並將最佳化結果應用於實際LED模組，再利用光譜儀進行實體量測驗證。 實驗結果顯示，PPO能於GPR所建立之近似環境中有效進行目標光譜之參數搜尋，並使光譜擬合誤差逐步降低。將逆推所得之控制參數應用於實際LED模組後，多組目標光譜皆能獲得與目標光譜相近之實測結果，顯示所建立之GPR輔助PPO架構具備由目標光譜逆推LED控制參數之可行性。 此外，本研究進一步以NIH3T3纖維母細胞進行初步光生物調節實驗，以觀察不同光照條件對細胞活性之影響。結果顯示，不同波長之光照對細胞活性呈現不同反應，其中綠光於本研究條件下呈現較高之細胞活性趨勢。由於本研究之細胞實驗屬初步驗證，相關結果主要用於觀察光照條件與細胞反應之關聯，並作為後續光療系統最佳化與生物效應研究之基礎。;Chronic wound care often requires prolonged and continuous treatment. Photobiomodulation therapy (PBMT) has the potential to support wound care by modulating cellular physiological responses through light irradiation at specific wavelengths. However， conventional phototherapy systems generally employ fixed wavelengths or predefined irradiation conditions， requiring manual adjustment of light-source parameters when treatment requirements change， thereby increasing the complexity of spectral design and control. This study aims to develop an intelligent phototherapy system capable of automatically searching for LED control parameters according to a target spectrum. A 3×3 RGB LED array was employed as the light source. Latin Hypercube Sampling (LHS) was used to establish experimental data relating LED control parameters to spectral outputs， and Gaussian Process Regression (GPR) was employed to construct an approximate optical-response model as a simulation environment for deep reinforcement learning. Proximal Policy Optimization (PPO) was further integrated to enable the agent to search for appropriate LED control parameters according to a target spectrum. The optimized parameters were subsequently applied to the actual LED module and experimentally validated using a spectrometer. The experimental results showed that PPO was capable of searching for LED control parameters within the approximate environment established by GPR， resulting in a gradual reduction in spectral fitting error. When the optimized control parameters were applied to the actual LED module， multiple target spectra achieved measured spectral profiles close to their corresponding targets. These results demonstrate the feasibility of the proposed GPR-assisted PPO framework for inversely determining LED control parameters from target spectra. In addition， preliminary photobiomodulation experiments were conducted using NIH3T3 fibroblasts to investigate the effects of different irradiation conditions on cellular activity. The results indicated that different wavelengths produced different cellular responses， with green light showing a relatively higher cellular activity trend under the experimental conditions of this study. Since the cell experiments were conducted as preliminary validation， the results are mainly used to investigate the relationship between irradiation conditions and cellular responses and to provide a basis for future optimization of the phototherapy system and investigation of its biological effects.
